Cheng-Few Lee

John C. Lee
Editors

Handbook of
Financial
Econometrics
and Statistics

1 3Reference
Handbook of Financial Econometrics
and Statistics
Cheng-Few Lee • John C. Lee
Editors

Handbook of
Financial Econometrics
and Statistics

With 281 Figures and 490 Tables

Editors
Cheng-Few Lee
Department of Finance and Economics, Rutgers Business School
Rutgers, The State University of New Jersey
Piscataway, NJ, USA
and
Graduate Institute of Finance
National Chiao Tung University
Hsinchu, Taiwan

John C. Lee
Center for PBBEF Research
North Brunswick, NJ, USA

ISBN 978-1-4614-7749-5 ISBN 978-1-4614-7750-1 (eBook)

ISBN 978-1-4614-7751-8 (print and electronic bundle)
DOI 10.1007/978-1-4614-7750-1
Springer New York Heidelberg Dordrecht London

Library of Congress Control Number: 2014940762

# Springer Science+Business Media New York 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts
in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being
entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication
of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from
Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center.
Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Financial econometrics and statistics have become very important tools for empirical
research in both finance and accounting. Econometric methods are important tools for
doing asset pricing, corporate finance, options and futures, and conducting financial
accounting research. Important econometric methods used in this research include:
single equation multiple regression, simultaneous regression, panel data analysis,
time series analysis, spectral analysis, non-parametric analysis, semi-parametric
analysis, GMM analysis, and other methods.
Portfolio theory and management research have used different statistical distri-
butions, such as normal distribution, stable distribution, and log normal distribu-
tion. Options and futures research have used binomial distribution, log normal
distribution, non-central chi square distribution, Poission distribution, and others.
Auditing research has used sampling survey techniques to determine the sampling
error and non-sampling error for auditing.
Based upon our years of experience working in the industry, teaching classes,
conducting research, writing textbooks, and editing journals on the subject of
financial econometrics and statistics, this handbook will review, discuss, and
integrate theoretical, methodological, and practical issues of financial econometrics
and statistics. There are 99 chapters in this handbook. Chapter 1 presents an
introduction of financial econometrics and statistics and shows how readers can
use this handbook. The following chapters, which have been contributed by
accredited authors, can be classified by the following 14 topics.
i. Financial Accounting (Chapters 2, 9, 10, 61, 97)
ii. Mutual Funds (Chapters 3, 24, 25, 68, 88)
iii. Microstructure (Chapters 4, 44, 96, 99)
iv. Corporate Finance (Chapters 5, 21, 30, 38, 42, 46, 60, 63, 75, 79, 95)
v. Asset Pricing (Chapters 6, 15, 22, 28, 34, 36, 39, 45, 47, 50, 81, 85, 87, 93)
vi. Options (Chapters 7, 32, 37, 55, 65, 84, 86, 90, 98)
vii. Portfolio Analysis (Chapters 8, 26, 35, 53, 67, 73, 80, 83)
viii. Risk Management (Chapters 11, 13, 16, 17, 23, 27, 41, 51, 54, 72, 91, 92)
ix. International Finance (Chapters 12, 40, 43, 59, 69)
x. Event Study (Chapters 14)
xi. Methodology (Chapters 18, 19, 20, 29, 31, 33, 49, 52, 56, 57, 58, 62, 74, 76,
77, 78, 82, 89)

v
vi Preface

xii. Banking Management (Chapters 64)

xiii. Pension Funds (Chapters 66)
xiv. Futures and Index Futures (Chapters 48, 70, 71, 94)
In addition to this classification, based upon the keywords of chapter 2-99, we
classify the information into a) finance and accounting topics and b) methodology
topics. This information can be found in chapter 1 of this handbook.
In the preparation of this handbook, first, we would like to thank the member of
advisory board and contributors of this handbook. In addition, we would like to
make note that we appreciate the extensive help from the Editor Mr. Brian Foster,
our research assistants Tzu Tai, Lianne Ng, and our secretary Ms. Miranda Mei-Lan
Luo. Finally, we would like to thank the financial support from the Wintek
Corporation and APEX International Financial Engineering Res. & Tech. Co. Ltd.
that allowed us to write the edition of this book.
There are undoubtedly some errors in the finished product, both typo-graphical
and conceptual. I would like to invite readers to send suggestions, comments,
criticisms, and corrections to the author Professor Cheng F. Lee at the Department
of Finance and Economics, Rutgers University at the email address lee@business.
rutgers.edu.

December 2012 Cheng-Few Lee

John C. Lee
Advisory Board

Ivan Brick Rutgers, The State University of New Jersey, USA

Stephen Brown New York University, USA
Charles Q. Cao Penn State University, USA
Chun-Yen Chang National Chiao Tung University, Taiwan
Wayne Ferson Boston College, USA
Lawrence R. Glosten Columbia University, USA
Martin J. Gruber New York University, USA
Hyley Huang Wintek Corporation, Taiwan
Richard E. Kihlstrom University of Pennsylvania, USA
E. H. Kim University of Michigan, USA
Robert McDonald Northwestern University, USA
Ehud I. Ronn The University of Texas at Austin, USA

vii
About the Editors

Cheng-Few Lee is a Distinguished Professor of Finance at Rutgers Business

School, Rutgers University and was chairperson of the Department of Finance
from 1988–1995. He has also served on the faculty of the University of Illinois
(IBE Professor of Finance) and the University of Georgia. He has maintained
academic and consulting ties in Taiwan, Hong Kong, China and the United States
for the past four decades. He has been a consultant to many prominent groups
including, the American Insurance Group, the World Bank, the United Nations,
The Marmon Group Inc., Wintek Corporation, and Polaris Financial Group.
Professor Lee founded the Review of Quantitative Finance and Accounting
(RQFA) in 1990 and the Review of Pacific Basin Financial Markets and Policies
(RPBFMP) in 1998, and serves as managing editor for both journals. He was also
a co-editor of the Financial Review (1985–1991) and the Quarterly Review of
Economics and Finance (1987–1989). In the past 39 years, Dr. Lee has written
numerous textbooks ranging in subject matters from financial management to
corporate finance, security analysis and portfolio management to financial analysis,
planning and forecasting, and business statistics. In addition, he edited two popular
books, Encyclopedia of Finance (with Alice C. Lee) and Handbook of Quantitative
Finance and Risk Management (with Alice C. Lee and John Lee). Dr. Lee has also
published more than 200 articles in more than 20 different journals in finance,
accounting, economics, statistics, and management. Professor Lee was ranked the
most published finance professor worldwide during the period 1953–2008.
Professor Lee was the intellectual force behind the creation of the new Masters
of Quantitative Finance program at Rutgers University. This program began in
2001 and has been ranked as one of the top ten quantitative finance programs in
the United States. These top ten programs are located at Carnegie Mellon Univer-
sity, Columbia University, Cornell University, New York University, Princeton
University, Rutgers University, Stanford University, University of California at
Berkley, University of Chicago, and University of Michigan.

John C. Lee is a Microsoft Certified Professional in Microsoft Visual Basic and

Microsoft Excel VBA. He has a Bachelor and Masters degree in accounting from
the University of Illinois at Urbana-Champaign.
John has worked over 20 years in both the business and technical fields as an
accountant, auditor, systems analyst and as a business software developer. He is the

ix
x About the Editors

author of the book on how to use MINITAB and Microsoft Excel to do statistical
analysis which is a companion text to Statistics of Business and Financial
Economics, 2nd and 3rd, of which he is one of the co-authors. In addition, he has
also coauthored the textbooks Financial Analysis, Planning and Forecasting,
2ed (with Cheng F. Lee and Alice C. Lee), and Security Analysis, Portfolio
Management, and Financial Derivatives (with Cheng F. Lee, Joseph Finnerty,
Alice C. Lee, and Donald Wort). John has been a Senior Technology Officer at
the Chase Manhattan Bank and Assistant Vice President at Merrill Lynch.
Currently, he is the Director of the Center for PBBEF Research.
Contents

Volume 1

1 Introduction to Financial Econometrics and Statistics . . . . . . . . . 1

Cheng-Few Lee and John C. Lee
2 Experience, Information Asymmetry, and Rational
Forecast Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
April Knill, Kristina L. Minnick, and Ali Nejadmalayeri
3 An Appraisal of Modeling Dimensions for Performance
Appraisal of Global Mutual Funds . . . . . . . . . . . . . . . . . . . . . . . . 101
G. V. Satya Sekhar
4 Simulation as a Research Tool for Market Architects . . . . . . . . . . 121
Robert A. Schwartz and Bruce W. Weber
5 Motivations for Issuing Putable Debt: An Empirical
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Ivan E. Brick, Oded Palmon, and Dilip K. Patro
6 Multi-Risk Premia Model of US Bank Returns: An Integration
of CAPM and APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Suresh Srivastava and Ken Hung
7 Nonparametric Bounds for European Option Prices .......... 207
Hsuan-Chu Lin, Ren-Raw Chen, and Oded Palmon
8 Can Time-Varying Copulas Improve the Mean-Variance
Portfolio? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Chin-Wen Huang, Chun-Pin Hsu, and Wan-Jiun Paul Chiou
9 Determinations of Corporate Earnings Forecast Accuracy:
Taiwan Market Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Ken Hung and Kuo-Hao Lee
10 Market-Based Accounting Research (MBAR) Models: A Test of
ARIMAX Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Anastasia Maggina

xi
xii Contents

11 An Assessment of Copula Functions Approach in Conjunction

with Factor Model in Portfolio Credit Risk Management . . . . . . . 299
Lie-Jane Kao, Po-Cheng Wu, and Cheng-Few Lee
12 Assessing Importance of Time-Series Versus Cross-Sectional
Changes in Panel Data: A Study of International Variations in
Ex-Ante Equity Premia and Financial Architecture . . . . . . . . . . . 317
Raj Aggarwal and John W. Goodell
13 Does Banking Capital Reduce Risk? An Application of
Stochastic Frontier Analysis and GMM Approach . . . . . . . . . . . . 349
Wan-Jiun Paul Chiou and Robert L. Porter
14 Evaluating Long-Horizon Event Study Methodology . . . . . . . . . . 383
James S. Ang and Shaojun Zhang
15 The Effect of Unexpected Volatility Shocks on Intertemporal
Risk-Return Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
Kiseok Nam, Joshua Krausz, and Augustine C. Arize
16 Combinatorial Methods for Constructing Credit Risk
Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Alexander Kogan and Miguel A. Lejeune
17 Dynamic Interactions Between Institutional Investors and
the Taiwan Stock Returns: One-Regime and Threshold
VAR Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
Bwo-Nung Huang, Ken Hung, Chien-Hui Lee, and Chin W. Yang
18 Methods of Denoising Financial Data ...................... 519
Thomas Meinl and Edward W. Sun
19 Analysis of Financial Time Series Using Wavelet Methods . . . . . . 539
Philippe Masset
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss
Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
Anna Chernobai, Svetlozar T. Rachev, and Frank J. Fabozzi
21 Effect of Merger on the Credit Rating and Performance of
Taiwan Security Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
Suresh Srivastava and Ken Hung
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese
Treasury Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
Rong Chen, Hai Lin, and Qianni Yuan
23 Factor Copula for Defaultable Basket Credit Derivatives . . . . . . . 639
Po-Cheng Wu, Lie-Jane Kao, and Cheng-Few Lee
Contents xiii

24 Panel Data Analysis and Bootstrapping: Application to China

Mutual Funds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
Win Lin Chou, Shou Zhong Ng, and Yating Yang

25 Market Segmentation and Pricing of Closed-End

Country Funds: An Empirical Analysis . . . . . . . . . . . . . . . . . . . . . 669
Dilip K. Patro

Volume 2

26 A Comparison of Portfolios Using Different Risk

Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
Jing Rung Yu, Yu Chuan Hsu, and Si Rou Lim
27 Using Alternative Models and a Combining Technique in Credit
Rating Forecasting: An Empirical Study . . . . . . . . . . . . . . . . . . . . 729
Cheng-Few Lee, Kehluh Wang, Yating Yang, and
Chan-Chien Lien
28 Can We Use the CAPM as an Investment Strategy?:
An Intuitive CAPM and Efficiency Test . . . . . . . . . . . . . . . . . . . . 751
Fernando Gómez-Bezares, Luis Ferruz, and Maria Vargas
29 Group Decision-Making Tools for Managerial Accounting and
Finance Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
Wikil Kwak, Yong Shi, Cheng-Few Lee, and Heeseok Lee
30 Statistics Methods Applied in Employee Stock Options ........ 841
Li-jiun Chen and Cheng-der Fuh
31 Structural Change and Monitoring Tests ................... 873
Cindy Shin-Huei Wang and Yi Meng Xie
32 Consequences for Option Pricing of a Long Memory
in Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903
Stephen J. Taylor
33 Seasonal Aspects of Australian Electricity Market . . . . . . . . . . . . 935
Vikash Ramiah, Stuart Thomas, Richard Heaney, and
Heather Mitchell
34 Pricing Commercial Timberland Returns in the
United States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
Bin Mei and Michael L. Clutter
35 Optimal Orthogonal Portfolios with Conditioning
Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 977
Wayne E. Ferson and Andrew F. Siegel
xiv Contents

36 Multifactor, Multi-indicator Approach to Asset Pricing:

Method and Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . 1003
Cheng-Few Lee, K. C. John Wei, and Hong-Yi Chen
37 Binomial OPM, Black–Scholes OPM, and Their Relationship:
Decision Tree and Microsoft Excel Approach . . . . . . . . . . . . . . . . 1025
John C. Lee
38 Dividend Payments and Share Repurchases of US Firms:
An Econometric Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Alok Bhargava
39 Term Structure Modeling and Forecasting Using the
Nelson-Siegel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1093
Jian Hua
40 The Intertemporal Relation Between Expected Return and
Risk on Currency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105
Turan G. Bali and Kamil Yilmaz
41 Quantile Regression and Value at Risk . . . . . . . . . . . . . . . . . . . . . 1143
Zhijie Xiao, Hongtao Guo, and Miranda S. Lam
42 Earnings Quality and Board Structure: Evidence from
South East Asia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1169
Kin-Wai Lee
43 Rationality and Heterogeneity of Survey Forecasts of the
Yen-Dollar Exchange Rate: A Reexamination . . . . . . . . . . . . . . . . 1195
Richard Cohen, Carl S. Bonham, and Shigeyuki Abe
44 Stochastic Volatility Structures and Intraday Asset Price
Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1249
Gerard L. Gannon
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock
Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1277
Ken Hung and Suresh Srivastava
46 Alternative Methods for Estimating Firm’s Growth Rate . . . . . . 1293
Ivan E. Brick, Hong-Yi Chen, and Cheng-Few Lee
47 Econometric Measures of Liquidity . . . . . . . . . . . . . . . . . . . . . . . . 1311
Jieun Lee
48 A Quasi-Maximum Likelihood Estimation Strategy for
Value-at-Risk Forecasting: Application to Equity Index
Futures Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1325
Oscar Carchano, Young Shin (Aaron) Kim, Edward W. Sun,
Svetlozar T. Rachev, and Frank J. Fabozzi
Contents xv

49 Computer Technology for Financial Service . . . . . . . . . . . . . . . . . 1341

Fang-Pang Lin, Cheng-Few Lee, and Huimin Chung

50 Long-Run Stock Return and the Statistical Inference . . . . . . . . . . 1381

Yanzhi Wang

Volume 3

51 Value-at-Risk Estimation via a Semi-parametric Approach:

Evidence from the Stock Markets . . . . . . . . . . . . . . . . . . . . . . . . . 1399
Cheng-Few Lee and Jung-Bin Su

52 Modeling Multiple Asset Returns by a Time-Varying t

Copula Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1431
Long Kang

53 Internet Bubble Examination with Mean-Variance Ratio . . . . . . . 1451

Zhidong D. Bai, Yongchang C. Hui, and Wing-Keung Wong

54 Quantile Regression in Risk Calibration . . . . . . . . . . . . . . . . . . . . 1467

Shih-Kang Chao, Wolfgang Karl Härdle, and Weining Wang

55 Strike Prices of Options for Overconfident Executives . . . . . . . . . 1491

Oded Palmon and Itzhak Venezia

56 Density and Conditional Distribution-Based Specification

Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1509
Diep Duong and Norman R. Swanson

57 Assessing the Performance of Estimators Dealing with

Measurement Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1563
Heitor Almeida, Murillo Campello, and Antonio F. Galvao

58 Realized Distributions of Dynamic Conditional Correlation

and Volatility Thresholds in the Crude Oil, Gold, and
Dollar/Pound Currency Markets . . . . . . . . . . . . . . . . . . . . . . . . . . 1619
Tung-Li Shih, Hai-Chin Yu, Der-Tzon Hsieh, and Chia-Ju Lee

59 Pre-IT Policy, Post-IT Policy, and the Real Sphere in Turkey . . . 1647
Ahmed Hachicha and Cheng-Few Lee

60 Determination of Capital Structure: A LISREL Model

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1669
Cheng-Few Lee and Tzu Tai

61 Evidence on Earning Management by Integrated Oil and Gas

Companies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1685
Raafat R. Roubi, Hemantha Herath, and John S. Jahera Jr.
xvi Contents

62 A Comparative Study of Two Models SV with MCMC

Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1697
Ahmed Hachicha, Fatma Hachicha, and Afif Masmoudi
63 Internal Control Material Weakness, Analysts Accuracy and
Bias, and Brokerage Reputation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1719
Li Xu and Alex P. Tang
64 What Increases Banks Vulnerability to Financial Crisis:
Short-Term Financing or Illiquid Assets? . . . . . . . . . . . . . . . . . . . 1753
Gang Nathan Dong and Yuna Heo
65 Accurate Formulas for Evaluating Barrier Options with
Dividends Payout and the Application in Credit
Risk Valuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1771
Tian-Shyr Dai and Chun-Yuan Chiu
66 Pension Funds: Financial Econometrics on the Herding
Phenomenon in Spain and the United Kingdom . . . . . . . . . . . . . . 1801
Mercedes Alda Garcı́a and Luis Ferruz
67 Estimating the Correlation of Asset Returns: A Quantile
Dependence Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1829
Nicholas Sim
68 Multi-criteria Decision Making for Evaluating Mutual Funds
Investment Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1857
Shin Yun Wang and Cheng-Few Lee
69 Econometric Analysis of Currency Carry Trade . . . . . . . . . . . . . . 1877
Yu-Jen Wang, Huimin Chung, and Bruce Mizrach
70 Evaluating the Effectiveness of Futures Hedging . . . . . . . . . . . . . 1891
Donald Lien, Geul Lee, Li Yang, and Chunyang Zhou
71 Analytical Bounds for Treasury Bond Futures Prices . . . . . . . . . . 1909
Ren-Raw Chen and Shih-Kuo Yeh
72 Rating Dynamics of Fallen Angels and Their Speculative
Grade-Rated Peers: Static vs. Dynamic Approach . . . . . . . . . . . . 1945
Huong Dang
73 Creation and Control of Bubbles: Managers Compensation
Schemes, Risk Aversion, and Wealth and Short Sale
Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1983
James S. Ang, Dean Diavatopoulos, and Thomas V. Schwarz
74 Range Volatility: A Review of Models and
Empirical Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2029
Ray Yeutien Chou, Hengchih Chou, and Nathan Liu
Contents xvii

75 Business Models: Applications to Capital Budgeting, Equity

Value, and Return Attribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2051
Thomas S. Y. Ho and Sang Bin Lee

Volume 4

76 VAR Models: Estimation, Inferences, and Applications . . . . . . . . 2077

Yangru Wu and Xing Zhou

77 Model Selection for High-Dimensional Problems . . . . . . . . . . . . . 2093

Jing-Zhi Huang, Zhan Shi, and Wei Zhong

78 Hedonic Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2119

Ben J. Sopranzetti

79 Optimal Payout Ratio Under Uncertainty and the Flexibility

Hypothesis: Theory and Empirical Evidence . . . . . . . . . . . . . . . . . 2135
Cheng-Few Lee, Manak C. Gupta, Hong-Yi Chen, and Alice C. Lee

80 Modeling Asset Returns with Skewness, Kurtosis,

and Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2177
Thomas C. Chiang and Jiandong Li

81 Does Revenue Momentum Drive or Ride Earnings or

Price Momentum? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2217
Hong-Yi Chen, Sheng-Syan Chen, Chin-Wen Hsin, and
Cheng-Few Lee

82 A VG-NGARCH Model for Impacts of Extreme Events on

Stock Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2263
Lie-Jane Kao, Li-Shya Chen, and Cheng-Few Lee

83 Risk-Averse Portfolio Optimization via Stochastic Dominance

Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2281
Darinka Dentcheva and Andrzej Ruszczynski

84 Implementation Problems and Solutions in Stochastic Volatility

Models of the Heston Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2303
Jia-Hau Guo and Mao-Wei Hung

85 Stochastic Change-Point Models of Asset Returns and Their

Volatilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2317
Tze Leung Lai and Haipeng Xing

86 Unspanned Stochastic Volatilities and Interest Rate Derivatives

Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2337
Feng Zhao
xviii Contents

87 Alternative Equity Valuation Models . . . . . . . . . . . . . . . . . . . . . . . 2401

Hong-Yi Chen, Cheng-Few Lee, and Wei K. Shih
88 Time Series Models to Predict the Net Asset Value (NAV) of
an Asset Allocation Mutual Fund VWELX . . . . . . . . . . . . . . . . . . 2445
Kenneth D. Lawrence, Gary Kleinman, and Sheila M. Lawrence
89 Discriminant Analysis and Factor Analysis: Theory
and Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2461
Lie-Jane Kao, Cheng-Few Lee, and Tzu Tai
90 Implied Volatility: Theory and Empirical Method . . . . . . . . . . . . 2477
Cheng-Few Lee and Tzu Tai
91 Measuring Credit Risk in a Factor Copula Model . . . . . . . . . . . . 2495
Jow-Ran Chang and An-Chi Chen
92 Instantaneous Volatility Estimation by Nonparametric
Fourier Transform Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2519
Chuan-Hsiang Han
93 A Dynamic CAPM with Supply Effect Theory and
Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2535
Cheng-Few Lee, Chiung-Min Tsai, and Alice C. Lee
94 A Generalized Model for Optimum Futures Hedge Ratio . . . . . . 2561
Cheng-Few Lee, Jang-Yi Lee, Kehluh Wang, and Yuan-Chung Sheu
95 Instrumental Variables Approach to Correct for Endogeneity
in Finance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2577
Chia-Jane Wang
96 Application of Poisson Mixtures in the Estimation of Probability
of Informed Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2601
Emily Lin and Cheng-Few Lee
97 CEO Stock Options and Analysts’ Forecast Accuracy
and Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2621
Kiridaran Kanagaretnam, Gerald J. Lobo, and Robert Mathieu
98 Option Pricing and Hedging Performance Under Stochastic
Volatility and Stochastic Interest Rates . . . . . . . . . . . . . . . . . . . . . 2653
Charles Cao, Gurdip S. Bakshi, and Zhiwu Chen
99 The Le Ch^atelier Principle of the Capital Market
Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2701
Chin W. Yang, Ken Hung, and Matthew D. Brigida
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2709
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2749
Reference Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2769
Contributors

Shigeyuki Abe Faculty of Policy Studies, Doshisha University, Kyoto, Japan

Raj Aggarwal University of Akron, Akron, OH, USA
Mercedes Alda Garcı́a Facultad de Economı́a y Empresa, Departamento de
Contabilidad y Finanzas, Universidad de Zaragoza, Zaragoza, Spain
Heitor Almeida University of Illinois at Urbana-Champaign, Champaign, IL, USA
James S. Ang Department of Finance, College of Business, Florida State
University, Tallahassee, FL, USA
Augustine C. Arize Texas A & M University-Commerce, Commerce, TX, USA
Zhidong D. Bai KLAS MOE & School of Mathematics and Statistics, Northeast
Normal University, Changchun, China
Department of Statistics and Applied Probability, National University of Singapore,
Singapore, Singapore
Gurdip S. Bakshi Department of Finance, College of Business, University of
Maryland, College Park, MD, USA
Turan G. Bali McDonough School of Business, Georgetown University,
Washington, DC, USA
Alok Bhargava School of Public Policy, University of Maryland, College Park,
MD, USA
Carl S. Bonham College of Business and Public Policy, University of Alaska
Anchorage, Anchorage, AK, USA
Ivan E. Brick Department of Finance and Economics, Rutgers, The State
University of New Jersey, Newark/New Brunswick, NJ, USA
Matthew D. Brigida Department of Finance, Clarion University of Pennsylvania,
Clarion, PA, USA
Murillo Campello Cornell University, Ithaca, NY, USA

xix
xx Contributors

Charles Cao Department of Finance, Smeal College of Business, Penn State

University, University Park, PA, USA
Oscar Carchano Department of Financial Economics, University of Valencia,
Valencia, Spain
Jow-Ran Chang National Tsing Hua University, Hsinchu City, Taiwan
Shih-Kang Chao Ladislaus von Bortkiewicz Chair of Statistics, C.A.S.E. – Center
for Applied Statistics and Economics, Humboldt-Universität zu Berlin, Berlin,
Berlin, Germany
An-Chi Chen KGI Securities Co. Ltd., Taipei, Taiwan
Hong-Yi Chen Department of Finance, National Central University, Taoyuan,
Taiwan
Li-jiun Chen Department of Finance, Feng Chia University, Taichung City,
Taiwan
Li-Shya Chen Department of Statistics, National Cheng-Chi University, Taipei
City, Taiwan
Ren-Raw Chen Graduate School of Business Administration, Fordham
University, New York, NY, USA
Rong Chen Department of Finance, Xiamen University, Xiamen, China
Sheng-Syan Chen National Central University, Zhongli City, Taiwan
Zhiwu Chen School of Management, Yale University, New Haven, USA
Anna Chernobai Department of Finance, M.J. Whitman School of Management,
Syracuse University, Syracuse, NY, USA
Thomas C. Chiang Department of Finance, Drexel University, Philadelphia, PA,
USA
Wan-Jiun Paul Chiou Department of Finance and Law College of Business
Administration, Central Michigan University, Mount Pleasant, MI, USA
Chun-Yuan Chiu National Chiao–Tung University, Taiwan, Republic of China
Institute of Information Management, National Chiao Tung University, Taiwan,
Republic of China
Hengchih Chou Department of Shipping and Transportation Management,
National Taiwan Ocean University, Keelung, Taiwan
Ray Yeutien Chou Institute of Economics, Academia Sinica and National Chiao
Tung University, Taipei, Taiwan
Win Lin Chou Department of Economics and Finance, City University of Hong
Kong, Hong Kong, China
Contributors xxi

Huimin Chung Graduate Institute of Finance, National Chiao Tung University,

Hsinchu, Taiwan
Michael L. Clutter Warnell School of Forestry and Natural Resources, University
of Georgia, Athens, GA, USA
Richard Cohen University of Hawaii Economic Research Organization and
Economics, University of Hawaii at Manoa, Honolulu, HI, USA
Tian-Shyr Dai National Chiao-Tung University, Taiwan, Republic of China
Huong Dang University of Canterbury, Christchurch, New Zealand
Darinka Dentcheva Department of Mathematical Sciences, Stevens Institute of
Technology, Hoboken, NJ, USA
Dean Diavatopoulos Finance, Villanova University, Villanova, PA, USA
Gang Nathan Dong Columbia University, New York, NY, USA
Diep Duong Department of Business and Economics, Utica College, Utica, NY,
USA
Frank J. Fabozzi EDHEC Business School, EDHEC Risk Institute, Nice, France
Luis Ferruz Facultad de Economı́a y Empresa, Departamento de Contabilidad y
Finanzas, Universidad de Zaragoza, Zaragoza, Spain
Wayne E. Ferson University of Southern California, Los Angeles, CA, USA
Cheng-der Fuh Graduate Institute of Statistics, National Central University,
Zhongli City, Taiwan
Antonio F. Galvao University of Iowa, Iowa City, IA, USA
Gerard L. Gannon Deakin University, Burwood, VIC, Australia
Fernando Gómez-Bezares Universidad de Deusto, Bilbao, Spain
John W. Goodell College of Business Administration, University of Akron,
Akron, OH, USA
Hongtao Guo Bertolon School of Business, Salem State University, Salem, MA,
USA
Jia-Hau Guo Institution of Finance, College of Management, National Chiao
Tung University, Hsinchu, Taiwan
Manak C. Gupta Temple University, Philadelphia, PA, USA
Ahmed Hachicha Department of Economic Development, Faculty of Economics
and Management of Sfax, University of Sfax, Sfax, Tunisia
Fatma Hachicha Department of Finance, Faculty of Economics and Management
of Sfax, Sfax, Tunisia
xxii Contributors

Chuan-Hsiang Han Department of Quantitative Finance, National Tsing Hua

University, Hsinchu, Taiwan, Republic of China
Wolfgang Karl Härdle Ladislaus von Bortkiewicz Chair of Statistics, C.A.S.E. –
Center for Applied Statistics and Economics, Humboldt–Universität zu Berlin,
Berlin, Berlin, Germany
Lee Kong Chian School of Business, Singapore Management University,
Singapore, Singapore
Richard Heaney Accounting and Finance, The University of Western Australia,
Perth, Australia
Yuna Heo Rutgers Business School, Rutgers, The State University of New Jersey,
Newark-New Brunswick, NJ, USA
Hemantha Herath Department of Accounting, Faculty of Business, Brock
University, St. Catharines, ON, Canada
Thomas S. Y. Ho Thomas Ho Company Ltd, New York, NY, USA
Der-Tzon Hsieh Department of Economics, National Taiwan University, Taipei,
Taiwan
Chin-Wen Hsin Yuan Ze University, Zhongli City, Taiwan
Chun-Pin Hsu Department of Accounting and Finance, York College, The City
University of New York, Jamaica, NY, USA
Yu Chuan Hsu National Chi Nan University, Nantou, Taiwan
Jian Hua Baruch College (CUNY), New York, NY, USA
Bwo-Nung Huang National Chung-Cheng University, Minxueng Township,
Chiayi County, Taiwan
Chin-Wen Huang Department of Finance, Western Connecticut State University,
Danbury, CT, USA
Jing-Zhi Huang Smeal College of Business, Penn State University, University
Park, PA, USA
Yongchang C. Hui School of Mathematics and Statistics, Xi’an Jiaotong
University, Xi’an, China
Ken Hung Division of International Banking & Finance Studies, Texas A&M
International University, Laredo, TX, USA
Mao-Wei Hung College of Management, National Taiwan University, Taipei,
Taiwan
John S. Jahera Jr. Department of Finance, College of Business, Auburn
University, Auburn, AL, USA
Contributors xxiii

Kiridaran Kanagaretnam Schulich School of Business, York University,

Toronto, ON, Canada
Long Kang Department of Finance, Antai College of Economics and
Management, Shanghai Jiao Tong University, Shanghai, China
The Options Clearing Corporation and Center for Applied Economics and Policy
Research, Indiana University, Bloomington, IN, USA
Lie-Jane Kao Department of Finance and Banking, Kainan University, Taoyuan,
ROC, Taiwan
Young Shin (Aaron) Kim College of Business, Stony Brook University, Stony
Brook, NY, USA
Gary Kleinman Montclair State University, Montclair, NJ, USA
April Knill The Florida State University, Tallahassee, FL, USA
Alexander Kogan Rutgers Business School, Rutgers, The State University of
New Jersey, Newark–New Brunswick, NJ, USA
Rutgers Center for Operations Research (RUTCOR), Piscataway, NJ, USA
Joshua Krausz Yeshiva University, New York, NY, USA
Wikil Kwak University of Nebraska at Omaha, Omaha, NE, USA
Tze Leung Lai Stanford University, Stanford, CA, USA
Miranda S. Lam Bertolon School of Business, Salem State University, Salem,
MA, USA
Kenneth D. Lawrence New Jersey Institute of Technology, Newark, NJ, USA
Sheila M. Lawrence Rutgers, The State University of New Jersey, New
Brunswick, NJ, USA
Alice C. Lee State Street Corp., USA
Cheng-Few Lee Department of Finance and Economics, Rutgers Business
School, Rutgers, The State University of New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
Chia-Ju Lee College of Business, Chung Yuan University, Chungli, Taiwan
Chien-Hui Lee National Kaohsiung University of Applied Sciences, Kaohsiung,
Taiwan
Geul Lee University of New South Wales, Sydney, Australia
Heeseok Lee Korea Advanced Institute of Science and Technology, Yuseong-gu,
Daejeon, South Korea
xxiv Contributors

Jang-Yi Lee Tunghai University, Taichung, Taiwan

Jieun Lee Economic Research Institute, Bank of Korea, Seoul, South Korea
John C. Lee Center for PBBEF Research, North Brunswick, NJ, USA
Kin-Wai Lee Division of Accounting, Nanyang Business School, Nanyang
Technological University, Singapore, Singapore
Kuo-Hao Lee Department of Finance, College of Business, Bloomsburg
University of Pennsylvania, Bloomsburg, PA, USA
Sang Bin Lee Hanyang University, Seong-Dong-Ku, Seoul, Korea
Miguel A. Lejeune George Washington University, Washington, DC, USA
Jiandong Li Chinese Academy of Finance and Development (CAFD) and Central
University of Finance and Economics (CUFE), Beijing, China
Chan-Chien Lien Treasury Division, E.SUN Commercial Bank, Taipei, Taiwan
Donald Lien The University of Texas at San Antonio, San Antonio, TX, USA
Si Rou Lim National Chi Nan University, Nantou, Taiwan
Emily Lin St. John’s University, New Taipei City, Taiwan
Fang-Pang Lin National Center for High Performance Computing, Hsinchu,
Taiwan
Hai Lin School of Economics and Finance, Victoria University of Wellington,
Wellington, New Zealand
Hsuan-Chu Lin Graduate Institute of Finance and Banking, National Cheng-
Kung University, Tainan, Taiwan
Nathan Liu Department of Finance, Feng Chia University, Taichung, Taiwan
Gerald J. Lobo C.T. Bauer College of Business, University of Houston, Houston,
TX, USA
Anastasia Maggina Business Consultant/Research Scientist, Avlona, Attikis,
Greece
Afif Masmoudi Department of Mathematics, Faculty of Sciences of Sfax, Sfax,
Tunisia
Philippe Masset Ecole Hôtelière de Lausanne, Le-Chalet-à-Gobet, Lausanne 25,
Switzerland
Robert Mathieu School of Business and Economics, Wilfrid Laurier University,
Waterloo, ON, Canada
Bin Mei Warnell School of Forestry and Natural Resources, University of
Georgia, Athens, GA, USA
Contributors xxv

Thomas Meinl Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany

Kristina L. Minnick Bentley University, Waltham, MA, USA
Heather Mitchell RMIT University, Melbourne, VIC, Australia
Bruce Mizrach Department of Economics, Rutgers, The State University of New
Jersey, New Brunswick, NJ, USA
Kiseok Nam Yeshiva University, New York, NY, USA
Ali Nejadmalayeri Department of Finance, Oklahoma State University,
Oklahoma, OK, USA
Shou Zhong Ng Hong Kong Monetary Authority, Hong Kong, China
Oded Palmon Department of Finance and Economics, Rutgers Business School –
Newark and New Brunswick, Piscataway, NJ, USA
Dilip K. Patro RAD, Office of the Comptroller of the Currency, Washington, DC,
USA
Robert L. Porter Department of Finance School of Business, Quinnipiac
University, Hamden, CT, USA
Svetlozar T. Rachev Department of Applied Mathematics and Statistics, College
of Business, Stony Brook University, SUNY, Stony Brook, NY, USA
FinAnalytica, Inc, New York, NY, USA
Vikash Ramiah School of Economics, Finance and Marketing, RMIT University,
Melbourne, Australia
Raafat R. Roubi Department of Accounting, Faculty of Business, Brock
University, St. Catharines, ON, Canada
Andrzej Ruszczynski Department of Management Science and Information
Systems, Rutgers, The State University of New Jersey, Piscataway, NJ, USA
G.V. Satya Sekhar Department of Finance, GITAM Institute of Management,
GITAM University, Visakhapatnam, Andhra Pradesh, India
Robert A. Schwartz Zicklin School of Business, Baruch College, CUNY, New
York, NY, USA
Thomas V. Schwarz Stetson University, DeLand, FL, USA
Yuan-Chung Sheu National Chiao-Tung University, Hsinchu, Taiwan
Yong Shi University of Nebraska at Omaha, Omaha, NE, USA
Chinese Academy of Sciences, Beijing, China
Zhan Shi Smeal College of Business, Penn State University, University Park, PA,
USA
xxvi Contributors

Tung-Li Shih Department of Hospitality Management, Ming Dao University,

Changhua Peetow, Taiwan
Wei K. Shih Bates White Economic Consulting, Washington, DC, USA
Andrew F. Siegel University of Washington, Seattle, WA, USA
Nicholas Sim School of Economics, University of Adelaide, Adelaide, SA,
Australia
Ben J. Sopranzetti Rutgers, The State University of New Jersey, Newark, NJ,
USA
Suresh Srivastava University of Alaska Anchorage, Anchorage, AK, USA
Jung-Bin Su Department of Finance, China University of Science and
Technology, Nankang, Taipei, Taiwan
Edward W. Sun KEDGE Business School and BEM Management School,
Bordeaux, France
Norman R. Swanson Department of Economics, Rutgers, The State University of
New Jersey, New Brunswick, NJ, USA
Tzu Tai Department of Finance and Economics, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Alex P. Tang Morgan State University, Baltimore, MD, USA
Stephen J. Taylor Lancaster University Management School, Lancaster, UK
Stuart Thomas RMIT University, Melbourne, VIC, Australia
Chiung-Min Tsai Central Bank of the Republic of China (Taiwan), Taipei,
Taiwan, Republic of China
Maria Vargas Universidad de Zaragoza, Zaragoza, Spain
Itzhak Venezia School of Business, The Hebrew University, Jerusalem, Israel
Bocconi University, Milan, Italy
Chia-Jane Wang Manhattan College, Riverdale, NY, USA
Cindy Shin-Huei Wang CORE, Université Catholique de Louvain and FUNDP,
Academie Louvain, Louvain-la-Neuve, Belgium
Department of Quantitative Finance, National TsingHwa University, Hsinchu City,
Taiwan
Kehluh Wang Graduate Institute of Finance, National Chiao Tung University,
Hsinchu, Taiwan
Shin Yun Wang National Dong Hwa University, Shou-Feng, Hualien, Taiwan
Contributors xxvii

Weining Wang Ladislaus von Bortkiewicz Chair of Statistics, C.A.S.E. – Center

for Applied Statistics and Economics, Humboldt-Universität zu Berlin, Berlin,
Berlin, Germany
Yanzhi Wang Yuan Ze University, Taiwan
Department of Finance, College of Management, National Taiwan University,
Taipei, Taiwan
Yu-Jen Wang Graduate Institute of Finance, National Chiao Tung University,
Hsinchu, Taiwan
Bruce W. Weber Lerner College of Business and Economics, University of
Delaware, Newark, DE, USA
K. C. John Wei Hong Kong University of Science and Technology, Kowloon,
Hong Kong
Wing-Keung Wong Department of Economics, Hong Kong Baptist University,
Kowloon, Hong Kong
Po-Cheng Wu Department of Finance and Banking, Kainan University, Taoyuan,
ROC, Taiwan
Yangru Wu Rutgers Business School – Newark and New Brunswick, Rutgers,
The State University of New Jersey, New Brunswick, NJ, USA
Zhijie Xiao Department of Economics, Boston College, Chestnut Hill, MA, USA
Yi Meng Xie School of Business and Administration, Beijing Normal University,
Beijing, China
Department of Economics, University of Southern California, Los Angeles, CA,
USA
Haipeng Xing SUNY at Stony Brook, Stony Brook, NY, USA
Li Xu Washington State University, Richland, WA, USA
Chin W. Yang Clarion University of Pennsylvania, Clarion, PA, USA
National Chung Cheng University, Chia–yi, Taiwan
Li Yang University of New South Wales, Sydney, Australia
Yating Yang Graduate Institute of Finance, National Chiao Tung University,
Hsinchu, Taiwan
Shih-Kuo Yeh Department of Finance, National Chung Hsing University,
Taichung 402, Taiwan, Republic of China
Hai-Chin Yu Department of International Business, Chung Yuan University,
Chungli, Taiwan
xxviii Contributors

Jing Rung Yu National Chi Nan University, Nantou, Taiwan

Qianni Yuan Department of Finance, Xiamen University, Xiamen, China
Kamil Yilmaz College of Administrative Sciences and Economics, Koc
University, Istanbul, Turkey
Shaojun Zhang School of Accounting and Finance, Faculty of Business, Hong
Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong
Feng Zhao The University of Texas at Dallas, Richardson, TX, USA
Wei Zhong Wang Yanan Institute for Studies in Economics and Department of
Statistics, School of Economics, Xiamen University, Xiamen, China
Chunyang Zhou Shanghai Jiaotong University, Shanghai, China
Xing Zhou Rutgers Business School – Newark and New Brunswick, Rutgers,
The State University of New Jersey, New Brunswick, NJ, USA
Introduction to Financial Econometrics
and Statistics 1
Cheng-Few Lee and John C. Lee

Contents
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Financial Econometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Single Equation Regression Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Simultaneous Equation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 Panel Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.4 Alternative Methods to Deal with Measurement Error . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.5 Time Series Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.6 Spectral Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Financial Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Important Statistical Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Principle Components and Factor Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Nonparametric and Semi-parametric Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.4 Cluster Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Applications of Financial Econometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Applications of Financial Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Overall Discussion of Papers in this Handbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 Summary and Conclusion Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Appendix 1: Brief Abstracts and Keywords for Chapters 2 to 99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

C.-F. Lee (*)

Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: lee@business.rutgers.edu
J.C. Lee
Center for PBBEF Research, North Brunswick, NJ, USA
e-mail: johnleejohnlee@yahoo.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1

DOI 10.1007/978-1-4614-7750-1_1,
# Springer Science+Business Media New York 2015
2 C.-F. Lee and J.C. Lee

Abstract
The main purposes of this introduction chapter are (i) to discuss important
financial econometrics and statistics which have been used in finance and
accounting research and (ii) to present an overview of 98 chapters which have
been included in this handbook. Sections 1.2 and 1.3 briefly review and
discuss financial econometrics and statistics. Sections 1.4 and 1.5 discuss
application of financial econometrics and statistics. Section 1.6 first classifies
98 chapters into 14 groups in accordance with subjects and topics. Then this
section has classified the keywords from each chapter into two groups:
finance and accounting topics and methodology topics. Overall, this chapter
gives readers of this handbook guideline of how to apply this handbook to
their research.

1.1 Introduction

Financial econometrics and statistics have become very important tools for empir-
ical research in both finance and accounting. Econometric methods are important
tools for asset-pricing, corporate finance, options, and futures, and conducting
financial accounting research. Important econometric methods used in this research
include: single equation multiple regression, simultaneous regression, panel data
analysis, time-series analysis, spectral analysis, nonparametric analysis, semi-
parametric analysis, GMM analysis, and other methods.
Portfolio theory and management research have used different statistics
distributions, such as normal distribution, stable distribution, and log-normal dis-
tribution. Options and futures research have used binomial distribution, log-normal
distribution, non-central Chi-square distribution, Poisson distribution, and others.
Auditing research has used sampling survey techniques to determine the sampling
error and non-sampling error for auditing. Risk management research has used
Copula distribution and other distributions.
Section 1.1 is the introduction. Section 1.2 discusses financial econometrics.
In this section, we have six subsections. These subsections include single equation
regression methods, simultaneous equation models, panel data analysis, as well
as alternative methods to deal with measurement error, time-series analysis,
and spectral analysis. In the next section, Sect. 1.3, we discuss financial statistics.
Within financial statistics, we discuss six subtopics, including statistical
distributions; principle components and factor analysis; nonparametric, semi-
parametric, and GMM analyses; and cluster analysis. After exploring these topics,
we discuss the applications of financial econometrics and financial statistics in
Sects. 1.4 and 1.5. In Sect. 1.6, we discuss the overview of all papers included
in this handbook in accordance with the subject and methodologies used in the
papers. Finally in Sect. 1.7, we summarize all the chapters in this handbook and add
our concluding remarks.
1 Introduction to Financial Econometrics and Statistics 3

As mentioned previously, Sect. 1.2 covers the topic of financial econometrics. We

divide this section into six subsections. Within Sect. 1.2.1, we talk about single equation
regression methods. We discuss some important issues related to single equation
regression methods, including Heteroskedasticity, Specification Error, Measurement
Error, Skewness and the Kurtosis Effect, Nonlinear Regression and Box-Cox transfor-
mation, Structural Change, the Chow Test and Moving Chow Test, Threshold Regres-
sion, Generalize Fluctuation Test, Probit and Logit Regression for Credit Risk Analysis,
Poisson Regression, and Fuzzy Regression. The next subsection, Sect. 1.2.2, analyzes
simultaneous equation models. Within the realm of simultaneous equation models, we
discuss two-stage least squares estimation (2SLS) method, seemly unrelated regression
(SUR) method, three-stage least squares estimation (3SLS) method, and disequilibrium
estimation method. In Sect. 1.2.3, we study panel data analysis, in which we go
over fixed effect model, random effect model, and clustering effect. The next subsection,
Sect. 1.2.3, explores alternative methods to deal with measurement error. The alternative
methods we look over in this section includes LISREL model, multifactor and
multi-indicator (MIMIC) model, partial least square method, and grouping method.
After we discuss alternative methods to deal with measurement error, we examine in
Sect. 1.2.4 time-series analysis. We include in our section about time-series analysis
some important models, including ARIMA, ARCH, GARCH, fractional GARCH, and
combined forecasting. In Sect. 1.2.5, we look into spectral analysis.
In the following section, Sect. 1.3, we discuss financial statistics, along with four
subsequent subtopics. In our first subsection, Sect. 1.3.1, we discuss some important
statistical distributions. This subsection will look into the different types of distri-
butions that are in statistics, including Binomial and Poisson distribution, normal
distribution, log-normal distribution, Chi-square distribution, and non-central
Chi-square distribution, Wishart distribution, symmetric and non-symmetric stable
distributions, and other known distributions. Then, we talk about principal compo-
nents and factor analysis in Sect. 1.3.2. In the following subsection, Sect. 1.3.3, we
examine nonparametric, semi-parametric, and GMM analyses. The last subsection,
Sect. 1.3.4, explores cluster analysis.
After discussing financial econometrics, we explore the applications of this topic
in different types of financial and accounting field research. In Sect. 1.4, we
describe these applications, including asset-pricing research, corporate finance
research, financial institution research, investment and portfolio research, option
pricing research, future and hedging research, mutual fund research, hedge fund
research, microstructure, earnings announcements, real option research, financial
accounting, managerial accounting, auditing, term structure modeling, credit risk
modeling, and trading cost/transaction cost modeling.
We also discuss applications of financial statistics into different types of
financial and accounting field research. Section 1.5 will include these applications
in asset-pricing research, investment and portfolio research, credit risk management
research, market risk research, operational risk research, option pricing research,
mutual fund research, hedge fund research, value-at-risk research, and auditing.
4 C.-F. Lee and J.C. Lee

1.2 Financial Econometrics

1.2.1 Single Equation Regression Methods

There are important issues related to single equation regression estimation method.
They are (a) Heteroskedasticity, (b) Specification error, (c) Measurement error,
(d) Skewness and kurtosis effect, (e) Nonlinear regression and Box-Cox transfor-
mation, (f) Structural change, (g) Chow test and moving Chow test, (h) Threshold
regression, (i) Generalized fluctuation, (j) Probit and Logit regression for credit risk
analysis, (k) Poisson regression, and (l) Fuzzy regression. These issues are briefly
discussed as follows:
(a) Heteroskedasticity
– White (1980) and Newvey and West (1987) are two important papers
discussing how the heteroskedasticity test can be performed. The latter
paper discusses heteroskedasticity when there are serial correlations.
(b) Specification error
– Specification error occurs when there is missing variable in a regression analysis.
To test the existence of specification error, we can refer to the papers by Thursby
(1985), Fok et al. (1996), Cheng and Lee (1986), and Maddala et al. (1996).
(c) Measurement error
– Management error problem is when there exists imprecise independent
variable in a regression analysis. Papers by Lee and Jen (1978), Kim
(1995, 1997, 2010), Miller and Modigliani (1966), and Lee and Chen
(2012) have explored how measurement error methods can be applied to
finance research. Lee and Chen have discussed alternative errors in variable
estimation methods and their application in finance research.
(d) Skewness and kurtosis effect
– Both skewness and kurtosis are two important measurement variables to pre-
pare stock variation analysis. Papers by Lee (1976a), Sears and Wei (1988), and
Lee and Wu (1985) discuss the skewness and kurtosis issue in asset pricing.
(e) Nonlinear regression and Box-Cox transformation
– Nonlinear regression and Box-Cox transformation are important tools for
finance, accounting, and urban economic researches. Papers by Lee (1976,
1977), Lee et al. (1990), Frecka and Lee (1983), and Liu (2006) have
discussed how nonlinear regression and Box-Cox transformation techniques
can be used to improve the specification of finance and accounting research.
Kau and Lee (1976), and Kau et al. (1986) have explored how Box-Cox
transformation can be used to conduct the empirical study of urban structure.
(f) Structural change
– Papers by Yang (1989), Lee et al. (2011b, 2013) have discussed how the
structural change model can be used to improve the empirical study of
dividend policy and the issuance of new equity.
(g) Chow test and Moving Chow test
– Chow (1960) has proposed a dummy variable approach to examine the
existence of structure change for regression analysis. Zeileis et al. (2002)
1 Introduction to Financial Econometrics and Statistics 5

have developed software programs to perform the Chow test and other
structural change models which has been frequently used in finance and
economic research.
(h) Threshold regression
– Hansen (1996, 1997, 1999, 2000a, and 2000b) have explored the issue of
threshold regressions and their applications in detecting structure change for
regression.
(i) Generalize fluctuation test
– Kuan and Hornik (1995) have discussed how the generalized fluctuation test
can be used to perform structural change to regression.
(j) Probit and Logit regression for credit risk analysis
– Probit and Logit regressions are frequently used in credit risk analysis.
Ohlson (1980) used the accounting ratio and macroeconomic data to do
credit risk analysis. Shumway (2001) has used accounting ratios and
stock rate returns for credit risk analysis in terms of Probit and
Logit regression techniques. Most recently, Hwang et al. (2008, 2009)
and Cheng et al. (2010) have discussed Probit and Logit regression for credit
risk analysis by introducing nonparametric and semi-parametric techniques
into this kind of regression analysis.
(k) Poisson regression
– Lee and Lee (2012) have discussed how the Poisson Regression can be
performed, regardless of the relationship between multiple directorships,
corporate ownership, and firm performance.
(l) Fuzzy regression
– Shapiro (2005), Angrist and Lavy (1999), and Van Der Klaauw (2002) have
discussed how Fuzzy Regression can be performed. This method has the
potential to be used in finance accounting and research.

1.2.2 Simultaneous Equation Models

In this section, we will discuss alternative methods to deal with simultaneous

equation models. There are (a) two-stage least squares estimation (2SLS) method,
(b) seemly unrelated regression (SUR) method, (c) three-stage least squares esti-
mation (3SLS) method, (d) disequilibrium estimation method, and (e) generalized
method of moments.
(a) Two-stage least squares estimation (2SLS) method
– Lee (1976a) has applied this to started market model; Miller and Modigliani
(1966) have used 2SLS to study cost of capital for utility industry; Chen
et al. (2007) have discuss the two-stage least squares estimation (2SLS)
method for investigating corporate governance.
(b) Seemly unrelated regression (SUR) method
– Seemly unrelated regression has frequently used in economic and financial
research. Lee and Zumwalt (1981) have discussed how the seemly unrelated
regression method can be applied in asset-pricing determination.
6 C.-F. Lee and J.C. Lee

(c) Three-stage least squares estimation (3SLS) method

– Chen et al. (2007) have discussed how the three-stage least squares estima-
tion (3SLS) method can be applied in corporate governance research.
(d) Disequilibrium estimation method
– Mayer (1989), Martin (1990), Quandt (1988), Amemiya (1974), and Fair
and Jaffee, (1972) have discussed how alternative disequilibrium estimation
method can be performed. Tsai (2005), Sealey (1979), and Lee et al. (2011a)
have discussed how the disequilibrium estimation method can be applied in
asset-pricing test and banking management analysis.
(e) Generalized method of moments
– Hansen (1982) and Hamilton (1994, ▶ Chap. 14) have discussed how GMM
method can be performed. Chen et al. (2007) have used the two-stage least
squares estimation (2SLS), three-stage squares method, and GMM method
to investigate corporate governance.

1.2.3 Panel Data Analysis

In this section, we will discuss important issues related to panel data analysis. They are
(a) fixed effect model, (b) random effect model, and (c) clustering effect model.
Three well-known textbooks by Wooldridge (2010), Baltagi (2008) and Hsiao
(2003) have discussed the applications of panel data in finance, economics, and
accounting research. Now, we will discuss the fixed effect, random effect, and
clustering effect in panel data analysis.
(a) Fixed effect model
– Chang and Lee (1977) and Lee et al. (2011a) have discussed the role of
the fixed effect model in panel data analysis of dividend research.
(b) Random effect model
– Arellano and Bover (1995) have explored the random effect model and its role
in panel data analysis. Chang and Lee (1977) have applied both fix effect and
random effect model to investigating the relationshipbetween price per share,
dividend per share, and retained earnings per share.
(c) Clustering effect model
– Papers by Thompson (2011), Cameron et al. (2006), and Petersen (2009)
review the clustering effect model and its impact on panel data analysis.

1.2.4 Alternative Methods to Deal with Measurement Error

In this section, we will discuss alternative methods of dealing with measurement

error problems. They are (a) LISREL model, (b) multifactor and multi-indicator
(MIMIC) model, and (c) partial least square method, and (d) grouping method.
(a) LISREL model
– Papers by Titman and Wessal (1988), Chang (1999), Chang et al. (2009),
Yang et al. (2010) have described the LISREL model and its way to resolve
the measurement error problems of finance research.
1 Introduction to Financial Econometrics and Statistics 7

(b) Multifactor and multi-indicator (MIMIC) model

– Chang et al. (2009) and Wei (1984) have applied in the multifactor and multi-
indicator (MIMIC) model in capital structure and asset-pricing research.
(c) Partial least square method
– Papers by Core (2000), Ittner et al. (1997), and Lambert and Lacker (1987)
have applied the partial least square method to deal with measurement error
problems in accounting research.
(d) Grouping method
– Papers by Lee (1973), Chen (2011), Lee and Chen (2013), Lee (1977b),
Black et al. (1972), Blume and Friend (1973), and Fama and MacBeth
(1973) analyze grouping method and its way to deal with measurement
error problem in capital asset-pricing tests.
There are other errors in variable method, such as (i) Classical method,
(ii) instrumental variable method, (iii) mathematical programming method,
(iv) maximum likelihood method, (v) GMM method, and (vi) Bayesian Statistic
Method. Lee and Chen (2012) have discussed all above-mentioned methods in
details.

1.2.5 Time Series Analysis

In this section, we will discuss important models in time-series analysis. They are
(a) ARIMA, (b) ARCH, (c) GARCH, (d) fractional GARCH, and (e) combined
forecasting.
– Two well-known textbooks by Anderson (1994) and Hamilton (1994) have
discussed the issues related to time-series analysis. We will discuss some
important topics in time-series analysis in the following subsections.
– Myers (1991) discloses ARIMA’s role in time-series analysis: Lien and Shrestha
(2007) discuss ARCH and its impact on time-series analysis: Lien (2010)
discusses GARCH and its role in time-series analysis: Leon and Vaello-Sebastia
(2009) further research into GARCH and its role in time series in a model called
Fractional GARCH.
– Granger and Newbold (1973), Granger and Newbold (1974), Granger and
Ramanathan (1984) have theoretically developed combined forecasting
methods. Lee et al. (1986) have applied combined forecasting methods to
forecast market beta and accounting beta. Lee and Cummins (1998) have
shown how to use the combined forecasting methods to perform cost of capital
estimates.

1.2.6 Spectral Analysis

Anderson (1994), Chacko and Viceira (2003), and Heston (1993) have discussed
how spectral analysis can be performed. Heston (1993) and Bakshi et al. (1997)
have applied spectral analysis in the evaluation of option pricing.
8 C.-F. Lee and J.C. Lee

1.3 Financial Statistics

1.3.1 Important Statistical Distributions

In this section, we will discuss different statistical distributions. They are:

(a) Poisson distribution, (c) normal distribution, (d) log-normal distribution,
(e) Chi-square distribution, (f) non-central Chi-square distribution.
Two well-known textbooks by Cox et al. (1979) and Rendleman and Barter
(1979) have used binomial, normal, and lognormal distributions to develop an
option pricing model. The following subsections note some famous authors
that provide studies on these different statistical distributions. Black and Sholes
(1973) have used lognormal distributions to derive the option pricing model.
Finally, Aitchison and Brown (1973) is a well-known book to investigate lognormal
distribution. Schroder (1989) has derived the option pricing model in terms of
non-central Chi-square distribution.
Fama (1971) has used stable distributions to investigate the distribution of stock
rate of returns. Chen and Lee (1981) have derived statistics distribution of Sharpe
performance measure and found that Sharpe performance measure can be described
by Wishart distribution.

1.3.2 Principle Components and Factor Analysis

Anderson’s (2003) book entitled “An Introduction to Multivariate Statistical

Analysis” has discussed principal components and factor analysis in detail. Chen
and Shimerda (1981), Pinches and Mingo (1973), and Kao and Lee (2012) discuss
how principal components and factor analyses can be used to do finance Lee et al.
(1989) and accounting research.

1.3.3 Nonparametric and Semi-parametric Analyses

Ait-Sahalia and Lo (2000), and Hutchison et al. (1994) have discussed how
nonparametric can be used in risk management and derivative securities evaluation.
Hwang et al. (2010), and Hwang et al. (2007) have used semi-parametric to conduct
credit risk analysis.

1.3.4 Cluster Analysis

The detailed procedures to discuss how cluster analysis can be used to find
groups in data can be found in the textbook by Kaufman and Rousseeuw (1990).
Brown and Goetzmann (1997) have applied cluster analysis in mutual fund
research.
1 Introduction to Financial Econometrics and Statistics 9

1.4 Applications of Financial Econometrics

In this section, we will briefly discuss how different methodologies of financial

econometrics will be applied to the topics of finance and accounting.
(a) Asset-pricing Research
– Methodologies used in asset-pricing research include (1) Heteroskedasticity,
(2) Specification error, (3) Measurement error, (4) Skewness and kurtosis
effect, (5) Nonlinear regression and Box-Cox transformation, (6) Structural
change, (7) Two-stage least squares estimation (2SLS) method, (8) Seemly
unrelated regression (SUR) method, (9) Three-stage least squares
estimation (3SLS) method, (10) Disequilibrium estimation method,
(11) Fixed effect model, (12) Random effect model, (13) Clustering effect
model of panel data analysis, (14) Grouping method, (15) ARIMA, (16) ARCH,
(17) GARCH, (18) Fractional GARCH, and (19) Wishart distribution.
(b) Corporate Finance Research:
– Methodologies used in Corporate finance research include (1) Heteroske-
dasticity, (2) Specification error, (3) Measurement error, (4) Skewness and
kurtosis effect, (5) Nonlinear regression and Box-Cox transformation,
(6) Structural change, (7) Probit and Logit regression for credit risk analysis,
(8) Poisson regression, (9) Fuzzy regression, (10) Two-stage least squares
estimation (2SLS) method, (11) Seemly unrelated regression (SUR) method,
(12) Three-stage least squares estimation (3SLS) method, (13) Fixed effect
model, (14) Random effect model, (15) Clustering effect model of panel data
analysis, and (16) GMM Analysis.
(c) Financial Institution Research
– Methodologies used in Financial Institution research include (1) Heteroske-
dasticity, (2) Specification error, (3) Measurement error, (4) Skewness and
kurtosis effect, (5) Nonlinear regression and Box-Cox transformation,
(6) Structural change, (7) Probit and Logit regression for credit risk analysis,
(8) Poisson regression, (9) Fuzzy regression, (10) Two-stage least squares
estimation (2SLS) method, (11) Seemly unrelated regression (SUR) method,
(12) Three-stage least squares estimation (3SLS) method, (13) Disequilib-
rium estimation method, (14) Fixed effect model, (15) Random effect model,
(16) Clustering effect model of panel data analysis, (17) Semiparametric
analysis.
(d) Investment and Portfolio Research
– Methodologies used in investment and portfolio research include
(1) Heteroskedasticity, (2) Specification error, (3) Measurement error,
(4) Skewness and kurtosis effect, (5) Nonlinear regression and Box-Cox
transformation, (6) Structural change, (7) Probit and Logit regression for
credit risk analysis, (8) Poisson regression, and (9) Fuzzy regression.
(e) Option Pricing Research
– Methodologies used in option pricing research include (1) ARIMA,
(2) ARCH, (3) GARCH, (4) Fractional GARCH, (5) Spectral analysis,
(6) Binomial distribution, (7) Poisson distribution, (8) normal distribution,
10 C.-F. Lee and J.C. Lee

(9) log-normal distribution, (10) Chi-square distribution, (11) non-central

Chi-square distribution, and (12) Nonparametric analysis.
(f) Future and Hedging Research
– Methodologies used in future and hedging research include (1) Heteroske-
dasticity, (2) Specification error, (3) Measurement error, (4) Skewness and
kurtosis effect, (5) Nonlinear regression and Box-Cox transformation,
(6) Structural change, (7) Probit and Logit regression for credit risk analysis,
(8) Poisson regression, and (9) Fuzzy regression.
(g) Mutual Fund Research
– Methodologies used in mutual fund research include (1) Heteroskedasticity,
(2) Specification error, (3) Measurement error, (4) Skewness and kurtosis
effect, (5) Nonlinear regression and Box-Cox transformation, (6) Structural
change, (7) Probit and Logit regression for credit risk analysis, (8) Poisson
regression, (9) Fuzzy regression, and (10) Cluster analysis.
(h) Credit Risk Modeling
– Methodologies used in credit risk modeling include (1) Heteroskedasticity,
(2) Specification error, (3) Measurement error, (4) Skewness and kurtosis effect,
(5) Nonlinear regression and Box-Cox transformation, (6) Structural change,
(7) Two-stage least squares estimation (2SLS) method, (8) Seemly unrelated
regression (SUR) method, (9) Three-stage least squares estimation (3SLS)
method, (10) Disequilibrium estimation method, (11) Fixed effect model,
(12) Random effect model, (13) Clustering effect model of panel data analysis,
(14) ARIMA, (15) ARCH, (16) GARCH, and (17) Semiparametric analysis.
(i) Other Application
– Financial econometrics is also important tools to conduct research in
(1) Trading cost/transaction cost modeling, (2) Hedge fund research,
(3) Microstructure, (4) Earnings announcement, (5) Real option research,
(6) Financial accounting, (7) Managerial accounting, (8) Auditing, and
(9) Term structure modeling.

1.5 Applications of Financial Statistics

Financial statistics is an important tool for research in (1) Asset-pricing research,

(2) Investment and portfolio research, (3) Credit risk management research,
(4) Market risk research, (5) Operational risk research, (6) Option pricing research,
(7) Mutual fund research, (8) Hedge fund research, (9) Value-at-risk research, and
(10) Auditing research.

1.6 Overall Discussion of Papers in this Handbook

In this section, we classify 98 papers (chapters 2–99) which have been presented in
Appendix 1 in accordance with (A) Chapter titles and (B) Keywords.
(A) Chapter title classification in terms of Chapter Titles
1 Introduction to Financial Econometrics and Statistics 11

Based on chapter titles, we classify 98 chapters into the following 14 topics:

(i) Financial Accounting (▶ Chaps. 2, 9, 10, 61, 97)
(ii) Mutual Funds (▶ Chaps. 3, 24, 25, 68, 88)
(iii) Microstructure (▶ Chaps. 4, 44, 47, 96)
(iv) Corporate Finance (▶ Chaps. 5, 21, 30, 38, 42, 46, 60, 63, 75, 79, 95)
(v) Asset Pricing (▶ Chaps. 6, 15, 22, 28, 34, 36, 39, 45, 50, 81, 85, 87, 93, 99)
(vi) Options (▶ Chaps. 7, 32, 37, 55, 65, 84, 86, 90, 98)
(vii) Portfolio Analysis (▶ Chaps. 8, 26, 35, 53, 67, 73, 80, 81, 83)
(viii) Risk Management (▶ Chaps. 11, 13, 16, 17, 23, 27, 41, 51, 54, 72, 91, 92)
(ix) International Finance (▶ Chaps. 12, 40, 43, 59, 69)
(x) Event Study (▶ Chap. 14)
(xi) Methodology (▶ Chaps. 18, 19, 20, 29, 31, 33, 46, 49, 52, 56, 57, 58, 62, 74,
76, 77, 78, 82, 89)
(xii) Banking Management (▶ Chap. 64)
(xiii) Pension Funds (▶ Chap. 66)
(xiv) Futures and Index Futures (▶ Chaps. 48, 70, 71, 94)
(B) Keywords classification
Based on the keywords in Appendix 1, we classify these keywords into two groups:
(i) finance and accounting topics and (ii) methodology topics. The number behind
each keyword is the chapter it is associated with.
(i) Finance and Accounting Topics
Abnormal earnings (87), Accounting earnings (87), Activity-based costing
system (27), Agency costs (5, 97), Aggregation bias (43), Analyst experience
(2), Analyst forecast accuracy (63), Analysts’ forecast accuracy (97), Analysts’
forecast bias (63, 97), Arbitrage pricing theory (APT) (6, 7, 36, 81), Asset (93),
Asset allocation (45), Asset allocation fund (88), Asset pricing (34, 81), Asset
return predictability (76), Asset returns (52), Asset-pricing returns (96), Asym-
metric information (5), Asymmetric mean reversion (15), Asymmetric stochas-
tic volatility (62), Asymmetric volatility response (15), Balanced scorecard
(29), Bank capital (13), Bank holding companies (13), Bank risks (13), Bank
stock return (6), Banks (12), Barrier option (65), Basket credit derivatives (23),
Behavioral finance (55, 66, 73), Bias (57), Bias reduction (92), Bid-ask spreads
(96, 99), Binomial option pricing model (37), Black-Scholes model (7, 90),
Black-Sholes option pricing model (37), Board structure (42), Bond ratings
(89), Bottom-up capital budgeting (75), Bounded complexity (85), Bounds
(71), Brier score (72), Brokerage reputation (63), Business cycle (67), Business
models (75), Business performance evaluation (29), Business value of firm,
Buy-and-hold return (50), Calendar-time (50), Calendar-time portfolio
approach (14), Call option (37), Capital asset-pricing model (CAPM) (6, 25,
28, 36, 81, 93), Capital budgeting (75, 29), Capital markets (25), Capital
structure (5, 60), Carry trade (69), Case-Shiller home price indices (19), CEO
compensation (97), CEO stock options (97), Change of measure (30), Cheapest-
to-deliver bond (71), Chicago board of trade, (71), Cholesky decomposition
(23), Closed-end Funds (25), Comparative financial systems (12), Composite
trapezoid rule (51), Comprehensive earnings (87), Compromised solution (89),
12 C.-F. Lee and J.C. Lee

Compustat database (38), Compound sum method (46), Conditioning informa-

tion (35), Constant/dynamic hedging (44), Contagious effect (11), Corner port-
folio (45), Corporate earnings (9), Corporate finance (5), Corporate merger (21),
Corporate ownership structure (42), Corporate policies (38), Corporation regu-
lation (9), Correlated defaults (11), Cost of capital (93), Country funds (25),
Credit rating (21), Credit rating (27), Credit risk (27, 65, 91), Credit risk index
(27), Credit risk rating (16), Credit VaR (91), Creditworthiness (16), Cumula-
tive abnormal return (50), Cumulative probability distribution (45), Currency
market (58), Cyberinfrastructure (49), Daily realized volatility (40), Daily stock
price (82), Debt maturity (64), Delivery options (71), Delta (45), Demand (33),
Demonstration effect (17), Deterioration of bank asset quality (64), Determinants
of capital structure (60), Discount cash flow model (46), Discretionary accruals
(61), Discriminant power (89), Disposition effect (22), Dividends (38, 65, 79),
Domestic investment companies (17), Double exponential smoothing (88), Dual-
ity (83), Dynamics (67), Earning management (61), Earnings change (10), Earn-
ings level (10), Earnings quality (42), Earnings surprises (81), Economies of scale
(21), Ederington hedging effectiveness (70), Effort allocation (97), Effort aversion
(55), EGB2 distribution (80), Electricity (33), Empirical Bayes (85), Empirical
corporate finance (95), Employee stock option (30), Endogeneity (38, 95),
Endogeneity of variables (13), Endogenous supply (93), Equity valuation models
(87), Equity value (75), European option (7), European put (5), Evaluation (34),
Evaluation of funds (3), Exactly identified (93), Exceedance correlation (52),
Exchange rate (43, 59), Executive compensation schemes (55), Exercise bound-
ary (30), Expected market risk premium (15), Expected stock return (80),
Expected utility (83), Experimental control (4), Experimental economics (4),
Extreme events (67), Fallen angel (72), Finance panel data (24), Financial
analysts (2), Financial crisis (64), Financial institutions (12), Financial leverage
(75), Financial markets (12), Financial modeling (3), Financial planning and
forecasting (87), Financial ratios (21), Financial returns (62), Financial service
(49), Financial simulation (49), Financial statement analysis (87), Financial
strength (16), Firm and time effects (24), Firm Size (9), Firm’s performance
score (21), Fixed operating cost (75), Flexibility hypothesis (79), Foreign
exchange market (40), Foreign investment (17), Fourier inversion (84), Fourier
transform (19), Free cash flow hypothesis (79), Frequentist segmentation (85),
Fund management (53), Fundamental analysis (87), Fundamental asset values
(73), Fundamental transform (84), Futures hedging (70), Gamma (45), Gener-
alized (35), Generalized autoregressive conditional heteroskedasticity (51),
Global investments (3), Gold (58), Green function (84), Grid and cloud com-
puting (49), Gross return on investment (GRI) (75), Group decision making
(29), Growth option (75), Growth rate (46), Hawkes process (11), Heavy-tailed
data (20), Hedge ratios (98), Hedging (98), Hedging effectiveness (94), Hedging
performance (98), Herding (66), Herding towards book-to-market factor (66),
Herding towards momentum factor (66), Herding towards size factor (66), Herding
towards the market (66), High end computing (49), High-dimensional data (77),
Higher moments (80), High-frequency data (40), High-order moments (57),
1 Introduction to Financial Econometrics and Statistics 13

Historical simulation (45), Housing (78), Illiquidity (30), Imitation (66),

Implied standard deviation (ISD) (90), Implied volatility (32, 90), Impression
management (61), Impulse response (76), Incentive options (55), Income
from operations (61), Independence screening (77), Index futures (44),
Index options (32), Inflation targeting (59), Information asymmetry (2, 96),
Information content (92), Information content of trades (76), Information
technology (49), Informational efficiency (76), Instantaneous volatility (92),
Institutional investors (17), Insurance (20), Intangible assets (38), Interest
rate risk (6), Interest rate volatility (86), Internal control material weakness
(63), Internal growth rate (46), Internal rating (16), International capital asset
pricing model (ICAPM) (25, 40), Internet bubble (53), Intertemporal risk-
return relation (15), Intraday returns (44), Investment (67), Investment equa-
tions (57), Investment risk taking (97), Investment strategies (68), Investment
style (68), Issuer default (23), Issuer-heterogeneity (72), Kernel pricing (7),
Laboratory experimental asset markets (73), Lead-lag relationship (17), Left-
truncated data (20), Legal traditions (12), Limited dependent variable model
(99), Liquidity (22, 99), Liquidity risk (64), Local volatility (92), Logical
analysis of data (16), Logical rating score (16), Long run (59), Long-run
stock return (50), Lower bound (7), Management earnings (9), Management
entrenchment (5), Management myopia (5), Managerial effort (55), Market
anomalies (44), Market efficiency (28, 73), Market microstructure (4, 96),
Market model (99), Market perfection (79), Market performance measure
(75), Market quality (76), Market segmentation (25), Market uncertainties
(67), Market-based accounting research (10), Markov property (72), Martin-
gale property (94), Micro-homogeneity (43), Minimum variance hedge ratio
(94), Mis-specified returns (44), Momentum strategies (81), Monetary policy
shock (59), Mutual funds (3), NAV of a mutual fund (88), Nelson-Siegel
curve (39), Net asset value (25), Nonrecurring items (61), Net present value
(NPV) (75), Oil (58), Oil and gas industry (61), OLS hedging strategy (70),
On-/off-the-run yield spread (22), Online estimation (92), Operating earnings
(87), Operating leverage (75), Operational risk (20), Opportunistic disclosure
management (97), Opportunistic earnings management (97), Optimal hedge
ratio (94), Optimal payout ratio (79), Optimal portfolios (35), Optimal trade-
offs (29), Option bounds (7), Option prices (32), Option pricing (49, 65),
Option pricing model (90), Optional bias (2), Options on S&P 500 index
futures (90), Oracle property (77), Order imbalance (96), Out-of-sample
return (8), Out-of-the-money (7), Output (59), Overconfidence (55),
Overidentifying restrictions (95), Payout policy (79), Pension funds (66),
Percent effective spread (99), Performance appraisal (3), Performance eval-
uation (8), Performance measures (28), Performance values (68), Persistence
(44), Persistent change (31), Poison put (5), Political cost (61), Portfolio
management (3, 35, 70), Portfolio optimization (8, 83), Portfolio selection
(26), Post-earnings-announcement drift (81), Post-IT policy (59), Predicting
returns (35), Prediction of price movements (3), Pre-IT policy (59),
Preorder (16), Price impact (99), Price indexes (78), Price level (59),
14 C.-F. Lee and J.C. Lee

Price on earnings model (10), Pricing (78), Pricing performance (98),

Probability of informed trading (PIN) (96), Property (78), Property rights (12),
Put option (37), Put-call parity (37), Quadratic cost (93), Quality options (71),
Random number generation (49), Range (74), Rank dependent utility (83), Rating
migration (72), Rational bias (2), Rational expectations (43), Real estate (78), Real
sphere (59), Realized volatility (74), Recurrent event (72), Recursive (85), Reflec-
tion principle (65), Regime-switching hedging strategy (70), Registered trading
firms (17), Relative value of equity (75), Research and development expense (61),
Restrictions (59), Retention option (75), Return attribution (75), Return models
(10), Reverse-engineering (16), Risk (83), Risk adjusted performance (3), Risk
aversion (55), Risk management (41, 49, 67, 74, 80), Risk measurement (26),
Risk premium (80), Risk-neutral pricing (32), Robust estimation (41), S&P
500 index (7), Sarbanes-Oxley act (63), SCAD penalty (77), Scale-by-scale
decomposition (19), Seasonality (33), Semi-log (78), Sentiment (30), Shape
parameter (82), Share prices (59), Share repurchases (38), Sharpe ratios (35, 53),
Short run (59), Short selling (26), Short-term financing (64), Signaling hypothesis
(79), Sigma (37), Smile shapes (32), Smooth transition (74), Special items (61),
Speculative bubbles (73), Spot price (33), Stationarity (10), Statistical learning
(77), Stochastic discount factors (25, 35), Stochastic interest rates (98), Stochastic
order (83), Stochastic volatility (44, 84, 85, 92, 98), Stock market overreaction
(15), Stock markets (67), Stock option (65), Stock option pricing (98), Stock
price indexes (62), Stock/futures (44), Strike price (55), Structural break (31),
Subjective value (30), Substantial price fluctuations (82), Sustainable growth
rate, synergy (21), Synthetic utility value (68), Systematic risk (3, 79), TAIEX
(45), Tail risk (67), Timberland investments (34), Time-varying risk (25), Time-
varying risk aversion (40), Time-varying volatility (15), Timing options (71),
Tobin’s model (99), Top-down capital budgeting (75), Total risk (79), Tourna-
ment (73), Trade direction (96), Trade turnover industry (9), Transaction costs
(99), Transfer pricing (29), Treasury bond futures (71), Trend extraction (18),
Trust (12), Turkish economy (59), U.S. stocks (52), Ultrahigh-dimensional data
(77), Unbiasedness (43), Uncertainty avoidance (12), Uncovered interest parity
(69), Unexpected volatility shocks (15), Unsystematic risk (3), Utility-based
hedging strategy (70), VaR-efficient frontier (45), Variability percentage adjust-
ment (21), Visual Basic for applications (37), Volatility index (VIX) (92),
Volatility (37, 80), Volatility co-persistence (44), Volatility daily effect (92),
Volatility dependencies (62), Volatility feedback effect (15), Weak efficiency
(43), Weak instruments (95), Wealth transfer (75), Write-downs (61), Yaari’s
dual utility (83), Yield curve (39), Zero-investment portfolio (50).
(ii) Methodology Topics
A mixture of Poisson distribution (98), Analyst estimation (ANOVA) (2, 28),
Analytic hierarchy process (29), Analysis of variance (19), Anderson-Darling
statistic (20), Anderson-Rubin statistic (95), ANST-GARCH model
(asymmetric nonlinear smooth transition- GARCH model) (15), Approxi-
mately normal distribution (28), ARCH (41, 44), ARCH models (32),
ARX-GARCH (autoregressive (AR) mean process with exogenous (X)
1 Introduction to Financial Econometrics and Statistics 15

variables- GARCH model) (85), Asset-pricing tests (35), Asset-pricing

regression (24) Asymmetric dependence (52), Asymptotic distribution (44),
Autocovariance (99), Autoregression (62), Autoregressive conditional jump
intensity (82), Autoregressive model (88), Autoregressive moving average
with exogenous variables (10), Autoregressive parameters (44), Bankruptcy
prediction (27), Bayesian updating (2), Binomial distribution (28), Block
bootstrap (56), Block granger causality (17), Bootstrap (8, 50), Bootstrap
test (14), Bootstrapped critical values (24), Boundary function (31), Box-Cox
(78), Bubble test (31), Change-point models (85), Clayton copula (8, 11),
Cluster standard errors (24), Custering effect (79), Co-integration (76),
Co-integration breakdown test (31), Combination of forecasts (88),
Combinatorial optimization (16), Combined forecasting (87), Combining
forecast (27), Complex logarithm (84), Conditional distribution (56), Condi-
tional market model (50), Conditional skewness (80), Conditional value-at-
risk (26, 83), Conditional variance (70), Conditional variance estimates
(44), Contemporaneous jumps (85), Contingency tables (28), Contingent
claim model (75), Continuous wavelet transform (19), Cook’s distance (63),
Copula (8, 41, 67, 74, 91), Correction method (92), Correlation (67, 73),
Correlation analysis (99), CoVar (54), Covariance decomposition (72),
Cox-Ingersoll-Ross (CIR) model (22, 71), Cross-sectional and time-series
dependence (42), CUSUM squared test (31), Data-mining (16), Decision
trees (37), Default correlation (23, 91), Dickey-Fuller test (10), Dimension
reduction (77), Discrete wavelet transform (19), Discriminant analysis (89),
Distribution of underlying asset (7), Double clustering (24), Downside risk
model (26), Dynamic conditional correlation (52, 58, 74), Dynamic factor
model (11), Dynamic random-effects models (38), Econometric methodology
(38), Econometric modeling (33), Econometrics (12), Error component
two-stage least squares (EC2SLS) (12), Error in variable problem (60, 96),
Estimated cross-sectional standard deviations of betas (66), Event study
methodology (5, 50), Ex ante probability (82), Excess kurtosis
(44), Exogeneity test (95), Expectation–maximization (EM) algorithm
(96), Expected return distribution (45), Explanatory power (89), Exponential
trend model (88), Extended Kalman filtering (86), Factor analysis (68, 89),
Factor copula (23, 91), Factor model (39, 50), Fama-French three-factor
model (14), Feltham and Ohlson model (87), Filtering methods (19), Fixed
effects (57, 63, 79), Forecast accuracy (2, 9), Forecast bias (2), Forecasting
complexity (97), Forecasting models (27), Fourier transform method (92),
Frank copula (11), GARCH (8, 40, 41, 96), GARCH hedging strategy (70),
GARCH models (48, 52), GARCH-in-mean (40), GARCH-jump model (82),
GARJI model (82), Gaussian copula (8), Generalized correlations (77), Gen-
eralized hyperbolic distribution (94), Generalized method of moments (13),
Gibbs sampler (62), Generalized least square (GLS) (36), Generalized
method of moments (GMM) (5, 25, 43, 57, 95), Generalized two-stage least
squares (G2SLS) (12), Goal programming (89), Goodness-of-fit test (82, 20),
Granger-causality test (76), Gumbel copula (8, 11), Hazard model (72),
16 C.-F. Lee and J.C. Lee

Heath-Jarrow-Morton model (86), Hedonic models (78), Heston (84),

Heterogeneity (43), Heteroskedasticity (57), Hidden Markov models (85),
Hierarchical clustering with K-Means approach (30), Hierarchical system
(68), Huber estimation (66), Hyperparameter estimation (85), Hypothesis
testing (53), Ibottson’s RATS (50), Infinitely divisible models (48), Instru-
mental variable (IV) estimation (95, 57), Johnson’s Skewness-adjusted t-test
(14), Joint-normality assumption (94), Jones (1991) model (61), Jump detec-
tion (18), Jump dilution model (30), Jump process (56), Kalman filter (66),
Kolmogorov-Smirnov statistic (20), Kupiec’s proportion of failures test (48),
Large-scale simulations (14), Latent variable (60), Least squares (78), Like-
lihood maximization (32), Linear filters (18), Linear trend model (88),
LISREL approach (36, 60), Locally linear quantile regression (54), Logistic
smooth transition regression model (69), Logit regression (27),
Log-likelihood function (99), Lognormal (65), Long-horizon event study
(14), Long memory process (31, 32), Loss distribution (20), Loss function
(51), MAD model (26), Matching procedure (63), Mathematical optimization
(55), MATLAB (90), Maximum likelihood (35, 38, 52), Maximum likelihood
estimation (MLE) (36, 44, 71), Maximum sharp measure (94), Markov Chain
Monte Carlo (MCMC) (62, 69, 85), Mean-variance ratio (53), Measurement
error (36, 57), Method of maximum likelihood (51), Method of moments (35),
A comparative study of two models SV with MCMC algorithm (62), Microsoft
Excel (37), Multiple indicator multiple causes (MIMIC) (36), Minimum gener-
alized semi-invariance (94), Minimum recording threshold (20), Minimum value
of squared residuals (MSE loss function) (10), Minimum variance efficiency
(35), Misspecification (44), Model formulation (38), Model selection (56, 77),
Monitoring fluctuation test (31), Monte Carlo simulation (11, 23, 32, 49, 57),
Moving average method (88), Moving estimates processes (79), MSE (62),
Multifactor diffusion process (56), Multifactor multi-indicator approach (36),
Multiple criteria and multiple constraint linear programming (29), Multiple
criteria decision making (MCDM) (68), Multiple indicators and multiple causes
(MIMIC) model (60), Multiple objective programming (26), Multiple regression
(6, 9), Multi-resolution analysis (19), Multivariate technique (89), Multivariate
threshold autoregression model (17), MV model (26), Nonlinear filters (18),
Nonlinear Kalman filter (22), Nonlinear optimization (38), Non-normality (41),
Nonparametric (7), Nonparametric density estimation (86), Nonparametric tests
(28), Normal copula (11), Normal distribution (45), Ohlson model (87), Order
flow models (4), Ordered logit (27), Ordered probit (27), Ordinary least-squares
regression (63, 73), Ordinary least-squares (OLS) (90, 39, 95, 36), Orthogonal
factors (6), Outlier (33), Out-of-sample forecasts (56), Panel data estimates
(12, 40, 38), Panel data regressions (42, 2), Parametric approach (51), Parametric
bootstrap (35), Partial adjustment (93), Partial linear model (54), Penalized least-
squares (77), Prediction test (31), Principle component analysis (89, 91), Principle
component factors (21), Probability density function (27), Quantile
autoregression (QAR) (41), Quadratic trend model (88), Quantile dependence
(67), Quantile regression (41, 54, 67), Quasi-maximum likelihood (22),
1 Introduction to Financial Econometrics and Statistics 17

Quasi-maximum likelihood estimation strategy (48), Random walk models (4),

Rank regressions (63), Realized distribution (58), Rebalancing model (26),
Recursive filters (85), Recursive programming (37), Reduced-form model
(23, 93), Regime-switch model (69), Regression models (85, 78), Revenue
surprises rotation-corrected angle (84), Ruin probability (20), Seemingly
unrelated regressions (SUR) (40, 93), Semi-parametric approach (51), Semi-
parametric model (54), Serial correlation (44), Shrinkage (77), Simple adjusted
formula (84), Simulations (55), Simultaneous equations (60, 93, 95, 87), Single
clustering (24), Single exponential smoothing (88), Skewed generalized student’s
t (51), Skewness (57), Specification test (56), Spectral analysis (19), Standard
errors in finance panel data (30), Standardized Z (21), State-space model (39, 66),
Static factor model (11), Stepwise discriminant analysis (89), Stochastic domi-
nance (83), Stochastic frontier analysis (13), Structural change model (79),
Structural equation modeling (SEM) (60), Structural VAR (17), Student’s
t-copula (8, 52), Student’s t-distribution (62), Seemingly unrelated regression
(SUR) (43), Survey forecasts (43), Survival analysis (72), SVECM models (59),
Stochastic volatility (SVOL) (62), Tail dependence (52), Taylor series expansion
(90), t-Copula (11), Tempered stable distribution (48), Term structure (32, 39,
71), Term structure modeling (86), Time-series analysis (18, 34, 41), Time-
heterogeneity (72), Time-series and cross-sectional effects (12), Time-varying
covariate (72), Time-varying dependence (8), Time-varying parameter (34),
Time-varying rational expectation hypothesis (15), Trading simulations (4),
Two-sector asset allocation model (45), Two-stage estimation (52), Two-stage
least square (2SLS) (95), Two-way clustering method of standard errors (42),
Unbounded autoregressive moving average model (88), Unconditional coverage
test (51), Unconditional variance (70), Uniformly most powerful unbiased test
(53), Unit root tests (10), Unit root time series (31), Unweighted GARCH
(44), Value-at-risk (VAR) (45, 54, 83, 20, 26, 41, 48, 51, 76), Variable selection
(77), Variance decomposition (76), Variance estimation (70), Variance reduction
methods (32), Variance-gamma process (82), VG-NGARCH model (82), Visual
Basic for applications (VBA) (37), Volatility forecasting (74), Volatility regime
switching (15), Volatility threshold (58), Warren and Shelton model (87), Wave-
let (18), Wavelet filter (19), Weighted GARCH, (44), Weighted least-squares
regression (14), Wilcoxon rank test (21), Wilcoxon two-sample test (9),
Wild-cluster bootstrap (24), and Winter’s method (88).

1.7 Summary and Conclusion Remarks

This chapter has discussed important financial econometrics and statistics which
have been used in finance and accounting research. In addition, this chapter has
presented an overview of 98 chapters which have been included in this handbook.
In Sect. 1.2 “Financial Econometrics,” we have six subsections which are: a single
equation regression methods, Simultaneous equation models, Panel data analysis,
Alternative methods to deal with measurement error, Time-series analysis, and
18 C.-F. Lee and J.C. Lee

Spectral Analysis. Section 1.3 “Financial Statistics” has four subsections: Important
Statistical Distributions, Principle components and factor analysis, Nonparametric
and Semi-parametric analyses, Cluster analysis review and discuss financial
econometrics and statistics. In Sect. 1.4 “Applications of financial econometrics,”
we briefly discuss how different methodologies of financial econometrics will
be applied to the topics of finance and accounting. These methods include: Asset-
pricing Research, Corporate Finance Research, Financial Institution Research,
Investment and Portfolio Research, Option Pricing Research, Future and Hedging
Research, Mutual Fund Research, and Credit Risk Modeling. Section 1.5, “Appli-
cations of Financial Statistics,” states that financial statistics is an important tool
to conduct research in the areas of (1) Asset-pricing Research, (2) Investment
and Portfolio Research, (3) Credit Risk Management Research, (4) Market
Risk Research, (5) Operational Risk Research, (6) Option Pricing Research,
(7) Mutual Fund Research, (8) Hedge Fund Research, (9) Value-at-risk Research,
and (10) Auditing. Section 1.6 is an “Overall Discussion of Papers in this
Handbook.” It classifies 98 chapters into 14 groups in accordance to Chapter title
and keywords.

Appendix 1: Brief Abstracts and Keywords for Chapters 2 to 99

Chapter 2: Experience, Information Asymmetry, and Rational

Forecast Bias

This chapter uses a Bayesian model of updating forecasts in which the bias in
forecast endogenously determines how the forecaster’s own estimates weigh into
the posterior beliefs. The model used in this chapter predicts a concave relationship
between accuracy in forecast and posterior weight that is put on the forecaster’s self-
assessment. This chapter then uses a panel regression to test the analytical findings
and find that an analyst’s experience is indeed concavely related to the forecast error.
Keywords: Financial analysts, Forecast accuracy, Information asymmetry,
Forecast bias, Bayesian updating, Panel regressions, Rational bias, Optional bias,
Analyst estimation, Analyst experience

Chapter 3: An Overview of Modeling Dimensions for Performance

Appraisal of Global Mutual Funds (Mutual Funds)

This paper examines various performance models derived by financial experts

across the globe. A number of studies have been conducted to examine investment
performance of mutual funds of the developed capital markets. The measure of
performance of financial instruments is basically dependent on three important
models derived independently by Sharpe, Jensen, and Treynor. All three models
are based on the assumptions that (1) all investors are averse to risk, and are single
period expected utility of terminal wealth maximizers, (2) all investors have
1 Introduction to Financial Econometrics and Statistics 19

identical decision horizons and homogeneous expectations regarding investment

opportunities, (3) all investors are able to choose among portfolios solely on the
basis of expected returns and variance of returns, (4) all trans-actions costs and taxes
are zero, and (5) all assets are infinitely divisible. Overall, this paper has examined
nine alternative mutual funds measure. The method used in this kind of research is
regression analysis.
Keywords: Financial modeling, Mutual funds, Performance appraisal,
Global investments, Evaluation of funds, Portfolio management, Systematic
risk, Unsystematic risk, Risk adjusted performance, Prediction of price movements

Chapter 4: Simulation as a Research Tool for Market Architects

This chapter uses simulation to gain insights into trading and market structure topic
by two statistical methods. The statistical methods we use include experimental
design, and careful controls over experimental parameters such as the instructions
given to participants. The first is discrete event simulation and the model of
computer-generated trade order flow that we describe in Sect. 3. To create a realis-
tic, but not ad hoc, market background, we use draws from a log-normal returns
distribution to simulate changes in a stock’s fundamental value, or P*. The model
uses price-dependent Poisson distributions to generate a realistic flow of computer-
generated buy and sell orders whose intensity and supply-demand balance vary over
time. The order flow fluctuations depend on the difference between the current
market price and the P* value. In Sect. 4, we illustrate the second method, which is
experimental control to create groupings of participants in our simulations that have
the same trading “assignment.” The result is the ability to make valid comparisons
of trader’s performance in the simulations.
Keywords: Trading simulations, Market microstructure, Order flow models,
Random walk models, Experimental economics, Experimental control

Chapter 5: Motivations for Issuing Putable Debt:

An Empirical Analysis

This paper is the first to examine the motivations for issuing putable bonds in which
the embedded put option is not contingent upon a company-related event. We find
that the market favorably views the issue announcement of these bonds that we
refer to as bonds with European put options or European putable bonds. This
response is in contrast to the response documented by the literature to other bond
issues (straight, convertible, and most studies examining poison puts), and to the
response documented in the current paper to the issue announcements of poison-put
bonds. Our results suggest that the market views issuing European putable bonds as
helping mitigate security mispricing. Our study is an application of important
statistical methods in corporate finance, namely, Event Studies and the use of
General Method of Moments for cross-sectional regressions.
20 C.-F. Lee and J.C. Lee

Keywords: Agency costs, Asymmetric information, Corporate finance, Capital

structure, Event study methodology, European put, General method of moments,
Management myopia, Management entrenchment, Poison put

Chapter 6: Multi Risk-Premia Model of U.S. Bank Returns:

An Integration of CAPM and APT

Interest rate sensitivity of bank stock returns has been studied using an augmented
CAPM: a multiple regression model with market returns and interest rate as
independent variables. In this chapter, we test an asset-pricing model in which
the CAPM is augmented by three orthogonal factors which are proxies for the
innovations in inflation, maturity risk, and default risk. The methodologies used in
this chapter are multiple regression and factor analysis.
Keywords: CAPM, APT, Bank stock return, Interest rate risk, Orthogonal
factors, Multiple regression

Chapter 7: Non-parametric Bounds for European Option Prices

This chapter derives a new nonparametric lower bound and provides an alternative
interpretation of Ritchken’s (1985) upper bound to the price of the European option.
In a series of numerical examples, our new lower bound is substantially tighter than
previous lower bounds. This is prevalent especially for out-of-the-money (OTM)
options where the previous lower bounds perform badly. Moreover, we present that
our bounds can be derived from histograms which are completely nonparametric in
an empirical study. We first construct histograms from realizations of S&P
500 index returns following Chen et al. (2006), calculate the dollar beta of the
option and expected payoffs of the index and the option, and eventually obtain our
bounds. We discover violations in our lower bound and show that those violations
present arbitrage profits. In particular, our empirical results show that out-of-the-
money calls are substantially overpriced (violate the lower bound). The methodol-
ogies used in this chapter are nonparametric, option pricing model, and histograms
methods.
Keywords: Option bounds, Nonparametric, Black-Scholes model, European
option, S&P 500 index, Arbitrage, Distribution of underlying asset, Lower bound,
Out-of-the-money, Kernel pricing

Chapter 8: Can Time-Varying Copulas Improve Mean-Variance

Portfolio?

This chapter evaluates whether constructing a portfolio using time-varying copulas

yields superior returns under various weight updating strategies. Specifically,
minimum-risk portfolios are constructed based on various copulas and the Pearson
1 Introduction to Financial Econometrics and Statistics 21

correlation, and a 250-day rolling window technique is adopted to derive a sequence

of time-varied dependences for each dependence model. Using daily data of the G-7
countries, our empirical findings suggest that portfolios using time-varying copulas,
particularly the Clayton-dependence, outperform those constructed using Pearson
correlations. The above results still hold under different weight updating strategies
and portfolio rebalancing frequencies. The methodologies used in this chapter are
Copulas, GARCH, Student’s t-Copula, Gumbel Copula, Clayton Copula, Time-
Varying Dependence, Portfolio Optimization, and Bootstrap.
Keywords: Copulas, Time-varying dependence, Portfolio optimization, Boot-
strap, Out-of-sample return, Performance evaluation, GARCH, Gaussian copula,
Student’s t-copula, Gumbel copula, Clayton copula

Chapter 9: Determinations of Corporate Earnings Forecast Accuracy:

Taiwan Market Experience

This chapter examines the accuracy of the earnings forecasts by the following test
methodologies. Multiple Regression Models are used to examine the effect of six
factors: firm size, market volatility, trading volume turnover, corporate earnings
variances, type of industry, and experience. If the two-sample groups are related,
Wilcoxon Two-Sample Test will be used to determine the relative earnings forecast
accuracy. Readers are well advised and referred to the chapter appendix for
methodological issues such as sample selection, variable definition, regression
model, and Wilcoxon tow-sample test.
Keywords: Multiple regression, Wilcoxon two-sample test, Corporate earnings,
Forecast accuracy, Management earnings, Firm size, Corporation regulation,
Volatility, Trade turnover, Industry

Chapter 10: Market-Based Accounting Research (MBAR) Models:

A Test of ARIMAX Modeling

This study uses standard models such as earnings level and earnings changes, among
others. Models that fit better to the data drawn from companies listed on the Athens
Stock Exchange have been selected employing autoregressive integrated moving
average with exogenous variables (ARIMAX) models. Models I (price on earnings
model) “II (returns on change in earnings divided by beginning-of-period price and
prior period)” V (returns on change in earnings over opening market value), VII
(returns deflated by lag of 2 years on earnings over opening market value), and IX
(differenced-price model) have statistically significant coefficients of explanatory
variables. These models take place with backward looking information instead of
forward looking information that recent literature is assessed. The methodologies
used in this chapter are price on earnings model, return models, autoregressive
moving average with exogenous variables (ARIMAX), minimum value of squared
residuals (MSE loss function), and Dickey-Fuller test.
22 C.-F. Lee and J.C. Lee

Keywords: Market-based accounting research (MBAR), Price on earnings

model, Earnings level, Earnings change, Return models, Autoregressive moving
average with exogenous variables (ARIMAX), Minimum value of squared resid-
uals (MSE loss function), Unit root tests stationarity, Dickey-Fuller test

Chapter 11: An Assessment of Copula Functions Approach in

Conjunction with Factor Model in Portfolio Credit Risk Management

This study uses a mixture of the dynamic factor model of Duffee (1999) and a
contagious effect in the specification of a Hawkes process, a class of counting
processes which allows intensities to depend on the timing of previous events
(Hawkes 1971). Using the mixture factor- contagious-effect model, Monte Carlo
simulation is performed to generate default times of two hypothesized firms. The
goodness-of-fit of the joint distributions based on the most often used copula
functions in literature, including the Normal, t-, Clayton, Frank, and Gumbel
copula, respectively, is assessed against the simulated default times. It is demon-
strated that as the contagious effect increases, the goodness-of-fit of the joint
distribution functions based on copula functions decreases, which highlights the
deficiency of the copula function approach.
Keywords: Static factor model, Dynamic factor model, Correlated defaults,
Contagious effect, Hawkes process, Monte Carlo simulation, Normal copula,
t-copula, Clayton copula, Frank copula, Gumbel copula

Chapter 12: Assessing Importance of Time-Series Versus

Cross-Sectional Changes in Panel Data: A Study of International
Variations in Ex-Ante Equity Premia and Financial Architecture

This chapter uses simultaneous equation modeling and uses Hausman test to
determine whether to report fixed or random-effects estimates. We first report
random-effects estimates based on the estimation procedure of Baltagi (Baltagi
1981; Baltagi and Li 1995; Baltagi and Li 1994). We consider that the error
component two-stage least squares (EC2SLS) estimator of Baltagi and Li (1995) is
more efficient than the generalized two-stage least squares (G2SLS) estimator
of Balestra and Varadharajan-Krishnakumar (1987). For our second estimation
procedure, for comparative purposes, we use the dynamic panel modeling estimates
recommended by Blundell and Bond (1998). We employ the model of Blundell and
Bond (1998), as these authors argue that their estimator is more appropriate than the
Arellano and Bond (1991) model for smaller time periods relative to the size
of the panels. We also use this two-step procedure, use as an independent
variable the first lag of the dependent variable, reporting robust standard errors of
Windmeijer (2005). Thus, our two different panel estimation techniques place
differing emphasis on cross-sectional and time-series effects, with the Baltagi-Li
1 Introduction to Financial Econometrics and Statistics 23

estimator emphasizing cross-sectional effects and the Blundell-Bond estimator

emphasizing time-series effects.
Keywords: Panel data estimates, Time-series and cross-sectional effects, Econo-
metrics, Financial institutions, Banks, Financial markets, Comparative financial sys-
tems, Legal traditions, Uncertainty avoidance, Trust, Property rights, Error component
two-stage least squares (EC2SLS), The generalized two-stage least squares (G2SLS)

Chapter 13: Does Banking Capital Reduce Risk?: An Application

of Stochastic Frontier Analysis and GMM Approach

This chapter employs stochastic frontier analysis to create a new type of instru-
mental variable. The unrestricted frontier model determines the highest possible
profitability based solely on the book value of assets employed. We develop
a second frontier based on the level of bank holding company capital as well as
the amount of assets. The implication of using the unrestricted model is that we are
measuring the unconditional inefficiency of the banking organization. This chapter
applies generalized method of moments (GMM) regression to avoid the problem
caused by departure from normality. To control for the impact of size on a bank’s
risk-taking behavior, the book value of assets is considered in the model.
The relationship between the variables specifying bank behavior and the use of
equity is analyzed by GMM regression.
Keywords: Bank capital, Generalized method of moments, Stochastic frontier
analysis, Bank risks, Bank holding companies, Endogeneity of variables

Chapter 14: Evaluating Long-Horizon Event Study Methodology

This chapter examines the performance of more than 20 different testing procedures
that fall into two categories. First, the buy-and-hold benchmark approach uses
a benchmark to measure the abnormal buy-and-hold return for every event firm,
and tests the null hypothesis that the average abnormal return is zero. Second, the
calendar-time portfolio approach forms a portfolio in each calendar month
consisting of firms that have had an event within a certain time period prior to the
month, and tests the null hypothesis that the intercept is zero in the regression of
monthly portfolio returns against the factors in an asset-pricing model. This chapter
also evaluates the performance of bootstrapped Johnson’s skewness-adjusted t-test.
This computation-intensive procedure is considered because the distribution of
long-horizon abnormal returns tends to be highly skewed to the right. The
bootstrapping method uses repeated random sampling to measure the significance
of relevant test statistics. Due to the nature of random sampling, the resultant
measurement of significance varies each time such a procedure is used. We also
evaluate simple nonparametric tests, such as the Wilcoxon signed-rank test or the
Fisher’s sign test, which are free from random sampling variation.
24 C.-F. Lee and J.C. Lee

Keywords: Long-horizon event study, Johnson’s Skewness-adjusted t-test,

Weighted least-squares regression, Bootstrap test, Calendar-time portfolio approach,
Fama-French three-factor model, Johnson’s skewness-adjusted t-statistic, Large-scale
simulations

Chapter 15: Effect of Unexpected Volatility Shocks on Intertemporal

Risk-Return Relation

This chapter employs the ANST-GARCH model that is capable of capturing the
asymmetric volatility effect of a positive and negative return shock. The key feature
of the model is the regime-shift mechanism that allows a smooth, flexible
transition of the conditional volatility between different states of volatility
persistence. The regime-switching mechanism is governed by a logistic transition
function that changes values depending on the level of the previous return shock.
With a negative (positive) return shock, the conditional variance process is
described as a high (low)-persistence-in-volatility regime. The ANST-GARCH
model describes the heteroskedastic return dynamics more accurately and generates
better volatility forecasts.
Keywords: Intertemporal risk-return relation, Unexpected volatility shocks,
Time-varying rational expectation hypothesis, Stock market overreaction,
Expected market risk premium, Volatility feedback effect, Asymmetric mean
reversion, Asymmetric volatility response, Time-varying volatility, Volatility
regime switching, ANST-GARCH model

Chapter 16: Combinatorial Methods for Constructing Credit

Risk Ratings

This chapter uses a novel method, the Logical Analysis of Data (LAD), to reverse-
engineer and construct credit risk ratings which represent the creditworthiness of
financial institutions and countries. LAD is a data-mining method based on com-
binatorics, optimization, and Boolean logic that utilizes combinatorial search tech-
niques to discover various combinations of attribute values that are characteristic of
the positive or negative character of observations. The proposed methodology is
applicable in the general case of inferring an objective rating system from archival
data, given that the rated objects are characterized by vectors of attributes taking
numerical or ordinal values. The proposed approaches are shown to generate
transparent, consistent, self-contained, and predictive credit risk rating models,
closely approximating the risk ratings provided by some of the major rating
agencies. The scope of applicability of the proposed method extends beyond the
rating problems discussed in this study, and can be used in many other contexts
where ratings are relevant. This study also uses multiple linear regression to derive
the logical rating scores.
1 Introduction to Financial Econometrics and Statistics 25

Keywords: Credit risk rating, Reverse-engineering, Logical analysis of data,

Combinatorial optimization, Data-mining, Creditworthiness, Financial strength,
Internal rating, Preorder, Logical rating score

Chapter 17: Dynamic Interactions in the Taiwan Stock Exchange:

A Threshold VAR Models

This chapter constructs a six-variable VAR model (including NASDAQ returns,

TSE returns, NT/USD returns, net foreign purchases, net domestic investment
companies (dic) purchases, and net registered trading firms (rtf) purchases) to
examine: (i) the interaction among three types of institutional investors, particularly
to test whether net foreign purchases lead net domestic purchases by dic and rtf (the
so-called demonstration effect); (ii) whether net institutional purchases lead market
returns or vice versa, and (iii) whether the corresponding lead-lag relationship is
positive or negative? Readers are well advised to refer to chapter appendix for
detailed discussion of the unrestricted VAR model, the structural VAR model, and
the threshold VAR analysis. The methodologies used in this chapter are multivar-
iate threshold autoregression model, structural VAR, and Block Granger Causality.
Keywords: Demonstration effect, Multivariate threshold autoregression model,
Foreign investment, Lead-lag relationship, Structural VAR, Block Granger causal-
ity, Institutional investors, Domestic investment companies, Registered trading
firms, Qualified foreign institutional investors

Chapter 18: Methods of Denoising Financial Data

This chapter uses denoising analysis which imposes new challenges for financial data
mining due to the irregularities and roughness observed in financial data, particularly,
for instantaneously collected massive amounts of tick-by-tick data from financial
markets for information analysis and knowledge extraction. Inefficient decomposition
of the systematic pattern (the trend) and noises of financial data will lead to erroneous
conclusions since irregularities and roughness of the financial data make the applica-
tion of traditional methods difficult. The methodologies used in this chapter are linear
filters, nonlinear filters, time-series analysis, trend extraction, and wavelet.
Keywords: Jump detection, Linear filters, Nonlinear filters, Time-series analy-
sis, Trend extraction, Wavelet

Chapter 19: Analysis of Financial Time: Series Using Wavelet

Methods

This chapter presents a set of tools, which allow gathering information about
the frequency components of a time-series. In the first step, we discuss spectral
26 C.-F. Lee and J.C. Lee

analysis and filtering methods. Spectral analysis can be used to identify and to
quantify the different frequency components of a data series. Filters permit to
capture specific components (e.g. trends, cycles, seasonalities) of the original
time-series. In the second step, we introduce wavelets, which are relatively new
tools in economics and finance. They take their roots from filtering methods and
Fourier analysis, but overcome most of the limitations of these two methods.
Their principal advantages derive from: (i) combined information from both
time-domain and frequency-domain and (ii) their flexibility as they do not
make strong assumptions concerning the data generating process for the series
under investigation.
Keywords: Filtering methods, Spectral analysis, Fourier transform, Wavelet
filter, Continuous wavelet transform, Discrete wavelet transform, Multi-resolution
analysis, Scale-by-scale decomposition, Analysis of variance, Case-Shiller home
price indices

Chapter 20: Composite Goodness-of-Fit Tests for Left Truncated

Loss Sample

This chapter derives the exact formulae for several goodness-of-fit statistics that
should be applied to loss models with left-truncated data where the fit of a distribution
in the right tail of the distribution is of central importance. We apply the proposed
tests to real financial losses, using a variety of distributions fitted to operational loss
and the natural catastrophe insurance claims data. The methodologies discussed in
this chapter are goodness-of-fit tests, loss distribution, ruin probability, value-at-risk,
Anderson-Darling statistic, Kolmogorov-Smirnov statistic.
Keywords: Goodness-of-fit tests, Left-truncated data, Minimum recording thresh-
old, Loss distribution, Heavy-tailed data, Operational risk, Insurance, Ruin probabil-
ity, Value-at-risk, Anderson-Darling statistic, Kolmogorov-Smirnov statistic

Chapter 21: Effect of Merger on the Credit Rating and Performance

of Taiwan Security Firms

This chapter identifies and defines variables for merger synergy analysis followed
by principal component factor analysis, variability percentage adjustment, and
performance score calculation. Finally, Wilcoxon sign rank test is used for hypoth-
esis testing. We extract principle component factors from a set of financial ratios.
Percentage of variability explained and factor loadings are adjusted to get
a modified average weight for each financial ratio. This weight is multiplied by
the standardized Z value of the variable, and summed a set of variables get a firm’s
performance score. Performance scores are used to rank the firm. Statistical signif-
icance of difference in pre- and post-merger rank is tested using the Wilcoxon
sign rank.
1 Introduction to Financial Econometrics and Statistics 27

Keywords: Corporate merger, Financial ratios, Synergy, Economies of scale,

Credit rating, Variability percentage adjustment, Principle component factors,
Firm’s performance score, Standardized Z, Wilcoxon rank test

Chapter 22: On-/Off-the-Run Yield Spread Puzzle: Evidence from

Chinese Treasury Markets

This chapter uses on-/off-the-run yield spread to describe “on-/off-the-run yield

spread puzzle” in Chinese treasury markets. To explain this puzzle, we introduce
a latent factor in the pricing of Chinese off-the-run government bonds and use this
factor to model the yield difference between Chinese on-the-run and off-the-run
issues. We use the nonlinear Kalman filter approach to estimate the model. The
methodologies used in this chapter are CIR model, nonlinear Kalman filter and
Quasi-maximum likelihood model.
Keywords: On-/off-the-run yield spread, Liquidity, Disposition effect, CIR
model, Nonlinear Kalman filter, Quasi-maximum likelihood

Chapter 23: Factor Copula for Defaultable Basket Credit Derivatives

This chapter uses a factor copula approach to evaluate basket credit derivatives with
issuer default risk and demonstrate its application in a basket credit linked note
(BCLN). We generate the correlated Gaussian random numbers by using the
Cholesky decomposition, and then, the correlated default times can be decided by
these random numbers and the reduced-form model. Finally, the fair BCLN coupon
rate is obtained by the Monte Carlo simulation. We also discuss the effect of
issuer default risk on BCLN. We show that the effect of issuer default risk cannot
be accounted for thoroughly by considering the issuer as a new reference entity in
the widely used one factor copula model, in which constant default correlation is often
assumed. A different default correlation between the issuer and the reference entities
affects the coupon rate greatly and must be taken into account in the pricing model.
Keywords: Factor copula, Issuer default, Default correlation, Reduced-form
model, Basket credit derivatives, Cholesky decomposition, Monte Carlo simulation

Chapter 24: Panel Data Analysis and Bootstrapping:

Application to China Mutual Funds

This chapter estimates double- and single-clustered standard errors by

wild-cluster bootstrap procedure. To obtain the wild bootstrap samples in
each cluster, we reuse the regressors (X), but modify the residuals by
transforming the OLS residuals with weights which follow the popular
two-point distribution suggested by Mammen (1993) and others. We then
28 C.-F. Lee and J.C. Lee

compare them with other estimates in a set of asset-pricing regressions. The

comparison indicates that bootstrapped standard errors from double clustering
outperform those from single clustering. They also suggest that bootstrapped
critical values are preferred to standard asymptotic t-test critical values to
avoid misleading test results.
Keywords: Asset-pricing regression, Bootstrapped critical values, Cluster stan-
dard errors, Double clustering, Firm and time effects, Finance panel data, Single
clustering, Wild-cluster bootstrap

Chapter 25: Market Segmentation and Pricing of Closed-End

Country Funds: An Empirical Analysis

This chapter finds that for closed-end country funds, the international CAPM can be
rejected for the underlying securities (NAVs) but not for the share prices. This
finding indicates that country fund share prices are determined globally, whereas
the NAVs reflect both global and local prices of risk. Cross-sectional variations in the
discounts or premiums for country funds are explained by the differences in the risk
exposures of the share prices and the NAVs. Finally, this chapter shows that the share
price and NAV returns exhibit predictable variation, and country fund premiums
vary over time due to time-varying risk premiums. The chapter employs Generalized
Method of Moments (GMM) to estimate stochastic discount factors and examines if
the price of risk of closed-end country fund shares and NAVs is identical.
Keywords: Capital markets, Country funds, CAPM, Closed-end funds, Market
segmentation, GMM, Net asset value, Stochastic discount factors, Time-varying
risk, International asset pricing

Chapter 26: A Comparison of Portfolios Using Different Risk

Measurements

This study uses three different risk measurements: the Mean-variance model, the
Mean Absolute Deviation model, and the Downside Risk model. Meanwhile short
selling is also taken into account since it is an important strategy that can bring
a portfolio much closer to the efficient frontier by improving a portfolio’s risk-
return trade-off. Therefore, six portfolio rebalancing models, including the MV
model, MAD model and the Downside Risk model, with/without short selling, are
compared to determine which is the most efficient. All models simultaneously
consider the criteria of return and risk measurement. Meanwhile, when short
selling is allowed, models also consider minimizing the proportion of short selling.
Therefore, multiple objective programming is employed to transform multiple
objectives into a single objective in order to obtain a compromising solution. An
example is used to perform simulation, and the results indicate that the MAD
model, incorporated with a short selling model, has the highest market value and
lowest risk.
1 Introduction to Financial Econometrics and Statistics 29

Keywords: Portfolio selection, Risk measurement, Short selling, MV model,

MAD model, Downside risk model, Multiple objective programming, Rebalancing
model, Value-at-risk, Conditional value-at-risk

Chapter 27: Using Alternative Models and a Combining Technique in

Credit Rating Forecasting: An Empirical Study

This chapter first utilizes the ordered logit and the ordered probit models. Then, we use
ordered logit combining method to weight different techniques’ probability measures,
as described in Kamstra and Kennedy (1998) to form the combining model.
The samples consist of firms in the TSE and the OTC market, and are divided into
three industries for analysis. We consider financial variables, market variables as well as
macroeconomic variables and estimate their parameters for out-of-sample tests. By
means of Cumulative Accuracy Profile, the Receiver Operating Characteristics, and
McFadden, we measure the goodness-of-fit and the accuracy of each prediction model.
The performance evaluations are conducted to compare the forecasting results, and we
find that combing technique does improve the predictive power.
Keywords: Bankruptcy prediction, Combining forecast, Credit rating, Credit
risk, Credit risk index, Forecasting models, Logit regression, Ordered logit,
Ordered probit, Probability density function

Chapter 28: Can We Use the CAPM as an Investment Strategy?:

An Intuitive CAPM and Efficiency Test

The aim of this chapter is to check whether certain playing rules, based on the
undervaluation concept arising from the CAPM, could be useful as investment
strategies, and can therefore be used to beat the Market. If such strategies work, we
will be provided with a useful tool for investors, and, otherwise, we will obtain
a test whose results will be connected with the efficient Market hypothesis (EMH)
and with the CAPM. The methodology used is both intuitive and rigorous: analyz-
ing how many times we beat the Market with different strategies, in order to check
whether when we beat the Market, this happens by chance.
Keywords: ANOVA, Approximately normal distribution, Binomial distribu-
tion, CAPM, Contingency tables, Market efficiency, Nonparametric tests, Perfor-
mance measures

Chapter 29: Group Decision Making Tools for Managerial

Accounting and Finance Applications

This chapter adopts an Analytic Hierarchy Process (AHP) approach to solve various
accounting or finance problems such as developing a business performance evalu-
ation system and developing a banking performance evaluation system. AHP uses
30 C.-F. Lee and J.C. Lee

hierarchical schema to incorporate nonfinancial and external performance mea-

sures. Our model has a broader set of measures that can examine external and
nonfinancial performance as well as internal and financial performance. While AHP
is one of the most popular multiple goals decision-making tools, Multiple Criteria
and Multiple Constraint (MC2) Linear Programming approach also can be used to
solve group decision-making problems such as transfer pricing and capital
budgeting problems. The methodologies used in this chapter are Analytic Hierarchy
Process, multiple criteria and multiple constraint linear programming, and balanced
scorecard and business performance evaluation.
Keywords: Analytic hierarchy process, Multiple criteria and multiple constraint
linear programming, Business performance evaluation, Activity-based costing sys-
tem, Group decision making, Optimal trade-offs, Balanced scorecard, Transfer
pricing, Capital budgeting

Chapter 30: Statistics Methods Applied in Employee Stock Options

This study provides model-based and compensation-based approaches to price

subjective value of employee stock options (ESOs). In model-based approach, we
consider a utility-maximizing model that the employee allocates his wealth
among the company stock, market portfolio, and risk-free bond, and then
derive the ESO formulae which take into account illiquidity and sentiment
effects. By using the method of change of measure, the derived formulae are
simply like that of the market values with altered parameters. To calculate
compensation-based subjective value, we group employees by hierarchical clus-
tering with K-Means approach and back out the option value in an equilibrium
competitive employment market. Further, we test illiquidity and sentiment effects
on ESO values by running the regressions which consider the problem of standard
errors in finance panel data.
Keywords: Employee stock option, Sentiment, Subjective value, Illiquidity,
Change of measure, Hierarchical clustering with K-Means approach, Standard
errors in finance panel data, Exercise boundary, Jump diffusion model

Chapter 31: Structural Change and Monitoring Tests

This chapter focuses on various structural change and monitoring tests for a class of
widely used time-series models in economics and finance, including I(0), I(1), I(d)
processes and the co-integration relationship. In general, structural change tests can be
categorized into two types: One is the classical approach to testing for structural
change, which employs retrospective tests using a historical data set of a given length;
the other one is the fluctuation-type test in a monitoring scheme, which means for
given a history period for which a regression relationship is known to be stable,
1 Introduction to Financial Econometrics and Statistics 31

we then test whether incoming data are consistent with the previously established
relationship. Several structural changes such as CUSUM squared tests, the QLR test,
the prediction test, the multiple break test, bubble tests, co-integration breakdown
tests, and the monitoring fluctuation test are discussed in this chapter, and we further
illustrate all details and usefulness of these tests.
Keywords: Co-integration breakdown test, Structural break, Long memory
process, Monitoring fluctuation test, Boundary function, CUSUM squared test,
Prediction test, Bubble test, Unit root time series, Persistent change

Chapter 32: Consequences of Option Pricing of a Long Memory

in Volatility

This chapter use conditionally heteroskedastic time-series models to describe the

volatility of stock index returns. Volatility has a long memory property in the
most general models and then the autocorrelations of volatility decay at a hyperbolic
rate; contrasts are made with popular, short memory specifications whose autocorrela-
tions decay more rapidly at a geometric rate. Options are valued for ARCH
volatility models by calculating the discounted expectations of option payoffs for an
appropriate risk-neutral measure. Monte Carlo methods provide the expectations. The
speed and accuracy of the calculations is enhanced by two variance reduction methods,
which use antithetic and control variables. The economic consequences of a long
memory assumption about volatility are documented, by comparing implied volatilities
for option prices obtained from short and long memory volatility processes.
Keywords: ARCH models, Implied volatility, Index options, Likelihood max-
imization, Long memory, Monte Carlo, Option prices, Risk-neutral pricing, Smile
shapes, Term structure, Variance reduction methods

Chapter 33: Seasonal Aspects of Australian Electricity Market

This chapter develops econometric models for seasonal patterns in both price returns
and proportional changes in demand for Australian electricity. Australian Electricity
spot prices differ considerably from equity spot prices in that they contain an extremely
rapid mean reversion process. The electricity spot price could increase to a market cap
price of AU$12,500 per Megawatt Hour (MWh) and revert back to a mean level
(AUD$30) within a half hour interval. This has implications for derivative pricing and
risk management. We also model extreme spikes in the data. Our study identifies both
seasonality effects and dramatic price reversals in the Australian electricity market.
The pricing seasonality effects include time-of-day, day-of-week, monthly, and yearly
effects. There is also evidence of seasonality in demand for electricity.
Keywords: Electricity, Spot price, Seasonality, Outlier, Demand, Econometric
modeling
32 C.-F. Lee and J.C. Lee

Chapter 34: Pricing Commercial Timberland Returns in the

United States

This chapter uses both parametric and nonparametric approaches to evaluate private-
and public-equity timberland investments in the United States. Private-equity timber-
land returns are proxied by the NCREIF Timberland Index, whereas public-equity
timberland returns are proxied by the value-weighted returns on a dynamic portfolio
of the US publicly traded forestry firms that had or have been managing timberlands.
Static estimations of the capital asset-pricing model and Fama-French three-factor
model are obtained by ordinary least squares, whereas dynamic estimations are
obtained by state-space specifications with the Kalman filter. In estimating the
stochastic discount factors, linear programming is used.
Keywords: Alternative asset class, Asset pricing, Evaluation, Fama-French
three-factor model, Nonparametric analysis, State-space model, Stochastic discount
factor, Timberland investments, Time series, Time-varying parameter

Chapter 35: Optimal Orthogonal Portfolios with Conditioning

Information

This chapter derives and characterizes optimal orthogonal portfolios in the presence
of conditioning information in the form of a set of lagged instruments. In this
setting, studied by Hansen and Richard (1987), the conditioning information is used
to optimize with respect to the unconditional moments. We present an empirical
illustration of the properties of the optimal orthogonal portfolios. The methodology in
this chapter includes regression and maximum likelihood parameter estimation, as
well as method of moments estimation. We form maximum likelihood estimates of
nonlinear functions as the functions evaluated at the maximum likelihood parameter
estimates.
Keywords: Asset-pricing tests, Conditioning information, Minimum variance
efficiency, Optimal portfolios, Predicting returns, Portfolio management, Stochas-
tic discount factors, Generalized, Method of moments, Maximum likelihood, Para-
metric bootstrap, Sharpe ratios

Chapter 36: Multi-factor, Multi-indicator Approach to Asset Pricing:

Method and Empirical Evidence

This chapter uses a multifactor, multi-indicator approach to test the capital asset-
pricing model (CAPM) and the arbitrage pricing theory (APT). This approach is
able to solve the measuring problem in the market portfolio in testing CAPM, and it
is also able to directly test APT by linking the common factors to the macroeco-
nomic indicators. We propose a MIMIC approach to test CAPM and APT. The beta
estimated from the MIMIC model by allowing measurement error on the market
1 Introduction to Financial Econometrics and Statistics 33

portfolio does not significantly improve the OLS beta, while the MLE estimator
does a better job than the OLS and GLS estimators in the cross-sectional regressions
because the MLE estimator takes care of the measurement error in beta. Therefore,
the measurement error problem on beta is more serious than that on the market
portfolio.
Keywords: Capital asset-pricing model, CAPM, Arbitrage pricing theory, Mul-
tifactor multi-indicator approach, MIMIC, Measurement error, LISREL approach,
Ordinary least square, OLS, General least square, GLS, Maximum likelihood
estimation, MLE

Chapter 37: Binomial OPM, Black-Scholes OPM and Their

Relationship: Decision Tree and Microsoft Excel Approach

This chapter will first demonstrate how Microsoft Excel can be use to create the
Decision Trees for the Binomial Option Pricing Model. At the same time, this
chapter will discuss the Binomial Option Pricing Model in a less mathematical
fashion. All the mathematical calculations will be taken care by the Microsoft Excel
program that is presented in this chapter. Finally, this chapter also uses the Decision
Tree approach to demonstrate the relationship between the Binomial Option Pricing
Model and the Black-Scholes Option Pricing Model.
Keywords: Binomial option pricing model, Decision trees, Black-Sholes option
pricing model, Call option, Put option, Microsoft Excel, Visual Basic for applica-
tions, VBA, Put-call parity, Sigma, Volatility, Recursive programming

Chapter 38: Dividend Payments and Share Repurchases

of U.S. Firms: An Econometric Approach

This chapter uses the econometric methodology to deal with the dynamic inter-
relationships between dividend payments and share repurchases and investigate
endogeneity of certain explanatory variables. Identification of the model parameters
is achieved in such models by exploiting the cross-equations restrictions on the
coefficients in different time periods. Moreover, the estimation entails using
nonlinear optimization methods to compute the maximum likelihood estimates of
the dynamic random-effects models and for testing statistical hypotheses using
likelihood ratio tests. This study also highlights the importance of developing
comprehensive econometric models for these interrelationships. It is common in
finance research to spell out “specific hypotheses” and conduct empirical research
to investigate validity of the hypotheses.
Keywords: Compustat database, Corporate policies, Dividends, Dynamic
random-effects models, Econometric methodology, Endogeneity, Maximum like-
lihood, Intangible assets, Model formulation, Nonlinear optimization, Panel data,
Share repurchases
34 C.-F. Lee and J.C. Lee

Chapter 39: Term Structure Modeling and Forecasting Using

the Nelson-Siegel Model

In this chapter, we illustrate some recent developments in the yield curve modeling
by introducing a latent factor model called the dynamic Nelson-Siegel model. This
model not only provides good in-sample fit, but also produces superior out-of-sample
performance. Beyond Treasury yield curve, the model can also be useful for other
assets such as corporate bond and volatility. Moreover, the model also suggests
generalized duration components corresponding to the level, slope, and curvature
risk factors. The dynamic Nelson-Siegel model can be estimated via a one-step
procedure, like the Kalman filter, which can also easily accommodate other variables
of interests. Alternatively, we could estimate the model through a two-step process by
fixing one parameter and estimating with ordinary least squares. The model is flexible
and capable of replicating a variety of yield curve shapes: upward sloping, downward
sloping, humped, and inverted humped. Forecasting the yield curve is achieved
through forecasting the factors and we can impose either a univariate autoregressive
structure or a vector autoregressive structure on the factors.
Keywords: Term structure, Yield curve, Factor model, Nelson-Siegel curve,
State-space model

Chapter 40: The Intertemporal Relation Between Expected Return

and Risk On Currency

The literature has so far focused on the risk-return trade-off in equity markets and
ignored alternative risky assets. This chapter examines the presence and signifi-
cance of an intertemporal relation between expected return and risk in the foreign
exchange market. This chapter tests the existence and significance of a daily risk-
return trade-off in the FX market based on the GARCH, realized, and range
volatility estimators. Our empirical analysis relies on the maximum likelihood
estimation of the GARCH-in-mean models, as described in Appendix A. We also
use the seemingly unrelated (SUR) regressions and panel data estimation to inves-
tigate the significance of a time-series relation between expected return and risk on
currency.
Keywords: GARCH, GARCH-in-mean, Seemingly unrelated regressions
(SUR), Panel data estimation, Foreign exchange market, ICAPM, High-frequency
data, Time-varying risk aversion, High-frequency data, Daily realized volatility

Chapter 41: Quantile Regression and Value-at-Risk

This chapter studies quantile regression (QR) estimation of Value-at-Risk (VaR).

VaRs estimated by the QR method display some nice properties. In this
chapter, different QR models in estimating VaRs are introduced. In particular,
VaR estimation based on quantile regression of the QAR models, Copula models,
1 Introduction to Financial Econometrics and Statistics 35

ARCH models, GARCH models, and the CaViaR models is systematically intro-
duced. Comparing the proposed QR method with traditional methods based on
distributional assumptions, the QR method has the important property that it is
robust to non-Gaussian distributions. Quantile estimation is only influenced by the
local behavior of the conditional distribution of the response near the specified
quantile. As a result, the estimates are not sensitive to outlier observations. Such
a property is especially attractive in financial applications since many financial data
like, say, portfolio returns (or log returns), are usually not normally distributed. To
highlight the importance of the QR method in estimating VaR, we apply the QR
techniques to estimate VaRs in International Equity Markets. Numerical evidence
indicates that QR is a robust estimation method for VaR.
Keywords: ARCH, Copula, GARCH, Non-normality, QAR, Quantile regres-
sion, Risk management, Robust estimation, Time series, Value-at-risk

Chapter 42: Earnings Quality and Board Structure:

Evidence from South East Asia

Using a sample of listed firms in Southeast Asia countries, this chapter examines the
association among board structure and corporate ownership structure in affecting
earnings quality. The econometric method employed is regressions of panel data.
In a panel data setting, I address both cross-sectional and time-series dependence.
Following Gow et al. (2010), I employ the two-way clustering method where the
standard errors are clustered by both firm and year in my regressions of panel data.
Keywords: Earnings quality, Board structure, Corporate ownership structure,
Panel data regressions, Cross-sectional and time-series dependence, Two-way
clustering method of standard errors

Chapter 43: Rationality and Heterogeneity of Survey Forecasts

of the Yen-Dollar Exchange Rate: A Reexamination

This chapter examines the rationality and diversity of industry-level forecasts of the
yen-dollar exchange rate collected by the Japan Center for International Finance.
We compare three specifications for testing rationality: the “conventional” bivariate
regression, the univariate regression of a forecast error on a constant and other
information set variables, and an error correction model (ECM). We extend the
analysis of industry-level forecasts to a SUR-type structure using an innovative
GMM technique (Bonham and Cohen 2001) that allows for forecaster cross-
correlation due to the existence of common shocks and/or herd effects. Our
GMM tests of micro-homogeneity uniformly reject the hypothesis that forecasters
exhibit similar rationality characteristics.
Keywords: Rational expectations, Unbiasedness, Weak efficiency, Micro-
homogeneity, Heterogeneity, Exchange rate, Survey forecasts, Aggregation bias,
GMM, SUR
36 C.-F. Lee and J.C. Lee

Chapter 44: Stochastic Volatility Structures and Intra-day

Asset Price Dynamics

This chapter uses conditional volatility estimators as special cases of a general

stochastic volatility structure. The theoretical asymptotic distribution of the measure-
ment error process for these estimators is considered for particular features observed
in intraday financial asset price processes. Specifically, I consider the effects of
(i) induced serial correlation in returns processes, (ii) excess kurtosis in the underly-
ing unconditional distribution of returns, (iii) market anomalies such as market
opening and closing effects, and (iv) failure to account for intraday trading patterns.
These issues are considered with applications in option pricing/trading strategies and
the constant/dynamic hedging frameworks in mind. The methodologies used in this
chapter are ARCH, maximum likelihood method, and unweighted GARCH.
Keywords: ARCH, Asymptotic distribution, Autoregressive parameters, Con-
ditional variance estimates, Constant/dynamic hedging, Excess kurtosis, Index
futures, Intraday returns, Market anomalies, Maximum likelihood estimates,
Misspecification, Mis-specified returns, Persistence, Serial correlation, Stochastic
volatility, Stock/futures, Unweighted GARCH, Volatility co-persistence

Chapter 45: Optimal Asset Allocation Under VaR Criterion:

Taiwan Stock Market

This chapter examines the riskiness of the Taiwan stock market by determining the
VaR from the expected return distribution generated by historical simulation.
Value-at-risk (VaR) measures the worst expected loss over a given time horizon
under normal market conditions at a specific level of confidence. VaR is determined
by the left tail of the cumulative probability distribution of expected returns. Our
result indicates the cumulative probability distribution has a fatter left tail, com-
pared with the left tail of a normal distribution. This implies a riskier market. We
also examined a two-sector asset allocation model subject to a target VaR con-
straint. The VaR-efficient frontier of the TAIEX traded stocks recommended,
mostly, a corner portfolio.
Keywords: Value-at-risk, Asset allocation, Cumulative probability distribution,
Normal distribution, VaR-efficient frontier, Historical simulation, Expected return dis-
tribution, Two-sector asset allocation model, Delta, gamma, Corner portfolio, TAIEX

Chapter 46: Alternative Methods for Estimating Firm’s Growth Rate

The most common valuation model is the dividend growth model. The growth rate is
found by taking the product of the retention rate and the return on equity. What is less well
understood are the basic assumptions of this model. In this paper, we demonstrate that the
model makes strong assumptions regarding the financing mix of the firm. In addition,
1 Introduction to Financial Econometrics and Statistics 37

we discuss several methods suggested in the literature on estimating growth rates and
analyze whether these approaches are consistent with the use of using a constant discount
rate to evaluate the firm’s assets and equity. The literature has also suggested estimating
growth rate by using the average percentage change method, compound-sum method,
and/or regression methods. We demonstrate that the average percentage change is very
sensitive to extreme observations. Moreover, on average, the regression method yields
similar but somewhat smaller estimates of the growth rate compared to the compound-
sum method. We also discussed the inferred method suggested by Gordon and Gordon
(1997) to estimate the growth rate. Advantages, disadvantages, and the interrelationship
among these estimation methods are also discussed in detail.
Keywords: Compound sum method, Discount cash flow model, Growth rate,
Internal growth rate, Sustainable growth rate

Chapter 47: Econometric Measures of Liquidity

A security is liquid to the extent that an investor can trade significant quantities of the
security quickly, at or near the current market price, and bearing low transaction costs.
As such, liquidity is a multidimensional concept. In this chapter, I review several
widely used econometrics or statistics-based measures that researchers have devel-
oped to capture one or more dimensions of a security’s liquidity (i.e., limited depen-
dent variable model (Lesmond et al. 1999) and autocovariance of price changes (Roll
1984)). These alternative proxies have been designed to be estimated using either
low-frequency or high-frequency data, so I discuss four liquidity proxies that are
estimated using low-frequency data and two proxies that require high-frequency data.
Low-frequency measures permit the study of liquidity over relatively long time
horizons; however, they do not reflect actual trading processes. To overcome this
limitation, high-frequency liquidity proxies are often used as benchmarks to determine
the best low-frequency proxy. In this chapter, I find that estimates from the effective
tick measure perform best among the four low-frequency measures tested.
Keywords: Liquidity, Transaction costs, Bid-ask spread, Price impact, Percent
effective spread, Market model, Limited dependent variable model, Tobin’s model,
Log-likelihood function, Autocovariance, Correlation analysis

Chapter 48: A Quasi-Maximum Likelihood Estimation Strategy

for Value-at-Risk Forecasting: Application to Equity Index
Futures Markets

The chapter uses GARCH model and quasi-maximum likelihood estimation

strategy to investigate equity index futures markets. We present the first empirical
evidence for the validity of the ARMA-GARCH model with tempered
stable innovations to estimate 1-day-ahead value-at-risk in futures markets for the
S&P 500, DAX, and Nikkei. We also provide empirical support that GARCH
38 C.-F. Lee and J.C. Lee

models based on the normal innovations appear not to be as well suited as infinitely
divisible models for predicting financial crashes. In our empirical analysis, we
forecast 1 % value-at-risk in both spot and futures markets using normal and
tempered stable GARCH models following a quasi-maximum likelihood estimation
strategy. In order to determine the accuracy of forecasting for each specific model,
backtesting using Kupiec’s proportion of failures test is applied.
Keywords: Infinitely divisible models, Tempered stable distribution, GARCH
models, Value-at-risk, Kupiec’s proportion of failures test, Quasi-maximum likeli-
hood estimation strategy

Chapter 49: Computer Technology for Financial Service

This chapter examines the core computing competence for financial services.
Securities trading is one of the few business activities where a few seconds of
processing delay can cost a company big fortune. Grid and Cloud computing
will be briefly described. How the underlying algorithm for financial analysis
can take advantage of Grid environment is chosen and presented. One of the
most popular practiced algorithms Monte Carlo Simulation is used in our cases
study for option pricing and risk management. The various distributed com-
putational platforms are carefully chosen to demonstrate the performance issue
for financial services.
Keywords: Financial service, Grid and cloud computing, Monte Carlo simula-
tion, Option pricing, Risk management, Cyberinfrastructure, Random number
generation, High end computing, Financial simulation, Information technology

Chapter 50: Long-Run Stock Return and the Statistical Inference

This chapter introduces the long-run stock return methodologies and their statistical
inference. The long-run stock return is usually computed by using a holding strategy
more than 1 year but up to 5 years. Two categories of long-run return methods are
illustrated in this chapter: the event-time approach and calendar-time approach. The
event-time approach includes cumulative abnormal return, buy-and-hold abnormal
return, and abnormal returns around earnings announcements. In former two methods,
it is recommended to apply the empirical distribution (from the bootstrapping method)
to examine the statistical inference, whereas the last one uses classical t-test.
In addition, the benchmark selections in the long-run return literature are introduced.
Moreover, the calendar-time approach contains mean monthly abnormal return,
factor models, and Ibbotson’s RATS, which could be tested by time-series volatility.
Keywords: Long-run stock return, Buy-and-hold return, Factor model, Event-
time, Calendar-time, Cumulative abnormal return, Ibottson’s RATS, Conditional
market model, Bootstrap, Zero-investment portfolio
1 Introduction to Financial Econometrics and Statistics 39

Chapter 51: Value-at-Risk Estimation via a Semi-Parametric

Approach: Evidence from the Stock Markets

This study utilizes the parametric approach (GARCH-based models) and the semi-
parametric approach of Hull and White (1998) (HW-based models) to estimate the
Value-at-Risk (VaR) through the accuracy evaluation of accuracy for the eight stock
indices in Europe and Asia stock markets. The measure of accuracy includes the
unconditional coverage test by Kupiec (1995) as well as two loss functions, quadratic
loss function and unexpected loss. As to the parametric approach, the parameters of
generalized autoregressive conditional heteroskedasticity (GARCH) model are esti-
mated by the method of maximum likelihood and the quantiles of asymmetric distri-
bution like skewed generalized student’s t (SGT) can be solved by composite trapezoid
rule. Sequentially, the VaR is evaluated by the framework proposed by Jorion (2000).
Turning to the semi-parametric approach of Hull and White (1998), before performing
the traditional historical simulation, the raw return series is scaled by a volatility ratio
where the volatility is estimated by the same procedure of parametric approach.
Keywords: Value-at-risk, Semi-parametric approach, Parametric approach,
Generalized autoregressive conditional heteroskedasticity, Skewed generalized
student’s t, Composite trapezoid rule, Method of maximum likelihood, Uncondi-
tional coverage test, Loss function

Chapter 52: Modeling Multiple Asset Returns by a Time-Varying

t Copula Model

This chapter illustrates a framework to model joint distributions of multiple asset

returns using a time-varying Student’s t copula model. We model marginal distri-
butions of individual asset returns by a variant of GARCH models and then use
a Student’s t copula to connect all the margins. To build a time-varying structure for
the correlation matrix of t copula, we employ a dynamic conditional correlation
(DCC) specification. We illustrate the two-stage estimation procedures for the
model and apply the model to 45 major US stocks returns selected from nine
sectors. As it is quite challenging to find a copula function with very flexible
parameter structure to account for difference dependence features among all pairs
of random variables, our time-varying t copula model tends to be a good working
tool to model multiple asset returns for risk management and asset allocation
purposes. Our model can capture time-varying conditional correlation and some
degree of tail dependence, while it also has limitations of featuring symmetric
dependence and inability of generating high tail dependence when being used to
model a large number of asset returns.
Keywords: Student’s t copula, GARCH models, Asset returns, U.S. stocks,
Maximum likelihood, Two-stage estimation, Tail dependence, Exceedance corre-
lation, Dynamic conditional correlation, Asymmetric dependence
40 C.-F. Lee and J.C. Lee

Chapter 53: Internet Bubble Examination with Mean-Variance Ratio

This chapter illustrates the superiority of the mean-variance ratio (MVR) test
over the traditional SR test by applying both tests to analyze the performance
of the S&P 500 index and the NASDAQ 100 index after the bursting of the
Internet bubble in 2000s. This shows the superiority of the MVR test statistic
in revealing short-term performance and, in turn, enables investors to make
better decisions in their investments. The methodologies used in this chapter
are mean-variance ratio, Sharpe ratio, hypothesis testing, and uniformly most
powerful unbiased test.
Keywords: Mean-variance ratio, Sharpe ratio, Hypothesis testing, Uniformly
most powerful unbiased test, Internet bubble, Fund management

Chapter 54: Quantile Regression in Risk Calibration

This chapter uses the CoVaR (Conditional VaR) framework to obtain accurate
information on the interdependency of risk factors. The basic technical ele-
ments of CoVaR estimation are two levels of quantile regression: one on
market risk factors; another on individual risk factor. Tests on the functional
form of the two-level quantile regression reject the linearity. A flexible semi-
parametric modeling framework for CoVaR is proposed. A partial linear model
(PLM) is analyzed. In applying the technology to stock data covering the crisis
period, the PLM outperforms in the crisis time, with the justification of the
backtesting procedures. Moreover, using the data on global stock markets
indices, the analysis on marginal contribution of risk (MCR) defined as the
local first order derivative of the quantile curve sheds some light on the source
of the global market risk.
Keywords: CoVAR, Value-at-risk, Quantile regression, Locally linear quantile
regression, Partial linear model, Semi-parametric model

Chapter 55: Strike Prices of Options for Overconfident Executives

This chapter uses Monte Carlo simulation to investigate the impacts of managerial
overconfidence on the optimal strike prices of executive incentive options.
Although it has been shown that optimally managerial incentive options should
be awarded in-the-money, in practice, most firms award them at-the-money. We
show that the optimal strike prices of options granted to overconfident executive are
directly related to their overconfidence level, and that this bias brings the optimal
strike prices closer to the institutionally prevalent at-the-money prices. The Monte
Carlo simulation procedure uses a Mathematica program to find the optimal effort
by managers and the optimal (for stockholders) contract parameters. An expanded
discussion of the simulations, including the choice of the functional forms and the
calibration of the parameters, is provided.
1 Introduction to Financial Econometrics and Statistics 41

Keywords: Overconfidence, Managerial effort, Incentive options, Strike price,

Simulations, Behavioral finance, Executive compensation schemes, Mathematica
optimization, Risk aversion, Effort aversion

Chapter 56: Density and Conditional Distribution Based

Specification Analysis

This chapter uses densities and conditional distributions analysis to carry out
consistent specification testing and model selection among multiple diffusion
processes. In this chapter, we discuss advances to this literature introduced by
Corradi and Swanson (2005), who compare the cumulative distribution (marginal
or joint) implied by a hypothesized null model with corresponding empirical
distributions of observed data. In particular, parametric specification tests in the
spirit of the conditional Kolmogorov test of Andrews (1997) that rely on block
bootstrap resampling methods in order to construct test critical values are discussed.
The methodologies used in this chapter are continuous time simulation methods,
single process specification testing, multiple process model selection, and multi-
factor diffusion process, block bootstrap, and jump process.
Keywords: Multifactor diffusion process, Specification test, Out-of-sample
forecasts, Conditional distribution, Model selection, Block bootstrap, Jump process

Chapter 57: Assessing the Performance of Estimators Dealing

with Measurement Errors

This chapter describes different procedures to deal with measurement error in linear
models, and assess their performance in finite samples using Monte Carlo simulations,
and data on corporate investment. We consider the standard instrumental variables
approach proposed by Griliches and Hausman (1986) as extended by Biorn
(2000) [OLS-IV], the Arellano and Bond (1991) instrumental variable estimator, and
the higher-order moment estimator proposed by Erickson and Whited (2000, 2002).
Our analysis focuses on characterizing the conditions under which each of these
estimators produces unbiased and efficient estimates in a standard “errors in variables”
setting. In the presence of fixed effects, under heteroscedasticity, or in the absence of
a very high degree of skewness in the data, the EW estimator is inefficient and
returns biased estimates for mismeasured and perfectly measured regressors. In con-
trast to the EW estimator, IV-type estimators (OLS-IV and AB-GMM) easily handle
individual effects, heteroskedastic errors, and different degrees of data skewness.
The IV approach, however, requires assumptions about the autocorrelation structure
of the mismeasured regressor and the measurement error. We illustrate the application
of the different estimators using empirical investment models.
Keywords: Investment equations, Measurement error, Monte Carlo simulations,
Instrumental variables, GMM, Bias, Fixed effects, Heteroscedasticity, Skewness,
High-order moments
42 C.-F. Lee and J.C. Lee

Chapter 58: Realized Distributions of Dynamic Conditional

Correlation and Volatility Thresholds in the Crude Oil, Gold,
and Dollar/Pound Currency Markets

This chapter proposes a modeling framework for the study of co-movements in

price changes among crude oil, gold, and dollar/pound currencies that are condi-
tional on volatility regimes. Methodologically, we extend the Dynamic Conditional
Correlation (DCC) multivariate GARCH model to examine the volatility and
correlation dynamics depending on the variances of price returns involving
a threshold structure. The results indicate that the periods of market turbulence
are associated with an increase in co-movements in commodity (gold and oil)
prices. The results imply that gold may act as a safe haven against major currencies
when investors face market turmoil.
Keywords: Dynamic conditional correlation, Volatility threshold, Realized
distribution, Currency market, Gold, Oil

Chapter 59: Pre-IT Policy, Post-IT Policy, and the Real Sphere
in Turkey

We estimate Two SVECM (Structural Vector Error Correction) Models for the
Turkish economy based on imposing short run and Long-run restrictions that
accounts for examining the behavior of the real sphere in the Pre-IT policy
(before Inflation-Targeting adoption) and Post-IT policy (after Inflation-
Targeting Adoption). Responses reveals that an expansionary interest policy
shock leads to a decrease in price level, a fall in output, an appreciation in the
exchange rate, an improvement in the share prices in the very short run for the
most of Pre-IT period.
Keywords: SVECM models, Turkish economy, Short run, Long run, Restric-
tions, Inflation targeting, Pre-IT policy, Post-IT policy, Share prices, Exchange rate,
Monetary policy shock, Output, Price level, Real sphere

Chapter 60: Determination of Capital Structure: A LISREL Model

Approach

In this chapter, we employ structural equation modeling (SEM) in LISREL system

to solve the measurement errors problems in the analysis of the determinants of
capital structure and find the important factors consistent with capital structure
theory by using date from 2002 to 2010. The purpose of this chapter is to investigate
whether the influences of accounting factors on capital structure change and
whether the important factors are consistent with the previous literature. The
methodologies discussed in this chapter are structural equation modeling (SEM),
multiple indicators and multiple causes (MIMIC) model, LISREL system, simul-
taneous equations, and SEM with confirmatory factor analysis (CFA) approach.
1 Introduction to Financial Econometrics and Statistics 43

Keywords: Capital structure, Structural equation modeling (SEM), Multiple indi-

cators and multiple causes (MIMIC) model, LISREL system, Simultaneous equations,
Latent variable, Determinants of capital structure, Error in variable problem

Chapter 61: Evaluating the Effectiveness of Futures Hedging

This chapter examines the Ederington hedging effectiveness (EHE) comparisons

between unconditional OLS hedge strategy and other conditional hedge strategies.
It is shown that OLS hedge strategy outperforms most of the optimal conditional
hedge strategies when EHE is used as the hedging effectiveness criteria. Before
concluding that OLS hedge is better than the others; however, we need to understand
under what circumstances the result is derived. We explain why OLS is the best hedge
strategy under EHE criteria in most cases, and how most conditional hedge strategies
are judged as inferior to OLS hedge strategy by an EHE comparison.
Keywords: Futures hedging, Portfolio management, Ederington hedging effec-
tiveness, Variance estimation, Unconditional variance, Conditional variance, OLS
hedging strategy, GARCH hedging strategy, Regime-switching hedging strategy,
Utility-based hedging strategy

Chapter 62: Evidence on Earning Management by Integrated Oil

and Gas Companies

This chapter uses Jones Model (1991) which projects the expected level of discre-
tionary accruals and demonstrates specific test methodology for detection of earnings
management in the oil and gas industry. This study utilized several parametric and
nonparametric statistical methods to test for such earnings management. By compar-
ing actuals versus projected accruals, we are able to compute the total unexpected
accruals. We also correlate unexpected total accruals with several difficult to manip-
ulate indicators that reflect company’s level of activities.
Keywords: Earning management, Jones (1991) model, Discretionary accruals,
Income from operations, Nonrecurring items, Special items, Research and devel-
opment expense, Write-downs, Political cost, Impression management, Oil and gas
industry

Chapter 63: A Comparative Study of Two Models SV with MCMC

Algorithm

This chapter examines two asymmetric stochastic volatility models used to describe
the volatility dependencies found in most financial returns. The first is the
autoregressive stochastic volatility model with Student’s t-distribution (ARSV-t),
and the second is the basic Svol of JPR (1994). In order to estimate these models,
our analysis is based on the Markov Chain Monte Carlo (MCMC) method. Therefore,
44 C.-F. Lee and J.C. Lee

the technique used is a Metropolishastings (Hastings 1970), and the Gibbs sampler
(Casella and George 1992; Gelfand and Smith 1990; Gilks et al. 1993). The empirical
results concerned on the Standard and Poor’s 500 composite Index (S&P), CAC40,
Nasdaq, Nikkei, and Dow-Jones stock price indexes reveal that the ARSV-t model
provides a better performance than the Svol model on the Mean Squared Error (MSE)
and the Maximum Likelihood function.
Keywords: Autoregression, Asymmetric stochastic volatility, MCMC, Metro-
polishastings, Gibbs sampler, Volatility dependencies, Student’s t-distribution,
SVOL, MSE, Financial returns, Stock price indexes

Chapter 64: Internal Control Material Weakness, Analysts’ Accuracy

and Bias, and Brokerage Reputation

This chapter uses the Ordinary Least-Squares (OLS) methodology in the main tests
to examine the impact of internal control material weaknesses (ICMW hereafter) on
sell side analysts. We match our ICMW firms with non-ICMWs based on industry,
sales, and assets. We re-estimate the models using rank regression technique to
assess the sensitivity of the results to the underlying functional form assumption
made by OLS. We use Cook’s distance to test the outliers.
Keywords: Internal control material weakness, Analyst forecast accuracy, Ana-
lyst forecast bias, Brokerage reputation, Sarbanes-Oxley act, Ordinary least squares
regressions, Rank regressions, Fixed effects, Matching procedure, Cook’s distance

Chapter 65: What Increases Banks’ Vulnerability to Financial Crisis:

Short-Term Financing or Illiquid Assets?

This chapter applies Logit and OLS econometric techniques to analyze the Federal
Reserve Y-9C report data. We show that short-term financing is a response to the adverse
economic shocks rather than a cause of the recent crisis. The likelihood of financial crisis
actually stems from the illiquidity and low creditworthiness of the investment. Our
results are robust to endogeneity concerns when we use a difference-in-differences
(DiD) approach with the Lehman bankruptcy in 2008 proxying for an exogenous shock.
Keywords: Financial crisis, Short-term financing, Debt maturity, Liquidity risk,
Deterioration of bank asset quality

Chapter 66: Accurate Formulae for Evaluating Barrier Options with

Dividends Payout and the Application in Credit Risk Valuation

This chapter approximates the discrete dividend payout by a stochastic continuous

dividend yield, so the post dividend stock price process can be approximated by
another log-normally diffusive stock process with a stochastic continuous payout
ratio up to the ex-dividend date. Accurate approximation analytical pricing
1 Introduction to Financial Econometrics and Statistics 45

formulae for barrier options are derived by repeatedly applying the reflection
principle. Besides, our formulae can be applied to extend the applicability of the
first passage model – a branch of structural credit risk model. The stock price falls
due to the dividend payout in the option pricing problem is analog to selling the
firm’s asset to finance the loan repayment or dividend payout in the first passage
model. Thus, our formulae can evaluate vulnerable bonds or the equity values given
that the firm’s future loan/dividend payments are known.
Keywords: Barrier option, Option pricing, Stock option, Dividend, Reflection
principle, Lognormal, Credit risk

Chapter 67: Pension Funds: Financial Econometrics on the Herding

Phenomenon in Spain and the United Kingdom

This chapter uses the estimated cross-sectional standard deviations of betas to

analyze if manager’s behavior enhances the existence of herding phenomena and
the impact of the Spanish and UK pension funds investment on the market effi-
ciency. We also estimate the betas with an econometric technique less applied in the
financial literature: state-space models and the Kalman filter. Additionally, in order
to obtain a robust estimation, we apply the Huber estimator. Finally, we apply
several models and study the existence of herding toward the market, size, book-to-
market, and momentum factors.
Keywords: Herding, Pension funds, State-space models, Kalman filter, Huber
estimation, Imitation, Behavioral finance, Estimated cross-sectional standard devi-
ations of betas, Herding toward the market, Herding toward size factor, Herding
toward book-to-market factor, and Herding toward momentum factor

Chapter 68: Estimating the Correlation of Asset Returns: A Quantile

Dependence Perspective

This chapter uses the Copula Quantile-on-Quantile Regression (C-QQR) approach

to construct the correlation between the conditional quantiles of stock returns. This
new approach of estimating correlation utilizes the idea that the condition of a stock
market is related to its return performance, particularly to the conditional quantile
of its return, as the lower return quantiles reflect a weak market while the upper
quantiles reflect a bullish one. The C-QQR approach uses the copula to generate
a regression function for modeling the dependence between the conditional
quantiles of the stock returns under consideration. It is estimated using a two-step
quantile regression procedure, where in principle, the first step is implemented to
model the conditional quantile of one stock return, which is then related in the
second step to the conditional quantile of another return.
Keywords: Stock markets, Copula, Correlation, Quantile regression, Quantile
dependence, Business cycle, Dynamics, Risk management, Investment, Tail risk,
Extreme events, Market uncertainties
46 C.-F. Lee and J.C. Lee

Chapter 69: Multi-criteria Decision Making for Evaluating Mutual

Funds Investment Strategies

This chapter uses the criteria measurements to evaluate investment style and
investigate multiple criteria decision-making (MCDM) problem. To achieve this
objective, first, we employ factor analysis to extract independent common factors
from those criteria. Second, we construct the evaluation frame using hierarchical
system composed of the above common factors with evaluation criteria, and then
derive the relative weights with respect to the considered criteria. Third, the
synthetic utility value corresponding to each investment style is aggregated by
the weights with performance values. Finally, we compare with empirical data and
find that the model of MCDM predicts the rate of return.
Keywords: Investment strategies, Multiple Criteria Decision Making (MCDM),
Hierarchical system, Investment style, Factor analysis, Synthetic utility value,
Performance values

Chapter 70: Econometric Analysis of Currency Carry Trade

This chapter investigates carry trade strategy in the currency markets whereby
investors fund positions in high interest rate currencies by selling low interest rate
currencies to earn the interest rate differential. In this chapter, we first provide an
overview of the risk and return profile of currency carry trade; second, we introduce
two popular models, the regime-switch model and the logistic smooth transition
regression model, to analyze carry trade returns because the carry trade returns are
highly regime dependent. Finally, an empirical example is illustrated.
Keywords: Carry trade, Uncovered interest parity, Markov chain Monte Carlo,
Regime-switch model, Logistic smooth transition regression model

Chapter 71: Analytical Bounds for Treasury Bond Futures Prices

This study employs a maximum likelihood estimation technique presented by Chen

and Scott (1993) to estimate the parameters for two-factor Cox-Ingersoll-Ross
models of the term structure. Following the estimation, the factor values are solved
for by matching the short rate with the cheapest-to-deliver bond price. Then, upper
bounds and lower bounds for Treasury bond futures prices can be calculated. This
study first shows that the popular preference-free, closed-form cost of carry model
is an upper bound for the Treasury bond futures price. Then, the next step is to
derive analytical lower bounds for the futures price under one- and two-factor
Cox-Ingersoll-Ross models of the term structure.
Keywords: Treasury bond futures, Delivery options, Cox-Ingersoll-Ross
models, Bounds, Maximum likelihood estimation, Term structure, Cheapest-to-
deliver bond, Timing options, Quality options, Chicago board of trade
1 Introduction to Financial Econometrics and Statistics 47

Chapter 72: Rating Dynamics of Fallen Angels and Their Speculative

Grade-Rated Peers: Static Versus Dynamic Approach

This study adopts the survival analysis framework (Allison 1984) to examine
issuer-heterogeneity and time-heterogeneity in the rating migrations of fallen
angels (FAs) and their speculative grade-rated peers (FA peers). Cox’s hazard
model is considered the pre-eminent method to estimate the probability that
an issuer survives in its current rating grade at any point in time t over the
time horizon T. In this study, estimation is based on two Cox’s hazard
models, including a proportional hazard model (Cox 1972) and a dynamic
hazard model. The first model employs a static estimation approach and time-
independent covariates, whereas the second uses a dynamic estimation
approach and time-dependent covariates. To allow for any dependence
among rating states of the same issuer, the marginal event-specific method
(Wei et al. 1989) was used to obtain robust variance estimates. For validation
purpose, the Brier score (Brier 1950) and its covariance decomposition
(Yates 1982) were applied to assess the forecast performance of estimated
models in forming time-varying survival probability estimates for issuers
out-of-sample.
Keywords: Survival analysis, Hazard model, Time-varying covariate, Recurrent
event, Brier score, Covariance decomposition, Rating migration, Fallen angel,
Markov property, Issuer-heterogeneity, Time-heterogeneity

Chapter 73: Creation and Control of Bubbles: Managers

Compensation Schemes, Risk Aversion, and
Wealth and Short Sale Constraints

This chapter takes an alternative approach of inquiry – that of using labora-

tory experiments – to study the creation and control of speculative bubbles.
The following three factors are chosen for analysis: the compensation scheme
of portfolio managers, wealth and supply constraints, and the relative risk
aversion of traders. Under a short investment horizon induced by
a tournament compensation scheme, speculative bubbles are observed in
markets of speculative traders and in mixed markets of conservative and
speculative traders. The primary method of analysis is to use live subjects
in a laboratory setting to generate original trading data, which are compared
to their fundamental values. Standard statistical techniques are used to sup-
plement analysis in explaining the divergence of asset prices from their
fundamental values.
Keywords: Speculative bubbles, Laboratory experimental asset markets,
Fundamental asset values, Tournament, Market efficiency, Behavioral finance,
Ordinary least squares regression, Correlation
48 C.-F. Lee and J.C. Lee

Chapter 74: Range Volatility: A Review of Models and Empirical

Studies

In this chapter, we survey the significant development of range-based volatility

models, beginning with the simple random walk model up to the conditional
autoregressive range (CARR) model. For the extension to range-based multivariate
volatilities, some approaches developed recently are adopted, such as the dynamic
conditional correlation (DCC) model, the double smooth transition conditional
correlation (DSTCC) GARCH model, and the copula method. At last, we introduce
different approaches to build bias-adjusted realized range to obtain a more efficient
estimator.
Keywords: Range, Volatility forecasting, Dynamic conditional correlation,
Smooth transition, Copula, Realized volatility, Risk management

Chapter 75: Business Models: Applications to Capital Budgeting,

Equity Value, and Return Attribution

This chapter describes a business model in a contingent claim modeling framework.

The chapter then provides three applications of the business model. Firstly, the
chapter determines the optimal capital budgeting decision in the presence of fixed
operating costs, and shows how the fixed operating cost should be accounted by in
an NPV calculation. Secondly, the chapter determines the values of equity value,
the growth option, the retention option as the building blocks of primitive firm
value. Using a sample of firms, the chapter illustrates a method in comparing the
equity values of firms in the same business sector. Thirdly, the chapter relates the
change in revenue to the change in equity value, showing how the combined
operating leverage and financial leverage may affect the firm valuation and risks.
Keywords: Bottom-up capital budgeting, Business model, Capital budgeting,
Contingent claim model, Equity value, Financial leverage, Fixed operating cost,
Gross return on investment (GRI), Growth option, Market performance measure,
NPV, Operating leverage, Relative value of equity, Retention option, Return
attribution, Top-down capital budgeting, Wealth transfer

Chapter 76: VAR Models: Estimation, Inferences, and Applications

This chapter provides a brief overview of the basic Vector autoregression (VAR)
approach by focusing on model estimation and statistical inferences. VAR models
have been used extensive in finance and economic analysis. Applications of VAR
models in some finance areas are discussed, including asset pricing, international
finance, and market microstructure. It is shown that such approach provides
a powerful tool to study financial market efficiency, stock return predictability,
exchange rate dynamics, and information content of stock trades and market
quality.
1 Introduction to Financial Econometrics and Statistics 49

Keywords: VAR, Granger-causality test, Impulse response, Variance decom-

position, Co-integration, Asset return predictability, Market quality, Information
content of trades, Informational efficiency

Chapter 77: Model Selection for High-Dimensional Problems

This chapter introduces penalized least squares, which seek to keep

important predictors in a model, while penalizing coefficients associated with
irrelevant predictors. As such, under certain conditions, penalized least
squares can lead to a sparse solution for linear models and achieve asymptotic
consistency in separating relevant variables from irrelevant ones. We then
review independence screening, a recently developed method for analyzing
ultrahigh-dimensional data where the number of variables or parameters can be
exponentially larger than the sample size. Independence screening selects relevant
variables based on certain measures of marginal correlations between candidate
variables and the response. Finally, we discuss and advocate multistage procedures
that combine independence screening and variable selection and that may be
especially suitable for analyzing high-frequency financial data.
Keywords: Model selection, Variable selection, Dimension reduction,
Independence screening, High-dimensional data, Ultrahigh-dimensional data, Gen-
eralized correlations, Penalized least squares, Shrinkage, Statistical learning,
SCAD penalty, Oracle property

Chapter 78: Hedonic Regression Models

The chapter examines three specific, different hedonic specifications: the linear,
semi-log, and Box-Cox transformed hedonic models and applies them to real estate
data. It also discusses recent innovations related to hedonic models and how these
models are being used in contemporary studies. This provides a basic overview of
the nature and variety of hedonic empirical pricing models that are employed in
the economics literature. It explores the history of hedonic modeling and summa-
rizes the field’s utility-theory-based, microeconomic foundations. It also provides
a discussion of and potential solutions for common problems associated with
hedonic modeling.
Keywords: Hedonic models, Regression, Real estate, Box-Cox, Pricing, Price
indexes, Semi-log, Least squares, Housing, Property

Chapter 79: Optimal Payout Ratio Under Uncertainty and the

Flexibility Hypothesis: Theory and Empirical Evidence

We theoretically extend the proposition of DeAngelo and DeAngelo’s (2006) opti-

mal payout policy in terms of the flexibility dividend hypothesis. We also
50 C.-F. Lee and J.C. Lee

introduce growth rate, systematic risk, and total risk variables into the theoretical
model. We use a panel data collected in the USA from 1969 to 2009 to empirically
investigate the impact of growth rate, systematic risk, and total risk on the optimal
payout ratio in terms of the fixed-effects model. Furthermore, we implement the
moving estimates process to find the empirical breakpoint of the structural change
for the relationship between the payout ratio and risks and confirm that the
empirical breakpoint is not different from our theoretical breakpoint. Our theo-
retical model and empirical results can therefore be used to identify whether
flexibility or the free cash flow hypothesis should be used to determine the
dividend policy.
Keywords: Dividends, Payout policy, Optimal payout ratio, Flexibility hypoth-
esis, Free cash flow hypothesis, Signaling hypothesis, Fixed effect, Clustering
effect, Structural change model, Moving estimates processes, Systematic risk,
Total risk, Market perfection

Chapter 80: Modeling Asset Returns with Skewness, Kurtosis,

and Outliers

This chapter uses an exponential generalized beta distribution of the second kind
(EGB2) to model the returns on 30 Dow-Jones industrial stocks. The model
accounts for stock return characteristics, including fat tails, peakedness
(leptokurtosis), skewness, clustered conditional variance, and leverage effect.
The goodness-of-fit statistic provides supporting evidence in favor of EGB2
distribution in modeling stock returns. The EGB2 distribution used in this
chapter is a four parameter distribution. It has a closed-form density function,
and its higher-order moments are finite and explicitly expressed by its parameters.
The EGB2 distribution nests many widely used distributions such as normal
distribution, log-normal distribution, Weibull distribution, and standard logistic
distribution.
Keywords: Expected stock return, Higher moments, EGB2 distribution, Risk
management, Volatility, Conditional skewness, Risk premium

Chapter 81: Does Revenue Momentum Drive or Ride Earnings or

Price Momentum?

This chapter performs dominance test to show that revenue surprises, earnings
surprises, and prior returns, each lead to significant momentum returns that cannot
be fully explained by the others, suggesting that each convey some exclusive
and unpriced information content. Also, the joint implications of revenue surprises,
earnings surprises, and prior returns are underestimated by investors, particularly
when information variables point in the same direction. Momentum
1 Introduction to Financial Econometrics and Statistics 51

cross-contingencies are observed in that momentum profits driven by firm funda-

mental information positively depend on the accompanying firm market informa-
tion, and vice versa. A three-way combined momentum strategy may offer monthly
return as high as 1.44%.
Keywords: Earnings surprises, Momentum strategies, Post-earnings-
announcement drift, Revenue surprises

Chapter 82: A VG-NGARCH Model for Impacts of Extreme

Events on Stock Returns

This chapter compares two types of GARCH models, namely, the VG-NGARCH
and the GARCH-jump model with autoregressive conditional jump intensity, i.e.,
the GARJI model, to make inferences on the log of stock returns when there are
irregular substantial price fluctuations. The VG-NGARCH model imposes
a nonlinear asymmetric structure on the conditional shape parameters in a variance-
gamma process, which describes the arrival rates for news with different degrees of
influence on price movements, and provides an ex ante probability for the occur-
rence of large price movements. On the other hand, the GARJI model, a mixed
GARCH-jump model proposed by Chan and Maheu (2002), adopts two indepen-
dent autoregressive processes to model the variances corresponding to moderate
and large price movements, respectively.
Keywords: VG-NGARCH model, GARCH-jump model, Autoregressive
conditional jump intensity, GARJI model, Substantial price fluctuations, Shape
parameter, Variance-gamma process, Ex ante probability, Daily stock price,
Goodness-of-fit

Chapter 83: Risk-Averse Portfolio Optimization via Stochastic

Dominance Constraints

This chapter presents a new approach to portfolio selection based on

stochastic dominance. The portfolio return rate in the new model is required to
stochastically dominate a random benchmark. We formulate optimality conditions
and duality relations for these models and construct equivalent optimization models
with utility functions. Two different formulations of the stochastic dominance
constraint, primal and inverse, lead to two dual problems which involve von
Neuman–Morgenstern utility functions for the primal formulation and rank depen-
dent (or dual) utility functions for the inverse formulation. We also discuss the
relations of our approach to value-at-risk and conditional value-at-risk.
Keywords: Portfolio optimization, Stochastic dominance, Stochastic order,
Risk, Expected utility, Duality, Rank dependent utility, Yaari’s dual utility,
Value-at-risk, Conditional value-at-risk
52 C.-F. Lee and J.C. Lee

Chapter 84: Implementation Problems and Solutions in Stochastic

Volatility Models of the Heston Type

This chapter compares three major approaches to solve the numerical instability
problem inherent in the fundamental solution of the Heston model. In this chapter,
we used the fundamental transform method proposed by Lewis to reduce the
number of variables from two to one and separate the payoff function from the
calculation of the Green function for option pricing. We show that the simple
adjusted-formula method is much simpler than the rotation-corrected angle method
of Kahl and Jäckel and also greatly superior to the direct integration method of
Shaw if taking computing time into consideration.
Keywords: Heston, Stochastic volatility, Fourier inversion, Fundamental trans-
form, Complex logarithm, Rotation-corrected angle, Simple adjusted formula,
Green function

Chapter 85: Stochastic Change-Point Models of Asset Returns and

Their Volatilities

This chapter considers two time-scales and uses the “short” time-scale to define
GARCH dynamics and the “long” time-scale to incorporate parameter jumps. This
leads to a Bayesian change-point ARX-GARCH model, whose unknown parameters
may undergo occasional changes at unspecified times and can be estimated by explicit
recursive formulas when the hyperparameters of the Bayesian model are specified.
Efficient estimators of the hyperparameters of the Bayesian model can be developed.
The empirical Bayes approach can be applied to the frequentist problem of partitioning
the time series into segments under sparsity assumptions on the change-points.
Keywords: ARX-GARCH, Bounded complexity, Contemporaneous jumps,
Change-point models, Empirical Bayes, Frequentist segmentation, Hidden Markov
models, Hyperparameter estimation, Markov chain Monte Carlo, Recursive filters,
Regression models, Stochastic volatility

Chapter 86: Unspanned Stochastic Volatilities and Interest Rate

Derivatives Pricing

This chapter first reviews the recent literature on the Unspanned Stochastic Volatil-
ities (USV) documented in the interest rate derivatives markets. The USV refers to
the volatilities factors implied in the interest rate derivatives prices that have little
correlation with the yield curve factors. We then present the result in Li and Zhao
(2006) that a sophisticated DTSM without USV feature can have serious difficulties
in hedging caps and cap straddles, even though they capture bond yields well.
Furthermore, at-the-money straddle hedging errors are highly correlated with
cap-implied volatilities and can explain a large fraction of hedging errors of all
caps and straddles across moneyness and maturities. We also present a multifactor
1 Introduction to Financial Econometrics and Statistics 53

term structure model with stochastic volatility and jumps that yields a closed-form
formula for cap prices from Jarrow et al. (2007). The three-factor stochastic volatility
model with Poisson jumps can price interest rate caps well across moneyness and
maturity. The econometric methods in this chapter include extended Kalman filter-
ing, maximum likelihood estimation with latent variables, local polynomial method,
and nonparametric density estimation.
Keywords: Term structure modeling, Interest rate volatility, Heath-Jarrow-
Morton model, Nonparametric density estimation, Extended Kalman filtering

Chapter 87: Alternative Equity Valuation Models

This chapter examines alternative equity valuation models and their ability to
forecast future stock prices. We use simultaneous equations estimation technique
to investigate the stock price forecast ability of Ohlson’s model, Feltham and
Ohlson’s Model, and Warren and Shelton’s (1971) model. Moreover, we use the
combined forecasting methods proposed by Granger and Newbold (1973) and
Granger and Ramanathan (1984) to form combined stock price forecasts from
individual models. Finally, we examine whether comprehensive earnings can
provide incremental price-relevant information beyond net income.
Keywords: Ohlson model, Feltham and Ohlson model, Warren and Shelton
model, Equity valuation models, Simultaneous equations estimation, Fundamental
analysis, Financial statement analysis, Financial planning and forecasting, Com-
bined forecasting, Comprehensive earnings, Abnormal earnings, Operating earn-
ings, Accounting earnings

Chapter 88: Time Series Models to Predict the Net Asset Value (NAV)
of an Asset Allocation Mutual Fund VWELX

This research examines the use of various forms of time-series models to predict
the total net asset value (NAV) of an asset allocation mutual fund. The first set of
model structures included simple exponential smoothing, double exponential
smoothing, and the Winter’s method of smoothing. The second set of predictive
models used represented trend models. They were developed using regression
estimation. They included linear trend model, quadratic trend model, and an
exponential model. The third type of method used was a moving average method.
The fourth set of models incorporated the Box-Jenkins method, including an
autoregressive model, a moving average model, and an unbounded autoregressive
and moving average method.
Keywords: NAV of a mutual fund, Asset allocation fund, Combination of
forecasts, Single exponential smoothing, Double exponential smoothing, Winter’s
method, Linear trend model, Quadratic trend model, Exponential trend model,
Moving average method, Autoregressive model, Moving average model,
Unbounded autoregressive moving average model
54 C.-F. Lee and J.C. Lee

Chapter 89: Discriminant Analysis and Factor Analysis:

Theory and Method

This chapter discusses three multivariate techniques in detail: discriminant analysis,

factor analysis, and principal component analysis. In addition, the stepwise
discriminant analysis by Pinches and Mingo (1973) is improved using a goal pro-
gramming technique. These methodologies are applied to determine useful financial
ratios and the subsequent bond ratings. The analysis shows that the stepwise discrim-
inant analysis fails to be an efficient solution as the hybrid approach using the goal
programming technique outperforms it, which is a compromised solution for the
maximization of the two objectives, namely, the maximization of the explanatory
power and the maximization of discriminant power.
Keywords: Multivariate technique, Discriminant analysis, Factor analysis, Prin-
ciple component analysis, Stepwise discriminant analysis, Goal programming,
Bond ratings, Compromised solution, Explanatory power, Discriminant power

Chapter 90: Implied Volatility: Theory and Empirical Method

This chapter reviews the different theoretical methods used to estimate implied
standard deviation and to show how the implied volatility can be estimated in
empirical work. The OLS method for estimating implied standard deviation is first
introduced and the formulas derived by applying a Taylor series expansion method
to Black-Scholes option pricing model are also described. Three approaches of
estimating implied volatility are derived from one, two, and three options, respec-
tively. Because of these formulas with the remainder terms, the accuracy of these
formulas depends on how an underlying asset is close to the present value of
exercise price in an option. The formula utilizing three options for estimating
implied volatility is more accurate rather than other two approaches. In this chapter,
we use call options on S&P 500 index futures in 2010 and 2011 to illustrate how
MATLAB can be used to deal with the issue of convergence in estimating implied
volatility of future options.
Keywords: Implied volatility, Implied standard deviation (ISD), Option pricing
model, MATLAB, Taylor series expansion, Ordinary least-squares (OLS), Black-
Scholes Model, Options on S&P 500 index futures

Chapter 91: Measuring Credit Risk in a Factor Copula Model

This chapter uses a new approach to estimate future credit risk on target portfolio based
on the framework of CreditMetricsTM by J.P. Morgan. However, we adopt the
perspective of factor copula and then bring the principal component analysis concept
into factor structure to construct a more appropriate dependence structure among
credits. In order to examine the proposed method, we use real market data instead of
a virtual one. We also develop a tool for risk analysis which is convenient to use,
1 Introduction to Financial Econometrics and Statistics 55

especially for banking loan businesses. The results show the fact that people assume
dependence structures are normally distributed will indeed lead to risks underestimate.
On the other hand, our proposed method captures better features of risks and shows the
fat-tail effects conspicuously even though assuming the factors are normally distributed.
Keywords: Credit risk, Credit VaR, Default correlation, Copula, Factor copula,
Principal component analysis

Chapter 92: Instantaneous Volatility Estimation by Nonparametric

Fourier Transform Methods

This chapter conducts some simulation tests to justify the effectiveness of the
Fourier transform method. Malliavin and Mancino (2009) proposed a nonparamet-
ric Fourier transform method to estimate the instantaneous volatility under the
assumption that the underlying asset price process is a semi-martingale. Two
correction schemes are proposed to improve the accuracy of volatility estimation.
By means of these Fourier transform methods, some documented phenomena such
as volatility daily effect and multiple risk factors of volatility can be observed.
Then, a linear hypothesis between the instantaneous volatility and VIX derived
from Zhang and Zhu (2006) is investigated.
Keywords: Information content, Instantaneous volatility, Fourier transform
method, Bias reduction, Correction method, Local volatility, Stochastic volatility,
VIX, Volatility daily effect, Online estimation

Chapter 93: A Dynamic CAPM with Supply Effect: Theory and

Empirical Results

This chapter first theoretically extends Black’s CAPM, and then uses price, dividend
per share, and earnings per share to test the existence of supply effect with US equity
data. A simultaneous equation system is constructed through a standard structural
form of a multi-period equation to represent the dynamic relationship between supply
and demand for capital assets. The equation system is exactly identified under our
specification. Then, two hypotheses related to supply effect are tested regarding
the parameters in the reduced-form system. The equation system is estimated by the
Seemingly Unrelated Regression (SUR) method, since SUR allows one to estimate
the presented system simultaneously while accounting for the correlated errors.
Keywords: CAPM, Asset, Endogenous supply, Simultaneous equations

Chapter 94: A Generalized Model for Optimum Futures Hedge Ratio

This chapter proposes the generalized hyperbolic distribution as the joint log-return
distribution of the spot and futures. Using the parameters in this distribution, we
derive several most widely used optimal hedge ratios: minimum variance,
56 C.-F. Lee and J.C. Lee

maximum Sharpe measure, and minimum generalized semivariance. To estimate

these optimal hedge ratios, we first write down the log-likelihood functions for
symmetric hyperbolic distributions. Then, we estimate these parameters by maxi-
mizing the log-likelihood functions. Using these MLE parameters for the general-
ized hyperbolic distributions, we obtain the minimum variance hedge ratio and the
optimal Sharp hedge ratio. Also based on the MLE parameters and the numerical
method, we can calculate the minimum generalized semivariance hedge ratio.
Keywords: Optimal hedge ratio, Generalized hyperbolic distribution, Martin-
gale property, Minimum variance hedge ratio, Minimum generalized semi-
invariance, Maximum Sharp measure, Joint-normality assumption, Hedging
effectiveness

Chapter 95: Instrument Variable Approach to Correct for

Endogeneity in Finance

This chapter reviews the instrumental variables (IV) approach to endogeneity from
the point of view of a finance researcher who is implementing instrumental variable
methods in empirical studies. This chapter is organized into two parts. Part I
discusses the general procedure of the instrumental variable approach, including
Two-Stage Least Square (2SLS) and Generalized Method of Moments (GMM), the
related diagnostic statistics for assessing the validity of instruments, which are
important but not used very often in finance applications, and some recent advances
in econometrics research on weak instruments. Part II surveys corporate finance
applications of instrumental variables. We found that the instrumental variables
used in finance studies are usually chosen arbitrarily, and very few diagnostic
statistics are performed to assess the adequacy of IV estimation. The resulting IV
estimates thus are questionable.
Keywords: Endogeneity, OLS, Instrumental variable (IV) estimation,
Simultaneous equations, 2SLS, GMM, Overidentifying restrictions, Exogeneity
test, Weak instruments, Anderson-Rubin statistic, Empirical corporate finance

Chapter 96: Application of Poisson Mixtures in the Estimation

of Probability of Informed Trading

This research first discusses the evolution of probability of informed trading in the
finance literature. Motivated by asymmetric effects, e.g., return and trading volume
in up and down markets, this study modifies a mixture of the Poisson distribution
model by different arrival rates of informed buys and sells to measure the
probability of informed trading proposed by Easley et al. (1996). By applying
the expectation–maximization (EM) algorithm to estimate the parameters of the
model, we derive a set of equations for maximum likelihood estimation and these
equations are encoded in a SAS Macro utilizing SAS/IML for implementation of
the methodology.
1 Introduction to Financial Econometrics and Statistics 57

Keywords: Probability of informed trading (PIN), Expectation–maximization

(EM) algorithm, A mixture of Poisson distribution, Asset-pricing returns, Order
imbalance, Information asymmetry, Bid-ask spreads, Market microstructure, Trade
direction, Errors in variables

Chapter 97: CEO Stock Options and Analysts’ Forecast Accuracy

and Bias

This chapter uses ordinary least squares estimation to investigate the relations
between CEO stock options and analysts’ earnings forecast accuracy and bias.
Our OLS models relate forecast accuracy and forecast bias (the dependent
variables) to CEO stock options (the independent variable) and controls for earn-
ings characteristics, firm characteristics, and forecast characteristics. In addition,
the models include controls for industry and year. We use four measures of options:
new options, existing exercisable options, existing unexercisable options, and total
options (sum of the previous three), all scaled by total number of shares outstand-
ing, and estimate two models for each dependent variable, one including total
options and the other including new options, existing exercisable options, and
existing unexercisable options. We also use both contemporaneous as well as
lagged values of options in our main tests.
Keywords: CEO stock options, Analysts’ forecast accuracy, Analysts’ forecast
bias, CEO compensation, Agency costs, Investment risk taking, Effort allocation,
Opportunistic earnings management, Opportunistic disclosure management,
Forecasting complexity

Chapter 98: Option Pricing and Hedging Performance Under

Stochastic Volatility and Stochastic Interest Rates

This chapter fills this gap by first developing an implementable option model in
closed form that admits both stochastic volatility and stochastic interest rates and that
is parsimonious in the number of parameters. Based on the model, both delta-neutral
and single-instrument minimum variance hedging strategies are derived analytically.
Using S&P 500 option prices, we then compare the pricing and hedging performance
of this model with that of three existing ones that, respectively, allow for (i) constant
volatility and constant interest rates (the Black-Scholes), (ii) constant volatility but
stochastic interest rates, and (iii) stochastic volatility but constant interest rates.
Overall, incorporating stochastic volatility and stochastic interest rates produces the
best performance in pricing and hedging, with the remaining pricing and hedging
errors no longer systematically related to contract features. The second performer in
the horse-race is the stochastic volatility model, followed by the stochastic interest
rates model and then by the Black-Scholes.
Keywords: Stock option pricing, Stochastic volatility, Stochastic interest rates,
Hedge ratios, Hedging, Pricing performance, and Hedging performance
58 C.-F. Lee and J.C. Lee

Chapter 99: The Le Châtelier Principle of the Capital Market

Equilibrium

This chapter purports to provide a theoretical underpinning for the problem of the
Investment Company Act. The theory of the Le Chatelier Principle is well known in
thermodynamics: The system tends to adjust itself to a new equilibrium as far as
possible. In capital market equilibrium, added constraints on portfolio investment in
each stock can lead to inefficiency manifested in the right-shifting efficiency
frontier. According to the empirical study, the potential loss can amount to millions
of dollars coupled with a higher risk-free rate and greater transaction and informa-
tion costs.
Keywords: Markowitz model, Efficient frontiers, With constraints, Without
constraints, Le Chatelier principle, Thermodynamics, Capital market equilibrium,
Diversified mutual funds, Quadratic programming, Investment company act

References
Aitchison, J., & Brown, J. A. C. (1973). Lognormal distribution, with special reference to its uses
in economics. London: Cambridge University Press.
Ait-Sahalia, Y., & Lo, A. W. (2000). Nonparametric risk management and implied risk aversion.
Journal of Econometrics, 94, 9–51.
Amemiya, T. (1974). A note on a fair and Jaffee model. Econometrica, 42, 759–762.
Anderson, T. W. (1994). The statistical analysis of time series. New York: Wiley-Interscience.
Anderson, T. W. (2003). An introduction to multivariate statistical analysis. New York: Wiley-
Interscience.
Angrist, J. D., & Lavy, V. (1999). Using Maimonides’ rule to estimate the effect of class size on
scholastic achievement. Quarterly Journal of Economics, 14(2), 533–575.
Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-
components models. Journal of Econometrics, 68(1), 29–51.
Bakshi, G., Cao, C., & Chen, Z. (1997). Empirical performance of alternative option pricing
models. Journal of Finance, 52(5), 2003–2049.
Baltagi, B. (2008). Econometric analysis of panel data (4th ed.). Chichester: Wiley.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81(3), 637–654.
Black, F., Jensen, M. C., & Scholes, M. (1972). The capital asset pricing model: Some empirical tests.
In M. C. Jensen (Ed.), Studies in the theory of capital markets. New York: Praeger.
Blume, M. E., & Friend, I. (1973). A new look at the capital asset pricing model. Journal of
Finance, 28, 19–33.
Brown, S. J., & Goetzmann, W. N. (1997). Mutual fund styles. Journal of Financial Economics,
43(3), 373–399.
Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2006). Robust inference with multiway clustering
(Working Paper).
Chacko, G., & Viceira, L. M. (2003). Spectral GMM estimation of continuous-time processes.
Journal of Econometrics, 116(1), 259–292.
Chang, C. F. (1999). Determinants of capital structure and management compensation: The
partial least squares approach. Ph.D. dissertation, Rutgers University.
Chang, H. S., & Lee, C. F. (1977). Using pooled time-series and cross section data to test the firm
and time effects in financial analysis. Journal of Financial and Quantitative Analysis,
12, 457–471.
1 Introduction to Financial Econometrics and Statistics 59

Chang, C., Lee, A. C., & Lee, C. F. (2009). Determinants of capital structure choice: A structural
equation modeling approach. Quarterly Review of Economic and Finance, 49(2), 197–213.
Chen, H. Y. (2011). Momentum strategies, dividend policy, and asset pricing test.
Ph.D. dissertation, Rutgers University.
Chen, S. N., & Lee, C. F. (1981). The sampling relationship between Sharpe’s performance
measure and its risk proxy: Sample size, investment horizon and market conditions. Manage-
ment Science, 27(6), 607–618.
Chen, K. H., & Shimerda, T. A. (1981). An empirical analysis of useful finance ratios. Financial
Management, 10(1), 51–60.
Chen, W. P., Chung, H., Lee, C. F., & Liao, W. L. (2007). Corporate governance and equity
liquidity: Analysis of S&P transparency and disclosure rankings. Corporate Governance: An
International Review, 15(4), 644–660.
Cheng, D. C., & Lee, C. F. (1986). Power of alternative specification errors tests in identifying
misspecified market models. The Quarterly Review of Economics and Business, 16(3), 6–24.
Cheng, K. F., Chu, C. K., & Hwang, R. C. (2010). Predicting bankruptcy using the discrete-time
semiparametric hazard model. Quantitative Finance, 10, 1055–1066.
Chow, G. C. (1960). Tests of equality between sets of coefficients in two linear regressions.
Econometrica, 28, 591–605.
Core, J. E. (2000). The directors’ and officers’ insurance premium: An outside assessment of the
quality of corporate governance. Journal of Law, Economics, and Organization, 16(2), 449–477.
Cox, J. C., Ross, S. A., & Rubinstein, M. (1979). Option pricing: A simplified approach. Journal of
Financial Economics, 7(3), 229–263.
Fair, R. C., & Jaffee, D. M. (1972). Methods of estimation of market in disequilibrium.
Econometrica, 40, 497–514.
Fama, E. F. (1971). Parameter estimates for symmetric stable distributions. Journal of the
American Statistical Association, 66, 331–338.
Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal of
Political Economy, 81, 607–636.
Fok, R. C. W., Lee, C. F., & Cheng, D. C. (1996). Alternative specifications and estimation
methods for determining random beta coefficients: Comparison and extensions. Journal of
Financial Studies, 4(2), 61–88.
Frecka, T. J., & Lee, C. F. (1983). Generalized financial ratio adjustment processes and their
implications. Journal of Accounting Research, 21, 308–316.
Gordon, J., & Gordon, M. (1997). The fiinite horizon expected return model. The Financial
Analyst Journal, 53, 52–61.
Granger, C. W. J., & Newbold, P. (1973). Some comments on the evaluation of economic
forecasts. Applied Economics, 5(1), 35–47.
Granger, C. W. J., & Newbold, P. (1974). Spurious regressions in econometrics. Journal of
Econometrics, 2, 111–120.
Granger, C. W. J., & Ramanathan, R. (1984). Improved methods of combining forecasts. Journal
of Forecasting, 3, 197–204.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica, 50(4), 1029–1054.
Hansen, B. E. (1996). Inference when a nuisance parameter is not identified under the null
hypothesis. Econometrica, 64(2), 413–430.
Hansen, B. E. (1997). Approximate asymptotic p-values for structural change tests. Journal of
Business and Economic Statistics, 15, 60–67.
Hansen, B. E. (1999). Threshold effects in non-dynamic panels: Estimation, testing, and inference.
Journal of Econometrics, 93, 345–368.
Hansen, B. E. (2000a). Sample splitting and threshold estimation. Econometrica, 68(3), 575–603.
Hansen, B. E. (2000b). Testing for structural change in conditional models. Journal of Economet-
rics, 97, 93–115.
60 C.-F. Lee and J.C. Lee

Heston, S. L. (1993). A closed-form solution for options with stochastic volatility with applica-
tions to bond and currency options. Review of Financial Studies, 6(2), 327–343.
Hsiao, C. (2003). Analysis of panel data (2nd ed.). Cambridge MA: Cambridge University Press.
Hutchinson, J. M., Lo, A. W., & Poggio, T. (1994). A nonparametric approach to pricing and
hedging derivative securities via learning networks. Journal of Finance, 49(3), 851–889.
Hwang, R. C., Cheng, K. F., & Lee, J. C. (2007). A semiparametric method for predicting
bankruptcy. Journal of Forecasting, 26, 317–342.
Hwang, R. C., Wei, H. C., Lee, J. C., & Lee, C. F. (2008). On prediction of financial distress using
the discrete-time survival model. Journal of Financial Studies, 16, 99–129.
Hwang, R. C., Cheng, K. F., & Lee, C. F. (2009). On multiple-class prediction of issuer crediting
ratings. Journal of Applied Stochastic Models in Business and Industry, 25, 535–550.
Hwang, R. C., Chung, H., & Chu, C. K. (2010). Predicting issuer credit ratings using
a semiparametric method. Journal of Empirical Finance, 17(1), 120–137.
Ittner, C. D., Larcker, D. F., & Rajan, M. V. (1997). The choice of performance measures in annual
bonus contracts. Accounting Review, 72(2), 231–255.
Jones, J. (1991). Earnings management during import relief investigations. Journal of Accounting
Research, 29, 193–228.
Kao, L. J., & Lee, C. F. (2012). Alternative method for determining industrial bond ratings: theory
and empirical evidence. International Journal of Information Technology & Decision Making,
11(6), 1215–1236.
Kau, J. B., & Lee, C. F. (1976). Functional form, density gradient and the price elasticity of
demand for housing. Urban Studies, 13(2), 181–192.
Kau, J. B., Lee, C. F., & Sirmans, C. F. (1986). Urban econometrics: Model developments and
empirical results. Greenwich: JAI Press. Research in Urban Economics, Vol. 6.
Kaufman, L., & Rousseeuw, P. J. (1990). Finding groups in data: An introduction to cluster
analysis (9th ed.). New York: Wiley-Interscience.
Kim, D. (1995). The errors in the variables problem in the cross-section of expected stock returns.
Journal of Finance, 50(5), 1605–1634.
Kim, D. (1997). A reexamination of firm size, book-to-market, and earnings price in the cross-
section of expected stock returns. Journal of Financial and Quantitative Analysis, 32(4),
463–489.
Kim, D. (2010). Issues related to the errors-in-variables problems in asset pricing tests. In
Handbook of quantitative finance and risk management (Part V, Chap. 70, pp. 1091–1108).
Kuan, C. M., & Hornik, K. (1995). The generalized fluctuation test: A unifying view. Econometric
Reviews, 14, 135–161.
Lambert, R., & Larcker, D. (1987). An analysis of the use of accounting and market measures of
performance in executive compensation contracts. Journal of Accounting Research, 25, 85–125.
Lee, C. F. (1973). Errors-in-variables estimation procedures with applications to a capital asset
pricing model. Ph.D. dissertation, The State University of New York at Buffalo.
Lee, C. F. (1976a). A note on the interdependent structure of security returns. Journal of Financial
and Quantitative Analysis, 11, 73–86.
Lee, C. F. (1976b). Functional form and the dividend effect of the electric utility industry. Journal
of Finance, 31(5), 1481–1486.
Lee, C. F. (1977a). Functional form, skewness effect and the risk-return relationship. Journal of
Financial and Quantitative Analysis, 12, 55–72.
Lee, C. F. (1977b). Performance measure, systematic risk and errors-in-variable estimation
method. Journal of Economics and Business, 39, 122–127.
Lee, A. (1996). Cost of capital and equity offerings in the insurance industry. Ph.D. dissertation,
The University of Pennsylvania in Partial.
Lee, C. F., & Chen, H. Y. (2013). Alternative errors-in-variable models and their application in
finance research (Working Paper). Rutgers University.
Lee, A. C., & Cummins, J. D. (1998). Alternative models for estimating the cost of capital for
property/casualty insurers. Review of Quantitative Finance and Accounting, 10(3), 235–267.
1 Introduction to Financial Econometrics and Statistics 61

Lee, C. F., & Jen, F. C. (1978). Effects of measurement errors on systematic risk and performance
measure of a portfolio. Journal of Financial and Quantitative Analysis, 13(2), 299–312.
Lee, K. W., & Lee, C. F. (2012). Multiple directorships, corporate ownership, firm performance
(Working Paper). Rutgers University.
Lee, C. F., & Wu, C. C. (1985). The impacts of kurtosis on risk stationarity: Some empirical
evidence. Financial Review, 20(4), 263–269.
Lee, C. F., & Zumwalt, J. K. (1981). Associations between alternative accounting
profitability measures and security returns. Journal of Financial and Quantitative Analysis,
16, 71–93.
Lee, C. F., Newbold, P., Finnerty, J. E., & Chu, C. C. (1986). On accounting-based, market-
based and composite-based beta predictions: Method and implications. Financial Review,
21, 51–68.
Lee, C. F., Wei, K. C. J., & Bubnys, E. L. (1989). The APT versus the multi-factor CAPM:
Empirical evidence. Quarterly Review of Economics and Business, 29, 6–16.
Lee, C. F., Wu, C. C., & Wei, K. C. J. (1990). Heterogeneous investment horizon and capital asset
pricing model: Theory and implications. Journal of Financial and Quantitative Analysis,
25, 361–376.
Lee, C. F., Tsai C. M., & Lee A. C. (2011a). Asset pricing with disequilibrium price adjustment:
Theory and empirical evidence. Quantitative Finance.
Lee, C. F., Gupta, M. C., Chen, H. Y., & Lee, A. C. (2011b). Optimal payout ratio under
uncertainty and the flexibility hypothesis: Theory and empirical evidence. Journal of Corpo-
rate Finance, 17(3), 483–501.
Lee, C. F., Chien, C. C., Lin, H. C., Lin, Y. C., & Lin, F. C. (2013). Tradeoff between reputation
concerns and economic dependence for auditors – Threshold regression approach (Working
Paper). Rutgers University.
Leon, A., & Vaello-Sebastia, A. (2009). American GARCH employee stock option valuation.
Journal of Banking and Finance, 33(6), 1129–1143.
Lien, D. (2010). An note on the relationship between the variability of the hedge ratio and hedging
performance. Journal of Futures Market, 30(11), 1100–1104.
Lien, D., & Shrestha, K. (2007). An empirical analysis of the relationship between
hedge ratio and hedging horizon using wavelet analysis. Journal of Futures Market, 27(2),
127–150.
Liu, B. (2006). Two essays in financial econometrics: I: Functional forms and pricing of country
funds. II: The term structure model of inflation risk premia. Ph.D. dissertation, Rutgers University.
Maddala, G. S., Rao, C. R., & Maddala, G. S. (1996). Handbook of statistics 14: Statistical
methods in finance. Amsterdam: Elsevier Science & Technology.
Martin, C. (1990). Corporate borrowing and credit constrains: Structural disequilibrium estimates
for the U.K. Review of Economics and Statistics, 72(1), 78–86.
Mayer, W. J. (1989). Estimating disequilibrium models with limited a priori price-adjustment
information. Journal of Econometrics, 41(3), 303–320.
Miller, M. H., & Modigliani, F. (1966). Some estimates of the cost of capital to the utility industry,
1954-7. American Economic Review, 56(3), 333–391.
Myers, R. J. (1991). Estimating time-varying optimal hedge ratios on futures markets. Journal of
Futures Markets, 11(1), 39–53.
Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica, 55(3), 703–708.
Ohlson, J. S. (1980). Financial ratios and the probabilistic prediction of bankruptcy. Journal of
Accounting Research, 19, 109–131.
Petersen, M. A. (2009). Estimating standard errors in finance panel data sets: Comparing
approaches. Review of Financial Studies, 22, 435–480.
Pinches, G. E., & Mingo, K. A. (1973). A multivariate analysis of industrial bond ratings. Journal
of Finance, 28(1), 1–18.
Quandt, R. E. (1988). The econometrics of disequilibrium. New York: Basil Blackwell Inc.
62 C.-F. Lee and J.C. Lee

Rendleman, R. J., Jr., & Barter, B. J. (1979). Two-state option pricing. Journal of Finance,
24, 1093–1110.
Rubinstein, M. (1994). Implied binomial trees. Journal of Finance, 49, 771–818.
Schroder, M. (1989). Computing the constant elasticity of variance option pricing formula.
Journal of Finance, 44(1), 211–219.
Sealey, C. W., Jr. (1979). Credit rationing in the commercial loan market: Estimates of a structural
model under conditions of disequilibrium. Journal of Finance, 34, 689–702.
Sears, R. S., & John Wei, K. C. (1988). The structure of skewness preferences in asset pricing
model with higher moments: An empirical test. Financial Review, 23(1), 25–38.
Shapiro, A. F. (2005). Fuzzy regression models (Working Paper). Penn State University.
Shumway, T. (2001). Forecasting bankruptcy more accurately: A simple hazard model. The
Journal of Business, 74, 101–124.
Thompson, S. B. (2011). Simple formulas for standard errors that cluster by both firm and time.
Journal of Financial Economics, 99(1), 1–10.
Thursby, J. G. (1985). The relationship among the specification error tests of Hausman, Ramsey
and Chow. Journal of the American Statistical Association, 80(392), 926–928.
Titman, S., & Wessels, R. (1988). The determinants of capital structure choice. Journal of
Finance, 43(1), 1–19.
Tsai, G. M. (2005). Alternative dynamic capital asset pricing models: Theories and empirical
results. Ph.D. dissertation, Rutgers University.
Van Der Klaauw, W. (2002). Estimating the effect of financial aid offers on college enrollment:
A regression-discontinuity approach. International Economic Review, 43(4), 1249–1287.
Wei, K. C. (1984). The arbitrage pricing theory versus the generalized intertemporal capital asset
pricing model: Theory and empirical evidence. Ph.D. dissertation, University of Illinois at
Urbana-Champaign.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica, 48(4), 817–838.
Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.).
Cambridge, MA: The MIT Press.
Yang, C. C. (1989). The impact of new equity financing on firms’ investment, dividend and debt-
financing decisions. Ph.D. dissertation, The University of Illinois at Urbana-Champaign.
Yang, C. C., Lee, C. F., Gu, Y. X., & Lee, Y. W. (2010). Co-determination of capital structure and
stock returns – A LISREL approach: An empirical test of Taiwan stock markets. Quarterly
Review of Economic and Finance, 50(2), 222–233.
Zeileis, A., Leisc, F., Hornik, K., & Kleiber, C. (2002). Strucchange: An R package for testing for
structural change in linear regression models. Journal of Statistical Software, 7, 1–38.
Experience, Information Asymmetry, and
Rational Forecast Bias 2
April Knill, Kristina L. Minnick, and Ali Nejadmalayeri

Contents
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.2 Theoretical Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.3 Empirical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5 Empirical Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.6 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Appendix 1: Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
The Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Proof of Proposition 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Proof of Proposition 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Appendix 2: Alternate Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Appendix 3: Alternate Proxies for Information Asymmetry for Table 2.5 . . . . . . . . . . . . . . . . . . . . 98
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

Abstract
We use a Bayesian model of updating forecasts in which the bias in forecast
endogenously determines how the forecaster’s own estimates weigh into the
posterior beliefs. Our model predicts a concave relationship between accuracy in
forecast and posterior weight that is put on the forecaster’s self-assessment.

A. Knill (*)
The Florida State University, Tallahassee, FL, USA
e-mail: aknill@cob.fsu.edu
K.L. Minnick
Bentley University, Waltham, MA, USA
e-mail: kminnick@bentley.edu
A. Nejadmalayeri
Department of Finance, Oklahoma State University, Oklahoma, OK, USA
e-mail: ali.nejadmalayeri@okstate.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 63

DOI 10.1007/978-1-4614-7750-1_2,
# Springer Science+Business Media New York 2015
64 A. Knill et al.

We then use a panel regression to test our analytical findings and find that an
analyst’s experience is indeed concavely related to the forecast error.
This study examines whether it is ever rational for analysts to post biased
estimates and how information asymmetry and analyst experience factor into the
decision. Using a construct where analysts wish to minimize their forecasting
error, we model forecasted earnings when analysts combine private information
with consensus estimates to determine the optimal forecast bias, i.e., the devi-
ation from the consensus. We show that the analyst’s rational bias increases with
information asymmetry, but is concavely related with experience. Novice ana-
lysts post estimates similar to the consensus but as they become more experi-
enced and develop private information channels, their estimates become biased
as they deviate from the consensus. Highly seasoned analysts, who have superior
analytical skills and valuable relationships, need not post biased forecasts.

Keywords
Financial analysts • Forecast accuracy • Information asymmetry • Forecast bias •
Bayesian updating • Panel regressions • Rational bias • Optional bias • Analyst
estimation • Analyst experience

2.1 Introduction

Extant evidence suggests an intimate link between an analyst’s experience and her
forecasting performance. Analysts who are experienced and highly specialized
often forecast better than others (Clement and Tse 2005; Bernhardt et al. 2006).
One way they do so is by posting an optimistic bias (Mest and Plummer 2003; Gu
and Xue 2007). Novice analysts with limited resources tend to herd with others,
which results in almost no bias (Bernhardt et al. 2006). In theory, superior fore-
casters produce better estimates either by resolving information asymmetry or by
offering a better assessment. Lim (2001) suggests that analysts can improve fore-
cast accuracy by strategically biasing their forecasts upwards, which placates
management, and in essence purchases additional information.1 Analysts with

1
Beyer (2008) argues that, even without incentives to appease management, analysts may still post
forecasts that exceed median earnings because managers can manipulate earnings upward to
prevent falling short of earnings forecasts. Moreover, Conrad et al. (2006) find support for the
idea that analysts’ “. . . recommendation changes are “sticky” in one direction, with analysts
reluctant to downgrade.” Evidence also indicates that analysts rarely post sell recommendations
for a stock, suggesting that losing a firm’s favor can be viewed as a costly proposition. At the
extreme, firms even pursue legal damages for an analyst’s unfavorable recommendations. In
a 2001 congressional hearing, president and chief executive officer of the Association for Invest-
ment Management and Research told the US House of Representatives Committee on Financial
Services, Capital Markets Subcommittee, that “. . .In addition to pressures within their firms,
analysts can also be, and have been, pressured by the executives of corporate issuers to issue
favorable reports and recommendations. Regulation Fair Disclosure notwithstanding, recent
history. . .has shown that companies retaliate against analysts who issue ‘negative’
2 Experience, Information Asymmetry, and Rational Forecast Bias 65

long histories of examining firms in a particular industry can also offer a unique
perspective, and they signal their ability by posting biased estimates to signal
superior ability (Bernhardt et al. 2006).
Of course, analysts do not indefinitely and indiscriminately bias forecasts to
appease the firm or signal their ability. This is mainly because analysts can learn
from the forecasts of other analysts (Chen and Jiang 2006). By incorporating
information from other forecasts, analysts can improve the accuracy of their own
forecasts without posting biased estimates. The important question then is: given
the analyst’s own assessment ability and her efficacy in procuring private informa-
tion versus using consensus information, how should she construct an optimal
forecast? Does that optimal forecast ever include bias, and how does information
asymmetry affect this decision? We address these questions by analytically
and empirically examining how an analyst’s experience and information asymme-
try affect her forecasting. In so doing, we also account for the role of the consensus
estimate in an analyst’s forecast. We begin by modeling the problem of
optimal forecasting. To be specific, we combine key features of current rational
forecasting models by Lim (2001) and Chen and Jiang (2006). As in Chen and Jiang
(2006), analysts in our model form rational (i.e., minimum squared error) forecasts
by weighing both public and private information.2 Following Lim (2001), our
analysts post rational forecasts that deviate from the consensus to purchase
private information from managers. Motivated by Lim (2001) and Bernhardt
et al. (2006), we also allow analysts to post biased forecasts because they have
more expertise than the consensus. The novelty of our approach is that we
directly model how information asymmetry and analyst experience combine to
affect the purchase of private information for use in the forecast deviation.
We are also able to model how analysts with different levels of experience, i.e.,
novice, moderately experienced, and highly seasoned analysts, construct their
forecast.
We analytically derive the optimal deviation from the consensus as one that
minimizes the mean squared error while allowing for rational Bayesian updating
based on public and private knowledge. Our analysis shows that even in a rational
forecast framework, analysts’ forecast deviation depends on the observed consen-
sus deviation of other analysts. When analysts observe others deviating from the
consensus (especially those with more experience), they gain enough insight to
avoid posting a large deviation themselves. Our results confirm the findings of both
Bernhardt et al. (2006) and Chen and Jiang (2006) – that analysts can rationally
herd. In the presence of the informative consensus, analysts choose to herd with
each other, rather than post estimates that are biased.
Our theory suggests that the likelihood of posting biased estimates, conditional
on the consensus, is significantly influenced by the analyst’s ability to process

recommendations by denying them direct access to company executives and to company-

sponsored events that are important research tools. Companies have also sued analysts personally,
and their firms, for negative coverage....”
2
See also Han et al. (2001).
66 A. Knill et al.

information. Consistent with Hong et al. 2000), we show that novice analysts
essentially herd. Without either honed analytical abilities or valuable relationships,
these analysts follow the premise that the consensus is more accurate than their own
private information. Second, we show that a moderately experienced analyst relies
more on her sources inside the firm than her analytical skills. This manifests itself as
a biased forecast since she must appease management to tap into those sources. This
link between experience and deviation is not monotonic. Highly seasoned analysts
do not purchase as much additional information (or they purchase the information at
a reduced price), either because they possess superior analytical skills or because
their valuable relationships with firms afford them beneficial information without
the optimistic forecast. This preferential treatment is akin to that which is afforded
to companies in relationship lending with banks (Petersen and Rajan 1994). Indeed,
Carey et al. (1998) argue that the relationship lending of banks can also be ascribed
to some nonbank financial intermediaries. Although the firms that an analyst covers
are not necessarily financial, the same relationship could certainly exist and is based
on information asymmetry.
We further demonstrate that as the analyst-firm information asymmetry
increases, so does the bias. Similar to the results found in Mest and Plummer
(2003), analysts find private channels and analytical ability valuable mitigating
factors when faced with information asymmetry. Moderately experienced analysts,
who begin to tap into reliable private channels without the valuable relationships
that might afford preferential treatment, post larger deviations with the hope of
ascertaining better information. This suggests that both information asymmetry
and experience interactively affect the way analysts balance public and private
information to form forecasts. Our model also shows that the effect of information
asymmetry and analyst experience on rational deviation depends on the
dispersion of and the correlation between public and private signals. The quality
of private information channels, the informativeness of consensus, and the con-
nectedness of public and private signals significantly affect how analysts form
forecasts.
To examine the validity of our analytical findings, we empirically investigate
how analyst experience and information asymmetry affect forecast deviation from
the consensus.3 Our empirical results confirm our theoretical predictions that
rational bias (i.e., deviation from the consensus) is concavely related to analyst
experience and positively associated with information asymmetry. Novice analysts
and highly seasoned analysts post forecasts with smaller bias, while moderately
seasoned analysts post estimates that deviate further from the consensus. Moder-
ately seasoned analysts can benefit from a positive bias if they have the confidence
to separate from the herd and access reliable private sources of information.

3
Clement and Tse (2005) are the closest to our analysis; however while they admit that the
observed link between inexperience and herding can be a complex issue that might have other
roots than just career concerns, they do not provide detailed insight as to what and how this
complexity develops.
2 Experience, Information Asymmetry, and Rational Forecast Bias 67

These results are stronger for earlier forecasts versus later ones. As analysts become
highly seasoned, with external information networks and superior skills in fore-
casting, they find appeasing management to purchase information less useful
(or less costly in the same way that relationship lending affords firms cheaper
capital). As one might expect, when information asymmetry increases, rational
deviation from the consensus also increases. The degree of information
asymmetry faced by seasoned analysts has a significant positive effect on the
forecast deviation (information asymmetry also affects novice analysts but not as
extensively).
This study contributes to a growing literature on rational bias in analyst fore-
casting. Motivated by recent works by Bernhardt et al. (2006), Beyer (2008), Chen
and Jiang (2006), Lim (2001), and Mest and Plummer (2003), we take an integrated
modeling approach to arrive at a rational forecast bias, which is based on Bayesian
updating of public consensus and endogenously acquired private information. Our
approach is unique in that the forecast bias affects how public and private infor-
mation are combined. Specifically, analysts can use bias to improve forecast
accuracy by purchasing private information but can also learn from other analysts
by following the consensus. Our analytical findings, which are empirically
confirmed, show that unlike behavioral models, analysts can rationally herd. Unlike
signaling models, seasoned analysts do not always post biased estimates. It also
shows the value of the relationships that both moderately experienced and
highly seasoned analysts leverage to gain reliable private information. These results
are of practical interest because they provide evidence that analysts can and may
optimally bias their earnings estimates and that this optimal bias differs across
both analyst experience and information asymmetry. This knowledge may be
useful to brokerage houses for training purposes and/or evaluation of analyst
performance, particularly across industries with different levels of information
asymmetry.

2.2 Theoretical Design

Recent studies show that earnings forecasts reflect public information and private
assessments (Boni and Womack 2006; Chen and Jiang 2006; Lim 2001; Ramnath
2002). As Bernhardt et al. (2006) note, analysts’ forecasts of public information
partly relate to the consensus, common information, and unanticipated market-wide
shocks. We thus model analyst earnings forecasts using two components of earn-
ings information: common and idiosyncratic, with some uncertainty about each
component. In doing so, we also assume, as does Lim (2001), that the idiosyncratic
component directly relates to the private information analysts can obtain from the
firm by posting an optimistic view of the firm’s prospects (Nutt et al. 1999). Our
typical analyst also observes a previously posted forecast whereby the analyst can
learn about both the common and idiosyncratic components of earnings by incor-
porating previously disclosed information. Since we assume analysts are Bayesians,
they can learn from previous forecasts as an alternative to posting a positive
68 A. Knill et al.

deviation to purchase private signals from the firm. Since these prior estimates
partially reflect private information, the analysts may heavily weigh them into their
assessment, depending on the perceived information asymmetry and experience of
the previous analysts.4
In this section, we model analyst earnings forecasting in the presence of forecast
uncertainty. Following recent studies (Lim 2001; Ramnath 2002), our analyst has
an unconditional estimate, E ¼ X + e, about earnings, X  N(0, s2), with
some uncertainty, e  N(0, te2). As in Chen and Jiang (2006), our analyst
also observes a noisy consensus forecast, Ec ¼ X + ec, with a consensus
uncertainty, ec  N(0, tc2). As a Bayesian, our analyst forms a conditional
forecast, F ¼ w E + (1 – w) Ec, by combining her unconditional estimate with the
consensus. The optimal weight in her conditional forecast minimizes her squared
forecast error.5
As in Lim (2001), the optimal forecasting, however, is endogenous to private
information acquisition. Our analyst’s forecast precision relates to the private
information analysts can obtain from the firm by posting an optimistic view of
the firm’s prospects. That is to say, while forecast precision, te, consists of a self-
accuracy component, t0, reflecting the analyst expertise, the forecast precision can
be improved by t(b) through placating managers by posting positively biased, b,
forecasts. The marginal precision per bias, ∂t/∂b, reflects the analyst’s information
asymmetry. This is because an analyst faced with greater information asymmetry
should derive a larger marginal benefit from a biased forecast. Since the conditional
forecast partially reflects private information, the analyst’s optimal rational forecast
(and forecast bias) depends on the analyst’s expertise and information asymmetry.
As noted before, the objective of the analyst is to minimize her squared forecast
error, F – X:

h i h i
min E ðF  XÞ2 ¼ min ðwE þ ð1  wÞEc Þ2 (2.1)
wjb wjb

where E[•] is the expectation operator. Let’s assume that the analyst’s estimate and
the consensus are correlated, E[e, ec] ¼ r (t tc)1. This correlation reflects the
extent to which the analyst uses common private channels. The analyst objective,
Eq. 2.1, can be rewritten as

4
Here, we focus only on the case of one-period sequential forecasting. However, we believe that
the main implications of our model hold true for a multi-period sequential forecasting setting.
Since we assume that the probabilistic characteristics of different components are known and
analysts can gauge each others’ experience and the amount of information asymmetry perfectly,
there would be no incentive to deviate from posting commensurate optimal, rational forecasts. If
expert analysts intentionally deviate from their optimal forecasts, no other analyst can compensate
for their experience or information asymmetry [for more discussion, see Trueman (1990)].
5
For more details on Bayesian methods of inference and decision making, see Winkler (1972).
2 Experience, Information Asymmetry, and Rational Forecast Bias 69

h i
min w2 ðt0 þ tðbÞÞ2 þ ð1  wÞ2 tc 2 þ 2rwð1  wÞðt0 þ tðbÞÞ1 t1
c : (2.2)
wjb

The optimal weight which solves the aforementioned objective function is

ðt0 þ tðbÞÞ2  rtc ðt0 þ tðbÞÞ

w¼ (2.3)
ðt0 þ tðbÞÞ2 þ t2c  2rtc ðt0 þ tðbÞÞ

Proposition 1 Assuming positively biasing forecasts have increasing but

diminishing returns, i.e., ∂t/∂b > 0 and ∂2t/∂b2 < 0, the optimal weight on
analyst’s own unconditional estimate is concavely related to the analyst’s expertise
(i.e., self-accuracy, t0). For analysts with self-accuracy less (more) than – t(b) +
0.5 r1 tc, the optimal weight on the analyst’s own unconditional estimate
increases (decreases) with the expertise. At the optimal weight, the analyst’s
conditional precision, t0 + t(b), equals 0.5 r1 tc.
Proof See the Appendix 1.
As the analyst’s expertise, t0, increases, her unconditional estimate precision
rises relative to the consensus estimate precision. The analyst would then gain
more accuracy by placing more weight on her own assessment, consistent with the
findings of Chen and Jiang (2006). Forecast bias, however, affects how the
analyst’s own unconditional estimate weighs into the optimal forecast. As fore-
cast bias increases, the analyst relies more on her own assessment because larger
bias improves precision through more privately acquired information. However,
this reliance on private information is limited. When the precision of private
information reaches a certain limit, 0.5 r1 tc, the analyst starts to reduce her
reliance on her own assessment. Since the consensus contains information, an
analyst does not need to rely solely on private information to improve her overall
accuracy.
Interestingly, the threshold on the reliance of private information is inversely
related to the correlation between analyst’s own and consensus precisions. When
the signals from the consensus and analyst are highly correlated, the analyst need
not post largely biased forecasts to improve her accuracy. At low correlations,
however, the analyst almost exclusively relies on her own assessment and uses
private channels heavily by posting biased forecasts. As Mest and Plummer (2003)
show, when uncertainty about the company is high, management becomes a more
important source of information. This confirms Bernhardt et al. (2006) contention
that signal “correlatedness” affects analyst forecasting.
Proposition 1 has an interesting testable implication. As noted, following
previous studies (Lim 2001; Chen and Jiang 2006), our analysts minimize the
squared error to arrive at the optimal bias. In a cross section of analysts, this
implies that as long as all analysts follow the squared-error-minimization rule,
then the implication of Proposition 1 holds empirically: there would be a concave
relation between experience and the weight an analyst places on her own
70 A. Knill et al.

unconditional assessment. However, since the weighted average of self-

assessment and consensus belief is theoretically identical with forecast error,
this then also implies that in the cross section, analysts’ forecast errors and their
experience are also related concavely.
Note that our analysts choose how to combine their own unconditional forecast
and the consensus forecast to arrive at their reported forecast. They do so by
accounting for the inherent error in each of these forecasts and choose a weight
that minimizes the squared forecast error: the difference between reported forecast
(i.e., conditional forecast) and the observed earnings. So for a novice analyst, the
choice is what weight to put on her unconditional forecast knowing that the fore-
casts reported by other analysts contain much more experience and perhaps less
information asymmetry. In the extreme case, where the novice analyst has no
confidence on her own assessment, the optimal weight for her unconditional
forecast is zero. She will fully herd. A highly seasoned analyst does the same
thing: she also chooses the weight she puts on her assessment vis-à-vis consensus.
In the alternate extreme case, where the highly seasoned analyst has utmost
confidence on her assessment, she puts 100 % weight on her unconditional forecast.
Moderately seasoned analysts thus fall somewhere in between; the weight they put
on their own unconditional forecasts is between zero and one. In such a setting, all
analysts arrive at their own squared-error-minimizing forecast. From the onset,
however, as econometricians, we can only observe their reported conditional
forecast. This means that we can only focus on the implication of analyst squared-
error-minimizing exploiting cross-sectional differences in the data. As noted, the
observed error is equal to the weighted average of unconditional forecast and
the consensus belief. This implies that observed error is directly linked with the
unobservable weight. However from Proposition 1, we know that the unobservable
optimal weight is concavely related to experience, which in turn, implies that, in
the cross section of analysts, the observed forecast error is concavely linked with
the experience as well.
Proposition 2 Assuming positively biasing forecasts have increasing but
diminishing returns, i.e., ∂t/∂b > 0 and ∂2t/∂b2 < 0, the optimal weight on an
analyst’s own unconditional estimate increases monotonically with information
asymmetry (i.e., the marginal accuracy for bias, ∂t/∂b).
Proof See the Appendix 1.
As the efficacy of the analyst’s private information acquisition increases, that is,
as ∂t/∂b rises, the analyst gains more precision with every cent of bias. More
resourceful analysts, such as those employed by large investment houses,
either have better private channels or can gain more from the same channels
(see, e.g., Chen and Jiang 2006; Clement and Tse 2005). As such, these analysts’
estimates become more accurate more rapidly as they bias their forecasts to purchase
information from the firm. From proposition 1, we know that at the optimal weight,
the sum of the analyst’s own estimate and consensus precision is constant.
2 Experience, Information Asymmetry, and Rational Forecast Bias 71

As information asymmetry increases, i.e., marginal precision per bias rises, a larger
bias is needed to obtain the optimal weight of the analyst’s estimate and the
consensus. Interestingly, as an analyst becomes more experienced, her optimal
weight on her own estimate becomes less sensitive to information asymmetry.

2.3 Empirical Method

We empirically test selected comparative statics to draw conclusions about the

validity of our analytical results. We focus our attention on analytical predictions
that result directly from the endogeneity of private information acquisition. Spe-
cifically, we test three of our model’s implications: (1) whether analyst experience
and forecast deviation are concavely related, (2) whether information asymmetry
and analyst forecast deviation are positively related, and (3) whether there is
interplay between the effects of experience and information asymmetry. We follow
Chen and Jiang (2006) and empirically define bias as the difference between an
analyst’s forecast and the consensus estimate. We calculate consensus using only
the most recent estimate from each analyst and include only those estimates that are
less than 90 days old (Lim 2001).
We first examine the impact of analyst experience on forecast deviation. As
analysts become more experienced, they post a more positive (i.e., increasing)
deviation to improve their information gathering, resulting in better forecasts.
Proposition 1 from the theoretical model suggests that this relationship is
nonlinear. Highly seasoned analysts achieve forecast accuracy on their own with-
out relying on procured information, or they possess valuable relationship with
management that does not necessitate purchasing information, causing this
nonlinearity. Empirically, we expect the relationship between deviation and expe-
rience to be concave (i.e., a positive coefficient on experience and a negative
coefficient on experience squared). To test our contentions, we estimate the
following OLS panel regression:

DFCi, t ¼ a0 þ Z0 Experiencei, t þ Z1 Experiencei, t 2 þ b0 Num Revisionsi, t

þ b1 Same Quarteri, t þ b2 Reg:FDi, t þ F0 Xt4 þ F1 I þ F2 t þ e,
(2.4)

where DFC is the Deviation from the consensus for analyst i in quarter t. We have
several analyst-specific controls: Experience, Experience squared, and the
NumRevisions. Following our analytical model, analysts with more experience
should be better at forecasting earnings and should have better channels of private
information. Following Leone and Wu (2007), we measure Experience in two
ways: using an absolute measure (the natural log of the quarters of experience)
and a relative measure (the natural log of the quarters of experience divided by the
72 A. Knill et al.

average of this measure for analysts following firms in the same industry). We
include Experience2 to test for a nonlinear relationship between Deviation and
Experience. To control for varying degrees of accuracy due to intertemporal
Bayesian updating, we include NumRevisions, which is the number of times the
analyst revises her estimate. Studies such as Givoly and Lakonishok (1979) suggest
that information has been acquired (i.e., purchased). This suggests a positive rela-
tionship between the number of revisions and the resulting deviation; thus we expect
the coefficient, b0, to be positive. Same quarter is an indicator variable for horizon
value, where the variable is equal to one if the estimate occurs in the same quarter as the
actual earnings report is released and zero otherwise. Including Same quarter can help
us understand whether analysts are more likely to be optimistic early in the period than
late in the period (Mikhail et al. 1997; Clement 1999; Clement and Tse 2005).6 We
expect the coefficient on this variable, b1, to be negative. Reg. FD is an indicator
variable equal to one if the quarter date is after Reg. FD was passed (October 23, 2000),
and zero otherwise.7 Although extant literature is divided on whether or not Reg. FD
has decreased the information asymmetry for analysts, if information asymmetry is
decreased and more public information available, analysts will be less likely to
purchase private information, and the coefficient on Reg. FD, b2, would be negative.
Indeed, Zitzewitz (2002) shows that although private information has decreased post
Reg. FD, the amount of public information has improved. Both Brown et al. (2004) and
Irani (2004) show that information asymmetry decreased following Reg. FD.8
We control for firm effects through the inclusion of firm-specific variables such as
Brokerage reputation, Brokerage size, Accruals, Intangible assets, and Return st. dev.
(vector X in Eq. 2.4). Following Barber et al. (2000), we use the number of companies
a brokerage house follows per year in total as a proxy for brokerage size. Brokerage
house reputation may play an integral role in accessing to private information
(Agrawal and Chen 2012). To calculate broker reputation, we start with the Carter
and Manaster (1990), Carter et al. (1998) and the Loughran and Ritter (2004)
rankings. When a firm goes public, the prospectus lists all of the firms that are in
the syndicate, along with their shares. More prestigious underwriters are listed higher
in the underwriting section. Based upon where the underwriting brokerage firm is
listed, they are assigned a value of 0–9, where nine is the highest ranking. As Carter
and Manaster (1990) suggest that prestigious financial institutions provide a lower
level of risk (i.e., lower information asymmetry), we include control variables for
these characteristics of the firm and expect the relationships to be negative.9

6
Horizon value and the NumRevisions are highly correlated at 65 %. We therefore orthogonalize
horizon value in the equation to ensure that multicollinearity is not a problem between these two
variables.
7
http://www.sec.gov/rules/final/33-7881.htm.
8
See Lin and Yang (2010) for a study of how Reg. FD affects analyst forecasts of restructuring
firms.
9
Brokerage reputation and Brokerage size are highly correlated at 67 %. We therefore orthogo-
nalize brokerage reputation in the equation to ensure that multicollinearity is not a problem
between these two variables.
2 Experience, Information Asymmetry, and Rational Forecast Bias 73

As suggested by Bannister and Newman (1996) and Dechow et al. (1998),

companies that consistently manage their earnings are easier to forecast. We
include Accruals (in $ millions) to control for the possibility that earnings are
easier to forecast for companies that manage their earnings, resulting in less
information “bought” from the company.10 We expect the coefficient on this
control variable to be negative. We include Intangible assets based on the fact
that companies with greater levels of intangible assets are more difficult to forecast
based on uncertainty of future performance (Hirschey and Richardson 2004).11
Thus, we expect the coefficient to be positive. We include Return standard devia-
tion to proxy for firm risk. More volatile firms are less likely to voluntarily issue
public disclosures, making it necessary for analysts to purchase private information
(Waymire 1986). The standard deviation is defined as the monthly standard devi-
ation of stock returns over a calendar year. All firm-level variables are lagged one
year (i.e., four quarters). We include both industry (one-digit SIC code) and time
indicators to control for industries/times where it is easier to forecast (see Kwon
(2002) and Hsu and Chiao (2010), e.g., to see how analyst accuracy differs across
industry).12 Finally, a formal fixed effects treatment around analysts is taken to
ensure that standard errors are not understated.
Our theoretical analysis also examines the effect of information asymmetry on
forecast deviation. In an environment with high information asymmetry, analysts
without adequate, reliable resources need to post a positive deviation to access
information, as shown in Proposition 2. We expect a positive coefficient on
Information asymmetry. We test this relationship by estimating the following
OLS panel regression:

DFCi, t ¼ a0 þ l0 Information Asymmetryi, t þ b0 Num Revisionsi, t þ b1 Same Quarteri, t

þ b2 Reg:FDi, t þ F0 Xt4 þ F1 I þ F2 t þ e,
(2.5)

where t is the quarter in which we measure deviation and i denotes the ith analyst.
We define Information asymmetry three different ways: (1) the inverse of analyst
coverage, i.e., 1/(number of brokerage houses following the firm), (2) the standard
deviation of the company’s forecasts (102), and (3) the relative firm size, i.e., the
difference between the firm’s assets and the quarterly median assets for
the industry. These definitions are constructed such that the direction of the
expected marginal coefficient is congruent with that of information asymmetry.

10
Following Stangeland and Zheng (2007), we measure Accruals as income before extraordinary
items (Data #237) minus cash flow from operations, where cash flow from operations is defined as
net cash flow from operating activities (Data #308) minus extraordinary items and discontinued
operations (Data #124).
11
Following Hirschey and Richardson (2004), we calculate Intangibles as intangible assets to total
assets (Data 33/Data #6).
12
As an alternate proxy for industry fixed effects, Fama-French 12-industry classifications (Fama
and French 1997) are used. Results using these proxies are available upon request.
74 A. Knill et al.

The first definition, the inverse of analyst coverage, is based on evidence that the
level of financial analyst coverage affects how efficiently the market processes
information (Bhattacharya 2001). Analyst coverage proxies for the amount of
public and private information available for a firm and, therefore, captures
a firm’s information environment (Zhang 2006). The second definition, Forecast
dispersion, is supported in papers such as Krishnaswami and Subramaniam (1998)
and Thomas (2002), among others. Lastly, we use Relative firm size as a proxy for
information asymmetry. This is equally supported by the literature, including but
not limited to Petersen and Rajan (1994) and Sufi (2007). The firm-specific control
variables are the same as in Eq. 2.4, and a fixed effects treatment around analysts is
taken.
As previously explained, our model allows for the interaction of experience and
information asymmetry; the concavity of the relationship between experience and
Deviation may change at different levels of information asymmetry. Thus, we
estimate Eq. 2.4 for analysts with both high and low levels of information asym-
metry (i.e., based on relation to the industry-quarter median forecast dispersion).
We expect the coefficients on Experience and Experience squared to increase when
there is more information asymmetry.
Similarly, we demonstrate that experience affects the link between information
asymmetry and deviation. While the deviation of the novice and highly seasoned
analysts is only marginally affected by information asymmetry, the deviation of the
moderately experienced analyst is highly affected by information asymmetry. Since
Proposition 2 suggests that information asymmetry will affect analysts differently
based on their experience, we segment our sample into three groups: novice,
moderately experienced, and highly seasoned analysts. To form these experience
groups, we create quarterly terciles of analysts based on their experience. The
bottom third comprises the novice analysts; the middle third, moderately seasoned
analysts; and the top third, highly seasoned analysts. We examine the model in
Eq. 2.5 separately for our three experience subsamples. The coefficient on infor-
mation asymmetry should be larger for experienced analysts.
It is possible that a resolution of idiosyncratic uncertainty makes deviation
insensitive to both analyst experience and information asymmetry. Regulations
such as Reg. FD were enacted to reduce the uncertainty around companies
(De Jong and Apilado 2008). If Reg. FD reduced firm-specific uncertainty more
than common uncertainty, then the coefficients on Experience, Experience squared,
and information asymmetry should be smaller after Reg. FD. In other words, if Reg.
FD was effective in “leveling the playing field” for analysts with regard to prefer-
ential treatment of some analysts over others, the implications of this model should
no longer exist.
Extant evidence suggests that by prohibiting exclusive private communication of
pertinent information, Reg. FD would cause overall earnings uncertainty to decline
(Baily et al. 2003). In Eqs. 2.4 and 2.5, we simply control for any effect that Reg.
FD might have. By using interactive terms, however, we can explore this relation-
ship more fully by offering a comparison of the relationship between Experience/
Information asymmetry and Deviation both with and without Reg. FD. If most of
2 Experience, Information Asymmetry, and Rational Forecast Bias 75

the decrease in uncertainty comes from improving the common component of

earnings (or equivalently, reducing the amount of private information), then the
enactment of Reg. FD should make any remaining private information even more
valuable and should lead to an increase in the importance of information asymmetry
on deviation. Further, this effect should be disparate based on analyst experience.
Specifically, experienced analysts (moderately experienced more so than highly
seasoned) should see a drop in their deviation from consensus based on an increase
in information asymmetry since their private information is no longer quite so
private, i.e., there is less information to buy. Novice analysts, on the other hand, will
likely be relatively unaffected since they depend mostly on consensus anyway. In
short, since the intent of the regulation is to even the playing field for analysts, the
effect of Reg. FD on the impact of experience could be a reduction in its impor-
tance. This would especially be the case with relative experience. If Reg. FD
achieved what it set out to achieve, we should see a reduction (i.e., flattening) of
the concavity of the experience relationship with deviation from consensus.

2.4 Data

Consistent with Brown and Sivakumar (2003) and Doyle et al. (2003), we define
earnings as the First Call reported actual earnings per share.13 Our sample includes
companies based in the United States with at least two analysts, consists of 266,708
analyst-firm-quarter forecasts from 1995 to 2007, and includes all analysts’ revi-
sions. Variables are winsorized at the 1 % level to ensure that results are not biased
by outliers.
First Call data is well suited for examining analyst revisions because most of the
analysts’ estimates in the data have the date that they were published by the broker.
These revisions are reflected daily, which aids in understanding the changes in
deviation based on changes in the information environment. One limitation with
First Call data is that it identifies only brokerage houses, not individual analysts.
Following Leone and Wu (2007), we make the assumption that for each firm quarter
there is only one analyst in each brokerage house following the firm. It is notewor-
thy that this biases our study against finding a nonlinear impact for analyst’s
experience.
Panel A of Table 2.1 shows the summary statistics of the variables used in our
analysis. The average Deviation from the consensus is 5.16¢. However, there is
wide dispersion, from 54.67¢ to 112.50¢, as compared to an average Forecast
error of 3.79¢, with a range of 131–72¢. The analysts in our sample have on
average 12.71 quarters of Experience (the most experienced analyst has 13 years
experience, and the least experienced analyst has no prior experience following the
firm). Analysts revise their estimates 2.49 times per quarter. The companies they

13
As a robustness test, we use I/B/E/S data. Results may be found in Appendix 2.
 76

Table 2.1 Data characteristics

Panel A. Summary statistics
Variable N Source Mean Median Std. dev. Min Max
DFC (¢) 266,708 First call; own calculation 5.16*** 1*** 20.43 54.67 112.5
Quarters follow 266,708 First call; own calculation 12.71*** 10*** 10.37 1 52
Experience 266,708 First call; own calculation 2.15*** 2.30*** 0.98 0 3.81
Relative experience 266,708 First call; own calculation 0.97*** 1*** 0.51 0.07 2.40
Analyst coverage1 266,708 First call; own calculation 0.21*** 0.13*** 0.22 0.03 1
Forecast dispersion (¢) 253,673 First call; own calculation 7.90*** 4.19*** 10.66 0 67.44
Relative firm size 217,126 COMPUSTAT; own calculation 9.68*** 0*** 60.80 1430.35 671.78
Forecast error (¢) 266,708 First call; own calculation 3.79*** 0*** 24.00 131 72
Same quarter 266,708 First call; own calculation 0.21*** 0*** 0.41 0 1
Broker reputation 266,708 Carter and Manaster (1990) 0.02*** 0.20*** 1.46 4.80 3.77
Broker size 266,708 First call; own calculation 6.36*** 6.63*** 1.03 0 7.79
Reg. FD 266,708 Securities Exchange Commission 0.63*** 1*** 0.48 0 1
Number of revisions 266,708 First call; own calculation 2.49*** 2*** 2.60 0 12
Accruals 266,708 COMPUSTAT 0.44*** 0.09*** 1.09 7.52 0.90
Intangible assets 266,708 COMPUSTAT 0.14*** 0.06*** 0.18 0 0.92
Return std. dev. 266,708 CRSP 0.13*** 0.11*** 0.08 0.03 0.47
Panel B. Correlation
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
DFC (1) 1.00
Quarters follow (2) 0.02 1.00
Experience (3) 0.01 0.89 1.00
Relative experience (4) 0.01 0.70 0.72 1.00
A. Knill et al.
 2

Analyst coverage1 (5) 0.01 0.19 0.18 0.01 1.00

Forecast dispersion (6) 0.32 0.06 0.05 0.00 0.11 1.00
Relative firm size (7) 0.01 0.10 0.08 0.01 0.10 0.03 1.00
Forecast error (8) 0.89 0.04 0.00 0.01 0.03 0.27 0.03 1.00
Same quarter (9) 0.10 0.14 0.26 0.13 0.02 0.03 0.00 0.09 1.00
Broker reputation (10) 0.02 0.01 0.03 0.04 0.02 0.01 0.01 0.02 0.00 1.00
Broker size (11) 0.00 0.30 0.32 0.19 0.18 0.03 0.05 0.01 0.05 0.08 1.00
Reg. FD (12) 0.05 0.33 0.28 0.01 0.17 0.02 0.06 0.07 0.05 0.18 0.19 1.00
NumRevisions (13) 0.14 0.21 0.27 0.12 0.15 0.21 0.03 0.11 0.00 0.03 0.10 0.09 1.00
Accruals (14) 0.01 0.17 0.14 0.02 0.19 0.11 0.41 0.00 0.01 0.03 0.08 0.09 0.06 1.00
Intangible assets (15) 0.00 0.02 0.03 0.01 0.02 0.10 0.06 0.01 0.02 0.02 0.07 0.15 0.04 0.01 1.00
Return std. dev. (16) 0.13 0.21 0.18 0.04 0.09 0.12 0.10 0.12 0.05 0.03 0.08 0.05 0.05 0.10 0.03 1.00
Analyst information comes from First Call for the term 1995–2007. Company information comes from Compustat/CRSP. DFC is the bias, defined as the
difference between an analyst’s forecast and the consensus estimate, where we calculate consensus using only the most recent estimate from each analyst and
include only those estimates that are less than 90 days old. Quarters follow is the number of quarters a brokerage firm has been following a company.
Experience is the natural log of the quarters of experience. Relative experience is the natural log of the quarters of experience divided by the average of this
measure for analysts following firms in the same industry. Analyst coverage1 is equal to 1/(number of brokerage houses following the firm), and forecast
dispersion is the standard deviation of the company’s forecasts (102). Relative firm size is the difference between the firm’s assets and the quarterly median
assets for the industry. Forecast error is the difference between the estimate and the actual earnings. Same quarter is a dummy variable equal to one if the
estimate is in the same quarter as the actual and zero otherwise. Brokerage reputation is a ranking of brokerage reputation, where 0 is the worst and 9 is the best.
Brokerage size is the natural log of the number of companies per quarter a brokerage house follows. Reg. FD is an indicator variable that takes on a value of 1 if
Experience, Information Asymmetry, and Rational Forecast Bias

Reg. FD is in effect and 0 otherwise. NumRevisions is the total number of revisions the analyst made during the quarter for the quarter-end earnings. Accruals
are income before extraordinary items minus cash flow from operations, where cash flow from operations is defined as net cash flow from operating activities
minus extraordinary items and discontinued operations. Intangible assets are intangible assets to total assets. Return std. dev. is the standard deviation of the
12-month returns. Panel A shows the univariate statistics and Panel B shows the pairwise correlations
Asterisks in Panel A represent statistical significance relative to zero
Bolded numbers in Panel B represent 5 % or 1 % significance
77
78 A. Knill et al.

follow have a Return standard deviation of 13 %, Intangibles of 14 %, and Accruals

of $0.44 (in millions) on average.
Panel B of Table 2.1 shows the correlation matrix for the variables used in the
analysis. There exists notable significant relation in the variables: Forecast error and
Experience with Deviation from consensus. The relation actually foreshadows one
of the main results of the paper, which is that private information can be used to
decrease forecast error. The correlation found in this table, 0.89, is not a problem
econometrically because fitted values are used in the specification found in
Table 2.5. The only other correlations that would be considered a problem
econometrically are for variables not used in the same specification, i.e., Experience
and Relative experience.

2.5 Empirical Findings

Table 2.2 presents the results of our estimation from Eq. 2.4, which examines how
analyst experience affects analyst deviation from the consensus. The coefficients on
our control variables all exhibit the expected signs. There is a negative relation
between the horizon value (Same quarter) and Deviation, which suggests that
analysts are more likely to be optimistic early in the period than late in the period.
Supporting the contentions of Carter and Manaster (1990), there is a negative
relationship between the brokerage characteristics – Reputation and Size – with
Deviation. There is also a negative relationship between Reg. FD and Deviation,
suggesting that the enactment of Reg. FD and its mandatory indiscriminant infor-
mation revelation have made forecasting easier. NumRevisions is positively related
to Deviation; revisions are made when valuable information content is received
(Givoly and Lakonishok 1979), suggesting that private information is paid for
through incremental deviation of their forecasts over time. We find that higher
Accruals lead to less deviation from the consensus. Firms who actively manage
earnings make forecasting easier; thus analysts do not have to purchase
private information. The positive sign on Intangible assets is expected as more
intangible assets make forecasting more difficult, necessitating the procurement of
private information. Finally, as the standard deviation of Returns increases, future
firm performance is more difficult to predict.
Turning to our variables of interest, we find that there is a highly significant
positive relationship between Experience and Deviation, as evidenced by a 0.829¢
increase in deviation for every additional unit of experience (specification 1) and
a 0.581¢ increase in deviation for every additional unit of relative experience
(specification 2).14 The results in specifications (3) and (4) confirm our contention
that this relationship is not linear. The squared Experience variable

14
Inasmuch as the Experience variable is transformed using the natural logarithm, one unit of
experience is approximately equal to two quarters of experience. For tractability, we refer to this as
a unit in the empirical results.
 2

Table 2.2 Impact of experience on analyst forecasting

All estimates Early estimates (horizon is longer) Late estimates (horizon is shorter) Estimates in low IA environment Estimates in high IA environment
1 2 3 4 5 6 7 8 9 10 11 12
Experience 0.829*** 3.024*** 6.751*** 1.945*** 0.950*** 4.745***
[0.049] [0.153] [0.571] [0.137] [0.109] [0.283]
Experience2 0.595*** 1.341*** 0.360*** 0.169*** 0.986***
[0.039] [0.118] [0.038] [0.028] [0.073]
Relative experience 0.581*** 4.250*** 2.961*** 2.712*** 1.781*** 5.572***
[0.081] [0.263] [0.301] [0.167] [0.186] [0.488]
RelExp2 1.735*** 0.953*** 0.764*** 0.712*** 2.267***
[0.118] [0.087] [0.055] [0.083] [0.221]
Same quarter 2.372*** 2.472*** 1.992*** 2.364*** 0.833*** 0.935*** 4.181*** 4.791***
[0.107] [0.107] [0.110] [0.107] [0.078] [0.076] [0.205] [0.200]
NumRevisions 1.204*** 1.232*** 1.189*** 1.227*** 0.958*** 0.965*** 1.545*** 1.569*** 0.499*** 0.512*** 1.388*** 1.444***
[0.016] [0.016] [0.016] [0.016] [0.025] [0.025] [0.030] [0.030] [0.013] [0.013] [0.027] [0.027]
Reg. FD 2.458*** 2.648*** 2.275*** 2.403*** 0.731* 0.591 1.687*** 1.467*** 0.368** 0.391** 3.394*** 3.613***
[0.227] [0.227] [0.228] [0.228] [0.388] [0.387] [0.406] [0.405] [0.167] [0.167] [0.408] [0.408]
Broker reputation 0.508*** 0.529*** 0.495*** 0.499*** 0.642*** 0.573*** 0.300*** 0.242** 0.205*** 0.206*** 0.714*** 0.714***
[0.103] [0.103] [0.103] [0.103] [0.185] [0.185] [0.104] [0.104] [0.075] [0.075] [0.186] [0.186]
Broker size 0.453*** 0.493*** 0.690*** 0.518*** 0.695** 0.696** 0.489*** 0.391** 0.519*** 0.476*** 0.759*** 0.451*
[0.151] [0.151] [0.152] [0.151] [0.275] [0.275] [0.153] [0.152] [0.112] [0.112] [0.270] [0.268]
Accruals 0.122***c 0.186*** 0.165*** 0.159*** 0.363*** 0.373*** 0.114*** 0.112*** 0.204*** 0.205*** 0.465*** 0.452***
[0.037] [0.037] [0.037] [0.037] [0.061] [0.061] [0.041] [0.041] [0.028] [0.028] [0.066] [0.066]
Experience, Information Asymmetry, and Rational Forecast Bias

Intangible assets 1.370*** 1.244*** 1.237*** 1.264*** 2.718*** 2.718*** 0.097 0.146 1.415*** 1.411*** 1.974*** 2.054***
[0.232] [0.232] [0.232] [0.231] [0.393] [0.393] [0.241] [0.242] [0.161] [0.161] [0.440] [0.440]
Return std. dev. 19.384*** 17.800*** 19.063*** 17.700*** 26.067*** 26.308*** 7.387*** 7.448*** 7.944*** 7.265*** 17.102*** 15.640***
[0.578] [0.571] [0.578] [0.571] [0.975] [0.976] [0.612] [0.612] [0.448] [0.441] [1.003] [0.991]
Constant 6.239*** 8.269*** 6.501*** 6.913*** 2.193 4.345* 6.268*** 6.633*** 6.286*** 6.423*** 13.895*** 14.185***
[1.439] [1.434] [1.438] [1.437] [2.682] [2.621] [1.467] [1.469] [0.975] [0.974] [3.041] [3.038]
Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

(continued)
79
 80

Table 2.2 (continued)

All estimates Early estimates (horizon is longer) Late estimates (horizon is shorter) Estimates in low IA environment Estimates in high IA environment
1 2 3 4 5 6 7 8 9 10 11 12
Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 266,708 266,708 266,708 266,708 134,296 134,296 132,412 132,412 133,574 133,574 133,134 133,134
# Brokerage houses 130 130 130 130 121 121 130 130 128 128 128 128
R-squared 0.06 0.06 0.06 0.06 0.07 0.07 0.04 0.04 0.04 0.04 0.08 0.08

Table 2.2 presents the results of our estimation from Eq. 2.4. Experience is the natural log of the number of quarters the analyst has followed the firm for which
forecast error is calculated. Relative experience is the analyst experience scaled by the average analyst experience (same industry). Information asymmetry
(IA) is proxied based on the forecast dispersion – the standard deviation of estimates. Same quarter is a dummy variable equal to one if the estimate is in the
same quarter as the actual and zero otherwise. Same quarter is orthogonalized (on NumRevisions) to ensure that multicollinearity is not a problem between
these two variables. NumRevisions is the total number of revisions the analyst made during the quarter for the quarter-end earnings. Reg. FD is an indicator
variable that takes on a value of 1 if Reg. FD is in effect and 0 otherwise. X is a vector of firm-specific variables including Broker reputation, Broker size,
Accruals, Intang. assets, and Return std dev. Brokerage reputation is a ranking of brokerage reputation, where 0 is the worst and 9 is the best. Brokerage size is
the natural log of the number of companies per quarter a Brokerage house follows. Brokerage reputation is orthogonalized (on brokerage size) to ensure that
multicollinearity is not a problem between these two variables. Accruals are the accrued revenue/liabilities utilized for earnings smoothing. Intang. assets are
the covered firm’s intangible assets value relative to its total assets. Return std. dev. is the standard deviation of the covered firm’s return. I is a vector of
one-digit SIC industry dummies. T is a vector of time dummies. A formal fixed effects treatment around analysts is employed. Standard errors are reported in
brackets
* **
, , and *** indicate significance levels of 10 %, 5 %, and 1 %, respectively (two-tailed test). In columns 1–4 we use the full sample. In columns 5–8, we use
the median value of horizon value to divide the firms into early (columns 5–6) and late estimates (7–8) groups to perform the analysis. In columns 9–12, we use
the median value of forecast dispersion to divide the firms into low information asymmetry (columns 9–10) and high information asymmetry (11–12) groups to
perform the analysis
A. Knill et al.
2 Experience, Information Asymmetry, and Rational Forecast Bias 81

(both definitions) is significantly and negatively related to Deviation (0.595¢ in

specification 3 for Experience and 1.735¢ in specification 4 for Relative experi-
ence). Confirming the analytical results, the empirical results suggest that there
exists a trade-off between public and private information. The results suggest that
when analysts first begin to follow a firm and gain experience, they are more likely
to post small deviations, indicating that they weigh public information (i.e., previ-
ous forecasts and the consensus), more heavily than private information. A possible
explanation is that they have not yet acquired the preferred access to information
that long-term relationships afford experienced analysts or the analytical expertise
to wade through noisy information. As analysts gain confidence in their analytical
ability and relationships with key personnel within the firm, however, they are more
likely to post large deviations to gain private information. Highly seasoned analysts
put almost no weight on the consensus in creating their earnings forecasts and rely
almost exclusively on their own private information. This information costs them
very little, either because they have honed their analytical ability so well that they
don’t need the information or because the cost of said information is reduced due to
their preferential relationship with the firm.
If our hypotheses are true, we would expect that analysts are more optimistic
early in the period and less so (and perhaps even pessimistic) later in the period
(shortly before earnings are announced).15 In order to test this, we segment our
sample by the median horizon value into two categories, early estimates (i.e., longer
horizon) and late estimates (i.e., shorter horizon). The results are substantially
stronger for earlier estimates as compared to late estimates. The coefficient on the
Experience variable is 6.751 when the horizon value in long and 1.945 when there is
a short horizon value (specification 5 vs. 7). The coefficients on the squared
Experience variable are also more extreme for the earlier analysts (1.341
vs. 0.360), showing that the concavity of the function is more pronounced early
in the forecasting period and less so later. This pronounced concavity suggests that
the distinction between moderately experienced analysts and novice/highly expe-
rienced analysts is more pronounced earlier in the estimation period. Essentially,
for every additional quarter of experience, there is a 5.410¢ increase in deviation
when horizon is long and a 1.585¢ increase in deviation when horizon is short.
Results for Relative experience are qualitatively identical, albeit with a reduced
difference between early and late. Results are qualitatively identical when we
divide horizon into terciles and define early as the highest quartile and late as the
lowest. We can conclude from these results that analysts are indeed more optimistic
earlier in the estimation period.
Next, we rerun the model in Eq. 2.4 on low and high information asymmetry
subsamples (using forecast dispersion and segmenting at the median), shown in
Table 2.2, columns 9–12. As Proposition 2 suggests, the empirical results indicate
that the impact of experience on optimal deviation is smaller for analysts when
information asymmetry is low. One additional quarter of Experience increases the

15
We are grateful to an anonymous referee for this point.
82 A. Knill et al.

analyst deviation by 0.950¢ (specification 9) in a low information asymmetry

environment versus 4.745¢ (specification 11) in a high information asymmetry
environment. One additional unit of Relative experience increases the analyst
Deviation by 1.781¢ (specification 10) in a low information asymmetry environ-
ment versus 5.572¢ (specification 12) in a high information asymmetry environ-
ment. We note once again that this relationship is not linear, as evidenced by the
statistically significant negative squared term. This implies that moderately expe-
rienced analysts post higher deviations from consensus than do their less and more
experienced colleagues.
Table 2.3 provides alternate specifications around the Reg. FD control variable.
Specifications (1) through (4) exclude the Reg. FD control to ensure that results are
not reliant on its inclusion. Overall, results from Table 2.2 remain consistent. That
said, we note that Relative experience in Specification (2) is no longer statistically
significant. The substantiation of the nonlinearity of this proxy for Experience in
Specification (4), albeit muted, suggests that this is likely due to the fact that we are
no longer controlling for the impact of Reg. FD on the deviation from consensus,
which we would expect would affect Relative Experience more so than experience
(i.e., no longer benefitting highly seasoned analysts for the preferred relationships
through better access to information) through the concavity of the function. In other
words, without controlling for Reg. FD independently, the “average” impact of the
proxy (i.e., before and after Reg. FD) is muted.
Specifications (4) through (8) explore further the effect Reg. FD has on the
relationship between Experience and Deviation. We first note that the nonlinearity
of the relationship between Experience and forecast deviation (Deviation) is intact,
indicating that leveling the playing field through Reg. FD has not (completely)
changed how forecasts are fundamentally linked with the information. Focusing on
the more potent results, we look to the Relative experience results (Specifications
7 and 8), and we note that Reg. FD is effective in reducing the private information
component, which in turn reduces the importance of relative experience. Empiri-
cally, this translates into a reduction in the concavity of the impact of Relative
experience; the “leveling of the playing field” flattens out that curve. We see this in
the Reg. FD interaction terms: specifically, a positive and significant marginal
effect on (Reg. FD * Relative experience) and negative and significant marginal
effect on (Reg. FD * Relative experience2). More concretely stated, when access to
private information is reduced, the relationship-based competitive advantage of
relative experience – preferential access to private information – is taken away.
These results highlight a distinction between precision and the value of prefer-
ential relationships between veteran analysts and management of the firm. Since all
analysts gain analytical expertise over time following the firm (though the marginal
effect of this would seem to wane for the most experienced, thus explaining the
maintained nonlinearity), but only some may gain preferential treatment by the
firm, one could argue that Experience is a better proxy for analytical precision from
the model and Relative experience is a better proxy for the preferential treatment
provided to veteran analysts that possess valuable relationships with top manage-
ment of the firms they are covering (i.e., management will provide private
 2

Table 2.3 Time impact of Reg. FD on the relationship between experience/relative experience and deviation
Dependent variable ¼ DFC (deviation from consensus)
1 2 3 4 5 6 7 8
Experience 0.302*** 2.116*** 1.528*** 2.764***
[0.050] [0.157] [0.088] [0.271]
Experience2 0.485*** 0.400***
[0.040] [0.082]
Reg. FD * experience 0.962*** 0.026
[0.100] [0.326]
Reg. FD * experience2 0.168*
[0.092]
Relative experience 0.072 0.511* 0.822*** 7.019***
[0.082] [0.269] [0.149] [0.503]
Relative experience2 0.274** 3.128***
[0.120] [0.243]
Reg. FD * relative experience 0.331* 3.715***
Experience, Information Asymmetry, and Rational Forecast Bias

[0.172] [0.586]
Reg. FD * relative experience2 1.831***
[0.277]
Reg. FD 0.319 2.308*** 1.398*** 0.975***
[0.318] [0.288] [0.362] [0.358]
(continued)
83
 84

Table 2.3 (continued)

Dependent variable ¼ DFC (deviation from consensus)
1 2 3 4 5 6 7 8
Firm FE Yes Yes Yes Yes Yes Yes Yes Yes
Industry FE Yes Yes Yes Yes Yes Yes Yes Yes
Time FE Yes Yes Yes Yes Yes Yes Yes Yes
Observations 266,708 266,708 266,708 266,708 266,708 266,708 266,708 266,708
# Brokerage houses 130 130 130 130 130 130 130 130
Model R2 0.06 0.06 0.06 0.06 0.06 0.06 0.06 0.06
Table 2.3 provides alternate specifications around the Reg. FD control variable. Experience is the natural log of the number of quarters the analyst has followed
the firm for which forecast error is calculated. Relative experience is the analyst experience scaled by the average analyst experience (same industry). Reg. FD
is an indicator variable that takes on a value of 1 if Reg. FD is in effect and 0 otherwise. Control variables included in the analysis but not shown above include
same quarter, NumRevisions, Reg. FD, Broker reputation, Broker size, Accruals, Intang. assets, and Return std. dev. Same quarter is a dummy variable equal
to one if the estimate is in the same quarter as the actual and zero otherwise. Same quarter is orthogonalized (on NumRevisions) to ensure that multicollinearity
is not a problem between these two variables. NumRevisions is the total number of revisions the analyst made during the quarter for the quarter-end earnings.
Brokerage reputation is a ranking of Brokerage reputation, where 0 is the worst and 9 is the best. Brokerage size is the natural log of the number of companies
per quarter a brokerage house follows. Brokerage reputation is orthogonalized (on brokerage size) to ensure that multicollinearity is not a problem between
these two variables. Accruals are the accrued revenue/liabilities utilized for earnings smoothing. Intang. assets are the covered firm’s intangible assets value
relative to its total assets. Return standard deviation is the standard deviation of the covered firm’s return. I is a vector of one-digit SIC industry dummies. T is
a vector of time dummies. A formal fixed effects treatment around analysts is employed. Standard errors are reported in brackets
* **
, , and *** indicate significance levels of 10 %, 5 %, and 1 %, respectively (two-tailed test)
A. Knill et al.
2 Experience, Information Asymmetry, and Rational Forecast Bias 85

information to only analysts with the oldest relationships). In short, Reg. FD

reduces the importance of the competitive advantage of preferential treatment
(proxied by Relative experience) while retaining the importance of analytical
ability (proxied by Experience). This is seen empirically in the insignificant change
in concavity of Experience and a decrease in the concavity of Relative experience.
In this light, we see that the impact of Reg. FD is what we would expect, both with
regard to Experience (seen in Specifications 5 and 6) and Relative experience.
Panel A of Table 2.4 shows the results of Eq. 2.5, which examines directly how
Information asymmetry affects Deviation, controlling for firm, analyst, and regu-
latory factors. Specifications (1) through (3) look at the effects of Information
asymmetry on the full sample of analysts, using the inverse of Analyst coverage,
Forecast dispersion, and Relative firm size, respectively. The first proxy
(specification 1) is the inverse of Analyst coverage. As predicted in the theoretical
section, Information asymmetry is positively related to Deviation. With fewer
resources at their disposal, analysts post greater deviations as a means to purchase
private information. This supplements the marginally valuable public information
available, leading to more accurate forecasts. When information asymmetry is
measured by Analyst coverage, a unit increase in the information asymmetry
measure leads to a 0.630¢ additional Deviation. When using Forecast dispersion
as the proxy, a one-unit increase in the information asymmetry measure leads to
a 0.630¢ increase in Deviation. Lastly, using Relative firm size as the proxy for
information asymmetry, a one-unit increase in the information asymmetry measure
leads to 0.006¢ additional Deviation. From the multivariate analysis, it is evident
that Forecast dispersion is the best proxy for Deviation from the consensus. The
specifications where information asymmetry is proxied by Forecast dispersion have
approximately two times the predictive power of the analyst coverage and size
proxy (R-Squared of 0.15 vs. 0.06 and 0.08, respectively).
In specifications (4)–(6) of Table 2.4, we include indicator variables for the three
experience terciles (suppressing the constant term in the model) to examine whether
information asymmetry affects analysts in a different manner based on experience.
We first note that the marginal effect on Information asymmetry remains positive as
in the first three specifications. Looking to the marginal effect of novice
analysts – the least experienced third of analysts in a given quarter – they find
sticking closer to the herd a better alternative (Zhou and Lai 2009). The cost of
acquiring precise information for an inexperienced analyst who lacks necessary
resources would be prohibitively large, i.e., the analyst would need to post an
outrageously optimistic (positively deviated) forecast. The marginal effect on
moderately experienced analysts – the middle third in experience in a given
quarter – suggests that they post deviations that are more positive than the average
analyst in an effort to purchase valuable private information. The marginal effect
for the tercile of highly seasoned analysts suggests that they post a deviation with
a smaller magnitude. This nonlinear relationship exists regardless of the proxy for
information asymmetry. Controlling for the information asymmetry in the estima-
tion environment, these results fall in line with our prediction of nonlinearity for
Experience and our conjecture that highly seasoned analysts possess both superior
 86

Table 2.4 Impact of information asymmetry on analyst forecasting

Panel A: All analysts
Dependent variable ¼ DFC (deviation from consensus)
1 2 3 4 5 6
1
Analyst coverage 0.630*** 0.627***
[0.191] [0.188]
Forecast dispersion 0.630*** 0.626***
[0.004] [0.004]
Relative firm size 0.006*** 0.007***
[0.001] [0.001]
Novice 5.015*** 2.756*** 2.305**
[1.079] [1.045] [1.156]
Moderately experienced 6.495*** 3.920*** 3.959***
[1.080] [1.046] [1.157]
Highly seasoned 6.100*** 3.320*** 3.824***
[1.082] [1.049] [1.159]
Firm FE Yes Yes Yes Yes Yes Yes
Industry FE Yes Yes Yes Yes Yes Yes
Time FE Yes Yes Yes Yes Yes Yes
Observations 266,708 253,673 217,126 266,708 253,673 217,126
A. Knill et al.
 2

# Brokerage houses 130 130 124 130 130 124

Model R2 0.06 0.15 0.08 0.12 0.21 0.14
Novice ¼ Moderate? 75*** 62*** 68***
Moderate ¼ Highly seasoned? 26*** 49*** 7***
Novice ¼ Moderate ¼ Highly seasoned? 38*** 39*** 35***
Panel B: Experience subsamples
Dependent variable ¼ DFC (deviation from consensus)
Information asymmetry¼ Coverage Forecast dispersion Relative firm size
1 2 3
Novice 0.591** 0.555*** 0.007***
[0.264] [0.006] [0.002]
Observations 91,993 85,906 73,888
# Brokerage houses 130 130 124
Model R2 0.05 0.14 0.06
Moderately experienced 1.075*** 0.720*** 0.014***
[0.334] [0.007] [0.002]
Observations 88,392 83,878 72,656
# Brokerage houses 115 114 109
Model R2 0.07 0.19 0.09
Experience, Information Asymmetry, and Rational Forecast Bias

Highly seasoned 0.456 0.624*** 0.001

[0.432] [0.007] [0.001]
Observations 86,323 83,889 70,582
# Brokerage houses 85 85 82
Model R2 0.08 0.15 0.09
(continued)
87
 88

Table 2.4 (continued)

Panel B: Experience subsamples
Dependent variable ¼ DFC (deviation from consensus)
Information asymmetry¼ Coverage Forecast dispersion Relative firm size
1 2 3
Novice ¼ Moderate? 2,780*** 7,709*** 2,421***
Moderate ¼ Highly seasoned? 1,978*** 5,862*** 3,742***
Table 2.4 shows the results of Eq. 2.5. We create three groups of experience: the bottom third is the novice analysts, the middle third are the moderately
seasoned analysts, and the top third are the highly seasoned analysts. Information asymmetry is defined in three ways: Analyst coverage1, Forecast
dispersion, and Relative firm size. Analyst coverage1 is the inverse of the number of analyst covering the firm. Forecast dispersion is the standard deviation of
estimates. Relative firm size is equal to the covered firm size minus industry-quarter median firm size (in $billions). Control variables included in the analysis
but not shown above include Same quarter, NumRevisions, Reg. FD, Broker reputation, Broker size, Accruals, Intang. assets, and Return std. dev. Same
quarter is a dummy variable equal to one if the estimate is in the same quarter as the actual and zero otherwise. Same quarter is orthogonalized
(on NumRevisions) to ensure that multicollinearity is not a problem between these two variables. NumRevisions is the total number of revisions the analyst
made during the quarter for the quarter-end earnings. Brokerage reputation is a ranking of brokerage reputation, where 0 is the worst and 9 is the best.
Brokerage size is the natural log of the number of companies per quarter a brokerage house follows. Brokerage reputation is orthogonalized (on brokerage size)
to ensure that multicollinearity is not a problem between these two variables. Accruals are the accrued revenue/liabilities utilized for earnings smoothing.
Intang. assets are the covered firm’s intangible assets value relative to its total assets. Return standard deviation is the standard deviation of the covered firm’s
return. I is a vector of one-digit SIC industry dummies. T is a vector of time dummies. A formal fixed effects treatment around analysts is employed. Panel
A Specifications 1–3 use the full sample and do not suppress the constant. Panel A Specifications 4–6 suppress the constant so we can include the three
experience indicator variables. To test whether the coefficients on the experience indicator variables are statistically different, we use a Wald test for
significance. Panel B shows the results of the same analysis using experience-based subsamples. We use a Hausman estimation to test whether there is
statistical significance between each experience estimation. Standard errors are reported in brackets
* **
, , and *** indicate significance levels of 10 %, 5 %, and 1 %, respectively (two-tailed test)
A. Knill et al.
2 Experience, Information Asymmetry, and Rational Forecast Bias 89

analytical abilities, which would suggest that perhaps they have to purchase less
information to achieve precision, and valuation relationships, which allow them to
purchase information at reduced “costs” (i.e., deviation from consensus). To ensure
that the marginal effects are significantly different, we perform Wald tests to
confirm that the differences between the coefficients are statistically significant.
As can be seen at the bottom of Panel A, the differences are statistically significant
in every case.
In Panel B, we examine experience-based subsamples to see more intricately
how experience affects the impact of Information asymmetry on Deviation. Results
confirm that Experience affects the link between Information asymmetry and
Deviation. The moderately experienced analyst is the most affected by an increase
in information asymmetry, regardless of the definition of information asymmetry. It
is interesting to note that the results for Analyst coverage and Forecast dispersion
are much stronger than Relative firm size. We use a Hausman test of statistical
significance and find that the experience estimations are all statistically different.
These empirical results nicely complement the theoretical predictions of our
model. Taken collectively, the results suggest that analysts’ rational forecast devi-
ation from the consensus is a result of analysts maximizing their objective function
of minimizing error while taking into consideration the information environment in
which they operate.
To provide evidence that the deviation from the consensus discussed in the rest
of the paper is indeed rationale, we test whether an analyst posting an estimate with
the optimal deviation from consensus, empirically represented by the fitted value of
either Eqs. 2.4 or 2.5, achieves a lower forecast error (defined as estimate minus
actual earnings). Here, a negative association between fitted Deviation from con-
sensus (DFC*) and Forecast error (FE) is indicative of optimal forecasting. Chen
and Jiang (2006) argue that a positive (negative) association between observed
Deviation from consensus (DFC) and Forecast error (FE) is indicative of an analyst
overweighting (underweighting) her own assessment. This is because if analysts
follow a Bayesian updating rule and minimize squared forecast error, there should
be no link between observed deviation from forecast and forecast error. They find
that analysts, on average, overweigh their own belief, and thus the link between
observed deviation from consensus and forecast error is positive. In this paper, we
extend Chen and Jiang’s model to allow for rational biasing of forecasts, similar to
Lim (2001). As such, we find that there should be a concave relationship between
Experience and Deviation from consensus. If our theory is correct and if our
empirical model of deviation from consensus truly captures the essence of rational
forecasting that occurs among analysts, the fitted Deviation from consensus should
be devoid of analysts’ overconfidence about their own information processing. In
fact, if fitted Deviation reflects the tendency toward more rational forecasting, we
should find a negative relationship between fitted Deviation and Forecast error.
Looking to the results, which are found in Table 2.5, Panel A, we see that this is
indeed what we find. In all specifications, the fitted Deviation (DFC) is negatively
related with Forecast error. All specifications lead to a significant decline in the
forecast error of analyst estimation. Specifically, we find that a 1¢ increase in bias
 90

Table 2.5 Optimal deviation from consensus and forecast error

Panel A: All analysts
Dependent variable ¼ forecast error
Experience Relative experience
Experience Relative experience Experience2 Relative experience2
Experience2 Relative experience2 Forecast dispersion Forecast dispersion Forecast dispersion
Reg. FD Reg. FD Reg. FD Reg. FD Reg. FD
NumRevisions NumRevisions NumRevisions NumRevisions NumRevisions
First-stage regressors included Controls Controls Controls Controls Controls
1 2 3 4 5
DFC* 1.081*** 1.079*** 1.029*** 1.030*** 1.029***
[0.009] [0.009] [0.006] [0.006] [0.006]
Constant 2.350*** 2.150** 2.586*** 2.845*** 2.425***
[0.851] [0.851] [0.850] [0.850] [0.850]
Observations 266,708 266,708 253,673 253,673 253,673
# Brokerage houses 130 130 130 130 130
Model R2 0.05 0.05 0.12 0.12 0.12
DFC* 0.456*** 0.444*** 0.427*** 0.433*** 0.427***
[0.016] [0.017] [0.010] [0.010] [0.010]
DFC*2 0.050*** 0.050*** 0.023*** 0.022*** 0.023***
[0.001] [0.001] [0.000] [0.000] [0.000]
A. Knill et al.
 2

Constant 1.765** 1.508* 1.487* 1.707** 1.378*

[0.840] [0.841] [0.791] [0.781] [0.790]
Observations 266,708 266,708 253,673 253,673 253,673
# Brokerage houses 130 130 130 130 130
Model R2 0.06 0.06 0.14 0.14 0.14
Panel B: Experience subsamples
Novice analysts 3.97 3.97 3.27 3.26 3.27
Moderately experienced 5.50 5.49 4.19 4.18 4.18
Highly seasoned 4.71 4.71 3.92 3.94 3.92
Table 2.5 Panel A presents the results of regressing forecast error on the regressors specified in each column. Experience is the natural log of the number of
quarters the analyst has followed the firm for which forecast error is calculated. Relative experience is the analyst experience scaled by the average analyst
experience (same industry). NumRevisions is the total number of revisions the analyst made during the quarter for the quarter-end earnings. Reg. FD is an
indicator variable that takes on a value of 1 if Reg. FD is in effect and 0 otherwise. Controls are vector of firm-specific variables including Same quarter, Broker
reputation, Broker size, Accruals, Intang. assets, and Return std dev. Same quarter is a dummy variable equal to one if the estimate is in the same quarter as the
actual and zero otherwise. Same quarter is orthogonalized (on NumRevisions) to ensure that multicollinearity is not a problem between these two variables.
Brokerage reputation is a ranking of brokerage reputation, where 0 is the worst and 9 is the best. Brokerage size is the natural log of the number of companies
per quarter a brokerage house follows. Brokerage reputation is orthogonalized (on brokerage size) to ensure that multicollinearity is not a problem between
these two variables. Accruals are the accrued revenue/liabilities utilized for earnings smoothing. Intang. assets are the covered firm’s intangible assets value
relative to its total assets. Return standard deviation is the standard deviation of the covered firm’s return. I is a vector of one-digit SIC industry dummies. T is
a vector of time dummies. Panel B presents the fitted values of DFC from each specification for each experience level. A formal fixed effects treatment around
analysts is employed. Standard errors are reported in brackets
* ** ***
Experience, Information Asymmetry, and Rational Forecast Bias

, , indicate significance levels of 10 %, 5 %, and 1 %, respectively (two-tailed test)

91
 92

Table 2.6 Alternate samples

Dependent variable ¼ DFC (deviation from consensus)
Without revisions With only one analyst Random effects I/B/E/S data
1 2 3 4 5 6 7 8 9 10 11 12
*** ***
Experience 0.817 2.822 2.158*** 1.396***
[0.141] [0.669] [0.155] [0.117]
Experience2 0.127*** 0.595*** 0.475*** 0.089**
[0.042] [0.186] [0.039] [0.036]
***
Relative Experience 1.990 0.157 0.749*** 1.938***
[0.267] [1.266] [0.267] [0.191]
Relative Experience2 0.829*** 0.365 0.351*** 0.658***
[0.123] [0.563] [0.120] [0.077]
Forecast dispersion 0.259*** 0.616*** 0.629*** 0.718***
[0.005] [0.157] [0.004] [0.003]
Same quarter 2.479*** 2.564*** 2.860*** 1.716*** 2.455*** 5.852 2.008*** 2.357*** 3.189*** 1.751*** 2.044*** 2.494***
[0.099] [0.097] [0.096] [0.515] [0.492] [3.657] [0.110] [0.107] [0.104] [0.090] [0.085] [0.077]
NumRevisions 1.598*** 1.721*** 0.053 1.186*** 1.226*** 0.773*** 1.076*** 1.141*** 0.700***
[0.089] [0.088] [0.899] [0.016] [0.016] [0.015] [0.016] [0.015] [0.014]
Reg. FD 4.923*** 5.004*** 4.349*** 4.747*** 5.084*** 46.26 2.223*** 2.412*** 2.560*** 3.243*** 3.490*** 3.268***
[0.366] [0.366] [0.368] [0.948] [0.948] [110.413] [0.228] [0.228] [0.222] [0.187] [0.187] [0.170]
Broker reputation 0.172 0.183* 0.275** 0.106 0.052 60.993 0.368*** 0.400*** 0.520***
[0.110] [0.110] [0.110] [0.430] [0.431] [65.397] [0.088] [0.090] [0.086]
Broker size 1.025*** 1.032*** 1.006*** 0.095 0.132 7.038 0.514*** 0.404*** 0.347***
[0.142] [0.141] [0.143] [0.458] [0.457] [21.599] [0.125] [0.127] [0.123]
Accruals 0.001 0.01 0.140*** 3.317*** 3.306*** 92.249 0.171*** 0.160*** 0.413*** 0.464*** 0.498*** 0.244***
[0.048] [0.047] [0.045] [0.593] [0.593] [212.638] [0.037] [0.037] [0.036] [0.030] [0.030] [0.027]
Intangible assets 0.153 0.121 0.870*** 0.38 0.419 3.35 1.216*** 1.268*** 4.191*** 0.881*** 0.861*** 2.196***
[0.263] [0.263] [0.260] [0.980] [0.981] [25.114] [0.232] [0.231] [0.227] [0.209] [0.210] [0.191]
A. Knill et al.
 2

Return std. dev. 10.012*** 9.187*** 2.975*** 10.716*** 9.805*** 0.384 18.238*** 17.662*** 5.043*** 19.153*** 17.969*** 1.710***
[0.666] [0.657] [0.664] [2.079] [2.063] [35.868] [0.579] [0.570] [0.568] [0.473] [0.472] [0.436]
Constant 10.593*** 11.004*** 9.734*** 8.532 10.407* 57.535 3.465*** 4.612*** 2.021* 2.682*** 0.828 2.010***
[1.409] [1.405] [1.392] [5.760] [5.770] [165.519] [1.258] [1.267] [1.219] [0.693] [0.694] [0.619]
Firm FE Yes Yes Yes Yes Yes Yes No No No Yes Yes Yes
Industry FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 66,550 66,550 62,014 13,181 13,181 146 266,708 266,708 253,673 295,717 295,717 295,717
# Brokerage houses{ 130 130 130 119 119 22 130 130 130 7708 7708 7708
Model R2 0.03 0.03 0.07 0.06 0.06 0.24 0.06 0.06 0.16 0.07 0.07 0.23

Analyst information comes from First Call (except in Specifications (10)–(12)). Company information comes from Compustat/CRSP. Experience is the
natural log of the number of quarters the analyst has followed the firm for which forecast error is calculated. Relative experience is the analyst experience
scaled by the average analyst experience (same industry). Same quarter is a dummy variable equal to one if the estimate is in the same quarter as the actual and
zero otherwise. Same quarter is orthogonalized (on NumRevisions) to ensure that multicollinearity is not a problem between these two variables. Forecast
dispersion is the standard deviation of estimates. X is a vector of firm-specific variables including Broker reputation, Broker size, Accruals, Intang. assets, and
Return std dev. Brokerage reputation is a ranking of brokerage reputation, where 0 is the worst and nine is the best. Brokerage size is the natural log of the
number of companies per quarter a brokerage house follows. Brokerage reputation is orthogonalized (on brokerage size) to ensure that multicollinearity is not
a problem between these two variables. Accruals are the accrued revenue/liabilities utilized for earnings smoothing. Intang. assets are the covered firm’s
intangible assets value relative to its total assets. Return standard deviation is the standard deviation of the covered firm’s return. I is a vector of one-digit SIC
industry dummies. T is a vector of time dummies. Specifications (1)–(3) include only original estimates (i.e., no revisions). Specifications (4)–(6) include only
firms with one analyst following the firm. Specifications (7)–(9) use random effects. Specifications (10)–(12) use I/B/E/S data. A formal fixed effects treatment
around analysts is employed unless otherwise noted. Standard errors are reported in brackets.
* **
, , and *** indicate significance levels of 10 %, 5 %, and 1 %, respectively
{
In sample derived from I/B/E/S data, this is actually the number of analysts
Experience, Information Asymmetry, and Rational Forecast Bias
93
 94

Table 2.7 Alternate proxies for information asymmetry

Experience2 Relative experience2 Experience2 Relative experience2

Inf. asymmetry Inf. Asymmetry Inf. asymmetry Inf. asymmetry Inf. asymmetry Inf. asymmetry
Reg. FD Reg. FD Reg. FD Reg. FD Reg. FD Reg. FD
NumRevisions NumRevisions NumRevisions NumRevisions NumRevisions NumRevisions
SameQtr SameQtr SameQtr SameQtr SameQtr SameQtr
Controls Controls Controls Controls Controls Controls
1 2 3 4 5 6
Panel A: Analyst coverage1
DFC 1.079*** 1.081*** 1.079*** 0.446*** 0.456*** 0.444***
[0.009] [0.009] [0.009] [0.017] [0.016] [0.017]
DFC squared 0.050*** 0.050*** 0.050***
[0.001] [0.001] [0.001]
Constant 2.171** 2.347*** 2.150** 1.523* 1.762** 1.508*
[0.851] [0.851] [0.851] [0.842] [0.840] [0.841]
Observations 266,708 266,708 266,708 266,708 266,708 266,708
# Brokerage houses 130 130 130 130 130 130
Panel B: Relative firm size
DFC 1.046*** 1.048*** 1.046*** 0.496*** 0.496*** 0.496***
[0.008] [0.008] [0.008] [0.015] [0.015] [0.015]
DFC squared 0.041*** 0.042*** 0.041***
[0.001] [0.001] [0.001]
Constant 1.505*** 1.513*** 1.504*** 1.100*** 1.122*** 1.100***
[0.066] [0.066] [0.066] [0.066] [0.066] [0.066]
Observations 217,126 217,126 217,126 217,126 217,126 217,126
# Brokerage houses 124 124 124 124 124 124
* **
, , and *** indicate significance levels of 10 %, 5 %, and 1 %, respectively
A. Knill et al.
2 Experience, Information Asymmetry, and Rational Forecast Bias 95

leads to a decrease in Forecast error from a low of 1.029 (specification 3) to 1.081¢

(specification 1), depending upon the specification. Considering that the sample
average Forecast error is 3.79¢, this is a significant decrease, suggesting that
posting such deviations is indeed rational.
Given the high marginal effect and considering the nonlinearity of the impact of
experience on deviation from consensus, we further test a squared term of the fitted
deviation from consensus variable. It would make sense that continually adding
a penny to one’s estimate would cease to increase the precision of the resulting
estimation at some point. In doing so, we see that the linear effect is reduced
considerably. The forecast error is decreased to less than half of a penny on average
and this reduction is decreasing with every marginal penny of bias.
Finally, we examine the fitted values of Deviation from consensus for the
different terciles of analyst experience, shown in Table 2.5, Panel B. For all five
specifications, we see evidence of nonlinearity with these fitted values. Once again,
the moderately experienced analyst is found to have the highest “optimal” devia-
tion. This confirms nicely both the theoretical and previous empirical results and
helps us to come full circle.

2.6 Robustness

To check for consistency in the results, we reexamine specifications (3) and

(4) from Table 2.2 and specification (2) from Table 2.4 on subsamples, altering
our empirical methodology and using a different data source. We analyze two
subsamples: (1) excluding revisions and (2) including only firms covered by one
analyst. We alter our empirical methodology by using random effects as opposed to
the fixed effect treatment used in the base specification of the paper. Lastly, we alter
our data source by using I/B/E/S data. Because the main empirical tests use First
Call data, which only looks at brokerage houses, one potential concern is that
perhaps we might not be able to control for analysts’ fixed effects effectively.
Given our specifications and variables we include in the model, we find that there
are only 130 unique brokerage houses. Since I/B/E/S reports analysts rather than
brokerage houses, we use their data and find results hold when individual analysts
fixed effect (7,708 unique analysts) are included. Due to the nature of IBES data, we
cannot, however, control for broker reputation and size. Results are qualitatively
identical across all robustness tests and are included in Appendix 2.
For brevity, we only use one information asymmetry proxy in Table 2.5. We
choose forecast dispersion as our proxy for information asymmetry since it provides
the best model fit (i.e., Model R2) of the three proxies. To show that our results are
robust to the other two information asymmetry proxies, we rerun Table 2.5 using
these alternate proxies. Our results are qualitatively identical and are shown in
Appendix 3.
In results not reported, we examine three final robustness tests. To address any
concerns about the power of our empirics based on the number of observations, we
examine the subsample of manufacturing firms. While this subsample is less than
96 A. Knill et al.

half of the original, the economic and statistical significance of coefficients

remains, providing support for our empirical results. We include alternate industry
dummies using Fama-French industry classifications (SIC codes are used for
classification in the base specifications). Results once again remain. Finally, to
address any concerns about sample selection problems and the dependence of the
power of our empirics on the number of brokerage houses, we run the estimations
not including Broker reputation (which reduces our sample to 130 houses) and have
315 brokerage houses. Our results remain. These results are available upon request.
The congruence of the results from both the included and unreported robustness
tests provides further credence to the conclusions of this study.

2.7 Conclusions

To understand the exact role experience and information asymmetry play in forming
a rational deviation, we offer an integrated framework in which an optimum forecast
is derived from weighing public and private information, where the resulting
forecast bias (i.e., deviation from the consensus) improves accuracy through the
acquisition of private information. We study how information asymmetry and
analyst experience affect the efficacy of the feedback from bias on private informa-
tion acquisition. Our construct enables a realistic optimization, where rational
forecast bias emerges as an outcome of a trade-off between public and private
information while minimizing forecast squared error. We show both analytically
and empirically that this trade-off and the resulting rational deviation are determined
by experience and information asymmetry. While the optimal bias and information
asymmetry are monotonically positively related, the optimal bias and experience are
concavely linked. Moderately experienced analysts find it optimal to purchase
private information with a positively biased deviation from the consensus. The
superior analytical skills and access to private information (at a lower cost) of highly
seasoned analysts leads them to optimally rely on private information.
We use Reg. FD as an experiment to further illuminate the relationship between
experience and forecast bias and to make a distinction between access to private
information and analytical expertise. We find that, as we would expect, the biases of
experienced analysts, who weigh private information more heavily, are affected
more. The extent to which information asymmetry and analyst experience play
a role in determining the analyst’s bias is directly affected by the dispersion of the
common and idiosyncratic signals of all analysts as well as the extent of their
correlation with each other.
Our results may help to paint a clearer picture of the types of information
produced by analysts. Institutions may be able to use these results to design
continuing education programs and to make hiring/firing decisions. Our results
can also help to explain why and how changes in the information environment
caused by regulatory changes (e.g., Reg. FD), index inclusions, and public disclo-
sures affect the role of analysts as information processors.
2 Experience, Information Asymmetry, and Rational Forecast Bias 97

Appendix 1: Proofs

The Objective Function

Given that the analyst’s forecast is a weighted average of the analyst’s uncondi-
tional estimate and the consensus, F ¼ wE + (1 – w)Ec, the objective function can be
expressed as
h i
min w2 ðt0 þ tðbÞÞ2 þ ð1  wÞ2 t2 1 1
c þ 2rwð1  wÞðt0 þ tðbÞÞ tc (2.6)
wjb

The first-order condition then is

2wðt0 þ tðbÞÞ2  2ð1  wÞt2 1 1

c þ 2rð1  wÞðt0 þ tðbÞÞ tc

 2rwðt0 þ tðbÞÞ1 t1

c  0

By collecting terms, we then have

n o
w ðt0 þ tðbÞÞ2 þ t2
c  2rð t 0 þ t ð b Þ Þ 1 1
t c ¼ t2 1 1
c  rðt0 þ tðbÞÞ tc

This means that the optimal weight is

ðt0 þ tðbÞÞ2  rtc ðt0 þ tðbÞÞ

w¼ (2.7)
ðt0 þ tðbÞÞ2 þ t2c  2rtc ðt0 þ tðbÞÞ

Proof of Proposition 1

By taking the derivative of Eq. 2.7 with respect to t0, we have

∂w 2t2c ðt0 þ tðbÞÞ  2rtc ðt0 þ tðbÞÞ2  rt3c

¼ h i2 (2.8)
∂t0
ðt0 þ tðbÞÞ2 þ t2c  2rtc ðt0 þ tðbÞÞ

Clearly, since the denominator of ∂w/∂t0 is positive, then the sign is only
a function of the numerator. This implies that the sign changes when the numerator,
2tc(t0 + t(b))  2r(t0 + t(b))2  rt2c , is at maximum. To find the maximum,
we solve for t0 that satisfies the first-order conditions of the numerator. The first-
order condition yields tc  2r(t0 + t(b))  0. Thus, at optimal weight
t0 + t(b) ¼ 0.5r1tc.
98 A. Knill et al.

Proof of Proposition 2

By taking the derivative of Eq. 2.7 with respect to bias, we have

h i
∂t
∂w 2t2c ðt0 þ tðbÞÞ  2rtc ðt0 þ tðbÞÞ2  rt3c ∂b
¼ h i2 (2.9)
∂b
ðt0 þ tðbÞÞ2 þ t2c  2rtc ðt0 þ tðbÞÞ

Clearly, since the denominator of ∂w/∂b is positive, then the sign is only
a function of the numerator. This implies (1) that since ∂t/∂b is positive, then the
optimal weight would be monotonically increasing with ∂t/∂b or information
asymmetry, and (2) that the optimal weight is nonlinearly, concavely related to
private information precision. Since the first term in the numerator is a quadratic
function of analyst’s own precision, the maximum in the function is the point at
which the numerator changes sign. This point, however, is exactly the same point at
which ∂w/∂t0 maximizes. For biases at which t0 + t(b) falls below 0.5r1tc., then
so long as bias increases so does the optimal weight.

Appendix 2: Alternate Samples

See Table 2.6.

Appendix 3: Alternate Proxies for Information Asymmetry for

Table 2.5

See Table 2.7.

Appendix 3 presents the results of regressing forecast error on the regressors
specified in each column. Inf. asymmetry is analyst coverage in Panel A and relative
firm size in Panel B. Same quarter is a dummy variable equal to one if the estimate is
in the same quarter as the actual and zero otherwise. Same quarter is orthogonalized
(on NumRevisions) to ensure that multicollinearity is not a problem between these
two variables. Controls is a vector of firm-specific variables including broker
reputation, broker size, accruals, intang. assets, and return std dev. Brokerage
reputation is a ranking of brokerage reputation, where 0 is the worst and 9 is the
best. Brokerage size is the natural log of the number of companies per quarter
a brokerage house follows. Brokerage reputation is orthogonalized (on brokerage
size) to ensure that multicollinearity is not a problem between these two variables.
Accruals are the accrued revenue/liabilities utilized for earnings smoothing. Intang.
assets are the covered firm’s intangible assets value relative to its total assets. Return
standard deviation is the standard deviation of the covered firm’s return. I is a vector
of one-digit SIC industry dummies. T is a vector of time dummies.
2 Experience, Information Asymmetry, and Rational Forecast Bias 99

References
Agrawal, A., & Chen, M. (2012). Analyst conflicts and research quality. Quarterly Journal of
Finance, 2, 179–218.
Baily, W., Li, H., Mao, C., & Zhong, R. (2003). Regulation fair disclosure and earnings informa-
tion: Market, analysts, and corporate response. Journal of Finance, 58, 2487–2514.
Bannister, J., & Newman, H. (1996). Accrual usage to manage earnings toward financial analysts’
forecasts. Review of Quantitative Finance and Accounting, 7, 259–278.
Barber, B., Lehavy, R., Trueman, B. (2000). Are all brokerage houses created equal? Testing
for systematic differences in the performance of brokerage house stock recommendations.
University of California at Davis and University of California at Berkeley (unpublished),
March.
Bernhardt, D., Campello, M., & Kutsoati, E. (2006). Who herds? Journal of Financial Economics,
80, 657–675.
Beyer, A. (2008). Financial analysts’ forecast revisions and managers’ reporting behavior. Journal
of Accounting and Economics, 46, 334–348.
Bhattacharya, N. (2001). Investors’ trade size and trading responses around earnings announce-
ments: An empirical investigation. Accounting Review, 76, 221–244.
Boni, L., & Womack, K. (2006). Analysts, industries, and price momentum. Journal of Financial
Quantitative Analysis, 41, 85–109.
Brown, L., & Sivakumar, K. (2003). Comparing the quality of two operating income measures.
Review of Accounting Studies, 4, 561–572.
Brown, S., Hillegeist, S., & Lo, K. (2004). Conference calls and information asymmetry. Journal
of Accounting and Economics, 37, 343–366.
Carey, M., Post, M., & Sharp, S. (1998). Does corporate lending by banks and finance companies
differ? Evidence on specialization in private debt contracting. Journal of Finance, 53, 845–878.
Carter, R., & Manaster, S. (1990). Initial public offerings and underwriter reputation. Journal of
Finance, 45, 1045–1068.
Carter, R., Dark, F., & Singh, A. (1998). Underwriter reputation, initial returns, and the long-run
performance of IPO stocks. Journal of Finance, 53, 285–311.
Chen, Q., & Jiang, W. (2006). Analysts’ weighting of private and public information. Review of
Financial Studies, 19, 319–355.
Clement, M. (1999). Analyst forecast accuracy: Do ability, resources, and portfolio complexity
matter? Journal of Accounting and Economics, 27, 285–303.
Clement, M., & Tse, S. (2005). Financial analyst characteristics and herding behavior in forecast-
ing. Journal of Finance, 60, 307–341.
Conrad, J., Cornell, B., Landsman, W., & Rountree, B. (2006). How do analyst recommendations
respond to major news? Journal of Financial and Quantitative Analysis, 41, 25–49.
De Jong, P., & Apilado, V. (2008). The changing relationship between earnings expectations and
earnings for value and growth stocks during Reg FD. Journal of Banking and Finance, 33,
435–442.
Dechow, P., Kothari, S., & Watts, R. (1998). The relation between earnings and cash flows.
Journal of Accounting and Economics, 25, 133–168.
Doyle, J., Lundholm, R., & Soliman, M. (2003). The predictive value of expenses excluded from
pro forma earnings. Review of Accounting Studies, 8, 145–174.
Fama, E., & French, K. (1997). Industry costs of equity. Journal of Financial Economics, 43,
153–193.
Givoly, D., & Lakonishok, J. (1979). The information content of financial analysts’ forecasts of
earnings. Journal of Accounting and Economics, 2, 165–185.
Gu, Z., & Xue, J. (2007). Do analysts overreact to extreme good news in earnings? Review of
Quantitative Finance and Accounting, 29, 415–431.
Han, B., Manry, D., & Shaw, W. (2001). Improving the precision of analysts’ earnings forecasts by
adjusting for predictable bias. Review of Quantitative Finance and Accounting, 17, 81–98.
100 A. Knill et al.

Hirschey, M., & Richardson, V. (2004). Are scientific indicators of patent quality useful to
investors? Journal of Empirical Finance, 11, 91–107.
Hong, H., Kubik, J., & Salomon, A. (2000). Security analysts’ career concerns and herding of
earnings forecasts. RAND Journal of Economics, 31, 121–144.
Hsu, D., & Chiao, C. (2010). Relative accuracy of analysts’ earnings forecasts over time:
A Markov chain analysis. Review of Quantitative Finance and Accounting, 37, 477–507.
Irani, A. (2004). The effect of regulation fair disclosure on the relevance of conference calls to
financial analysts. Review of Quantitative Finance and Accounting, 22, 15–28.
Krishnaswami, S., & Subramaniam, V. (1998). Information asymmetry, valuation, and the corpo-
rate spin-off decision. Journal of Financial Economics, 53, 73–112.
Kwon, S. (2002). Financial analysts’ forecast accuracy and dispersion: High-tech versus low-tech
stocks. Review of Quantitative Finance and Accounting, 19, 65–91.
Leone, A., & Wu, J. (2007). What does it take to become a superstar? Evidence from institutional
investor rankings of financial analysts. Simon School of Business Working Paper No. FR 02-12.
Available at SSRN: http://ssrn.com/abstract=313594 or http://dx.doi.org/10.2139/ssrn.313594.
Lim, T. (2001). Rationality and analysts; forecast deviation. Journal of Finance, 56, 369–385.
Lin, B., & Yang, R. (2010). Does regulation fair disclosure affect analysts’ forecast performance?
The case of restructuring firms. Review of Quantitative Finance and Accounting, 38, 495–517.
Loughran, T., & Ritter, J. (2004). Why has IPO underpricing changed over time? Financial
Management, 33(3), 5–37.
Mest, D., & Plummer, E. (2003). Analysts’ rationality and forecast bias: Evidence from analysts’
sales forecasts. Review of Quantitative Finance and Accounting, 21, 103–122.
Mikhail, M., Walther, D., & Willis, R. (1997). Do security analysts improve their performance
with experience? Journal of Accounting Research, 35, 131–157.
Nutt, S., Easterwood, J., & Easterwood, C. (1999). New evidence on serial correlation in analyst
forecast errors. Financial Management, 28, 106–117.
Petersen, M., & Rajan, R. (1994). The benefits of lending relationships: Evidence from small
business data. Journal of Finance, 49, 3–37.
Ramnath, S. (2002). Investor and analyst reactions to earnings announcements of related firms: An
empirical analysis. Journal of Accounting Research, 40, 1351–1376.
Stangeland, D., & Zheng, S. (2007). IPO underpricing, firm quality, and analyst forecasts.
Financial Management, 36, 1–20.
Sufi, A. (2007). Information asymmetry and financing arrangements: Evidence from syndicated
loans. Journal of Finance, 62, 629–668.
Thomas, S. (2002). Firm diversification and asymmetric information: Evidence from analysts’
forecasts and earnings announcements. Journal of Financial Economics, 64, 373–396.
Trueman, B. (1990). Theories of earnings-announcement timing. Journal of Accounting and
Economics, 13, 285–301.
Waymire, G. (1986). Additional evidence on the accuracy of analyst forecasts before and after
voluntary managed earnings forecasts. Accounting Review, 61, 129–141.
Winkler, R. L. (1972). An introduction to Bayesian inference and decision. New York: Holt,
Rinehart and Winston.
Zhang, F. (2006). Information uncertainty and stock returns. Journal of Finance, 61, 105–137.
Zhou, T., & Lai, R. (2009). Herding and information based trading. Journal of Empirical Finance,
16, 388–393.
Zitzewitz, E. (2002). Regulation fair disclosure and the private information of analysts. Available
at SSRN: http://ssrn.com/abstract=305219 or http://dx.doi.org/10.2139/ssrn.305219.
An Appraisal of Modeling Dimensions for
Performance Appraisal of Global 3
Mutual Funds

G.V. Satya Sekhar

Contents
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2 Performance Evaluation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3 A Review on Various Models for Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.3.1 Jensen Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.3.2 Fama Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.3.3 Treynor and Mazuy Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.4 Statman Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.5 Choi Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.6 Elango Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.3.7 Chang, Hung, and Lee Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.3.8 MM Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.3.9 Meijun Qian’s Stage Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

Abstract
A number of studies have been conducted to examine investment performance of
mutual funds of the developed capital markets. Grinblatt and Titman (1989,
1994) found that small mutual funds perform better than large ones and that
performance is negatively correlated to management fees but not to fund size or
expenses. Hendricks, Patel, and Zeckhauser (1993), Goetzmann and Ibbotson
(1994), and Brown and Goetzmann (1995) present evidence of persistence
in mutual fund performance. Grinblatt and Titman (1992) and Elton, Gruber,

G.V. Satya Sekhar

Department of Finance, GITAM Institute of Management, GITAM University, Visakhapatnam,
Andhra Pradesh, India
e-mail: gudimetlavss@yahoo.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 101
DOI 10.1007/978-1-4614-7750-1_3,
# Springer Science+Business Media New York 2015
102 G.V. Satya Sekhar

and Blake (Journal of Financial Economics 42:397–421, 1996) show that past
performance is a good predictor of future performance. Blake, Elton, and
Grubber (1993), Detzler (1999), and Philpot, Hearth, Rimbey, and Schulman
(1998) find that performance is negatively correlated to fund expense, and
that past performance does not predict future performance. However, Philpot,
Hearth, and Rimbey (2000) provide evidence of short-term performance persis-
tence in high-yield bond mutual funds. In their studies of money market mutual
funds, Domian and Reichenstein (1998) find that the expense ratio is the most
important factor in explaining net return differences. Christoffersen (2001) shows
that fee waivers matter to performance. Smith and Tito (1969) conducted a study
into 38 funds for 1958–1967 and obtained similar results. Treyner (1965)
advocated the use of beta coefficient instead of the total risk.

Keywords
Financial modeling • Mutual funds • Performance appraisal • Global investments •
Evaluation of funds • Portfolio management • Systematic risk • Unsystematic
risk • Risk-adjusted performance • Prediction of price movements

3.1 Introduction

Performance of financial instruments is basically dependent on three important

models derived independently by Sharpe, Jensen, and Treynor. All three models are
based on the assumptions that (1) all investors are averse to risk and are single-
period expected utility of terminal wealth maximizers, (2) all investors have
identical decision horizons and homogeneous expectations regarding investment
opportunities, (3) all investors are able to choose among portfolios solely on the
basis of expected returns and variance of returns, (4) all transactions costs and taxes
are zero, and (5) all assets are infinitely divisible.

3.2 Performance Evaluation Methods

The following paragraphs indicate a brief description of the studies on “perfor-

mance evaluation of mutual funds.”
Friend et al. (1962) offered the first empirical analysis of mutual funds perfor-
mance. Sharpe (1964), Treynor and Mazuy (1966), Jensen (1968), Fama (1972),
and Grinblatt and Titman (1989, 1994) are considered to be classical studies in
performance evaluation methods. Sharpe (1964) made a significant contribution in
the methods of evaluating mutual funds. His measure is based on capital asset
prices, market conditions with the help of risk and return probabilities. Sharpe
(1966) developed a theoretical measure better known as reward to variability ratio
that considers both average return and risk simultaneously in its ambit. It tested
efficacy through a sample of 34 open-ended funds considering annual returns and
standard deviation of annual return risk surrogate for the period for 1954–1963.
3 An Appraisal of Modeling Dimensions 103

The average reward to variability ratio of 34 funds was considerably smaller

than Dow Jones portfolio and considered enough to conclude that average
mutual funds performance was distinctly inferior to an investment in Dow Jones
Portfolio.
Treynor (1965) advocated the use of beta coefficient instead of the total risk.
He argues that using only naı̈ve diversification, the unsystematic variability of
returns of the individual assets in a portfolio typically average out of zero. So he
considers measuring a portfolio’s return relative to its systematic risk more
appropriate.
Treynor and Mazuy (1966) devised a test of ability of the investment managers
to anticipate market movements. The study used the investment performance out-
comes of 57 investment managers to find out evidence of market timing abilities
and found no statistical evidence that the investment managers of any of the sample
funds had successfully outguessed the market. The study exhibited that the invest-
ment managers had no ability to outguess the market as a whole but they could
identify under priced securities.
Michael C. Jensen (1967) conducted an empirical study of mutual funds during
the period 1954–1964 for 115 mutual funds. His results indicate that these funds are
not able to predict security prices well enough to outperform a buy-the-market-and-
hold policy. His study ignores the gross management expenses to be free. There was
very little evidence that any individual fund was able to do significantly better than
which investors expected from mere random chance. Jensen (1968) measured the
performance as the return in excess of equilibrium return mandated by capital asset
pricing model. Jensen’s measure is based on the theory of the pricing of capital
assets by Sharpe (1964), Linter (1965), and Treynor.
Smith and Tito (1969) conducted a study into 38 funds for 1958–1967 and
published results relating to performance of mutual funds. However, Mc Donald
(1974) examined 123 mutual funds for 1960–1969 measures to be closely corre-
lated; more importantly, he found that on an average, mutual funds perform about
as well as native “buy and hold” strategy.
Fama (1972) suggested alternative methods for evaluating investment perfor-
mance with somewhat finer breakdowns of performance on the stock selection,
market timing, diversification, and risk bearing. It devised mechanism for segrega-
tion part of an observed investment return due to managers’ ability to pick up the
best securities at a given level of risk from part that is due to the prediction of
general market price movements.
Dunn and Theisen (1983) study is about ranking by the annual performance of
201 institutional portfolios for the period 1973 through 1982 without controlling for
fund risk. They found no evidence that funds performed within the same quartile
over the 10-year period. They also found that ranks of individual managers based on
5-year compound returns revealed no consistency.
Eun et al. (1991) reported similar findings. The benchmarks used in their study
were the Standard and Poor’s 500 Index, the Morgan Stanley Capital International
World Index, and a self-constructed index of US multinational firms. For the
period 1977–1986, the majority of international funds outperformed the US market.
104 G.V. Satya Sekhar

However, they mostly failed to outperform the world index. The sample consisted
of 19 US-based international funds, and the Sharpe measure was used to assess
excess returns.
Barua and Varma (1993b) have examined the relationship between the NAV and
the market price on Mastershares. They conclude that market prices are far more
volatile than what can be justified by volatility of NAVs. The prices also show
a mean reverting behavior, thus perhaps providing an opportunity for discovering
a trading rule to make abnormal profits in the market. Such a rule would basically
imply buying Mastershares whenever the discount from NAV was quite high and
selling Mastershares whenever the discount was low.
Droms and Walker (1994) used a cross-sectional/time-series regression meth-
odology. Four funds were examined over 20 years (1971–1990), and 30 funds were
analyzed for a 6-year period (1985–1990). The funds were compared to the
Standard and Poor’s 500 Index, the Morgan Stanley Europe, Australia, and Far
East Index (EAFE) which proxies non-US stock markets, and the World Index.
Applying the Jensen, Sharpe, and Treynor indices of performance, they found
that international funds have generally underperformed the US market and the
international market. Additionally, their results indicated that portfolio turnover,
expense ratios, asset size, load status, and fund size are unrelated to fund
performance.
Bauman and Miller (1995) studied the persistence of pension and investment
fund performance by type of investment organization and investment style. They
employed a quartile ranking technique, because they noted that “investors pay
particular attention to consultants’ and financial periodicals’ investment perfor-
mance rankings of mutual funds and pension funds.” They found that portfolios
managed by investment advisors showed more consistent performance (measured
by quartile rankings) over market cycles and that funds managed by banks and
insurance companies showed the least consistency. They suggest that this result
may be caused by a higher turnover in the decision-making structure in these less
consistent funds. This study controls for the effects of turnover of key decision
makers by restricting the sample to those funds with the same manager for the entire
period of study.
Volkman and Wohar (1995) extend this analysis to examine factors that impact
performance persistence. Their data consists of 322 funds over the period
1980–1989 and shows performance persistence is negatively related to size and
negatively related to levels of management fees.
Elton et al. (1996) examined the predictability of stock mutual funds perfor-
mance based on risk-adjusted future performance. It also demonstrated application
of modern portfolio techniques on past data to improve selection, which permitted
construction of portfolio funds that significantly outperformed a rule based on the
past rank alone. The portfolio so selected was reported to have small, but statisti-
cally significant, positive risk-adjusted returns during a period when mutual funds
in general had negative risk-adjusted returns.
Jayadeve (1996) paper enlightens performance evaluation based on monthly
returns. His paper focuses on performance of two growth-oriented mutual funds
3 An Appraisal of Modeling Dimensions 105

(Mastergain and Magnum Express) on the basis of monthly returns compared to

benchmark returns. For this purpose, risk-adjusted performance measures
suggested by Jensen and Treynor and Sharpe are employed.
Carhart (1997) shows that expenses and common factors in stock returns such as
beta, market capitalization, 1-year return momentum, and whether the portfolio is
value or growth oriented “almost completely” explain short-term persistence in
risk-adjusted returns. He concludes that his evidence does not “support the exis-
tence of skilled or informed mutual fund portfolio managers.”
Yuxing Yan (1999) examined performance of 67 US mutual funds and the S&P
500 Index with 10-year daily return data from 1982 to 1992. The S&P index was
used as benchmark index. Daily data are transformed into weekly data for compu-
tational reasons. In the calculations, it was assumed that the S&P 500 market index
is a good one, i.e., it is efficient and its variance is constant.
Redmand et al.’s (2000) study examines the risk-adjusted returns using Sharpe’s
Index, Treynor’s Index, and Jensen’s alpha for five portfolios of international
mutual funds during 1985–1994. The benchmarks for competition were the US
market proxied by the Vanguard Index 500 mutual fund and a portfolio of funds that
invest solely in US stocks. The results show that for 1985 through 1994 the portfolio
of international mutual funds outperformed the US market and the portfolio of US
mutual funds.
Rahul Bhargava et al. (2001) evaluated the performance of 114 international
equity managers over the January 1988 to December 1997 period. Performance
tests are conducted using Sharpe and Jensen performance methodologies. Three
major findings are reported. First, international equity managers, on an average,
were unable to outperform the MSCI world market proxy during the sample period.
Second, geographic asset allocation and equity style allocation decisions enhanced
the performance of international managers during the sample period. Third, sepa-
rately managed funds were outperformed mutual funds.
Sadhak’s (2003) study is an attempt to evaluate the performance of Indian
mutual funds with the help of data pertaining to (a) trends in income and expenses,
(b) investment yield and risk-associated returns, and (c) returns of Indian mutual
funds vis-à-vis returns of other emerging markets.
Bala Ramasamy and Yeung’s (2003) survey focused on Malaysia where the
mutual fund industry started in the 1950s but only gained importance in the 1980s
with the establishment of government-initiated program. The sample size
consisting of 56 financial advisors representing various life insurance and mutual
fund companies resulted in 864 different profiles of mutual funds. The cojoint
analysis was employed to generate the questionnaire and analyze its results. The
results of this survey point to three important factors which dominate the choice of
mutual funds. These are consistent past performance, size of funds, and costs of
transaction.
Chang et al. (2003) identified hedging factor in the equilibrium asset pricing
model and used this benchmark to construct a new performance measure. Based
on this measure, they are able to evaluate mutual fund managers hedging timing
ability in addition to more traditional security selectivity and timing. While security
106 G.V. Satya Sekhar

selectivity performance involves forecasts of price movements of selected

individual stock, market timing measures the forecasts of next period realizations
of the market portfolio. The empirical evidence indicates that the selectivity
measure is positive on average and the market timing measure is negative on
average.
Obeid (2004) has suggested a new dimension called “modified approach for risk-
adjusted performance of mutual funds.” This method can be considered as more
powerful, because it allows not only for an identification of active resources but also
for identification of risk. He observed two interesting results: first, it can be shown
that in some cases, a superior security selection effect is largely dependent on
taking higher risks. Second, even in the small sample analyzed in the study,
significant differences appear between each portfolio manager’s styles of selection.
Gupta OP and Amitabh Gupta (2004) published their research on select Indian
mutual funds during a 4-year period from 1999 to 2003 using weekly returns based
on NAVs for 57 funds. They found that fund managers have not outperformed the
relevant benchmark during the study period. The funds earned an average return of
0.041 per week against the average market return of 0.035 %. The average risk-free
rate was 0.15 % per week, indicating that the sample funds have not earned even
equivalent to risk-free return during the study period.
Subash Chander and Japal Singh (2004) considered selected funds during the
period from November 1993 to March 2003 for the purpose of their study. It was
found that the Alliance Mutual Fund and Prudential ICICI Mutual Funds have
posted better performance for the period of study in that order as compared to other
funds. Pioneer ITI, however, has shown average performance and Templeton India
mutual fund has staged a poor show.
Amit Singh Sisodiya (2004) makes comparative analysis of performance of
different mutual funds. He explains that a fund’s performance when viewed on
the basis of returns alone would not give a true picture about the risk the fund would
have taken. Hence, a comparison of risk-adjusted return is the criteria for analysis.
Bertoni et al. (2005) analyzed the passive role that, implicitly, would
place institutional investors in such a context. The study was conducted in
Italy using empirical evidence from the Italian stock exchange (Comit Index).
This study finds that three factors reduce the freedom of institutional investors
to manage their portfolio – the market target size, the fund structure, and the
benchmarking.
Sudhakar and Sasi Kumar (2005) made a case study of Franklin Templeton
mutual fund. The sample consists of a total of ten growth-oriented mutual funds
during the period from April 2004 to March 2005. NIFTY based on NSE Index was
used as the proxy for the market index, and each scheme is evaluated with respect to
the NSE index to find out whether the schemes were able to beat the market or not.
It was found that most of the growth-oriented mutual funds have been able
to deliver better returns than the benchmark indicators. In the sample study, all
the funds have positive differential returns indicating better performance and
diversification of the portfolio, except two funds with negative differential returns,
viz., Franklin India Bluechip Fund and Templeton India Income Fund.
3 An Appraisal of Modeling Dimensions 107

Martin Eling (2006) made a remarkable contribution to the theory of “performance

evaluation measures.” In this study, data envelopment analysis (DEA) is presented as
an alternative method for hedge fund performance measurement. As an optimization
result, DEA determines an efficiency score, which can be interpreted as
a performance measure. An important result of the empirical study is that completely
new rankings of hedge funds compared to classic performance measures.
George Comer (2006) examined the stock market timing ability of two samples
of hybrid mutual funds. The results indicate that the inclusion of bond indices and
a bond timing variable in a multifactor Treynor-Mazuy model framework leads to
substantially different conclusion concerning the stock market timing performance
of these funds relative to the traditional Treynor-Mazuy model find less stock
timing ability over the 1981–1991 time period provide evidence of significant
stock timing ability across the second fund sample during the 1999–2000 period.
Yoon K. Choi (2006) proposed an incentive-compatible portfolio performance
evaluation measure. In this model, a risk-averse portfolio manager is delegated to
manage a fund, and his portfolio construction (and information-gathering) effort is
not directly observable to investors, in which managers are to maximize investors’
gross returns net of managerial compensation. He considers the effect of
organizational elements such as economics of scale on incentive and thus on
performance.
Ramesh Chander (2006) study examined the investment performance of man-
aged portfolios with regard to sustainability of such performance in relation to fund
characteristics, parameter stationarity, and benchmark consistency. The study
under consideration is based on the performance outcome of 80 investment
schemes from public as well as private sectors for the 5-year period encompassing
January 1998 through December 2002. The sample comprised 33.75 % of small,
26.75 % of medium, 21.25 % of large, and 18.75 % of the giant funds.
Ramesh Chander (2006a) study on market timing abilities enables us to under-
stand how well the manager has been able to achieve investment targets and how
well risk has been controlled in the process. The results reported were unable to
generate adequate statistical evidence in support of manager’s successful market
timing. It persisted across measurement criteria, fund characteristics, and the
benchmark indices. However, absence of performance is noted for alternative
sub-periods signifying the negation of survivorship bias.
Beckmann et al. (2007) found that Italian female professionals do not only assess
themselves as more risk averse than their male colleagues, they also prefer a more
passive portfolio management compared to the level they are allowed to. Besides,
in a competitive tournament scenario near the end of the investment period, female
asset managers do not try to become the ultimate top performer when they have
outperformed the peer group. However in case of underperformance, the risk of
deviating from the benchmark makes female professionals more willing than their
male colleagues to seize a chance of catching up.
Gajendra Sidana (2007) made an attempt to classify hundreds of mutual funds
employing cluster analysis and using a host of criteria like the 1-year-old return,
2-year annualized return, 3-year annualized return, 5-year annualized return, alpha,
108 G.V. Satya Sekhar

and beta. The data is obtained from value research. The author finds inconsistencies
between investment style/objective classification and the return obtained by
the fund.
Coates and Hubbard (2007) reviewed the structure, performance, and dynamics
of the mutual fund industry and showed that they are consistent with competition. It
was also found that concentration and barriers to entry are low, actual entry is
common and continuous, pricing exhibits no dominant long-term trend, and market
shares fluctuate significantly. Their study also focused on “effects of competition on
fee” and “pricing anomalies.” They suggested legal interventions are necessary in
setting fee in mutual funds of United States.
Subha and Bharati’s (2007) study is carried out for open-ended mutual fund
schemes and 51 schemes are selected by convenient sampling method. NAVs are
taken for a period of 1 year from 1 October 2004 to 30 September 2005. Out of the
51 funds, as many as 18 schemes earned higher returns than the market return. The
remaining 33 funds however generated lower returns than the market.
Sondhi’s (2007) study analyzes the financial performance of 36 diversified
equity mutual funds in India, in terms of rates of return, comparison with risk-
free return, benchmark comparison, and risk-adjusted returns of diversified equity
funds. Fund size, ownership pattern of AMC, and type of fund are the main factors
considered in this study. The study reveals that private sector is dominating public
sector.
Cheng-Ru Wu et al.’s (2008) study adopts modified Delphi method and the
analytical hierarchy process to design an assessment method for evaluating mutual
fund performance. The most important criteria for mutual fund performance should
be “mutual fund style” followed by “market investment environment.” This result
indicates investor’s focus when they evaluate the mutual fund performance.
Eleni Thanou’s (2008) study examines the risk-adjusted overall performance of
17 Greek Equity Mutual Funds between the years 1997 and 2005. The study
evaluated performance of each fund based on the CAPM performance methodol-
ogy, calculating the Treynor and Sharpe Indexes for the 9-year period as well as for
three sub-periods displaying different market characteristics. The results indicated
that the majority of the funds under examination followed closely the market,
achieved overall satisfactory diversification, and some consistently outperformed
the market, while the results in market timing are mixed, with most funds
displaying negative market timing capabilities.
Kajshmi et al. (2008) studied a sample of schemes in the 8-year period. This study
considers performance evaluation and is restricted to the schemes launched in the
year 1993 when the industry was thrown open to private sector under the regulated
environment by passing the SEBI (Mutual Funds) Regulations 1993. The perfor-
mance of the sample schemes were in line with that of the market as evident from the
positive beta values. All the sample schemes were not well diversified as depicted by
the differences in the Jensen alpha and Sharpe’s differential return.
Massimo Masa and Lei Zhang (2008) found the importance of organizational
structure on Asset Management Company of mutual fund. Their study found that
more hierarchical structures invest less in firms located close to them and deliver
3 An Appraisal of Modeling Dimensions 109

lower performance. An additional layer in hierarchical structure reduces the

average performance by 24 basis points per month. At the same time, more
hierarchical structures leads to herd more and to hold less concentrated portfolios.
Manuel Ammann and Michael Verhofen (2008) examined the impact of prior
performance on the risk-taking behavior of mutual fund managers. Their sample
taken from US funds started in January 2001 and ended in December 2005. The
study found that prior performance in the first half of the year has, in general,
a positive impact on the choice of the risk level in the second half of the year.
Successful fund managers increase the volatility and the beta and assign a higher
proportion of their portfolio to value stocks, small firms, and momentum stocks in
comparison to unsuccessful fund managers.
Onur et al. (2008) study evaluates the performance of 50 large US-based
international equity funds using risk-adjusted returns during 1994–2003. This
study provides documentation on the risk-adjusted performance of international
mutual funds. The evaluation is based on objective performance measures grounded
in modern portfolio theory. Using the methodology developed by Modigliani and
Miller in 1997, the study reports the returns that would have accrued to these mutual
funds for a 5-year holding period as well as a 10-year holding period. It is evident
from the empirical results of this study that the funds with the highest average
returns may lose their attractiveness to investors once the degree of risk embedded
in the fund has been factored into the analysis.
Qiang Bu and Nelson Lacey (2008) examined the determinants of US mutual
fund terminations and provided estimates of mutual fund hazard functions. Their
study found that mutual fund termination correlates with a variety of fund-specific
variables as well as with market variables such as the S&P 500 Index and the short-
term interest rate. This was tested with the underlying assumptions of the semi-
parametric Cox model and reject proportionality. They also found that different
fund categories exhibit distinct hazard functions depending on the fund’s invest-
ment objectives.
David M. Smith (2009) discussed the size and market concentration of the
mutual fund industry, the market entry and exit of mutual funds, the benefits and
costs of mutual fund size changes, the principal benefits and costs of ownership
from fund shareholders’ perspective, etc. This study is based on data from
Morningstar (2009) about US mutual fund industry, which was composed of
607 fund families.
Baker et al. (2010) investigated the relation between the performance and
characteristics of 118 domestic actively managed institutional equity mutual
funds. The results showed that the large funds tend to perform better, which
suggests the presence of significant economies of scale. The evidence indicates a
positive relation between cash holding and performance. They also found evidence
in a univariate analysis that expense ratio class is an important determinant of
performance, and the results are significant in a multivariate setting using Miller’s
active alpha as a performance metric.
Khurshid et al. (2009) studied the structure of the mutual fund industry in India
and analyzed the state of competition among all the mutual funds in private sector
110 G.V. Satya Sekhar

and public sector. The levels of competition and their trends have been obtained for
the periods March 2003–March 2009. This study found overall mutual fund indus-
try is facing a high competitive environment. An increasing trend of competition
was observed within bank institution, private sector foreign, and private sector joint
venture mutual funds.
Mohit Gupta and Aggarwal’s (2009) study focused on the portfolio creation and
industry concentration of 18 ELSS schemes during April 2006 to April 2007.
Mutual fund industry concentration was the variable used in classification or cluster
creation. This exercise was repeated each month for the period under study. Finally
portfolio performance was compared with index fund, portfolio of three randomly
picked funds of the previous month, and the return and risk parameters of ELSS
category as a whole.
Talat Afza and Ali Rauf’s (2009) study aims to provide guidelines to the
managers of open-ended Pakistani mutual funds and benefit small investors by
pointing out the significant variables influencing the fund performance. An effort
has been made to measure the fund performance by using Sharpe ratio with the help
of pooled time-series and cross-sectional data and focusing on different fund
attributes such as fund size, expenses, age, turnover, loads, and liquidity. The
quarterly sample data are collected for all the open-ended mutual funds listed on
Mutual Fund Association of Pakistan (MUFAP), for the years 1999–2006. The
results indicate that among various funds attributes are: lagged return, liquidity and
had significant impact on fund performance.
Amar Ranu and Depali Ranu (2010) critically examined the performance of
equity funds and found out the top 10 best performing funds among 256 equity
mutual fund schemes in this category. They considered three factors for selection:
(a) mutual funds having 5 years of historical performance, (b) fund schemes having
a minimum of Rs.400 crore of assets under management, and (c) funds which have
average return more than 22.47. They found that HDFC TOP 200 (Growth) option
was outperforming among the top 10 best performing equity funds.
Sunil Wahal and Albert Wang (2010) found impact of the entry of new mutual
funds on incumbents using the overlap in their portfolio holdings as a measure of
competitive intensity. Their study revealed that funds with high overlap also
experience quantity competition through lower investor flows, have lower alphas,
and higher attrition rates. These effects only appeared after the late 1990s, at which
point there appears to be endogenous structural shift in the competitive environ-
ment. Their concluding remark is that “the mutual fund market has evolved into one
that displays the hallmark features of a competitive market.”
Sukhwinder Kaur Dhanda et al.’s (2012) study considered the BSE-30 as
a benchmark to study the performance of mutual funds in India. The study period
has been taken from 1 April 2009 to 31 March 2011. The findings of the study
reveal that only three schemes have performed better than benchmark. In the year
2009, HDFC Capital Builder has the top performer. It was 69.18 returns and 26.37
SD and 0.78 beta. HDFC Capital Builder scheme has given the reward for
variability and volatility. HDFC Top 200 Fund and Birla Sun Life Advantage
Funds are on second and third position in terms of return. HDFC Top 200 Fund
3 An Appraisal of Modeling Dimensions 111

has shown better performance than Birla Sun Life Advantage Fund in terms of SD,
beta, Sharpe ratio, and Treynor ratio. Birla Sun Life Advantage Fund has more risk
than the benchmark. Kotak Select Focus Fund has the poorer performer in terms of
risk and return. Except two schemes all other schemes have performed better than
benchmark. Except Kotak Select Focus Fund all other schemes are able to give
reward for variability and volatility.

3.3 A Review on Various Models for Performance Evaluation

3.3.1 Jensen Model

Given the additional assumption that the capital market is in equilibrium, all three
models yield the following expression for the expected one-period return on any
security (or portfolio) j:


E R j ¼ R F þ bJ ½ E ð R m Þ  R F  (3.1)

RF ¼ the one-period risk-free interest rate.

bJ ¼ Cov(j RJ, RM)/s2 RM ¼ the measure of risk (hereafter called systematic risk)
which the asset pricing model implies is crucial in determining the prices of risky
assets.
E(RM) ¼ the expected one-period return on the “market portfolio” which
consists of an investment in each asset in the market in proportion to its fraction
of the total value of all assets in the market. It implies that the expected
return on any asset is equal to the risk-free rate plus a risk premium given by
the product of the systematic risk of the asset and the risk premium on the market
portfolio.

3.3.2 Fama Model

In Fama’s decomposition performance evaluation measure of portfolio, overall

performance can be attributed to selectivity and risk. The performance due to
selectivity is decomposed into net selectivity and diversification. The difference
between actual return and risk-free return indicates overall performance:

Rp  Rf (3.2)

wherein
Rp is actually return on the portfolio, which is monthly average return of fund
and
Rf is monthly average return on treasury bills 91 days.
The overall performance further can be bifurcated into performance due to
selectivity and risk.
112 G.V. Satya Sekhar

Thus,


Rp  Rf ¼ Rp  Rp bp þ Rp bp  Rf (3.3)

In other words, overall performance ¼ selectivity + risk

3.3.3 Treynor and Mazuy Model

Treynor and Mazuy developed a prudent and exclusive model to measure invest-
ment managers’ market timing abilities. This formulation is obtained by adding
squared extra return in the excess return version of the capital asset pricing model as
given below:



2þ
Rpt  Rft ¼ a þ bp Rmt  Rft þ yp Rmt  Rft ept (3.4)

where Rpt is monthly return on the fund, Rft is monthly return on 91 days treasury
bills, Rmt is monthly return on market index, and Ept is error term.
This model involves running a regression with excess investment return as
dependent variable and the excess market return and squared excess market return
as independent variables. The value of coefficient of squared excess return acts as
a measure of market timing abilities that has been tested for significance of using
t-test. Significant and positive values provide evidence in support of the investment
manager’s successful market timing abilities.

3.3.4 Statman Model

Statman measured mutual funds using the following equation (Statman 2000):
eSDAR (excess standard deviation and adjusted return)



¼ Rf þ R p  Rf Sm =Sp  Rm (3.5)

In this formulae, Rf ¼ monthly return on 3-month treasury bills, Rp ¼ monthly

return on fund portfolio, Rm ¼ monthly return on the benchmark index,
Sp ¼ standard deviation of portfolio p’s return, and Sm ¼ standard deviation of
return on the benchmark index.
This model is used for short-term investment analysis. The performance is
compared with it benchmark on monthly basis.

3.3.5 Choi Model

Choi provides a theoretical foundation for an alternative portfolio performance

measure that is incentive-compatible. In this model, a risk-averse portfolio manager
3 An Appraisal of Modeling Dimensions 113

is delegated to manage a fund, and his portfolio construction (and information-

gathering) effort is not directly observable to investors. The fund manager is paid on
the basis of the portfolio return that is a function of effort, managerial skill, and
organizational factors. In this model, the effect of institutional factors is described
by the incentive contractual form and disutility (or cost) function of managerial
efforts in fund operations. It focuses on the cost function as an organizational
factor (simply, scale factor). It was assumed that the disutility function of each
fund is determined by the unique nature of its operation (e.g., fund size) and is an
increasing function of managerial effort at an increasing rate.

3.3.6 Elango Model

Elango’s model also compares the performance of public sector funds vs private
sector mutual funds in India. In order to examine the trend in performance of NAV
during the study period, growth rate in NAV was computed. The growth rate was
computed based on the following formula (Elango 2003):
Growth rate: Rg ¼ ðY t  Y 0 =Y 0 Þ  100 (3.6)

Rg: growth rate registered during the current year

Yt: yield in current year
Y0: yield in previous year
In order to examine whether past is any indicator of future growth in the NAV,
six regression analyses were carried out. NAV of base year was considered as the
dependent variable and current year as in the independent variable.

Equation: Y ¼ A þ b X (3.7)

Dependent variable: Y ¼ NAV of 1999–2000

Independent variable: X ¼ NAV of 2000–2001
In the same way, the second regression equation computed using NAVs of
2000–2001 and 2001–2002, as dependent and independent variables.

3.3.7 Chang, Hung, and Lee Model

The pricing model adopted by Jow-Ran Chang, Nao-Wei Hung, and Cheng-Few
Lee is based on competitive equilibrium version of intemporal asset pricing model
derived in Campbell. The dynamic asset pricing model incorporates hedging risk as
well as market. This model uses a log-linear approximation to the budget constraint
to substitute out consumption from a standard intertemporal asset pricing model.
Therefore, asset risk premia are determined by the covariances of asset returns with
the market return and with news about the discounted value of all future market
returns. Formally, the pricing restrictions on asset i imported by the conditional
version of the model are
114 G.V. Satya Sekhar

Et r i , tþ1  r f , tþ1 ¼ V iI =2 þ gV im þ ðg  1ÞV ih (3.8)

where
Etri, t + 1, log return on asset; rf, t + 1, log return on riskless asset; Vii denotes Vart (ri,t + 1);
g is the agent’s coefficient of relative risk aversion; Vim denotes Covt (ri, t + 1, rm,t + 1)
and Vih ¼ Covt (ri,t + 1, (Et + 1  Et), _1j ¼ 1 rj rm,t + 1 + j); the parameter,
r ¼ 1  exp(c  w); and c  w is the mean log consumption to wealth ratio.
This states that the expected excess log return in an asset, adjusted for a Jensen’s
inequality effect, is a weighted average of two covariances: the covariance with the
return from the market portfolio and the covariance with news about future returns
on invested wealth. The intuition in this equation that assets are priced using their
covariances with the return on invested wealth and future returns on invested wealth.

3.3.8 MM Approach

Leah Modigliani and Franco Modigliani are better known as M2 in the investment
literature. This measure is developed adjusting portfolio return. This adjustment is
carried on the uncommitted (cash balances) part of the investment portfolio at the
riskless return so as to enable all portfolio holdings to participate in the return
generation process. This adjustment is needed to bring out the level playing field for
portfolio risk-return and vis-à-vis market return. The effect of this adjustment is
reported below (Modigliani and Modigliani 1997):

M2 ¼ Rp  Rm (3.9)

Rp ¼ ðRf  ð1  Sdm=SdpÞÞ þ ðRp Sdm=SdpÞ (3.10)

In this formulae * Rp ¼ expected return, Rf ¼ risk-free return, Sdm ¼ standard

deviation of market portfolio, and Sdp ¼ standard deviation of managed portfolio.
In case the managed portfolio has twice the standard deviation of the market,
then, the portfolio would be half invested in the managed portfolio and the
remaining half would be invested at the riskless rate. Likewise, in case the managed
portfolio has lower standard deviation than the market portfolio, it would be levered
by borrowing money and investing the money in managed portfolio. Positive
M2 value indicates superior portfolio performance, while negative indicates
actively managed portfolio manager’s inability to beat the benchmark portfolio
performance.

3.3.9 Meijun Qian’s Stage Pricing Model

Meijun Qian’s (2009) study reveals about the staleness, which is measured prices
imparts a positive statistical bias and a negative dilution effect on mutual fund
performance. First, evaluating performance with non-synchronous data generates
3 An Appraisal of Modeling Dimensions 115

Table 3.1 Overview of different measures

Measures Description Interpretation
Sharpe ratio Sharpe ratio ¼ fund return in excess The higher the Sharpe ratio, the better
of risk-free return/standard deviation the fund returns relative to the amount
of fund. Sharpe ratios are ideal for of risk taken
comparing funds that have a mixed
asset classes
Treynor ratio Treynor ratio ¼ fund return in The higher the Treynor ratio shows
excess of risk-free return/beta of higher returns and lesser market risk
fund. Treynor ratio indicates relative of the fund
measure of market risk
Jensen measure This shows relative ratio between Jensen measure is based on systematic
alpha and beta risk. It is also suitable for evaluating
a portfolio’s performance in
combination with other portfolios
M2 measure It matches the risk of the market A high value indicates that the
portfolio and then calculate portfolio has outperformed and vice
appropriate return for that portfolio versa
Jensen model E(Rj) ¼ RF + bJ[E(Rm)  RF] The expected one-period return on the
“market portfolio” which consists of an
investment in each asset in the market
in proportion to its fraction of the total
value of all assets in the market
Fama model Rp–Rf ¼ [Rp  Rp(bp + Rp(bp–Rf)] Overall performance ¼ selectivity + risk
Treynor and (Rpt–Rft) ¼ a + bp (Rmt  Rft) This model involves running
Mazuy model + yp (Rmt  Rft)2+ept a regression with excess investment
return as dependent variable and the
excess market return and squared
excess market return as independent
variables
Statman model eSDAR ¼ Rf + (Rp  Rf)(Sm/Sp)–Rm This model used for short-term
investment analysis. The performance
is compared with it benchmark on
monthly basis
Elango model Rg ¼ (Yt  Y0/Y0)  100 In order to examine whether past is any
indicator of future growth in the NAV,
six regression analyses were carried
out. NAV of base year was considered
as the dependent variable and current
year as in the independent variable

a spurious component of alpha. Second, stale prices create arbitrage opportunities

for high-frequency traders whose trades dilute the portfolio returns and hence fund
performance. This paper introduces a model that evaluates fund performance while
controlling directly for these biases. Empirical tests of the model show that alpha
net of these biases is on average positive although not significant and about 40 basis
points higher than alpha measured without controlling for the impacts of stale
pricing. The difference between the net alpha and the measured alpha consists of
three components: a statistical bias, the dilution effect of long-term fund flows, and
116 G.V. Satya Sekhar

the dilution effect of arbitrage flows. Thus, assuming that information generated
in time t is not fully incorporated into prices until one period later, the observed
fund return becomes a weighted average of true returns in the current and last
periods:
r t ¼ a þ br mt þ et , (3.11)

r t  ¼  r t1 þ ð1  Þr t , (3.12)

where rt denotes the true excess return of the portfolio with mean m and variance s2
and rmt denotes the excess market return with mean mm and variance sm. Both rt and
rmt are i.i.d, and the error term et is independent of rmt. Rt* is the observed excess
return of the portfolio with zero flows, while  is the weight on the lagged true
return. That is, the higher the , the staler the prices. Assumedly, arbitrage traders
can earn the return rt*, by trading at the fund’s reported net assets values
(Table 3.1).

3.4 Conclusion

This paper is intended to examine various performance models derived by financial

experts across the globe. A number of studies have been conducted to examine
investment performance of mutual funds of the developed capital markets.
The measure of performance of financial instruments is basically dependent on
three important models derived independently by Sharpe, Jensen, and Treynor. All
three models are based on the assumption that (1) all investors are averse to risk and
are single-period expected utility of terminal wealth maximizers, (2) all investors
have identical decision horizons and homogeneous expectations regarding
investment opportunities, (3) all investors are able to choose among portfolios
solely on the basis of expected returns and variance of returns, (4) all transactions
costs and taxes are zero, and (5) all assets are infinitely divisible.

References
Afza, T., & Rauf, A. (2009). Performance Evaluation of Pakistani mutual funds. Pakistan
Economic and Social Review, 47(2), 199–214.
Ammann, M., & Verhofen, M. (2008). The impact of prior performance on the risk-taking of
mutual fund manager. Annals of Finance, (5), 69–90.
Arugaslan, O., Edwards, E., & Samant, A. (2008). Risk-adjusted performance of international
mutual funds. Managerial Finance, 34(1), 5–22.
Baker, M., Litov, L., Wachter, J., & Wurgler, J. (2010). Can mutual fund managers pick stocks?
Evidence from their trades prior to earning announcements. Journal of Financial and Quan-
titative Analysis, 45, 1111–1131.
Baker Kent, H., Haslem, J. A., & Smith, D. M. Performance and Characteristics of Actively
Managed Institutional Equity Mutual Funds [Electronic Copy]. http://ssrn.com/
abstract=1124577
3 An Appraisal of Modeling Dimensions 117

Barua, S. K., & Varma, J. R. (1993). Speculative dynamics: The case of master shares (Advances
in financial planning and forecasting, Vol. 5). Greenwitch: Jai Press.
Bauman, W. S., & Miller, R. E. (1995). Portfolio performance rankings in stock market cycles.
Financial Analysts Journal, 51, 79–87.
Beckmann, D., Lutje, T., & Rebeggiani, L. (2007). Italian asset managers’ behavior: Evidence on
overconfidence, risk taking, and gender (Discussion Paper No. 358). Hannover: Leibniz
Universitat Hannover, Department of Economics, ISSN: 0949-9962.
Bertoni, A., Brtmetti, G., & Cesari, C. (2005). Mutual fund bench marking and market bubbles-a
behavioural approach. Transition Studies Review, 12(1), 36–43.
Bhargava, R., Gallo, J., & Swason, P. T. (2001). The performance, asset allocation and investment
style of international equity manager. Review of Quantitative Finance and Planning, 17,
377–395.
Born, E. K. (1983). International banking in the 19th and 20th centuries. New York: St. Martin’s
Press.
Brown, S. J., & Goetzmann, W. N. (1995). Performance persistence. The Journal of Finance,
50(2), 679–698.
Bu, Q., & Lacey, N. (2008). On understanding mutual fund terminations. Journal of Economics
and Finance, 33, 80–99.
Carhart, M. M. (1997). Persistence in mutual fund performance. Journal of Finance, 52, 57–82.
Chander, R. (2006). Informational efficiency, parameter stationarity and bench mark consis-
tency of investment performance. The ICFAI Journal of Applied Finance. March 2006,
p. 29–45.
Chander, R. (2006b). Investment manager’s market timing abilities: Empirical evidences. The
ICFAI Journal of Applied Finance, 12(8), 15–31.
Chander, S., & Singh, J. (2004). Performance of mutual funds in India: An empirical evidence. The
ICFAI Journal of Applied Finance, 10, 45–63.
Cheng-Ru Wu, Hsin-Yuan Chang, Li-Syuan Wu (2008). A framework of assessable mutual fund
performance. Journal of Modeling in Management, 3(2), 125–139.
Choi, Y. K. (2006). Relative portfolio performance evaluation and incentive structure. Journal of
Business, 79(2), 903–921.
Christofeersen, S. E. (2001). Why do money fund managers voluntarily waive their fee? Journal of
Finance, 74, 117–140.
Coates, J. C., & Hubbard, G. R. (2007). Competition in the mutual fund industry: Evidence and
implications for policy (Discussion Paper No.592). Source: http://ssrn.com/abstract¼1005426
Comer, G. (2006). Hybrid mutual funds and market timing performance. Journal of Business,
79(2), 771–797.
Detzler, L. M. (1999). The performance of global bond mutual funds. Journal of Banking and
Finance, 23, 1195–1217.
Dhanda, S. K., Batra, G. S., & Anjum, B. (2012). Performance evaluation of selected open ended
mutual funds in India. International Journal of Marketing, Financial Services & Management
Research, 1(1), January 2012, ISSN 2277 3622. Online Available at http://indianresearch-
journals.com
Domain, D. L., & Rechenstein, W. (1998). Performance and persistence in money market mutual
fund returns. Financial Services Review, 6, 169–183.
Droms, W. G., & Walker, D. A. (1994). Investment performance of international mutual funds.
Journal of Financial Research, 17, 1–14.
Dunn, P. C., & Theisen, R. D. (1983). How consistently do active managers win? Journal of
Portfolio Management, 9, 47–51.
Elango, R. (2003). Which fund yields more returns? A comparative analysis on the NAV
performance of select public v/s private/foreign open-ended mutual fund schemes in India.
Mutual Funds. The Management Accountant, Vol. 39(4), (2004), p283–290.
Eling, M. (2006). Performance measurement of hedge funds using data envelopment analysis.
Financial Markets and Portfolio Management, 2o, 442–471.
118 G.V. Satya Sekhar

Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). Market timing ability and volatility implied in
investment newsletters’ asset allocation recommendations. Journal of Financial Economics,
42, 397–421.
Eun, C. S., Kolodny, R., & Resnick, B. G. (1991). U.S. based international mutual funds:
A performance evaluation. The Journal of Portfolio Management, 17, 88–94.
Fama, E. F. (1972). Components of investment performance. Journal of Finance, 27, 551–567.
Friend, I., Brown, F. E., Herman, E. S., & Vickers, D. (1962). A study of mutual funds. Washing-
ton, D.C.: U.S. Government Printing Office.
Goetzman, W. N., & Ibbotson, R. G. (1994). Do winners repeat; predicting mutual fund perfor-
mance. The Journal of Portfolio Management, 20, 9–18.
Grinblatt, M., & Titman, S. (1989). Mutual fund performance – An analysis of quarterly holdings.
Journal of Business, 62, 393–416.
Grinblatt, M., & Titman, S. (1992). The persistence of mutual fund performance. Journal of
Finance, 47(5), 1977–1984.
Grinblatt, M., & Titman, S. (1994). Momentum investment strategies – Portfolio performance and
herding: A study of mutual fund behavior. American Economic Review, 85, 1088–1105.
Gupta, M., & Aggarwal, N. (2009). Mutual fund portfolio creation using industry concentration.
The ICFAI Journal of Management Research, Viii(3), 7–20.
Gupta, O. P., & Gupta, A. (2004). Performance evaluation of select Indian mutual fund schemes:
An empirical study. The ICFAI Journal of Applied Finance, 10, 81–97.
Hendricks, D., Patel, J., & Zeckhauser, R. (1993). Hot hands in mutual funds: A short run
persistence of relative performance. Journal of Finance, 48, 93–130.
Jayadeve, M. (1996). Mutual fund performance: An analysis of monthly returns. Finance India,
X(1), 73–84.
Jensen, M. C. (1967). The performance of mutual funds in the period 1945–1964. Journal of
Finance, 23, 389–416.
Jensen, M. C. (1968). The performance of mutual funds in the period 1945–1964. Journal of
Finance, 23, 389–416.
Jow-Ran Chang, Mao-Wei Hung, Cheng-few Lee (2003). An intemporal CAPM approach to evaluate
mutual fund performance. Review of Quantitative Finance and Accounting, 20, 425–433.
Kajshmi, N., Malabika, D., & Murugesan, B. (2008). Performance of the Indian mutual funds;
A study with special reference to growth schemes. Asia-Pacific Business Review, Iv(3), 75–81.
Khurshid, S. M. Z., Rohit, & Sing, G. P. (2009). Level and trends of competition among the mutual
funds in India. Research Journal of Business Management, 3(2), 47–67.
Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets. The Review of Economics and Statistics, 47(1), 13–37.
Massa, M., & Lei Zhang. (2008). The effects of organizational structure on asset management.
http://ssrn.com/abstract¼1328189
Meijun Qian (2009). Stale prices and the performance evaluation of mutual funds. Journal of
Financial and Quantitative Analysis.
Mc Donald, J. G. (1974). Objectives and performance of mutual funds 1960–1967. Journal of
Financial and Quantitative Analysis, 9, 311–333.
Modigliani, F., & Modigliani, L. (1997). Risk adjusted performance. Journal of Portfolio Man-
agement, 23, 45–54.
Obeid, A. T. (2004). A modified approach for risk-adjusted performance attribution. Financial
Markets and Portfolio Management, 18(3), 285–305.
Philpot, J., Hearth, D., Rimbey, J., & Schulman, C. (1998). Active management, fund size and
bond mutual fund returns. The Financial Review, 33, 115–126.
Ramasamy, B., & Yeung, C. H. M. (2003). Evaluating mutual funds in an emerging market: Factors
that matter to financial advisors. International Journal of Bank Marketing, 21, 122–136.
Ranu, A., & Ranu, D. (2010, January). The best performing equity funds. The Analyst.
Redmand, A. L., Gullett, N. S., & Manakyan, H. (2000). The performance of global and
international mutual funds. Journal of Financial and Strategic Decisions, 13(1), 75–85.
3 An Appraisal of Modeling Dimensions 119

Sadhak, H. (2003). Mutual funds in India: Marketing strategies and investment practices.
New Delhi: Response Books.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance, 19, 225–242.
Sharpe, W. F. (1966). Mutual fund performance. Journal of Business, 39, 119–138.
Sidana, G. (2007). Classifying mutual funds in India: Some results from clustering. Indian Journal
of Economics and Business, II(2).
Sisodiya, A. S. (2004, July). Mutual fund industry in India – Coming of age. Chartered Financial
Analyst, pp. 17–22.
Smith, D. M. (2009). The economics of mutual funds. In J. A. Haslem (Ed.), A companion to
mutual funds (Chap. 3). New York: Wiley.
Smith, K. & Tito, D. (1969). Risk-return measures of post-portfolio performance. Journal of
Financial and Quantitative Analysis, 4, 449–471.
Sondhi, H. J. (2007). Financial performance of equity mutual funds in India. New Delhi: Deep &
Deep.
Statman, M. (2000). Socially responsible mutual funds. Financial Analysts Journal, 56, 30–38.
Subha, M. V., & Bharati, J. S. (2007). An empirical study on the performance of select mutual fund
schemes in India. Journal of Contemporary Research in Management, I(1).
Sudhakar, A., & Sasi, K. K. (2005, November). Performance evaluation of mutual funds: A case
study. Southern Economist, pp. 19–23.
Thanou, E. (2008). Mutual fund evaluation during up and down market conditions: The case of
Greek equity mutual funds. International Research Journal of Finance and Economics, (13).
Treynor, J. L. (1965). How to rate management of investment funds. Harvard Business Review, 43,
63–75.
Treynor, J. L., & Mazuy, K. K. (1066). Can mutual funds outguess the markets. Harvard Business
Review, 44, 131–136.
Treynor, J. L., & Mazuy, K. K. (1966). Can mutual funds outguess the markets. Harvard Business
Review, 44, 131–136.
Volkman, D. A., & Wohar, M. E. (1995). Determinants of persistence in relative performance of
mutual funds. Journal of Financial Research, 18, 415–430.
Wahal, S., & Alber (Yan) Wang (2010). Competition among mutual funds. Journal of Financial
Economics. Source: http://ssrn.com/abstract¼1130822
Yuxing Yan (1999). Measuring the timing ability of mutual fund managers. Annals of Operations
Research, 87, 233–243.
Simulation as a Research Tool for
Market Architects 4
Robert A. Schwartz and Bruce W. Weber

Contents
4.1 Studying Market Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.2 The Interaction Between Theoretical Modeling, Empirical Analysis,
and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.2.1 An Application: Call Auction Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.2.2 An Application: The Search for an Equilibrium Price and Quantity . . . . . . . . . . . 125
4.2.3 The Realism of Computer-Generated Data Versus Canned Data . . . . . . . . . . . . . . . 128
4.3 An Equity Market Trading Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.3.1 Informed Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.3.2 Liquidity Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.3.3 Momentum Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.4 Simulation in Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.4.1 Trading Instructions: Simulation A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.4.2 Trading Instructions: Simulations B-1 and B-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.4.3 Trading Instructions: Simulation C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Modeling Securities Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Further Details on Simulation Model and Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Components of the Market Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

R.A. Schwartz (*)

Zicklin School of Business, Baruch College, CUNY, New York, NY, USA
e-mail: robert.Schwartz@baruch.cuny.edu
B.W. Weber
Lerner College of Business and Economics, University of Delaware, Newark, DE, USA
e-mail: bweber@udel.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 121
DOI 10.1007/978-1-4614-7750-1_4,
# Springer Science+Business Media New York 2015
122 R.A. Schwartz and B.W. Weber

Abstract
Financial economists have three primary research tools at their disposal: theoretical
modeling, statistical analysis, and computer simulation. In this chapter, we focus on
using simulation to gain insights into trading and market structure topics, which are
growing in importance for practitioners, policy-makers, and academics. We show
how simulation can be used to gather data on trading decision behavior and to
analyze performance in securities markets under controlled yet competitive condi-
tions. We find that controlled simulations with participants are a flexible and
reliable research tool when it comes to studying issues involving traders and market
architecture. The role of the discrete event simulation model we have developed is
to create a backdrop, or a controlled stochastic environment, for running market
experiments with live subjects. Simulations enable us to gather data on trading
participants’ decision making and to ascertain the ability of incentives and market
structures to influence outcomes. The statistical methods we use include experi-
mental design and careful controls over experimental parameters such as the
instructions given to participants. Furthermore, results are assessed both at the
individual level to understand how participants respond to incentives in a trading
setting and also at the market level to know whether the predicted outcomes are
achieved and how well the market operated.
There are two statistical methods described in the chapter. The first is discrete
event simulation and the model of computer-generated trade order flow that we
describe in Sect. 4.3. To create a realistic, but not ad hoc, market background, we
use draws from a log-normal returns distribution to simulate changes in a stock’s
fundamental value, or P*. The model uses price-dependent Poisson distributions to
generate a realistic flow of computer-generated buy and sell orders whose intensity
and supply-demand balance vary over time. The order flow fluctuations depend on
the difference between the current market price and the P* value. In Sect. 4.4, we
illustrate the second method, which is experimental control to create groupings of
participants in our simulations that have the same trading “assignment.” The result
is the ability to make valid comparisons of traders’ performances in the simulations.

Keywords
Trading simulations • Market microstructure • Order flow models • Random
walk models • Experimental economics • Experimental control

4.1 Studying Market Structure

Financial economists have three major research methods at their disposal:

theoretical modeling, statistical analysis, and simulation. We will focus on using
simulation to gain insights into trading and market structure. In so doing, we show
how simulation can be used to analyze participant behavior in a security market.
We find that controlled simulations with participants are a flexible and reliable
research tool when it comes to studying trading behavior and market architecture.
4 Simulation as a Research Tool for Market Architects 123

Using good experimental design we can draw statistically valid conclusions from
simulations at both the individual level to understand how participants respond to
incentives in a trading setting and also at the market level to know whether the
predicted outcomes are achieved and how well the market operated. We begin by
considering the interaction between the three tools.

4.2 The Interaction Between Theoretical Modeling, Empirical

Analysis, and Simulation

A theoretical formulation based on a limited number of abstract assumptions

enables complex reality to be translated into a simplified representation that can
be rigorously analyzed. As such, theoretical modeling typically provides the
underpinnings for both empirical and simulation analysis. Theory alone, however,
can take us only so far. One cannot expect that every detailed aspect of reality can
be analyzed from a theoretical vantage point (Clemons and Weber 1997).
Moreover, in light of the literature on behavioral economics, it is clear that not all
human behavior follows the dictates of rational economic modeling.
We turn to empirical analysis both to provide confirmation of a theoretical model
and to describe aspects of reality that theoretical analysis has not been able to
explain (Zhang et al. 2011). But necessary data may not be available and empirical
analysis, like theoretical modeling, has its limitations. Variability in important
variables, such as differences across time or traded instruments, is difficult for
empiricists to control (Greene 2011; Kennedy 2008). Moreover, empirical variables
may change in correlated ways, yet this should not be mistaken for a causal
mechanism. When we seek insights that neither theory nor empirical analysis
can provide, simulation has an important role to play (Parker and Weber 2012).
An example will help explain.

4.2.1 An Application: Call Auction Trading

Consider an electronic call auction. At a call, submitted orders are batched together
for simultaneous execution at a single clearing price. For the batching, buy orders
are cumulated from the highest price to the lowest to produce a function that
resembles a downward sloping demand curve, and sell orders are cumulated from
the lowest price to the highest to produce a function that resembles an upward
sloping supply curve. The algorithm generally used for determining the clearing
price at the time that the market is called finds the price which maximizes the
number of shares that execute. In an abstract, theoretical model, the number of
shares that trade is maximized at the price where the downward sloping buy curve
crosses the upward sloping sell curve. Consequently, with continuous order
functions, the clearing price which maximizes the number of shares that trade in
a call auction is uniquely determined by the point where the two curves cross, and
this price, in economics parlance, is an equilibrium value.
124 R.A. Schwartz and B.W. Weber

All told, this call auction procedure has excellent theoretic properties, and one
might expect that designing a well-functioning call would be a straightforward task.
This is not the case, however. In reality, as the saying goes, “the devil is in the details.”
For theoretical modeling, we might for analytic convenience assume a large
enough number of participants so that no one individual has any market power. We
might further assume that the cumulated buy and sell curves are continuous
functions and that the call auction is the only trading facility available. In reality,
however, the buy and sell curves are step functions and some players’ orders will be
large enough to impact a call’s clearing price.
To illustrate, assume an exact match of 40,000 shares to buy and 40,000 shares
to sell at a price of $50 and that, at the next higher price of $50.10, sell orders totaling
50,000 shares and buy orders totaling 30,000 shares exist. A buyer could move the
price up by entering more than 10,000 shares at $50.10 or greater.
Notice also that the real-world buy and sell curves are step functions
(neither price nor quantity is a continuous variable) and thus that the cumulated
buy orders may not exactly match the cumulated sell orders at the price where the
two curves cross. Moreover, many exchange-based call auctions are offered
along with continuous trading in a hybrid environment. These realities of the
marketplace affect participants’ order placement decisions and, consequently,
impact market outcomes for both price and the number of shares that trade.
In response, participants will enter their orders strategically when coming to
the market to trade.
These strategic interactions and the decisions that market participants make
depend on the call auction’s rules of order disclosure (i.e., its transparency) and its
rules of order execution which apply when an exact cross is not obtained. What
guidance do market architects have in dealing with questions such as these other than
their own, hopefully educated, speculation? The questions being raised may be too
context specific for theory to address and, if a new market structure is being consid-
ered, the data required for empirical analysis will not yet exist. This is when
simulation analysis can be used to good advantage.
Regarding transparency, alternative call auction structures include full
transparency (i.e., display the complete set of submitted orders), partial
transparency (e.g., display an indicated clearing price and any order imbalance at
that price), or no transparency at all (i.e., be a dark pool). Regarding the procedure
for dealing with an inexact cross, the alternatives for rationing orders on the
“heavy” side of the market include pro rata execution, the application of time
priority to orders at the clearing price exactly, and the application of time priority
to all orders at the clearing price and better (i.e., to higher priced buys or to lower
priced sells). These and other decisions have been debated in terms of, for
instance, the ability to game or to manipulate an auction, the incentive to enter
large orders, and the incentive to submit orders early in the book building
period before the auction is called. But definitive answers are difficult to
come by. This is when valuable guidance can be (and has been) obtained via the
use of simulation analysis.
4 Simulation as a Research Tool for Market Architects 125

4.2.2 An Application: The Search for an Equilibrium

Price and Quantity

In this section, we consider another application of simulation as a research tool: the

search for an equilibrium price and quantity in a competitive marketplace.
Determining equilibrium values for a resource’s unit price and quantity traded is
a keystone of economic analysis. In the standard formulation, equilibrium is
determined by the intersection of market demand and supply curves, with
scant consideration given to just how the buy and sell orders of participants actually
meet in a marketplace. In fact, the “marketplace” is typically taken to be
nothing more than a mystical, perfectly frictionless environment, and the
actual discovery of equilibrium values for price and quantity is implicitly
assumed to be trivial.
Real-world markets are very different. In the non-frictionless environment,
a panoply of transaction costs interact to make price and quantity discovery an
imperfect process. In this far more complex setting, two further issues need to be
analyzed: (1) the trading decisions that market participants make when
confronted by the imperfections (frictions) that characterize real-world markets
and (2) the determination of actual prices and quantities based on the
decisions of the individual participants. Both issues are ideally suited for
simulation analysis.
We consider this further in this section of the chapter, with particular reference
to one specific market: the secondary market for trading equity shares of already
issued stock. Underlying the simulation analysis is an economic model that is based
on the following assumptions:
1. The decision maker qua investor is seeking to maximize his/her expected utility
of wealth as of the end of a single holding period.
2. The investor is risk averse.
3. There are just two assets, one risk-free (cash) and one risky (equity shares).
4. There are no explicit trading cost (i.e., there are no commissions or borrowing
costs, and short selling is unrestricted).
5. Share price and share holdings are continuous variables.
6. A brief trading period (T0 to T1) that is followed by a single investment period
(T1 to T2).
7. The participant’s expectation of price at T2 is exogenous (i.e., independent of the
stock’s current price).
8. The investor is a perfect decision maker when it comes to knowing his/her
demand curve to hold shares of the risky asset.
From this set of assumptions, and following the derivation described in
Ho et al. (1985) and Francioni et al. (2010), we can obtain a participant’s demand
to hold shares of the risky asset. The participant’s demand is described by two
simple, linear functions:

P0 ¼ a  2bN ðan ordinary demand curveÞ

126 R.A. Schwartz and B.W. Weber

and

PR ¼ a  bN ða reservation price demand curveÞ

where P0 denotes price with respect to the ordinary curve, PR denotes a reservation
price, and N is the number of shares held.
The ordinary curve shows that if, for instance, the price of shares is P01 , the
participant maximizes expected utility by holding N1 shares; if alternatively price is
P02 , the participant maximizes expected utility by holding N2 shares; and so on.
Values given by the reservation curve show that at a quantity N1, the maximum the
participant would pay is PR1 when the alternative is to hold no shares at all; or, at
a quantity N2, the maximum the participant would pay is PR2 when the alternative is to
hold no shares at all; and so on. Identifying the reservation price demand curve
enables us to obtain easily a monetary measure of the gains from trading.
To facilitate the exposition, assume for the moment that the participant initially
holds no shares. Then if, for instance, N1 shares are acquired, we have


Surplus ¼ N1 PR  P

where “Surplus” denotes the gains from buying N1 shares and P is the price at which
the N1 shares were bought (note that, for a purchase, we have Surplus > 0 for
P < PR). The participant controls P via the price of his/her order but knows neither the
price at which the order will execute (if price improvement is possible) nor whether
or not the order will, in fact, execute. Because P is not known with certainty, Surplus
is not known with certainty and thus the investor seeks to maximize the expected
value of Surplus. It follows from the derivations cited above that the maximization of
the expected value of Surplus is consistent with the maximization of the expected
utility of the end of the holding period wealth (i.e., at time T2).
With the demand curves to hold shares established, it is straightforward to obtain
linear functions that describe the investor’s propensity to buy and to sell shares in
relation to both ordinary and reservation prices. If the investor’s initial position is
zero shares, the buy curve is the same as the downward sloping demand curve. By
extending the demand curve up and through the price intercept (a) into the negative
quadrant, we see that at prices higher than the intercept, a, the investor will want to
hold a negative number of shares (i.e., establish a short position by selling shares
that he/she does not currently own). By flipping the portion of the demand to hold
curve that is in the negative quadrant into the positive quadrant (and viewing
a negative shareholdings adjustment as a positive sell), we obtain a positively
inclined (vertical) mirror image that is the sell function.
The procedure just described can easily be replicated for any initial share
holdings (either a long or a short position). The individual buy and sell curves
can be aggregated across investors and, following standard economic theory, the
intersection of the aggregate buy curve with the aggregate sell curve establishes
the equilibrium values of share price and the number of shares that trade.
This equilibrium would be achieved if all participants were, simultaneously,
4 Simulation as a Research Tool for Market Architects 127

to submit their complete demand functions to the market, as they presumably would
in a perfect, frictionless environment.
Continue to assume that investors know their continuous, negatively inclined
demand curves to hold shares of the risky asset, but let us now consider how they
might operate in a non-frictionless marketplace where they cannot all simulta-
neously submit their complete and continuous, downward sloping buy functions
and upward sloping sell functions. What specific orders will they send to the
market, and how will these orders interact so as to be turned into trades? Will
equilibrium values for price and quantity be achieved? This highly complex issue
might best be approached by observing participant behavior and, to this end,
simulation can be used as the research tool. Here is how one might go about it.
Let participants compete with each other in a networked, simulated environment.
Each participant is given a demand curve to hold shares of a risky stock, and each is
asked to implement that demand curve by submitting buy or sell orders to the market.
The participants are motivated to place their orders strategically given their demand
curves, the architectural structure of the marketplace, and the objective against which
their performance is assessed – namely, the maximization of expected surplus.
The simulation can be structured as follows. Give all of the participants in
a simulation run the same demand curve,

P ¼ 20  0:5N0

where N0 represents the number of shares initially held. Divide the participants into
two equal groups, A and B, according to the number of shares they are initially
holding, with N0A ¼ 4 for group A players and N0B ¼ 8 for group B players.
Accordingly, the buy curve for each individual in group A is

P ¼ 18  0:5Q

and the sell curve for each individual in group B is

P ¼ 16 þ 0:5Q

where Q is the number of shares bought or sold. The associated reservation curves are

PR ¼ 18  0:25Q ðfor the buyersÞ

and

PR ¼ 16 þ 0:25Q ðfor the sellersÞ

From the ordinary (as opposed to the reservation) buy and sell curves, and
recalling that the groups A and B are of equal size, we obtain an equilibrium
price of 17 and an equilibrium quantity traded of 2 (per participant). From the
128 R.A. Schwartz and B.W. Weber

reservation buy and sell curves, we see that if each participant bought or sold two
shares, the surplus for each would be

2ð17:50  17:00Þ ¼ $1:00 ðfor the buyersÞ

and

2ð17:00  16:50Þ ¼ $1:00 ðfor the sellersÞ

Participants, however, do not know the distribution of initial shareholdings

across the other investors, and thus they know neither the buy and sell functions
of all the other traders, nor the equilibrium price and quantity for the market. It is up
to each of them individually to submit their orders wisely, and it is up to all of them
collectively to find the equilibrium price along with the quantity to trade at that
price. How do they operate? How well do they do? How quickly and successfully
can they collectively find equilibrium values for P and Q? And how are participant
decisions and market outcomes affected by different market structures? With regard
to each of these issues, important insights can be obtained through the use of
simulation as a research tool.

4.2.3 The Realism of Computer-Generated Data Versus

Canned Data

A live equity trading simulation can either depend exclusively on person-to-person

interaction or add computer-driven order flow; we focus on the latter. Most exper-
imental economics research is based on the former. Computer-driven simulations
can be based on either canned data or on data that the computer itself generates; we
focus on the latter. Canned data have two major limitations. First, participants in the
simulation cannot affect the stream of prices that they are trading against and,
consequently, the dynamic, two-way interaction between live participants and the
marketplace is absent. Second, canned data cannot be used to analyze new market
structure because, quite simply, it is exclusively the product of the actual market
within which it was generated.
On the other hand, a simulation model based on computer-generated data faces
a formidable challenge: capturing the dynamics of a real-world market. Canned
data does not face this problem – it is, after all, generated in a real marketplace. We
discuss the challenge of realism in this section of the chapter with specific reference
to an equity market.
It has been well established in the financial economics literature that, in an equity
market which is fully efficient, security prices follow random walks. “Efficiency” in
this financial markets context is generally understood as referring to “informational
efficiency,” by which we mean that market prices reflect all existing information. In
a nutshell, the raison d’être of a stock’s price following a random walk in a fully
informationally efficient market can be understood as follows. If all (and we mean all)
4 Simulation as a Research Tool for Market Architects 129

information about a security is reflected in a stock’s market price, then only totally
new, totally unanticipated information can cause the stock’s price to change. But
totally new and thus totally unanticipated information can be either bullish or bearish
with equal probability and thus, with equal probability, can lead to either positive or
negative price changes (returns). Thus, the argument goes, in an informationally efficient
market, returns are not predictable and stock prices follow random walks.
It would be relatively straightforward to structure an equity market simulation
based on machine-driven prices that follow a random walk. One would start
a simulation run with an arbitrarily selected seed price and have that price evolve
as the simulation progresses according to random draws from a (log-normal) returns
distribution with arbitrary variance and a zero mean. Real-world prices do not
evolve in this fashion, however. In a world characterized by trading costs, imperfect
information, and divergent (i.e., nonhomogeneous) expectations based on publicly
available information, prices do not follow simple random walks. Rather, price
changes (returns) in relatively brief intervals of time (e.g., intraday) evolve in
dynamic ways that encompass complex patterns of first-order and higher-order
correlations. Structuring a computer simulation to produce prices that capture this
dynamic property of real-world markets is the objective. In the next session of this
chapter, we set forth the major properties of a machine-driven trading simulation,
TraderEx, which we have formulated so as to achieve this goal.

4.3 An Equity Market Trading Simulation Model

In this section, we focus on a key conceptual foundation of the TraderEx simulation

model: how the machine-generated order flow is structured. Above all else, it is this
modeling that captures the dynamic property of trades and prices and, in so doing,
that enables our software to compete with canned data in terms of realism. First, we
present a brief overview of the functions the computer performs.
Looking under the hood, the TraderEx software can be thought of, first and
foremost, as a package of statistical distributions. The prices and sizes of the
machine-driven orders that power the TraderEx simulation are determined by
draws from these distributions. The software also maintains and displays an order
book or set of dealer quotes, drives a ticker tape that displays trade prices and sizes,
computes summary statistics for each simulation run (e.g., a stock’s volume-
weighted average price and participant performance measures such as profit or
loss), and provides post trade analytics (both statistics and graphs). The simulation
game can be played either individually (i.e., one person interacting with the
machine-driven order flow in a solitaire environment) or as a group (i.e., in
a networked environment that also includes machine-driven order flow).
While our machine-driven order flow, as we have said, is powered by random
draws from distributions, we are able to tell meaningful economic stories about
these statistical draws. These stories involve exogenous information change and
three different economic agents: informed traders, liquidity traders, and noise
traders. Interestingly, while the simulation requires this tripartite division, it has
130 R.A. Schwartz and B.W. Weber

also been established in the academic microstructure literature that, for an equity
market not to fail, informed traders must interact with liquidity traders, and noise
traders must also be part of the mix.

4.3.1 Informed Orders

In TraderEx, orders submitted by informed traders are specified with reference to an

exogenous variable we call P* that can be viewed as an equilibrium price. That is, at the
value P*, the aggregate flow of buy and sell orders is in balance (much as, in economic
analysis, buy and sell orders are in balance at the price where a demand curve crosses
a supply curve). More specifically, the TraderEx market is in equilibrium when the
lowest posted offer is greater than P* and the highest posted bid is less than P*. When
this equilibrium is achieved, no informed orders are submitted to the market, and an
incoming (liquidity) order can be from a buyer or a seller with equal probability. On the
other hand, if P* is greater than the lowest posted offer, informed orders kick in and
the probability of an incoming order being from a buyer is raised to 0.6 (a parameter
that can, of course, be adjusted). Equivalently, if P* is lower than the highest posted bid,
informed orders again kick in and the probability of an incoming order being
from a seller is raised to 0.6. This asymmetry between the buy and sell orders
that exists when P* is not within the quotes keeps the quotes loosely linked to P*.
P* evolves as the simulation progresses according to a Poisson arrival process.
Each jump in P* symbolizes informational change. The size of the change in P* at
each new arrival is determined by a draw from a log-normal returns distribution
with a zero mean and a variance that is a controllable parameter.

4.3.2 Liquidity Orders

The second component of the order flow, liquidity orders, is also modeled as
a Poisson arrival process, but with one important difference: at any point of time,
the probability of the newly arriving liquidity order being a buy equals the
probability of its being sell equals 0.5. All liquidity orders are priced, with the
price determined by a draw from a double triangular distribution that is located with
reference to the best posted bid and offer quotes.
A new liquidity order is entered on the book as a limit order if it is a buy
with a price lower than the best posted offer or if it is a sell with a price higher
than the best posted bid. A new liquidity order with a price equal to or more
aggressive than the best posted offer (for a buy) or the best posted bid (for a sell) is
executed immediately as a market order. Liquidity orders can (randomly) cause the
market’s bid and ask quotes to drift away from the equilibrium value, P*. When this
occurs, informed orders that are entered as market orders pull market prices
back towards P*.
4 Simulation as a Research Tool for Market Architects 131

4.3.3 Momentum Orders

The third component of the order flow is orders entered by noise traders. TraderEx
activity includes just one kind of noise trader – a momentum player – and it operates
as follows: whenever three or more buy orders (or sell orders) arrive sequentially,
the conditional probability is increased that the next arriving order will also be
a buy (or a sell).
As in the microstructure literature, noise traders are needed in the simulation
model to keep participants from too easily identifying price movements that have
been caused by informed orders responding to a change in P*. This is what our
momentum orders achieve. For instance, assume that P* jumps several ticks above
the best posted offer. An accelerated arrival of informed buy orders would be
triggered and prices on the TraderEx book would rise over a sequence of trades,
causing a pattern of positively autocorrelated price movements that can, with
relative ease, be detected by a live participant. But, to obscure this, the momentum
orders create faux price trends that mimic, and therefore obfuscate, the information-
induced trends.
Momentum orders play a further role in the TraderEx simulations. They sys-
tematically cause transaction prices to overshoot P*. Then, as informed orders kick
in, prices in the simulation mean revert back to P*. This mean reversion and its
associated accentuated short-run volatility encourage the placement of limit orders.
This is because overshooting causes limit orders to execute, and limit order placers
profit when price then mean reverts. To see this, assume that the stock is currently
trading at the $23.00 level and that P* jumps from $23.00 to $24.00. As price starts
to tick up to $24.00, momentum orders join the march and carry price beyond
$24.00–$24.20. Assume a limit order to sell is on the book at $24.20 and that it
executes. The limit order placer then benefits from having sold at $24.20 when the
momentum move ends and when a P* of $24.00 exerts its influence and price mean
reverts back towards $24.00.

4.4 Simulation in Action

Since the earliest version of TraderEx was developed in 1995, we have run
hundreds of “live market” simulations with students in our trading and market
microstructure electives and with executive education participants. In addition, we
have developed training modules on trading for new hires at a number of global
banks. We have also run controlled experimental economics studies of trading
decision making and alternative market structures (Schwartz and Weber 1997).
To illustrate the potential for research from using market simulation, we will
examine the data generated by simulation participants’ behavior in a number of
settings. In a January 2011 class session of the “Trading and Financial Market
Structure” elective at London Business School, a simulation was run which covered
132 R.A. Schwartz and B.W. Weber

Fig. 4.1 Initial order book at

start of simulated trading day.
Limit orders to buy are on the
left and limit orders to sell on
the right. Participants enter
buy and sell market orders in
the white boxes at the top of
the book. Limit orders are
entered by clicking on the
gray rectangles at the price
level the user selects

1 day of trading in an order-driven market structure. In it, 42 graduate students were

divided into 21 teams of two. The order book maintained price and time priority
over limit orders. The price increment was 10 cents, and the order book at the open
looked similar to Fig. 4.1 below.

4.4.1 Trading Instructions: Simulation A

Eight teams were each given the instruction to sell 1,500 units, and seven teams
were each asked to buy 1,300. Five other teams had the role of either day traders or
proprietary trading desks. Three of these five teams were instructed to buy 900 then
sell 900 and to have a closing position of 0. Two teams were asked to sell 800, then
buy 800, and to close with a flat position. A trial simulation was run, and perfor-
mance metrics were discussed before the simulation began. The teams with a sell
instruction were told they would be assessed on the basis of the highest average
selling price, while buying teams competed on the basis of the lowest average
buying price. The “prop” teams were told that maximizing closing P&L was their
objective but that they should finish flat and have no “overnight risk.” A screen
similar to the one the participants saw is shown in Fig. 4.2.
 4
Simulation as a Research Tool for Market Architects
133

Fig. 4.2 End of a trading day in a networked order book simulation from the screen used by the instructor or simulation administrator
134 R.A. Schwartz and B.W. Weber

1,500
Closing
position
1,000 goals: +1300,
-1,500, or 0

500

0
LBS18

LBS14

LBS12

LBS10

LBS16

LBS11

LBS19

LBS13

LBS15

LBS20

LBS21
LBS4

LBS6

LBS2

LBS5

LBS3

LBS1

LBS8

LBS7

LBS9
500

−1,000

−1,500

−2,000

Fig. 4.3 Final positions of 21 trading teams. Teams were supposed to end with holdings of either
1,500 (sellers), +1,300 (buyers), or flat (0, prop-day traders)

One of the first lessons of behavioral studies of trading done via simulation is
that following instructions is not simple for participants. As Fig. 4.3 shows, seven of
the 21 teams did not end the simulation with the position they were instructed to
have. Three of the selling teams sold more than instructed and three of the
proprietary trading teams had short, nonzero positions at the end of the day. Of
course, the noncompliant teams had excuses – “the end of the day came too fast,” or
“there were not enough willing buyers in the afternoon,” or “we didn’t want to pay
more than the VWAP price on the screen.” These complaints could, of course, also
apply in a real equity market.
In the simulation, the market opened at £20.00. The day’s high was £23.60, the
low £18.80, and the last trade of the day was at £21.50. The £4.80 high-low range
(24 %) reflects a volatile day. The day’s volume-weighted average price (VWAP)
was £20.03. Trading volume was 42,224 units and 784 trades took place. Although
the teams were in the same market, and had the same buying or selling instructions
with the same opportunities, there were vast differences in performance. As Fig. 4.4
shows, the best buying team paid £19.68, or £0.35 less than both the worst team and
VWAP, adding nearly 2 % to the investment return. The best selling team received
£0.95 more per share than VWAP and £1.23 per share more than the worst selling
team. This outcome would add almost 5 % to one selling investor’s return relative
to the average. The conclusion is clear: trading performance has a substantial
impact on investors’ returns.
4 Simulation as a Research Tool for Market Architects 135

100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
£19.683
LBS11
Buy 1,300
LBS5

LBS3

LBS19

LBS1

LBS8

LBS13

Sell 1,500 £20.977

LBS18

LBS14

LBS4

LBS6

LBS2

LBS12

LBS10

LBS16
VWAP = £20.03
£19.60 £19.70 £19.80 £19.90 £20.00 £20.10 £20.20

Fig. 4.4 Performance of trading teams as measured by average buying or selling price (scale
shown on bottom of figure reading left to right) and the percentage of teams’ trading done via limit
orders (scale shown on top of figure reading right to left)

Also shown in Fig. 4.4 is the team’s use of limit orders. Again, there is
substantial variation, with the best buying team trading exclusively with limit
orders and the best selling team completing 79 % of its trading with limit orders.
Note that, as the chart shows, a higher use of limit orders did not assure good trading
outcomes.
Note in Fig. 4.4 that the buying teams all matched or improved on VWAP,
while only three of the selling teams were able to for more than VWAP. Teams’
trading was completed with a varied combination of limit orders and market
orders, with the best team, for instance, selling its 1,500 with 77 % limit orders
and 23 % market orders. No significant correlation existed between order choice
and performance.
As Fig. 4.5 shows, among the five proprietary trading teams, the largest loss was
generated by the team (LBS21) that also had the largest risk as measured by average
absolute value of their inventory position during the trading day. The greatest profit
came from LBS15, a team that only took a moderate level of risk. Again, the
simulation reveals substantial variation and behavioral differences across traders.
136 R.A. Schwartz and B.W. Weber

Average Position
-1,345 P&L LBS21

LBS20

LBS15

LBS9

LBS7

−750 −500 −250 0 250 500 750

Fig. 4.5 Performance of 5-day trading teams. Collectively the proprietary traders lost money, and
only two teams returned to a zero position

4.4.2 Trading Instructions: Simulations B-1 and B-2

A quote-driven market structure was used in two other simulations with the
same graduate students. The underlying order flow and P* model simulated by
the computer is the same as in the order-driven market. The students were put into
22 teams, and two markets (i.e., two separate networks) with 11 teams each
were run. In each of the networks, seven teams were market makers, and the
other teams were given large orders to fill over the course of the day.
In B-1, the market opened at 20.00, and the day’s high and low prices were 21.00
and 18.70, respectively. The last trade was 19.50, and the VWAP was 20.00, with
706 trades generating a volume of 39,875 units. Participants were told that a unit
represented 1,000 shares and that those with buy or sell instructions were handling
institutional-sized orders that could affect prices.
In their trading, four of the market makers generated a positive profit; as
a group they earned 1.4 pence per share, or seven basis points, in
their trading. As Fig. 4.6 shows, only three of the dealers (LBS1, LBS3, LBS2)
ended the day with a flat (either zero or no more than 35 units) closing
position. The chart shows the seven dealers’ average inventory position, closing
profit or loss, and their closing inventory position. LBS4, for instance, had an
average position over the day of +797, a loss of 674, and a closing
position of +1,431.
4 Simulation as a Research Tool for Market Architects 137

Average Position=> LBS37

P&L=>
Closing Position=>
LBS9

LBS1

LBS3

LBS8

LBS4

LBS2

−1000 −500 0 500 1000 1500

Fig. 4.6 Performance of seven market maker trading teams. The teams that controlled risk by
keeping their average absolute value of their positions below 350 were able to make profits. Teams
LBS4 and LBS2 incurred large losses and held large risk positions

In the B-2 market, the simulation started at 20.00, and the day’s high and low
prices were 22.30 and 19.00, respectively. The last trade was 22.20, and the VWAP
was 20.22, with 579 trades generating volume of 36,161 units. Reflecting
higher volatility, only two market makers generated a positive profit and, as
a group, they lost 2.7 pence per share, or 13 basis points, in their trading.
Figure 4.7 shows that three of the six dealers (LBS31, LBS25, LBS3) had
a flat position at the end of the day. The other dealers had short positions
reflecting the P* (equilibrium price) increases and buying pressure that drove
the prices up during the simulation.

4.4.3 Trading Instructions: Simulation C

An order-driven market structure was used to study trading when participants

are given either an upward sloping supply curve or a downward sloping demand
curve. Rather than being instructed to buy a fixed quantity of shares, buyers and
sellers are asked to maximize their surplus or profit, which is the number of shares
bought or sold times the amount their reservation price differs from the average
price they paid. The reservation prices for buyers are decreasing in the quantity they
hold in their position, a demand curve. The reservation prices for sellers are increasing
in the quantity they have sold, a supply curve. Our interest is in whether a group of
trading participants will trade the optimal quantity in the market or whether in the
absence of explicit order quantities they overtrade, or trade too little.
138 R.A. Schwartz and B.W. Weber

Average Position=> LBS31

P&L=>
Closing Position=>
LBS26

LBS30

LBS25

LBS3

LBS28

−1200 −900 −600 −300 0 300 600

Fig. 4.7 Performance of six market maker trading teams. Only two teams, LBS #31 and #26, were
able to make profits, and the teams that allowed their average position to exceed 300 had the
largest losses

In the example below, we provide the buyers’ reservation price function, which starts
at $26.00 but decreases by one for each 1,000 shares bought. If a participant buys 6,000
shares over the course of the simulation, for instance, the reservation price is $20.00.
If they paid on average $18, then the surplus generated is 12,000.

Reservation BUY curve: PR = $26 – (X units/1,000)

# shares bought = _________
PR = 26 – (# shares bought /1,000) = _________
Average buy price = _________
Surplus = (# shares) × (PR – Avg buy price) = _________

The sellers’ reservation price function is below. It starts at $12.00 but increases
by one for each 1,000 shares bought. If a participant sells 2,000 shares, the
reservation price is $14.00. If the sellers’ average selling price is $18, for instance,
the surplus is 16,000.

Reservation SELL curve: PR = $12 + (X units/1,000)

# shares sold = _________
PR = 12 + (# shares sold/1,000) = _________
Average sell price = _________
Surplus = (# shares) × (Avg sell price – PR) = _________
4 Simulation as a Research Tool for Market Architects 139

Fig. 4.8 Trading experiment with supply and demand functions and surplus calculations based on
a participant trading at the market average price over the day of $19.11. The optimal trading
strategy was to end with a position of 3,500 long (buyer) or short (seller)

We provided the surplus functions above to a group of 28 participants. The group

was split into teams of two, with seven buying teams and seven selling teams.
Although the teams were not aware of the other teams’ curves, the equilibrium price
was $19, and the maximum surplus was achieved by participants that built
a position of 3,500 and bought at the lowest possible prices or sold at the highest
prices (see Fig. 4.8). In our experiment, average per share trade price was $19.11.
In the experiment, participants in the 14 teams built positions ranging from 1,100
to 7,500 (see Fig. 4.9). There was a substantial dispersion of performance, yet
the market’s trade prices converged to within $0.11 of the $19.00 equilibrium price.
The average ending position was 4,001, so participants were within 15 % of the
optimal position implicit in the reservation price functions. The teams with the
greatest surpluses were able to come close to the optimal position of 3,500 and to
buy below the average price and sell at greater than the average price.
As these three simulation exercises show, a number of insights can be gained
into the effect of market structure on participant behavior and market outcomes
(the quality of price and quantity discovery) by running “live market” simulations.
First, participants’ performance varied widely despite being in the same environ-
ment with the same trading instructions. Second, markets are complicated and not
 140

Fig. 4.9 Even when provided with reservation price functions, the live participants in the markets traded near the equilibrium and showed little inclination to
“overtrade.” The average position was only 501 units away from the optimal closing position of 3,500
R.A. Schwartz and B.W. Weber
4 Simulation as a Research Tool for Market Architects 141

perfectly liquid; filling large orders, while competing to outperform price bench-
marks, is challenging. Participants often did not complete their instructions even
though completion was part of their assessment. Third, trading can add (or subtract)
value for investment managers. Even in fairly simple simulation games, perfor-
mance varied widely across participant teams. Finally, risk was not well controlled
by the (admittedly inexperienced) trading participants. Although given specific
position limits as guidelines, many participants had large average positions and
suffered substantial losses from adverse price movements.
Market architects today combine different trading systems designs (Zhang et al.
2011). Beyond studies of trading decision making, the live simulations provide
a method for comparing alternative market structures. For instance, implicit trading
costs incurred in handling a large order over a trading day can be contrasted in
markets with and without dealers, and with and without a dark liquidity pool. Live
simulations, by extending what can be learned with analytical modeling and
empirical data analytics, provide a laboratory for examining a broad set of questions
about trading behavior and market structures.

4.5 Conclusion

Simulation is a powerful research tool that can be used in conjunction with

theoretical modeling and empirical research. While simulation can enable
a researcher to delve into issues that are too detailed and specific for theory to
handle, a simulation structure must itself be based on a theoretical model. We have
illustrated this with reference to TraderEx, the simulation that we have developed
and used to analyze details of market structure and trading in an equity market.
One further research methodology incorporates simulation: experimental economics.
This application applies the simulated market in a laboratory where multiple players are
networked together. With a well-defined performance measure and carefully crafted
alternative market structure and/or information environments, a simulation-based exper-
imental application can yield valuable insights into the determinants of market outcomes
when market structure affects individual behavior and when behavioral economics
along with theoretically rational decision making characterizes participant actions.

Appendix

Modeling Securities Trading

The simulation methodology was chosen for its ability to accommodate critical
institutional features of the market mechanism and off-exchange dealers’ opera-
tions. While higher level abstractions and simplifications could yield an analytically
tractable model, it is not consistent with the goals of this chapter. These complex-
ities included here are the specialist’s role, the use of large and small order sizes,
and the limit and market orders.
142 R.A. Schwartz and B.W. Weber

Earlier market structure research provides useful insights, but missing institu-
tional details prevent it from determining the effect of third markets. Garman’s
(1976) model of an auction market identifies a stochastic order arrival process and
a market structure consistent with negative serial autocorrelations or the tendency
of price changes to reverse themselves. Garman’s model has no specialist role, and
an analytic solution is obtained only for the case with one possible price. He notes
“the practical difficulties of finding analytic solutions in the general case are
considerable, and numerical techniques such as Monte Carlo methods suggest
themselves.” Mendelson (1987) derives analytic results that provide a comparison
of market consolidation or fragmentation on market performance. The work
provides numerous insights into market design trade-offs. As a simplification,
however, all orders from traders are for one unit of the security.
Simulation has yielded useful results in other microstructure research. Garbade
(1978) investigated the implications of interdealer brokerage (IDB) operations in
a competing dealer market with a simulation model and concluded that there
are benefits to investors from IDBs through reduced dispersion of quotes and trans-
action prices closer to the best available in the market. Cohen et al. (1985) used
simulation to analyze market quality under different sets of trading priority rules. They
showed that systems that consolidate orders and that maintain trading priorities by an
order’s time of arrival in the market increase the quality of the market. Hakannson
et al. (1985) studied the market effects of alternative price-setting and own-inventory
trading policies for an NYSE-style specialist dealer using simulation. They found that
pricing rules “independent of the specialist’s inventories break down.”

Further Details on Simulation Model and Environment

Our simulation model has been used in experiments with subject-traders to test
hypotheses concerning market structures. The simulation model is dynamic, with
informational changes occurring in a way that creates the possibility of realizing
trading profits from close attention to price changes and careful order handling. In
our human-machine interactive environment, the computer generates orders from
an unlimited number of “machine-resident” traders and investors. This enables us
easily to satisfy the conditions for an active, competitive market.
To be useful and valid, simulation models must reflect real-world dynamics
without being burdened by unnecessary real-world detail. A simulation model also
requires a strong theoretical foundation. The advantage of simulation over theoretical
modeling is that “theorizing” requires abstracting away from some of the very details
of market structure that exchange officials and regulators wish to study. Consequently,
theoretical modeling can give only limited insight into the effects of market design
changes on the behavior of market participants. The advantage of simulation vis-à-vis
empirically testing of new market structures is that the simulated experiments can be
run at much lower cost and across a broader range of alternatives.
The objective of our computer simulation is to provide a backdrop for assessing
the decisions of live participants. Trading in the model is in a single security and is
4 Simulation as a Research Tool for Market Architects 143

the result of “machine-generated” order flow interacting with the order placement
decisions of the live participants. Assumptions are made about the arrival process
of investors’ orders, changes to a price, p*, which is an “equilibrium value” at which
the expected arrival rate of buy orders equals the expected arrival rate of sell orders.
P* follows a random walk jump process. In other words, the equilibrium value
jumps randomly from one level at interarrival times based on sampling an exponential
distribution. After a shift in p*, the orders of the informed traders pull the quotes and
transaction prices up or down, causing them to trend towards the new p* level.
Occasionally, market prices can also trend away from p* because of the orders of
momentum traders or the chance arrival of a string of buy or sell liquidity orders.
However, movements away from p* are unsustainable; eventually the imbalanced
order flow causes a price reversal and market prices gravitate back towards p*.
Little work in experimental economics has used computers to create background
order flow into which participants individually enter orders (Smith and Williams
1992; Friedman 1993). Our test environment does. We use discrete event computer
simulation to do the following:
• Generate a background public order flow that can (1) be placed on a public limit
order book for later execution, or (2) execute as market orders immediately
against the public limit order book.
• Give the live participants the opportunity to trade a quantity of stock. Depending
on the market structure, the live participants can (i) place them in a public limit
order book and wait for a trade to occur or (ii) execute them against the public
limit order book immediately. Variants of the simulation model include market
makers, or call auctions, or dark liquidity pools to facilitate transactions.
• Maintain the screen which displays (i) orders on the public limit order book,
(ii) a time-stamped record of all transaction sizes and prices for each trading
session, and (iii) the users’ position, risk, and profit performance data.
• Capture information concerning (i) the live participants’ decisions and
(ii) market quality measures such as bid-ask spreads.
In summary, with this realistic and theory-based model and the ability in the
simulation to control the level of transparency provided, we have a rigorous
environment to assess trading decisions and the effects of different market rules.

Components of the Market Model

In the simulation model, assumptions are made about the arrival process of
investors’ orders, elasticity of supply and demand, and order placement strategies,
price volatility, and the proportions of market and limit orders.
Order Arrival. Orders arrive according to a price-dependent Poisson function.
Using time-stamped transactions data on six stocks traded on the London Stock
Exchange, a Kolmogorov-Smirnov goodness-of-fit test fails to reject the null
hypothesis of exponential interarrival in 17 out of 22 sample periods at the 0.10
level of significance. We would expect to reject in just over two cases due to
random realizations. The fit is not perfect in part because transactions tend to
144 R.A. Schwartz and B.W. Weber

Fig. 4.10 Buy and sell order $35.00

arrival rates. At market prices
greater than p*, sell orders
will arrive with greater $34.00 S(p(t))
intensity than buy orders. At
market prices less than p*,

Price
buy orders will arrive with Stochastic
$33.00
greater intensity than sell equilibrium,
orders p*(t)
$32.00 D(p(t))

$31.00
5.5 6.5 7.5 8.5 9.5
Arrival Rate: Orders per Hour

cluster somewhat more than predicted by the theoretical model (Weber 1991).
Given the shortcoming of using empirical distributions in simulations (Law and
Kelton 1989), the Poisson assumption appears sufficiently justified for capturing the
typical behavior of the order arrival process.
The Poisson interarrival time, T, is exponentially distributed with bt equal to the
mean interarrival time at time t. The mean interarrival time is set at the beginning of
each experiment and assumed to hold constant. A realization at time t is thus
Tt ¼ C(b). The supply and demand structure follows closely those previously
developed in the market microstructure literature (Garbade and Silber 1979), in
which buy and sell arrival rates are step functions of the difference between the
quoted price and the equilibrium value of the security Fig. 4.10.
Garman (1976) termed the intersection of the supply and demand functions
a “stochastic equilibrium.”
Demand/buy orders, D(p):

lB ðpi ; p Þ ¼ a1 for pt  < pi

lB ðpi ; p Þ ¼ a1 þ a2 ðpt   pi Þ for pi ¼ pt  þ a
with a ¼ tick size, 2ðtick sizeÞ, 3ðtick sizeÞ, . . . d
lB ðpi ; p Þ ¼ a1 þ a2 d for pt   pi > d

Supply/sell orders, S(p):

 
lS pj ; p ¼ a1 for pt  > pi
   
lS pj ; p ¼ a1 þ a2 pj  pt  for pj ¼ pt   a
with a ¼ tick size, 2ðtick sizeÞ, 3ðtick sizeÞ, . . . d
 
lS pj ; p ¼ a1 þ a2 d for pj  pt  > d
4 Simulation as a Research Tool for Market Architects 145

The constant a1 reflects the proportion of arrivals that are market orders.
The coefficient a2 determines the arrival rate of limit orders with reservation prices.
Limit order traders are sensitive to discrepancies between available prices and the
equilibrium value. The parameter, d, is the range around the equilibrium
value from which limit prices for limit orders are generated. At a price pi lower
than the equilibrium value at the time, pt*, the arrival rate of buy orders will
exceed the rate of sell order arrivals. The resulting market buy orders and limit
order bids will exceed the quantity of sell orders for below-equilibrium values.
The arrival rate discrepancy will cause prices to rise since in expectation,
orders will trade against the lowest offer quotes, and add new, higher priced
bid quotes.
Order Size. Orders are between one and 250 units of the security. This reflects
a convenient normalization that is consistent with the empirically observable range
of order sizes. A unit may represent, for instance, three round lots, or 300 shares.
Beyond 250 units, we assume the trade would be handled as a block trade, and
negotiated outside of the standard market design, or arrive in the market in smaller
broken-up pieces. Large orders can have “market impact,” and can move prices up
for large buyers, and force them down for larger sellers. The functioning of the
market for large orders is consistent with observed trade discounts for large sell
orders and premiums for large buy orders.
Order Placement Strategies. The machine-generated order flow consists of
liquidity, informed, and momentum trading orders. The liquidity orders are either
limited price orders or market (immediately executable) orders. Market orders
execute on arrival but are “priced” to reflect a maximum acceptable premium or
discount to the current bid or offer. If the market order is large enough, its
price impact (the need to hit successive lower priced bids or lift higher priced
offers) will exceed the acceptable discount or premium, and the remaining
order quantity will become a limit order after partially executing against the
limit order book.
Information Generation. Idiosyncratic information events occur that change
the share value, p*, at which buying and selling order arrival rates are
balanced. Information event occurs according to a Poisson arrival process.
When an information innovation occurs, the price will have a random walk jump.
Price Random Walk. Idiosyncratic information events occur that change
the share value, p*, at which the arrival rate of buy orders equals the arrival rate
of sell orders. The time between information change is assumed to be exponentially
distributed with mean, 12 h. Empirical validation is difficult, because informatio-
n affects the share values in unobservable ways. When there is a change in
information that will shift the “balance price,” p* evolves according to a random
walk without return drift. To assure nonnegative prices, the natural log of price is
used, yielding a log-normal distribution for the equilibrium price. The white noise
term, et, is normally distributed with variance linear in the time since the
last observation. This is consistent with the price diffusion models used in
the financial economics literature (Cox and Rubinstein 1985):
146 R.A. Schwartz and B.W. Weber

ln pt * ¼ ln pt  T * + et where, et  N(0, Ts2)

where pt *  LN(ln pt  T *, Ts2)
The natural logarithm of the current price is an unbiased estimator of the natural
logarithm of any subsequent price:

Eðln pt þ T j pt Þ ¼ ln pt 
Empirical validation for the random walk model comes from numerous tests,
whose results “are remarkably consistent in their general finding of randomness . . .
serial correlations are found to be small” (Malkiel 1987).
Information Effects. If the bid and offer quotes straddle p*, there is no informed
order flow and buying and selling order arrival rates will be equal. When p* is
outside of the bid-offer range, additional one-sided market orders will be generated
according to a Poisson process.

References
Clemons, E., & Weber, B. (1997). Information technology and screen-based securities trading:
Pricing the stock and pricing the trade. Management Science, 43(12), 1693–1708.
Cohen, K., Conroy, R., & Maier, S. (1985). Order flow and the quality of the market. In
Y. Amihud, T. Ho, & R. Schwartz (Eds.), Market making and the changing structure of the
securities industry (pp. 93–109). Lanham, MD: Lexington Books.
Cox, J., & Rubinstein, M. (1985). Options markets. Englewood Cliffs: Prentice-Hall.
Francioni, R., Hazarika, S., Schwartz, R., & Reck, M. (2010). Security market microstructure: The
analysis of a non-frictionless market. In C. F. Lee & A. C Lee (Eds.), Handbook of quantitative
finance and risk management (pp. 333–353). Germany: Springer New York & Heidelberg.
Friedman, D. (1993). How trading institutions affect financial market performance: Some labora-
tory evidence. Economic Inquiry, 31, 410–435.
Garbade, K. (1978). The effect of interdealer brokerage on the transactional characteristics of
dealer markets. Journal of Business, 51(3), 477–498.
Garbade, K., & Silber, W. (1979). Structural organization of secondary markets: clearing fre-
quency, dealer activity and liquidity risk. The Journal of Finance, 34, 577–593.
Garman, M. (1976). Market microstructure. Journal of Financial Economics, 3, 257–275.
Greene, W. H. (2011). Econometric analysis (7th ed.). Upper Saddle River, NJ: Prentice Hall.
Hakansson, N., Beja, A., & Kale, J. (1985). On the feasibility of automated market making by
a programmed specialist. Journal of Finance, 40, 1–20.
Ho, T., Schwartz, R., & Whitcomb, D. (1985). The trading decision and market clearing under
transaction price uncertainty. Journal of Finance, 40, 21–42.
Kennedy, P. (2008). A guide to econometrics (6th ed.). Hoboken, New Jersey: Wiley-Blackwell.
Law, A., & Kelton, W. (1989) Simulation modeling and analysis, McGraw-Hill, New York, NY.
Malkiel, B. (1987). Efficient market hypothesis. In B. Malkiel (Ed.), The New Palgrave:
A dictionary of economics (pp. 127–134). London: Macmillan Press.
Mendelson, H. (1987) Consolidation, Fragmentation, and Market Performance, Journal of
Financial and Quantitative Analysis, 22, 189–207.
Parker, C., & Weber, B. (2012). How IT can disrupt markets: A simulation analysis (Working
Paper). London: London Business School.
Schwartz, R., & Weber, B. (1997). Next-generation securities market systems: An experimental
investigation of quote-driven and order-driven trading. Journal of Management Information
Systems, 14, 57–79.
4 Simulation as a Research Tool for Market Architects 147

Smith, V., & Williams, A. (1992, December). Experimental market economics. Scientific
American, pp. 116–121.
Weber, B. (1991) Information technology and securities markets: Feasibility and desirability of
alternative electronic trading systems, Unpublished dissertation, University of Pennsylvania.
Zhang, S. S., Storkenmaier, A., Wagener, M., & Weinhardt, C. (2011). The quality of electronic
markets. In Forty Fourth Hawaii International Conference on System Sciences, Kauai, Hawaii,
pp. 1–10.
Zhang, X. F. (2010). High frequency trading, stock volatility, and price discovery (Working
Paper). New Haven, CT: Yale University.
Motivations for Issuing Putable Debt:
An Empirical Analysis 5
Ivan E. Brick, Oded Palmon, and Dilip K. Patro

Contents
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.3 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.4 Data and Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Appendix 1: Sample of Firms Issuing Putable Bonds and the Put Bond Characteristics . . . . 175
Appendix 2: Sample of Firms Issuing Poison Put Bonds and the Put Bond
Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Appendix 3: Estimating the Standard Abnormal Returns and the White t-Statistic . . . . . . . . . . 181
Appendix 4: Generalized Method of Moments (GMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

Abstract
This paper examines the motivations for issuing putable bonds in which the
embedded put option is not contingent upon a company-related event. We find

We would like to thank Sanjay Deshmukh and conference participants at the Seattle FMA
meetings, October 2000, for helpful comments on an earlier version of the paper circulated under
a different title. The views expressed in this paper are strictly that of the authors and not of the
OCC or the Comptroller of the Currency.
I.E. Brick (*)
Department of Finance and Economics, Rutgers, The State University of New Jersey,
Newark/New Brunswick, NJ, USA
e-mail: ibrick@business.rutgers.edu; ibrick@andromeda.rutgers.edu
O. Palmon
Department of Finance and Economics, Rutgers Business School Newark and New Brunswick,
Piscataway, NJ, USA
e-mail: palmon@business.rutgers.edu
D.K. Patro
RAD, Office of the Comptroller of the Currency, Washington, DC, USA
e-mail: dilip.patro@occ.treas.gov

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 149
DOI 10.1007/978-1-4614-7750-1_5,
# Springer Science+Business Media New York 2015
150 I.E. Brick et al.

that the market favorably views the issue announcement of these bonds that we
refer to as bonds with European put options or European putable bonds. This
response is in contrast to the response documented by the literature to other bond
issues (straight, convertible, and most studies examining poison puts) and to the
response documented in the current paper to the issue announcements of poison
put bonds. Our results suggest that the market views issuing European putable
bonds as helping mitigate security mispricing. Our study is an application of
important statistical methods in corporate finance, namely, event studies and the
use of general method of moments for cross-sectional regressions.

Keywords
Agency costs • Asymmetric information • Corporate finance • Capital structure •
Event study methodology • European put • General method of moments •
Management myopia • Management entrenchment • Poison put

5.1 Introduction

This paper examines the motivations for issuing putable bonds in which the
embedded option is not contingent upon company-related events. The option
entitles bondholders to sell the bond back to the firm on the exercise date (usually
3–10 years after the bond is issued) at a predetermined price (usually at par). We
refer to these bonds as bonds with European put options or European putable bonds.
Unlike the poison put bonds (i.e., bonds with event risk covenants) that have been
studied by the literature,1 the exercise of the option in a putable bond is not
contingent upon a company-related event. This distinction is important because
a poison put protects the bondholder from a specific event (e.g., a takeover) and
may be designed to help prevent that event. In contrast, putable debt provides
protection to the bondholder against any deterioration in the value of her claim.
This distinction is important also because the two types of embedded put options
may serve different purposes. Crabbe and Nikoulis (1997) provide a good overview
of the putable bond market.
Corporate managers determine which contingent claims the company issues to
finance its activities. This choice includes the debt-equity mix and the specific
design of the debt. The design of the debt includes its maturity, seniority, collateral,
and the type of embedded options included in the bond contract. The theoretical
corporate finance literature indicates that including convertible, callable, and/or
putable bonds in the capital structure may help mitigate agency costs and reduce
asymmetric information. Haugen and Senbet (1981) theoretically demonstrate that
the optimal combination of embedded call and put options should eliminate the

1
See for example, Crabbe (1991), Bae et al. (1994, 1997), Cook and Easterwood (1994), and Roth
and McDonald (1999), Nash et al. (2003). Billett et al. (2007) find that 5 % of corporate bonds of
their sample have a non-poison putable option.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 151

asset substitution problem. Bodie and Taggart (1978) and Barnea et al. (1980) show
that bonds with call options can mitigate the underinvestment problem. Robbins
and Schatzberg (1986) demonstrate that the increased interest cost of callable bond
can be used to convey the true value of the firm to the market.2
The literature has not empirically examined the motivation and equity valuation
impact of issuing debt with European put options. We fill this gap in the literature in
several ways. First, we examine the stock price reaction around the announcement
dates of the two types of putable bond issues: bonds with European put options and
bonds with poison puts. Further, we consider four alternative motivations for
incorporating either a European or a poison put option in a bond contract and test
the hypotheses that putable debt is issued to reduce security mispricing, agency
costs of debt, management entrenchment, and myopia. The first possible motivation
for issuing putable debt is asymmetric information. Consider a company that is
undervalued by the market. The market should also undervalue its risky straight-
debt issue. However, if this firm were to issue putable bonds, the put option would
be overvalued. (Recall that put options are negatively related to the value of the
underlying asset.) Consequently, the market value of bonds with a put option is less
sensitive to asymmetric information than straight bonds, minimizing market
mispricing of these debt securities. Additionally, if the market overestimates the
risk of the bond, it may overvalue the embedded put option at issuance, thereby
increasing bond proceeds and benefiting the shareholders.3 The second possible
motivation is mitigating agency costs of debt. For example, the existence of a put
option mitigates the advantage to stockholders (and the loss to bondholders) from
risk shifting, thereby reducing the incentive to shift risk. Risk shifting hurts putable
bondholders less than holders of straight bonds because the value of the put option
(which is held by bondholders) is an increasing function of the firm’s risk. The third
possible motivation is the relatively low coupon rate (a myopic view that ignores
the potential liability to the firm due to the put option). The fourth possible
motivation is that the put option may serve to entrench management. This fourth
motivation is most relevant for firms issuing poison puts since these put options are
exercisable contingent upon company-related events that are usually related to
a change in ownership.
We find that the market reacts favorably to the issue announcement of European
put bonds. We also examine the relationship between the abnormal returns around
the put issue announcement date and firm characteristics that proxy for asymmetric
information problems, potential agency costs (i.e., risk-shifting ability and the level
of free cash flow), and management myopia. The empirical evidence is consistent
with the view that the market considers issuing putable bonds as mitigating security
mispricing caused by asymmetric information. The results do not support the idea

2
Other benefits posited by the literature include minimizing tax liabilities. See for example, Brick
and Wallingford (1985), and Brick and Palmon (1993).
3
Brennan and Schwartz (1988) offer a similar argument to explain the benefits of issuing
convertible bonds.
152 I.E. Brick et al.

that putable bonds are issued to obtain lower coupon rates (i.e., management
myopia). Our empirical findings are robust to a number of alternate specifications.
In contrast to European putable bonds, and consistent with the management
entrenchment hypothesis (see Cook and Easterwood (1994) and Roth and
McDonald (1999)), we find that the market reacts unfavorably to the issue
announcement of poison put bonds. However, consistent with Bae et al. (1994)
who argue that poison put bonds are useful in mitigating agency cost problems, we
find that the abnormal returns around the issue announcement of poison put bonds
are positively related to the protection level of the event risk covenant. Thus, our
results are consistent with the view that European put bonds are effective in
mitigating security mispricing problems, but, in contrast, poison put bonds are
related to management entrenchment or mitigating agency cost problems.
The paper’s organization is as follows. In the next section we summarize the
previous literature. Section 5.3 develops the empirical hypotheses. We describe the
data and empirical methodology in Sect. 5.4. The empirical results are summarized
in Sect. 5.5. We offer concluding remarks in Sect. 5.6.

5.2 Literature Review

The theoretical corporate finance literature concludes that the firm’s financing
decision may affect its equity value for several reasons. First, issuing debt
increases firm value because it decreases its tax liabilities.4 Second, issuing debt
may be, in part, a signaling mechanism that informs the market of private
information. For example, Ross (1977) and Ravid and Sarig (1991) demonstrate
that the manager of a firm with better prospects than the market perceives has an
incentive to signal her firm’s quality by issuing a greater amount of debt than
issued in a symmetric information environment. Third, prudent level of debt can
reduce the agency costs arising from the conflict of interest between managers and
shareholders as demonstrated by Jensen and Meckling (1976). For example,
Jensen (1986) demonstrates that leverage can minimize the deleterious effect of
free cash flow on the firm.5 However, leverage is also shown to generate other

4
See, for example, Modigliani and Miller (1963), Scott (1976) and Kim (1978). In contrast, Miller
(1977) suggests that the tax benefit of interest is marginal. However, Mackie-Mason (1990)
empirically demonstrates the significant impact of corporate taxes upon the observed finance
choices of firms.
5
Hence, empirically, we would expect that as firms announce increased levels of debt, the stock
price should increase. However, studies by Dann and Mikkelson (1984), Mikkelson and Partch
(1986), Eckbo (1986), and Shyam-Sunder (1991) indicate that there is no systematic relationship
between the announcement of firm’s debt financing and its stock price or that this relationship is
weakly negative. One potential explanation for this result is that the market can predict future debt
offerings as argued by Hansen and Chaplinsky (1993). Another potential explanation, as suggested
by Miller and Rock (1985) and documented by Hansen and Crutchley (1990), is that raising
external capital may indicate a cash shortfall.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 153

agency problems because of the conflict of interest between stockholders and

bondholders. These agency problems include underinvestment and risk shifting
(or asset substitution).6
Other studies indicate that these agency problems can be mitigated or eliminated
by optimal security design. For example, Haugen and Senbet (1981) demonstrate
that the optimal combination of embedded call and put options completely elimi-
nates the asset substitution problem. Bodie and Taggart (1978) and Barnea
et al. (1980) show that the call options of bonds can mitigate the underinvestment
problem. Similarly, Green (1984) demonstrates that including convertible bonds in
the capital structure may also mitigate the underinvestment problem.
The literature has also demonstrated that an appropriate debt security design
may alleviate asymmetric information problems. For example, Robbins and
Schatzberg (1986) demonstrate that the increased interest cost of callable bond
can be used to signal the value of the firm. Moreover, as similarly stated for the case
of convertible bonds by Brennan and Schwartz (1988), the inclusion of a put option
can minimize the mispricing of debt securities if the market overestimates the risk
of default. Chatfield and Moyer (1986) find that putable bonds may be issued by
financial institutions for asset-liability management in a period of volatile interest
rates. Tewari and Ramanlal (2010) find that callable-putable bonds provide protec-
tion to bondholders and improved returns to stockholders.
The literature examines the equity valuation impact of a special type of putable
bond known as the poison put bond. In poison put bonds, the put option is
exercisable contingent upon a company-related event such as a leveraged
restructuring, takeover, or downgrading the debt credit rating to speculative
grade. David (2001) shows that puts may have higher strategic value than intrinsic
value. Crabbe (1991) finds that such event-related covenant put bonds reduce the
cost of borrowing to the firm. Bae et al. (1994) conclude that stockholders benefit
from the inclusion of event-related (risk) covenants. Furthermore, Bae
et al. (1997) empirically document that the likelihood of a firm to include event
risk covenants is positively related to the firm’s agency costs of debt. In contrast,
Cook and Easterwood (1994) and Roth and McDonald (1999) find that the
inclusion of poison put bonds benefits both management and bondholders at the
expense of stockholders.
In contrast to these studies of poison put options, our study examines debt issues
in which the embedded put option is equivalent to a European put with a fixed
exercise date that is usually 3–10 years after the issue date. That is, bondholders
may exercise the put option only at the exercise date, and their ability to do so is not
contingent upon any particular company-related event. Thus, we believe that
issuing bonds with an embedded put option that is not contingent upon a company-
related event (such as a change in ownership) is not likely to be motivated by

6
Myers (1977) demonstrates that shareholders may avoid some profitable net present value pro-
jects because the benefits accrue to the bondholders. Jensen and Meckling (1976) demonstrate that
leverage increases the incentives for managers, acting as agents of the stockholders, to increase the
risk of the firm.
154 I.E. Brick et al.

management entrenchment. Consequently, we consider the security mispricing

motivation and two other possible motivations: mitigating agency costs and the
myopic behavior on the part of the management.

5.3 Hypotheses

In this section, we describe four alternative motivations for incorporating a put

option in a bond contract and outline their empirical implications.
The first motivation is reduction in the level of security mispricing due to
asymmetric information. Consider a company that is undervalued by the market.7
The market should also undervalue its risky straight-debt issue. However, if this
firm were to issue putable bonds, the put option would be overvalued. Conse-
quently, the market value of bonds with put option is less sensitive to asymmetric
information than the market value of straight bonds, minimizing market mispricing
of these debt securities. Additionally, the overvaluation of the implicit put option
increases the debt proceeds at the issuance thereby benefiting the shareholders. This
possible motivation has the following empirical implications. First, issuing bonds
with put option should be associated with an increase in equity value. Second, this
increase in firm value should be negatively related to the accuracy with which the
firm’s value has been estimated prior to the bond issue. Third, because the value of
the put option is directly related to the magnitude of firm undervaluation, the
increase in equity value should be directly related to the value of the put option.
The second possible motivation is mitigating the bondholder-stockholder
agency costs.8 The value of a put option is an increasing function of the firm’s
risk. Thus, its existence mitigates the gains to stockholders from, and hence their
incentives for, risk shifting. This possible motivation has the following empirical
implications. First, the benefit to a firm from incorporating a put option in the bond
contract should be directly related to the firm’s ability to shift risk in a way that
increases the value of stockholders claims at the expense of bondholders. Second,
the inclusion of a put option should help restrain management from taking on
negative net present value projects in the presence of Jensen’s (1986) free cash
flow problem.9 In the absence of a put option, undertaking negative net present
value projects should reduce security prices for both stockholders and bondholders.

7
We assume that managers of companies that are overvalued have no incentive to resolve security
mispricing. Consequently, managers of undervalued firms could send a credible signal to the
market of the firm’s undervaluation with the inclusion of a put feature. Overvalued firms should
not be able to mimic since the put option represents a potential liability that is greater to these firms
than of the undervalued firms.
8
The bondholders-stockholders agency conflict has been found to be important in security design
by Bodie and Taggart (1978), Barnea et al. (1980), Haugen and Senbet (1981), Jung et al. (1996),
and Lewis et al. (1998).
9
Equivalently, the inclusion of a put option should restrain management from consuming
a suboptimal amount of perquisites or nonpecuniary benefits.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 155

In contrast, giving bondholders the option to put the bond back to the firm at face
value shifts more of the negative price impact of undertaking negative net present
value projects to the stockholders. This negative stock price impact may reduce
management compensation that is tied to equity performance and/or induce tender
offers that will ultimately help replace the current management. These increased
costs to stockholders and management should induce management to refrain from
undertaking negative net present value projects. Thus, we hypothesize that the
benefit to the stockholders from incorporating a put option in the bond contract
should be directly related to the level of free cash flow. Third, the gain in firm value
should be related to the magnitude of the agency cost problem that the firm faces.
The valuation of the put option by the market (assuming market efficiency and
symmetric information) should be positively related to the magnitude of the agency
cost problem faced by the company. Thus, the benefit to the firm should also be
directly related to the aggregate value of the implied put option of the issue (scaled
by firm size). Fourth, for our sample of poison put bonds, the benefit to shareholders
should increase with the strength of the event risk covenant, since the greater the
strength of the event risk covenant, the less likely management will engage in value
decreasing activities. Bae et al. (1994) tested a similar set of hypotheses for
a sample of poison put bonds.10
The third possible motivation is the low coupon rate (compared to the coupon
rates of straight or callable debt issues). This motivation reflects management
myopia as it ignores the potential liability to the firm due to the put option.11 That
is, myopic management may not fully comprehend the increased risk to the firm’s
viability that is posed by the put option. In particular, bondholders would have the
right to force the firm to (prematurely) retire its debt at a time that is most
inconvenient to the firm, which in turn can precipitate a financial crisis. Further,
if the cost of financial distress is significant, given rational markets and myopic
management, issuing putable debt may negatively impact the value of equity.
This possible motivation has the following empirical implications. First, because
management myopia implies that management pursues suboptimal policies, the
issuance of putable debt should be associated with a decline in equity value.12
Second, the decline should be more severe the larger is the aggregate value of
the implied put option. Third, because expected costs of financial distress are

10
Although bonds become due following formal default, hence all bonds in a sense become
putable, bondholders usually recover substantially less than face value following formal default
because equity value has already vanished. In contrast, a formal inclusion of put option may allow
bondholders to recover the face value at the expense of shareholders when financial distress is not
imminent but credit deterioration has occurred.
11
Other studies that have examined managerial myopia include Stein (1988), Meulbroek
et al. (1990), Spiegel and Wilkie (1996), and Wahal and McConnell (2000).
12
Haugen and Senbet (1978, 1988) theoretically demonstrate that the organizational costs of
bankruptcy are economically insignificant. However, if the putable bonds increase the potential
of technical insolvency, then it will increase potential agency costs that are not necessarily
insignificant.
156 I.E. Brick et al.

negatively related to the financial stability of the company, the decline should be
more severe the lower the credit rating of the bond.
The fourth possible motivation is that the use of put options enhances manage-
ment entrenchment. This hypothesis is relevant for firms issuing poison but not
European put bonds. In particular, many bonds with event risk covenants (i.e.,
poison puts) place restrictions on the merger and acquisition activity of the issuing
firms, thereby strengthening the hands of management to resist hostile takeover
bids. This possible motivation has the following empirical implications. First,
issuing bonds with put options should be followed by a decrease in equity value.
Second, in contrast to the mitigating agency costs hypothesis, the abnormal
returns of equity around the issue announcement date should be inversely related
to the level of event-related covenant protection offered to the bondholder.

5.4 Data and Methodology

Our sample of bonds with put options is taken from the Warga Fixed Income
Database. The database includes pricing information on bonds included in the
Lehman Brothers Bond Indices from January 1973. Using this database we select
all bonds with fixed exercise date put option (i.e., not event contingent), issued
between January 1973 and December 1996, excluding issues by public utilities,
multilateral agencies (such as the World Bank), and sovereign issues. In our sample
we keep only those issues for which we find announcement dates in the Dow Jones
News Retrieval Service. This resulted in a sample of 158 bonds of which the earliest
bond is issued in 1979. Our final sample of bonds is selected according to the
following criteria:
(a) The issuing company’s stock returns are available on the CRSP tapes. For
20 companies CRSP data are not available, resulting in 138 issues. To reduce
confounding effects, all repeat issues by the same company within a year of
a sample issue are eliminated. We also eliminated observations for which we
find other contemporaneous corporate events. This further reduced our sample
to 104 issues.
(b) Furthermore, we eliminate from the sample companies that do not have suffi-
cient accounting data, credit rating, issue size, and industry code in either the
Compustat tapes, Moody’s Industrial Manuals, or the Warga’s Fixed Income
database. This reduces our sample size by 13 firms.
(c) We eliminate one more company for which we found no I/B/E/S data, thus
yielding a final sample of 90 firms.
The list of these firms and the characteristics of these bonds are reported in
Appendix 1. The maturity of these put bonds ranges from 5 to 100 years, with an
average initial maturity of 24 years. The put exercise dates ranges from 2 to
30 years, with an average put expiration period of 7 years.13

13
Four bonds have multiple put exercise dates.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 157

For the sake of comparison, we also construct a sample poison put bonds.14 The
issue announcements of poison put bonds are taken from Dow Jones Interactive.
We also searched LexisNexis but did not find any new announcements.15 This
search resulted in 67 observations. Our final sample of bonds with poison put
feature is further refined using criteria (a), (b), and (c) described above. Criterion
(a) reduces the sample size to 57 observations, and applying criterion (b) further
reduces it to 47 observations. The list of these firms and the characteristics of these
bonds are reported in the Appendix 2.
For these issues, we collect CRSP daily returns for our event study.16 We
calculate the abnormal returns (AR) on equity for each firm around the date of the
issue announcement, assuming that the stochastic process of returns is generated
by the market model. We define the announcement date to be the earlier of the
date on which the bond is issued and the date on which the issue information
appears in the news wires, or published in the Wall Street Journal, as depicted by
the Dow Jones Interactive. Our final sample includes only bonds whose issue
announcement explicitly mentions the inclusion of a put option. We estimate the
market model coefficients using the time period that begins 200 trading days
before and ends 31 trading days before the event, employing the CRSP value-
weighted market index as the benchmark portfolio. We use these coefficients to
estimate abnormal returns for days
30 to +30. We calculate the t-statistics
for the significance of the abnormal and cumulative abnormal returns using
the methodology employed by Mikkelson and Partch (1986). See Appendix 3
for details.
We examine the abnormal returns and their determinants for the sample of
issuers of bonds with European put options. First, we test whether issuing these
bonds significantly affects equity values. The stock price should, on average, react
positively to the issue announcement if either reducing debt security mispricing due
to asymmetric information or mitigating agency costs is a major motivation for
issuing putable bonds. On the other hand, the stock price should, on average, react
negatively to the issue announcement if management myopia (i.e., the relatively
low coupon rate compared to straight or callable debt issues) is a major motivation
of management in issuing putable bonds and if costs of financial distress are
significant.
Second, we estimate the following cross-sectional regression equation that
relates the cumulative abnormal return of the issuing firm’s equity to firm

14
Unlike the European put bond sample which we were able to obtain from Warga’s Fixed Income
Database, we obtained our sample of poison puts from Dow Jones Interactive by using keywords
such as poison put, event risk, and covenant ranking. Warga’s database does not include an
identifiable sample of poison put bonds.
15
We were only able to find a sample of poison put bonds with issue announcements using Dow
Jones Interactive and LexisNexis for the period between 1986 until 1991. We did use these news
services for the periods between 1979 and 2000.
16
The event study methodology was pioneered by Fama et al. (1969).
158 I.E. Brick et al.

characteristics that proxy for agency costs, asymmetric information problems, and
managerial myopia:

CAR3i ¼ b0 þ b1 FCFi þ b2 RISK i þ b3 SIZEi þ b4 INTSAVEDi

þ b5 ANALYSTSi þ b6 FINSi þ ei, (5.1)

where CAR3i is the 3-day (i.e., t ¼
1, 1) cumulative abnormal return for firm I.
The variables, FCF, RISK, SIZE, INTSAVED, and ANALYSTS proxy for agency
costs, information asymmetry, or management myopia. FINS is a dummy variable
that indicates whether the company is a financial institution.
FCF is the level of free cash flow of the firm for the fiscal year prior to the issue
announcement of putable bonds. We construct two alternative measures of FCF
which are similar to the definition employed by Lehn and Poulsen (1989)
and Bae et al. (1994). FCF1 is defined as [Earnings before Interest and
Taxes – Taxes – Interest Expense – Preferred Dividend Payments – Common
Stock Dividend Payments]/Total Assets. FCF2 is defined as [Earnings before
Interest and Taxes – Taxes – Interest Expense]/Total Assets.17 The inclusion of
a put option in the bond contract should help restrain management from misusing its
cash flow resources. We therefore expect b1 to be positive if reducing agency cost is
a major motivation for issuing putable bonds.
RISK is a dummy variable that equals 1 if the putable bond is rated by Standard
and Poor’s as BBB+ or below and zero otherwise.18 The impact of issuing putable
bonds on equity value may be associated with the variable RISK because of two
alternative hypotheses. First, the level of agency costs due to the asset substitution
problem is positively related to the probability of financial distress. We use this
dummy variable as a proxy for the firm’s ability and incentive to shift risk.
According to the agency cost motivation, we expect b2 to be positive. Second,
the greater the probability of financial distress, the more likely are bondholders to
exercise their put option and force the firm to prematurely retire the debt at a time
that is most inconvenient to the firm. Thus, according to the management myopia
hypothesis, we expect b2 to be negative.
SIZE is the natural logarithm of the total asset level of the issuing firm at the end
of the fiscal year preceding the issue announcement date. SIZE may have two
contrasting impacts on CAR3. First, SIZE may be interpreted as a proxy for the
level of asymmetric information. We expect that the degree of asymmetric infor-
mation is inversely related to SIZE. Hence, according to the debt security
mispricing motivation, we expect the putable bond issue announcement to have
a larger positive impact on small firms than on large firms. Consequently, we expect

17
Note that FCF1 is the actual free cash flow of the firm while FCF2 is the (maximum) potential
free cash flow if dividends are not paid. We do not subtract capital expenditures from the free cash
flow variables because they maybe discretionary and maybe allocated suboptimally to satisfy
management’s interests.
18
Had we assigned the value of 1 to only junk bonds, the number of observations with the risk
variable equal to one would be too small for statistical inference.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 159

b3 to be negative. On the other hand, we also expect that the risk of default is
inversely related to firm size. The higher the probability of default, the lower the
probability that the firm will be in existence and be able to pay its obligations
(including the par value of the bond if the put option is exercised) at the exercise
date of these European put options. Thus, size may be an indirect proxy of the
aggregate value of the put option. In this case, we expect b3 to be positive.
INTSAVED is the scaled (per $100 of assets for the fiscal year prior to the issue
announcement date) annual reduction in interest expense due to the incorporation of
a European put option in the bond contract. The annual interest expense reduction is
calculated as the product of the dollar amount of the putable bond issue and the
yield difference between straight and putable bonds. These yield differences are
calculated by subtracting the yields to maturity of the putable bond from the yield to
maturity of an equivalent non-putable bond, also taken from Warga’s Fixed Income
Database. The equivalent bond has similar maturity, credit rating, and call feature
as the putable bond. The equivalent non-putable bond is selected from the issuing
firm if, at the issue date of the putable debt, the issuing firm has an appropriate
(in terms of maturity, credit rating, and call feature) outstanding non-putable bond.
Otherwise, we choose an appropriate equivalent bond from a firm of the same
industry code as given by the Warga’s Fixed Income database. We posit that value
of the embedded put option is directly related to the (present) value of INTSAVED.
If the benefits of mitigating either agency costs or security mispricing are a major
reason for issuing European put bonds, and if these benefits are directly related to
the value of the embedded option, then b4 should be positive. In contrast, if
managerial myopia is a major reason for issuing European put bonds, and if the
expected reduction in equity value due to the cost of financial distress is related to
the value of the option, then b4 should be negative. The b4 coefficient may be
negative also because it proxies for a loss of interest tax shield due issuing putable
rather than straight bonds.
The variable ANALYSTS, defined as the natural logarithm of one plus the number
of analysts who follow the issuing firm for the quarter prior to the putable bond
issue announcement date, is another proxy for the degree of asymmetric informa-
tion.19 We obtain the number of analysts following each firm, for the year prior to
the bond announcement, from the I/B/E/S tapes. We hypothesize that the degree of
asymmetric information is negatively related to ANALYSTS. Thus, if asymmetric
information motivation is a major motivation for issuing putable debt, we expect
b5 to be negative.
We note that the European put and the poison put samples vary in their
proportion of financial service companies. The European put sample contains
25 (out of 90) financial service companies, while the poison put contains only
one. To control for the potential sector impact, we introduce the dummy variable
FINS which equals one if the company is a financial institution and zero otherwise

19
An alternative measure of the degree of asymmetric information is the number of analysts. The
empirical results are robust to this alternative measure.
160 I.E. Brick et al.

Table 5.1 A summary of expected signs of abnormal returns and their determinants for issuers of
bonds with European put options
Agency cost Security mispricing Managerial myopia
Abnormal returns Positive Positive Negative
FCF Positive No prediction No prediction
RISK Positive No prediction Negative
SIZE No prediction Ambiguous No prediction
INTSAVED Positive Positive Negative
ANALYSTS No prediction Negative No prediction
The determinants of abnormal returns are: RISK is a dummy variable equal to one if the bond issue
has an S&P bond rating of BBB+ or below and zero otherwise. FCF is defined in two separate
ways. FCF1 and FCF2 are the two free cash flow measures as described in the text. In particular,
FCF1 is defined as [Earnings before Interest and Taxes – Taxes – Interest Expense – Preferred
Dividend Payments – Common Stock Dividend Payments]/Total Assets. FCF2 is defined as
[Earnings before Interest and Taxes – Taxes – Interest Expense]/Total Assets. INTSAVED mea-
sures the relative amount of aggregate interest expense saved per $1,000 of total assets of the
issuing firm. ANALYSTS is the natural logarithm of one plus the number of analysts who follow the
issuing firm for the year prior to the putable bond issue announcement date. Those predictions that
are confirmed by our empirical study are in bold letters, and those that are significantly different
from zero are also underlined

for our regression analysis of the European put sample.20 Table 5.1 provides
a summary of the expected signs of the regression coefficients for the sample of
European put bonds.
We estimate our regressions using the general method of moments (GMM)
procedure. See Appendix 4. The procedure yields unbiased White t-statistics
estimates that are robust to heteroscedasticity.21 Note that because our instruments
are the regressors themselves, the parameter estimates from OLS and GMM are
identical. This is also discussed in Appendix 3.
We test the robustness and appropriateness of our specification by estimating
alternative specifications that include additional variables. First, we include
a dummy variable for the existence of a call feature to take into account the
effectiveness of the call feature to mitigate agency costs. Second, we include
a dummy variable to indicate that the bond’s put option expires within 5 years
because effectiveness of the put bond in mitigating agency costs may be related to

20
In the next section, we report the regression results when we exclude financial service companies
from our European put sample. Essentially, the basic results of our paper are not affected by the
inclusion or exclusion of financial services company.
21
The violation of the homoscedastic assumption for OLS does not lead to biased regression
coefficients estimators but potentially biases the computed t-statistic. The GMM procedure provides
an asymptotically unbiased estimation of the t-statistics without specifying the heteroscedastic
structure of the regression equation. The t-statistics obtained using GMM are identical to those
obtained using ordinary least squares (OLS) in the absence of heteroscedasticity.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 161

time to expiration of the put option. Third, we also include interaction variables
between these dummy variables and FCF, SPRISK, INTSAVE, and ANALYSTS.
In alternative specifications, we include interest rate volatility measures to take
into account the potential sensitivity of the value of the put option to interest rate
volatility. It should be noted that the value of the put option depends on the
volatility of the corporate bond yield. This volatility may be due to factors specific
to the company (such as corporate mismanagement and asymmetric information
about firm’s value) and macroeconomic factors relating to the stability and level of
market-wide default-free interest rates. Our independent variables in Eqs. 5.1 and
5.2 control for firm-specific yield volatility. To incorporate the possible impact of
market-wide interest rate volatility, we include several alternative variables. We
measure interest rate volatility as the standard deviation of the monthly 5-year
Fama-Bliss discount rate from CRSP. The standard deviation is alternatively
measured during a period of 60 months and 24 months immediately prior and
after the announcement date. Alternatively, we use the level of the Fama-Bliss
discount rate as a proxy for interest rate volatility.
We repeat the analysis for the sample of issuers of poison put bonds. As with the
sample of bonds with an embedded European put option, we first examine the
abnormal returns and their determinants for the sample of issuers of poison putable
bonds. In essence, we test whether issuing these bonds significantly affects equity
value. The stock price should, on average, react positively to the issue announce-
ment if either mitigating agency costs or reducing debt security mispricing due to
asymmetric information is a major motivation for issuing poison bonds. On the
other hand, the stock price should, on average, react negatively to the issue
announcement if management myopia (i.e., the relatively low coupon rate com-
pared to straight or callable debt issues) or management entrenchment is a major
motivation of management in issuing putable bonds.
We use two alternative specifications for the cross-sectional study that relates
the abnormal returns of the poison putable bond sample to firm characteristics.
The first is described by Eq. 5.1 The second includes a variable, COVRANK, that
equals the S&P event risk ranking on a scale of 1–5. S&P event risk ranking of one
implies that the embedded put option provides the most protection to bondholders
against credit downgrade events. Event risk ranking of five offers the least
protection to bondholders against credit downgrade events. Thus, the second
specification is:

CAR3i ¼ b0 þ b1 FCFi þ b2 RISK i þ b3 SIZEi þ b4 INTSAVEDi

þ b5 ANALYSTSi þ b6 COVRANK i þ ei: (5.2)

If mitigating agency costs was a major motivation for issuing poison put bonds
as suggested by Bae et al. (1994), then we would expect regression coefficients
b1, b2, and b4 to be positive. We also expect b6 to be negative because
COVRANK is negatively related to extent of bondholder protection, and this
protection should deter management from asset substitution activities.
162 I.E. Brick et al.

In contrast, if, as Cook and Easterwood (1994) and Roth and McDonald (1999)
argue, management entrenchment is a major motivation behind for issuing
putable bonds, then we expect b1 to be negative since we expect the potential
loss of value due to management entrenchment to be positively related to free
cash flow and b6 to be positive.22

5.5 Empirical Results

This section presents the empirical results for tests of our hypotheses discussed
above. Panel A of Table 5.2 provides summary statistics of issuers of bonds with
a European put option. The average annual sales, long-term debt, and total assets for
the fiscal year prior to the issue announcement date are $14.66 billion, $4.21 billion,
and $30.79 billion, respectively. The mean leverage ratio, defined as the long-term
debt to total assets, is 18.6 %. The average issue size of the putable bond is $190
million and represents, on average, 2.7 % of the firm’s assets. As measured by
RISK, 24 % of putable bonds are rated below A-. The average free cash flow of the
firm as a percentage of the firm’s total assets is below 4 %. The average amount of
interest saved (INTSAVED) due to the inclusion of a put option feature in the bond
issue is approximately 0.03 % of the total assets of the issuing firm. The maximum
interest expense saved is as high as $1.23 per $100 of total assets.23 However,
INTSAVED is negative for ten observations, which may be due to the lack of
a closely matched straight bond. To guard against this error-in-variable problem,
we estimate our regression Eq. 5.1 using two alternative samples: the entire sample
and a sample comprising of positive INTSAVED observations. The number of
analysts following a company ranges from 1 to 41. Panel B of Table 5.2 provides
the corresponding summary statistics for the sample of issuers of poison put bonds.
We note that these firms are smaller than the issuers of bonds with European put
options. Additionally, issuers of poison puts have a smaller number of following
analysts, and the bonds tend to be somewhat riskier than the issuers of bonds with
European put options.
Panel A of Table 5.3 presents the average daily abnormal performance of the
equity of issuers of bonds with European put options in our sample for t ¼
30 to
t ¼ 30. These abnormal returns are obtained from a market model. Please note that
the t-statistics presented in Table 5.3 are based on standardized abnormal returns. In
an efficient market, we expect the market to impound the economic informational
impact of the new bond issue on the day of the announcement (t ¼ 0). The average
abnormal return at t ¼ 0 is almost 0.33 % and is significantly positive at the 1 % level.

22
If a major motivation for issuing putable bonds is to enhance management entrenchment, then
we expect that the value of entrenchment is directly related to the level of free cash flow which
management can misappropriate.
23
Crabbe (1991) demonstrates that bonds with poison puts reduce the cost of borrowing for firms.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 163

Table 5.2 Summary statistics of putable bond issuing firms

Variable Mean Std dev Minimum Maximum
Panel A: Sample of 90 issuers of bonds with European put options
Sales 14,656.250 23,238.860 89.792 124,993.900
Long-term debt 4,206.870 7,568.520 51.522 50,218.300
Total assets 30,789.070 41,623.450 182.281 174,429.410
Leverage ratio 0.186 0.122 0.012 0.457
Issue size 190.162 106.701 14.300 500.000
Ebit 1,818.390 2,647.810 10.457 13,275.700
ISSUE 0.027 0.056 0.001 0.439
RISK 0.244 0.432 0.000 1.000
FCF1 0.020 0.093
0.082 0.823
FCF2 0.034 0.095
0.080 0.824
Number of analysts 21.922 8.388 1.000 41.000
INTSAVED 0.029 0.130
0.028 1.226
Panel B: Sample of 47 issuers of bonds with poison put options
Sales 5,033.849 5,921.869 406.360 34,922.000
Long-term debt 1,203.060 2,148.865 25.707 13,966.000
Total assets 5,080.781 7,785.288 356.391 51,038.000
Leverage ratio 0.208 0.112 0.031 0.476
Issue size 158.114 80.176 50.000 350.000
Ebit 481.121 639.842
30.800 3,825.000
ISSUE 0.062 0.048 0.006 0.251
RISK 0.314 0.469 0.000 1.000
FCF1 0.034 0.032
0.075 0.111
FCF2 0.051 0.036
0.050 0.132
Number of analysts 18.803 7.792 2.000 39.000
INTSAVED 0.011 0.044
0.070 0.216
Sales, long-term debt, total assets, and EBIT are in millions of dollars and are for the fiscal year
prior to the putable bond issue announcement. Leverage ratio is the ratio of the long-term debt to
total assets. Issue size is the dollar amount of putable bond issue in millions of dollars. ISSUE is
the ratio of issue size to the total assets of the issuing firm as of the fiscal year prior to the issue
announcement date. RISK is a dummy variable equal to one if the bond issue has an S&P bond
rating of BBB+ or below and zero otherwise. FCF1 and FCF2 are the two free cash flow measures
as described in the text. In particular, FCF1 is defined as [Earnings before Interest and
Taxes – Taxes – Interest Expense – Preferred Dividend Payments – Common Stock Dividend
Payments]/Total Assets. FCF2 is defined as [Earnings before Interest and Taxes – Taxes – Interest
Expense]/Total Assets. Number of analysts is the number of analysts who follow the issuing firms
for the year prior to the putable bond issue. INTSAVED measures the relative amount of aggregate
interest expense saved per $100 of total assets of the issuing firm

This suggests that the market views favorably the announcement of putable bonds.
This may be due to mitigating agency costs or resolving asymmetric information.
However, for the 3-day window (t ¼
1, 1), the abnormal return is 0.19 % but is not
statistically significant from zero.
164 I.E. Brick et al.

Table 5.3 The average daily abnormal returns for firms issuing putable bonds from 30 days prior
to the putable bond issuance announcement to 30 days after the announcement. The t-statistics are
based on standardized abnormal returns. CAR is the cumulative abnormal return
Event day Abnormal return t-statistic CAR
Panel A: Sample of 90 issuers of bonds with European put options

30 0.0006 0.1988 0.0006

29 0.0004 0.2305 0.0010

28 0.0011 0.8169 0.0021

27
0.0003
0.1248 0.0017

26 0.0002
0.2028 0.0019

25 0.0003 0.0263 0.0022

24 0.0003 0.2174 0.0025

23
0.0002
0.1306 0.0023
22 0.0007 0.2573 0.0030

21
0.0003
0.0367 0.0026

20
0.0024
1.6700 0.0003

19 0.0000 0.1348 0.0002

18
0.0025
1.6932
0.0022

17 0.0029 1.7141 0.0007

16
0.0005 0.0154 0.0002

15 0.0022 1.4378 0.0024

14 0.0022 1.4701 0.0046

13
0.0002
0.1052 0.0043

12 0.0007 0.4783 0.0050

11
0.0014
1.2803 0.0036

10 0.0003 0.1904 0.0039

9
0.0014
0.5100 0.0025

8
0.0006
0.6052 0.0019

7 0.0001
0.4020 0.0021

6 0.0005 0.1268 0.0026

5 0.0027 1.3209 0.0053

4
0.0001
0.0050 0.0052

3 0.0020 1.4407 0.0071

2
0.0028
1.8590 0.0043

1
0.0021
1.4323 0.0022
0 0.0033 2.6396 0.0055
1 0.0007 0.3534 0.0061
2 0.0007 0.3177 0.0068
3 0.0002 0.2628 0.0070
4
0.0010
0.7364 0.0060
5 0.0000
0.2821 0.0059
6
0.0036
2.4342 0.0023
7
0.0002
0.4355 0.0022
8
0.0004
0.4253 0.0018
(continued)
5 Motivations for Issuing Putable Debt: An Empirical Analysis 165

Table 5.3 (continued)

Event day Abnormal return t-statistic CAR
9
0.0007
0.5236 0.0011
10 0.0010 1.0509 0.0021
11 0.0039 2.8180 0.0060
12 0.0010 0.7853 0.0070
13
0.0002
0.0036 0.0068
14
0.0013
0.7663 0.0055
15 0.0006
0.0920 0.0061
16 0.0014 1.2984 0.0075
17 0.0002 0.0847 0.0077
18
0.0001
0.2386 0.0076
19 0.0002
0.3055 0.0078
20 0.0005 0.3692 0.0083
21
0.0006
0.0998 0.0077
22
0.0016
1.1099 0.0061
23
0.0014
0.8924 0.0047
24 0.0015 1.1956 0.0061
25 0.0001
0.3050 0.0062
26
0.0001 0.0257 0.0061
27 0.0004 0.4372 0.0065
28
0.0011
0.4242 0.0054
29 0.0004 0.8785 0.0058
30
0.0010
0.5929 0.0048
Panel B: Sample of 47 issuers of bonds with poison put options

30
0.0002
0.2385
0.0002

29
0.0029
1.7698
0.0032

28
0.0005
0.1247
0.0036

27 0.0015 0.2635
0.0021

26
0.0010
0.4019
0.0031

25 0.0013 0.6437
0.0018

24 0.0003
0.1287
0.0015

23
0.0007
0.3447
0.0022

22
0.0021
1.3803
0.0044

21 0.0023 0.5731
0.0020

20
0.0028
1.2905
0.0048

19
0.0009 0.1682
0.0058

18
0.0024
1.1787
0.0081

17 0.0034 1.4394
0.0048

16 0.0001
0.0490
0.0047

15
0.0002
0.2326
0.0049

14 0.0035 1.8138
0.0014

13 0.0010 0.7301
0.0005

12
0.0026
1.9908
0.0031
(continued)
166 I.E. Brick et al.

Table 5.3 (continued)

Event day Abnormal return t-statistic CAR

11 0.0002
0.0182
0.0029

10 0.0026 1.3253
0.0003

9
0.0034
1.5412
0.0038

8
0.0002 0.2221
0.0040

7 0.0003 0.0028
0.0037

6 0.0019 1.0805
0.0018

5
0.0009
0.7397
0.0026

4
0.0006
0.5318
0.0032

3
0.0001
0.2118
0.0033

2
0.0003
0.0870
0.0036

1
0.0028
1.5751
0.0064
0
0.0028
1.3777
0.0092
1 0.0004
0.5639
0.0088
2 0.0019 0.8618
0.0069
3 0.0027 1.4826
0.0041
4 0.0022 1.4571
0.0019
5 0.0034 1.2458 0.0014
6
0.0006
0.5491 0.0009
7
0.0002
0.1866 0.0006
8
0.0036
2.2457
0.0030
9 0.0026 1.4036
0.0003
10
0.0017
1.0479
0.0020
11
0.0016
1.1626
0.0036
12
0.0008
0.7991
0.0043
13
0.0020
1.0808
0.0064
14
0.0004
0.4107
0.0068
15 0.0002
0.2023
0.0067
16
0.0005
0.2077
0.0072
17
0.0003
0.3234
0.0075
18 0.0005 0.5278
0.0070
19
0.0018
0.9564
0.0087
20 0.0014 0.7079
0.0073
21
0.0022
1.0614
0.0095
22
0.0015
0.4152
0.0110
23
0.0004
0.2990
0.0114
24
0.0015
0.7324
0.0129
25
0.0006
0.2189
0.0134
26 0.0004 0.4731
0.0130
27 0.0015 0.2512
0.0115
28 0.0014 0.2703
0.0102
29
0.0022
1.1175
0.0124
30
0.0002 0.2668
0.0126
 5

Abnormal Returns Around Annoucement of European and Poiosn Put Bonds

0.0100

0.0050

0.0000

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

−9
−8
−7
−6
−5
−4
−3
−2
−1

−30
−29
−28
−27
−26
−25
−24
−23
−22
−21
−20
−19
−18
−17
−16
−15
−14
−13
−12
−11
−10

CAR
−0.0050

−0.0100
Motivations for Issuing Putable Debt: An Empirical Analysis

−0.0150
Event Date

CAR (European puts) CAR (Poison puts)

Fig. 5.1 Cumulative abnormal returns from t ¼
30 to t ¼ +30 around the issue announcement of bonds with European or poison put features
167
168 I.E. Brick et al.

Table 5.4 Cross-sectional regression results for the European put sample: base case
Sample of 80 bonds with positive
Full sample of 90 firms INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Intercept
0.0388
1.69 0.095
0.03657
1.58 0.1193
FCF1
0.0016
0.09 0.9289
0.00392
0.22 0.8277
RISK
0.0016
0.25 0.8026
0.00245
0.37 0.7117
SIZE 0.0097 2.93 0.0044 0.010637 3.03 0.0034
INTSAVED 0.0269 2.59 0.0115 0.023819 2.27 0.0259
ANALYSTS
0.0153
2.48 0.0152
0.01848
2.82 0.0062
FINS
0.0203
2.79 0.0065
0.02062
2.75 0.0075
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6¼0 26.38 0.0002 29.19 0.000
Adj. R2 0.1185 0.1314
Coef. t-ratio p-value Coef. t-ratio p-value
Intercept
0.0390
1.71 0.0915
0.03642
1.58 0.1191
FCF2
0.0004
0.02 0.9808
0.00392
0.21 0.8309
RISK
0.0016
0.24 0.8075
0.00247
0.37 0.7094
SIZE 0.0097 2.93 0.0043 0.010633 3.03 0.0034
INTSAVED 0.0270 2.59 0.0112 0.023751 2.26 0.0266
ANALYSTS
0.0152
2.47 0.0157
0.01849
2.8 0.0065
FINS
0.0202
2.73 0.0078
0.02066
2.7 0.0085
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6¼0 26.73 0.0002 29.25 0.000
Adj. R2 0.1184 0.1314
This table reports the regression coefficients and their t-statistics of the following regression
equation:
CAR3i ¼ b0 + b1FCFi + b2RISKi + b3SIZEi + b4INTSAVEDi + b5ANALYSTSi + b6FINS + ei
CAR3 is the cumulative abnormal return measured from the day before the announced issue to the
day after the announced issue; FCF1 and FCF2 are the two free cash flow measures as described in
the text; RISK is a dummy variable equal to one if the bond issue has an S&P bond rating of BBB+
or below and zero otherwise; SIZE is the natural logarithm of the total assets of the issuing firm at
the end of the fiscal year prior to the issue announcement; INTSAVED measures the relative
amount of aggregate interest expense saved per $100 of total assets of the issuing firm; ANALYSTS
is the natural logarithm of one plus the number of analysts following the firm; and FINS is equal to
one if the parent company is a financial institution and is equal to zero otherwise. The p-values
assume a two-tail test. All the t-ratios are heteroscedasticity consistent

Panel B of Table 5.3 presents the corresponding results for the issuers of poison put
bonds. Note that the abnormal return at t ¼ 0 is
0.28 % with a t-statistic of
1.38.
Furthermore, during the 3-day (i.e., t ¼
1, 1) window, the abnormal return is
negative (
0.52 %) and is significantly different than zero (t-statistic of
2.03).
This negative abnormal return is consistent with the view that poison put bonds help
entrench the current management. The cumulative abnormal returns of the two sample
types of embedded put option are also depicted in Fig. 5.1. The patterns of CARs
5 Motivations for Issuing Putable Debt: An Empirical Analysis 169

around the issue announcement dates clearly show the positive trend of CARs
for bonds with European puts versus the negative trend of CARs for bonds with
poison puts.
Next, to test our hypotheses, we examine what determines the cross-sectional
variations in the CARs for the sample of European and poison put bonds. Table 5.4
reports the parameter and t-statistic estimates for regression Eq. 5.1 for the sample
of issuers of bonds with European put options. The four regressions reported in
Table 5.4 differ by the sample that is used and by the specification used for the free
cash flow variable. The regression in the top left panel uses the entire sample of
90 observations and the variable FCF1. The regression in the top right panel reports
the results using only the 80 observations in which the variable INTSAVED is
positive. The regressions reported in the bottom panels use FCF2 instead of FCF1.
First note that our estimates are robust to these alternative specifications. The
Wald test of the hypothesis that all five independent variables are jointly zero is
rejected at the 1 % significance level for all four regression specifications. The
regression coefficients on the ANALYSTS variable, b5, are negative and significantly
different from zero for all four regressions. These estimates are consistent with the
reduction in security mispricing motivation for issuing putable bonds. In contrast,
the regression coefficients for FCF and RISK are not significantly different from
zero, indicating a lack of empirical support for the mitigating agency cost hypoth-
esis. The regression coefficients for the INTSAVED variables are positive and are
significantly different from zero in all four regressions. This result is again consis-
tent with the security mispricing motivation and inconsistent with the management
myopia motivation. Additionally, the regression coefficient b3 for the size variable
is also positive and significantly different from zero, a result more consistent with
size being related to the probability of survivorship of the firm. In summary,
announcement of bonds with European puts is associated with positive abnormal
returns which are related to our proxies for potential benefits from mitigating
security mispricing.
Panel A of Table 5.5 presents the corresponding estimates for the sample of
issuers of poison put bonds, while panel B of Table 5.5 presents the estimates of
Eq. 5.2 for the same sample. The estimates reported in Table 5.5 are different from
those in Table 5.4. The coefficients of the ANALYSTS variable, b5, are significantly
positive for the full sample. The coefficients for Size are significantly negative for
the full sample. All the other variables are not significantly different than zero. The
estimates of Eq. 5.2 indicate that the coefficients of the variable COVRANK are
negative and significantly different from 0.
Our results so far indicate that the equity abnormal returns around the issue
announcement dates of poison put bonds are negative, consistent with the mana-
gerial entrenchment evidence in Cook and Easterwood (1994) and Roth and
McDonald (1999). Additionally, the positive and significant coefficient of the
ANALYSTS variable may also be consistent with the management entrenchment
hypothesis because it appears that the market negative response to the issuance of
poison putable bonds arises from the less followed firms, where the management
170 I.E. Brick et al.

Table 5.5 Cross-sectional regression results for the poison put sample
Sample of 28 bonds with
Full sample of 47 firms positive INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Panel A: Sample of issuers of bonds with poison put options
Intercept 0.015 0.530 0.602 0.030 0.680 0.506
FCF1
0.170
1.330 0.191
0.082
0.510 0.617
RISK 0.010 1.550 0.129 0.001 0.090 0.932
SIZE
0.007
2.250 0.030
0.006
1.090 0.288
INTSAVED 0.022 0.370 0.714
0.024
0.280 0.782
ANALYSTS 0.014 2.100 0.042 0.006 0.370 0.717
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼0 11.927 0.036 1.906 0.861
Adj. R2 0.015
0.181
Intercept 0.018 0.730 0.472 0.026 0.640 0.532
FCF2
0.153
1.280 0.208
0.045
0.270 0.788
RISK 0.009 1.400 0.168 0.000 0.030 0.978
SIZE
0.007
2.480 0.017
0.005
1.020 0.318
INTSAVED 0.027 0.460 0.649
0.016
0.180 0.858
ANALYSTS 0.014 1.980 0.054 0.006 0.330 0.744
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼0 11.538 0.042 1.857 0.868
Adj. R2 0.007
0.192
Sample of 23 bonds with
Full sample of 40 firms positive INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Panel B: Sample of issuers of bonds with poison put options with COVRANK
Intercept 0.052 1.200 0.239 0.136 2.550 0.021
FCF1
0.158
1.100 0.281
0.271
1.330 0.201
RISK 0.011 1.140 0.264 0.006 0.370 0.720
SIZE
0.008
1.800 0.081
0.014
2.370 0.031
INTSAVED 0.029 0.430 0.670
0.071
0.720 0.484
COVRANK
0.008
3.900 0.000
0.012
3.520 0.003
ANALYSTS 0.013 2.090 0.045 0.009 0.700 0.497
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6¼0 48.723 0.000 20.513 0.002
Adj. R2 0.114 0.058
Intercept 0.053 1.440 0.160 0.117 3.030 0.008
FCF2
0.132
1.000 0.327
0.159
0.890 0.386
RISK 0.010 1.110 0.275 0.000 0.010 0.995
SIZE
0.009
2.030 0.051
0.013
2.590 0.020
INTSAVED 0.032 0.510 0.616
0.030
0.240 0.810
COVRANK
0.008
4.100 0.000
0.010
3.610 0.002
(continued)
5 Motivations for Issuing Putable Debt: An Empirical Analysis 171

Table 5.5 (continued)

Sample of 23 bonds with
Full sample of 40 firms positive INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Panel B: Sample of issuers of bonds with poison put options with COVRANK
ANALYSTS 0.014 1.970 0.057 0.010 0.710 0.488
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6 ¼ 0 53.590 0.000 31.077 0.000
Adj. R2 0.102
0.006
Panels A reports the regression coefficients and their t-statistics of the following regression
equation:
CAR3i ¼ b0 + b1FCFi + b2RISKi + b3SIZEi + b4INTSAVEDi + b5ANALYSTSi + ei
Panel B reports the regression coefficients the regression coefficients and their t-statistics of the
following regression equation:
CAR3i ¼ b0 + b1FCFi + b2RISKi + b3SIZEi + b4INTSAVEDi + b5ANALYSTSi + b7COVRANKi + ei
CAR3 is the cumulative abnormal return measured from the day before the announced issue
to the day after the announced issue; FCF1 and FCF2 are the two free cash flow measures
as described in the text; RISK is a dummy variable equal to one if the bond issue has a S&P
bond rating of BBB+ or below, and zero otherwise; SIZE is the natural logarithm of the total
assets of the issuing firm at the end of the fiscal year prior to the issue announcement;
INTSAVED measures the relative amount of aggregate interest expense saved per $100
of total assets of the issuing firm; ANALYSTS is the natural logarithm of one plus the number
of analysts following the firm; and COVRANK equals the S&P Event Risk Ranking
on a scale of 1–5. The p-values assume a two-tail test. All the t-ratios are heteroscedasticity
consistent

strategy may not be as well known prior to the bond issuance. However, these
returns are negatively related to the event risk covenant ranking consistent with the
agency cost evidence in Bae et al. (1994). The negative abnormal returns experi-
enced by issuers of poison puts and the different coefficients on SIZE, INTSAVED,
and ANALYSTS indicate that the European put bonds are viewed differently than
poison put bonds.
In summary, the results reported in Tables 5.2–5.5 are consistent with the view
that European put bonds, where the put option does not depend on a specific
company event, are more effective in mitigating problems that are associated
with security mispricing. Furthermore, there is no empirical support for the hypoth-
esis that European put bonds are used as a vehicle for management entrenchment. In
contrast, the evidence for poison put bonds is consistent with both management
entrenchment and mitigating agency costs.
We now discuss several robustness checks of our basic results. Note that, as
reported in Table 5.4, the regression coefficient of the financial industry dummy
variable, b6, is negative and significantly different from zero. To verify that the
differing results between the European put bond sample (where 25 out of 90 firms
172 I.E. Brick et al.

Table 5.6 Cross-sectional regression results for the European put sample excluding financial
service companies
Sample of 61 bonds with positive
Full sample of 65 firms INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Intercept
0.0410
1.27 0.2087
0.0416
1.25 0.2182
FCF1 0.0005 0.04 0.9715
0.0016
0.10 0.9200
RISK
0.0028
0.41 0.6821
0.0033
0.49 0.6279
SIZE 0.0108 2.79 0.0071 0.0114 2.80 0.0071
INTSAVED 0.0248 1.85 0.0694 0.0242 1.79 0.0785
ANALYSTS
0.0178
2.54 0.0138
0.0191
2.70 0.0093
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼0 32.66 0.0000 32.10 0.0000
Adj. R2 0.1450 0.1524
Intercept
0.0409
1.28 0.2073
0.0414
1.25 0.2181
FCF2 0.0000 0.00 0.9992
0.0031
0.17 0.8676
RISK
0.0028
0.41 0.6805
0.0034
0.50 0.6206
SIZE 0.0108 2.78 0.0072 0.0114 2.79 0.0071
INTSAVED 0.0248 1.85 0.0686 0.0241 1.79 0.0784
ANALYSTS
0.0178
2.54 0.0137
0.0191
2.70 0.0091
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼0 32.66 0.0000 32.10 0.0000
Adj. R2 0.1450 0.1525
This table reports the regression coefficients and their t-statistics of the following regression
equation:
CAR3i ¼ b0 + b1FCFi + b2RISKi + b3SIZEi + b4INTSAVEDi + b5ANALYSTSi + ei
CAR3 is the cumulative abnormal return measured from the day before the announced issue to the
day after the announced issue; FCF1 and FCF2 are the two free cash flow measures as described in
the text; RISK is a dummy variable equal to one if the bond issue has a S&P bond rating of BBB+
or below, and zero otherwise; SIZE is the natural logarithm of the total assets of the issuing firm at
the end of the fiscal year prior to the issue announcement; INTSAVED measures the relative
amount of aggregate interest expense saved per $100 of total assets of the issuing firm; ANALYSTS
is the natural logarithm of one plus the number of analysts following the firm. The p-values assume
a two-tail test. All the t-ratios are heteroscedasticity consistent

are from the financial service sector) and the poison put sample (where only one
firm is from the financial service sector), we replicate the regressions of Tables 5.4
and 5.5 excluding all financial service companies. The estimates for the subsample
of European put bonds are reported in Table 5.6. Note that the coefficients and
significance levels are very similar to those reported in Table 5.4. Because only one
poison put company is from the financial sector and because the estimates of the
corresponding subsample of poison put bonds are very similar to those reported in
Table 5.5, we do not report these estimates. Thus, we conclude that the differences
between poison and European put bonds are not due to the different sector compo-
sition of our samples.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 173

Table 5.7 The cross-sectional regression results for the European put sample: impact of time to
expiration
Sample of 80 bonds with
Full sample of 90 firms positive INTSAVED
Coef. t-ratio p-value Coef. t-ratio p-value
Intercept
0.0288
1.27 0.2063
0.0282
1.22 0.2279
FCF1 0.0037 0.24 0.8147 0.0007 0.05 0.9627
RISK
0.0010
0.16 0.8703
0.0021
0.32 0.7527
SIZE 0.0088 2.74 0.0075 0.0098 2.88 0.0052
INTSAVED 0.0143 1.35 0.1814 0.0131 1.22 0.2249
ANALYSTS
0.0165
2.68 0.0090
0.0193
2.95 0.0043
FINS
0.0185
2.60 0.0112
0.0194
2.61 0.0109
EXPLT5 0.0097 1.91 0.0601 0.0083 1.37 0.1739
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6¼ b7¼0 34.82 0.0000 37.88 0.000
Adj. R2 0.1413 0.1420
Intercept
0.0291
1.3 0.1981
0.0283
1.22 0.2248
FCF2 0.0051 0.33 0.7397 0.0010 0.06 0.9506
RISK
0.0010
0.15 0.8794
0.0021
0.31 0.7542
SIZE 0.0088 2.74 0.0075 0.0098 2.88 0.0052
INTSAVED 0.0144 1.36 0.1784 0.0132 1.22 0.2257
ANALYSTS
0.0164
2.67 0.0092
0.0193
2.93 0.0045
FINS
0.0184
2.53 0.0133
0.0194
2.57 0.0122
EXPLT5 0.0097 1.91 0.0593 0.0083 1.38 0.1733
w2 p-value w2 p-value
H0:b1¼b2¼b3¼b4¼b5¼b6¼b7¼0 34.85 0.0000 37.82 0.000
Adj. R2 0.1415 0.1420
This table reports the regression coefficients and their t-statistics of the following regression equation:
CAR3i ¼ b0 + b1FCFi + b2RISKi + b3SIZEi + b4INTSAVEDi + b5ANALYSTSi + b6FINS + b7EXPLT5 + ei
CAR3 is the cumulative abnormal return measured from the day before the announced issue to the
day after the announced issue; FCF1 and FCF2 are the two free cash flow measures as described in
the text; RISK is a dummy variable equal to one if the bond issue has a S&P bond rating of BBB+
or below and zero otherwise; SIZE is the natural logarithm of the total assets of the issuing firm at
the end of the fiscal year prior to the issue announcement; INTSAVED measures the relative
amount of aggregate interest expense saved per $100 of total assets of the issuing firm; ANALYSTS
is the natural logarithm of one plus the number of analysts following the firm; FINS is equal to one
if the parent company is a financial institution and is equal to zero otherwise; and EXPLT5 is equal
to one if the expiration of the embedded option is less than 5 years from the issue date and is equal
to zero otherwise. The p-values assume a two-tail test. All the t-ratios are heteroscedasticity
consistent

Next, we examine whether the term to expiration of European put bonds affects
our estimates.24 Table 5.7 repeats the regressions of Table 5.4 when we introduce
an additional explanatory dummy variable, EXPLT5, that is equal to one if the

24
Recall that by definition the option of poison put bonds does not have an expiration date.
174 I.E. Brick et al.

expiration date of the embedded European put option is less than 5 years from
the issue date and zero otherwise. We find that INTSAVE is still positive but is
no longer significant and the regression coefficient for EXPLT5 is positive and
is significantly different from zero at the 10 % level for the full sample.
The estimates and the significance of the other coefficients are largely
unaffected. We conclude that these results still confirm the security mispricing
hypothesis because the regression coefficients for ANALYSTS are significantly
negative.
Our results are robust to alternative specifications we described at the end of the
previous section. In particular, including a dummy variable for the existence of
a call feature, interaction variables and alternative measures of interest rate vola-
tility do not affect our results. These results are not reported here but are available
upon request.
Finally, we test for multicollinearity by examining the eigenvalues of
the correlation matrix for (non-dummy) independent variables. For orthogonal
data the eigenvalue, l, for each variable should equal 1, and S1/l ¼ the
number of regressors (i.e., five for our study). For our sample of bonds with an
embedded European put option, this sum is 7.14 when FCF1 is used, and
this sum ¼ 7.19 when FCF2 is used, indicating a lack of significant
multicollinearity. Similar results are obtained for the poison put sample. There-
fore, our findings are robust to alternate specifications and are not driven by
multicollinearity.

5.6 Concluding Remarks

This paper examines the motivations and equity valuation impact of issuing Euro-
pean putable bonds, a bond containing an embedded European put option held by
bondholders. The option entitles them to sell the bond back to the firm on the
exercise date at a predetermined price. Unlike a poison put bond which has been
studied by the literature, the exercise of the put option in a European putable bond is
not contingent upon a company-related event.
We find that the market reacts favorably to the issue announcement of such
putable bonds. We consider three alternative motivations for incorporating
a European put option in a bond contract: reducing the security mispricing
impact of asymmetric information, mitigating agency costs, and the relatively
low coupon rate (a myopic view that ignores the potential liability to the
firm due to the put option). We test these hypotheses by conducting a
cross-sectional empirical study of the impact of putable debt issue announce-
ments on the equity value of the issuing companies. Our results indicate that the
market favorably views putable bonds as a means to reduce security mispricing.
We also find that the market reaction to poison put announcements differs
from the market reaction to issue announcements for bonds with European
put options.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 175

Appendix 1: Sample of Firms Issuing Putable Bonds and the Put

Bond Characteristics

Put S&P
Amount Issue Maturity expiration YTM credit
Name Coupon ($000’s) date (years) (years) Callability (%) rating
Air Products 7.34 100,000 06/15/ 30 12 N 7.20 7
and Chemicals 1996
Inc
American 8.50 150,000 06/09/ 10 5 N 8.56 5
Express 1989
Anadarko 7.25 100,000 03/17/ 30 5 N 6.90 9
Petroleum 1995
Corp
Anadarko 7.73 100,000 09/19/ 100 30 N 7.21 9
Petroleum 1996
Corp
Bankamerica 7.65 150,000 04/26/ 10 2 C 12.23 10
Corp 1984
Bausch & 6.56 100,000 08/12/ 30 5 N 6.71 8
Lomb Inc 1996
Baxter 8.88 100,000 06/14/ 30 5 C 8.96 9
International 1988
Inc
Burlington 7.29 200,000 05/31/ 40 12 N 7.39 10
Northern 1996
Santa Fe
Champion 15.63 100,000 07/22/ 12 3 C 15.68 9
International 1982
Corp
Champion 6.40 200,000 02/15/ 30 10 N 6.69 10
International 1996
Corp
Chase 7.55 250,000 06/12/ 12 5 C 8.87 5
Manhattan 1985
Corp – old
Chrysler Corp 12.75 200,000 11/01/ 15 5 N 12.75 10
1984
Chrysler Corp 9.65 300,000 07/26/ 20 5 C 9.65 10
1988
Chrysler Corp 9.63 200,000 09/06/ 20 2 C 9.65 11
1988
Circus Circus 6.70 150,000 11/15/ 100 7 N 6.66 9
Enterprise Inc 1996
Citicorp 9.40 250,000 12/06/ 12 2 C 11.05 4
1983
Citicorp 10.25 300,000 12/19/ 10 2 C 10.31 4
1984
(continued)
176 I.E. Brick et al.

Put S&P
Amount Issue Maturity expiration YTM credit
Name Coupon ($000’s) date (years) (years) Callability (%) rating
Citicorp 8.75 250,000 12/12/ 15 3 C 8.84 4
1985
Coca-Cola 7.00 300,000 09/27/ 30 10 N 7.00 5
Enterprises 1996
Commercial 8.50 100,000 02/05/ 10 5 N 8.84 8
Credit 1988
Commercial 8.70 150,000 06/09/ 20 10 N 9.06 8
Credit 1989
Commercial 8.70 100,000 06/11/ 20 3 N 7.54 7
Credit 1990
Commercial 7.88 200,000 02/01/ 30 10 N 7.30 6
Credit 1995
Conagra Inc 7.13 400,000 10/02/ 30 10 N 6.95 9
1996
Corning Inc 7.63 100,000 07/31/ 30 10 N 7.81 6
1994
Deere & Co 8.95 199,000 06/08/ 30 10 C 8.99 7
1989
Diamond 7.65 100,000 06/25/ 30 10 N 7.37 10
Shamrock Inc 1996
Dow 8.48 150,000 08/22/ 30 14 C 9.35 7
Chemical 1985
Eastman 7.25 125,000 04/08/ 10 5 N 8.48 8
Kodak Co 1987
Eaton Corp 8.00 100,000 08/18/ 20 10 N 7.90 7
1986
Eaton Corp 8.88 38,000 06/14/ 30 15 N 9.00 7
1989
Eaton Corp 6.50 150,000 06/09/ 30 10 N 6.64 7
1995
Enron Corp 9.65 100,000 05/17/ 12 7 N 9.50 10
1989
First Chicago 8.50 98,932 05/13/ 12 7 N 8.70 7
Corp 1986
First Interstate 7.35 150,000 08/23/ 15 6 C 13.70 4
Bancorp 1984
First Interstate 9.70 100,000 07/08/ 15 5 C 9.91 4
Bancorp 1985
First Union 7.50 250,000 04/25/ 40 10 N 7.96 8
Corp (NC) 1995
First Union 6.82 300,000 08/01/ 30 10 N 7.43 8
Corp (NC) 1996
Ford Motor Co 7.50 250,000 10/30/ 15 3 N 9.63 6
1985
(continued)
5 Motivations for Issuing Putable Debt: An Empirical Analysis 177

Put S&P
Amount Issue Maturity expiration YTM credit
Name Coupon ($000’s) date (years) (years) Callability (%) rating
Ford Motor Co 9.95 300,000 02/10/ 40 3 N 8.56 6
1992
General 6.75 250,000 11/06/ 25 5 C 6.80 2
Electric Co 1986
General 8.25 500,000 04/26/ 30 3 C 8.41 2
Electric Co 1988
General 8.38 200,000 04/30/ 10 5 N 8.38 5
Motors Corp 1987
General 8.63 400,000 06/09/ 10 5 N 8.59 5
Motors Corp 1989
General 8.88 500,000 05/31/ 20 5 N 8.80 5
Motors Corp 1990
Harris Corp 6.65 100,000 08/01/ 10 5 N 6.90 8
1996
Ingersoll- 6.48 150,000 06/01/ 30 10 N 6.64 7
Rand Co 1995
Intl Business 13.75 125,000 03/09/ 12 3 C 14.05 2
Machines 1982
Corp
ITT Industries 8.50 100,000 01/20/ 10 5 N 8.50 7
Inc 1988
ITT Industries 8.55 100,000 06/12/ 20 9 N 9.20 7
Inc 1989
ITT Industries 3.98 100,000 02/15/ 15 3 N 8.60 7
Inc 1990
Johnson 7.70 125,000 02/28/ 20 10 N 7.83 7
Controls Inc 1995
K N Energy 7.35 125,000 07/25/ 30 10 N 7.47 9
Inc 1996
Litton 6.98 100,000 03/15/ 40 10 N 7.01 9
Industries Inc 1996
Lockheed 7.20 300,000 05/01/ 40 12 N 7.29 9
Martin Corp 1996
Marriott Corp 9.38 250,000 06/11/ 20 10 N 9.74 8
1987
Merrill Lynch 11.13 250,000 03/26/ 15 5 C 13.36 4
& Co 1984
Merrill Lynch 9.38 125,000 06/04/ 12 6 C 9.48 4
& Co 1985
Merrill Lynch 8.40 200,000 10/25/ 30 5 N 8.67 6
& Co 1989
Motorola Inc 8.40 200,000 08/15/ 40 10 N 8.36 4
1991
Motorola Inc 6.50 400,000 08/31/ 30 10 N 6.55 4
1995
(continued)
178 I.E. Brick et al.

Put S&P
Amount Issue Maturity expiration YTM credit
Name Coupon ($000’s) date (years) (years) Callability (%) rating
Occidental 9.25 300,000 08/03/ 30 15 N 9.37 10
Petroleum 1989
Corp
Penney 6.90 200,000 08/16/ 30 7 N 7.07 7
(JC) Co 1996
Philip Morris 7.00 150,000 07/15/ 5 2 N 7.30 7
Cos Inc 1986
Philip Morris 9.00 350,000 05/09/ 10 6 N 9.39 7
Cos Inc 1988
Philip Morris 6.95 500,000 06/01/ 10 5 N 6.91 7
Cos Inc 1996
Pitney Bowes 8.63 100,000 02/10/ 20 10 N 8.70 4
Inc 1988
Pitney Bowes 8.55 150,000 09/15/ 20 10 Y 8.64 4
Inc 1989
Regions 7.75 100,000 09/15/ 30 10 N 8.07 7
Financial Corp 1994
Ryder System 9.50 125,000 07/01/ 15 2 C 9.69 7
Inc 1985
Seagram Co 4.42 250,000 08/03/ 30 15 N 9.90 7
Ltd 1988
Security 12.50 150,000 09/27/ 12 6 C 12.62 3
Pacific Corp 1984
Security 7.50 150,000 04/03/ 15 3 C 7.54 3
Pacific Corp 1986
Service Corp 7.00 300,000 05/26/ 20 7 N 6.80 9
International 1995
Southtrust 1.62 100,000 05/09/ 30 10 N 7.28 8
Corp 1995
State Street 7.35 150,000 06/15/ 30 10 N 7.23 5
Corp 1996
Suntrust 6.00 200,000 02/15/ 30 10 N 6.33 7
Banks Inc 1996
Triad Systems 14.00 71,500 08/09/ 8 3 C 13.99 16
Corp 1989
TRW Inc 9.35 100,000 05/31/ 30 10 N 9.28 7
1990
Union Carbide 6.79 250,000 06/01/ 30 10 N 6.83 10
Corp 1995
United 8.50 150,000 09/22/ 30 10 N 8.60 9
Dominion 1994
Realty Trust
Westinghouse 11.88 100,000 03/19/ 12 3 C 13.38 6
Electric 1984
(continued)
5 Motivations for Issuing Putable Debt: An Empirical Analysis 179

Put S&P
Amount Issue Maturity expiration YTM credit
Name Coupon ($000’s) date (years) (years) Callability (%) rating
Westinghouse 8.88 150,000 05/31/ 24 4 N 8.84 13
Electric 1990
Whitman Corp 7.29 100,000 09/19/ 30 8 N 7.26 9
1996
WMX 8.75 250,000 04/29/ 30 5 C 8.79 4
Technology 1988
WMX 7.65 150,000 03/15/ 20 3 N 7.71 6
Technology 1991
WMX 6.22 150,000 05/09/ 10 3 N 7.41 6
Technology 1994
WMX 6.65 200,000 05/16/ 10 5 N 6.43 6
Technology 1995
WMX 7.10 450,000 07/31/ 30 7 N 7.16 7
Technology 1996
Xerox Corp 11.25 100,000 08/25/ 15 3 C 11.44 5
1983

Appendix 2: Sample of Firms Issuing Poison Put Bonds and the

Put Bond Characteristics

S&P
Amount Years to credit
Name Coupon ($million) Issue date maturity Callability YTM (%) rating
Aar Corp 9.500 65 10/27/89 12 N 9.425 10
AMR 9.750 200 03/15/90 10 N 9.85 7
Anheuser- 8.750 250 12/01/89 10 N 8.804 5
Busch
Armstrong 9.750 125 08/18/89 19 N 9.5 5
World
Ashland Oil 11.125 200 10/08/87 30 C 10.896 7
Becton 9.950 100 03/13/89 10 N 10 6
Dickinson
Bowater Inc 9.000 300 08/02/89 20 N 9.331 7
Chrysler 10.300 300 06/15/90 2 N 10.43 11
Financial
Coastal 10.250 200 12/06/89 15 N 10.02 11
Corporation
Consolidated 9.125 150 08/17/89 10 N 9.202 6
Freightways
Corning Inc 8.750 100 07/13/89 10 N 8.655 7
CPC Intl 7.780 200 12/15/89 15 N 7.780 8
(continued)
180 I.E. Brick et al.

S&P
Amount Years to credit
Name Coupon ($million) Issue date maturity Callability YTM (%) rating
Cummins 9.750 100 03/21/86 30 C 10.276 9
Engine
Cyprus 10.125 150 04/11/90 12 N 10.172 10
Minerals
Dresser 9.373 68.8 03/10/89 11 C 9.650 8
Industries
Eaton Corp 9.000 100 03/18/86 30 C 9.323 7
Federal 9.200 150 11/15/89 5 N 9.206 9
Express
General 10.125 115 03/22/90 12 N 10.269 8
American
Transportation
Georgia Pacific 10.000 300 06/13/90 7 N 10.053 9
Grumman Corp 10.375 200 01/05/89 10 C 10.375 9
Harris Corp 10.375 150 11/29/88 30 C 10.321 8
Harsco 8.750 100 05/15/91 5 N 8.924 8
International 9.700 150 03/21/90 10 N 9.823 8
Paper
Kerr-Mcgee 9.750 100 04/01/86 30 C 9.459 8
Knight-Rydder 9.875 200 04/21/89 20 N 10.05 5
Lockhee Corp 9.375 300 10/15/89 10 N 9.329 7
Maytag 8.875 175 07/10/89 10 N 9.1 8
Monsanto 8.875 100 12/15/89 20 N 8.956 7
Morton 9.250 200 06/01/90 30 N 9.358 5
International
Parker- 9.750 100 02/11/91 30 C 9.837 7
Hannifin
Penn Central 9.750 200 08/03/89 10 N 9.358 11
Penn Central 10.875 150 05/01/91 20 N 11.016 11
Potlatch 9.125 100 12/01/89 20 N 9.206 8
Questar 9.875 50 06/11/90 30 C 9.930 6
Ralston Purina 9.250 200 10/15/89 20 N 9.45 8
Rite-Aid 9.625 65 09/25/89 27 C 9.99 6
Rohm And 9.373 100 11/15/89 30 C 9.618 7
Haas
Safety-Kleen 9.250 100 09/11/89 10 N 9.678 9
Sequa Corp 9.625 150 10/15/89 10 N 9.574 11
Stanley Works 8.250 75 04/02/86 10 C 8.174 7
Strawbridge 8.750 50 10/24/89 7 C 9.374 8
And Clothier
Union Camp 10.000 100 04/28/89 30 C 10.185 6
Unisys 10.300 300 05/29/90 7 N 10.794 10
(continued)
5 Motivations for Issuing Putable Debt: An Empirical Analysis 181

S&P
Amount Years to credit
Name Coupon ($million) Issue date maturity Callability YTM (%) rating
United Airlines 12.500 150 06/03/88 7 C 12.500 15
United 8.875 300 11/13/89 30 N 9.052 5
Technologies
VF Corp 9.500 100 10/15/89 10 C 9.500 7
Weyerhaeuser 9.250 200 11/15/90 5 N 9.073 6
The S&P ratings are based on a scale from 1 (AAA+) to 24 (Not rated). In the Callability column,
C denotes callable bond and N denotes non-callable bond

Appendix 3: Estimating the Standard Abnormal Returns and the

White t-Statistic

Fama et al. (1969) introduced the event study methodology when they analyzed
the impact of stock dividend announcements upon stock prices. Essentially,
they used the Market Model to estimate the stochastic relationship between
stock returns and the market portfolio. In particular, we estimate the following
regression:

Ri, t ¼ ai þ bi Rm, t þ ei, t (5.3)

We estimate the market model coefficients using the time period that begins
200 trading days before and ends 31 trading days before the event, employing the
CRSP value-weighted market index as the benchmark portfolio. We use these
coefficients to estimate abnormal returns for days
30 to +30. The abnormal return
is defined as

ARi, t ¼ Ri, t
ai
bi Rm, t : (5.4)

The mean abnormal return for the sample of i firms, ARt, at time t is found by

X
n
ARt ¼ ARi, t =n: (5.5)
i¼1

In order to conduct tests of significance, we must ensure that ARi,t has identical
standard deviation. Assuming that the forecast values are normally distributed, we
can scale ARi,t by the standard deviation of the prediction, Sit, given by Eq. 5.4 In
particular,
n h 2  2 io1=2
Sit ¼ s2i þ ð1=EDÞ þ Rmt
Rm =S Rmi
Rm (5.6)
182 I.E. Brick et al.

where s2i is the variance of the error term of Eq. 5.3, ED is the number of days
to estimate the market model for firm I, Rmt is the return of the market portfolio, and
Rm is the mean market return in the estimation period.
Hence the standardized abnormal return, SARi,t, is equal to ARi,t/Sit. SARi,t is
distributed normally with a mean of zero and a standard deviation equal to one.
The mean standardized abnormal return for time t, SARt, is the sum of the SARi,t
divided by n. SARt is normally distributed with a mean of zero and
a standard deviation of the square root of l/n. The cumulative abnormal returns for
days 1 to k, CARk, is the sum of mean abnormal returns for t ¼ 1 to k.
The standardized cumulative mean excess returns for the k days after month
t ¼ 0, SCARt, is equal to the sum of SARt for days 1 to k. SCART is normally
distributed with a standard deviation of square root of k/n. Please see Campbell
et al. (1996) for a detailed discussion of these tests as well as event studies in general.
In our cross-sectional regressions where we regress the CARs on firm-specific
variables, we estimate the model using GMM. However, since we are not
interested in conditional estimates, our regressors are the instruments. Therefore
our parameters estimates are the same that would be obtained by OLS. However,
we use the White (1980) heteroscedastic-consistent estimator. In the standard
regression model,

Y ¼ Xb þ e (5.7)

The OLS estimator of b ¼ (X 0 X)
1(X 0 y) with the covariance matrix

1
1
VaRðbÞ ¼ ðX0 XÞ ðX0 OXÞðX0 XÞ (5.8)

When the errors are homoscedastic, O ¼ s2I and the variance reduces to s2(X0 X)
1.
However when O is unknown as shown by White (1980), Eq. 5.8 can be used
using a consistent estimator of O. White showed that can be done using the residual
from Eq. 5.7.
So when the heteroscedasticity is of the unknown form,

1    
1
Using ðX0 XÞ X0 diag e2 X ðX0 XÞ (5.9)

gives us heteroscedastic-consistent White standard errors and White t-statistics.

Appendix 4: Generalized Method of Moments (GMM)

GMM is a generalization of the method of moments developed by Hansen (1982).

The moment conditions are derived from the model. Suppose Yt is a multivariate
independently and identically distributed (i.i.d.) random variable. The econometric
model specifies the relationship between Yt and the true parameters of the model
(y0). To use GMM there must exist a function f(Yt, y0) so that
5 Motivations for Issuing Putable Debt: An Empirical Analysis 183

mðy0 Þ  E½f ðYt ; y0 Þ ¼ 0: (5.10)

In GMM, the theoretical expectations are replaced by sample analogs:

X
gðy; Yt Þ ¼ 1=T f ðYt ; yÞ: (5.11)

The law of large numbers ensures that the RHS of above equation is the same as

E½f ðYt ; y0 Þ: (5.12)

The sample GMM estimator of the parameters may be written as (see Hansen
1982)
h X i0 X
Y ¼ arg min 1=T f ðYt ; yÞ WT 1=T f ðYt ; yÞ: (5.13)

So essentially GMM finds the values of the parameters so that the sample
moment conditions are satisfied as closely as possible. In our case for the regression
model,

yt ¼ Xt 0 b þ et : (5.14)

The moment conditions include

E½ðyt
Xt 0 bÞxt  ¼ E½et xt  ¼ 0 for all t: (5.15)

So the sample moment condition is

X
1=T ðyt
Xt 0 bÞxt

and we want to select b so that this is as close to zero as possible. If we select b as

(X0 X)
1(X0 y), which is the OLS estimator, the moment condition is exactly satisfied.
Thus, the GMM estimator reduces to the OLS estimator and this is what we estimate.
For our case the instruments used are the same as the independent variables. If,
however, there are more moment conditions than the parameters, the GMM estimator
above weighs them. These are discussed in detail in Greene (2008, Chap. 15).
The GMM estimator has the asymptotic variance
 
1

1
X0 Z ðZ0 OZÞ Z0 X (5.16)

In our case Z ¼ X since we use the independent variables as the instruments Z.

The White robust covariance matrix may be used for O as discussed in
Appendix 3 when heteroscedasticity is present. Using this approach, we estimate
GMM with White heteroscedasticity consistent t-stats.
184 I.E. Brick et al.

References
Bae, S. C., Klein, D. P., & Padmaraj, R. (1994). Event risk bond covenants, agency costs of debt
and equity, and stockholder wealth. Financial Management, 23, 28–41.
Bae, S. C., Klein, D. P., & Padmaraj, R. (1997). Firm characteristics and the presence of event risk
covenants in bond indentures. Journal of Financial Research, 20, 373–388.
Barnea, A., Haugen, R. A., & Senbet, L. W. (1980). A rationale for debt maturity structure and call
provisions in the agency theoretic framework. Journal of Finance, 25, 1223–1234.
Billett, M. T., King, T. D., & Mauer, D. C. (2007). Growth opportunities and the choice of
leverage, debt maturity, and covenants. Journal of Finance, 62, 697–730.
Bodie, Z., & Taggart, R. A. (1978). Future investment opportunities and the value of the call
provision on a bond. Journal of Finance, 33, 1187–1200.
Brennan, M. J., & Schwartz, E. S. (1988). The case for convertibles. Journal of Applied Corporate
Finance, 1, 55–64.
Brick, I. E., & Palmon, O. (1993). The tax advantages of refunding debt by calling, repurchasing,
and putting. Financial Management, 22, 96–106.
Brick, I. E., & Wallingford, B. A. (1985). The relative tax benefits of alternative call features in
corporate debt. Journal of Financial and Quantitative Analysis, 20, 95–106.
Campbell, J., Lo, A. W., & MacKinlay, C. (1996). The econometrics of financial markets.
Princeton: Princeton University Press.
Chatfield, R. E., & Moyer, R. C. (1986). Putting away bond risk: An empirical examination of the
value of the put option on bonds. Financial Management, 15(2), 26–33.
Cook, D. O., & Easterwood, J. C. (1994). Poison put bonds: An analysis of their economic role.
Journal of Finance, 49, 1905–1920.
Crabbe, L. (1991). Event risk: An analysis of losses to bondholders and ‘super poison put’ bond
covenants. Journal of Finance, 46, 689–706.
Crabbe, L., & Nikoulis, P. (1997). The putable bond market: Structure, historical experience, and
strategies. The Journal of Fixed Income, 7, 47–60.
Dann, L. Y., & Mikkelson, W. H. (1984). Convertible debt issuance, capital structure change and
financing-related information: Some new evidence. Journal of Financial Economics, 13,
157–186.
David, A. (2001). Pricing the strategic value of putable securities in liquidity crises. Journal of
Financial Economics, 59, 63–99.
Eckbo, B. E. (1986). Valuation effects of corporate debt offerings. Journal of Financial Econom-
ics, 5, 119–152.
Fama, E., Fisher, L., Jensen, M., & Roll, R. (1969). The adjustment of stock prices to new
information. International Economic Review, 10, 1–21.
Green, R. (1984). Investment incentives, debt, and warrants. Journal of Financial Economics, 13,
115–136.
Greene, W. (2008). Econometric analysis. Boston: Pearson.
Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica, 50, 1029–1054.
Hansen, R. S., & Chaplinsky, S. (1993). Partial anticipation, the flow of information and the
economic impact of corporate debt sales. Review of Financial Studies, 6, 709–732.
Hansen, R. S., & Crutchley, C. (1990). Corporate earnings and financings: An empirical analysis.
Journal of Business, 63, 347–371.
Haugen, R. A., & Senbet, L. W. (1978). The insignificance of bankruptcy costs to the theory of
optimal capital structure. Journal of Finance, 33, 383–393.
Haugen, R. A., & Senbet, L. W. (1981). Resolving the agency problems of external capital through
options. Journal of Finance, 26, 629–647.
Haugen, R. A., & Senbet, L. W. (1988). Bankruptcy and agency costs: Their significance
to the theory of optimal capital structure. Journal of Financial and Quantitative Analysis,
23, 27–38.
5 Motivations for Issuing Putable Debt: An Empirical Analysis 185

Jensen, M. C. (1986). Agency costs of free cash flow, corporate finance, and takeovers. American
Economic Review, 76, 323–329.
Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs
and ownership structure. Journal of Financial Economics, 3, 305–360.
Jung, K., Kim, Y., & Stulz, R. M. (1996). Timing, investment opportunities, managerial discretion,
and the security issue decision. Journal of Financial Economics, 42, 159–185.
Kim, E. H. (1978). A mean-variance theory of optimal capital structure and corporate debt
capacity. Journal of Finance, 33, 45–63.
Lehn, K., & Poulsen, J. (1989). Free cash flow and stockholders gain in going private transactions.
Journal of Finance, 44, 771–788.
Lewis, C., Rogalski, R., & Seward, J. K. (1998). Industry conditions, growth opportunities and the
market reactions to convertible debt financing decisions. (Unpublished mimeo, Vanderbilt
University).
Mackie-Mason, J. K. (1990). Do taxes affect corporate financing decisions? Journal of Finance,
45, 1471–1493.
Meulbroek, L. K., Mitchell, M. L., Mulherin, J. H., Netter, J. M., & Poulsen, A. B. (1990). Shark
repellants and managerial myopia: An empirical test. Journal of Political Economy, 98,
1108–1117.
Mikkelson, W. H., & Partch, M. M. (1986). Valuation effects of security offerings and the issuance
process. Journal of Financial Economics, 15, 31–60.
Miller, M. H. (1977). Debt and taxes. Journal of Finance, 32, 261–275.
Miller, M. H., & Rock, K. (1985). Dividend policy under asymmetric information. Journal of
Finance, 40, 1031–1051.
Modigliani, F., & Miller, M. H. (1963). Corporate income taxes and the cost of capital:
A correction. American Economic Review, 63, 433–443.
Myers, S. (1977). Determinants of corporate borrowing. Journal of Financial Economics, 5,
147–175.
Nash, R. C., Netter, J. M., & Poulsen, A. B. (2003). Determinants of contractual relations between
shareholders and bondholders: Investment opportunities and restrictive covenants. Journal of
Corporate Finance, 9, 201–232.
Ravid, S. A., & Sarig, O. H. (1991). Financial signaling by committing to cash outflows. Journal of
Financial and Quantitative Analysis, 26, 165–180.
Robbins, E. H., & Schatzberg, J. D. (1986). Callable bonds: A risk-reducing signaling mechanism.
Journal of Finance, 41, 935–950.
Ross, S. (1977). The determination of financial structure: The incentive signaling approach. Bell
Journal of Economics, 8, 23–40.
Roth, G., & McDonald, C. G. (1999). Shareholder-management conflict and event risk covenants.
Journal of Financial Research, 22, 207–225.
Scott, J. (1976). A theory of optimal capital structure. Bell Journal of Economics, 7, 33–54.
Shyam-Sunder, L. (1991). The stock price effect of risky versus safe debt. Journal of Financial
and Quantitative Analysis, 26, 549–558.
Spiegel, Y., & Wilkie, S. (1996). Investment in a new technology as a signal of firm value under
regulatory opportunism. Journal of Economics and Management Strategy, 5, 251–276.
Stein, J. C. (1988). Takeover threats and managerial myopia. Journal of Political Economy, 96,
61–80.
Tewari, M., & Ramanlal, P. (2010). Is the put option in U.S. structured bonds good news for both
bondholders and stockholders? International Research Journal of Finance and Economics, 52,
50–59.
Wahal, S., & McConnell, J. J. (2000). Do institutional investors exacerbate managerial
myopia. Journal of Corporate Finance: Contracting, Governance and Organization, 6,
307–329.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica, 48, 817–838.
Multi-Risk Premia Model of US Bank
Returns: An Integration of CAPM and APT 6
Suresh Srivastava and Ken Hung

Contents
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.2 Multiple Regression Model of Bank Return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
6.2.1 Pricing of Interest Rate Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.3 Multi-Risk Premia Asset-Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.4 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
6.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.5.1 Two-Variable Regression Model and Pricing of Interest Rate Risk . . . . . . . . . . . . 195
6.5.2 Multi-risk Premia Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Appendix 1: An Integration of CAPM and APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Principal Component Factor Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Appendix 2: Interest Rate Innovations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Orthogonalization Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Univariate ARMA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Vector ARMA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Abstract
Interest rate sensitivity of bank stock returns has been studied using an
augmented CAPM, a multiple regression model with market returns and interest
rate as independent variables. In this paper, we test an asset-pricing model in
which the CAPM is augmented by three orthogonal factors which are proxies for
the innovations in inflation, maturity risk, and default risk. The model proposed

S. Srivastava (*)
University of Alaska Anchorage, Anchorage, AK, USA
e-mail: afscs@uaa.alaska.edu
K. Hung
Division of International Banking & Finance Studies, Texas A&M International University,
Laredo, TX, USA
e-mail: ken.hung@tamiu.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 187
DOI 10.1007/978-1-4614-7750-1_6,
# Springer Science+Business Media New York 2015
188 S. Srivastava and K. Hung

is an integration of CAPM and APT. The results of the two models are compared
to shed light on sources of interest rate risk.
Our results using the integrated model indicate the inflation beta to be
statistically significant. Hence, innovations in short-term interest rates contain
valuable information regarding inflation premium; as a result the interest rate risk
is priced with respect to the short-term interest rates. Further, it also indicates that
innovations in long-term interest rates contain valuable information regarding
maturity premium. Consequently the interest rate risk is priced with respect to
the long-term interest rates. Using the traditional augmented CAPM, our investi-
gation of the pricing of the interest rate risk is inconclusive. It shows that interest
rate risk was priced from 1979 to 1984 irrespective of the choice of interest rate
variable. However, during the periods 1974–1978 and 1985–1990, bank stock
returns were sensitive only to the innovations in the long-term interest rates.

Keywords
CAPM • APT • Bank stock return • Interest rate risk • Orthogonal factors •
Multiple regression

6.1 Introduction

Interest rate sensitivity of commercial bank stock returns has been the subject of
considerable academic research. Stone (1974) proposed a multiple regression
model incorporating both the market return and interest rate variables as return-
generating independent variables. While some studies have found the interest rate
variable to be an important determinant of common stock returns of banks (Fama
and Schwert 1977; Lynge and Zumwalt 1980; Christie 1981; Flannery and James
1984; Booth and Officer 1985), others have found the returns to be insensitive
(Chance and Lane 1980) or only marginally explained by the interest rate factor
(Lloyd and Shick 1977). A review of the early literature can be found in Unal and
Kane (1988). Sweeney and Warga (1986) used the APT framework and concluded
that the interest rate risk premium exists but varies over time. Flannery et al. (1997)
tested a two-factor model for a broad class of security returns and found the effect
of interest rate risk on security returns to be rather weak. Bae (1990) examined the
interest rate sensitivity of depository and nondepository firms using three different
maturity interest rate indices. His results indicate that depository institutions’ stocks
are sensitive to actual and unexpected interest rate changes, and the sensitivity
increases for longer-maturity interest rate variables. Song (1994) examined the
two-factor model using time-varying betas. His results show that both market
beta and interest rate beta varied over the period 1977–1987. Yourougou (1990)
found the interest rate risk to be high during a period of great interest rate volatility
(post-October 1979) but low during a period of stable interest rates (pre-October
1979). Choi et al. (1992) tested a three-factor model of bank stock returns using
market, interest, and exchange rate variables. Their findings about interest rate risk
are consistent with the observations of Yourougou (1990).
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 189

The issue of interest rate sensitivity remains empirically unresolved. Most of the
studies use a variety of short-term and long-term bond returns as the interest rate
factor without providing any rationale for their use. The choice of bond market
index seems to affect the pricing of the interest rate risk. Yet, there is no consensus
on the choice of the interest rate factor that should be used in testing the two-factor
model. In this paper, we provide a plausible explanation of why pricing of interest
rate risk differs with the choice of interest rate variable. We also suggest a hybrid
return-generating model for bank stock returns in which the CAPM is augmented
by three APT-type factors to account for unexpected changes in the inflation
premium, the maturity-risk premium, and the default-risk premium. The use of
three additional factors provides a better understanding of the interest rate sensi-
tivity and offers a plausible explanation for the time-varying interest rate risk
observed by other investigators. Our empirical investigation covers three distinc-
tion economic and bank regulatory environments: 1974–1978, a period of increas-
ing but only moderately volatile interest rates in a highly regulated banking
environment; (2) 1979–1984, a period characterized by high level of interest rates
with high volatility, in which there was gradual deregulation of the banking
industry; and (3) 1985–1990, a low interest rate and low-volatility period during
which many regulatory changes were made in response to enormous bank loan
losses and bankruptcies. The results of the multifactor asset-pricing model are
compared with those from the two-factor model in order to explain the time-
varying interest rate risk.
The rest of this paper is divided into five sections. In Sect. 6.2, we
describe the two-factor model of the bank stock return and the pricing of the
interest rate risk. The multi-risk premia model and the specification of the
factors are discussed in Sect. 6.3. The data for this analysis is described in
Sect. 6.4. Section 6.5 presents empirical results, and Sect. 6.6 concludes the
paper.

6.2 Multiple Regression Model of Bank Return

Stone (1974) proposed the following two-variable bank stock return-generating

model:

Rjt ¼ aj þ b1j Rmt þ b2j RIt þ ejt (6.1)

where Rjt is the bank common stock return, Rmt is the market return, and RIt is the
innovation in the interest rate variable. Coefficients aj and b1j are analogous to the
alpha and beta coefficients of the market model, and b2j represents interest rate risk.
Since then, numerous researchers have studied the pricing of interest rate risk with
varying results. While Stone (1974) and others did not place an a priori restriction
on the sign of b2j, the nominal contracting hypothesis implies that it should
be positive. This is because the maturity of bank assets is typically longer than
190 S. Srivastava and K. Hung

that of liabilities.1 Support for this hypothesis was found by Flannery and James
(1984) but not by French et al. (1983).
An important issue in the empirical investigation of the two-factor model is the
specification of an appropriate interest rate factor. Theoretical consideration of
factor analysis requires that two factors, Rmt and RIt, be orthogonal whereby choice
of the second factor (RIt) would not influence the first factor loading (blj). The
resolution of this constraint requires a robust technique for determining the unex-
pected changes in the interest rate that is uncorrelated with the market return. There
are three approaches to specify the interest rate factor (Appendix 2). In the first
approach, the expected change in the interest rate is estimated using the high
correlation between the observed interest rate and the market rate. The residual –
difference between observed and estimated rates – is used as the interest rate factor.
The second approach is to identify and estimate a univariate ARMA model for the
interest rate variable and use the residuals from the ARMA model as the second
factor. In the third approach, the interest rate variable (RIt) and the market return
(Rmt) are treated as the components of a bivariate vector, which is modeled as
a vector ARMA process. The estimated model provides the unanticipated change in
interest rate variable to be used as the second factor in the augmented CAPM,
Eq. 6.1. Srivastava et al. (1999) discuss the alternate ways of specifying the
innovations in the interest rate variable and its influence on the pricing of the
interest rate risk. In this paper, the error term from the regression of interest rates
on market returns is used as the orthogonal interest rate factor in Eq. 6.1.

6.2.1 Pricing of Interest Rate Risk

In addition to changes in the level of expected or unexpected inflation, changes in other

economic conditions produce effects on interest rate risk. For example, according to the
intertemporal model of the capital market (Merton 1973; Cox et al. 1985), a change in
interest rates alters the future investment opportunity set; as a result, investors require
additional compensation for bearing the risk of such changes. Similarly, changes in the
investor’s degree of risk aversion, default risk, or maturity risk of bank financial assets
cause additional shifts in the future investment opportunities for the bank stockholders.
The specific choice of the bond market index for the two-variable model determines
what unexpected change is captured by the coefficient b2j.
The nominal return on a debt security, R, is expressed as

R ¼ Rrf þ MRP þ DRP þ LP (6.2)

where Rrf is the real risk-free rate plus an inflation premium, DRP is the default-risk
premium, MRP is the maturity-risk premium, and LP is the liquidity-risk premium.

1
The sign of b2j is negative when changes in bond yields and not the bond market return are used as
the interest rate variable (see Sweeney and Warga 1986).
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 191

A change in nominal return consists of changes in the risk-free, liquidity-risk,

default-risk, and maturity-risk rate. A change in the short-term risk-free rate can
be attributed to changes in real rate or short-term inflation. A sizeable change in the
real rate takes place over a longer time horizon. Therefore, it should not signifi-
cantly impact monthly stock returns and cause the interest rate risk.2 However,
Fama and Gibbons (1982) suggested that changes in real rate may cause changes in
the magnitude of interest risk from period to period. Fama and Schwert (1977)
argued that, in equilibrium, the risk premium in the short-term interest rate is
compensation to the investor for changes in the level of expected inflation. How-
ever, French et al. (1983) pointed out that the interest rate risk is due to the
unexpected changes in inflation. More specifically, the interest rate risk can be
viewed as the compensation for expected or unexpected changes in the level of
short-term inflation. Maturity risk is associated with the changes in the slope of the
yield curve that caused a number of economic factors, such as supply and demand
of long-term credit, long-term inflation, and risk aversion. Estimates of b2j using
short-term T-bill returns as the second factor account for the changes in short-term
inflation, whereas estimates of b2j using long-term T-note returns account for the
unexpected changes in long-term inflation as well as maturity risk. The inflation
expectation horizon approximates that of the maturity of the debt security. Esti-
mating coefficient b2j using BAA bond returns as the interest rate factor explains
the fluctuations in bank stock return due to unexpected changes in long-term
inflation, maturity-risk premium, default-risk premium, or a combination of risk
premia.3
As the risk premia are additive, in Eq. 6.2, the magnitude b2j depends on
the choice of the interest rate index. A priori expectations about the relative size
of b2j are:

b2jf6-month T-billg > b2jf3-month T-billg

b2j f7-year T-noteg > b2j f6-month T-billg (6.3)
b2j fBAA-rated bondg > b2j f7-year T-noteg

where the debt security within the brackets identifies the choice of the bond market
index. A number of researchers have indicated that bank stock returns are sensitive
to unexpected changes in long-term but not short-term interest rates. This observa-
tion is consistent with the expectations expressed in Eq. 6.3. However, we are
unable to isolate and identify which component of the interest rate risk is being

2
Federal Reserve’s intervention changes the short-term interest rates. These rate changes take
place after considerable deliberation and are often anticipated by financial markets. Hence, it
neither generates interest rate innovations nor produces interest rate risk.
3
Liquidity premium is ignored in our discussion and subsequent model construction because our
sample does not include financially distressed banks, so there are insufficient variations in the
liquidity premium.
192 S. Srivastava and K. Hung

priced when a long-term bond return is used. It is feasible to observe

a significant change in stock returns due to unexpected changes in default-
risk premium (or maturity-risk premium), whether or not nominal contracting
hypothesis is valid. The two-factor model can ascertain the pricing of interest
rate risk without identifying the source. Commercial banks have a variety of
nominal assets and liabilities with different sensitivities to unexpected changes
in short-term inflation, maturity risk, and default risk. In the next section, we
propose an asset-pricing model in which the interest rate factor of the
two-factor model is replaced by three factors, each of which represents
a different risk component of the bond return.4

6.3 Multi-Risk Premia Asset-Pricing Model

We propose a hybrid asset-pricing model to investigate the interest rate sensitivity

of bank stock returns.5 The traditional CAPM is augmented by three additional
APT-type factors to account for unexpected changes in the inflation premium,
the maturity-risk premium, and the default-risk rating. Hence, the proposed
return-generating model is written as

Rjt ¼ aj þ b1j Rmt þ b2Pj DPt þ b2Mj DMRt

(6.4)
þ b2Dj DDRt þ ejt

where DPt, DMRt, and DDRt are the proxies for the innovations in inflation,
maturity risk, and default risk, respectively (specification of these principal
component factors consistent with APT is discussed in the Appendix). Coeffi-
cients b1, b2P, b2M, and b2D are the measures of market risk, inflation risk,
maturity risk, and default risk, respectively. The expected return is given by
CAPM:


E Rj ¼ aj þ b1j EðRmÞ (6.5)

However, systematic risk is determined by the integrated model:

Total systematic risk ¼ market risk þ inflation risk þ maturity risk

þ default risk (6.6)

4
An appropriate interest rate variable that should be used to examine the pricing of the interest rate
risk in the two-variable framework is not easily available. However, one could construct an index
composed of long-term US bonds and corporate bonds with duration equal to the net duration of
bank assets and appropriate default risk. This interest rate variable will identify the true pricing of
the interest rate risk.
5
This model’s conceptual framework is provided by Chen et al. (1986). However, their factors are
not orthogonal.
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 193

The three APT-type factors have managerial implications regarding bank’s asset
and liability management. The first factor, DPt, has implications pertaining to the
management of the treasury securities portfolio held by the bank. The value of this
portfolio is sensitive to the changes in the short-term interest rate. In addition, these
changes impact the short-maturity funding gap. The null hypothesis b2P ¼ 0 implies
that management has correctly anticipated future changes in short-term inflation
and has taken steps to correctly hedge the pricing risk through the use of derivative
contracts and minimize the short-maturity funding gap.
The second factor, DMRt, has implications regarding the bank’s long-term assets
and liabilities. The market value of a bank’s net worth is very sensitive to changes
in the slope of the term structure. The null hypothesis b2M ¼ 0 conjectures that
management has correctly anticipated future changes in the slope of the term
structure and has immunized the institution’s net worth by a sensible allocation
assets in the loan portfolio, sale of loans to the secondary markets, securitization of
loans, and the use of derivative contracts. The third factor, DDRt, relates to the
management of loan losses and overall default-risk rating of bank assets. The null
hypothesis b2D ¼ 0 infers that management has correctly anticipated future loan
losses due to changes in exogenous conditions, and subsequent loan losses,
however large, will not adversely affect the stock returns.

6.4 Data Description

Month-end yield for 3-month T-bill, 6-month T-bill, 1-year T-note, 7-year T-note,
and BAA corporate bonds, for the period January 1974 to December 1990, were
obtained, and monthly returns were calculated.6 Return on the CRSP equally
weighted index of NYSE stocks was used as the market return. Month-end closing
prices and dividends for a sample of 88 banks were obtained from Compustat’s
Price, Dividend, and Earning (PDE) data tape, and monthly returns were calculated.
Availability of continuous data was the sole sample selection criteria. This selection
criterion does introduce a survivorship bias. However, it was correctly pointed out
by Elyasiani and Iqbal (1998) that the magnitude of this bias could be small and not
affect the pricing of interest risk. Equally weighted portfolios of bank returns were
calculated for this study. The total observation period is divided into three
contrasting economic and bank regulatory periods: (1) an increasing but moderately
volatile interest rate period from January 1974 to December 1978 in a highly
regulated environment; (2) a high interest rate and high-volatility period from
January 1979 to December 1984, during which there was gradual deregulation of
the industry; and (3) a low interest rate and low-volatility period from January 1985
to December 1990, during which many regulatory changes were made in response
to banks loan loss problems. The descriptive statistics of the sample are

6
The return on a bond index is calculated from the yield series as RIt ¼  (YIt  YI,t1)/YIt
where YIt is the bond index yield at time t.
194 S. Srivastava and K. Hung

Table 6.1 Summary statistics of portfolio returns and interest rate yields
1974–1978 1979–1984 1985–1990
Bank portfolio return 9.66%a 22.28% 6.48%
(5.95%)b (4.93%) (6.17%)
Portfolio betac 0.7091 0.7259 1.0102
CRSP market return 22.05% 22.05% 8.58%
(7.63%) (5.39%) (5.36%)
3-month T-bill yield 6.21% 10.71% 6.92%
(1.27%) (2.41%) (1.00%)
6-month T-bill yield 6.48% 10.80% 7.02%
(1.27%) (2.21%) (0.95%)
1-year T-note yield 7.05% 11.70% 7.62%
(1.25%) (2.27%) (1.00%)
7-year T-note yield 7.23% 11.91% 8.67%
(0.54%) (1.79%) (1.09%)
BAA bond yield 9.66% 14.04% 10.84%
(0.68%) (1.97%) (0.98%)
Monthly inflationd 0.78% 0.46% 0.25%
(0.69%) (0.50%) (0.23%)
Maturity-risk premiume 0.68% 0.20% 1.05%
(0.88%) (1.32%) (0.72)
Default-risk premiumf 1.94% 2.13% 2.17%
(0.77%) (0.76%) (0.46%)
a
Average returns and yields are reported on annualized basis
b
Standard deviation is in parenthesis
c
Estimated using single-index market model
d
Measured by the relative change in consumer price index
e
Yield differential between 7-year T-note and 1-year T-note
f
Yield differential between BAA-rated corporate bond and 7-year T-note

summarized in Table 6.1. The average monthly inflation, as measured by the

relative change in the consumer price index, was highest during 1974–1978 and
lowest during 1985–1990. The average default-risk premium was of the same order
of magnitude for all the three periods. The average maturity-risk premium was high
during 1985–1990 indicating a steeper yield curve. For the period 1979–1984, the
average maturity-risk premium was low, but its standard deviation was high. This
indicates a relatively less steep but volatile yield curve.
The bank portfolio’s average return (9.66 %), for the period 1974–1978,
was much smaller than the average market return (22.05 %). For the period
1979–1984, both returns increased dramatically; the portfolio’s average return
(22.28 %) was about the same as the average market return (22.05 %). For the
period 1985–1990, the portfolio and markets average return both dropped
dramatically to 6.48 % and 8.58 %, respectively. The estimated portfolio beta
increased from about 0.7 in the two earlier periods to about 1.0 in the latest
period.
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 195

6.5 Empirical Results

6.5.1 Two-Variable Regression Model and Pricing of Interest

Rate Risk

The estimated coefficients of the two-variable regression model are presented in

Table 6.2.7 For the period 1974–1978, a period characterized by low and stable
interest rates, the interest rate beta was statistically insignificant with all the
interest rate factors except the 7-year T-note returns. For the period 1979–1984,

Table 6.2 Two-variable model of the bank stock returns: Rjt ¼ aj + b1jRmt + b2jRIt + ejt
Interest rate variablea
3-month 6-month 1-year 7-year BAA
Period Estimates T-bill T-bill T-note T-note Bond
1974–1978 Constant, a 0.0046 0.0045 0.0043 0.0033 0.0042
(0.18)b (0.17) (0.16) (0.14) (0.16)
Market beta, b1 0.6921 0.6886 0.6897 0.6845 0.6987
(15.05) (14.67) (15.17) (16.47) (15.76)
Interest rate beta, b2 0.0566 0.0608 0.0700 0.2893 0.2248
(1.01) (1.07) (1.24) (2.77) (0.93)
R2 0.828 0.828 0.830 0.845 0.828
F-statistic 137.3 137.7 138.8 156.3 136.9
1979–1984 Constant, a 0.0052 0.0052 0.0052 0.0052 0.0052
(1.49) (1.53) (1.57) (1.70) (1.62)
Market beta, b1 0.7261 0.7258 0.7262 0.7262 0.7264
(11.73) (12.01) (12.37) (13.33) (12.74)
Interest rate beta, b2 0.1113 0.1425 0.1714 0.3582 0.6083
(3.47) (3.99) (4.41) (5.97) (5.15)
R2 0.684 0.699 0.716 0.755 0.732
F-statistic 74.8 80.1 87.2 106.6 94.4
1985–1990 Constant, a 0.0018 0.0018 0.0018 0.0018 0.0018
(0.51) (0.51) (0.52) (0.53) (0.52)
Market beta, b1 1.010 1.010 1.010 1.010 1.010
(15.31) (15.30) (15.45) (15.70) (15.57)
Interest rate beta, b2 0.0508 0.0374 0.1036 0.1752 0.2741
(0.51) (0.40) (1.22) (1.95) (1.62)
R2 0.773 0.772 0.777 0.794 0.780
F-statistic 117.3 117.1 120.0 125.1 122.5
a
The error term from the regression of interest rate on market return is used as the appropriate
orthogonal interest rate variable
b
t-statistics are in the parenthesis

7
Test of two-variable model for the period 1991–2007 indicated that interest rate risk is not priced.
Tables can be provided to interested readers.
196 S. Srivastava and K. Hung

the interest rate beta was statistically significant with all the interest factors, and
its magnitude varied substantially. This period was characterized by a relatively
flat but volatile yield curve. For the period 1985–1990, the interest rate beta was
statistically significant only with the 7-year T-note returns as the interest rate
factor. In general, the estimated value of b2j for the periods 1974–1978 and
1985–1990 is smaller and less statistically significant than the value for the period
1979–1984.
The magnitude of interest rate beta, whether significant or not, varies with the
choice of the interest rate variable. The size of b2j increases with the choice of
securities in the following order: 3-month T-bill, 6-month T-bill, 1-year T-note,
7-year T-note, and BAA-rated bond (except b2j estimated using BAA bond return
for the period 1974–1978 and the 6-month T-bill return in 1985–1990). This
observation validates inequality (6.3) in all the periods and suggests the expanding
nature of the investment opportunity set with increased horizon (Merton 1973). As
stated earlier, the difference between b2j estimated using 7-year T-note returns and
that using 3-month T-bill returns measures the effect on the bank stock returns due
to the unexpected changes in the maturity-risk premium. Further, difference
between b2j estimated using BAA bond returns and that using 7-year T-note returns
measures the effect on the bank stock returns due to the unexpected changes in the
default-risk premium.
The fact that bank stock returns are more sensitive to the long-term interest rates
than to short-term interest rates is consistent with our expectation about the size of
b2 expressed in inequalities (6.3). Similar results were reported by other researchers
(such as Unal and Kane (1988) and Chen et al. (1986)). A shift of focus from short-
term to long-term inflation expectation could explain this result. An alternative
explanation is that bank balance sheet returns are better approximated by long-term
than by short-term bond returns. To the extent that balance maturity mismatches
occur, they should be related to long-term bond returns. The reason is simply that
long-term bond returns include the present value of more future period returns than
do short-term bond returns. That is, long-term bond returns include price changes
not included in short-term bond returns. The price changes in the long-term bond
returns represent price changes in long-term bank contracts. The most representa-
tive term equals the term of assets or liabilities, whichever is longer. If the maturity
mismatch is a net asset (liability) position, then the long-term bond maturity would
reflect the asset (liability) maturity. The estimate of a large, positive coefficient (b2j)
for 7-year T-notes and BAA bonds implies that banks mismatch in favor of assets
with relatively long maturities.
The plausible causes for the change in the interest risk from period to period are
(1) changes in real returns, as reported by Fama and Gibbons (1982), (2) unexpected
changes in short-term and long-term inflation expectation, (3) shift of focus from
short-term to long-term inflation expectation, (4) unexpected changes in risk aver-
sion, (5) unexpected changes in the default risk of bank’s nominal contracts, and
(6) structural instability of the systematic interest rate equation used to extract
interest rate innovations. The estimated coefficients of multi-risk premia model will
shed some light on this issue.
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 197

6.5.2 Multi-risk Premia Model

The estimates of the first systematic interest rate Eq. 6.11, which specifies innovations
in the short-term default-free bond returns, are reported in Table 6.3. The ordinary
least square (OLS) estimates were rejected because of the presence of serial auto-
correlation as indicated by the Durbin-Watson statistic (see Woolridge 2009).
A Durbin-Watson statistic equal to 2 indicates the absence of any serial autocorre-
lation. The generalized least square (GLS) estimates seemed more appropriate
because the Durbin-Watson statistic was approximately equal to 2 and the value of
R2 was higher. Results reported in Table 6.3 exhibit a significant correlation between
market return and short-term interest rate variable in the period 1974–1978, but not in
the periods 1979–1984 and 1985–1990. However, a low value of R2 indicates that the
relationship expressed in Eq. 6.4 is not as robust as one would have preferred. This
approach of extracting changes in short-term inflation was used by Fama and Schwert
(1977) and French et al. (1983). Their estimation period overlapped our estimation
period 1974–1978 but not the later ones. In Table 6.4 we present the results of the
second systematic interest rate Eqs. 6.12a and 6.12b. The estimated coefficient, y1,
determined the average slope of the yield curve during the estimation period. The OLS
estimates were rejected because of the presence of serial correlation indicated by the
Durbin-Watson statistics. The term structure coefficient y1 was significant in all the
periods for Specifications a and b. However, coefficients y2 of Specification b was

Table 6.3 Estimating a proxy for ex-post unexpected inflation: RST,t ¼ d0 + d1Rmt + et
Estimated coefficients
1974–1978 1979–1984 1985–1990
Ordinary least square
Constant, d0 0.0068 0.0070 0.0037
(0.87)a (0.45) (0.89)
Market linkage coefficient, d1b 0.2987 0.5791 0.0891
(2.97) (2.19) (1.13)
R2 0.132 0.076 0.018
Durbin-Watson statistic 1.90 1.19 1.12
Generalized least square
Constant, d0 0.0075 0.0013 0.0044
(0.82) (0.08) (0.83)
Market linkage coefficient, d1b 0.3173 0.3136 0.0806
(3.16) (1.44) (1.15)
R2 0.145 0.306 0.242
Durbin-Watson statistic 1.99 1.97 1.92
Ordinary Least Square estimates were rejected because of the presence of serial correlation
indicated by Durbin-Watson statistic
RST,t and Rmt are 3-month T-bill and CRSP equally-weighted market returns
a
t-statistics are in the parenthesis
b
Coefficient d1 measures the stock–bond market linkage
198 S. Srivastava and K. Hung

Table 6.4 Estimating alternate proxies for maturity-risk premia: Specification a : RLT, t ¼ y0 þ
y1 RST, t þ et , Specification b : RLT, t ¼ y0 þ y1 RST, t þ y2 Rmt þ et
1974–1978 1979–1984 1985–1990
Specification a: generalized least square
Constant, y0 0.0028 0.0053 0.0017
(1.02)a (2.11) (0.66)
Term structure coefficient, y1b 0.3988 0.5201 0.8183
(10.30) (16.29) (14.96)
R2 0.675 0.846 0.820
Durbin-Watson statistic 1.98 1.83 1.84
Specification b: generalized least square
Constant, y0 0.0024 0.0064 0.0011
(0.84) (2.47) (0.48)
Term structure coefficient, y1b 0.4070 0.5122 0.8105
(10.02) (15.78) (15.34)
Market linkage coefficient, y2c 0.0216 0.0576 0.0944
(0.68) (1.07) (2.56)
R2 0.677 0.849 0.835
Durbin-Watson statistic 1.98 1.82 1.80
Ordinary least square estimates were rejected because of the presence of serial correlation
indicated by Durbin-Watson statistic
RLT,t and RST,t are 7-year T-note and 1-year T-note returns
a
t-statistics are in the parenthesis
b
Coefficient y1 measures the average slope of the yield curve
c
Coefficient y2 accounts for the stock–bond market linkage

significant only during 1985–1990. The errors from Specification a for periods
1974–1978 and 1979–1984 and from Specification b for the period 1985–1990 were
used to specify the second risk factor DMRt. Estimates of the third systematic interest
rate Eqs. 6.13a and 6.13b are presented in Table 6.5. As before, the GLS estimates
were deemed appropriate. The errors from Specification a for the periods 1974–1978
and 1979–1984 and from Specification b for the period 1985–1990 were used to
specify the unexpected change in default risk, DDRt.
Results of the multi-risk premia model are presented in Table 6.6. The inflation
beta is statistically significant in the periods 1979–1984 and 1985–1990 but not in
the period 1974–1978. It was shown in Table 6.3 that quasi-differenced short-term
interest rates were correlated with the market return for the period 1974–1978 but
not for the periods 1979–1984 and 1985–1990. Hence, one could argue that when
short-term interest rates are correlated with the market return (i.e., 1974–1978), the
error term from Eq. 6.8 contains no systematic information. This results in the
inflation beta being insignificant and the interest rate risk not priced with respect to
the short-term rates within the context of the two-factor model. A corollary is that,
when short-term interest rates are uncorrelated with the market return (i.e.,
1979–1984 and 1985–1990), the error term from Eq. 6.11 contains valuable infor-
mation leading to the inflation beta being significant. The maturity beta was found
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 199

Table 6.5 Estimating alternate proxies for the default-risk premia: Specification a : RBAA, t ¼
j0 þ j1 RLT, t þ et , Specification b : RBAA, t ¼ j0 þ j1 RLT, t þ j2 Rmt þ et
1974–1978 1979–1984 1985–1990
Specification a: generalized least square
Constant, j0 0.0021 0.0039 0.0010
(0.60)a (1.12) (0.66)
Default-risk coefficient, j1b 0.0842 0.4129 0.4734
(2.27) (12.27) (13.47)
R2 0.553 0.800 0.772
Durbin-Watson statistic 1.88 1.97 1.99
Specification a: generalized least square
Constant, j0 0.0019 0.0051 0.0006
(0.54) (1.41) (0.48)
Term structure coefficient, j1b 0.0861 0.3976 0.4637
(2.30) (11.35) (13.83)
Market linkage coefficient, j2c 0.0066 0.0491 0.0577
(0.49) (1.49) (2.39)
R2 0.555 0.806 0.785
Durbin-Watson statistic 1.87 1.97 2.00
Ordinary least square estimates were rejected because of the presence of serial correlation
indicated by Durbin-Watson statistic
RBAA,t and RLT,t are BAA corporate bond and 7-year T-note returns
a
t-statistics are in the parenthesis
b
Coefficient j1 measures the average yield differential between RBAA,t and RLT,t
c
Coefficient j2 accounts for the stock–bond market linkage

to be statistically significant in the periods 1974–1978 and 1979–1984 but not for
the period 1985–1990. Results in Table 6.4 (Specification b) showed that long-term
interest rates were correlated with the market return for the period 1985–1990 but
not for the periods 1974–1978 and 1979–1984. Hence, we posit that when long-
term interest rates are correlated with the market return (i.e., 1985–1990), the error
term from Eq. 6.12b contains no systematic information. This results in the maturity
beta being insignificant. A corollary is that, when long-term interest rates are
uncorrelated with the market return (i.e., 1974–1978 and 1979–1984), the error
term from Eq. 6.12b contains valuable information producing a significant maturity
beta and the interest rate risk is priced with respect to the long-term rates within the
context of the two-factor model. The default beta was found to be statistically
significant in the period 1985–1990 but not for the periods 1974–1978 and
1979–1984. The economic factors that lead to significant correlation between
market returns and long-term interest rates (Eq. 6.12b) or between market returns
and BAA-rated bond returns (Eq. 6.13b) caused the interest rate risk to be priced
with respect to the long-term rates within the context of the two-factor model
(1985–1990). Since the correlation between market return and interest rate changes
over time, the interest rate risk also changes over time.
200 S. Srivastava and K. Hung

Table 6.6 Multi-risk premia model of the bank stock returns: Rjt ¼ a0j + b1jRmt + b2QjDPt +
b2MjDMRt + b2DjDDRt + ejt
Estimated coefficientsa
1974–1978 1979–1984 1985–1990
Constant, a0 0.0056 0.0041 0.0047
(1.84)b (1.04) (1.31)
Market beta, b1 0.7397 0.6833 0.7185
(16.94) (10.55) (17.55)
Inflation beta, b2P 0.0698 0.1292 0.2157
(1.27) (3.45) (2.29)
Maturity beta, b2M 0.5350 0.4881 0.2476
(2.87) (2.93) (1.38)
Default beta, b2D 0.1973 0.1153 0.5696
(0.56) (0.43) (1.94)
R2 0.849 0.754 0.851
Durbin-Watson statistic 2.00 2.04 1.98
Ordinary least square estimates were rejected because of the presence of serial correlation
indicated by Durbin-Watson statistic
Portfolio beta estimated using single-index model are 0.7091, 0.7259, and 1.0102 for the periods
1974–1978, 1979–1984, and 1985–1990, respectively
a
Generalized least square estimates
b
t-statistics are in the parenthesis

Negative inflation and maturity betas for the period 1985–1990 need some
explanation because it is contrary to a priori expectation. For the period
1985–1990, the bank portfolio and market return both dropped dramatically to
6.48 % and 8.58 %, respectively. However, the estimated portfolio beta increased
from about 0.7 in the two earlier periods to about 1.0 in this period (beta estimated
independently by the single-index model). Consequently, the market factor alone
will overestimate the bank portfolio’s expected return. The negative values of
inflation beta and maturity beta (though insignificant) correct the overestimation.
Economic factors and regulatory changes that fuelled M&A activities during this
period must have been such that they increased the portfolio beta without increasing
the ex-post portfolio return. Some of these unidentifiable factors are negatively
correlated with the interest rates. One of the shortcomings of the factor analytic
approach is that factors are at times unidentifiable. In spite of difficulties in
explaining some of the results for the period 1985–1990, the multi-risk premia
model does provide greater insight into the pricing of interest rate risk.

6.6 Conclusions

In this paper, we examine the interest rate sensitivity of commercial bank returns
covering three distinct economic and regulatory environments. First, we investigate
the pricing of the interest rate risk within the framework of the two-factor model.
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 201

Our results indicate that interest rate risk was priced during 1979–1984 irrespective
of the choice of interest rate variable. However, during the periods 1974–1978 and
1985–1990, bank stock returns were sensitive only to the unexpected changes in the
long-term interest rates. Next, we tested to an asset-pricing model in which the
traditional CAPM is augmented by three additional factors to account for unex-
pected changes in the inflation, the maturity premium, and default premium. Our
results show that the inflation beta was significant for the periods 1979–1984 and
1985–1990, but not for the period 1974–1978; the maturity beta was significant for
the periods 1974–1978 and 1979–1984 but not for the period 1985–1990; and the
default beta was significant for the period 1985–1990 but not for the periods
1974–1978 and 1978–1984.
We can infer that when short-term interest rates are correlated with the
market return, the innovations in short-term interest rate are indeed white
noise. However, innovations in short-term interest rates contain valuable
information when short-term interest rates are uncorrelated with the market
return. This will lead to a significant inflation beta and the interest rate risk
will be priced with respect to the short-term rates within the context of the
two-factor model. We can also infer that when long-term interest rates are
correlated with the market return, the innovations in long-term interest rate
are indeed white noise. However, innovations in long-term interest rates
contain valuable information when long-term interest rates are uncorrelated
with the market return. This results in a significant maturity beta and priced
interest rate risk with respect to the long-term rates within the context of the
two-variable model.

Appendix 1: An Integration of CAPM and APT

The traditional CAPM is augmented by three additional APT-type principal

component factors (Johnson and Wichern 2007) to account for unexpected changes
in the inflation premium, the maturity-risk premium, and the default-risk rating.
Hence, the proposed return-generating model is written as

Rjt ¼ aj þ b1j Rmt þ b2Pj DPt þ b2Mj DMRt þ b2Dj DDRt þ ejt (6.7)

where DPt, DMRt, and DDRt are the proxies for the innovations in inflation,
maturity risk, and default risk, respectively. Coefficients b1, b2P, b2M, and b2D are
the measures of systematic market risk, inflation risk, maturity risk, and default risk,
respectively, and are consistent with APT.

Principal Component Factor Specification

An important issue in the empirical investigation of this model is the specification of

an appropriate inflation and maturity-risk and default-risk factors. Being innovations
202 S. Srivastava and K. Hung

in economic variables, these factors cannot be predicted using past information.

Hence, they must meet the following conditions:
EðDPt jt  1Þ ¼ 0
EðDMRt jt  1Þ ¼ 0 (6.8)
EðDDRt jt  1Þ ¼ 0
where E(.jt1) is the expectation based on available information at t1. Theoret-
ical consideration dictates that the choice of a factor (say DMRt) should not
influence any other factor loadings (say ∃1j). Hence, there should not be any
contemporaneous cross-correlations between the factors. Hence,
EðDPt :DMRt Þ ¼ 0
EðDPt :DDRt j ¼ 0 (6.9)
EðDRt :DDRt Þ ¼ 0

In addition, there are no common market shocks that may influence any of the
risk factors. So the following conditions

EðRmt :DPt Þ ¼ 0
EðRmt :DMRt Þ ¼ 0 (6.10)
EðRmt :DDRt Þ ¼ 0

must be satisfied. We use a generated regressor approach to construct orthogonal

inflation-risk, maturity-risk, and default-risk factors.
Economic factors that lead to changes in the equity market return also induce
term structure movements. This leads to a significant correlation between the stock
market index and the bond market index. Hence, we use the stock market return to
forecast systematic changes in the short-term interest rates. The innovations in the
short-term default-free bond return specify the unexpected changes in inflation.
Hence, the first systematic interest rate equation is written as

RST, t ¼ d0 þ d1 Rmt þ eIt (6.11)

where Rmt is the return on the stock market index and RST,t is the return on the short-
term default-free bond return. The error term in Eq. 6.11, eIt, specifies the unex-
pected change in inflation and is the generated regressor which serves as a proxy for
factor DPt. Fama and Schwert (1977) and French et al. (1983) employ similar
approaches to specify a proxy for changes in inflation.
The yield differential between short-term and long-term default-free bond
represents the slope of the yield curve. The relationship used to construct the
maturity-risk factor is given by the second systematic interest rate equation:

RLT, t ¼ y0 þ y1 RST, t þ eMt (6.12a)

6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 203

where y1 measures the average slope of the yield curve. Alternately, we will also
test the following systematic interest rate equation:

RLT, t ¼ y0 þ y1 RST, t þ y2 Rmt þ eMt (6.12b)

Inclusion of the independent variable, Rmt, will control for the stock–bond
market linkage. The estimated coefficient y1 measures the slope of the yield
curve, and it determines the maturity-risk premium on long-term assets and liabil-
ities. The error term in Eqs. 6.12a or 6.12b, eMt, specifies the unexpected change in
maturity risk and will be used as the generated regressor which serves as a proxy for
factor DMRt.
The portfolio of commercial banks used for this study consisted of money
center and large national and regional banks. Most of these banks were well
capitalized and had acquired a balanced portfolio of assets. The average
default-risk rating of the portfolio of banks should be close to the default-
risk rating of a high-grade corporate bond. Hence, the yield differential
between BAA corporate bond and long-term treasury security is used to
construct the default-risk factor. Our third systematic interest rate equation is
written as

RBAA, t ¼ j0 þ j1 RLT, t þ eDt (6.13a)

Alternately, we will also test the following systematic interest rate

equation:

RBAA, t ¼ j0 þ j1 RLT, t þ j2 Rmt þ eDt (6.13b)

where j1 measures the average default risk on BAA corporate bond. Inclusion of the
independent variable, Rmt, will control for the stock–bond market linkage. The error
term in Eqs. 6.13a or 6.13b, eDt, specifies the unexpected change in default risk
and serves as the generated regressor which will be used as the proxy for
factor DDRt.

Appendix 2: Interest Rate Innovations

An important issue in the empirical investigation of the two-factor model is the

specification of an appropriate interest rate factor. Theoretical consideration of
factor analysis requires that two factors, Rmt and RIt, be orthogonal whereby
choice of the second factor (RIt) would not influence the first factor loading
(blj). The resolution of this constraint requires a robust technique for determin-
ing the unexpected changes in the interest rate that is uncorrelated with the
market return. The three approaches to specify the interest rate factor are
presented here.
204 S. Srivastava and K. Hung

Orthogonalization Procedure

Economic factors producing changes in the market return also induced term
structure movements. This leads to a high correlation between the market factor
and the interest rate factor. Hence, market return can be used as an instrument
variable to forecast the expected changes in the interest rates. To find the unex-
pected component of interest rates, the expected interest rate is purged by
regressing RIt on Rmt and using the residuals. The systematic interest rate risk
equation is

RIt ¼ d0 þ d1 Rmt þ eit (6.14)

The residuals, eit, are the unsystematic interest rates and are used to replace RIt
in Eq. 6.14. The validity of this approach has been questioned on methodological
grounds. It is pointed out that this orthogonalization procedure produces biased
estimates of coefficients (intercept and b1j in Eq. 6.14) and that the deficiency of b2j
is not improved. On the other hand, use of an unorthogonal interest rate factor leads
to the errors-in-variable problem, i.e., the estimated coefficient b2j also captures
some of the effects responsible for changing the market factor, and hence, it is not
a true measure of the interest rate risk. Another problem using an unorthogonal
interest rate factor stems from the fact that interest rate (RIt) is usually
autoregressive. Therefore, residuals, eit, from Eq. 6.14 are autocorrelated unless
GLS parameter estimation procedure is employed. To use the GLS procedure,
a variance-covariance matrix has to be specified, which is not an easy task.

Univariate ARMA Model

The second approach is to identify and estimate an ARMA model for the interest
rate variable, RIt (Flannery and James (1984)). The unanticipated change in interest
rate from the estimated model (i.e., residuals) is used to replace RIt in Eq. 6.14. In
general, the ARMA model of order (p, q) for the univariate time series, RIt, is
written as

RIt ¼ ф1 RIt1 þ ф2 RIt2 þ . . . þ фp RItp þ m  y1 eI, t1  . . .  yq e1, tq þ eIt

(6.15)

where eIt, eI,t1, . . . are identically and independently distributed random errors
with mean zero. The ARMA procedure for the modeling of time series data is
outlined in Box and Jenkins (1976). The modeling is usually done in three steps.
First, a tentative parsimonious model is identified. Second, the parameters are
estimated, and diagnostic (Box-Pierce Q) statistics and residual auto correlation
plots are examined. The model is acceptable if the time series of residuals is white
noise and the Box-Pierce Q statistics are significant.
6 Multi-Risk Premia Model of US Bank Returns: An Integration of CAPM and APT 205

This approach leads to unbiased estimates of all the coefficients in Eq. 6.14. The
shortcoming of the univariate ARMA approach is that valuable information
contained in the stock and bond market linkage is ignored. Consequently, coeffi-
cient b2j captures some of the effects of economic factors producing stock market
changes.

Vector ARMA Model

In recent years vector autoregressive moving average (VARMA) models have

proved to be useful tools to describe the dynamic relationship between economic
variables. A vector autoregressive moving average model of order (p, q) for
k-dimensional time series, Rt ¼ (R1t, R2t, . . . Rkt)T, is generated by the following
equation:

FðBÞRt ¼ YðBÞet (6.16)

 
FðBÞ ¼ I  ф1 B  ф2 B2  . . .  фp Bp (6.17)

YðBÞ ¼ I  y1 B  y2 B2  . . .  yq Bq (6.18)

where B is the back shift operator, I is kxk unit matrix, and et is a sequence of an
independent k-dimensional vector with zero mean and positive definite covariance
matrix. The interest rate variable RIt and the market return Rmt are treated as the
components of a bivariate vector. Then vector (RIt, Rmt)T is modeled as a vector AR
process. The estimated model provides the unanticipated change in interest rate
variable to be used as the second factor in the augmented CAPM model.

References
Bae, S. C. (1990). Interest rate changes and common stock returns of financial institutions:
Revised. Journal of Financial Research, 13, 71–79.
Booth, J. R., & Officer, D. T. (1985). Expectations, interest rates, and commercial bank stocks.
Journal of Financial Research, 8, 51–58.
Box, G. E. P., & Jenkins, G. M. (1976). Time series analysis, forecasting and control. San Fran:
Holden-Day.
Chance, D. M., & Lane, W. R. (1980). A Re-examination of interest rate sensitivity in the common
stock of financial institutions. Journal of Financial Research, 3, 49–55.
Chen, N., Roll, R., & Ross, S. (1986). Economic forces and the stock market. Journal of Business,
59, 383–404.
Choi, J. J., Elyasiani, E., & Kopecky, K. (1992). The sensitivity of bank stock returns to
market, interest, and exchange rate risks. Journal of Banking and Finance, 16, 983–1004.
Christie, A. A. (1981). The stochastic behavior of common stocks variances: Value, leverage and
interest rate effects. Journal of Financial Economics, 10, 407–432.
206 S. Srivastava and K. Hung

Cox, J. C., Ingersoll, J. E., & Ross, S. A. (1985). An intertemporal general equilibrium model of
asset prices. Econometrica, 53, 363–84.
Elyasiani, E., & Iqbal, M. (1998). Sensitivity of the bank stock returns distribution to changes in
the level and volatility of interest rate: A GARCH-M model. Journal of Banking & Finance,
22, 535–563.
Fama, E. F., & Gibbons, M. R. (1982). Inflation, real returns and capital investment. Journal of
Monetary Economics, 9, 297–323.
Fama, E. F., & Schwert, W. G. (1977). Asset returns and inflation. Journal of Financial
Economics, 4, 115–146.
Flannery, M. J., & James, C. M. (1984). The effect of interest rate changes on the common stock
returns of financial institutions. Journal of Finance, 39, 1141–1153.
Flannery, M. J., Hameed, A. S., & Harjes, R. H. (1997). Asset pricing time-varying risk premia and
interest rate risk. Journal of Banking and Finance, 21, 315–335.
French, K. R., Ruback, R. C., & Schwert, G. W. (1983). Effects of nominal contracting hypothesis
on stock returns. Journal of Political Economy, 91, 70–96.
Johnson, R., & Wichern, D. (2007). Applied multivariate statistical analysis (6th ed.). Upper
Saddle River: Pearson Education.
Llyod, W. P., & Shick, R. (1977). A test of Stone’s two index models of returns. Journal of
Financial and Quantitative Analysis, 12, 363–376.
Lynge, M., & Zumwalt, J. (1980). An empirical study of interest rate sensitivity of commercial
bank returns: A multi index approach. Journal of Financial and Quantitative Analysis, 15,
731–742.
Merton, R. C. (1973). An intertemporal capital asset pricing model. Econometrica, 41, 867–887.
Song, F. (1994). A two factor ARCH model for deposit-institution stock returns. Journal of Money,
Credit, and Banking, 26, 323–340.
Srivastava, S. C., Hamid, S., & Choudhury, A. H. (1999). Stock and bond market linkage in the
empirical study of interest rate sensitivity of bank returns. Journal of Applied Business
Research, 15, 47–58.
Stone, B. (1974). Systematic interest rate risk in a two – index model of returns. Journal of
Financial and Quantitative Analysis, 9, 709–721.
Sweeney, R. J., & Warga, A. D. (1986). The pricing of interest-rate risk: Evidence from the stock
market. Journal of Finance, 41, 393–410.
Unal, H., & Kane, E. J. (1988). Two approaches to assessing the interest-rate sensitivity
of deposit-institutions’ equity returns. In C. Andrew (Ed.), Research in finance
(Vol. 7, pp. 113–132). Greenwich Conn.: Jai Press.
Woolridge, J. (2009). Introductory econometrics: A modern approach (4th ed.). Mason: Cengage
Learning.
Yourougou, P. (1990). Interest rate and the pricing of depository financial intermediary common
stock: Empirical evidence. Journal of Banking and Finance, 14, 803–320.
Nonparametric Bounds for European
Option Prices 7
Hsuan-Chu Lin, Ren-Raw Chen, and Oded Palmon

Contents
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
7.2 The Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
7.3 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
7.4 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.5 Empirical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230

Abstract
There is much research whose efforts have been devoted to discovering the
distributional defects in the Black-Scholes model, which are known to cause severe
biases. However, with a free specification for the distribution, one can only find
upper and lower bounds for option prices. In this paper, we derive a new nonpara-
metric lower bound and provide an alternative interpretation of Ritchken’s (1985)
upper bound to the price of the European option. In a series of numerical examples,
our new lower bound is substantially tighter than previous lower bounds.

The financial support of National Science Council, Taiwan, Republic of China (NSC 96-2416-H-
006-039-), is gratefully acknowledged.
H.-C. Lin (*)
Graduate Institute of Finance and Banking, National Cheng-Kung University, Tainan, Taiwan
e-mail: hsuanchu@mail.ncku.edu.tw
R.-R. Chen
Graduate School of Business Administration, Fordham University, New York, NY, USA
e-mail: rchen@fordham.edu
O. Palmon
Department of Finance and Economics, Rutgers Business School – Newark and New Brunswick,
Piscataway, NJ, USA
e-mail: palmon@business.rutgers.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 207
DOI 10.1007/978-1-4614-7750-1_7,
# Springer Science+Business Media New York 2015
208 H.-C. Lin et al.

This is prevalent especially for out-of-the-money (OTM) options where the previ-
ous lower bounds perform badly. Moreover, we present that our bounds can be
derived from histograms which are completely nonparametric in an empirical
study. We first construct histograms from realizations of S&P 500 index returns
following Chen, Lin, and Palmon (2006); calculate the dollar beta of the option and
expected payoffs of the index and the option; and eventually obtain our bounds. We
discover violations in our lower bound and show that those violations present
arbitrage profits. In particular, our empirical results show that out-of-the-money
calls are substantially overpriced (violate the lower bound).

Keywords
Option bounds • Nonparametric • Black-Scholes model • European option • S&P
500 index • Arbitrage • Distribution of underlying asset • Lower bound • Out-of-
the-money • Kernel pricing

7.1 Introduction

In a seminal paper, Merton (1973) presents for the first time the no-arbitrage bounds
of European call and put options. These bounds are nonparametric and do not rely on
any assumption.1 Exact pricing formulas such as the Black and Scholes (1973) model
and its variants, on the other hand, rely on strong assumptions on the asset price
process and continuous trading. Due to the discreteness of actual trading opportuni-
ties, Perrakis and Ryan (1984) point out that option analyses in continuous time limit
the accuracy and applicability of the Black-Scholes and related formulas. Relying on
Rubinstein’s (1976) approach, the single-price law, and arbitrage arguments, they
derive upper and lower bounds for option prices with both a general price distribution
and discrete trading opportunities. Their lower bound is tighter than that of Merton.
Levy (1985) applies stochastic dominance rules with borrowing and lending at the
risk-free interest rate to derive upper and lower option bounds for all unconstrained
utility functions and alternatively for concave utility functions. The derivation of these
bounds can be applied to any kinds of stock price distribution as long as the stock is
“nonnegative beta,” which is identical to the assumption of Perrakis and Ryan (1984).
Moreover, Levy claims that Perrakis and Ryan’s bounds can be obtained by applying
the second-degree stochastic dominance rule. However, Perrakis and Ryan do not
cover all possible combinations of the risky asset with the riskless asset, and their
bounds are therefore wider than those of Levy. Levy also applies the first-degree
stochastic dominance rule (FSDR) with riskless assets to prove that Merton’s bounds
are in fact FSDR bounds and applies the second-degree stochastic dominance rule to
strengthen Merton’s bounds on the option value. At the same time, Ritchken (1985)
uses a linear programming methodology to derive option bounds based on primitive
prices in incomplete markets and claims that his bounds are tighter than those of
Perrakis and Ryan (1984).

1
The only assumption is that both option and its underlying stock are traded securities.
7 Nonparametric Bounds for European Option Prices 209

With an additional restriction that the range of the distribution of the one-period
returns per dollar invested in the optioned stock is finite and has a strictly positive
lower limit, Perrakis (1986) extends Perrakis and Ryan (1984) to provide bounds for
American options. Instead of assuming that no opportunities exist to revise positions
prior to expiration in Levy (1985) and Ritchken (Ritchken 1985), Ritchken and Kuo
(1988) obtain tighter bounds on option prices under an incomplete market by
allowing for a finite number of opportunities to revise position before expiration
and making more restrictive assumptions on probabilities and preferences. The
single-period linear programming option model is extended to handle multiple
periods, and the stock price is assumed to follow a multiplicative multinomial
process. Their results show that the upper bounds are identical to those of Perrakis,
while the lower bounds are tighter. Later, Ritchken and Kuo (1989) also add suitable
constraints to a linear programming problem to derive option bounds under higher
orders of stochastic dominance preferences. Their results show that while the upper
bounds remain unchanged beyond the second-degree stochastic dominance, the
lower bounds become sharper as the order of stochastic dominance increases.2
Claiming that Perrakis and Ryan (1984), Levy (1985), Ritchken (1985), and
Perrakis (1986) are all parametric models, Lo (1987) derives semi-parametric
upper bounds for the expected payoff of call and put options. These upper bounds
are semi-parametric because they depend on the mean and variance of the stock price
at maturity but not on its entire distribution. In addition, the derivation of
corresponding semi-parametric upper bounds for option prices is shown by adopting
the risk-neutral pricing approach of Cox and Ross (1976).3 To continue the work of
Lo (1987), Zhang (1994) and De La Pena et al. (2004), both of which assume that the
underlying asset price must be continuously distributed, sharpen the upper option
bounds of Lo (1987). Boyle and Lin (1997) extend the results of Lo (1987) to
contingent claims on multiple underlying assets. Under an intertemporal setting,
Constantinides and Zariphopoulou (2001) derive bounds for derivative prices with
proportional transaction costs and multiple securities. Frey and Sin (1999) examine
the sufficient conditions of Merton’s bounds on European option prices under random
volatility. More recently, Gotoh and Konno (2002) use the semi-definite programming
and a cutting plane algorithm to study upper and lower bounds of European call option
prices. Rodriguez (2003) uses a nonparametric method to derive lower and upper

2
To further explore the research work of Ritchken and Kuo (1989) under the decreasing absolute
risk aversion dominance rule, Basso and Pianca (1997) obtain efficient lower and upper option
pricing bounds by solving nonlinear optimization problem. Unfortunately, neither model provides
enough information of their numerical examples for us to compare our model with. The Ritchken-
Kuo model provides no Black-Scholes comparison, and the Basso-Pianca model provides only
some partial information on the Black-Scholes model (we find the Black-Scholes model under 0.2
volatility to be 13.2670 and under the 0.4 volatility to be 20.3185, which are different from what
are reported in their paper (12.993 and 20.098, respectively)) which is insufficient for us to provide
any comparison.
3
Inspired by Lo (1987), Grundy (1991) derives semi-parametric upper bounds on the moments of
the true, other than risk-neutral, distribution of underlying assets and obtains lower bounds by
using observed option prices.
210 H.-C. Lin et al.

bounds and contributes a new tighter lower bound than previous work. Huang (Huang
2004) puts restrictions on the representative investor’s relative risk aversion and
produces a tighter call option bound than that of Perrakis and Ryan (1984). Hobson
et al. (2005) derive arbitrage-free upper bounds for the prices of basket options. Peña
et al. (2010) conduct static-arbitrage lower bounds on the prices of basket options via
linear programming. Broadie and Cao (2008) introduce new and improved methods
based on simulation to obtain tighter lower and upper bounds for pricing American
options. Lately, Chung et al. (2010) also use an exponential function to approximate
the early exercise boundary to obtain tighter bounds on American option prices.
Chuang et al. (2011) provide a more complete review and comparison of theoretical
and empirical development on option bounds.4
In this paper we derive a new and tighter lower bound for European option prices
under a nonparametric framework. We show that Ritchken’s (1985) upper bound is
consistent with our nonparametric framework. Both bounds are nonparametric
because the price distribution of underlying asset is totally flexible, can be arbitrarily
chosen, and is consistent with any utility preference.5 We compare our lower bound
with those in previous studies and show that ours dominate those models by a wide
margin. We also present the lower bound result on the model with random volatility
and random interest rates (Bakshi et al. 1997; Scott 1997) to demonstrate how easily
our model can be made consistent with any parametric structure.6 Finally, we
present how our bounds can be derived from histograms which are nonparametric
in an empirical study. We discover violations of our lower bound and show that those
violations present arbitrage profits.7 In particular, our empirical results show that
out-of-the-money calls are substantially overpriced (violate the lower bound).

7.2 The Bounds

A generic and classical asset pricing model with a stochastic kernel is


S t ¼ Et M t , T S T , (7.1)

where Mt,T is the marginal rate of substitution, also known as the pricing kernel
that discounts the future cash flow at time T; Et[.] is the conditional expectation

4
Since our paper only provides a nonparametric method on examining European option bounds,
our literature review is much limited. For a more complete review and comparison on prior studies
of option bounds, please see Chuang et al. (2011).
5
Christoffersen et al. (2010) provide results for the valuation of European-style contingent claims
for a large class of specifications of the underlying asset returns.
6
Given that our upper bound turns out to be identical to Ritchken’s (1985), we do not compare with
those upper bound models that dominate Ritchken (e.g., Huang (2004), Zhang (1994) and De La
Pena et al. (2004)). Also, we do not compare our model with those models that require further
assumptions to carry out exact results (e.g., Huang (2004) and Frey and Sin (1999)), since it is
technically difficult to do.
7
For the related empirical studies of S&P 500 index options, see Constantinides et al. (2009, 2011).
7 Nonparametric Bounds for European Option Prices 211

under the physical measure ℙ taken at time t; and St is the value of an arbitrary
asset at time t. The standard kernel pricing theory (e.g., Ingersoll (1989))
demonstrates that


Dt, T ¼ Et Mt, T , (7.2)

where Dt,T is the risk-free discount factor that gives the present value of $1 over the
period (t,T). The usual separation theorem gives rise to the well-known risk-neutral
pricing result:


St ¼ Et Mt, T St

 ðT Þ
¼ Et Mt, T E^t ½ST , (7.3)
ðT Þ
¼ Dt, T E^t ½ST 

If the risk-free interest rate is stochastic, then E^t(T)[.] is the conditional expec-
tation under the T-forward measure P ^ ðT Þ . When the risk-free rate is non-stochastic,
then the forward measure reduces to the risk-neutral measure P ^ and will not depend
ðT Þ ^
upon maturity time, i.e., Et ½ ! E t ½. 8

Note that Eqs. 7.1 and 7.3 can be applied to both the stock and the option prices.
This leads to the following theorem which is the main result of this paper.
Theorem 7.1 The following formula provides a lower bound for the European call
option Ct:
 
Ct ¼ Dt, T Et ½CT  þ bC St  Dt, T Et ½ST  (7.4)
cov½CT ; ST 
where bC ¼ var½ST  :
Proof By Eq. 7.1, the option price must follow Ct ¼ Et[Mt,T CT], and hence


Ct ¼ Et Mt, T CT



¼ Et Mt, T Et ½CT  þ cov Mt, T , CT (7.5)


¼ Dt, T Et ½CT  þ cov Mt, T , CT
or


Ct  Dt, T Et ½CT  ¼ cov Mt, T ; CT : (7.6)

Similarly,


St  Dt, T Et ½ST  ¼ cov Mt, T ; ST : (7.7)

8
Without loss of generality and for the ease of exposition, we take non-stochastic interest rates and
proceed with the risk-neutral measure P ^ for the rest of the paper.
212 H.-C. Lin et al.

Hence, to prove Eq. 7.4, we only need to prove



cov Mt, T ; CT  bC cov Mt, T ; ST : (7.8)

Note that cov[Mt,T,CT] ¼ cov[Mt,T, max{ST  K, 0}] is monotonic in strike

price K and has a minimum value when K ¼ 0 in which case cov[Mt,T,CT] ¼ cov
[Mt,T,ST] and a maximum value when K!1 in which case cov[Mt,T,
CT] ¼ 0. Hence, cov[Mt,T,ST]  cov[Mt,T,CT] which is less than 0. Note that
0 < bC < 1 (see the proof in the following Corollary). In Appendix 1, it is proved
that cov[Mt,T,CT]  bC cov[Mt,T,ST].
The put lower bound takes the same form and is provided in the following corollary:
Corollary 7.1 The lower bound of the European put option Pt can be obtained by the
put-call parity and satisfies the same functional form in Theorem 7.1:
 
Pt ¼ Dt, T Et ½PT  þ bP St  Dt, T Et ½ST  (7.9)
cov½PT ; ST 
where bP ¼ var½ST  :
[Proof] By the Put-Call Parity:

Pt ¼ Ct þ Dt, T K  St  Ct þ Dt, T K  St ¼ Pt : (7.10)

We then substitute in the result of Theorem 7.1 to get

 
Pt ¼ Dt, T Et ½CT  þ bC St  Dt, TEt ½ST  þ Dt, T K St
¼ Dt, T Et ½ST þ PT  K  þ bC S t   Dt, T Et ½ST  þ Dt, T K  St
¼ Dt, T Et ½PT  þ bP St  Dt, T Et ½ST  (7.11)

where bP ¼ bC  1. Note that bP < 0 < bC. This also implies that bc<1. Finally, it is
straightforward to show that bP ¼ covvar½P½ST T;ST  using the put-call parity.
Therefore, for both call and put options, since the relationship between the
pricing kernel M and stock price S is convex, the theorem provides a lower
bound.9 Merton (1973) shows that the stock price will be the upper bound for
the call option and the strike price should be the upper bound for the put
option; otherwise, arbitrage should occur. Ritchken (1985) provides a tighter
upper bound10 than that of Merton, which is stated in the following theorem,
although we provide an alternative proof to Ritchken that is consistent with our
derivation of the lower bound.

In the Appendix, e > 0.

9

10
Perrakis and Ryan (1984) and Ritchken (1985) obtain the identical upper bound.
7 Nonparametric Bounds for European Option Prices 213

Theorem 7.2 The following formulas provide upper bounds for the European call
and put options (Ritchken (1985))11:

St
Ct ¼ Et ½CT 
Et ½ST    (7.12)
St St
Pt ¼ Et ½PT  þ Dt, T  K:
Et ½ST  E t ½ ST 

Proof Similar to the proof of the lower bound, the upper bound of the call option is
provided as follows:


St Et Mt, T ST
Et ½CT  ¼ Et ½CT 
Et ½ST  E t ½ ST 



Et Mt, T ST CT  cov Mt, T ST , CT
¼
Et ½ST 


Et Mt, T ST CT (7.13)
>
Et ½ S T 
ðCÞ
Et ½ST 
¼ Ct
Et ½ S T 
> Ct :

The third line of the above equation is a result from the fact that cov[Mt,TST,
CT] < 0. The fourth line of the above equation is a change of measure with the call
option being the numeraire. The last line of the above equation is a result based upon
t [ST]/Et[ST] > 1.
E(C) 12

By the put-call parity, we can show that the upper bound of the put option
requires an additional term:
St
Pt ¼ Ct þ Dt, T K  St ¼ Et ½PT þ ST  K  þ Dt, T K  St
Et ½ S T 
  (7.14)
St St
¼ Et ½PT  þ Dt, T  K:
Et ½ S T  Et ½ S T 
The lower and upper bounds we show in this paper have two important advantages
over the existing bounds. The bounds will converge to the true value of the option if:
• The expected stock return, EtS½St T , approaches the risk-free rate.
• The correlation between the stock and the call or put option (rSC or rSP)
approaches 1 or 1.

11
This is same as Proposition 3-i (Eq. 7.26) in Ritchken (1985).
By the definition of measure change, we have Et[CTST] ¼ Et[CT]E(C)
12 (C)
t [ST] which implies Et [ST]/
E[ST] ¼ Et[CTST]/{Et[CT]Et[ST]} > 1.
214 H.-C. Lin et al.

These advantages help us identify when the bounds are tight and when they are
not. The first advantage indicates that the bounds are tight for low-risk stocks and
not tight for high-risk stocks. The second advantage indicates that the bounds are
tighter for in-the-money options than out-of-the-money options.

7.3 Comparisons

The main purpose of this section is to compare our lower bound to several
lower bounds in previous studies, namely, Merton (1973), Perrakis and Ryan
(1984), Ritchken (1985), Ritchken and Kuo (1988), Gotoh and Konno (2002),
and Rodriguez (2003) using the Black-Scholes model as the benchmark for its
true option value. We also compare Ritchken’s upper bound (which is also our
upper bound) with more recent works by Gotoh and Konno (2002) and
Rodriguez (2003).
The Black-Scholes model has five variables: stock price, strike price, volatility
(standard deviation), risk-free rate (constant), and time to maturity. In addition to
the five variables, the lower bound models need the physical expected stock return.
The following is the base case for the comparison:

Current stock S0 50
Strike K 50
Volatility s 0.2
Risk-free rate r 0.1
Time to maturity T 1
Stock expected return m 0.2

In the Black-Scholes model, stock price (S) evolution follows a log normal
process:

dS ¼ mSdt þ sSdW (7.15)

where instantaneous expected rate of stock return m and volatility of stock price s
are assumed to be constants and where dW is a wiener process. The call option price
is computed as

C ¼ S0 N ðdþ Þ  erT KN ðd Þ (7.16)

where

ln S0  ln K þ ðr  0:5s2 ÞT
d ¼ pffiffiffi :
s T
7 Nonparametric Bounds for European Option Prices 215

The lower bound is computed by simulating the stock price using Eq. 7.15 via
a binomial distribution approximation:

Sj ¼ S0 uj dnj : (7.17)

As n approaches infinity, Sj approaches the log normal distribution, and the

binomial model converges to the Black-Scholes model. Under the risk-neutral
measure, the probability associated with the jth state is set as
 
b ½j ¼ n j
Pr p^ ð1  p^Þnj (7.18)
j

where

erDt  d
p^ ¼ ,
ud
pffiffiffiffi
u ¼ es Dt
,
pffiffiffiffi
d ¼ es Dt
,

and Dt ¼ T/n represents the length of the partition. Under the actual measure, the
formula will change to

 
n j
Pr½j ¼ p ð1  pÞnj (7.19)
j

where

emDt  d
p¼ :
ud

Finally, the pricing kernel in our model is set as

b ½j rT
Pr
M0T ½j ¼ e : (7.20)
Pr½j

In our results, we let n be great enough so that the binomial model price is 4-digit
accurate to the Black-Scholes model. We hence set n to be 1,000. The results are
reported in Table 7.1. The first panel presents the results for various moneyness
levels, the second panel presents the results for various volatility levels, the third
panel presents the results for various interest rates, the fourth panel presents the
results for various maturity times, and the last presents the results for various stock
216 H.-C. Lin et al.

expected returns. In general, the lower bounds are tighter (for all models) when the
moneyness is high (in-the-money), volatility is high, risk-free rate is high, time to
maturity is short, and the expected return of stock is low.
This table presents comparisons between our lower bound and existing lower
bounds by Merton (1973), Perrakis and Ryan (1984), and Ritchken (1985). The
base case parameter values are:
• Stock price ¼ 50
• Strike price ¼ 50
• Volatility ¼ 0.2
• Risk-free rate ¼ 10 %
• Time to maturity ¼ 1 year
• Stock expected return (m) ¼ 20 %
The highlighted rows represent the base case.
As we can easily see, universally our model for the lower bound is tighter than any
of the comparative models. One result particularly worth mentioning is that our lower
bound performs better than the other lower bound models in out-of-the-money
options. For example, our lower bound is much better than Ritchken’s (1985)
lower bound when the option is 20 % out-of-the-money and continues to show
value when Ritchken’s lower bound returns to 0 (see the first panel of Table 7.1).
While Ritchken and Kuo (1988) claim to obtain tighter lower bounds than
Perrakis (1986) and Perrakis and Ryan (1984), they do not show direct comparisons
in their paper. Rather, they present through a convergence plot (Figure 3 on page
308 in Ritchken and Kuo (1988)) of a Black-Scholes example with the true value
being $5.4532 and the lower bound approaching roughly $5.2. The same parameter
values with our lower bound show a lower bound of $5.4427, which demonstrates
a substantial improvement over the Ritchken and Kuo model.
The comparisons with more recent studies of Gotoh and Konno (2002) and
Rodriguez (2003) are given in Tables 7.2 and 7.3.13 Gotoh and Konno use
semi-definite programming and a cutting plane algorithm to study upper and lower
bounds of European call option prices. Rodriguez uses a nonparametric method to
derive lower and upper bounds. As we can see in Tables 7.2 and 7.3, except for very
few upper bound cases, none of the bounds under the Gotoh and Konno’s model and
Rodriguez’s model are very tight, compared to our model. Furthermore, note that our
model requires no moments of the underlying distribution.14
The base case parameter values are:
• Stock price ¼ 40
• Risk-free rate ¼ 6 %
• Stock expected return (m) ¼ 20 %

13
We also compare with the upper bound by Zhang (1994), which is an improved upper bound by
Lo (1987), and show overwhelming dominance of our upper bound. The results (comparison to
Tables 7.1, 7.2, and 7.3 in Zhang) are available upon request.
14
The upper bounds by the Gotoh and Konno model perform well in only in-the-money, short
maturity, and low volatility scenarios, and these scenarios are where the option prices are close to
their intrinsic values, and hence the percentage errors are small.
7 Nonparametric Bounds for European Option Prices 217

Table 7.1 Lower bound comparison

S Blk-Sch Merton PR Ritch. Our $error %error

20 0.0000 0 0 0 0 0.0000
25 0.0028 0 0 0 0 0.0028
30 0.0533 0 0 0 0 0.0533
35 0.3725 0 0 0 0.0008 0.3717 99.78
40 1.3932 0 0 0.272 0.8224 0.5708 40.97
45 3.4746 0 1.441 2.278 2.8957 0.5789 16.66
50 6.6322 4.758 5.449 5.791 6.1924 0.4398 6.63
55 10.6248 9.758 10.017 10.143 10.3535 0.2714 2.55
60 15.1288 14.758 14.849 14.888 14.9851 0.1437 0.95
65 19.9075 19.758 19.788 19.797 19.8395 0.0680 0.34
70 24.8156 24.758 24.768 24.767 24.7860 0.0296 0.12
75 29.7794 29.758 29.761 29.76 29.7673 0.0121 0.04
80 34.7658 34.758 34.759 34.754 34.7611 0.0047 0.01

Vol Blk-Sch Merton PR Ritch. Our $error %error

0.1 5.1526 4.7580 4.7950 4.8630 4.8930 0.2596 5.04
0.15 5.8325 4.7580 5.0290 5.2400 5.4367 0.3958 6.79
0.2 6.6322 4.7580 5.4490 5.7890 6.1924 0.4398 6.63
0.25 7.4847 4.7580 5.9700 6.4360 7.0247 0.4600 6.15
0.3 8.3633 4.7580 6.5370 7.1010 7.8864 0.4769 5.70
0.35 9.2555 4.7580 7.1180 7.7760 8.7601 0.4954 5.35
0.4 10.1544 4.7580 7.6990 8.4570 9.6382 0.5162 5.08
0.45 11.0559 4.7580 8.2680 9.1250 10.5170 0.5389 4.87
0.5 11.9574 4.7580 8.8220 9.7780 11.3945 0.5629 4.71
0.55 12.8569 4.7580 9.3570 10.4240 12.2692 0.5877 4.57

Rate Blk-Sch Merton PR Ritch. Our $error %error

0.02 4.4555 0.9901 1.7380 2.7680 3.0243 1.4313 32.12
0.04 4.9600 1.9605 2.6940 3.5010 3.8402 1.1198 22.58
0.06 5.4923 2.9118 3.6310 4.2490 4.6400 0.8523 15.52
0.08 6.0504 3.8442 4.5490 5.0200 5.4240 0.6264 10.35
0.1 6.6322 4.7581 5.4490 5.7970 6.1924 0.4398 6.63
0.12 7.2355 5.6540 6.3310 6.5810 6.9456 0.2900 4.01
0.14 7.8578 6.5321 7.1960 7.3670 7.6839 0.1739 2.21
0.16 8.4965 7.3928 8.0430 8.1470 8.4076 0.0889 1.05
0.18 9.1488 8.2365 8.8740 8.9350 9.1169 0.0319 0.35
0.2 9.8122 9.0635 9.6880 9.7170 9.8122 0.0000 0.00
0.22 10.4841 9.8741 10.4870 10.4830 10.4841 0.0000 0.00

T Blk-Sch Merton PR Ritch. Our $error %error

0.1 1.5187 0.4975 1.2850 1.3600 1.4976 0.0211 1.39
0.2 2.3037 0.9901 1.8870 2.0240 2.2469 0.0568 2.47
0.3 2.9693 1.4777 2.4000 2.5940 2.8696 0.0997 3.36
0.4 3.5731 1.9605 2.8730 3.0990 3.4266 0.1465 4.10
0.5 4.1371 2.4385 3.3250 3.5830 3.9416 0.1955 4.72
0.6 4.6726 2.9118 3.7640 4.0510 4.4272 0.2454 5.25
0.7 5.1862 3.3803 4.1940 4.5020 4.8909 0.2953 5.69

(continued)
218 H.-C. Lin et al.

Table 7.1 (continued)

0.8 5.6822 3.8442 4.6170 4.9460 5.3375 0.3447 6.07
0.9 6.1635 4.3034 5.0350 5.3710 5.7705 0.3930 6.38
1 6.6322 4.7581 5.4490 5.7890 6.1924 0.4398 6.63

Mu (µ) Blk-Sch Merton PR Ritch. Our $error %error

0.1 6.6322 4.7580 6.3740 6.4310 6.6322 0.0000 0.00
0.15 6.6322 4.7580 5.8380 6.0610 6.4703 0.1620 2.44
0.2 6.6322 4.7580 5.4490 5.7890 6.1924 0.4398 6.63
0.25 6.6322 4.7580 5.1800 5.5820 5.8689 0.7633 11.51
0.3 6.6322 4.7580 5.0040 5.4360 5.5572 1.0750 16.21
0.35 6.6322 4.7580 4.8940 5.3090 5.2936 1.3386 20.18
0.4 6.6322 4.7580 4.8300 5.2130 5.0929 1.5393 23.21
0.45 6.6322 4.7580 4.7940 5.1330 4.9536 1.6786 25.31

Note: S is the stock price; vol is the volatility; rate is the risk-free rate; Mu (m) is the expected rate
of return of stock; Blk-Sch is the Black-Scholes (1973) solution; Merton is the Merton (1973)
model; PR is Perrakis and Ryan (1984) model; Ritch. is the Ritchken (1985) model; Our is our
model; $error is error in dollar; %error is error in percentage

The base case parameter values are:

• Strike price ¼ 50
• Volatility ¼ 0.2
• Risk-free rate ¼ 10 %
• Time to maturity ¼ 1 year
• Stock expected return (m) ¼ 20 %

7.4 Extensions

In addition to a tight lower bound, another major contribution of our model is

that it makes no assumption on the distribution of the underlying stock (unlike Lo
(1984) and Gotoh and Konno (2002) who require moments of the underlying distri-
bution) or any assumption on interest rates and volatility (unlike Rodriguez (2003)
who requires constant interest rates). As a result, our lower bound can be used with
models that assume random volatility and random interest rates or any arbitrary
specification of the underlying stock. Note that our model needs only the dollar beta
of the option and expected payoffs of the stock and the option.15 In this section, we
extend our numerical experiment to a model with random volatility and random
interest rates.
Option models with random volatility and random interest rates can be derived
with closed form solutions under the Scott (1997) and Bakshi et al. (1997)

15
The term “dollar beta” is originally from Page 173 of Black (1976). Here we mean bc and br.
7 Nonparametric Bounds for European Option Prices 219

Table 7.2 Comparison of upper and lower bounds with the Gotoh and Konno (2002) model
Lower bound Upper bound
Stk Our GK Blk-Sch Our GK
S ¼ 40; rate ¼ 6 %; vol ¼ 0.2; t ¼ 1 week
30 10.0346 10.0346 10.0346 10.1152 10.0349
35 5.0404 5.0404 5.0404 5.1344 5.0428
40 0.4628 0.3425 0.4658 0.5225 0.5771
45 0.0000 0.0000 0.0000 0.0000 0.0027
50 0.0000 0.0000 0.0000 0.0000 0.0003
S ¼ 40; rate ¼ 6 %; vol ¼ 0.8; t ¼ 1 week
30 10.0400 10.0346 10.0401 10.1202 10.1028
35 5.2644 5.0404 5.2663 5.3483 5.4127
40 1.7876 1.2810 1.7916 1.8428 2.2268
45 0.3533 0.0015 0.3548 0.3717 0.5566
50 0.0412 0.0000 0.0419 0.0444 0.1021
S ¼ 40; rate ¼ 6 %; vol ¼ 0.8; t ¼ 12 week
30 11.9661 10.4125 12.0278 12.7229 12.8578
35 8.7345 6.2980 8.8246 9.4774 9.7658
40 6.2141 3.8290 6.3321 6.8984 7.5165
45 4.3432 2.5271 4.4689 4.9421 6.8726
50 2.9948 1.5722 3.1168 3.4990 4.5786
Note: S is the stock price; Stk is the strike price; vol is the volatility; rate is the risk-free rate; Blk-
Sch is the Black-Scholes (1973) solution; GK is the Gotoh and Konno (2002) model; Our is our
model; $error is error in dollar; %error is error in percentage

specifications. However, here, given that there is no closed form solution to the
covariance our model requires, we shall use Monte Carlo to simulate the lower
bound. In order to be consistent with the lower bound, we must use the same Monte
Carlo paths for the valuation of the option. For the ease of exposition and simplicity,
we assume the following joint stochastic processes of stock price S, interest rate r,
and volatility V under the actual measure, respectively:
pffiffiffiffi
dS ¼ mSdt þ V SdW ^2
^2
dr ¼ aðy  r Þdt þ vd W (7.21)
dV ¼ Vd W^2

where dW is a wiener process, dWidWj ¼ 0, m, a, y, u, and  are constants. The

processes under the actual measure are used for simulating the lower and upper
bounds. The Monte Carlo simulations are performed under the risk-neutral measure
in order to compute the option price:
pffiffiffiffi
dS ¼ rSdt þ ^1
V Sd W
^2
dr ¼ aðy  r Þdt þ vd W (7.22)
dV ¼ Vd W^ 3:
220 H.-C. Lin et al.

Table 7.3 Comparison of upper and lower bounds with the Rodriguez (2003) model
Lower bound Upper bound
S Our Rodriguez Blk-Sch Our Rodriguez
30 0.0221 0 0.0538 0.1001 0.1806
32 0.0725 0.0000 0.1284 0.2244 0.3793
34 0.1828 0.0171 0.2692 0.4451 0.7090
36 0.3878 0.1158 0.5072 0.7973 1.2044
38 0.7224 0.3598 0.8735 1.3100 1.8900
40 1.2177 0.7965 1.3950 2.0044 2.7767
42 1.8982 1.4521 2.0902 2.8927 3.8619
44 2.7711 2.3329 2.9676 3.9697 5.1315
46 3.8319 3.4286 4.0255 5.2211 6.5640
48 5.0709 4.7177 5.2535 6.6302 8.1337
50 6.4703 6.1724 6.6348 8.1753 9.8149
52 8.0072 7.7635 8.1494 9.8327 11.5835
54 9.6574 9.4631 9.7758 11.5794 13.4187
56 11.3974 11.2462 11.4933 13.3950 15.3032
58 13.2067 13.0922 13.2832 15.2621 17.2235
60 15.0683 14.9845 15.1292 17.1674 19.1693
62 16.9716 16.9101 17.0179 19.1024 21.1328
64 18.9035 18.8593 18.9384 21.0573 23.1086
66 20.8559 20.8250 20.8822 23.0262 25.0926
68 22.8239 22.8020 22.8429 25.0056 27.0822
70 24.8018 24.7867 24.8157 26.9915 29.0754
Note: S is the stock price; Blk-Sch is the Black-Scholes (1973) solution; Rodriguez is the
Rodriguez (2003) model; Our is our model; $error is error in dollar; %error is error in percentage

To simplify the problem without loss of generosity, we assume that investors

^2
charge no risk premiums on interest rate risk and volatility risk, i.e., dW 2 ¼ dW
^
and dW 3 ¼ dW 3 .
The simulations are done by the following integrals under the actual (top
equation) and risk-neutral (bottom equation) measures:
ð t ðt ðt 
pffiffiffiffiffiffi
St ¼ S0 exp mu du  1⁄2V u du þ V u dW u
0 0 0
ð t ðt ðt  (7.23)
pffiffiffiffiffiffi
S^t ¼ S0 exp r u du  1
⁄2V u du þ ^
V u dW u :
0 0 0

The no-arbitrage price of the call option is computed under the risk-neutral
measure as

 ðt 

 
^
C ¼ E exp  r u du max S t  K, 0 : (7.24)
0
7 Nonparametric Bounds for European Option Prices 221

Table 7.4 Lower bound under the random volatility and random interest rate model
S BCC/Scott Our $error %error
25 1.9073 1.1360 0.7713 40.44
30 3.2384 2.4131 0.8253 25.49
35 5.2058 4.1891 1.0167 19.53
40 7.5902 6.5864 1.0039 13.23
45 10.2962 9.5332 0.7630 7.41
50 13.6180 12.8670 0.7510 5.52
55 17.1579 16.5908 0.5671 3.30
60 21.1082 20.5539 0.5543 2.63
65 25.0292 24.7251 0.3040 1.21
70 29.3226 29.0742 0.2484 0.85
Note: S is the stock price; BCC/Scott are Bakshi et al. (1997) and Scott (1997) models; Our is our
model; $error is error in dollar; %error is error in percentage

The bounds are computed under the actual measure. For example,

E½Ct  ¼ E½maxfSt  K, 0g: (7.25)

Given that dWidWj ¼ 0, we can simulate the interest rates and volatility
separately and then simulate the stock price. That is, conditional on known interest
rates and volatility, under independence, the stock price is log normally distributed.
We perform our simulations using 10,000 paths over 52 weekly periods. The
parameters are given as follows:

Strike K 50
Time to maturity T 1
Stock expected return m 0.2
Reverting speed a 0.5
Reverting level y 0.1
Interest rate volatility u 0.03
Initial interest rate r0 0.1
Initial variance V0 0.0416
Volatility on variance  0.2

Note that implicitly we assume the price of risk for both interest rate process and
volatility process to be 0, for simplicity and without loss of generality. The results
are shown in Table 7.4. Compared to the model of the Black-Scholes (i.e., the first
panel of Table 7.4), the lower bound performs similarly in the random volatility
and random interest rate model. Take the base case as an example where the
Black-Scholes price is 6.6322, the Bakshi-Cao-Chen/Scott price is 13.6180 as
a result of extra uncertainty in the stock price due to random volatility and interest

16
This is so because the initial volatility is 0.2.
222 H.-C. Lin et al.

rates. The error of the lower bound of our model is 0.7510 in the Bakshi-Cao-Chen/
Scott case as opposed to 0.4398 in the Black-Scholes case.17 The percentage error is
5.52 % in the Bakshi-Cao-Chen/Scott case versus 6.63 % in the Black-Scholes case.
The in-the-money options have larger percentage errors than those of the out-of-
the-money options.
The parameters are given as follows:

Strike price 50
Time to maturity 1
Stock expected return 0.2
Reverting speed 0.5
Reverting level 0.1
Interest rate volatility 0.03
Price of risk 0
Initial interest rate 0.1
Initial volatility 0.2
Volatility on volatility 0.2
The Monte Carlo paths are 10,000. The stock price, volatility, and interest rate processes are
assumed to be independent

7.5 Empirical Study

In this section, we test the lower and upper bounds against data. Charles Cao has
generously provided us with the approximated prices of S&P 500 index call option
contracts, matched levels of S&P 500 index, and approximated risk-free 90-day
T-Bill rates for the period of June 2, 1988 through December 31, 1991.18 For each
day, the approximated option prices are calculated as the average of the last bid and
ask quotes. Index returns are computed using daily closing levels for the S&P
500 index that are collected and confirmed using data obtained from Standard and
Poor’s, CBOE, Yahoo, and Bloomberg.19
The dataset contains 46,540 observations over 901 days (from June 2, 1988, to
December 31, 1991). Hence, on average there are over 50 options for various
maturities and strikes. The shortest maturity of the dataset is 7 days and the longest
is 367 days. 15 % of the data are less than 30 days to maturity, 32 % are between

17
This Black-Scholes case is from the highlighted row in the first panel of Table 7.1.
18
The data are used in Bakshi et al. (1997).
19
The (ex-dividend) S&P 500 index we use is the index that serves as an underlying asset for the option.
For option evaluation, realized returns of this index need not be adjusted for dividends unless the timing
of the evaluated option contract is correlated with lumpy dividends. Because we use monthly
observations, we think that such correlation is not a problem. Furthermore, in any case, this should
not affect the comparison of the volatility smile between our model and the Black-Scholes model.
7 Nonparametric Bounds for European Option Prices 223

30 and 60 days, 30 % are between 60 and 180 days, and 24 % are more than
180 days to maturity. Hence, these data do not have maturity bias.
The deepest out-of-the-money option is 18.33 %, and the deepest
in-the-money option is 47.30 %. Roughly half of the data are at-the-money options
(46 % of the data are within 5 % in-the-money and out-of-the-money). 10 % are
deep-in-the-money (more than 15 % in-the-money), but less than 1 % of the data are
deep-out-of-the-money (more than 15 % out-of-the-money). Hence, the data have
disproportional fraction of in-the-money options. This is clearly a reflection of the
bull market in the sample period.
The best way to test the lower and upper bounds derived in this paper is to use
a nonparametric, distribution-free model. Note that the lower bound in Theorem 7.1
requires only the expected return of the underlying stock and the covariance
between the stock and the option. There is no further requirement for the lower
and the upper bounds. In other words, our lower bound model can permit any
arbitrary distribution of the underlying stock and any parametric specification of the
underlying stock such as random volatility, random interest rates, and jumps.
Hence, to best test the bounds with a parsimonious empirical design, we adopt
the histogram method introduced by Chen et al. (2006) where the underlying asset
is modelled by past realizations, i.e., histogram.
We construct histograms from realizations of S&P 500 index (SPX) returns. We
calculate the price on day t of an option that settles on day T using a histogram of
S&P 500 index returns for a holding period of T–t, taken from a 5-year window
immediately preceding time t.20 For example, an
x-calendar-day
 option price on
any date is evaluated using a histogram of round 252365 x -trading-day holding period
returns where round [.] is rounding the nearest integer.21 The index levels used to
calculate these returns are taken from a window that starts on the 1260th ( 5
252) trading day before the option trading date and
ends  1 day before the trading
date. Thus, this histogram contains 1, 260  round 252 365 x -trading-day return reali-
zations. Formally, we compute histogram of the (unannualized) returns by the
following equation:

Rt, tþx, i ¼ ln Stix  ln Sti , (7.26)

where each i is an observation in time and t is the last i. For example, if t is 1988/06/
02 and x is 15 calendar days (or ten business days). We further choose our
histogram horizon to be 5 years or 1,260 business days. Fifteen business days
after 1988/06/02 is 1988/06/17. To estimate a distribution of the stock return for
1988/06/17, we look back a series of 10-business-day returns. Since we choose
a 5-year historical window, or 1,260-business-day window, the histogram will contain

20
We use three alternative time windows, 2-year, 10-year, and 30-year, to check the robustness of
our procedure and results.
21
The conversion is needed because we use trading-day intervals to identify the appropriate return
histograms and calendar-day intervals to calculate the appropriate discount factor.
224 H.-C. Lin et al.

1,260 observations. The first return in the histogram, R1933/06/02,1933/06/07,1, is the differ-
ence between the log of the stock price on 1988/06/01, ln St–1, and the log of the stock
price 15 calendar days (ten business days) earlier on 1988/05/17, ln St–1–a. The second
observation in the histogram, R1933/06/02/1933/06/07,2 is computed as ln St–2–ln St–2–a.
After, we complete the histogram of returns, we then convert it to the histogram of
prices by multiplying every observation in the histogram by the current stock price:

Stþx, i ¼ St Rt, tþx, i : (7.27)

The expected option payoff is calculated as the average payoff where all the
realizations in the histogram are given equal weights. Thus, Et[CT,T,K] and Et[ST]
are calculated as
8
>
 1 XN  
< Et CT , T , K ¼ max S T , i  K, 0
N i¼1
> XN , (7.28)
: Et ½ S T  ¼ 1 S T , i
N i¼1

where N is the total number of realized returns and Ct,T,K is the price observed at time t,
of an option that expires at time T with strike price K. Substituting the results in Eq.
7.28 in the approximation pricing formula of Eq. 7.4, we obtain our empirical model:

  
C t ¼ P t , T E t C T , T , K þ bC S t  P t , T E t ½ S T 
 
1 XN   1 XN
¼ Pt , T max ST , i  K, 0 þ bC St  Pt, T S
i¼1 T , i
(7.29)
N i¼1 N

where the dollar beta is defined as bC ¼ covvar½C½ST T;ST  as defined in Eq. 7.4.
Note that option prices should be based upon projected future volatility levels
rather than historical estimates. We assume that investors believe that the distribu-
tion of index returns over the time to maturity follows the histogram of a particular
horizon with a projected volatility. In practice, traders obtain this projected vola-
tility by calibrating the model to the market price. We incorporate the projected
volatility, n*t,T,K, into the histogram by adjusting its returns:
v
t, T , K 
R
t, T , K , i ¼ Rt, T , i  Rt, T þ Rt, T ::, i ¼ 1,   , N, (7.30)
vt, T

where the historical volatility nt,T is calculated as the standard deviation of the
historical returns as follows:

1 XN 2
v2t, T ¼ R t , T , i  R t, T (7.31)
N1 i¼1

where Rt,T,i ¼ ST,i/St and Rt, T ¼ N1 SNi¼1 Rt, T , i is the mean return.
7 Nonparametric Bounds for European Option Prices 225

Note that the transformation from R to R* changes the standard deviation

from nt,T to n*t,T,K, but does not change the mean, skewness, or kurtosis. In our
empirical study, we approximate the true volatility by the Black-Scholes
implied volatility. For the upper bound calculations, we also need an
expected mean return of the stock. In our empirical study, we simply use
the histogram mean for it.
The selection of the time horizon is somewhat arbitrary. Empiricists know
well that too long horizons reduce the impact of recent events and yet too
short understate the impact of long time effects. Given that there is no
consensus on a most proper horizon, we perform our test over a variety of choices,
namely, 5-year, 10-year, and 30-year, and the results are similar. To
conserve space, we provide the results on the 10-year and leave the others
available on request.
The results are shown in Table 7.5. Columns (1) and (2) define the maturity
and moneyness buckets. Short maturity is less than 30 days to maturity, medium
is between 31 and 90 days, long is between 91 and 180 days, and real long is
over 180 days to maturity. At-the-money is between 5 % in-the-money and 5 %
out-of-the-money (or 5 %), near-in-the-money/near-out-of-the-money is
between 5 % and 15 %, and deep-in-the-money/deep-out-of-the-money is over
15 %. Moneyness is defined as S/K–1. Column (3) is a frequency count of the
number of observations in each bucket. Column (4) represents the average value of
the ratios of the lower bound over the market price of the option.
Column (5) shows the number of violations when the lower bound is higher than
the market price.
Out of the entire sample (46,540 observations), on average, the lower bound is
9.57 % below the market value and the upper bound is 9.28 % above the market
value. When we look into subsamples, the performances vary. In general, the lower
bound performs better in-the-money than out-of-the-money and medium maturity
than other maturities. The best lower bound performance is when the option is near-
in-the-money and short-term maturity (2.83 % below market value).
To visualize the upper and lower bounds, we plot selected contracts in Fig. 7.1.
In Fig. 7.1, we plot at-the-money (ATM) options from four maturities, 1 month
(Fig. 7.1a), 2 months (Fig. 7.1b), 3 months (Fig. 7.1c), and 6 months (Fig. 7.1d).
As we move from short maturity to long maturity, the number of observations
drops (51, 38, 21, 15 observations, respectively). The bounds are wider as we
move from short maturity to long maturity, consistent with the analysis in
the previous sections.
As we can see, in general, the lower bound using histograms is best for
in-the-money and short-dated options and worst for out-of-the-money options.
On average the lower bound is 9.57 % below the market value (the ratio is
90.43 %). However, there are violations. For example, medium-term, near-out-of-
the-money options have five violations of the lower bound, and there are a total
of 4,233 violations. Theoretically, there should be arbitrage opportunities when the
bounds are violated. To test the lower bound does imply arbitrage opportunities, we
226 H.-C. Lin et al.

Table 7.5 Empirical results on the lower bound

(1) (2) (3) (4) (5)
Maturity Money Total % lbdd Violation
Short Deep out 0 0
Medium Deep out 0 0
Long Deep out 84 57.26 0
Real long Deep out 295 61.18 0
Short Near out 145 65.33 0
Medium Near out 1,925 69.27 5
Long Near out 3,554 82.29 30
Real long Near out 3,317 86.88 271
Short At 4,106 89.73 492
Medium At 7,732 92.45 962
Long At 5,650 92.82 380
Real long At 3,857 92.61 584
Short Near in 2,220 97.17 218
Medium Near in 3,924 97.08 785
Long Near in 2,951 94.50 234
Real long Near in 2,018 90.31 130
Short Deep in 660 92.48 0
Medium Deep in 1,296 94.52 29
Long Deep in 1,352 93.14 46
Real long Deep in 1,454 89.42 67

Total 46,540 90.43 4,233

Note:
1. This is based upon 2,520 business-day (10 years) horizon. Results of other horizons are similar
and are available on request
2. Columns (1) and (2) define the maturity and moneyness buckets. Short maturity is less than
30 days to maturity, medium is between 31 and 90 days, long is between 91 and 180 days, and real
long is over 180 days to maturity. At-the-money is between 5 % in-the-money and 5 % out-of-the-
money (or 5 %), near-in-the-money/near-out-of-the-money is between 5 % and 15 %, and deep-
in-the-money/deep-out-of-the-money is over 15 %. Moneyness is defined as S/K  1. Column
(3) is a frequency count of the number of observations in each bucket. Column (4) represents the
average value of the ratios of the lower bound over the market price of the option. Column
(5) shows the number of violations when the lower bound is higher than the market price
3. The best lower bound performance is when the option is near-in-the-money and short-term
maturity (2.83 % below market value)
The underlying stock return distribution is a 10-year historical return histogram with the volatility
replaced by the Black-Scholes implied volatility

perform a simple buy and hold trading strategy. If the lower bound is violated, we will
buy the option and hold it till maturity. For the 4,233 (out of 46,540) violations, the
buy and hold strategy generated $22,809 or an average of $5.39 per contract.
Given that the buy and hold strategy can be profitable simply due to the bull market,
we compute those that had no violation of the lower bound. For the 42,307 cases that
violated no lower bound, the average profit is $1.83. Hence, the options that violated
the lower bound imply a trading profit 200 % above the average.
7 Nonparametric Bounds for European Option Prices 227

a 30-day Maturity
14 actual ATM
lbdd
12 ubdd

10

0
02/27/88 09/14/88 04/02/89 10/19/89 05/07/90 11/23/90 06/11/91 12/28/91 07/15/92

b 60-day Maturity
18 actual ATM
lbdd
16
ubdd
14
12
10
8
6
4
2
0
02/27/88 09/14/88 04/02/89 10/19/89 05/07/90 11/23/90 06/11/91 12/28/91 07/15/92

91-day Maturity
c 25 ATM
actual
lbdd
ubdd
20

15

10

0
02/27/88 09/14/88 04/02/89 10/19/89 05/07/90 11/23/90 06/11/91 12/28/91 07/15/92

Fig. 7.1 (continued)

228 H.-C. Lin et al.

182-day Maturity
d 40 ATM
actual
lbdd
35 ubdd
30

25

20

15

10

0
02/27/88 09/14/88 04/02/89 10/19/89 05/07/90 11/23/90 06/11/91 12/28/91 07/15/92

Fig. 7.1 (a) Plot of 30-day maturity actual ATM option prices and its upper and lower bound values.
(b) Plot of 60-day maturity actual ATM option prices and its upper and lower bound values. (c) Plot
of 91-day maturity actual ATM option prices and its upper and lower bound values. (d) Plot of
182-day maturity actual ATM option prices and its upper and lower bound values

7.6 Conclusion

In this paper, we derive a new and tighter lower bound for European option
prices comparing with those of previous studies. We further reinterpret
Ritchken’s (1985) upper bound under a nonparametric framework. Our
model contributes to the literature in two different ways. First, our bounds
require no parametric assumption of the underlying stock or the moments of
the distribution. Furthermore, our bounds require no assumptions on interest
rates or volatility. The only requirements of our model are the dollar beta
of the option and expected payoffs of the stock and the option. Hence, our
bounds can be applied to any model such as the random volatility and
random interest rate model by Bakshi et al. (1997) and Scott (1997).
Second, despite of much looser and flexible assumptions, our bounds are
significantly tighter than the existing upper and lower bound models. Most
importantly, our bounds are tighter for the out-of-the-money options that
cannot be bounded efficiently by previous models. Finally, we apply our model to
real data using histograms of the realized stock returns. The results show that
nearly 10 % of the observations violate the lower bound. These violations are
shown to generate significant arbitrage profits, after correction of the bull market in
the sample period.
7 Nonparametric Bounds for European Option Prices 229

Appendix 1

In this appendix, we prove Theorem 7.1. Without loss of generality, we prove

Theorem 7.1 by a three-point convex function. The extension of the proof to
multiple points is straightforward but tedious. Let the distribution be trinomial
and the relationship between the pricing kernel M and the stock price S be convex.
In the following table, x  0; y > e  0.

Probability M S C ¼ max{S  K, 0}
p2 M+x S+y>K S+yK
2p(1  p) M Se<K 0
(1  p)2 Mx Sy<K 0

When e ¼ 0, the relationship between the pricing kernel M and stock price S is
linear, and we obtain equality. When e > 0, the relationship is convex. We first
calculate the mean values:
E½M ¼ M  x þ 2px 
E½S ¼ S  y þ 2pðy 
 e þ 2p e
2

E½C ¼ p ðS þ y  K :
2

The three covariances are computed as follows:

 
cov½M; S ¼ 2pð1  p xðy þ eð2p  1
cov½M; C ¼ 2p2 ð1  p xðS þ y  K 
cov½S; C ¼ 2p2 ð1  p ðS þ y  K ðy þ pe :

The variance of the stock price is more complex:

var½S ¼ 2pð1  pÞz

where
 
z ¼ y2 þ e2 1  2p þ 2p2 þ 2eyð2p  1Þ > 0:
As a result, it is straightforward to show that

cov½S; C 2p2 ð1  pÞðS þ y  K Þðy þ peÞ  

cov½M; S ¼ 2pð1  p xðy þ eð2p  1
var½S 2pð1  pÞz
 ðy þ peÞðy þ eð2p  1ÞÞ
¼ 2p2 ð1  pÞxðS þ y  K
z

 e ð 1  p Þðy  eÞ
¼ 2p ð1  pÞxðS þ y  K 1 þ
2
z

 2p ð1  pÞxðS þ y  K ¼: cov½M, C:
2
230 H.-C. Lin et al.

The fourth line is obtained because cov[M,C] < 0 and 1 þ eð1pzÞðyeÞ > 1. Note that
the result is independent of p since all it needs is 0 < p < 1 for 1 þ eð1pzÞðyeÞ to be
greater than 1. Also note that when e ¼ 0 the equality holds.

References
Bakshi, G., Cao, C., & Chen, Z. (1997). Empirical performance of alternative option pricing
models. Journal of Finance, 52, 2003–2049.
Basso, A., & Pianca, P. (1997). Decreasing absolute risk aversion and option pricing bounds.
Management Science, 43, 206–216.
Black, F. (1976). The pricing of commodity contacts. Journal of Financial Economics, 3, 167–179.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81, 637–654.
Boyle, P., & Lin, X. (1997). Bounds on contingent claims based on several assets. Journal of
Financial Economics, 46, 383–400.
Broadie, M., & Cao, M. (2008). Improved lower and upper bound algorithms for pricing American
options by simulation. Quantitative Finance, 8, 845–861.
Chen, R. R., Lin, H. C., & Palmon, O. (2006). Explaining the volatility smile: Reduced form versus
structural option models. 2006 FMA Annual Meeting. Salt Lake City, UT.
Christoffersen, P., Elkamhi, R., Feunou, B., & Jacobs, K. (2010). Option valuation
with conditional heteroskedasticity and nonnormality. Review of Financial Studies,
23, 2139–2189.
Chuang, H., Lee, C. F. & Zhong, Z. K. (2011). Option bounds: A review and comparison
(Working paper).
Chung, S. L., Hung, M. W., & Wang, J. Y. (2010). Tight bounds on American option prices.
Journal of Banking and Finance, 34, 77–89.
Constantinides, G., & Zariphopoulou, T. (2001). Bounds on derivative prices in an intertemporal
setting with proportional transaction costs and multiple securities. Mathematical Finance,
11, 331–346.
Constantinides, G. M., Jackwerth, J. C., & Perrakis, S. (2009). Mispricing of S&P 500 index
options. Review of Financial Studies, 22, 1247–1277.
Constantinides, G. M., Czerwonko, M., Jackwerth, J. C., & Perrakis, S. (2011). Are options on
index futures profitable for risk-averse investors? Empirical evidence. Journal of Finance, 66,
1407–1437.
Cox, J., & Ross, S. (1976). The valuation of options for alternative stochastic process. Journal of
Finance Economics, 3, 145–166.
De La Pena, V., Ibragimov, R., & Jordan, S. (2004). Option bounds. Journal of Applied Proba-
bility, 41, 145–156.
Frey, R., & Sin, C. (1999). Bounds on European option prices under stochastic volatility.
Mathematical Finance, 9, 97–116.
Gotoh, J. Y., & Konno, H. (2002). Bounding option prices by semidefinite programming: A cutting
plane algorithm. Management Science, 48, 665–678.
Grundy, B. (1991). Option prices and the underlying assets return distribution. Journal of Finance,
46, 1045–1069.
Hobson, D., Laurence, P., & Wang, T. H. (2005). Static-arbitrage upper bounds for the prices of
basket options. Quantitative Finance, 5, 329–342.
Huang, J. (2004). Option pricing bounds and the elasticity of the pricing kernel. Review of
Derivatives Research, 7, 25–51.
Ingersoll, J. (1989). Theory of financial decision making. Totowa, New Jersey: Rowman &
Littlefield.
7 Nonparametric Bounds for European Option Prices 231

Levy, H. (1985). Upper and lower bounds of put and call option value: Stochastic dominance
approach. Journal of Finance, 40, 1197–1217.
Lo, A. (1987). Semi-parametric upper bounds for option prices and expected payoff. Journal of
Financial Economics, 19, 373–388.
Merton, R. (1973). The theory of rational option pricing. The Bell Journal of Economics and
Management Science, 4, 141–183.
Peña, J., Vera, J. C., & Zuluaga, L. F. (2010). Static-arbitrage lower bounds on the prices of basket
options via linear programming. Quantitative Finance, 10, 819–827.
Perrakis, S. (1986). Option pricing bounds in discrete time: Extensions and the pricing of the
American put. Journal of Business, 59, 119–141.
Perrakis, S., & Ryan, J. (1984). Option pricing in discrete time. Journal of Finance, 39, 519–525.
Ritchken, P. (1985). On option pricing bounds. Journal of Finance, 40, 1219–1233.
Ritchken, P., & Kuo, S. (1988). Option bounds with finite revision opportunities. Journal of
Finance, 43, 301–308.
Ritchken, P., & Kuo, S. (1989). On stochastic dominance and decreasing absolute risk averse
option pricing bounds. Management Science, 35, 51–59.
Rodriguez, R. (2003). Option pricing bounds: Synthesis and extension. Journal of Financial
Research, 26, 149–164.
Rubinstein, M. (1976). The valuation of uncertain income stream and the pricing of options. The
Bell Journal of Economics, 7, 407–425.
Scott, L. (1997). Pricing stock options in a jump diffusion model with stochastic volatility and
interest rates: Applications of Fourier inversion methods. Mathematical Finance, 7, 413–424.
Zhang, P. (1994). Bounds for option prices and expected payoffs. Review of Quantitative Finance
and Accounting, 4, 179–197.
Can Time-Varying Copulas Improve the
Mean-Variance Portfolio? 8
Chin-Wen Huang, Chun-Pin Hsu, and Wan-Jiun Paul Chiou

Contents
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
8.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
8.3 Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.3.1 Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.3.2 Copula Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
8.3.3 Gaussian Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
8.3.4 Student’s t-Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
8.3.5 Archimedean Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
8.3.6 Portfolio Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
8.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.5.1 Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.5.2 Average Portfolio Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
8.5.3 Testing Significance of Return Difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
8.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250

C.-W. Huang
Department of Finance, Western Connecticut State University, Danbury, CT, USA
e-mail: huangc@wcsu.edu
C.-P. Hsu (*)
Department of Accounting and Finance, York College, The City University of New York,
Jamaica, NY, USA
e-mail: chsu@york.cuny.edu
W.-J.P. Chiou
Department of Finance and Law College of Business Administration, Central Michigan
University, Mount Pleasant, MI, USA
e-mail: Chiou1P@cmich.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 233
DOI 10.1007/978-1-4614-7750-1_8,
# Springer Science+Business Media New York 2015
234 C.-W. Huang et al.

Abstract
Research in structuring asset return dependence has become an indispensable
element of wealth management, particularly after the experience of the recent
financial crises. In this paper, we evaluate whether constructing a portfolio using
time-varying copulas yields superior returns under various weight updating
strategies. Specifically, minimum-risk portfolios are constructed based on vari-
ous copulas and the Pearson correlation, and a 250-day rolling window tech-
nique is adopted to derive a sequence of time-varied dependencies for each
dependence model. Using daily data of the G7 countries, our empirical findings
suggest that portfolios using time-varying copulas, particularly the Clayton
dependence, outperform those constructed using Pearson correlations. The
above results still hold under different weight updating strategies and portfolio
rebalancing frequencies.

Keywords
Copulas • Time-varying dependence • Portfolio optimization • Bootstrap • Out-
of-sample return • Performance evaluation • GARCH • Gaussian copula •
Student’s t-copula • Gumbel copula • Clayton copula

8.1 Introduction

The return of a portfolio depends heavily on its asset dependence structure. Over the
past decade, copula modeling has become a popular alternative to Pearson corre-
lation modeling when describing data with an asymptotic dependence structure and
a non-normal distribution.1 However, several critical issues attached to the appli-
cations of copulas emerge: Do portfolios using time-varying copulas outperform
those constructed with Pearson correlations? How does the risk return of copula-
based portfolios change over the business cycle? The estimation of parameters has
become particularly critical for finance academics and professionals on the heels of
the recent financial crises. In this chapter, we model the time-varying dependence
of an international equity portfolio using several copula functions and the Pearson
correlation. We investigate whether a portfolio constructed with copula dependence
yields superior returns as compared to a portfolio constructed using a Pearson
correlation under various weight updating strategies.
This paper extends the existing literature in two ways. First, we estimate our
time-varying copulas using a rolling window of the latest 250 trading days. It is well
accepted that the dependencies between asset returns are time varying (Kroner and
Ng 1998; Ang and Bekaert 2002). Differing from the regime-switching type used in
Rodriguez (2007) and Okimoto (2008) or the time-evolving type GARCH model
used in Patton (2006a), we estimate time-varying copulas via a rolling window

1
See Chan et al. (1999), Dowd (2005), Patton (2006a), Engle and Sheppardy (2008), Chollete
et al. (2009), Bauer and Vorkink (2011).
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 235

based on daily data from the previous year. The rolling window method has several
benefits. First, it is a method frequently adopted by practitioners. Second, the rolling
window method considers only the past year’s information when forming depen-
dencies, thus avoiding disturbances that may have existed in the distant past.
Several studies have applied this technique, such as Aussenegg and Cech (2011).
However, Aussenegg and Cech (2011) considered only the daily Gaussian and
Student’s t-copulas in constructing models, and it is reasonable to consider monthly
and quarterly frequencies, given that portfolio managers do not adjust their portfo-
lios on a daily basis. Our research also extends Aussenegg and Cech’s (2011) study
by including the Archimedean copulas to govern the strength of dependence.
Second, our study investigates how the choice of copula functions affects
portfolio performance during periods of economic expansion and recession. The
expansion and recession periods we define are based on announcements from the
National Bureau of Economic Research (NBER). While the use of copula functions
in financial studies has grown enormously, little work has been done in comparing
copula dependencies under different economic states.
Using daily US dollar-denominated Morgan Stanley Capital International
(MSCI) indices of the G7 countries, our empirical results suggest that the
copula-dependence portfolios outperform the Pearson-correlation portfolios. The
Clayton-dependence portfolios, for most scenarios studied, deliver the highest
portfolio returns, indicating the importance of lower-tail dependence in building
an international equity portfolio. Moreover, the choice of weight updating fre-
quency matters. As we increase the weight updating frequency from quarterly to
monthly, the portfolio returns for the full sample and recession periods also
increase, regardless of the choice of dependence measure. Our finding supports
the value of active portfolio reconstruction during recession periods.
This paper is organized as follows. Section 8.2 reviews the literature on copula
applications in portfolio modeling. Section 8.3 describes the empirical models.
Section 8.4 presents the data used. The main empirical results are reported in
Sect. 8.5. Section 8.6 concludes.

8.2 Literature Review

Copulas, implemented in either static or time-varying fashion, are frequently seen in

options pricing, risk management, and portfolio selection. In this section, we review
some copula applications in portfolio selection. Patton (2006a) pioneered time-
varying copulas by modifying the copula functional form to allow its parameters to
vary. Patton (2006a) used conditional copulas to examine asymmetric dependence in
daily Deutsche mark (DM)/US dollar (USD) and Japanese yen (Yen)/US dollar
(USD) exchange rates. His empirical results suggest that the correlation between
DM/USD and Yen/USD exchange rates is stronger when the DM and Yen are
depreciating against the dollar. Hu (2006) adopted a mixture of a Gaussian copula,
a Gumbel copula, and a Gumbel survival copula to examine the various dependence
structures of four stock indices. His results demonstrate the underestimation problem
236 C.-W. Huang et al.

due to multivariate normality correlations and the importance of incorporating both

the structure and the degree of dependence into the portfolio evaluation. Kole
et al. (2007) compared the Gaussian, Student’s t, and Gumbel copulas to illustrate
the importance of selecting an appropriate copula to manage the risk of a portfolio
composed of stocks, bonds, and real estate. Kole et al. (2007) empirically demon-
strated that the Student’s t-copula, which considers the dependence both in the center
and the tail of the distribution, provides the best fit for the extreme negative returns of
the empirical probabilities under consideration.
Rodriguez (2007) studied financial contagions in emerging markets using
switching Frank, Gumbel, Clayton, and Student’s t-copulas. Rodriguez (2007)
found evidence that the dependence structures between assets change during
a financial crisis and that a good asset allocation strategy should allow dependence
to vary with time. Chollete et al. (2009) modeled asymmetric dependence in
international equity portfolios using a regime-switching, canonical vine copula
approach, which is a branch of the copula family first described by Aas
et al. (2007). Chollete et al. (2009) documented that the canonical vine copula
provides better portfolio returns and that the choice of copula dependencies affects
the VaR of the portfolio return. Chollete et al. (2011) investigated international
diversification benefits using the Pearson correlation and six copula functions. Their
results show that dependence increases over time and that the intensity of the
asymmetric dependence varies in different regions of the world.
While some existing studies have applied copulas to the optimization of portfo-
lio selection, most have tended to focus on portfolio risks, i.e., value at risk, rather
than portfolio returns. Empirically, however, investors pay at least equal attention
to portfolio returns. Our study is among the few that have focused on equity
portfolio returns using time-varying copulas.

8.3 Empirical Methods

8.3.1 Copulas

A copula C is a function that links univariate distribution functions into a multivariate

distribution function. Let F be an n-dimensional joint distribution function and let
U ¼ (u1, u2, . . ., un)T be a vector of n random variables with marginal distributions
F1, F2, . . ., Fn. According to Sklar’s (1959) theorem, if the marginal distributions
F1, F2, . . ., Fn are continuous, then a copula C exists, where C is a multivariate
distribution function with all uniform (0,1) marginal distributions.2 That is,

Fðu1 ; u2 ; . . . ; un Þ ¼ CðF1 ðu1 Þ, F2 ðu2 Þ, . . . , Fn ðun ÞÞ, for all u1 , u2 , . . . , un 2 ℝn : (8.1)

2
For detailed derivations, please refer to Cherubini et al. (2004), Demarta and McNeil (2005),
Embrechts et al. (2003), Embrechts et al. (2005), Franke et al. (2008), Nelson (2006), and
Patton (2009).
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 237

Table 8.1 The characteristics of different copulas

Dependence model Tail dependence Parameter range
Pearson correlation No r ϵ (1, 1)
Gaussian copula No r ϵ (1, 1)
Student’s t-copula Yes (symmetry) r ϵ (1, 1), v > 2
Gumbel copula Yes (upper tail) d ϵ (0, 1)
Clayton copula Yes (lower tail) a ϵ [1, 1)\{0}

For a bivariate case, the model can be defined as

Fðx; yÞ ¼ CðFX ðxÞ, FY ðyÞÞ (8.2)

8.3.2 Copula Specifications

In this paper, we consider four copula functions: the Gaussian, Student’s t, Gumbel,
and Clayton. The Gaussian copula focuses on the center of the distribution and
assumes no tail dependence. The Student’s t-copula stresses both the center of the
distribution and symmetric tail behaviors. The Clayton copula emphasizes lower-
tail dependence, while the Gumbel copula focuses on upper-tail dependence.
Table 8.1 summarizes the characteristics of each copula in detail.

8.3.3 Gaussian Copula

The Gaussian copula is frequently seen in the finance literature due to its close
relationship to the Pearson correlation. It represents the dependence structure of two
normal marginal distributions. According to Nelson (2006), the bivariate Gaussian
copula can be expressed as

ð F1 ðxÞ ð F1 ðyÞ  2 

1 s  2rst þ t2
Cðx; yÞ ¼ ds pffiffiffiffiffiffiffiffiffiffiffiffiffi exp 
dt
1 1 2p 1  r2 2 ð 1  r2 Þ
 1 
¼ Fr F ðxÞ, F1 ðyÞ , (8.3)

where F denotes the univariate standard normal distribution function and Fr is the
joint distribution function of the bivariate standard normal distribution with correla-
tion coefficient –1  r  1. The Gaussian copula has no tail dependence unless r ¼ 1.

8.3.4 Student’s t-Copula

Unlike the Gaussian copula, which fails to capture tail behaviors, the Student’s t-copula
depicts the dependence in the center of the distribution as well as in the tails. The
Student’s t-copula is defined using the multivariate t distribution and can be written as
238 C.-W. Huang et al.

ð t1 ð 1  vþ2
v ðxÞ tv ðyÞ 1 s2  2rst þ t2 2

Ctv, r ðx; yÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi 1 þ dsdt

1 1 2p 1  r2 v ð 1  r2 Þ
 
¼ t2v, r t1 1
v ðxÞ, tv ðyÞ , (8.4)

where t2v, r indicates the bivariate joint t distribution, tv1 is the inverse of
the distribution of a univariate t distribution, v is the degrees of freedom, and r
is the correlation coefficient of the bivariate t distribution when v > 2.

8.3.5 Archimedean Copulas

According to Embrechts et al. (2005), the coefficient of upper-tail dependence (lu)

of two series, X and Y, can be defined as
h  
lu ðx; yÞ ¼ limq1 P Y > Fy ðqÞX > Fx ðqÞ : (8.5)

The upper-tail dependence presents the probability that Y exceeds its qth
quantile given that X exceeds its qth quantile, considering the limit as q goes to
its infinity. If the limit lu∈[0,1] exists, then X and Y are said to show upper-tail
dependence. In the same manner, the coefficient of lower-tail dependence (ll) of
X and Y is described as
h  
ll ðx; yÞ ¼ limq0þ P Y  Fy ðqÞX  Fx ðqÞ : (8.6)

Since both FU1 and FU2 are continuous density functions, the upper-tail
dependence can be presented as
 
P Y > Fy ðqÞX > Fx ðqÞ
lu ¼ lim  : (8.7)
q0 P X > Fx ð qÞ

For lower-tail dependence, it can be described as

 
P Y  Fy ð qÞ  X  Fx ð qÞ
ll ¼ limþ  : (8.8)
q0 P X  Fx ð qÞ

8.3.5.1 Gumbel Copula

The Gumbel copula is a popular upper-tail dependence measure as suggested in
Embrechts et al. (2005). The Gumbel copula can be written as
 n
o 
1 1 d
cðx; yÞ ¼ exp  lnðxÞ þ ðlnðyÞ
d d , (8.9)
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 239

where 0 < d  1 measures the degree of dependence between X and Y. When

d ¼ 1, X and Y do not have upper-tail dependence (i.e., X and Y are independent at
upper tails), and when d ! 0, X and Y have perfect dependence.

8.3.5.2 Clayton Copula

The Clayton copula is used to measure lower-tail dependence. The Clayton copula
is defined as
h i
cðx; yÞ ¼ max ðxa þ ya  1Þa ; 0 ,
1
(8.10)

where a describes the strength of dependence. If a ! 0, X and Y do not have lower-

tail dependence. If a ! 1, X and Y have perfect dependence.

8.3.6 Portfolio Construction

The selection of optimal portfolios draws on the seminal work of Markowitz

(1952). Specifically, we adopt the variance minimization strategy with no short
selling and transaction cost assumptions.3 An optimal portfolio allocation can be
found by solving the following optimization problem:
0
Minfwg w Vw
X
Subject to wi ¼ 1, wi  0, (8.11)

where wi is the weight of asset i, and V is the covariance matrix of asset returns.
Because dependence is a time-varying parameter, the data from a subset of
250 trading days prior to the given sample date t is used to derive the dependence
for date t. With 1,780 daily data points in our sample, we calculate a total of 1,531
dependencies for each copula method and the Pearson correlation. With these
dependencies, optimal portfolio weightings can be obtained by solving
a quadratic function subject to specified constraints. The optimal weightings for
time t are used to calculate the realized portfolio returns for (t + 1).4
In practice, portfolio managers periodically reexamine and update the optimal
weights of their portfolios. If the asset allocation of an existing portfolio has
deviated from the target allocation to a certain degree and if the benefit of updating
exceeds its costs, a portfolio reconstruction action is executed. In this paper,
we construct a comprehensive study of portfolio returns by varying the state of

3
Short selling usually involves other service fees, which vary depending on the creditability of the
investors. Because the focus of this study is on the effect of the dependence structure on portfolio
performance, we assume that short selling is not allowed to simplify the comparison.
4
For example, we use return data from t1 to t250 to calculate the optimal portfolio weights with
dependencies estimated from the copulas and the Pearson correlation. The optimal portfolio
weights are applied to the return data at t251 to calculate the realized portfolio returns.
240 C.-W. Huang et al.

Table 8.2 The summary statistics of the G7 indices

Canada France Germany Italy Japan UK USA
Mean (%) 0.0301 0.0064 0.0082 0.0003 0.0017 0.0039 0.0082
Std. dev. 0.0164 0.0171 0.0178 0.0161 0.0155 0.0159 0.0144
Skewness 0.8781 0.0740 0.0666 0.0477 0.1475 0.0535 0.1365
Kurtosis 14.1774 10.7576 8.6920 12.9310 7.4592 12.9143 12.1182
Jarque-Bera 9494 4465 2404 7315 1481 7290 6171
JB P-value 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Observations 1780 1780 1780 1780 1780 1780 1780
The results indicate that the daily returns of the G7 indices are not normally distributed

the economy (i.e., expansion or recession), the dependence structure, and the
frequency of weight updating, i.e., quarterly, monthly, and daily. Quarterly
updating allows investors to update the optimal weights on the first trading days
of March, June, September, and December; monthly updating allows investors to
update the optimal weights on the first trading days of each month. Under daily
updating, investors update the optimal weights every trading day.

8.4 Data

The data are the US dollar-denominated daily returns of the Morgan Stanley Capital
International (MSCI) indices for the G7 countries, including Canada, France,
Germany, Italy, Japan, the UK, and the USA. The sample period covers from the
first business day in June 2002 to the last business day in June 2009, for a total of
1,780 daily observations. Based on the definitions provided by the National Bureau
of Economic Research, we separate the data into an expansion period from June
2002 to November 2007 and a recession period from December 2007 to June 2009.
Table 8.2 presents the descriptive statistics. Among the G7 countries, Canada
had the highest daily returns, while the USA had the lowest. Germany, however,
experienced the most volatile returns. All return series exhibit high kurtosis,
suggesting fat tails on return distributions. The results of the Jarque-Bera test reject
the assumption that the G7 indices have normal distributions.

8.5 Empirical Results

8.5.1 Dependence

Using 1,780 daily data points from the G7 countries, for each dependence model,
we estimate 21 dependence pairs, each containing a sequence of 1,531 dependen-
cies. The parameters for the Gaussian, Student’s t, Gumbel, and Clayton copula
functions are estimated using the two-stage inference for the margins (IFM) method
proposed by Joe and Xu (1996) and Joe (1997).
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 241

The dependencies from the Pearson correlation are calculated using the standard
method. Appendix 1 shows the maximum and the minimum of the 21 dependence
pairs of each dependence model.
The graphs in Fig. 8.1 show the dependencies between the USA and other
countries as estimated via the various copulas and the Pearson correlation. In
general, the Gaussian copula estimation is similar to that of the corresponding
Pearson correlation, but the Student’s t-copulas show significant jumps over time.
For our sample period, Japan shows a low dependence with the US market as
compared to other economies.

8.5.2 Average Portfolio Returns

Table 8.3 presents the average portfolio returns for the full sample period,
the expansion period, and the recession period for the quarterly, monthly, and
daily weight updating strategies. Under the quarterly weight updating, the
Clayton-dependence portfolios have the highest average returns at 6.07 % in the
expansions and 12.52 % in the recessions, and the Pearson-correlation portfolios
have the lowest average returns, 5.48 % in the expansions and 14.25 % in the
recessions. The order of portfolio performance, in the form of its dependence
model regardless of the state of economy, is as follows: the Clayton copula, the
Gumbel copula, the Student’s t-copula, the Gaussian copula, and the Pearson
correlation. Because both the Clayton and Gumbel copulas highlight the tail
dependence between assets, the empirical evidence suggests that with
a quarterly weighting strategy, tail dependence, particularly lower-tail depen-
dence, is important for obtaining superior average portfolio returns across differ-
ent states of the economy.
As we increase the updating frequency from quarterly to monthly, similar empir-
ical results are observed. That is, the Clayton-copula portfolios yield the highest
average returns, while the Pearson-correlation portfolios provide the lowest average
returns. In the expansion periods, the order of portfolio performance, in the form of its
dependence model, is as follows: the Clayton copula, the Student’s t-copula, the
Gaussian copula, the Gumbel copula, and the Pearson correlation. In recession
periods, the order of portfolio performance, in the form of its dependence model, is
as follows: the Clayton copula, the Gumbel copula, the Student’s t- copula, the
Gaussian copula, and the Pearson correlation. According to Kole et al. (2007), the
Gaussian copula, which does not consider lower-tail dependence, tends to be overly
optimistic on the portfolio’s diversification benefits, and the Gumbel copula, which
focuses on the upper tail and pays no attention to the center of the distribution, tends
to be overly pessimistic on the portfolio’s diversification benefits. We verify this
argument by observing that the Gumbel-copula portfolio outperforms only the
Pearson-correlation portfolio in the expansion periods, while the Gaussian-copula-
dependence portfolio outperforms the Pearson-correlation portfolio only in recession
periods. Interestingly, as we increase the weight updating frequency from quarterly to
monthly, the average portfolio returns for the full sample and recession periods also
242 C.-W. Huang et al.

The Dependence using Different Copulas and Pearson Correlation: US vs. Canada
0.90
a
0.80
0.70
0.60
0.50
0.40
0.30 Clayton
Gaussian
0.20 Gumbel
0.10 t
Pearson
0.00
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09

The Dependence using Different Copulas and Pearson Correlation: US vs. France
0.800
0.700
b
0.600
0.500
0.400
0.300 Clayton
0.200 Gaussian
Gumbel
0.100 t
Pearson
0.000
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09

The Dependence using Different Copulas and Pearson Correlation: US vs. Germany
0.80
c
0.70
0.60
0.50
0.40
0.30 Clayton
0.20 Gaussian
Gumbel
0.10 t
Pearson
0.00
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09
The Dependence using Different Copulas and Pearson Correlation: US vs. Italy
0.08
0.70
d
0.60
0.50
0.40
0.30 Clayton
0.20 Gaussian
Gumbel
0.10 t
0.00 Pearson
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09

Fig. 8.1 (continued)

8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 243

The Dependence using Different Copulas and Pearson Correlation: US vs. Japan
0.40 Clayton
e Gaussian
0.30 Gumbel
t
0.20 Pearson

0.10

0.00
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09
−0.10

−0.20
The Dependence using Different Copulas and Pearson Correlation: US vs. UK
0.80
0.70
f
0.60
0.50
0.40
0.30 Clayton
0.20 Gaussian
Gumbel
0.10 t
Pearson
0.00
Jan-03 May-04 Sep-05 Jan-07 May-08 Sep-09

Fig. 8.1 The dependence using different copulas and Pearson correlation. Panel (a) USA versus
Canada. Panel (b) USA versus France. Panel (c) USA versus Germany. Panel (d) USA versus
Italy. Panel (e) USA versus Japan. Panel (f) USA versus UK

increase, regardless of the choice of dependence measures. Thus, the empirical results
seem to support the need for active portfolio reconstruction during recessions.
As the weight updating frequency increases to daily, the Clayton copula delivers
only the highest average portfolio returns during the expansion period. The
Student’s t-copula, by contrast, generates the highest portfolio average returns
for the full sample and recession periods. The influence of the lower-tail
dependence seems to diminish under daily weight reconstruction. The Gaussian-
copula portfolio delivers the worst portfolio performance in both expansion and
recession periods.

8.5.3 Testing Significance of Return Difference

The results reported in the previous section show the average portfolio returns for
different dependencies and weight updating frequencies. One issue with average
returns is that if extreme values exist over the examined period, the empirical
results may be biased and relatively high-standard deviations will be reported.
Previous methods of examining the robustness of portfolio performance usually
build on the data normality assumption (Jobson and Korkie 1981; Memmel
2003), which goes against the empirical facts.
244 C.-W. Huang et al.

Table 8.3 Average portfolio returns

Clayton Gaussian Gumbel Student’s t Pearson
Panel A: quarterly adjustments
Full sample returns 1.44 % 0.91 % 1.16 % 1.11 % 0.57 %
(1.1903) (1.1749) (1.1922) (1.1763) (1.0644)
Expansion returns 6.07 % 5.63 % 5.70 % 5.66 % 5.48 %
(0.6946) (0.0798) (0.6962) (0.7014) (0.6613)
Recession returns 12.52 % 13.35 % 12.56 % 12.64 % 14.25 %
(2.0549) (2.0025) (2.0578) (1.7922) (2.0153)
Panel B: monthly adjustments
Full sample returns 1.63 % 1.08 % 1.22 % 1.22 % 0.75 %
(1.1764) (1.1812) (1.1835) (1.1736) (1.0199)
Expansion returns 6.15 % 5.67 % 5.66 % 5.69 % 5.23 %
(0.6800) (0.6928) (0.6870) (0.6864) (0.6401)
Recession returns 12.01 % 12.78 % 12.19 % 12.30 % 12.80 %
(2.0375) (2.0355) (2.0474) (2.0247) (1.1708)
Panel C: daily adjustments
Full sample returns 1.34 % 0.85 % 1.16 % 1.47 % 1.03 %
(1.1737) (1.1651) (1.1749) (1.1733) (1.0253)
Expansion returns 5.84 % 5.35 % 5.59 % 5.70 % 5.39 %
(0.6733) (0.6859) (0.6839) (0.6766) (0.6367)
Recession returns 12.27 % 12.74 % 12.22 % 11.34 % 12.14 %
(2.0382) (2.0053) (2.0277) (2.0346) (1.7280)
The average portfolio returns are presented in an annualized, percentage format. Three weight
updating frequencies are considered: quarterly, monthly, and daily. Within each frequency, we
report the returns for the full sample period, the expansion period, and the recession period. The
numbers in the parentheses are standard errors

To cope with this problem, Ledoit and Wolf (2008) proposed an alternative
testing method using the inferential studentized time-series bootstrap. Ledoit and
Wolf’s (2008) method is as follows.5 Let a and b be two investment strategies, and
let rat and rbt be the portfolio returns for strategies a and b, respectively, at time t,
where t ranges from 1 to i. The mean vector m and the covariance matrix S for the
return pairs (ra1,rb1)’,. . .,(rat,rbt)’ are denoted by
 X s2 s 
ma
m¼ and ¼ a ab
: (8.12)
mb sab s2b

The performance of strategies a and b can be examined by checking whether the

difference between the Sharpe ratios for strategies a and b are statistically different
from 0. That is,
m m
D ¼ Sa  Sb ¼ a  b (8.13)
sa sb

5
For detailed derivations and computer codes, please refer to Ledoit and Wolf (2008).
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 245

and

^ ¼ Sba  Sbb ¼ mba  mbb ,

D (8.14)
sba sbb

where D is the difference between the two Sharpe ratios, and Sa and Sb are the
Sharpe ratios for strategies a and b, respectively.
Let the second moments of the returns from strategies a and b be denoted by ga
and gb. Then ga ¼ E(gat2) and gb ¼ E(gbt2). Let u and ^u be (ma, mb, ga, gb)’ and
^ can be expressed as
ðm^a , m^b , g^a , g^a Þ0 , respectively. Then D and D

^ ¼ f ð^u Þ
D ¼ f ðuÞ and D (8.15)

ma mb
pffi d
where f ðuÞ ¼ pffiffiffiffiffiffiffiffiffi  pffiffiffiffiffiffiffiffiffi
ffi and ið^u  uÞ! N ð0; CÞ:
ga ma
2 gb mb
2

For time-series data, Ledoit and Wolf

(2008) have argued
XW
that C can be evaluated by the studentized bootstrap as C b ¼ 1 x x0 ,
# j¼1 j j
Xb
where xj ¼ p1ffiffib y t ¼ 1, . . . , #:# is the integer part of the fraction of
t¼1 ðj1Þbþt
the total observations divided by the blocks b. Also,

 
yt ¼ r ta  m^a , r tb  m^b , r 2 ^ 2 ^
ta  ga , r tb  gb t ¼ 1, . . . , i: (8.16)

Following Ledoit and Wolf’s (2008) method, we examine the significance of

60 pairs of portfolio performance. The size of the bootstrap iteration is 10,000 to
ensure a sufficient sample.6 Table 8.4 presents the results from Ledoit and Wolf’s
(2008) portfolio performance test.
The results indicate that during the recession periods and using quarterly
weight updating, the Pearson correlation underperforms all the copula dependencies
at a confidence level of 90 % or greater. During recession periods and adopting
monthly weight updating, the superiority of the copula dependencies jumps to a 99 %
confidence level. Moreover, during the recession periods and assuming daily
updating, the Student’s t-copula outperforms the Pearson correlation at the 99 %
confidence level.
Overall, Ledoit and Wolf’s (2008) empirical tests illustrate the superiority of
copulas during recession periods, regardless of the frequency of weight updating.
During a bullish market, this advantage seems not as statistically significant as
during a bearish market.

6
Ledoit and Wolf (2008) suggested that 5,000 iterations guarantee a sufficient sample. We adopt
a higher standard of 10,000 iterations to strengthen our testing results.
246 C.-W. Huang et al.

Table 8.4 Ledoit and Wolf portfolio performance test

Panel A: quarterly adjustments
CL-GA CL-GU CL-PE CL-t GA-GU
Expansion 0.821 0.812 0.788 0.677 0.916
Recession 0.060a 0.987 0.054a 0.839 0.0760a
GA-PE GA-t GU-PE GU-t PE-t
Expansion 0.892 0.930 0.766 0.912 0.778
Recession 0.0030c 0.0727a 0.0267b 0.943 0.026b
Panel B: monthly adjustments
CL-GA CL-GU CL-PE CL-t GA-GU
Expansion 0.193 0.415 0.568 0.744 0.881
Recession 0.249 0.803 0.021c 0.295 0.092a
GA-PE GA-t GU-PE GU-t PE-t
Expansion 0.850 0.892 0.795 0.896 0.809
Recession 0.001c 0.318 0.019c 0.475 0.008c
Panel C: daily adjustments
CL-GA CL-GU CL-PE CL-t GA-GU
Expansion 0.389 0.814 0.732 0.929 0.913
Recession 0.000c 0.119 0.173 0.371 0.000c
GA-PE GA-t GU-PE GU-t PE-t
Expansion 0.928 0.915 0.460 0.831 0.301
Recession 0.718 0.001c 0.045c 0.778 0.019c
The performance tests are conducted using the approach suggested by Ledoit and Wolf (2008).
The tests examine whether the returns from two portfolios are significantly different at the
95 % level
CL stands for the Clayton copula, GA stands for the Gaussian copula, GU stands for the gumbel
copula, PE stands for Pearson correlation, and t stands for the Student’s t-copula
a
Represents 90 % statistical significance
b
Represents 95 % statistical significance
c
Represents 99 % statistical significance

8.6 Conclusions

In this paper, we study whether adopting time-varying copulas as a measure of

dependence of asset returns can improve portfolio performance. This study was
motivated by the fact that the traditional Pearson correlation is inadequate in
describing most financial returns. Moreover, the robustness of copula functions
has not been fully examined under different states of the economy and weight
updating scenarios. We evaluate the effectiveness of various copulas in managing
portfolios while considering portfolio rebalance frequencies and the business cycle.
The significance of return difference is tested using the studentized time-series
bootstrap method suggested by Ledoit and Wolf (2008).
The main findings are as follows: first, an international equity portfolio modeled
using the Pearson correlations underperforms those modeled using copula-based
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 247

dependencies, especially during recession periods. Our findings remain robust

regardless of the rebalancing frequency. Second, the importance of lower-tail
behaviors in portfolio modeling is highlighted by the higher average portfolio
returns of the Clayton-dependence portfolios. Third, the choice of weight updating
frequency affects portfolio returns. The portfolios using monthly weight updating
frequency provide better portfolio returns than those using quarterly and daily weight
adjustments.
We add to the current literature by thoroughly evaluating the effectiveness
of asymmetric conditional correlations in managing portfolio risk. This paper
synthesizes the major concepts and modi operandi of previous research and max-
imizes the practicality of applying copulas under a variety of scenarios. Future
research into copulas can be extended to the contagion of various asset classes and
interest rates and evaluations of the impact of certain economic events.

Appendix 1

Appendix 1 illustrates the dependence of the G7 countries from different depen-

dence models. Note that to ease the comparison between dependencies, we trans-
form Gumbel dependencies by (1  d). Therefore, the range for the Clayton and the
Gumbel copulas is between 0 and 1, with 0 meaning no dependence and 1 standing
for perfect dependence. The range for the Gaussian copula, the Student’s t-copula,
and the Pearson correlation is 1 to 1, with 0 meaning no dependence and 1 or 1
standing for complete dependence.

CA FR DE IT JP UK USA
Panel A: Gaussian dependence
CA
Max
Min
FR
Max 0.7064
Min 0.3914
DE
Max 0.6487 0.9686
Min 0.2016 0.7856
IT
Max 0.6320 0.9407 0.9274
Min 0.2143 0.2303 0.7001
JP
Max 0.1814 0.2761 0.2478 0.4023
Min 0.2115 0.2238 0.2229 0.0119
UK
Max 0.6393 0.9100 0.8665 0.8575 0.4664
Min 0.1945 0.2384 0.2008 0.2050 0.0329
(continued)
248 C.-W. Huang et al.

CA FR DE IT JP UK USA
USA
Max 0.7221 0.5031 0.5231 0.4661 0.5831 0.5671
Min 0.1864 0.2127 0.2087 0.2414 0.2241 0.1997
Panel B: Student’s t dependence
CA
Max
Min
FR
Max 0.7509
Min 0.3921
DE
Max 0.4578 0.9810
Min 0.1070 0.7476
IT
Max 0.4367 0.9810 0.9586
Min 0.1089 0.7476 0.6937
JP
Max 0.1093 0.1679 0.1522 0.6846
Min 0.1284 0.1511 0.1574 0.0093
UK
Max 0.4478 0.8055 0.7380 0.7316 0.7335
Min 0.1094 0.1546 0.1359 0.1230 0.0284
USA
Max 0.5733 0.8055 0.3457 0.3176 0.4375 0.6614
Min 0.1107 0.1546 0.1343 0.1360 0.1519 0.2687
Panel C: Gumbel dependence
CA
Max
Min
FR
Max 0.5947
Min 0.3220
DE
Max 0.3744 0.9063
Min 0.0000 0.5928
IT
Max 0.3666 0.7286 0.8544
Min 0.0000 0.0000 0.5516
JP
Max 0.0961 0.1384 0.1222 0.5356
Min 0.0000 0.0000 0.0000 0.0200
(continued)
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 249

CA FR DE IT JP UK USA
UK
Max 0.3650 0.6946 0.6156 0.6086 0.5855
Min 0.0000 0.0000 0.0000 0.0000 0.0234
USA
Max 0.4340 0.2816 0.2952 0.2649 0.3261 0.5967
Min 0.0000 0.0000 0.0000 0.0000 0.0000 0.2493
Panel D: Clayton dependence
CA
Max
Min
FR
Max 0.6763
Min 0.2899
DE
Max 0.3585 0.9327
Min 0.0000 0.6696
IT
Max 0.3463 0.7635 0.9004
Min 0.0000 0.0000 0.6006
JP
Max 0.0038 0.0408 0.0287 0.6476
Min 0.0000 0.0000 0.0000 0.0000
UK
Max 0.3657 0.7193 0.6567 0.6506 0.6794
Min 0.0000 0.0000 0.0000 0.0000 0.0000
USA
Max 0.5248 0.2450 0.2476 0.1987 0.3472 0.6622
Min 0.0000 0.0000 0.0000 0.0000 0.0000 0.1982
Panel E: Pearson correlation
CA
Max
Min
FR
Max 0.7002
Min 0.3966
DE
Max 0.6944 0.9726
Min 0.3645 0.7890
IT
Max 0.6833 0.9596 0.9477
Min 0.3877 0.8204 0.6965
(continued)
250 C.-W. Huang et al.

CA FR DE IT JP UK USA
JP
Max 0.3549 0.4594 0.4702 0.4073
Min 0.0411 0.0251 0.0170 0.0072
UK
Max 0.7080 0.9573 0.9298 0.9181 0.4612
Min 0.3676 0.7791 0.6572 0.6990 0.0154
USA
Max 0.7586 0.6096 0.7443 0.5871 0.2078 0.5480
Min 0.3764 0.2647 0.2921 0.2481 0.1562 0.1913

References
Aas, K., Czado, C., Frigessi, A., & Bakken, H. (2007). Pair-copula constructions of multiple
dependence. Insurance: Mathematics and Economics, 2, 1–25.
Ang, A., & Bekaert, G. (2002). International asset allocation with regime shifts. Review of
Financial Studies, 15, 1137–1187.
Ang, A., & Chen, J. (2002). Asymmetric correlations of equity portfolios. Journal of Financial
Economics, 63, 443–494.
Aussenegg, W., & Cech, C. (2011, forthcoming). Simple time-varying copula estimation. In
A. S. Barczak, E. Dziwok (Ed.), Mathematical econometrical and computational methods in
finance and insurance. Katowice: University of Economics in Katowice.
Bauer, G. H., & Vorkink, K. (2011). Forecasting multivariate realized stock market volatility.
Journal of Econometrics, 160, 93–101.
Chan, L. K., Karceski, J., & Lakonishok, J. (1999). On portfolio optimization: Forecasting
covariances and choosing the risk model. Review of Financial Studies, 12, 937–974.
Cherubini, U., Luciano, E., & Vecchiato, W. (2004). Copula methods in finance. Hoboken: Wiley.
Chollete, L., Heinen, A., & Valdesogo, A. (2009). Modeling international financial
returns with a multivariate regime-switching copula. Journal of Financial Econometrics, 7,
437–480.
Chollete, L., Peña, V., & Lu, C. (2011). International diversification: A copula approach. Journal
of Banking and Finance, 35, 403–417.
Demarta, S., & McNeil, A. (2005). The t copula and related copulas. International Statistical
Review, 73, 111–129.
Dowd, K. (2005). Copulas and coherence. Journal of Portfolio Management, Fall, 32(1), 123–127.
Embrechts, P., Frey, R., & McNeil, A. (2005). Quantitative risk management. Princeton: Princeton
University Press.
Embrechts, P., Lindskog, F., & McNeil, A. (2003). Modelling dependence with copulas and
applications to risk management. In S. T. Rachev (Ed.), Handbook of heavy tailed distributions
in finance (pp. 329–384). Amsterdam: Elsevier/North-Holland.
Engle, R. F., & Sheppardy K. (2008). Evaluating the specification of covariance models for large
portfolios (Working Paper). New York University.
Franke, J., Hardle, W., & Hafner, C. (2010). Statistics of financial markets: An introduction. New
York: Springer.
Hu, L. (2006). Dependence patterns across financial markets: A mixed copula approach. Applied
Financial Economics, 16, 717–729.
8 Can Time-Varying Copulas Improve the Mean-Variance Portfolio? 251

Jobson, J., & Korkie, B. (1981). Performance hypothesis testing with the Sharpe and Treynor
measures. Journal of Finance, 36, 889–908.
Joe, H. (1997). Multivariate models and dependence concepts. New York: Chapman & Hall/CRC.
Joe, H., & Xu, J. J. (1996). The estimation method of inference functions for margins for
multivariate models (Technical Report 166). Department of Statistics, University of British
Columbia.
Kole, E., Koedijk, K., & Verbeek, M. (2007). Selecting copulas for risk management. Journal of
Banking and Finance, 31, 2405–2423.
Kroner, K. E., & Ng, V. (1998). Modeling asymmetric comovements of asset returns. Review of
Financial Studies, 11, 817–844.
Ledoit, O., & Wolf, M. (2008). Robust performance hypothesis testing with the Sharpe ratio.
Journal of Empirical Finance, 15, 850–859.
Longin, F., & Solnik, B. (2001). Extreme correlation of international equity markets. Journal of
Finance, 56, 649–676.
Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7, 77–91.
Memmel, C. (2003). Performance hypothesis testing with the Sharpe ratio. Finance Research
Letters, 1, 21–23.
Nelson, R. (2006). An introduction to copulas. New York: Springer.
Okimoto, T. (2008). New evidence of asymmetric dependence structures in international equity
markets. Journal of Financial and Quantitative Analysis, 43, 787–816.
Patton, A. (2006a). Modelling asymmetric exchange rate dependence. International Economic
Review, 47, 527–556.
Patton, A. (2006b). Estimation of multivariate models for time series of possibly different lengths.
Journal of Applied Econometrics, 21, 147–173.
Patton, A. (2009). Copula-based models for financial time series. In T. Mikosch, J. P. Kreiss,
T. G. Andersen, R. A. Davis, J. P. Kreiss, & T. Mikosch (Eds.), Handbook of financial time
series. New York: Springer.
Rodriguez, J. C. (2007). Measuring financial contagion: A copula approach. Journal of Empirical
Finance, 14, 401–423.
Sklar, A. (1959) “Fonctions de Répartition à n Dimensions et Leurs Marges,” Publications de
l’Institut de Statistique de l’Université de Paris 8, 229–231.
Determinations of Corporate Earnings
Forecast Accuracy: Taiwan Market 9
Experience

Ken Hung and Kuo-Hao Lee

Contents
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
9.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
9.3 Testable Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.3.1 Firm Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.3.2 Volatility of Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.3.3 Volume Turnover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
9.3.4 Corporate Earnings Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
9.3.5 Type of Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
9.3.6 Forecasting Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
9.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
9.4.1 Comparison of Management and Analyst’s Earnings Forecast . . . . . . . . . . . . . . . . . 260
9.4.2 Factors Influencing the Absolute Errors of Earnings Forecast . . . . . . . . . . . . . . . . . . 263
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Methodology Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Variable Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Market Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Wilcoxon Two-Sample Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276

K. Hung (*)
Division of International Banking & Finance Studies, Texas A&M International University,
Laredo, TX, USA
e-mail: ken.hung@tamiu.edu
K.-H. Lee
Department of Finance, College of Business, Bloomsburg University of Pennsylvania,
Bloomsburg, PA, USA
e-mail: klee@bloomu.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 253
DOI 10.1007/978-1-4614-7750-1_9,
# Springer Science+Business Media New York 2015
254 K. Hung and K.-H. Lee

Abstract
Individual investors are actively involved in stock market and are making
investment decision based on publicly available and nonproprietary information,
such as corporate earnings forecasts from the management and the financial
analyst. Also, the management forecast is another important index investors
might use.
To examine the accuracy of the earnings forecasts, the following test meth-
odology have been conducted. Multiple regression models are used to examine
the effect of six factors: firm size, market volatility, trading volume turnover,
corporate earnings variances, type of industry, and experience. If the two-sample
groups are related, Wilcoxon two-sample test will be used to determine the
relative earnings forecast accuracy.
The results indicate that firm size has no effect on management forecast,
voluntary management forecast, mandatory management forecast, and analysts’
forecast. There are some indications that forecasting accuracy is affected by
market ups and downs. The results also reveal that relative accuracy of earnings
forecasts is not a function of trading volume turnover. However, management’s
earnings forecast and analysts’ forecasts are sensitive to earnings variances.
Readers are well advised and referred to the chapter appendix for methodo-
logical issues such as sample selection, variable definition, regression model,
and Wilcoxon two-sample test.

Keywords
Multiple regression • Wilcoxon two-sample test • Corporate earnings • Forecast
accuracy • Management earnings • Firm size • Corporation regulation •
Volatility • Trade turnover • Industry

9.1 Introduction

In recent times, individual investors are actively involved in stock market and are
making investment decision based on publicly available and nonproprietary infor-
mation. Corporate earnings forecasts are an important investment tool for investors.
Corporate earnings forecasts come from two sources: the company management
and financial analyst. As an insider, the management has the advantage of
possessing more information and hence provides a more accurate earnings forecast.
However, because of the existing relationship of the company with its key investor
group, the management may have a tendency to take an optimistic view and
overestimate its future earnings. In contrast, the financial analysts are less informed
about the company and often rely on management briefings. They have more
experiences in the overall market and economies and are expected to analyze
companies objectively. Hence, analysts should provide reliable and more accurate
earnings forecast.
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 255

Whether investors should rely on the earnings forecast made by the management
or by the analyst is a debatable issue. Many researchers have examined the accuracy
of such earnings forecasts. Because there are differences in methodologies, sample
selections and time horizons, the findings and conclusions from the previous studies
are conflicting and inconclusive. This motivated us to do a new analysis by using
a different methodology.
According to Regulations for Publishing Corporate Earnings Forecast imposed
by the Department of Treasury,1 publicly traded Taiwanese companies have to
publish their earnings forecasts under the following situations:
1. To issue new stocks or acquired liabilities;
2. When more than one third of the board has been changed;
3. When one of the situations as listed in section 185 of the corporation regulations
happens;
4. Merger and acquisitions;
5. Profit gain/loss up to one third of its annual revenue due to an unexpected event;
6. Revenue loss over 30 % compared to last year;
7. Voluntarily publish its earnings forecast
Since management earnings forecasts are mandatory or voluntary, the focus of
this research is to examine the accuracy of management’s overall earnings forecast,
management voluntary earnings forecast, management mandatory earnings fore-
cast, and financial analyst earnings forecast.

9.2 Literature Review

Jaggi (1980) examined the impact of company size on forecast accuracy using
management’s earnings forecasts from the Wall Street Journal and analysts’
earnings forecasts from Value Line Investment Service from 1971 to 1974. He
argued that because a larger company has strong financial and human capital
resources, its management’s earnings forecast would be more accurate than the
analyst’s. The sample data were classified into six categories based on the size of
the firms’ total revenue to examine the factors that attribute to the accuracy of
management’s earnings forecast with the analyst’s. The result of his research did
not support his hypothesis that management’s forecast is more accurate than the
analyst’s.
Bhushan (1989) assumed that it is more profitable trading large companies’
stocks because large companies have better liquidity than the small ones. Therefore,
the availability of information is related to company size. His research results
support his hypothesis that the larger the company size, the more information is
available to financial analysts and the more accurate their earnings forecasts are.
Kross and Schreoder (1990) proposed that brokerage firm’s characteristics influ-
ence analysts’ earnings forecasting accuracy. In their analysis, sample analysts’

1
Securities Regulation Committee, Department of Treasury, series 00588, volume #6, 1997
256 K. Hung and K.-H. Lee

earnings forecasts from 1980 to 1981 were obtained from Value Line Investment,
and the market value of a firm is used as the size of the firm. The results of this study
on analysts’ earnings forecasts did not find a positive relation between the company
size and the analyst’s forecast accuracy. Xu (1990) used the data range from 1986 to
1990 and the logarithm of average total revenue as a proxy of company earnings
and examined factors associated with the accuracy of analysts’ earnings forecast.
The hypothesis that the larger the firm size is, the more accurate the analysts’
earnings forecast would be was supported.
Su (1996) focuses on comparison of relative accuracy of management and
analysts’ earnings forecasts by using cross-sectional design method. Samples
selection includes forecast data during the time period from 1991 to 1994. The
company’s “book value” of its total assets is used as a proxy for the size of the
company in the regression analysis. The author believes that analysts are more
attracted to larger companies, and there are more incentives for them to follow large
companies than small companies in their forecasting. Therefore, the size of
a company will affect the relative accuracy of analyst earnings forecast. On the
other hand, large companies possess excessive human and financial resources and
information which analysts have no access to, to allow managers to anticipate the
corporate future earnings with high accuracy. The study results show that analyst
and voluntary earnings forecast accuracy for larger companies are higher than
forecast accuracy for small companies.
Yie (1996) examines the factors influencing management and financial analyst
earnings forecasting accuracy. Data used in this study are the earnings forecasts
during the years 1991–1995. She uses the company’s total assets and market value
of the company’s equity as proxies for company size. The finding of this research
reveals that the relative earnings forecast accuracy (management, voluntary man-
agement, mandatory management, and analyst) is not affected by the size of the
company when the company’s total assets are used as the proxy of company size.
The result also indicates that mandatory management’s earnings forecast and
analysts’ earnings forecasts are influenced by company size if market value of
company’s equity is used.
Xu (1990) examines the relative accuracy of analysts’ earnings forecasts,
a hypothesis that market volatility is one of factors that influence the relative
accuracy of analyst earnings forecast. In upmarket situation, a vast amount of
information regarding corporate earnings and overwhelming trading activities
may hinder the analyst from getting realistic and objective information; thus,
overoptimistic forecast might be a result. In contrast, when market is experiencing
a downturn, individual investors are less speculative and more rational; thus,
information about corporate earnings tends to be more accurate. Under these
circumstances, the analyst tends to provide earnings forecasts with a higher level
of accuracy. The results of this study support the author’s hypothesis. Jiang (1993)
examines the relative accuracy between management’s earnings forecast and ana-
lyst earnings forecast. He hypothesizes that analysts’ earnings forecast has a higher
degree of accuracy in down market compared to upmarket situation. He uses
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 257

samples forecast data from years 1991 to 1993 in his analysis and finds that the
result of this research supports his argument.
Das et al. (1998) used a cross-sectional approach to study the optimistic
behavior of financial analysts. Especially, they focused on the predicative accu-
racy of past information analysts’ earnings forecast associated with magnetite of
the bias in analysts’ earnings forecasts. The sample selection covers the time
period from 1989 to 1993 with 274 companies’ earnings forecasts information.
A regression method was used in this research. The term “optimistic behavior” is
referred to as the optimistic earnings forecasts made by financial analysts. The
authors hypothesize the following scenario: there is higher demand for nonpublic
information for firms whose earnings are more difficult to predict than for firms
whose earnings can be accurately forecasted using public information. Their
finding supports the hypothesis that analysts will make more optimistic forecasts
for low-predictability firms with an assumption that optimistic forecast facilitates
access to management’s nonpublic information. Clement (1999) studies the rela-
tion between the quality and forecast accuracy of analysts’ reports. It also identifies
systematic and time persistence in analysts’ earnings forecast accuracy and exam-
ines the factors associated with degree of accuracy. Using the I/B/E/S database, the
author has found that earnings forecast accuracy is positively related with analysts’
experience (a surrogate for analyst ability and skill) and employer size (a surrogate
for resources available) and inversely related with the number of firms and
industries followed by the analyst. The sample selection covers the time horizon
from 1983 to 1994 with earnings forecasts of 9,500 companies and 7,500 analysts.
The author believes that as the analyst’s experience increases, his earnings forecast
accuracy will increase, which implies that the analyst has a better understanding of
the idiosyncrasies of a particular firm’s reporting practices or he might establish
a better relationship with insiders and therefore gain better access to the managers’
private information. An analyst’s portfolio complexity is also believed to have
association with his earnings forecast accuracy. He hypothesizes that forecast
accuracy would decrease with the number of industries/firms followed. The effect
of available resources impacts analyst’s earnings forecast in such a way that
analysts employed by a larger broker firm supply more accurate forecasts than
smaller ones. The rationale behind this hypothesis is that the analyst hired
by a large brokerage firm has better access to the private information of managers
at the companies he follows. Large firms have more advanced networks that
allow the firms to better disseminate their analyst’s recommendations into the
capital markets. The results of this research support the hypothesis made by the
author.
Xiu (1992) studies the relative accuracy of management and analysts’ earn-
ings forecasts using Taiwan database covering the period 1986–1990. The
management and analyst’s earnings forecasts used in the study are from Busi-
ness News, Finance News, Central News Paper, and The United Newspaper. The
research methodology is to examine management’s earnings forecast accuracy
with prior and posterior analyst’s earnings forecasts. The result reveals that
258 K. Hung and K.-H. Lee

a management’s forecast is superior to prior analyst’s forecast, but less accurate

than posterior analyst’s earnings forecasts.
Jiang (1993) examined the determinants associated with analysts’ earnings
forecast and management and analyst’s earnings accuracy under different assump-
tions. A sample of Taiwan corporations is collected from the Four Seasons news-
paper. Jiang uses cross-sectional regression analysis to investigate the relations
between the forecast accuracy and firm’s size, rate of earnings deviation, forecast-
ing time horizon, market situation, and rate of annual trading volume turnover. His
results show that earnings forecasts provided by analysts are more accurate than
management earnings forecasts. Chia (1994) focuses a study on mandatory man-
agement earnings forecasts and the rate of trading volume turnover in an
unpublished thesis.

9.3 Testable Hypotheses

9.3.1 Firm Size

The size of a firm is believed to have influence on the accuracy of analyst’s and
management’s earnings forecast. Jaggi (1980), Bhushan (1989), and Clement (1999)
found that the larger the company is, the more accurate the earnings forecast will
be. They believe that holding other factors constant, larger companies have more
financial and human resources available that allow the management to draw more
precise earnings forecast than smaller companies. Thus, forecasts and recommenda-
tions supplied by larger firms are more valuable and accurate than the smaller firms:
H1: The accuracy of management’s earnings forecast increases with the size of
the firm.
H2: The accuracy of management’s voluntary earnings forecast increases with the
size of the firm.
H3: The accuracy of management’s mandatory earnings forecast increases with the
size of the firm.
H4: The accuracy of analysts’ earnings forecast increases with the size of the firm.

9.3.2 Volatility of Market

The accuracy of earnings forecast will be affected by market situation. When

market is very volatile and unstable, investors who are looking for the opportunities
to profit will act more, speculative about what would be the next for the market. In
this situation, it is more difficult for analysts to figure out the real useful information
for their forecasts; they might have a tendency to overoptimistically forecast the
earnings and provide recommendations. When a market is in a relative stable
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 259

period, investors tend to be rational about the next movement of the market; there is
less biased information regarding corporate earnings among the general investors;
thus, the information accessed by analysts will allow them to be more objective in
the earnings forecast. In contrast, the management has the insights on what is really
happening in the aspects of operation, finance, top management changes, and
profitability of the business. Even if they are less vulnerable regardless of what
the market situation is, voluntary management’s earnings forecast might be affected
by market volatility to some extent:
H5: Management’s earnings forecast will not be affected by volatility of market.
H6: Voluntary management’s earnings forecast is a function of the of market
volatility.
H7: Mandatory management’s earnings forecast is not affected by market volatility.
H8: Accuracy of analysts’ earnings forecast is affected by market volatility.

9.3.3 Volume Turnover

The relationship between trading volume turnover and accuracy of earnings

forecast can be examined based on the hypothesis that daily stock trading volume
represents the public investors’ perception about a company. Larger trading volume
during a day for a particulate stock reflects a higher degree of divergence on
confidence about the company’s stock, and vice versa. This public perception on
a stock might distract management’s and analysts’ judgment; they need more time
and strive extra efforts in order to prove accurate earnings forecasts:
H9: Trading volume turnover affects the accuracy of management’s earnings
forecast.
H10: Trading volume turnover affects the accuracy of voluntary management’s
earnings forecast.
H11: Trading volume turnover affects the accuracy of mandatory management’s
earnings forecast.
H12: Trading volume turnover affects the accuracy of analysts’ earnings forecast.

9.3.4 Corporate Earnings Variance

Corporate earnings surprises are an important aspect of analysts’ earnings forecast.

The larger the earnings surprise is, the less useful the past information will be in
earnings forecasting and the harder it is to make accurate forecasts. Corporate
earnings variances represent the earnings surprises a company has in the past; it
would affect the accuracy of management and analysts’ earnings forecasts:
H13: Corporate earnings variances affect the accuracy of management’s earnings
forecast.
H14: Corporate earnings variances affect the accuracy of voluntary management’s
earnings forecast.
260 K. Hung and K.-H. Lee

H15: Corporate earnings variances affect the accuracy of mandatory manage-

ment’s earnings forecast.
H16: Corporate earnings variances affect the accuracy of analysts’ earnings
forecast.

9.3.5 Type of Industry

There may exist a relationship between the type of industry and earnings forecast
accuracy. They hypothesize that the difference between different industries may
result in different levels of accuracy on earnings forecast. Some analysts may not
possess adequate knowledge necessary in the forecasting in a particular industry;
therefore, their forecast may not be as accurate as management’s earnings forecast.
Hence, the following hypotheses can be tested:
H17: Type of industries influences the accuracy of management‘s earnings forecast.
H18: Type of industries affects the accuracy of voluntary management’s earnings
forecast.
H19: Type of industries affects the accuracy of mandatory management’s earnings
forecast.
H20: Type of industries affects the accuracy of analysts’ earnings forecast.

9.3.6 Forecasting Experience

Analysts’ accuracy of earnings forecast will improve as their experience and

knowledge about companies increase. They learn from their previous forecasts
and make the next forecast more accurate. A similar argument can be made about
the management’s earnings forecast. Hence, the following hypotheses can be
tested:
H21: Forecasting experience influences the accuracy of management’s earnings
forecast.
H22: Forecasting experience affects the accuracy of voluntary management’s
earnings forecast.
H23: Forecasting experience affects the accuracy of mandatory management’s
earnings forecast.
H24: Forecasting experience affects the accuracy of analysts’ earnings forecast.

9.4 Empirical Results

9.4.1 Comparison of Management and Analyst’s Earnings Forecast

To compare the relative accuracy of management and analysts’ earnings forecasts,

we focus on four major aspects regarding the relative accuracy of earnings fore-
casts. First, management versus analysts’ earnings forecasts is made to compare the
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 261

relative accuracy of the forecasts. Secondly, a comparison is made between volun-

tary management’s forecasts and analysts’ forecasts. Thirdly, mandatory manage-
ment’s forecasts are compared with analysts’ forecasts to determine the relative
accuracy of the forecasts. Finally, tests of hypothesis have been made to further
prove the relative accuracy of management and analysts’ earnings forecasts.
Table 9.1 provides descriptive statistics of management and financial analysts’
earnings forecasts based from 1987 to 1999. It can be observed that absolute errors
of management earnings forecasts are less than the analyst’s in the years 1992,
1993, 1995, 1996, 1997, and 1997. It indicates that management’s earnings fore-
casts are superior to analysts during that time period. But, in other time periods, the
absolute errors for management’s earnings forecast are higher than analysts’,
indicating analysts provide higher forecast productions. Overall, it is less obvious
to conclude who has higher predicate ability for providing more precise earnings
forecast.
The last three rows of Table 9.1 list the mean absolute errors of earnings
forecasts by management and analysts during three different time periods. From
1988 to 1992, management’s forecast absolute mean error is 1.796, whereas
analyst’s earnings forecast absolute mean error is 1.503. From 1993 to 1999,
management’s earnings forecast absolute mean error is 2.031, while analysts’
forecast absolute mean error shows higher value of 2.236. If we look at the entire
time period from 1987 to 1999, the absolute mean error for management’s forecast
is less than the absolute mean error for analysts’ earnings forecast, which is 1.969
and 2.043 respectively. A conclusion can be drawn from the above results that
management’s earnings forecasts are more accurate than analysts’ forecasts from
the early 1990s, but less accurate during the late 1980s.
Table 9.2 shows the results of the Wilcoxon signed-rank test used to test the
relative accuracy of managements’ forecast and analysts’ forecast. Comparing the
negative ranks and positive ranks in Table 9.2, management’s forecasts are less
accurate than analysts’ forecasts in the years 1987, 1988, 1989, 1990, and 1999, but
more accurate in the years 1995 and 1998. There is no significant difference in the
absolute errors between management’s forecasts and analysts’ forecasts.
If we examine the z-values of the test for the entire time period (1986–1999), the
z-value for Wilcoxon signed-rank test is
0.346, which is not significant enough to
tell the difference between the two samples. This supports the hypothesis H1 that
there are no significant differences between management’s forecast accuracy and
analysts’ forecast accuracy. This also agrees with the findings suggested by Imhoff
and Pare (1982) and Baretley and Cameron (1991). They believe the reason for that
is due to the similar abilities of forecasters and comparable networks to access
company information (public/private) between management and analysts; it is
possible that both can provide relative accurate earnings forecasts.
If the entire time period is divided into two subsamples, one is from 1987 to 1992
and the other is from 1993 to 1999, the latter subsample shows a significant level of
0.05 with a z-value of
2.138, which indicates that management’s forecasts are less
reliable than analysts’ forecasts. For the former subsample, it shows no contradic-
tion with the results by Imhoff and Pare (1982) and Baretley and Cameron (1991).
 262

Table 9.1 Descriptive statistics of management and analyst’s earnings forecast errors
Management’s forecast error Analyst’s forecast error
Year Sample size Mean Standard error Maximum value Minimum value Mean Standard error Maximum value Minimum value
1999 402 1.15 3.64 38.55 0.0006 1.07 3.18 35.76 0.0006
1998 360 1.61 5.80 63.48 0.0017 2.10 7.10 75.81 0.0037
1997 317 0.74 2.20 29.87 0.0040 0.78 3.17 53.23 0.0002
1996 267 1.12 5.41 55.45 0.0007 1.33 6.49 71.34 0.0009
1995 226 0.86 2.93 33.27 0.0001 0.87 2.31 28.42 0.0018
1994 178 0.62 2.17 18.8 0.0015 0.62 2.17 23.86 0.0025
1993 157 12.71 149.1 1869.1 0.0011 13.80 164.1 2057.4 0.0003
1992 126 0.65 1.48 11.58 0.0013 0.73 1.48 8.14 0.0005
1991 144 0.84 2.99 32.93 0.0014 0.65 1.10 5.67 0.0079
1990 120 5.98 32.60 334.94 0.0017 4.86 20.92 187.3 0.0075
1989 116 1.11 2.28 18.66 0.0009 0.91 1.59 10.56 0.0056
1988 96 1.25 3.29 19.57 0.0001 1.08 2.74 15.43 0.0006
1987 78 0.66 1.88 15.73 0.0171 0.57 1.78 14.62 0.0025
1993–1999 1,907 2.031 0.983 1869.1 0.0001 2.236 1.083 2057. 0.0002
1987–1993 680 1.796 0.535 334.93 0.0001 1.503 0.34 187.368 0.0005
1987–1999 2,587 1.969 37.57 1869.1 0.0001 2.043 40.88 2057.44 0.0002
K. Hung and K.-H. Lee
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 263

Table 9.2 Wilcoxon signed-rank test for earnings forecast accuracy of management and analyst’s
forecasts errors
Year Sample size Negative ranks Positive ranks Ties Z-value Sig.
1999 402 150 230 22
3.119 0.002***
1998 360 209 140 11
5.009 0***
1997 317 147 140 30
0.5 0.96
1996 267 143 119 5
1.567 0.17
1995 226 117 89 20
2.857 0.004***
1994 178 82 90 6
0.339 0.734
1993 157 74 79 4
0.502 0.616
1992 126 60 62 4
1.136 0.256
1991 144 69 69 6
0.407 0.684
1990 120 43 76 1
2.941 0.003***
1989 116 46 67 3
2.7 0.007***
1988 96 28 60 8
3.503 0***
1987 78 27 51 0
3.315 0.001***
1993–1999 1,299 626 607 66
1.592 0.111
1987–1993 1,288 569 665 54
2.138 0.033**
1987–1999 2,587 1,195 1,272 120
0.346 0.73
Negative ranks: absolute error of management’s earnings forecast < absolute error of analyst
earnings forecast
Positive ranks: absolute error of management’s earnings forecast > absolute error of analyst
earnings forecast
Tie: absolute error of management’s earnings forecast ¼ absolute error of analyst earnings forecast
*
Significant level ¼ 0.1, ** significant level ¼ 0.05, *** significant level ¼ 0.01

9.4.2 Factors Influencing the Absolute Errors of Earnings Forecast

9.4.2.1 Firm Size

We argue that the management possesses the relative advantage of having private
insights that the analyst cannot access, and that a larger company has stronger
human and financial resources. Therefore, the management forecasts of corporate
earnings are much more precise. On the other hand, a larger company tends to draw
attentions and is more likely to attract and be followed by financial analysts;
analysts’ forecasts can be objective and accurate as well; however, the results
from our research do not support this argument. Table 9.4 shows that the p-value
of t-parameter for the size of a company (0.478) does not reach a significant level,
indicating the size of a company is not associated with the accuracy of manage-
ment’s earnings forecast. This result does not support the hypothesis H1: manage-
ment’s earnings forecast accuracy increases with the size of company. Tables 9.5
and 9.6 show results of the regression analysis for two subsamples representing the
time period of 1987–1992 and 1993–1999 to investigate the relationship between
company size and accuracy of management’s earnings forecast.
264 K. Hung and K.-H. Lee

Table 9.3 Taiwan stock market volatility from 1987 to 1999

Year Rm (%) Rf (%) Rm
Rf (%) Market volatility
1999 27 4 23 Upmarket
1998
13 6
19 Down market
1997 14 5 9 Upmarket
1996 39 5 34 Upmarket
1995
24 5
29 Down market
1994
13 5
18 Down market
1993 55 6 49 Upmarket
1992
28 6
34 Down market
1991 13 6 7 Upmarket
1990
59 8
67 Down market
1989 51 7 44 Upmarket
1988 124 4 120 Upmarket
1987 138 4 134 Upmarket
Market volatility measure, Rm, suggested by Pettengill et al. (1995), is the last month’s market
return minus the first month’s market return divided by the first month’s market return in a given
year. Rf is the risk-free rate

9.4.2.2 Volatility of Market

In Table 9.4, the p-value of t-parameter in column 4 for the volatility of market of
a company (0.075) does not reach a significant level, which implicates the
accuracy of management’s earnings forecast is not positively associated with
the volatility of market. This result supports the hypothesis H1: management’s
earnings forecast accuracy will not change with market volatility. Further exam-
ining the two sub-tables of Tables 9.4, 9.5, and 9.6, p-value of parameter for the
volatility of market of a company (0.310) indicates that management’s earnings
forecast accuracy will not change with market volatility during 1987–1999. But
the t-parameter for the volatility of market is
2.569, indicating the management
can provide accurate forecast during upmarket, but less accurate forecast during
down market.

9.4.2.3 Trading Volume Turnover

The p-values of t-parameter in column 4 for the trading volume turnover of
a company in all three tables do not reach a significant level. The regression
analysis does not support the hypothesis H8 that trading volume turnover will
affect management earnings forecast accuracy.

9.4.2.4 Corporate Earnings Variances

In Table 9.4, the p-values of t-parameter (2.74) in column 4 for the rate of earnings
divination of a company is 0.01, which shows corporate earnings variances affect
management’s earnings forecast accuracy. Positive value of t-parameter means
management earnings forecast accuracy decreases as corporate earnings variance
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 265

Table 9.4 Regression model for the absolute errors of management earnings forecasts dating
from 1987 to 1999
Independent variable Correlation coefficient t-statistic p-value of t-statistic
Intercept
11.59
0.52 0.604
Size 0.69 0.71 0.478
I1: Market volatility 2.86
1.55 0.122
TR: Rate of trading volume turnover 0.09 0.10 0.923
CV: Corporate earnings variances 2.74 3.50 0.00***
E: Forecasting experience 1.00 1.25 0.212
I2: Cement 0.71 0.11 0.196
I3: Food
0.38
0.07 0.943
I4: Plastic
1.13
0.20 0.844
I5: Textile
0.87
0.19 0.853
I6: Electrical machinery 0.85 0.14 0.888
I7: Electronic equipment and cable
0.78
0.13 0.895
I8: Chemical 14.83 2.70 0.007***
I9: Glass and ceramic
1.36
0.17 0.867
I10: Paper manufacturing
0.84
0.11 0.912
I11: Steel
0.79
0.14 0.891
I12: Rubber
0.46
0.68 0.946
I13: Auto
2.74
0.28 0.782
I14: Electronics
0.82
0.17 0.865
I15: Construction
1.01
0.18 0.856
I16: Transportation
0.53
0.08 0.933
I17: Travel
0.76
0.10 0.924
I18: Insurance
0.92
0.14 0.886
I19: Grocery 0.65 0.09 0.925
R-square 0.016
I2–I19: dummy variables for industry
*
Significant level ¼ 0.10, ** significant level ¼ 0.05, *** significant level ¼ 0.01

increases. This supports the hypothesis H12 that corporate earnings variance has an
effect on management earnings forecast accuracy. From the examination for
Tables 9.5 (1993–1999) and 9.6 (1987–1992), management earnings forecast
accuracy is affected by corporate earnings variances during the recent years
(1993–1999).

9.4.2.5 Type of Industry

To determine whether and which industry will influence the forecast accuracy,
18 industries are selected and represented by a dummy variable Ij. From Table 9.4,
I8 is the only industry that has a significant level for the p-values of t-parameter of
14.73. According to our assumption, I8 represents the chemical industry. Thus, we
conclude that management’s forecasts are reliable for most of the industries studied
in this research, except for the chemical industry.
266 K. Hung and K.-H. Lee

Table 9.5 Regression model for the absolute errors of management earnings forecasts dating
from 1993 to 1999
Independent variable Correlation coefficient t-statistic p-value of t-statistic
Intercept
18.16
0.59 0.552
Size 1.00 0.76 0.450
I1: Market volatility
2.54
1.02 0.310
TR: Rate of trading volume turnover 0.04 0.03 0.976
CV: Corporate earnings variances 3.48 3.42 0.001***
E: Forecasting experience 1.26 1.23 0.218
I2: Cement 1.94 0.21 0.836
I3: Food
0.96
0.15 0.885
I4: Plastic
2.08
0.27 0.789
I5: Textile
1.39
0.23 0.815
I6: Electrical machinery 1.20 0.14 0.875
I7: Electronic equipment and cable
1.42
0.18 0.859
I8: Chemical 20.36 2.84 0.005***
I9: Glass and ceramic
1.91
0.17 0.854
I10: Paper manufacturing
2.22
0.19 0.851
I11: Steel
1.13
0.16 0.875
I12: Rubber
0.74
0.08 0.935
I13: Auto
3.29
0.26 0.798
I14: Electronics
3.01
0.50 0.617
I15: Construction
1.55
0.23 0.822
I16: Transportation
0.37
0.05 0.964
I17: Travel
0.36
0.03 0.974
I18: Insurance
1.99
0.23 0.815
I19: Grocery
0.70
0.08 0.939
R-square 0.016
I2–I19:dummy variables for industry
*
Significant level ¼ 0.10, ** significant level ¼ 0.05, *** significant level ¼ 0.01

9.4.2.6 Forecasting Experience

The results of regression analyses for investigating the relationship of forecasting
experience and management’s earnings forecast accuracy are shown in Tables 9.4,
9.5, and 9.6. All three p-values of t-parameter in column 4 for forecasting experi-
ence indicate that management earnings forecast accuracy is not affected by
previous forecasting experiences. This conclusion does not support the hypothesis
H20 that forecasting experiences affect management’s forecast accuracy. Test of
other hypotheses indicates similar results.

9.5 Conclusions

The results of our research indicate that company size has no effect on any of the
following: management forecast, voluntary management forecast, mandatory
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 267

Table 9.6 Regression model for the absolute errors of management earnings forecasts dating
from 1978 to 1992
Independent variable Correlation coefficient t-statistic p-value of t-statistic
Intercept 15.55 0.87 0.386
Size
0.68
0.87 0.383
I1: Market volatility
2.57
1.78 0.075*
TR: Rate of trading volume turn over) 0.25 0.47 0.636
CV: Corporate earnings variances
0.39
0.59 0.555
E: Forecasting experience 0.57 0.76 0.450
I2: Cement
0.59
0.11 0.914
I3: Food 0.54 0.11 0.913
I4: Plastic
0.16
0.03 0.974
I5: Textile 1.17 0.26 0.799
I6: Electrical machinery
0.68
0.12 0.905
I7: Electronic equipment and cable 0.40 0.08 0.936
I8: Chemical 2.05 0.41 0.680
I9: Glass and ceramic
0.40
0.06 0.955
I10: Paper manufacturing 2.05 0.37 0.709
I11: Steel 0.56 0.10 0.923
I12: Rubber
0.14
0.02 0.981
I13: Auto 1.94 0.24 0.814
I14: Electronics 7.98 1.64 0.102
I15: Construction 0.04 0.01 0.994
I16: Transportation
0.36
0.06 0.949
I17: Travel
0.96
0.16 0.876
I18: Insurance 1.63 0.29 0.770
I19: Grocery 3.78 0.66 0.509
R-square 0.035
I2–I19: dummy variables for industry
*
Significant level ¼ 0.10, ** significant level ¼ 0.05, *** significant level ¼ 0.01

management forecast, and analysts’ forecast. This result agrees with Jaggi
(1980) and Kross and Schreoder (1990) that analyst’s earnings forecast accuracy
is not related with the size of company, but differs from the results suggested by
Bhushan (1989), Das et al. (1998), Clement (1999), Xu (1990), and Jiang (1993)
that company size does influence the relative precision of management or ana-
lysts’ earnings forecasts.
It can be seen that the relative accuracy of management’s earnings forecast and
analyst’s earnings forecast is not affected by market situation across the entire
range of sampled forecasts. There are some indications that forecasting accuracy
is affected by market ups and downs. For instance, the relative accuracy of
voluntary management’s earnings forecast during the entire time period, accuracy
of management’s forecast, and analysts’ earnings forecasts during the years 1978
through 1992 are more accurate when market is up and are less accurate during the
268 K. Hung and K.-H. Lee

down market. This result agrees with what Su has suggested – earnings forecast
accuracy is affected by market volatility, but in different ways. We believe that
due to the fact that more individual investors who are most likely to chase the
market when it is up saturate the Taiwan stock market. They examine corporate
earnings with more caution, so their expectations for companies in general are
more realistic and rational. Therefore, overall earnings forecast accuracy is
increased, and vice versa.
The results of this study reveal that relative accuracy of all four kinds of earnings
forecasts is not the functions of trading volume turnover. This agrees with the
results obtained by Chai, but it disagrees with the results of Jiang (1993). Results of
regression analysis indicate that management’s earnings forecast and analysts’
forecasts are sensitive to the corporate earnings variances. This conclusion proves
the hypothesis supports H13 through H16 and supports the theories of Kross and
Schreoder (1990) and Jiang (1993). We postulate that corporate earnings variance
of earnings surprises is an important indicator for a company’s profitability and its
earnings in the future. The management and analysts use past year’s earnings
surprises to forecast that future earnings, with an assumption of higher forecast
inaccuracy, are a result of a high degree of earnings deviation. Therefore, they will
need to exercise their highest ability in making earnings forecast more accurate. But
we found that the higher the corporate earnings variance is, the lower the forecast
accuracy will be for both the management and for analysts. Corporate earnings
variances should not be used as an important indicator as how a company is
operating, but it represents a complicated business environment it operates
in. The higher the complicity of the business environment is, the less accurate the
prediction/forecast will be.
Analyst earnings forecast and management’s earnings forecast are biased for the
chemical industry over the entire time period of sampled forecast; voluntary
management’s earnings forecasts for textile, electrical machinery, and paper
manufacturing industries are inaccurate during 1993–1999. Mandatory manage-
ment’s earnings forecasts are very inaccurate for the food industry, textile industry,
and travel industry in the time period of 1993–1999. This supports Kross and
Schreoder (1990) who concluded that analyst’s earnings forecast is affected by
the type of industry he/she follows.
The results reveal that the relative accuracy of management’s earnings forecast
and analyst’s earnings forecast do not respond to the differences of forecasters’
previous experiences. But, the relative accuracy of mandatory management’s
earnings forecast for forecasters affect the entire time period and the subsampled
voluntary management’s forecast previous earnings forecast experiences. This
conclusion agrees with Clement’s (1999) finding.
We rationalize that forecast accuracy is positive related to the forecasting
experiences as hypotheses H21 through H24 state. The results from this study
indicate otherwise. The forecast accuracy of mandatory and voluntary manage-
ment’s earnings forecast has a negative relationship with previous forecasting
experiences. We argue that it is because of (1) mis-quantifying variable as
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 269

a proxy of forecasting experiences and (2) only using the past mandatory man-
agement’s earnings forecasts as the base of future focusing without paying
attention to ways on how to reduce the forecasting errors in those forecasts.
Therefore, the more forecasting experiences the forecaster has, the less accurate
the forecast will be.

Methodology Appendix

Sample Selection

This research uses cross-sectional design to examine the relative accuracy of

management and analysts’ earnings forecast. Due to the disclosure regulation in
Taiwan, management’s earnings forecast is classified as two categories: manda-
tory earnings forecast and voluntary earnings forecast1. Samples used are
management’s and analysts’ earnings forecasts at all publicly traded companies
during the time period of 1987–1999. The forecasts are compared with actual
corporate earnings on an annual basis. Voluntary and mandatory management’s
forecast and analysts’ forecast are then used to compare with actual corporate
earnings to evaluate the effects of management motivation and behaviors on their
earnings forecasts.
Management and analysts’ pretax earnings forecast data are collected from
“Taiwan Business & Finance News” during the time period of 1987–1999. Actual
corporate earnings are collected from the Department of Education’s “AREMOS”
database. Only those firms were included in the samples whose stocks were traded on
the Taiwan Stock Exchange before December 31, 1999. Also, forecasts made after
accounting year and before announcement of earnings were excluded from the sample.
Management’s earnings forecast and analysts’ earnings forecast samples for this
research are selected to cover the time period from 1987 to 1999. Available database,
over the 13-year period, consisted of 5,594 management’s earnings forecasts, in
which 2,894 management forecasts are voluntary and 2,700 management’ forecasts
are mandatory. A total of 17,783 analysts’ forecasts are in the database. The selected
samples, presented in Table 9.7, consist of 2,941 management earnings forecasts, of
which 2,046 are voluntary and 1,679 mandatory forecasts and 3,210 analysts’
earnings forecasts. Table 9.7 shows that the average number of analysts’ earnings
forecasts is more than the number of management’s earnings forecasts. A higher
frequency of analysts’ earnings forecasts is expected as an analyst may cover more
than one firm. Most of management’s earnings forecasts are made after 1991; it may
be attributed by the amendment of “Regulation of Financial Report of Stock Issuer”
imposed by the Taiwanese government in 1991. In the new regulation, a new section
dealing with earnings forecast was added requiring company’s management to
disclose its earnings forecasts to the general public. Comparing the number of
management voluntary forecasts and mandatory forecasts, the latter is about 1.5
times more than the former except during the years 1991–1993.
270 K. Hung and K.-H. Lee

Table 9.7 Sample of earnings forecasts by management and analysts from Taiwan database
selected for the study
Year Management voluntary Management mandatory Management Analysts
1999 407 236 319 479
1998 408 238 335 430
1997 384 227 288 376
1996 328 201 223 322
1995 281 193 218 279
1994 219 135 112 247
1993 172 137 84 225
1992 139 78 84 200
1991 156 154 16 175
1990 125 125 NAa 154
1989 123 123 NA 130
1988 112 112 NA 106
1987 87 87 NA 87
Total 2,941 2,046 1,679 3,210
a
Mandatory forecast requirement was introduced in 1991

Variable Definition

Absolute Earnings Errors

The mean value of total corporate earnings before tax is used as proxy of earnings
forecast. Using earnings before tax in the analysis will eliminate other factors, such
as raising cash for capital investment, earnings retention for capital investment, and
stock distribution from paid in capital that might impact the accuracy of the
analysis. Absolute earnings (before tax) forecast error is used to compare the
relative accuracy of management and analysts’ earnings forecasts.
Management’s forecasts errors are calculated as follows:

1 Xn
MFm, i, t ¼  FEm, i, j, t
N j¼1

1 Xn
MF1m, i, t ¼  FE1m, i, j, t
N j¼1

1 Xn
MF2m, i, t ¼  FE2m, i, j, t
N j¼1
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 271

  
 
AFEm, i, t ¼  MFm, i, t
AEi, t =AEi, t 

  
 
AFE1m, i, t ¼  MF1m, i, t
AEi, t =AEi, t 

  
 
AFE2m, i, t ¼  MF2m, i, t
AEi, t =AEi, t 

where,
MFm,i,t: mean management’s pretax earnings forecast for company i at time t
MF1m,i,t: mean management’s voluntary pretax earnings forecast for company
i at time t
MF2m,i,t: mean management’s mandatory pretax earnings forecast for company
i at time t
FEm,i,j,t: management’s jth pretax earnings forecast for company i at time t
FE1m,i,j,t: voluntary management’s jth pretax earnings forecast for company i at
time t
FE2m,i,j,t: mandatory management’s jth pretax earnings forecast for company i in
year t
AFEm,i,t: absolute error of management’s pretax earnings forecast for company
i at time t
AFE1m,i,t: absolute error of voluntary management’s pretax earnings forecast for
company i at time t
AFE2m,i,t: absolute error of mandatory management’s pretax earnings forecast
for company i at time t
AEi,t: actual pretax EPS for company i at time t
Analysts forecast errors are calculated as follows:

1 Xn
MFf, i, t ¼  FEf, i, j, t
n j¼1

  
 
AFEf, i, t ¼  MFf, i, t
AEi, t =AEi, t 

where
MFf,i,t: mean analysts’ pretax earnings forecast for company at time t
FEf,i,j,t: analysts’ jth pre-tax earnings forecast for company i at time t;
AFEf,i,t: absolute error of analyst’s pre-tax earnings forecast for company i at
time t;
AEi,t: actual pre-tax EPS for company i at time t.
272 K. Hung and K.-H. Lee

Company Size
Unlike other previous researchers who used market value of company’s equity
as a indication of size of a company, in this study, a company’s last year’s
total revenue is used as the size of the company. The reason is because
Taiwanese market is not efficient and investors are not informed fully with
information they need during their investment decision-making process; specu-
lations among individual investors are the main cause of the stock market
volatility; thus, market value of company’s equity cannot fully represent
a company’s real size.In order to better control the company size for our
regression analysis, a logarithm of company’s last year’s total revenue is used
as the following:

 
SIZEi, t ¼ ln TAi, t
1

where
SIZEi,t: the size of company i at time t;
TAi,t
1: total revenue of company i at time t
1.

Market Volatility

This study adapts what Pettengill, Glenn N, Sundaram, Sridhar, and Mathur
and Ike used in their research to measure market volatility. Market volatility
is measured as upmarket or down market by using market-adjusted return.
This return is calculated as Rm
Rf, in which Rm is the last month’s
market return minus the last first month’s market return divided by the first
month’s market return in a given year. Rf is the risk-free interest rate in the
same year:

ReturnðMarket
adjustedÞ ¼ Rm
Rf

where
Upmarket if Return (Market-adjusted) >0
Down market if Return (Market-adjusted) <0
A dummy variable is used to identify market volatility. Market volatility is set to
1 if a year’s Return (Market – adjusted) is greater than 0 and set to 0 otherwise. Table 9.3
reports the market volatility of Taiwan market.

Trading Volume Turnover

Trading volume turnover is defined as the value of a company’s stock daily trading
volume divided by the company’s number of shares outstanding. To make this
proxy better fit in the regression analysis, a logarithm is applied to the value and
multiplies by 1,000:
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 273

1X n
AVi, t ¼ V i, j, t
N j¼1

 
AVi, t
1
TRi, t ¼ ln 1, 000 
CSi, t
1

where
Vi,j,t: daily trading volume in day j at time t for company i
AVi,t: mean daily trading volume at time t for company i
CSi,t
1: number of shares outstanding at time t
1 for company i
TRi,t: rate of trading volume turnover at time t for company i

Corporate Earnings Variance

In this research, we only consider the past 3 years’ historical earnings surprises as
a proxy of a company’s corporate earnings variances. Thus, the corporate earnings
variance is defined as the following:

!
sðXÞ
CVi, t ¼ LN  
X

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX n  2
u
u Xt
X
t
t¼1
sðXÞ ¼
n
1

1X 3
X¼ Yt
3 t¼1

where
CVi,t: corporate earnings variance at time t for company i
s(x): actual corporate earnings variance for company i
Xt: actual earnings at time t for company i
X: mean EPS (before tax) for company i
Yt: actual EPS at time t for company i

Type of Industry
There are two major ways to classify industries:
(i) “Industry classification of Republic of China” by State Council in 1987
(ii) Industry classification used by stock exchange house
274 K. Hung and K.-H. Lee

In this research, we use the latter one to classify industries, and a variable Ij is set
to represent nineteen different industries: cement, food, plastic, textile, electrical
machinery, electronic equipment and cable, chemical, glass and ceramic, paper
manufacturing, steel, rubber, auto, electronics, construction, transportation, travel,
insurance, and others.

Proxy for Experience

According to research done by previous researchers, although earnings forecast
accuracy is positively related to management and analysts’ previous forecasting
experiences, it is difficult to quantify the experiences. In this research, we argue that
the accuracy of nth management and analyst’s earnings forecast depends on their
(n
1)th forecasting experience. Therefore, the proxy of experience is defined as
the following:

X
n
Em , i , t ¼ Em, i, j, t
1
j¼1

X
n
E1m, i, t ¼ E1m, i, j, t
1
j¼1

X
n
E2m, i, t ¼ E2m, i, j, t
1
j¼1

X
n
Ef , i , t ¼ Ef , i, j, t
1
j¼1

where
Em,i,t: total number of times of management’s earnings forecasting experience at
time t for company i
E1m,i,t: total number of times of voluntary management’s earnings forecasting
experience at time t for company i
E2m,i,t: total number of times of mandatory management’s earnings forecasting
experience at time t for company i
Ef,i,t: total number of times of analysts’ earnings forecasting experience at time
t for company i

Regression Model

A multiple regression model is used to examine the effect of six factors: firm size,
market volatility, trading volume turnover, corporate earnings variances, type of
industry, and experience.
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 275

Regression model for management’s forecast absolute error percentage:

AFEm, i ¼ a0 þ a1 ðSIZEi Þ þ a2 Ii, 1

 
þ a3 ðTRi Þ þ a4 ðCVi Þ þ a5 Em, i
(9.1)
X
23
þ ak Ii, k
4 þ ei
k¼6

Regression model for voluntary management’s forecast absolute error

percentage:

AFE1m, i ¼ b0 þ b1 ðSIZEi Þ þ b2 Ii, 1

 
þ b3 ðTRi Þ þ b4 ðCVi Þ þ b5 E1m, i
(9.2)
X
23
þ bk Ii, k
4 þ ei
k¼6

Regression model for mandatory management’s forecast absolute error

percentage:
AFE2m, i ¼ c0 þ c1 ðSIZEi Þ þ c2 Ii, 1
 
þ c3 ðTRi Þ þ c4 ðCVi Þ þ c5 E2m, i
(9.3)
X
23
þ ck Ii, k
4 þ ei
k¼6

Regression model for analysts’ forecast absolute error percentage:

AFEf , i ¼ d0 þ d1 ðSIZEi Þ þ d2 Ii, 1

 
þ d3 ðTRi Þ þ d4 ðCVi Þ þ d5 Ef , i
(9.4)
X
23
þ dk Ii, k
4 þ ei
k¼6

where
AFEm,i: absolute error percentage of management’s forecast for company i
AFE1m,i: absolute error percentage of voluntary management’s forecast for
company i
AFE2m,i: absolute error percentage of mandatory management’s forecast for
company i
AFEf,i: absolute error percentage of analysts’ forecast for company i
SIZEi: size of company i
Ii,1: market volatility (1 if market is upmarket, 0 if market is down market)
276 K. Hung and K.-H. Lee

TRi: rate of trading volume turn over for company i

CVi: corporate earnings variances for company i
Em,i: management’s earnings forecasting experience for company i
E1m,i: voluntary management’s earnings forecasting experience for company i
E2m,i: mandatory management’s earnings forecasting experience for company i
Ef,i: analyst earnings forecasting experience for company i
Ii,2
19: type of industry for company i

Wilcoxon Two-Sample Test

If the two-sample groups are related, Wilcoxon two-sample test will be used to
determine the relative earnings forecast accuracy:

W
Eð W Þ
Z ¼ pffiffiffiffiffiffiffiffiffiffiffiffi
V ðW Þ

where
W1: rank sum of absolute error percentage for management’s earnings forecasts
W2: rank sum of absolute error percentage for analysts’ earnings forecasts
W: smaller value between W1 and W2
E(W): expected values of W-distribution
pffiffiffiffiffiffiffiffiffiffiffiffi
VðWÞ: deviation of W-distribution

References
Bartley, J. W., & Cameron, A. B. (1991). Long-run earnings forecasts by managers and financial
analysts. Journal of Business Finance & Accounting, 18, 21–41.
Bhushan, R. (1989). Firm characteristics and analyst following. Journal of Accounting & Eco-
nomics, 11, 255–274.
Chia, Yuxuan (1994). Examination of mandatory management earnings forecasts and the rate
of trading turnover. Unpublished thesis, Accounting Department, National Chengchi
University.
Clement, B. M. (1999). Analyst forecast accuracy: Do ability, resources, and portfolio complexity
matter? Journal of Accounting & Economics, 27, 285–303.
Das, S., Carolyn, B., Levine, K. S. (1998). Earnings predictability and bias in analysts’ earnings
forecasts. Accounting Review, 73, 277–294.
Imhoff, E. A., & Pare, P. (1982). Analysis and comparison of earnings forecasts agents. Journal of
Accounting Research, 20, 429–439.
Jaggi, B. (1980). Further evidence on the accuracy of management forecasts vis-à-vis analyst’
forecasts. The Accounting Review, 55, 96–101.
Jiang, Jianquan (1993). Factors associated with analyst’s earnings forecast accuracy
under different assumptions in Taiwan. Unpublished thesis, School of Business, Taiwan
University.
Kross, W., Ro, B., & Schreoder, D. (1990). Earnings expectations: The analysts information
advantage. The Accounting Review, 65, 461–476.
9 Determinations of Corporate Earnings Forecast Accuracy: Taiwan Market Experience 277

Pettengill, G. N., Sundaram, S., & Mathur, I. (1995). The conditional relation between beta and
returns, Journal of Financial and Quantitative Analysis, 30(1), 101–116
Su, Yongru (1996). Comparison of relative accuracy of management and analysts earnings
forecast. Unpublished thesis, Accounting Department, Soochow University.
Xu, Xiubin (1990). A examination of analysts earnings forecasts accuracy. Unpublished thesis,
Accounting Department, National Chengchi University.
Xu, Jinjuan (1992). The effect of management earnings forecast on portfolio management.
Unpublished thesis, Accounting Department, National Chengchi University, Taiwan.
Yie, Meiling (1996). Examination on factors in management and financial analyst earnings
forecasting accuracy. Unpublished thesis, Accounting Department, National Chung Cheng
University.
Market-Based Accounting Research
(MBAR) Models: A Test of ARIMAX 10
Modeling

Anastasia Maggina

Contents
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
10.2 Review of the Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
10.3 Methodology and Models Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
10.4 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
10.5 Empirical Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
10.5.1 Unit Root Tests: Testing for Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
10.5.2 Forecasting Dependent Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
10.6 Conclusions and Suggestions for Further Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296

Abstract
The purpose of this study is to provide evidence drawn from publicly traded
companies in Greece as far as some of the standard models of accounting earnings
and returns relations mainly collected through the literature. Standard models such
as earnings level and earnings changes have been investigated in this study. Models
that fit better to the data drawn from companies listed on the Athens Stock
Exchange have been selected employing autoregressive integrated moving average
with exogenous variables (ARIMAX) models. Models I (price on earnings model),
II (returns on change in earnings divided by beginning-of-period price and prior
period), V (returns on change in earnings over opening market value),VII (returns
deflated by lag of 2 years on earnings over opening market value), and IX
(differenced-price model) have statistically significant coefficients of explanatory
variables. In addition, model II (returns on change in earnings divided by
beginning-of-period price and prior period with MSE (minimum squared error)

A. Maggina
Business Consultant/Research Scientist, Avlona, Attikis, Greece
e-mail: anastasiamaggina@yahoo.gr; a.maggina@yahoo.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 279
DOI 10.1007/978-1-4614-7750-1_10,
# Springer Science+Business Media New York 2015
280 A. Maggina

loss function in ARIMAX (2,0,2)) is prevalent. These models take place with
backward-looking information instead of forward-looking information that recent
literature is assessed. Application of generalized autoregressive conditional
heteroscedasticity (GARCH) models is suggested for further future research.

Keywords
Market-based accounting research • Price on earnings model • Earnings level •
Earnings change • Return models • Autoregressive-moving average with exog-
enous variables • Minimum value of squared residuals (MSE loss function) •
Unit root tests • Stationarity • Dickey-Fuller test

10.1 Introduction

The relationship between accounting earnings and stock prices has been discussed both
in the accounting and financial literature. In a consideration of market-based account-
ing research, the association between returns and earnings has been very low due to
(i) poor specification of the estimating equation, (ii) poor informational properties of
reported earnings, (iii) inappropriate choice of the assumed proxy for expected earn-
ings, and (iv) the availability of more timely sources of the value-relevant information
in earnings statement (Strong and Walker 1993). This has been resolved by allowing
for time-series and cross-sectional variation in the regression parameters, by including
an earnings yield and partitioning all-inclusive earnings into pre-exceptional, excep-
tional, and extraordinary components. Yet, empirical evidence on returns-earnings
association in certain applications is very strong (Easton and Harris 1991; Kothari
and Zimmerman 1995). Standard empirical models, that is, price-earnings association,
and returns-earnings association (earnings level and earnings changes) have been
investigated in the USA and UK with no empirical evidence from at least the rest of
Europe. The purpose of this paper is to develop an empirical background by investi-
gating whether some of the standard models that have been collected through the
literature would be relevant for evaluating accounting earnings/returns associations in
a stock market being in transition from an emerging to a matured one. In other words,
we purport to select those models that better fit to available data. The rationale of the
Greek stock market response is one major objective, while the other is to make some
comparisons with other findings presented in the literature.
The Athens stock exchange has been established on 1876. It was an emerging
market some years ago and has run up to a mature one. It is a normal market with no
surprises, and investments are not only a place for corporations to get financing;
more importantly it is a place to create wealth. Registered companies are young,
growth-oriented to long-established enterprises. The Greek stock market has played
a great role in the economic development of the country in the last half of the
twentieth century. Facts that have influenced the ASE is the inclusion of the country
in the Economic and Monetary Union and the crash of 1999 that affected the life of
many Greek families. The ASE operates in a country which has been in deep
recession since 2008.
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 281

The paper is organized as follows: Literature is discussed in Sect. 10.2.

Methodology and model building process are described in Sect. 10.3.
Section 10.4 discusses the sample design. Section 10.5 presents empirical findings.
Conclusions are summarized in last Sect. 10.6.

10.2 Review of the Literature

Common interest of both accounting and finance scholars and users of financial
statements (primarily, financial analysts and investors) is to obtain concrete and
increased knowledge of the association between accounting earnings and stock
prices for a more or less predictive ability explanation. An evaluation of the
informational content of, at least, the most basic accounting numbers contributes
to the improvement of models. Among the various treatments of the relation
between accounting earnings and stock prices, the seminal work of Ball and
Brown (1968) is precedent. Their work indicates that accounting earnings and
some of its components exhibit an information content drawn through stock prices.
Later on, Brown et al. (1985) worked on the relative ability of the current innova-
tion in annual earnings and the revision in next year’s earnings forecasts to explain
changes in stock prices. However, it has empirically been tested (Fama et al. 1969;
Fama 1970). Successive changes in individual common stocks are very nearly
independent because of an efficient market in which adjustments to new informa-
tion are rapidly made (Mandelbrot 1966; Samuelson 1965). Beaver (1970) and
Rosenberg and Marathe (1975) worked to discover financial statement components
that are related to risk and, thus, predict expected stock returns. They maintain
that financial statements constitute an assured source of information (and can be
prepared in a short time period) and, consequently, the direction of 1-year-ahead
earnings changes affects stock prices. Their results indicate that the summary
measure robustly predicts future stock returns.
Beaver et al. (1980) developed a model known as BLM theory. They followed
the traditional price-earnings relation and tested for the information content of
prices with respect to future earnings. BLM regress percentage change in price
which may contain information about future earnings not reflected in the current
earnings. Lev and Ohlson (1982) describe price and return models as complemen-
tary. Given that stock prices are associated with accounting earnings, analysts
forecast the reported accounting numbers, and as Brown et al. (1985) maintain,
analysts use their 1-year-ahead forecasts to convey their expectations about perma-
nent earnings. In all the above treatments of the subject matter referring to account-
ing earnings and stock prices, the more and more prevailing role of managerialism
could be more or less emphatic. Trueman (1986) explained why managers would be
willing to disclose forecasts of higher or lower than expected earnings. Undoubt-
edly, the disclosure of favorable accounting numbers or the disclosure of forecasts
of favorable expected earnings contributes to an increase of the market value of the
firm. Hirst et al. (2008) assert that managers often issue earnings forecasts to correct
information asymmetry problems and, thus, influence their firm’s stock price.
282 A. Maggina

Lipe and Kormendi (1987) found that stock returns were considered as a function of
the revisions in expectations of earnings. They show that the stock return reaction to
earnings is a function of (1) the time-series properties of earnings, (2) the interest
rate used to discount expected future earnings, and (3) the relative ability of
earnings. Furthermore, Freeman (1987) maintains that the relation between security
prices and firm-specific information is associated with the market value of a firm’s
common equity and concludes that the magnitude of the abnormal returns related to
earnings is a decreasing function of firm size. Of much interest is a hypothesis
tested by Freeman (1987) which implies that (i) the abnormal security returns
related to accounting earnings occur earlier for large than for small firms and
(ii) abnormal returns are lower for large firms and higher for small firms. The
information content theorists consider size as an important conditioning variable
when testing the information content of prices with respect to future earnings and
contemporaneous price changes.
Beaver et al. (1987), Collins et al. (1987), Collins and Kothari (1989), and many
others show that unexpected earnings for a year are correlated with returns
from a prior year. Christie (1987) concludes that while return and price models
are economically equivalent, return models are econometrically less
problematic. Meanwhile, Holthausen and Verrecchia (1988) worked on price
changes when information is announced. Landsman and Magliolo (1988) argue
that price models are superior to return models for certain applications. According
to Cornell and Landsman (1989), stock prices respond to earnings announcements.
Alternatively stated, unexpected increases in earnings are associated with a rise in
stock prices, and unexpected decreases in earnings are associated with a fall in stock
prices. Relating unexpected accounting earnings and security prices aims to assess
the information content of the latter (Collins and Kothari 1989). Conservative
proponents support the view that financial statement ratios are the basic tools for
evaluating and predicting accounting earnings and, consequently, security prices.
Ou and Penman (1989) maintain that the relationship of financial statement char-
acteristics to value is not apparent. Easton and Zmijewski (1989) and that of Board
and Walker (1990) analyze a coefficient that measures the response of stock prices
to accounting earnings coefficient. They measure the response of stock prices to
accounting earnings announcements, and they empirically show that the higher this
coefficient, the smaller the stock price changes.
Lipe (1990) worked on the relation between stock returns and accounting earnings,
assuming that market observes current-period information other than earnings. In the
process of relating stock prices to accounting earnings, (i) a coefficient which mea-
sures the stock-return response to a one dollar earnings changes as a function of both
“predictability” and “persistence” of earnings, and (ii) the variance of stock price
changes during the reporting of earnings has been tested in the literature (Cho and
Jung 1991; etc.). Easton and Harris (1991) presented models relating earnings vari-
ables and security returns, concluding that both current earnings level and the earnings
change variables play a role in the security valuation. They are in fact correlated.
Strong and Walker (1993) used a panel regression approach to examine the association
between annual stock price returns and reported earnings figures of industrial
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 283

companies in the UK and proved through an analysis of earnings components that

models of the relation between earnings and returns that focus exclusively on the
deflated first difference of earnings are misspecified. Kahya et al. (1993) showed that
earnings growth rates are predictable using past earnings growth rates, stock price
returns are predictable using past earnings growth rates as well as stock price returns,
and firm size has no incremental explanatory power with respect to equity prices.
Kothari and Zimmerman (1995) provide empirical results confirming that price
models’ earnings response coefficients are less biased and argue that the price specifi-
cation suffers more from heteroscedasticity/misspecification problems than the return
model. They conclude that in some research contexts, the combined use of both price
and return models may be useful. Chen et al. (2001) obtain evidence of value relevance
of accounting information in China based on a return and a price model according to
both the pooled cross section and time-series regressions or the year-by-year regressions,
and find that value relevance is higher for companies using only A shares to domestic
investors despite of the lack of alternative information sources (i.e., earnings forecasts,
financial analysts), the lack of a sufficient level of corporate governance, and the even
recent new phenomenon of independent auditing in China. Lundholm and Myers (2002)
examined how a firm’s disclosures affect the mix of earnings information reflected in its
annual stock return and found that increased disclosure activity “brings the future
forward” into current stock returns. Chen et al. (2011) examined the properties of
accounting numbers of listed firms in China by investigating the interplay between
accounting earnings and stock prices. They found that core earnings (or operating
income) have a greater association with contemporaneous stock returns than
nonoperating revenues and expenses, and they also found that different types of firm
ownership may have different impacts on the information content of earnings
components.

10.3 Methodology and Models Used

Literature concerning accounting earnings and stock prices has been formulated
in the framework of an earnings-based valuation model expanded with the
dividend irrelevance proposition. These standard models which have also been
tested by Easton and Harris (1991) and Zimmerman and Kothari (1995) are as
follows:
I. Price on earnings

Pi, j ¼ a þ bAi, j þ ei, j  (10.1)

II. Returns on change in earnings divided by beginning-of-period price and prior

period earnings divided by the price at the beginning of the return period


 


Ai, j =Pi, j1 ¼ ai, j þ bi, j Ai, j  Ai, j1 =Pi, j1 þ Ai, j1 =Pi, j1 (10.2)
284 A. Maggina

III. Returns on earnings over opening market value


 

Pi, j  Pi, j1 þ di, j =Pi, j1 ¼ ai, j þ bi, j Ai, j =Pi, j1 þ ei, j (10.3)

IV. Returns on prior earnings model over opening market value


 

Pi, j  Pi, j1 þ di, j =Pi, j1 ¼ ai, j þ bi, j Ai, j1 =Pi, j1 þ ei, j (10.4)

V. Returns on change in earnings over opening market value


 

Pi, j  Pi, j1 þ di, j =Pi, j1 ¼ ai, j þ bi, j Ai, j  Ai, j1 =Pi, j1 þ ei, j (10.5)

VI. Returns on change in earnings over opening market value and on earnings over
opening market value

 

Pi, j  Pi, j1 þ di, j =Pi, j1 ¼ a þ b1i, j Ai, j  Ai
, j1 =Pi, j1
þ b2i, j Ai, j =Pi, j1 þ ei, j (10.6)

VII. Returns (deflated by lag of 2 years) on earnings over opening market value

 

Pi, j  Pi, j1 þ di, j =Pi, j2 ¼ ai, j þ bi, j Ai, j =Pi, j1 þ ei, j (10.7)

VIII. Return model regressed on earnings over opening market value

Pi, j =Pi, j1 ¼ a þ b1i, j Ai, j =Pi, j1 (10.8)

IX. Differenced-price model

Pi, j  Pi, j1 ¼ Ai, j  Ai, j1 (10.9)

where
Pi,j ¼ stock price (per share) of firm i in period j
Ai,j ¼ earnings per share of firm i in period j
di,j ¼ dividend per share of firm i in period j
a ¼ a constant in a linear relationship(intercept parameter)
b1,b2 ¼ a slope parameter or a coefficient in a linear regression
i ¼ cross-selection item, j ¼ time-series item
To be familiar and consistent with the existing literature, some requirements are
stressed. For example, earnings per share divided by price at the beginning of the
return period (Ai,j/Pi,j  1) refers to current earnings level variable. Change in
earnings divided by beginning-of-period price refers to earnings change variable
[(Ai,j  Ai,j  1)/Pi,j  1]. Thus far, the models that have been selected to be tested
express the following:
Model I: Expresses price as a multiple of earnings.
Model II: Expresses historical price-earnings ratios with an earnings change versus
earnings levels explanation form.
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 285

Model III: From an earnings valuation perspective, earnings levels(divided by

beginning-of-period price) will be associated with returns.
Model IV: Illustrates a linear relation between prior earnings divided by beginning-
of-period price and security returns over that period.
Model V: Regresses annual security returns on change in earnings divided by
beginning-of-period price.
Model VI: Expresses the contribution of change in earnings versus earnings levels
in the explanation of stock returns.
Model VII: Returns (deflated by lag of 2 years) on earnings over opening market value.
Model VIII: Expresses returns on earnings level divided by beginning-of-period price.
Model IX: Expresses differenced-price model.
As in Easton and Harris (1991), the models under investigation have been based
on either the book value valuation model or the earnings valuation model.
The book value valuation model indicates that

Pij ¼ BVij þ uij (10.10)

Taking first differences we have

DPij ¼ DBVij þ uij (10.11)

But in general

DBVij ¼ Aij  dij (10.12)

Substituting (3) into (2), rearranging and dividing by Pij1 yields

(DPij + dij)/Pij1 ¼ Aij/Pij1 + uij (model III)
On the other hand,

Pij ¼ rAij þ uij (10.13)

Given the dividends irrelevance proposition, we have

Pij þ dij ¼ rAij þ uij (10.14)

It follows that (DPij + dij)/Pij1 ¼ r(DAij/Pij1) + uij (model V)

ARIMAX (autoregressive integrated moving average with exogenous variables) as
a suitable technique for nonstationarity time-series modeling is employed in this study.

10.4 Sample Selection

The whole population containing all Greek-listed companies in the Athens

Stock Exchange is investigated in this study. The total number of companies
amounts to 513 companies. The main source of data is the Athens Stock
286 A. Maggina

Exchange Annual Yearbook, the annual statistical bulletin, and the Internet.
Total number of companies refers to the time period 1974 up to 2005 (the most
recently available data when writing the paper). The full sample (1974–2005) is
separated in two samples (1974–1999 and 2000–2005) which correspond to
two-time periods, that is, before the Euro currency and the Euro era. As in
Kothari and Zimmerman (1995), to avoid any undue influence of extreme
observations, the largest and the lowest 1 % of observations is excluded from
the sample. EPS take positive or zero values. All firms have a December fiscal
year-end. Annual earnings include those from discontinued operations and
extraordinary items.

10.5 Empirical Findings

For each year, a nonconstant number of companies is available. We define the total
number of companies at each year t as nt. We would like to estimate the following
nine models for the period from 1974 to 2005:
Model I


Pt ¼ a þ bAt þ et , for et  N 0; s2 , (10.15)
X
nt X
nt
where Pt ¼ nt 1 Pi, t and At ¼ nt 1 Ai, t:
i¼1 i¼1
Model II

At At  At1 At1 

¼ a þ b1 þ b2 þ et , for et  N 0; s2 , (10.16)
Pt1 Pt1 Pt1

At Xnt
Ai, t At  At1 Xnt
Ai, t  Ai, t1 At1
where ¼ nt 1 , ¼ nt 1 , and ¼ nt 1
Pt1 i¼1
P i , t1 P t1 i¼1
P i , t1 P t1
Xnt
Ai, t1
.
i¼1
Pi, t1

Model III

Pt  Pt1 þ dt At 

¼ a þ b1 þ et , for et  N 0; s2 , (10.17)
Pt1 Pt1
Pt  Pt1 þ dt Xnt
Pi, t  Pi, t1 þ di, t At Xnt
Ai, t
where ¼ nt 1 and ¼ nt 1 .
Pt1 i¼1
P i, t1 P t1 i¼1
P i, t1
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 287

Model IV

Pt  Pt1 þ dt At1 

¼ a þ b1 þ et , for et  N 0; s2 , (10.18)
Pt1 Pt1
Pt  Pt1 þ dt Xnt
Pi, t  Pi, t1 þ di, t Xnt
Ai, t1
where ¼ nt 1 and APt1 ¼ n t
1
.
Pt1 i¼1
P i, t1
t1
i¼1
Pi, t1
Model V

Pt  Pt1 þ dt At  At1 

¼ a þ b1 þ et , for et  N 0; s2 , (10.19)
Pt1 Pt1

Pt  Pt1 þ dt Xnt
Pi, t  Pi, t1 þ di, t
where ¼ nt 1 and
Pt1 i¼1
Pi, t1
At  At1 X Ai, t  Ai, t1
nt

¼ nt 1 .
Pt1 i¼1
Pi, t1
Model VI

Pt  Pt1 þ dt At  At1 At 

¼ a þ b1 þ b2 þ et , for et  N 0; s2 , (10.20)
Pt1 Pt1 Pt1

Pt  Pt1 þ dt X nt
Pi, t  Pi, t1 þ di, t
where ¼ nt 1 ,
Pt1 i¼1
Pi, t1
At  At1 Xnt
Ai, t  Ai, t1 At Xnt
Ai , t
¼ nt 1 , and ¼ nt 1 .
Pt1 i¼1
P i, t1 P t1 i¼1
P i, t1
Model VII

Pt  Pt1 þ dt At 

¼ a þ b1 þ et , for et  N 0; s2 , (10.21)
Pt2 Pt1
Pt  Pt1 þ dt Xnt
Pi, t  Pi, t1 þ di, t At Xnt
Ai, t
where ¼ nt 1 and ¼ nt 1 .
Pt2 i¼1
Pi, t2 Pt1 i¼1
Pi, t1
Model VIII

Pt At 

¼ a þ b1 þ et , for et  N 0; s2 , (10.22)
Pt1 Pt1
Pt X Pi , t
nt
At Xnt
Ai , t
where ¼ nt 1 and ¼ nt 1 .
Pt1 i¼1
Pi, t1 Pt1 i¼1
Pi, t1
288 A. Maggina

Table 10.1 The ARIMAX (p,d,q) specification for each of the nine models and the relative MSE
loss functions
Model I ARIMAX(1,0,1) 2.766371
Model II ARIMAX(2,0,2) 0.003763
Model III ARIMAX(2,0,2) 0.741482
Model IV ARIMAX(2,0,2) 0.719591
Model V ARIMAX(0,1,1) 0.732002
Model VI ARIMAX(2,0,2) 0.584564
Model VII ARIMAX(2,1,1) 0.888239
Model VIII ARIMAX(2,0,2) 0.762231
Model IX ARIMAX(2,1,2) 2.245854

Model IX



ðPt  Pt1 Þ ¼ a þ b1 ðAt  At1 Þ þ et , for et  N 0; s2 , (10.23)

X
nt


where ðPt  Pt1 Þ ¼ nt 1 Pi, t  Pi, t1 and
i¼1
X
nt


ðAt  At1 Þ ¼ nt 1 Ai, t  Ai, t1 .
i¼1
The nine aforementioned models are characterized by autocorrelated and
heteroscedastic residuals. Thus, the models are expanded in the ARIMAX
(autoregressive integrated moving average) framework (for details about ARIMAX
modeling, the interested reader is referred to Box and Jenkins (1976)) in order to
model the autocorrelated residuals. Moreover, we take into consideration White’s
(1980) heteroscedasticity-consistent covariance matrix estimator which provides
correct estimates of the coefficient covariances in the presence of heteroscedasticity
of unknown form.
The ARIMAX (p,d,q) models are estimated in the following form:

ð1  L!Þd yt ¼ Xt b þ et !
X
p X q
1 ci L et ¼ 1 þ
i
d i L et
i (10.24)
i¼1 i¼1
et  N ð0; s2 Þ,

where yt is the dependent variable, Xt is the vector of explanatory variables, L is the

lag operator, and b is a vector of parameters to be estimated. ci, for i ¼ 1,. . .,p, and
di, for i ¼ 1,. . .,q, are also parameters to be estimated.
For each of the nine models, the ARIMAX (p,d,q) specification is estimated for
p ¼ 0,1,2, d ¼ 0,1,2 and q ¼ 0,1,2. Therefore, for each model, 27 ARIMAX
specifications are estimated. The ARIMAX (p,d,q) specification with the minimum
value of squared residuals (MSE loss function) is selected as the most appropriate.
The following Table presents the selected specifications for each model (Table 10.1).
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 289

Hence, the estimated models are the following:

Model I
Pt ¼ 8:5  5:1At þ et
et ¼ 1:05et1 þ et  1:46et1 (10.25)
et  N ð0; s2 Þ:

Model II
At At  At1 At1
¼ 0:004 þ 0:99 þ 1:04 þ et
Pt1 Pt1 Pt1 (10.26)
et ¼ 1:5et1 0:84et2 þ et  1:14et1 þ 0:52et2
et  N ð0; s2 Þ:

Model III

Pt  Pt1 þ d t At
¼ 0:28  0:21 þ et
Pt1 Pt1 (10.27)
et ¼ 0:32et1 þ 0:46et2 þ et þ 0:92et1  1:16et2
et  N ð0; s2 Þ:
Model IV
Pt  Pt1 þ d t At1
¼ 0:47 þ 2:48 þ et
Pt1 Pt1 (10.28)
et ¼ 0:21et1 þ 0:36et2 þ et  0:68et1  1:42et2
et  N ð0; s2 Þ:

Model V

Pt  Pt1 þ d t At  At1
ð 1  LÞ ¼ 0:055  0:5 þ et
Pt1 Pt1 (10.29)
et ¼ et  1:52et1
et  N ð0; s2 Þ:

Model VI

Pt  Pt1 þ d t At  At1 At
¼ 0:33  2:6 þ 2:4 þ et
Pt1 Pt1 Pt1 (10.30)
et ¼ 0:4et1 þ 0:3et2 þ et 0:37et1  1:9et2
et  N ð0; s2 Þ:

Model VII

Pt  Pt1 þ d t At
ð1  LÞ ¼ 0:2  0:76 þ et
Pt2 Pt1 (10.31)
et ¼ 0:1et1 þ 0:3et2 þ et  1:7et1
et  N ð0; s2 Þ:
290 A. Maggina

Model VIII

Pt At
¼ 1:3  0:22 þ et
Pt1 Pt1 (10.32)
et ¼ 0:31et1 þ 0:46et2 þ et þ 0:94et1  1:1et2
et  N ð0; s2 Þ:
Model IX

ð1  LÞðPt  Pt1 Þ ¼ 0:16  5:54ðAt  At1 Þ þ et

et ¼ 0:25et1 þ 0:13et2 þ et  1:2et1  0:64et2 (10.33)
et  N ð0; s2 Þ:

In the sequel, for the models being estimated, the estimates of the coefficients,
their standard errors, the t-ratios, and their p-values are presented. According to the
t-statistics, computed as the ratio of the coefficients to their standard errors, the
coefficients of the explanatory variables are statistically significant in models
I (price on earnings), II (returns on change in earnings divided by beginning-of-
period price and prior period earnings divided by the price at the beginning of the
return period), V (returns on change in earnings over opening market value), VII
(returns (deflated by lag of 2 years) on earnings over opening market value, and IX
(differenced-price model).
Thus, we conclude that these models explain the relationship between dependent
and explanatory variables, whereas the models III (returns on earnings over opening
market value), IV (returns on prior earnings model over opening market value), VI
(returns on change in earnings over opening market value, and on earnings over
opening market value), and VIII (return model regressed on earnings over opening
market value) fail to explain any strong relationship for the variables under
investigation.
Model I

Coefficient Std. error t-statistic Prob.

C 8.533220 14.06649 0.606635 0.5494
A 5.132325 1.427713 3.594788 0.0013
AR(1) 1.059546 0.183145 5.785297 0.0000
MA(1) 1.461802 0.439241 3.328021 0.0026

Model II

Coefficient Std. error t-statistic Prob.

C 0.004505 0.001576 –2.857574 0.0092
A_A_1_TO_P_1 0.999564 0.000720 1388.375 0.0000
A_1_TO_P_1 1.045894 0.006215 168.2873 0.0000
AR(1) 1.586004 0.418002 3.794251 0.0010
(continued)
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 291

Coefficient Std. error t-statistic Prob.

AR(2) –0.846287 0.326826 –2.589408 0.0167
MA(1) –1.149986 0.488050 –2.356289 0.0278
MA(2) 0.523007 0.324461 1.611925 0.1212

Model III

Coefficient Std. error t-statistic Prob.

C 0.283162 0.469953 0.602534 0.5527
A_P_1 0.215456 0.245653 0.877076 0.3895
AR(1) 0.328559 0.487341 0.674188 0.5069
AR(2) 0.461880 0.150453 3.069922 0.0054
MA(1) 0.922833 0.503703 1.832097 0.0799
MA(2) 1.168622 0.757600 1.542532 0.1366

Model IV

Coefficient Std. error t-statistic Prob.

C 0.474203 0.250825 1.890572 0.0713
A_1_TO_P_1 2.482914 2.237444 1.109710 0.2786
AR(1) 0.210692 0.313382 0.672316 0.5081
AR(2) 0.363890 0.330713 1.100322 0.2826
MA(1) 0.683117 0.316795 2.156340 0.0418
MA(2) 1.427611 0.387230 3.686727 0.0012

Model V

Coefficient Std. error t-statistic Prob.

C 0.055179 0.014928 3.696284 0.0010
A_A_1_TO_P_1 0.508631 0.167841 3.030438 0.0053
MA(1) 1.529472 0.264051 5.792342 0.0000

Model VI

Coefficient Std. error t-statistic Prob.

C 0.333214 0.401790 0.829324 0.4158
A_A_1_TO_P_1 2.616280 1.897450 1.378840 0.1818
A_P_1 2.428167 1.979060 1.226929 0.2328
AR(1) 0.442328 0.194141 2.278386 0.0328
AR(2) 0.322828 0.140587 2.296282 0.0316
MA(1) 0.373195 0.331781 1.124822 0.2728
MA(2) 1.954250 0.408247 4.786934 0.0001
292 A. Maggina

Model VII

Coefficient Std. error t-statistic Prob.

C 0.238146 0.139477 1.707424 0.1018
A_P_1 0.760945 0.282609 2.692572 0.0133
AR(1) 0.094551 0.236234 0.400242 0.6928
AR(2) 0.326603 0.207761 1.572016 0.1302
MA(1) 1.739517 0.430607 4.039685 0.0005

Model VIII

Coefficient Std. error t-statistic Prob.

C 1.336294 0.475802 2.808510 0.0100
A_P_1 0.226643 0.247267 0.916591 0.3689
AR(1) 0.313565 0.480801 0.652171 0.5208
AR(2) 0.462875 0.157081 2.946732 0.0072
MA(1) 0.937106 0.494468 1.895180 0.0707
MA(2) 1.106481 0.744072 1.487061 0.1506

Model IX

Coefficient Std. error t-statistic Prob.

C 0.169538 0.371999 0.455748 0.6530
A_MINUS_A_1 5.545373 0.888799 6.239178 0.0000
AR(1) 0.254505 0.276560 0.920251 0.3674
AR(2) 0.135901 0.243004 0.559251 0.5816
MA(1) 1.192648 0.472557 2.523820 0.0193
MA(2) 0.640639 0.667506 0.959750 0.3476

10.5.1 Unit Root Tests: Testing for Stationarity

We estimate the augmented Dickey and Fuller (1979) test in order to investigate the
null hypothesis of a unit root or nonstationarity of the time series under investiga-
tion, i.e., yt. The Dickey-Fuller test is carried out by estimating:
X
l
Dyt ¼ a0 þ a1 yt1 þ a2 t þ bi Dyti þ et (10.34)
i¼1

where et  N(0,s2), and D is the difference operator, or D ¼ (1  L). The lag order
Xl
l of bi Dyti is selected based on the Schwarz (1978) information criterion.
i¼1
The null and the alternative hypotheses may be written as
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 293

H 0 : a1 ¼ 0
(10.35)
H A : a1 6¼ 0:

A rejection of the null hypothesis indicates that the series yt is

stationary. The Eq. 10.20 is estimated in three different versions: (i) for
a0 ¼ a2 ¼ 0, (ii) for a2 ¼ 0, and (iii) for a0 6¼ 0 and a2 6¼ 0. The variables we are
interested in are
Pt in model I; At in model I; PAt1t in models II, III, VI, VII, and VIII; AtPA t1
t1
in
At1 Pt Pt1 þd t
models II, V, VI; Pt1 in models II and IV; Pt1 in models III, IV, V, and VI;
Pt Pt1 þd t
Pt2 in model VII; PPt1t in model VIII; (Pt  Pt1) in model IX; and (At  At1)
in model IX.
The test statistics and the relative p-values from testing 21 for the Eq. 10.20 for
a0 ¼ a2 ¼ 0 are the following:

Variable Test statistic p-value

Pt 2.256635 0.0254
At 3.414078 0.0013
At
Pt1
5.218131 0.00
At At1 6.010163 0.00
Pt1
At1
Pt1
3.238033 0.00
Pt Pt1 þ d t
Pt1
4.363369 0.00
Pt Pt1 þ dt
Pt2
5.920676 0.00
Pt
Pt1
2.698176 0.0087
(PtPt1) 7.122593 0.00
(At  At  1) 7.294227 0.00

The test statistics and the relative p-values from testing 21 for the Eq. 10.20 for
a2 ¼ 0 are the following:

Variable Test statistic p-value

Pt 3.625190 0.0109
At 5.722834 0.00
At
Pt1
5.587381 0.00
At  At1 6.107705 0.00
Pt1
At1
Pt1
5.774024 0.00
Pt  Pt1 þ d t
Pt1
4.399238 0.0016
Pt  Pt1 þ d t
Pt2
6.043400 0.00
Pt
Pt1
4.345041 0.0018
(Pt  Pt1) 6.998199 0.00
(At  At1) 7.157803 0.00
294 A. Maggina

The test statistics and the relative p-values from testing 21 for the Eq. 10.20 for
a0 6¼ 0 and a2 6¼ 0 are the following:
Variable Test statistic p-value
Pt 3.565939 0.0497
At 5.698735 0.0003
At
Pt1
5.667882 0.0004
At  At1
Pt1
6.213167 0.0001
At1
Pt1
5.673596 0.0003
Pt  Pt1 þ d t
Pt1
4.332965 0.0092
Pt  Pt1 þ d t
Pt2
5.931949 0.0002
Pt
Pt1
4.290153 0.0102
(Pt  Pt1) 6.875067 0.00
(At  At1) 7.024534 0.00

In all the cases, the p-values are less than a 5 % level of significance. Therefore,
the null hypothesis is rejected at any case. Hence, the series are defined, by the
augmented Dickey-Fuller tests, to be stationary.

10.5.2 Forecasting Dependent Variables

For the following models, the dependent variable is predicted for the year 2006,
after adding in the dataset the values of the explanatory variables for the year 2006:
Model I

Pt ¼ 8:5  5:1At þ et
et ¼ 1:05et1 þ et  1:46et1 (10.36)
et  N ð0; s2 Þ:

Model II

At At  At1 At1
¼ 0:004 þ 0:99 þ 1:04 þ et
Pt1 Pt1 Pt1 (10.37)
et ¼ 1:5et1 0:84et2 þ et  1:14et1 þ 0:52et2
et  N ð0; s2 Þ:

Model III
Pt  Pt1 þ d t At
¼ 0:28  0:21 þ et
Pt1 Pt1 (10.38)
et ¼ 0:32et1 þ 0:46et2 þ et þ 0:92et1  1:16et2
et  N ð0; s2 Þ:
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 295

Table 10.2 The 1-year-ahead forecasts (year 2006) of the dependent variables for the nine
models
Number of model Best model Forecast of the dependent variable
Model I ARIMAX(1,0,1) 1.066527
Model II ARIMAX(2,0,2) 0.010287
Model III ARIMAX(2,0,2) 0.094187
Model IV ARIMAX(2,0,2) 0.236392
Model V ARIMAX(0,1,1) 0.715401
Model VI ARIMAX(2,0,2) 0.059651
Model VII ARIMAX(2,1,1) 0.372343
Model VIII ARIMAX(2,0,2) 0.833744
Model IX ARIMAX(2,1,2) 4.755135

Model IV

Pt  Pt1 þ d t At1
¼ 0:47 þ 2:48 þ et
Pt1 Pt1 (10.39)
et ¼ 0:21et1 þ 0:36et2 þ et  0:68et1  1:42et2
et  N ð0; s2 Þ:
Model V

Pt  Pt1 þ d t At  At1
ð 1  LÞ ¼ 0:055  0:5 þ et
Pt1 Pt1 (10.40)
et ¼ et  1:52et1
et  N ð0; s2 Þ:

Model VI

Pt  Pt1 þ d t At  At1 At
¼ 0:33  2:6 þ 2:4 þ et
Pt1 Pt1 Pt1 (10.41)
et ¼ 0:4et1 þ 0:3et2 þ et 0:37et1  1:9et2
et  N ð0; s2 Þ:
Model VII
Pt  Pt1 þ d t At
ð1  LÞ ¼ 0:2  0:76 þ et
Pt2 Pt1 (10.42)
et ¼ 0:1et1 þ 0:3et2 þ et  1:7et1
et  N ð0; s2 Þ:

Model VIII
Pt At
¼ 1:3  0:22 þ et
Pt1 Pt1 (10.43)
et ¼ 0:31et1 þ 0:46et2 þ et þ 0:94et1  1:1et2
et  N ð0; s2 Þ:
296 A. Maggina

Model IX

ð1  LÞðPt  Pt1 Þ ¼ 0:16  5:54ðAt  At1 Þ þ et

et ¼ 0:25et1 þ 0:13et2 þ et  1:2et1  0:64et2 (10.44)
et  N ð0; s2 Þ:

See Table 10.2

10.6 Conclusions and Suggestions for Further Future Research

This study denotes whether some standard models of market-based accounting

research can explain stock returns in a different country than the USA and UK. In
other words, in the Greek stock market, there is a memory of earnings in stock returns.
Models I (price on earnings), II (returns on change in earnings divided by beginning-of-
period price and prior period earnings divided by the price at the beginning of the return
period), V (returns on change in earnings over opening market value), VII (returns
deflated by lag of 2 years on earnings over opening market value), and IX (differenced-
price model) have statistically significant coefficients of explanatory variables. In
addition, model II with MSE (minimum squared error) loss function in ARIMAX
(2,0,2) is prevalent. ARIMAX (2,0,2) is a representation of the ARIMAX (p,d,q) with
two lags in autoregressive, after being differenced 0 times, with two lags in residuals.
Unlike the US market, in the Athens Stock Exchange, there is a far lower ratio of
professional analysts per registered company, and management earnings forecasts are
only recently and very rarely made publicly available which as fact makes earnings
change and earnings level as determinant factor in stock returns. Further analysis in
earnings components and revenues from sales may explain more satisfactorily the
returns and earnings association. This is in the due course. The issuance of a law on
1985 (Presidential Decree 360/1985) for the publication of semiannual reports and
the issuance of a law (2533/1997) for the publication of quarterly reports seem to
have no effect on the main accounting moments employed in this study. Even the
institutional changes such as liberalization of the auditing profession, corporate tax
cuts, and the change in currency seem to have no effect on the models under
investigation. The different business environment that is going to be formed with
forecasted financial statements and the IAS(IFRS) may make the model selection
a different task. Besides application of generalized autoregressive conditional
heteroscedasticity (GARCH) models is suggested for further future research.

References
Ball, R., & Brown, P. (1968). An empirical evaluation of accounting income numbers. Journal of
Accounting Research, 6, 159–178.
Beaver, W. (1970). The time series behavior of earnings variables, empirical research in
accounting. Selected studies. Supplement to Journal of Accounting Research, 62–99.
10 Market-Based Accounting Research (MBAR) Models: A Test of ARIMAX Modeling 297

Beaver, W., Lambert, R., & Morse, D. (1980). The information content of security prices. Journal
of Accounting and Economics, 2, 3–28.
Board, J.L.G., & Walker, M. (1990). Intertemporal and cross sectional variation in the association
between unexpected accounting rates of return and abnormal returns. Journal of Accounting
Research, 182–192.
Box, G. E. P., & Jenkins, G. M. (1976). Time series analysis: Forecasting and control, revised
edition. Oakland: Holden-Day.
Brown, P., Foster, G., & Noreen, E. (1985). Security analyst multiyear earnings forecasts and the
capital market. Monograph, American Accounting Association, 11, 1–45.
Chen, C., Chen, S., & Su, X. (2001). Is accounting information value-relevant in the emerging
Chinese stock market? Journal of International Accounting, Auditing and Taxation, 10, 1–22.
Chen, G., Firth, M., & Ning Gao, D. (2011). The information content of earnings components:
Evidence from the Chinese stock market. European Accounting Review, 20, 669–692.
Cho, J. Y., & Jung, K. (1991). Earnings response coefficients: A synthesis of theory and empirical
evidence. Journal of Accounting Literature, 10, 85–116.
Christie, A. (1987). On cross-sectional analysis in accounting research. Journal of Accounting and
Economics, 9, 231–258.
Collins, D., & Kothari, S. (1989). An analysis of international and cross-sectional determinants of
earnings response coefficients. Journal of Accounting and Economics, 11, 143–181.
Collins, D., Kothari, S., & Rayburn, J. (1987). Firm size and the information content of prices with
respect to earnings. Journal of Accounting and Economics, 9, 111–138.
Cornell, B., & Landsman, W. (1989). Security price response to quarterly earnings announcements
and analyst forecast revisions. The Accounting Review, 64, 680–692.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series
with a unit root. Journal of the American Statistical Association, 74, 427–431.
Easton, P., & Harris, T. (1991). Earnings as an explanatory variable for returns. Journal of
Accounting Research, 29, 19–36.
Easton, P., & Zmijewski, M. (1989). Cross-sectional variation in the stock market response to
accounting earnings announcements. Journal of Accounting and Economics, 29, 19–36.
Fama, E. (1970). Efficient capital markets: A review of theory and empirical work. Journal of
Finance, 25(2), 383–417.
Fama, E., Fisher, L., Jensen M., & Roll, R. (1969). The adjustment of stock prices to new
information. International Economic Review, 10(1), 1–21.
Freeman, R. (1987). The association between accounting earnings and stock returns for large and
small firms. Journal of Accounting and Economics, 9, 195–228.
Greene, W.H. (2007). Econometric analysis. Prentice Hall, New Jersey.
Hirst, E., Koonce, L., & Venkataraman, S. (2008). Management earnings forecasts: A review and
framework. Accounting Horizons, 22, 315–338.
Holthausen, R. W., & Verrecchia, R. E. (1988). The effect of informedness and consensus on price
and volume behavior. The Accounting Review, 63, 191–208.
Kahya, E., Maggina, A., & Theodossiou, P. (1993). The information content of accounting earnings in
Greece: An empirical investigation. Journal of Multinational Financial Management, 3, 143–156.
Kormendi, R., & Lipe, R. (1987). Earnings innovations, earnings persistence and stock returns.
Journal of Business, 60, 323–345.
Kothari, S. P., & Zimmerman, J. L. (1995). Price and return models. Journal of Accounting and
Economics, 20, 155–192.
Landsman, W. R., & Magliolo, J. (1988). Cross-sectional capital market research and model
specification. The Accounting Review, 64, 586–604.
Lev, B., & Ohlson, J. A. (1982). Market based empirical research in accounting:
A review, interpretations, and extensions. Supplement to Journal of Accounting Research, 20,
249–322.
Lipe, R. (1990). The relation between stock returns and accounting earnings given alternative
information. The Accounting Review, 65, 49–71.
298 A. Maggina

Lundholm, R., & Myers, L. (2002). Bringing the future forward: The effect of disclosure on the
returns-earnings relation. Journal of Accounting Research, 40, 809–839.
Mandelbrot, B. (1966). Forecasts of future prices, unbiased markets and martingale models.
Journal of Business, 39, 242–255.
Ou, J., & Penman, S. (1989). Financial statement analysis and the prediction of stock returns.
Journal of Accounting and Economics, 11, 295–329.
Rosenberg, B., & Marathe, V. (1975). The prediction of investment risk: Systematic and residual
risk. In Proceedings of the seminar on the analysis of security prices (pp. 185–226). Chicago:
Center for Research in Security Prices, University of Chicago.
Samuelson, P. (1965). Proof that properly anticipated prices fluctuate randomly. Industrial Man-
agement Review, 31, 41–49.
Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics, 6, 461–464.
Strong, N., & Walker, M. (1993). The explanatory power of earnings for stock returns. The
Accounting Review, 68, 385–399.
Trueman, B. (1986). Why do managers voluntarily release earnings forecasts. Journal of Account-
ing and Economics, 8, 53–71.
Wei, W.W. (2006). Time series analysis: Univariate and multivariate methods. Pearson Education.
White, H. (1980). A heteroskedasticity-consistent covariance matrix and a direct test for
heteroskedasticity. Econometrica, 48, 817–838.
An Assessment of Copula Functions
Approach in Conjunction with 11
Factor Model in Portfolio Credit Risk
Management

Lie-Jane Kao, Po-Cheng Wu, and Cheng-Few Lee

Contents
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
11.2 Dependence Structure in Static Factor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
11.3 Dependence Structure in Dynamic Factor Model Using Intensity Function . . . . . . . . . . 303
11.4 Contagious Model: Mutually Exciting Intensity Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
11.5 Copula Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
11.6 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
11.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Appendix 1: Point Process and Its Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Appendix 2: Copula Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315

Abstract
In credit risk modeling, factor models, either static or dynamic, are often used to
account for correlated defaults among a set of financial assets. Within the realm of
factor models, default dependence is due to a set of common systemic factors.
Conditional on these common factors, defaults are independent. The benefit of
a factor model is straightforward coupling with a copula function to give an
analytic formulation of the joint distribution of default times. However, factor
models fail to account for the contagion mechanism of defaults in which a firm’s

L.-J. Kao (*) • P.-C. Wu

Department of Finance and Banking, Kainan University, Taoyuan, ROC, Taiwan
e-mail: ljkao@mail.knu.edu.tw; pcwu@mail.knu.edu.tw
C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 299
DOI 10.1007/978-1-4614-7750-1_11,
# Springer Science+Business Media New York 2015
300 L.-J. Kao et al.

default risk increases due to their commercial or financial counterparties’ defaults.

This study considers a mixture of the dynamic factor model of Duffee (Review of
Financial Studies 12, 197–226, 1999) and a contagious effect in the specification
of a Hawkes process, a class of counting processes which allows intensities to
depend on the timing of previous events (Hawkes. Biometrika 58(1), 83–90,
1971). Using the mixture factor-contagious-effect model, Monte Carlo simulation
is performed to generate default times of two hypothesized firms.
The goodness-of-fit of the joint distributions based on the most often used
copula functions in literature including the normal, t-, Clayton, Frank, and
Gumbel copula, respectively, is assessed against the simulated default times. It
is demonstrated that as the contagious effect increases, the goodness-of-fit of the
joint distribution functions based on copula functions decreases, which high-
lights the deficiency of the copula function approach.

Keywords
Static factor model • Dynamic factor model • Correlated defaults • Contagious
effect • Hawkes process-Monte • Carlo simulation • Normal copula • t-copula •
Clayton copula • Frank copula • Gumbel copula

11.1 Introduction

An understanding of correlated defaults of the underlying assets is fundamental to

portfolio management and the pricing of credit derivatives such as a CDS contract,
a CDO contract, or a basket default swap contract (Hull and White 2001; Zhou
2001; Das et al. 2006). There exist, however, different ways of introducing corre-
lations when modeling assets’ defaults. In Schönbucher (2003), a good modeling
framework for correlated defaults should include (1) being able to produce default
correlations of a realistic magnitude, (2) being able to model the timing of defaults,
(3) being capable of reproducing default clustering periods, and (4) easily calibrated
and implemented by keeping the number of parameters under control without
growing dramatically with the number of assets.
In responding to these properties, two types of factor models coupled with a set
of small number of common systemic factors have been followed in literature to
account for default correlations. The first type of factor model is static in the sense
that the modeling of default correlations among a set of I assets is based on the
creditworthiness indices X1, . . ., XI at a specific time point T, which is analogous to
Merton’s structural model (1974) in which an asset defaults when its creditworthi-
ness index falls below some threshold at a specific time point. The static factor
model has been widely used for the computation of joint default and loss probability
distribution in a portfolio (Vasicek 1997; Koyuoglu and Hickman 1998; Finger
2000; Schönbucher 2003). Portfolio credit risk models fit within this framework
include CreditMetrics (1997) of JP Morgan, Credit Risk+ of Credit Suisse Financial
Products (1997), and KMV’s portfolio manager (Kealhofer 1995).
11 An Assessment of Copula Functions Approach 301

In the static factor model, default dependence among a set of I assets is

introduced by the correlations between the creditworthiness X1, . . ., XI and a set
of common systemic economic factors. Conditional on these common systemic
factors, the creditworthiness X1, . . ., XI are independent. Depending on the distri-
butions imposing on these common systemic factors, a specific type of copula
function can be used to obtain an analytical expression for the joint distribution
of the default times for the I assets (Li 1999, 2000; Schönbucher 2003; Gregory and
Laurent 2005).
The second type of factor models uses a more consistent framework to model
correlated defaults via assets’ default intensities that describe the evolution of
instantaneous default probabilities dynamically over time. In parallel to the static
factor model, the dynamic factor model generates default dependency among assets
through a set of common factors. Conditioned to the realization of these common
factors, the assets’ defaults are independent. However, being different from the
static factor model in which common factors affect assets’ creditworthiness at
a specific time horizon, the common factors in dynamic factor models affect assets’
default intensities dynamically over time (Duffee 1999; Yu 2005; Driessen 2005;
Elizalde 2005).
The benefit of a factor model, either static or dynamic, is its straightforward
conjunction with a copula function to give an analytic formulation of the joint
distribution of default times (Li 1999, 2000; Schönbucher and Schubert 2001;
Schönbucher 2003; Gregory and Laurent 2005). In this way, model parameters
can be readily inferred from the common systemic factors and individual assets’
default times. The main drawback of a factor model, however, is that the likelihood
of an asset’s default does not change due to the defaults of any other assets, i.e., the
default of one asset does not trigger the defaults of other related assets. While this
might be adequate for production firms, it may be inadequate for studying the
default risk of a financial institution with large positions in a few assets whose
default can trigger the failure of other financial institutions (Lucas 1995; Jarrow and
Yu 2001; Nagpal and Bahar 2001; Das et al. 2006). Under such circumstance,
correlated defaults are due to counterparties’ default risk. For this reason, the
contagious default models in which the default intensity of an asset depends on
the status (default/not default) of the other assets are proposed (Davis and Lo 2001;
Jarrow and Yu 2001). These contagious default models allow extra-default depen-
dence to be introduced compared to the factor models.
This study uses a mixture of a dynamic factor model by Duffee (1999) with
a contagious effect as a benchmark model to generate simulated default times of
two assets. Various copula functions that are often used in credit risk modeling,
including the normal copula, t-copula, Clayton copula, Frank copula, and Gumbel
copula, are employed to calibrate the simulated default times. As the contagious
effect increases, the simulated default times show that the goodness-of-fit of the
aforementioned copula functions decreases.
The paper is organized as follows. Section 11.2 introduces the static factor
model and its link to the copula function approach. Section 11.3 introduces the
302 L.-J. Kao et al.

dynamic factor model in a point process. Section 11.4 introduces the contagious
effect model based on mutually exciting point processes. The theoretical founda-
tions of the copula function approach are given in Sect. 11.5. A simulation study is
given in Sect. 11.6. Section 11.7 concludes.

11.2 Dependence Structure in Static Factor Model

In a static factor model, the creditworthiness indices X1, . . ., XI of I assets at

times 0 < t1, . . ., tI
T are explained by 1 or a set of common systemic risk factors
Y1, . . ., YK in the form

X
K
Xi ¼ rij Y j þ riðKþ1Þ ei
j¼1

where rij determines the relative importance of asset i to the common factor Yj and
ri(K+1) the relative weight of idiosyncratic factor ei for asset i. To the extent that the
creditworthiness indices are related to the common systemic factors Y1, . . ., YK, the
likelihood of joint default events across assets varies accordingly. Here the common
factors Y1, . . ., YK and the idiosyncratic factors e1,. . ., eI are independent. In the
following, for simplicity, we will consider a single factor Y and the ith asset’s
creditworthiness index:
qffiffiffiffiffiffiffiffiffiffiffiffiffi
Xi ¼ ri Y þ 1  r2i ei : (11.1)

Equation 11.1 is the one-factor model adopted in CreditMetrics (Gupton

et al. 1997). Since the default environment is completely determined by creditwor-
thiness indices X1, . . ., XI at times t1, . . ., tI, respectively, the factor model is static in
that the dynamics of the firms’ credit quality that evolves during the time horizon
[0, T] is ignored.
Suppose default of the ith asset occurs if the creditworthiness index
Xi falls below a certain threshold ui at time ti and the marginal distribution
functions

F1 ðt1 Þ ¼ Prðt1
t1 Þ, . . . , FI ðtI Þ ¼ PrðtI
tI Þ

of the default times t1, . . ., tI are given. The threshold ui and the joint distribution of
the default times t1, . . . , tI satisfy Fi(ti) ¼ Pr(Xi
ui) and F(t1, . . . , tI) ¼ Pr(X1
u1,
. . ., XI
uI), respectively. By noting that conditional on the common factor Y, the
creditworthiness indices X1, . . . , XI are independent and therefore
ð I 
PrðX1
U 1 , . . . , XI
U I Þ ¼ P PðXi
ui jyÞ f ðyÞdy:
i¼1
11 An Assessment of Copula Functions Approach 303

If assuming the latent common factors Y and idiosyncratic factors e1, . . ., eI are
standard normally distributed, then the creditworthiness indices Xi are also standard
normally distributed. Therefore, the conditional probability
!
F1 ðFi ðti ÞÞ  ri y
PðXi
ui jyÞ ¼ F pffiffiffiffiffiffiffiffiffiffiffiffiffi
1  r2i

And the joint distribution F(t1, . . ., tI) of default times t1, . . ., tI is linked to
a one-factor normal copula function in the form

ð" !#
I F1 ðFi ðti ÞÞ  ri y
Fðt1 ; . . . ; tI Þ ¼ PF pffiffiffiffiffiffiffiffiffiffiffiffiffi fðyÞdy (11.2)
i¼1 1  r2i

where f is the standard normal density function. Alternatively, one can assume that
the latent common factor Y is Gamma distributed with parameter 1/y and let

 
PðXi
ui jyÞ ¼ exp y’1 ðFi ðti ÞÞ

where ’1(s) ¼ sy  1 and ’ is the generator. Then the joint distribution F(t1, . . ., tI)
is linked to a one-factor Clayton copula function in the form
ð 
I 1

Fðt1 ; . . . ; tI Þ ¼ P exp y’ ðFi ðti ÞÞ fðyÞdy: (11.3)
i¼1

The Clayton copula function and its generator ’ are given in Definition 2.7 of
Appendix 2.

11.3 Dependence Structure in Dynamic Factor Model Using

Intensity Function

In the dynamic factor model, as in the reduced-form model put forward

by Jarrow and Turnbull (1995), Lando (1998), and Duffie and Singleton
(1999a), default is treated as an unpredictable jump of a firm’s value and the
default time t is treated as the time of the unpredictable jump. Often the
unpredictable jump or default is considered as triggered by an exogenous
event that occurs with an instantaneous likelihood specified by the intensity
function of the first jump in a point process. The formal definition of a point
process and properties of its intensity function are given in Definitions 1.1–1.2
of Appendix 1.
In the dynamic factor model, the intensity function of a default depends on
certain exogenously determined stochastic common systemic factors Xt which
304 L.-J. Kao et al.

induce correlated defaults among assets (Duffee 1999; Driessen 2005; Elizalde
2005; Yu 2005). In contrast to the static factor model, the evolution of these
common factors over time is modeled, and conditional on the evolution of these
common factors Xt, defaults are independent. The formulation of the default
intensity as such is referred as a doubly Poisson point process (Lando 1998). In
Definition 1.3, the properties of a point process and a (doubly) Poisson process are
given. An example of a dynamic factor model is given in Duffee (1999), in which
the default intensity of asset i is

   
li , t ¼ a0, i þ a1, i X1, t  X1 þ a2, i X2, t  X2 þ li, t c

where a0,i, a1,i, and a2,i are constants and X1,t and X2,t are two latent factors
interpreted as the slope and level of the default-free yield curve, i.e., the risk-free
interest rate

r t ¼ ar þ X1, t þ X2, t :

The asset’s specific intensity li;t , independent across assets, is assumed to obey
a mean-reversion process

  qffiffiffiffiffiffi
dli,t ¼ ki yi  li,t dt þ sli li,t dW i,t (11.4)

where W1,t, . . ., WI,t are independent Brownian motions. To introduce more default
correlation, Duffie and Singleton (1999b) incorporate joint as well as idiosyncratic
jumps in the default intensity li,t.
In Das et al. (2007), the hypothesis whether default events can be modeled as
a doubly Poisson point process that solely depends on “exogenous” factors is tested.
Based on a time series of US corporate defaults, they strongly rejected the hypoth-
esis that defaults can be modeled as a doubly Poisson point process. Instead of
a factor model, in the following, the contagious model in which default status of
other firms will affect the default intensity of the underlying asset will be
considered.

11.4 Contagious Model: Mutually Exciting Intensity Function

Examples of contagious models include the infectious default model by Davis and
Lo (2001) in which the intensity function follows a piecewise deterministic Markov
process and the propensity model by Jarrow and Yu (2001). In the propensity
model, firms are divided into K primary and I-K-1 secondary firms: the default
intensities l1,t, . . ., lK,t of primary firms are determined by some exogenously
11 An Assessment of Copula Functions Approach 305

common factors, while those of the secondary firms depend on the status (default or
not) of the primary firms but not the other way around (asymmetric dependence).
Specifically, the default intensity of the ith secondary firm is given by

X
K
li , t ¼ l i , t þ aji, t 1ftj
tg (11.5)
j¼1

where K + 1
i
I. Here li,t represents the part of secondary firm i’s hazard rate
independent of the default status of other firms.
Recently, Hawkes process, a class of counting processes which allow intensities
to depend on the timing of previous events (Hawkes 1971), had been adopted to
model the aggregate default intensity of a portfolio (Azizpour and Giesecke 2008;
Errais et al. 2010). In Errais et al. (2010), the cumulated default intensity of n firms
is specified by

X
n
lðtÞ ¼ l0 þ aebðttj Þ 1ftj
tg : (11.6)
j¼1

In Lando and Nielsen (2010), a total of 2,557 firms with an average of 1,142 and
a minimum of 1,007 firms at any time throughout the sample period from January
1982 to December 2005 are used to calibrate the goodness-of-fit of an extended
dynamic factor model that includes exogenously determined common factors as
well as a contagion effect in terms of a Hawkes process. During the sampling
period, 370 firms default and a contagious effect is concluded.
In this study, we simulate the default times of two firms using a mixture model
based on a dynamic factor model of Duffee (1999) together with a contagious effect
in the specification of a Hawkes process. Various commonly used copula functions
are calibrated against simulated default times to demonstrate the copula functions in
the modeling of credit default risk.

11.5 Copula Functions

Copulas, introduced by Sklar (1959), have been extensively applied in areas such as
actuarial science using survival data. It was adopted by Li (2000) and Gregory and
Laurent (2005) to the application in modeling the joint distribution of the default
times t1, . . ., tI of a set of I firms. According to Sklar (1959), for any continuous
joint distribution F(t1, . . ., tI), there exists a uniquely determined I-dimensional
copula CF(u1, . . ., uI), where ui ¼ F(ti), 1
i
I, such that for all (t1, . . ., tI) in RI,

Fðt1 ; . . . ; tI Þ ¼ CF ðF1 ðt1 Þ, . . . , FI ðtI ÞÞ: (11.7)

In Definition 2.1, the formal definition of a copula is given. The theorem by Sklar
(1959) is given in Appendix 2.
306 L.-J. Kao et al.

According to Sklar (1959), as the existence of the copula function CF is

guaranteed, the joint distribution F of default times t1, . . ., tI can be obtained via
Eq. 11.7 given the marginal distributions F1, . . ., FI of individual default times.
However, the analytic formulation of the specific copula function CF is usually
unknown and even intractable. For this reason, some copula functions are chosen in
an ad hoc way in credit risk modeling. The most often used copulas are the normal
copulas (Li 2000; Frey et al. 2001; Gregory and Laurent 2005). However, normal
copula presents no tail dependency and has been criticized for not assigning enough
probability for the occurrence of extreme events.
To account for tail dependency, the double-t-copula or a member of
the Archimedean-type copulas can be used. The t-copula is radically
symmetric in that its lower and upper tail dependence is the same. In case
lower tail dependence is desired, the Clayton copula should be used, while
a Gumbel copula should be used for upper tail dependence. However, the Gumbel
copula does not allow for negative dependence. To compare these copulas,
three indices to measure the dependence structure between two random variables
introduced by the copula function are given in Definitions 2.3–2.4, respectively.
The three dependence measures include the global dependent measure,
i.e., Kendall’s tau, and two tail dependent measures, i.e., the upper/lower tail
dependence coefficients.
Instead of tail dependence, this study focused on the dependent structure of the
firms’ default times due to contagious effects. Schönbucher and Schubert (2001)
study the dynamics of default intensities and show that a Clayton copula, a member
of the Archimedean copula family, is related to the contagious models of Davis and
Lo (2001) and Jarrow and Yu (2001). In the following simulation study based on
a mixture of the dynamic factor model of Duffee (1999) and a contagious effect
specified by a Hawkes process, the aforementioned copula functions are calibrated
against the simulated default times generated by the mixture model to test the
goodness-of-fit of the copulas when contagious effect is present.

11.6 Simulation Study

A simulation based on a mixture of the dynamic factor model of Duffee (1999)

incorporated with contagious effect specified by a Hawkes process is performed here.
In Duffee (1999), a firm’s default intensity is determined by two factors X1,t and X2,t,
where X1,t and X2,t are the two latent components of the risk-free interest rate rt ¼ ar +
X1,t + X2,t. The latent components Xj,t, j ¼ 1, 2, obeys the mean-reversion process:
  pffiffiffiffiffiffiffi
dXj, t ¼ j mj  Xj, t dt þ sXj Xj, t dW Xj, t (11.8)

where WX1;t and WX2;t are two independent Wiener processes. To account for coun-
terparty default risk, however, a modification is made on the default intensity in
Eq. 11.8 by including the default status of the other firm(s). For the two-firm case,
the default intensities are, respectively,
11 An Assessment of Copula Functions Approach 307

Table 11.1 Parameter specification for mixture model

Panel a:
i ki yi sjl
1 0.023 0.0036 0.051
2 0.600 0.1407 0.104
Panel b:
j ar jmj sjX
1 10.474 1.003 0.0134
2 10.032 0.060 0.0449
Panel c:
i a0,i a1,i a2,i
1 0.0132 –0.142 0.001
2 0.0196 0.001 0.062
This table reports the parameter values of the proposed mixture model. In Panel a, the parameter
values in the mean-reversion process in Eq. 11.4 for the two firms (i ¼ 1,2), respectively, are given.
In Panel b, parameter values of the latent components Xj,t (j ¼ 1,2) in Eq. 11.8 are given. In Panel c,
parameter values of default intensities li(i ¼ 1,2) in Eqs. 11.9 and 11.10, respectively, are given

   
l1, t ¼ a0,1 þ a1,1 X1, t  X1 þ a2,1 X2,t  X2
(11.9)
þ l1,t þ aebðtt2 Þ 1ft2
tg
   
l2, t ¼ a0, 2 þ a1, 2 X1, t  X1 þ a2, 2 X2, t  X2
(11.10)
þ l2, t þ aebðtt1 Þ 1ft1
tg

where t1 and t2 are the default times of the 1st and 2nd firm, respectively. The
firm’s specific default intensities l1;t and l2;t are assumed to follow the mean-
reversion processes in Eq. 11.4.
The values of the parameters ki, yi, j, mj, a0,i, a1,i, a2,i, sXj , 1
i
2, 1
j
2 are
from Duffee (1999), in which month-end prices of noncallable corporate bonds
from January 1985 to December 1995 across 161 firms, with majority of investment
grade bonds, are used to calibrate default intensity process. In Table 11.1, the
parameter values used are given. Two cases with different contagious effects are
considered in the study. In the first case, the parameter a that specifies the conta-
gious effect is set to 0.1, while the a value is 0.25 in the second case. For each case,
10,000 pairs of simulated default times for the two firms are generated.
Five different copula functions are considered to fit the empirical joint distribu-
tion of the default times t1 and t2. They are the normal, t-, Clayton, Frank, and
Gumbel copula, respectively. For the normal and t-copula, the parameter r is set to
the empirical correlation coefficient of the simulated default times. Minimum mean
squares error is used to obtain the estimated parameter values of the degree of
freedom v in t-copula and y in Clayton, Frank, and Gumbel copula, respectively.
In Table 11.2, the estimated parameters for the five copula functions and the
corresponding sum squares of errors (SSE) are given. As can be seen in Table 11.2,
308 L.-J. Kao et al.

Table 11.2 Parameter estimation of copula functions

Case I: a ¼ 0.1
Copula type SSE Parameter estimation
Normal 0.2394 r ¼ 0.2137
t-copula 0.1650 r ¼ 0.2137, d.f. ¼ 10
Clayton 0.6043 y ¼ 0.51
Frank 0.1695 y ¼ 1.70
Gumbel 0.1591 y ¼ 1.15
Case II: a ¼ 0.25
Copula type SSE Parameter estimation
Normal 0.3867 r ¼ 0.4687
t-copula 0.2213 r ¼ 0.4687, d.f. ¼ 3
Clayton 1.4439 y ¼ 1.51
Frank 0.4525 y ¼ 3.75
Gumbel 0.1431 y ¼ 1.45
This table reports the estimated parameters of the normal, t-, Clayton, Frank, and Gumbel copulas.
For the parameter r, empirical correlation of the simulated default times is used. SSE is used to
obtain the estimates of the degree of freedom (d.f.) for the t-copula and the parameter y for the
Clayton, Frank, and Gumbel copula, respectively

the normal copula is outperformed by the t-, Frank, and Gumbel copulas when
a ¼ 0.10, while outperformed by the t- and Gumbel copulas only when a ¼ 0.25. It
can also be seen that as the contagious effect parameter a increases from 0.10 to
0.25, except the Gumbel copula, the SSEs also increase significantly for the normal,
t-, Clayton, Frank, and Gumbel copulas. In both cases, i.e., a ¼ 0.10 and a ¼ 0.25,
the Gumbel copula performs the best while the Clayton copula performs the worst
among the five copula approximations.
In Figs. 11.1 and 11.2, the contour plots of the differences between the empirical
and copula-based joint distribution functions when the contagious effect parameter
a ¼ 0.10 and a ¼ 0.25, respectively, are given. For the normal copula, as can be seen in
Fig. 11.1 (panel a), larger deviations from the empirical joint distribution occur when
the default times t1 and t2 are in the range from 5 to 10 years when a ¼ 0.10. When
a ¼ 0.25, as illustrated in Fig. 11.2 (panel a), larger deviations occur in the range from
5 to 15 years. For the t-copula, larger deviations from the empirical joint distribution
occur when the default times t1 and t2 are in the range from 5 to 10 years when
a ¼ 0.10. However, in panel b of Fig. 11.2 where a ¼ 0.25, in addition to the range
between 5 and 10 years, the t-copula fails to approximate the empirical joint distri-
bution well for small default times t1 and t2. This implies that as the contagious effect
increases, the t-copula does not explain the lower tail dependence well. In panel d of
Figs. 11.1 and 11.2, when using the Gumbel copula, larger deviations occur when the
default times t1 and t2 are in the range from 2 to 7 years. This indicates that the
Gumbel copula does not approximate well for smaller default times.
Taken together, the goodness-of-fit of the normal, t-, Clayton, Frank, and Gumbel
copula, respectively, decreases as the contagious effect parameter a increases.
This phenomenon is more apparent for the Frank copula. The contour plots of the
11 An Assessment of Copula Functions Approach 309

20 20
a: normal copula b: t-copula
15 15

τ2
τ2

10 10

5 5

0 0
0 5 10 15 20 0 5 10 15 20
τ1 τ1
20 20
c: Clayton copula d: Gumbel copula
15 15

τ2
10
τ2

10
5 5

0 0
0 5 10 15 20 0 5 10 15 20
τ1 τ1

Fig. 11.1 Contours of Differences between Empirical Joint Distribution and Copula Approxi-
mation. Note Panels a-d show the contours of differences with absolute values exceeding 0.01 for
normal, t-Clayton, and Gumbel copula, respectively. The contagious effect parameter a is 0.1

20 20
a: normal copula b: t-copula
15 15
τ2

10 10
τ2

5 5

0 0
0 5 10 15 20 0 5 10 15 20
τ1 τ1
20 20
c: Clayton copula d: Gumbel copula
15 15
τ2

10
τ2

10
5 5
0 0
0 5 10 15 20 0 5 10 15 20
τ1 τ1

Fig. 11.2 Contours of Differences between Empirical Joint Distribution and its Copula Approx-
imation. Note Panels a-d show the contours of differences with absolute values exceeding 0.01 for
normal, t-Clayton, and Gumbel copula, respectively. The contagious effect parameter a is 0.25

differences between the empirical and Frank copula-based joint distribution functions
when the contagious effect parameter a ¼ 0.10 and a ¼ 0.25, respectively, are given
in Panel a–b of Fig. 11.3. As can be seen in Fig. 11.3, when a ¼ 0.10, the Frank
copula fails to approximate the empirical distribution well only for small default
310 L.-J. Kao et al.

20
a: contagious effect = 0.1
15
τ2

10

0
0 2 4 6 8 10 12 14 16 18 20
τ1
20
b: contagious effect = 0.25
15

10
τ2

0
0 2 4 6 8 10 12 14 16 18 20
τ1

Fig. 11.3 Contours of Differences between Empirical Joint Distribution and its Frank Copula Approx-
imation. Note Panels a–d show the contours of differences with absolute values exceeding 0.01 for
normal, t-Clayton, and Gumbel copula, respectively. The contagious effect parameter a is 0.1 and 0.25

times t1 and t2. As the contagious effect parameter a ¼ 0.25, however, the Frank
copula fails to approximate the empirical distribution well not only for small default
times t1 and t2 but also larger t1 and t2. This can also be seen in Table 11.2, where
the SSE increases from 0.1695 to 0.4525 as a increases from 0.10 to 0.25.

11.7 Conclusion

As a firm’s default determines the economic opportunities available, it is very likely

that the status of firms’ defaults will affect one another. That is, a firm’s default
might have contagious effect on other firms’ defaults. As a result, one often
observes firms’ default simultaneously. This study demonstrates that the copula
functions from literature do not necessarily, and most unlikely, match the joint
distribution F of the correlated default times t1, . . ., tI of a set of I firms as the
contagious effect is increasing.
Five of the most often used copula functions in literature, namely, the normal,
t-, Clayton, Frank, and Gumbel copulas, are under study. Among the five copula
functions, the Clayton copula performs the worst, whereas the Gumbel copula
performs the best. Except the Gumbel copula, the goodness-of-fit of the other
four copulas decreases as the contagious effect increases. This suggests further
advanced statistical tool for the modeling of contagious defaults in a more consis-
tent and accurate way is in need.
11 An Assessment of Copula Functions Approach 311

Appendix 1: Point Process and Its Intensity

Let (O, H, P) be a probability space, where P is a physical measure. A formal

definition of a point process is given below.
Definition 1.1 A point process N with stopping times t1, t2, . . . . 2[0,T] is a counting
measure on the probability space (O, H, P) in that for any Borel subset E[0,T], the
counting measure N(E) represents the number of time points t1, t2, . . . , in E.
Let {Ht: t2[0, T]} be a filtration on (O, P) so that Ht contains the accumulated
information generated by a point process N till time t. The notion of the compen-
sator of the point process N is given below.
Definition 1.2 Define the compensator A of the point process N as the unique
random measure on (O, P) such that:
i. A is Ht-predictable.

ii. For every nonnegative Ht-predictable process H,

0T 1 0T 1
ð ð
E@ HdN A ¼ E@ HdAA:
0 0

The characterization of the compensator A is essential to the statistical infer-

ence of the point process N in that the compensated process M(t) ¼ N(t)  A(t), t 2
[0,T] is a martingale under P. In this case a positive, Ht-predictable process l(t)
exists so that

ðt
AðtÞ ¼ lðsÞds:
0

Then l(t) is called the intensity process of the point process N. The Ht-predictable
intensity has the x dt, the probability one of the stopping time tk that occurs during
(t-dt,t] is

P½N ðt  dt, t ¼ 1jHt  ¼ lðtÞdt þ oðdtÞ:

Definition 1.3 If the compensator process A(t) is continuous and deterministic, then
the point process N(t) is a Poisson process. A doubly stochastic Poisson process
M(t) is a point process in which the intensity process l(t) is Ft-predictable, where Ft
contains the accumulated information generated by the point process M till time
t and the entire trajectory {l(s): s > 0}. Conditioning on one realization of the
intensity process {l(s): s > 0}, the point process M(t) is an inhomogeneous Poisson
process.
312 L.-J. Kao et al.

Appendix 2: Copula Functions

Basically, a copula function C links the joint distribution F(y1, . . ., yI) of I multivariate
random variables Y1, . . ., YI to their univariate marginal distributions F1(y1), . . .,
FI(yI). A formal definition of a copula function is given as follows.
Definition 2.1 A function C: [0, 1]N![0, 1] is a copula if there are uniform random
variables U1, . . ., UN taking values in [0, 1] such that C is their joint distribution
function that satisfies:
1. C(1, . . ., 1, ui, 1, . . ., 1) ¼ ui for all i ¼ 1, . . ., N and ui2[0, 1].
2. For all u2[0, 1]N, C(u) ¼ 0 if at least one coordinate ui ¼ 0, i ¼ 1, . . ., N.
3. For all u2[0, 1]N, v2[0, 1]N with ui
vi, i ¼ 1, . . ., N, the C volume of the
hypercube
i1 þ...þiN
X 2 X2
... ð1Þ Cðwi1 ; . . . ; wiN Þ  0
i1 ¼1 i¼1

where wik = uk if ik ¼ 1 else wik = vk .

The theoretical groundwork of applying a copula function is based on Sklar’s
theorem (Sklar 1959) in the following:
Theorem 2.2 Let t1, . . ., tI be random variables with marginal distribution func-
tions F1, . . ., FI and joint distribution function F. Then there exists an I-dimensional
copula C such that

Fðt1 ; . . . ; tI Þ ¼ Cðu1 ; . . . ; uI Þ ¼ CðF1 ðt1 Þ, . . . , FI ðtI ÞÞ

for all (t1, . . ., tI) in RI, where ui ¼ F(ti), 1
i
I. Moreover, if each Fi is continuous,
then the copula C is unique.
In order to calibrate a copula, indices that measure the dependence structure
introduced by the choice of the copula function are used. We focus on three
dependence measures that depend only on the copula function, not in the marginal
distributions. Among the three dependence measures, Kendall’s tau is the measure
of global dependence, while upper/lower tail dependence coefficients are two local
measures of dependence.
Definition 2.3 Let F be the joint distribution of two random variables X1 and X2
with marginal F1 and F2, respectively. Kendall’s tau is the probability of con-
0
cordance minus the probability of discordance. Specifically, if (X1, X2) and (X1 ,
0
X2 ) are two realizations of F, then Kendall’s tau is defined as

h 0
 0
 i h 0
 0
 i
P X 1  X 1 X2  X 2 > 0  P X 1  X 1 X2  X2 < 0 (11.11)

If C is the copula of the joint distribution F, i.e., C(u1,u2) ¼ C(F1(x1), F2(x2)),

Kendall’s tau is
11 An Assessment of Copula Functions Approach 313

ð1 ð1
tau ¼ 4 Cðu; vÞdCðu; vÞ  1:
0 0

Definition 2.4 Let F(X1, X2) be the joint distribution of two random variables X1 and
X2, with marginal F1 and F2. The coefficients of upper and lower tail dependence
are defined, respectively, as


lU ¼ lim Pr X1 > F1 1
1 ð uÞ X 2 > F2 ð uÞ
u"1


lL ¼ lim Pr X1 < F1 1
1 ðuÞ X2 < F2 ðuÞ :
u#0

Definition 2.5. (Normal Copula) The I-dimensional normal copula is expressed as

 
Cðu1 ; . . . ; uI Þ ¼ FIS F1 ðu1 Þ, . . . , F1 ðuI Þ ,

where S is a positive-definite correlation matrix, FIS is the distribution function of

an I-dimensional multivariate normal random vector with correlation matrix S, and
F1 is the inverse of the distribution function of a standard normal random variable.
The density function c of C(u1, . . ., uI) is

1 1 0  1 
cðu1 ; . . . ; uI Þ ¼ 1=2 exp z S  III z
jSj 2

where z 0 ¼ (F1(u1), . . ., F1(uI)), and III is the unity matrix. For a normal
copula, the relationship between the linear correlation coefficient r and Kendall’s
tau is

r ¼ sin ð2p  tauÞ: (11.12)

On the other hand, the coefficients of upper and lower tail dependence are both
zero, i.e., a normal copula is tail independent.
Definition 2.6 (t-Copula) The I-dimensional t-copula is expressed as

 
Cðu1 ; . . . ; uI Þ ¼ tIn, S t1 1
n ðu1 Þ, . . . , tn ðuI Þ

where tn1 denotes the inverse of the distribution function of a univariate t-student
random variable with n degrees of freedom and tIn;S denotes the distribution function
of a multivariate t-distribution with n degrees of freedom and positive-definite
dispersion matrix S. Its density function is

 nI1 nþI 0 1  2
nþI

G 2 G 2 1 þ z Sn z
 ðnþ1 Þ
 nþ1I 12 z2i 2

G 2 jSj I
1þ
i¼1 n
314 L.-J. Kao et al.

where z 0 ¼ (tn1(u1), . . ., tn1(uI)). For a t-copula, the relationship between the

linear correlation coefficient r and Kendall’s tau is the same as Eq. 11.12. The
coefficients of upper and lower tail dependence are

ðn þ 1Þð1  rÞ 1=2
lU ¼ lL ¼ 2  2tn :
1þr

Definition 2.7 (Archimedean Copula) The I-dimensional Archimedean copula can

be expressed as
!
X
I
1
Cðu1 ; . . . ; uI Þ ¼ f fðui Þ (11.13)
i¼1

where the generator f is a continuous strictly decreasing function from [0, 1] to

[0, 1] satisfying f(0) ¼ 1 and f(1) ¼ 0. In particular, when the generator

uy  1
fðuÞ ¼ ,y  0
y
then Eq. 11.13 is called a Clayton copula. When the generator
 yu
e 1
fðuÞ ¼ ln y , y 6¼ 0
e 1

Equation 11.13 is called a Frank copula. When the generator f(u) ¼ (lnu)y,
y  1, Eq. 11.13 is a Gumbel copula. In particular, the density function of a Clayton
copula is
!I1=y
X
I I

cðu1 ; ...; uI Þ ¼ 1  I þ uy
i P uiy1 ðði  1Þy þ 1Þ : (11.14)
i¼1
i¼1

For a Clayton copula, Kendall’s tau and tail dependency measures are

ð1
uy  1 y
tau ¼ 1 þ 4 y1
du ¼
yu y2
0
0
f ð2vÞ
lU ¼ 2  2 lim 0 ¼0
v!0 f ðvÞ
0
f ð2vÞ
lL ¼ 2 lim 0 ¼2
v!0 f ðvÞ

where the function ’(u) ¼ f1(v) is the inverse of the generator f.

11 An Assessment of Copula Functions Approach 315

References
Azizpour, S., & Giesecke, K. (2008). Self-exciting corporate defaults: Contagion vs. Frailty.
Working paper, Stanford University.
Credit Suisse Financial Products. (1997). CreditRisk+: A credit risk management framework.
New York.
CreditMetrics. (1997). The benchmark for understanding credit risk. Technical document.
New York: JP Morgan.
Das, S. R., Freed, L., Geng, G., & Kapadia, N. (2006). Correlated default risk. Journal of Fixed
Income, 16(2), 7–32.
Das, S., Duffie, D., Kapadia, N., & Saita, L. (2007). Common failings: How corporate defaults are
correlated. Journal of Finance, 62, 93–117.
Davis, M., & Lo, V. (2001). Infectious defaults. Quantitative Finance, 1, 382–387.
Driessen, J. (2005). Is default event risk priced in corporate bonds? Review of Financial Studies,
18, 165–195.
Duffee, G. R. (1999). Estimating the price of default risk. Review of Financial Studies, 12, 197–226.
Duffie, D., & Singleton, K. J. (1999a). Modeling term structures of defaultable bonds. Review of
Financial Studies Special, 12(4), 687–720.
Duffie, D., Singleton, K. J. (1999b). Simulating correlated defaults. Working paper, Graduate
School of Business, Stanford University.
Elizalde, A. (2005). Do we need to worry about credit risk correlation? Journal of Fixed Income,
15(3), 42–59.
Errais, E., Giesecke, K., & Goldberg, L. R. (2010). Affine point processes and portfolio credit risk.
SIAM Journal on Financial Mathematics, 1, 642–665.
Finger, C. (2000). A comparison of stochastic default rate models. Risk Metrics Journal, 1(2), 49–73.
Frey, R., McNeil, A., & Nyfeler, M. (2001). Copulas and credit models. RISK, 111–114.
Gregory, J., & Laurent, J. P. (2005). Basket default swaps, CDO’s and factor copulas. Journal of
Risk, 7(4), 103–22.
Gupton, G., Finger, C., & Bhatia, M. (1997). Credit metrics technical document. New York:
Morgan Guaranty Trust.
Hawkes, A. G. (1971). Spectra of some self-exciting and mutually exciting point processes.
Biometrika, 58(1), 83–90.
Hull, J., & White, A. (2001). Valuing credit default swaps II: Modeling default correlations.
Journal of Derivatives, 8(3), 12–22.
Jarrow, R., & Turnbull, S. (1995). Pricing derivatives on financial securities subject to credit risk.
Journal of Finance, 50, 53–86.
Jarrow, R., & Yu, F. (2001). Counterparty risk and the pricing of defaultable securities. The
Journal of Finance, 56, 1765–1799.
Kealhofer, S. (1995). Managing default risk in portfolios of derivatives. In: David M. Rowe (ed.)
Derivative Credit Risk: Advances in Measurement and Management. London: Risk Publications.
Koyuoglu, H. U., & Hickman, A. (1998). Reconcilable differences. Risk, 11(10), 56–62.
Lando, D. (1998). On cox processes and credit risky securities. Review of Derivative Research, 2,
99–120.
Lando, D., & Nielsen, M. S. (2010). Correlation in corporate default: Contagion or conditional
independence? Journal of Financial Intermediation, 19(3), 355–372.
Li, D. X. (1999). On default correlation: A copula function approach. Working paper 99–07,
The Risk Metrics Group.
Li, D. X. (2000). On default correlation: A copula approach. Journal of Fixed Income, 9, 43–54.
Lucas, D. (1995). Default correlation and credit analysis. Journal of Fixed Income, 9(4), 76–87.
Merton, R. (1974). On the pricing of corporate debt: The risk structure of interest rates. Journal of
Finance, 29, 449–470.
Nagpal, K., & Bahar, R. (2001). Measuring default correlation. Risk, 14(3), 129–132.
316 L.-J. Kao et al.

Schönbucher, P. J. (2003). Credit derivatives pricing models. Wiley Finance, UK.

Schönbucher, P., & Schubert, M. (2001). Copula dependent default risk in intensity models.
Working paper, Department of Statistics, Bonn University.
Sklar, A. (1959). Fonctions de repartition a n dimensions et leurs marges. Public Institute Statistic
University Paris, 8, 229–231.
Vasicek, O. A. (1997). The loan loss distribution. Technical report, KMV Corporation.
Yu, F. (2005). Correlated default in reduced-form models. Journal of Investment Management,
3(4), 16–27.
Zhou, C. (2001). An analysis of default correlations and multiple defaults. Review of Financial
Studies, 14(2), 555–576.
Assessing Importance of Time-Series
Versus Cross-Sectional Changes in Panel 12
Data: A Study of International Variations
in Ex-Ante Equity Premia and Financial
Architecture

Raj Aggarwal and John W. Goodell

Contents
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
12.2 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
12.2.1 Panel Data Estimation Procedures: Time Series Versus Cross-Sectional
Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
12.2.2 Financial Architecture, Transactions Costs, and Risks . . . . . . . . . . . . . . . . . . . . . . . 321
12.2.3 Equity Premium as Measure of Equity Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
12.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
12.3.1 Estimating the Equity Premium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
12.3.2 Estimating the Cross-Border Determinants of Financial Architecture . . . . . . . 326
12.3.3 Estimation Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
12.3.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
12.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
12.4.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
12.4.2 Results of Baltagi-Li EC2SLS Modeling and Blundell-Bond Modeling . . . . 328
12.4.3 Additional Econometric Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
12.4.4 Discussion of Initial Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
12.4.5 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
12.4.6 Discussion of Robustness Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
12.4.7 Discussion of Differences in the Results of the Two
Different Modelings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
12.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Appendix 1: Pearson Correlation Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Appendix 2: Estimation Procedure of Baltagi-Li EC2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
Appendix 3: Estimation Procedure of Blundell and Bond (1998) System
GMM Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347

R. Aggarwal (*)
University of Akron, Akron, OH, USA
e-mail: aggarwa@uakron.edu
J.W. Goodell
College of Business Administration, University of Akron, Akron, OH, USA
e-mail: johngoo@uakron.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 317
DOI 10.1007/978-1-4614-7750-1_12,
# Springer Science+Business Media New York 2015
318 R. Aggarwal and J.W. Goodell

Abstract
In the study of economic and financial panel data, it is often important to
differentiate between time series and cross-sectional effects. We present
two estimation procedures that can do so and illustrate their application by
examining international variations in expected equity premia and financial archi-
tecture where a number of variables vary across time but not cross-sectionally,
while other variables vary cross-sectionally but not across time. Using two
different estimation procedures, we find a preference for market financing to be
negatively associated with the size of expected premia. However, we also find that
US corporate bond spreads negatively determine financial architecture according
to the first procedure but not according to the second estimation as US corporate
bond spreads change value each year but have the same value across countries.
Similarly some measures that change across countries but do not change across
time, such as cultural dimensions as well as the index of measures against self-
dealing, are significant determinants of financial architecture according second
estimation but not according to the first estimation. Our results show that using
these two estimation procedures together can assess time series versus cross-
sectional variations in panel data. This research should be of considerable interest
to empirical researchers.
We illustrate with simultaneous-equation modeling. Following a Hausman
test to determine whether to report fixed or random-effects estimates, we first
report random-effects estimates based on the estimation procedure of Baltagi
(Baltagi 1981; Baltagi and Li 1995; Baltagi and Li 1994). We consider that the
error component two-stage least squares (EC2SLS) estimator of Baltagi and Li
(1995) is more efficient than the generalized two-stage least squares (G2SLS)
estimator of Balestra and Varadharajan-Krishnakumar (1987). For our second
estimation procedure, for comparative purposes we use the dynamic panel
modeling estimates recommended by Blundell and Bond (1998). We employ
the model of Blundell and Bond (1998), as these authors argue that their
estimator is more appropriate than the Arellano and Bond (1991) model for
smaller time periods relative to the size of the panels. We also use this
two-step procedure and use as an independent variable the first lag of the
dependent variable, reporting robust standard errors of Windmeijer (2005).
Thus, our two different panel estimation techniques place differing emphases
on cross-sectional and time series effects, with the Baltagi-Li estimator
emphasizing cross-sectional effects and the Blundell-Bond estimator
emphasizing time series effects.

Keywords
Panel data estimates • Time series and cross-sectional effects • Econometrics •
Financial institutions • Banks • Financial markets • Comparative financial
systems • Legal traditions • Uncertainty avoidance • Trust • Property rights •
Error component two-stage least squares (EC2SLS) • Generalized two-stage
least squares (G2SLS)
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 319

12.1 Introduction

Recently, there has been a veritable explosion of studies in a wide variety of areas
that employ panel data econometric methodology. Analysis of panel data allows
independent variables to vary both cross-sectionally and across time, while
panel data econometrics correct for cross-correlations between time series and
cross-sectional error terms. In recent years there has been the introduction of
a number of new refinements in analyzing panel data (Petersen 2009; Wooldridge
2010), all maintaining the goal of accounting for cross-correlation between time
series and cross-sectional error terms when assessing coefficient significance.
However, in the study of economic and financial panel data, it is often important to
assess the differential impact of time series versus cross-sectional effects; but panel data
techniques are unclear how this may be accomplished. In other words, panel data
methodologies typically do not inform us fully regarding which effect (time series or
cross-sectional) is more important or more dominant within particular data sets or
contexts. In this chapter we employ two contrasting estimation procedures, which,
respectively, emphasize cross-sectional versus time series differences, to clarify the
impacts of these two influences. We undertake this comparison of econometric methods
within a finance-related context which takes into account possible endogeneity.
We illustrate with simultaneous-equation modeling (outlined below). Following
a Hausman test to determine whether to report fixed or random-effects estimates,
we first report random-effects estimates based on the estimation procedure of
Baltagi (1981; Baltagi and Li 1994, 1995). We consider that the error component
two-stage least squares (EC2SLS) estimator of Baltagi and Li (1995) is more
efficient than the generalized two-stage least squares (G2SLS) estimator of Balestra
and Varadharajan-Krishnakumar (1987).
For our second estimation procedure, for comparative purposes we use the
dynamic panel modeling estimates recommended by Blundell and Bond (1998).
We employ the model of Blundell and Bond (1998), as these authors argue that their
estimator is more appropriate than the Arellano and Bond (1991) model for smaller
time periods relative to the size of the panels. We also use the two-step procedure,
reporting robust standard errors of Windmeijer (2005). Within this modeling we use
as an independent variable the first lag of the dependent variable. Thus, our two
different panel estimation techniques place differing emphases on cross-sectional
and time series effects, with the Baltagi-Li estimator emphasizing cross-sectional
effects and the Blundell-Bond estimator emphasizing time series effects.
Regarding the context of this econometric study, recent research suggests that both
national variations in the structures of financial intermediation and equity premia are
determined by many similar socioeconomic factors so that national variations in equity
premia can be expected to influence national variations in the structure of financial
intermediation and vice versa. But, until very recently prior literatures in these two
areas have ignored each other and the resulting possible endogeneity problems. Recent
exceptions include Aggarwal and Goodell (2011a, b), both of which examine the role
of financial architecture in determining nations’ equity premia. In this chapter we use
this economic context to illustrate two differing estimation procedures.
320 R. Aggarwal and J.W. Goodell

We find that a number of independent variables used in this study significantly

determine financial architecture across nations and across time. While a number of
our results are consistent across differing estimation procedures, we also document
a number of results that differ according to the two estimation procedures. After
controlling for outliers and serial correlation in further robustness checks, we find
that some significant differences remain, differences that help assess the relative
influence of time series versus cross-sectional variations.
We document, using the two different estimation procedures, a preference for
market financing to be negatively associated with the size of expected premia,
However, after controlling for multicollinearity and serial correlation in robustness
checks, we also find that US corporate bond spreads negatively determine financial
architecture according to the first procedure but not according to the second estima-
tion as US corporate bond spreads change value each year but have the same value
across countries. Similarly some measures that change across countries but do not
change across time, such as cultural dimensions as well as the index of measures
against self-dealing, are significant determinants of financial architecture according
to the second estimation but not according to the first estimation. We conclude from
our presented example that the two estimation procedures can produce results with
different emphases with regard to cross-sectional and time series effects.

12.2 Literature

12.2.1 Panel Data Estimation Procedures: Time Series Versus

Cross-Sectional Effects

In recent years, the development of panel data econometrics has facilitated a large
increase in scholarship where panel data models are applicable. This is particularly
the case in international finance where data can be described across countries and time
as well as across industries and time. The development of panel data methods has
followed from the introduction by Hansen (1982) of Generalized Method of Moments
(GMM). GMM, including the use of instrumental variables, allows the implementa-
tion of consistent estimations based on conditional expectations which are inconsis-
tent with the use of earlier methods such as ordinary least squares regression.
Dynamic effects can render the fixed-effects estimator of panel models biased and
inconsistent, especially for data covering finite and short time periods. Among
alternative estimators that control for persistence is the system Generalized
Method of Moments (GMM) estimator proposed by Blundell and Bond (1998).
This procedure addresses econometric problems such as regressor endogeneity,
measurement error, and weak instruments while controlling for time-invariant,
country-specific effects such as distance or common language. Arellano and Bond
(1991) suggest transforming the model, either in the first differences or in orthogonal
deviations, to eliminate the fixed effects and to estimate it by using the two-step
GMM estimator. The second and higher lags of the endogenous variable in levels are
suitable instruments to overcome the estimation problem. However, when data are
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 321

highly persistent, Blundell and Bond (1998) argue that this procedure can be
improved by using the system – GMM estimation – which supplements the equations
in the first differences with equations in levels. For the equations in the first differ-
ences, the instruments used are the lagged levels, and for the equations in levels, the
instruments are the lagged differences.
In addition, the use of simultaneous-equation modeling on the same data
set allows us to assess the results using the Blundell and Bond estimation procedure
against an alternative. Following a Hausman test to determine whether to report
fixed or random-effects estimates, we report random-effects estimates based on the
estimation procedure of Baltagi (1981; Baltagi and Li 1994, 1995). We believe that
the error component two-stage least squares (EC2SLS) estimator of Baltagi and Li
(1995) is more efficient than the generalized two-stage least squares (G2SLS)
estimator of Balestra and Varadharajan-Krishnakumar (1987) because of
a broader set of transformations of the instruments.
It is very useful to examine the differences obtained with these two estimation
procedures. These two panel data estimation procedures have different emphases
with regard to cross-sectional versus time series effects. Because of the nature of
their construction, while the Baltagi-Li estimator emphasizes cross-sectional
effects, the Blundell-Bond estimator emphasizes time series effects.

12.2.2 Financial Architecture, Transactions Costs, and Risks

The channeling of funds from savers to investors, or financial intermediation, is

a necessary function in all countries and is generally undertaken primarily through
financial institutions and/or through financial markets. Either financing channel must
resolve the issues of asymmetric information, adverse selection, and agency costs
involved in financing contracts that cover the monitoring and collection of funds provided
by savers to investors. Given that all optimal contracts are incomplete, the efficacy and
efficiency of overcoming contracting costs depends on the nature of “hold-up” costs in
a country, i.e., the ability and willingness of the contracting parties to try and take
advantage of each other.
This ability and willingness to take advantage of the other party in incomplete
contracts depend not only on industrial structure and the legal environment
(reflecting the relative power of the contracting agents and legal constraints on
their behavior) but also on ethical and other informal conventions that depend on
social and cultural values. As these differ from country to country and given that
institutions and markets differ in how they enforce incomplete contracts, financial
institutions may be optimal in some combinations of ethical, cultural, and social
conditions, while financial markets may be optimal in other conditions, and financial
institutions may be favored in some countries, while financial markets are favored in
other countries. Recent research notes that national preferences for market financing
increase with political stability, societal openness, economic inequality, and equity
market concentration and decrease with regulatory quality and ambiguity aversion
(e.g., Modigliani and Perotti 2000; Ergungor 2004; Kwok and Tadesse 2006;
Aggarwal and Goodell 2009).
322 R. Aggarwal and J.W. Goodell

Of course, as noted by Coase (1960), in a theoretically ideal and perfect financial

system, it would make no difference whether financial intermediation was privately done
through banks or publicly through markets. However, in reality many factors must be
considered. According to North (1990), the costliness of information needed for mea-
surement and enforcement of exchanges creates “transaction costs.” Transaction costs
involve costs of defining property rights and costs of enforcing contracts – including costs
of information. “Transformation costs” are the costs associated with using technology
and the efficiency of factor and product markets and are reflected in transactions costs.
Whether institutions lower or raise overall transactions costs has to do in part with the
ability of participants to be informed and to understand the nature of the particular
institutional environment. This includes not just understanding the nature of contracts and
their enforceability but also the temperament and motivations of other participants.
Additional transactions costs may also be associated with market transactions. As
noted by Williamson (1988) and others more recently (e.g., Aggarwal and Zhao
2009), Transaction Cost Economics (TCE) suggests that when the costs of market
exchange are sufficiently high, firms can obtain cheaper financing through some
other means. The alternative to market financing is typically through some sort of a
prescribed arrangement, such as a bank loan or, more broadly, through a prescribed
transfer of resources through a horizontal or vertical network. Hart (1995, 2001)
recognize that the primary transaction costs of market exchanges stem from the
uncertainties of contracts. From the point of view of the equity investor, obtaining
reliable information about firms is innately costly and, to some degree, fallible.
These costs will be shared with the supplier of equity, causing equity financing to be
more costly for the firm. This view is supported by Bhattacharya and Thakor (1993)
who suggest that a unifying thread among a great number of papers on banking is
that “intermediation is a response to the inability of market-mediated mechanisms to
efficiently resolve informational problems.”
Modigliani and Perotti (2000) theorize that when societies’ enforcement regimes
are not adequate, bank financing is favored. In this instance the binding of transactions
becomes more private than public and reflects longer-term reputations and relation-
ships between the parties, such as those between firms and their banks. Modigliani and
Perotti (2000) suggest that when the rights of minority (or outside) investors are not
adequate, less equity investment will be available for new enterprises (also see Myers
1977). According to Modigliani and Perotti (2000), in such societies, there will be
more bank lendings instead of financing with public equity.
Modigliani and Perotti (2000) also suggest that banks, because of an emphasis
on collateral, are less likely or able to differentiate firms with good future prospects
versus those with poor future prospects. Alternatively, markets with good gover-
nance are better able to distinguish between these types of firms, a view supported
by recent literature.1 However, Rajan (1992) notes that a higher emphasis on

1
For instance, Shirai (2004) reports that, because of improvements in official oversight for the
period 1997–2001, Indian capital markets improved significantly in being able to differentiate
high-quality firms from low-quality firms.
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 323

collateral can also lead to such an informational advantage for banks such that they
can charge excessively high interest rates which can weaken economic develop-
ment. Underlying Rajan and Zingales (1998) is the notion that when reliable
information about firms is too difficult or costly for the general public, banks
provide delegated monitoring (see Diamond 1984).
As this brief review of the relevant literature indicates, the international
determinants of financial architecture must include national characteristics such
as quality of investor protection, cultural and legal variables, and the equity
premium. However, there may be an endogeneity problem as financial architecture
seems to depend on a number of similar variables and the two variables, financial
architecture and equity premia, may influence each other. Our research design
includes a resolution of this issue.

12.2.3 Equity Premium as Measure of Equity Risk

We can expect variations in the ex ante equity premia across countries. Equity
premia can be expected to reflect the price of risk in equity investments. Depending
on national characteristics such as the nature of their institutional structures and
their levels of financial development, countries may differ with regard to both the
risk involved in equity investments and in the price of such risk. One way to think
about this extra risk and its price is to think about how the supply and demand for
equity investments may differ across countries, especially as most countries have
less than perfect capital markets.
Ibbotson et al. (2006) suggest that because of many obstacles and limitations, the
supply and demand for equity in markets may not respond to market forces as
would be expected from a theoretical view of efficient markets.2 For example, the
supply of equity may be restricted as bureaucratic rules and regulations may deter the
formation and market listing of corporate shares. Similarly, the demand for equity
may be limited due to uncertain property rights and the unreliability of public
information on potential investments. As the equity premium is the price of equity
risk, it is determined by the balance between supply and demand for equity. In order
to understand the nature and size of equity premia, it is important to account for the
nature of equity demand and supply in actual, imperfect, markets.
While there are no perfect measures of national supply of equity, a number
of variables could be used as indirect indicators. For example, stock market
capitalization as a ratio of GDP is one such widely used measure. But this ratio
by itself is an inadequate proxy because this measure will rise with a rise in
valuation (as well as with new listings) and valuation is itself affected by supply.
Other arguable measures of equity supply would include the number of shares

2
Ibbotson et al. (2006) actually start with somewhat of an alternative view to the commonly held
notion that prices in capital markets are set by the supply and demand for capital. Instead, they
focus on the viewpoint of the supplier of capital (an investor) and suggest that there is a supply and
demand for returns and that it is returns that are priced in the marketplace.
324 R. Aggarwal and J.W. Goodell

listed per unit of population or the number of new shares listed. A more extensive
evaluation of how developed an equity market is might include R2, the degree
to which individual stocks move synchronously with the overall equity market in
that country (Morck et al. 2000), or the degree to which market capitalization
is concentrated in a few firms. Turner (2003) associates the development of
nations’ private bond markets with the quality of their local investor bases.
Regulatory restrictions and lack of accounting standards can inhibit bond trading
by institutional investors. So if the quality of local bond markets in some measure
reflects the narrowness of the local investor base, then the quality of the bond
market might partially and indirectly determine the nonpecuniary (e.g., for control
rights) demand for equity.
Similarly, the demand for equity returns is likely to be influenced by a great
variety of factors that influence the risk level of equity and society’s perceptions,
tolerance, or price of equity risk. The nature of legal protection for investors,
disclosure requirements, the level of social trust that a particular society
believes can be placed in strangers, and the political stability of a country certainly
are some factors that come to mind. However, it is also reasonable to suppose that
there is correlation among many of the social, cultural, legal, and governance factors
that might affect the demand for equity. Ibbotson et al. (2006) suggest that the
demand for equity return is potentially also affected by concern for real returns as
opposed to nominal returns. Further, Moerman and van Dijk (2010) document
evidence that inflation risk is priced in international asset returns – so our investiga-
tion of the demand for equity returns also ought to control for inflation variability.
The supply and demand for equity in a country are also likely to be affected by
its financial architecture, i.e., the relative importance of the banking sector versus
financial markets in a country. It is now well recognized that some countries like
Japan and Germany are bank oriented, while other countries, like the Anglo-Saxon
countries, depend more on financial markets. For example, it can be expected that
equity premia are likely to be lower in countries with well-developed financial and
equity markets with a less restricted, and so greater, supply of equity than in bank-
oriented countries.

12.3 Methodology

As discussed briefly above, it seems that both national variations in the structures of
financial intermediation and equity premia are determined by many similar socio-
economic factors so that national variations in equity premia can be expected to
influence national variations in the structure of financial intermediation and vice
versa. But, until very recently, prior literatures in these two areas have ignored each
other and the resulting possible endogeneity problem. Recent exceptions include
Aggarwal and Goodell (2011a, b), both of which papers examine the role of financial
architecture in partial determining nations’ equity premia. This chapter overcomes
this limitation of prior research and examines the determinants of international
variations in financial architecture, accounting for the relationship between financing
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 325

architecture and ex ante equity risk premia with different estimation techniques.
One estimation technique uses simultaneous-equation estimates that include the
equity premium as an instrumental variable along with a number of other relevant
institutional, governance, and cultural factors. Another estimation technique uses
a dynamic panel data modeling which includes as an independent variable the ex ante
equity premium as a predetermined variable. Specifically, this chapter examines the
nature of international variations in national financial architecture with a special
emphasis on exploring the role of ex ante equity premia.

12.3.1 Estimating the Equity Premium

This chapter uses the abnormal earnings growth (AEG) model to estimate the ex
ante equity premium. This model is chosen not just because of its efficacy at
predicting ex post values but rather also because it is regarded as a fundamental
model upon which other previously used models are based (Penman 2005). In
addition, we use AEG estimations rather than alternative estimations in order to
avoid assumptions of future payout ratios. Further, the AEG model avoids the
problem with some estimation procedures when composite market-to-book ratios
for particular country/years are less than one.
Our estimates of ex ante equity premiums are based on data from the
period 1996–2006. Consensus of all available individual earnings forecasts
are obtained from Institutional Brokers Estimate System (I/B/E/S) of Thomson
Financial, as well as data for actual earnings per share, dividends per share, share
prices, and the number of shares outstanding. In the interests of consistency, forecasts
are collected as of April each year. Most firms end their fiscal years in December,
and, typically, values are reported within 90 days of the fiscal year end. If a firm has
I/B/E/S earnings forecasts for +1 and +2 years and a 5-year growth forecast, it is
retained in the sample. Firm-level data are then aggregated for each year. This paper
uses national treasury bill rates as a measure of the risk-free rate. This data is obtained
from International Financial Statistics of the International Monetary Fund (IMF).
This chapter estimates that abnormal earnings grow at a constant rate after
5 years at a rate equal to the rate of expected nominal GDP growth. Expected
GDP growth is modeled according to the following equation:

expNomGDPgrowth ¼ ð1 þ expInflationÞ  ð1 þ expRealGDPgrowthÞ  1

(12.1)

Expected inflation is modeled as the arithmetic average of the current, preceding,

and subsequent years’ inflation as reported by IFS. Expected real GDP growth is
modeled as the arithmetic average of the current, preceding and subsequent years’ real
GDP growth as reported by World Development Indicators. If national treasury
bill rates are not available, a 1-year money market rate or a similar country-specific
short-term rate is used.
326 R. Aggarwal and J.W. Goodell

This chapter uses the following AEG model:

!
eps1 gst þ k eps1  glt
d1
p0 ¼ (12.2)
k k  glt

where eps1 is the next-period forecasted earnings per share (when all firms aggre-
gated, this becomes next year’s forecasted earnings), p0 is the price which when all
firms are aggregated becomes aggregate market values, d1 are the dividends per
share which become aggregate dividends, glt is the long-term growth rate, and gst is
the short-term growth rate. We proxy the short-term rate as the geometric average
of growth in earnings forecasts up to the fifth forecast. The long-term growth rate
is proxied as the rate of expected nominal GDP growth. We do not incorporate
the 5-year growth forecast from I/B/E/S in the estimate of the short-term growth
rate because of concerns that, by using an average of the 5-year growth
forecast from I/B/E/S, a bias will be introduced as a consequence of equal
weighting across years.

12.3.2 Estimating the Cross-Border Determinants

of Financial Architecture

Besides the use of expected equity premia as an instrumental variable, we also use
a number of other independent variables. We control for inflation variability
(INFLATION_VOLATILITY) using the variance of the preceding 5 years as
reported by IFS. To capture time variation in risk appetite, we include the annual
difference from Moody’s Baa corporate bond yield and 10-year Treasury yield as
reported by the St Louis Federal Reserve for the middle of April for each year.
STOCK_VOLATILITY is the annualized standard deviation of monthly equity
returns of the respective Morgan Stanley Country Index. To account for relative
firm size and equity market concentration, we establish a Herfindahl index for each
country for each year (CONCENTRATION). A value of CONCENTRATION
close to 1 would suggest that most market capitalization for a particular country
in a given year is due to a small number of firms. Market capitalization data is
obtained from I/B/E/S. Based on Aggarwal and Goodell (2009), we speculate in
advance of empirical findings that country/years with high market capitalization
concentration will be more market based.
We control for regional differences by including as an independent variable
a dummy variable and REGION_EUROPE that receive “1” if the country is Europe.
We consider that the European region has historically had a unique relationship with
banking (see Rajan and Zingales 2003). We also control for cross-national differ-
ences in wealth and wealth inequality by including real GDP and the Gini coefficient
from World Development Indicators.
In order to assess the cultural impact on financial architecture, we include
as independent variables four cultural dimensions of Hofstede (2001):
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 327

uncertainty avoidance (UAI), power distance (PDI), masculinity (MAS), and

individualism (IDV). Kwok and Tadesse (2006) and others have found that
national culture is an important partial determinant of nations’ preference for
markets versus banks.
We control for regulatory quality using a governance indicator from Djankov
et al. (2008) his is the comprehensive index of cross-national differences in
protection of investors from self-dealing, the anti-self-dealing index
(ANTI_SELF_DEAL). Fligstein (2001) suggests that market participants primarily
value stable worlds and that markets are state-directed societal solutions to
competition. More efficient regulation and control of corruption will
improve transparency and so lower costs of resolving the asymmetry of
information inherent with markets. Aggarwal and Goodell (2009), Kwok and
Tadesse (2006), and Ergungor (2004) have found that governance impacts
financing preferences.

12.3.3 Estimation Methodologies

In order to use the estimator of Baltagi and Li (1995), we design a set of simulta-
neous equations. With regard to our modeling of simultaneous equations, we
suggest that the risk factors and costs generally involved with resolving asymmetric
information in markets is reflected in the equity premium. Therefore we model
financial architecture as partially determined by the respective country/year equity
premium as well as political factors. Our empirical estimation models are based on
this set of equations:

X
FIN_ARCHit ¼ ait þ b1it  EQ_PREMit þ b2it X2it þ eit (12.3)
X
EQ_PREMit ¼ ait þ b1it FIN_ARCHit þ b2it X1it þ eit (12.4)

In Eqs. 12.3 and 12.4, FIN_ARCH is domestic stock market capitalization

divided by domestic assets of deposit money banks; this measure is
constructed from the data from Beck et al. (2000). EQ_PREM is our estimate of
the equity premium, and X1 and X2 represent a number of independent
variables, including risk, political, and social factors. Alternatively, as noted earlier,
we also use, for the same respective sets of independent variables, the
results of Blundell and Bond (1998) dynamic panel estimation procedures.

12.3.4 Data

We analyze data for 41 countries over a recent 11-year period using panel data
methods. Specifically, we assess data for the 11-year period, 1996–2006, from
328 R. Aggarwal and J.W. Goodell

41 countries: Argentina, Australia, Austria, Belgium, Brazil, Canada, Chile, Czech

Republic, Denmark, Finland, France, Germany, Greece, India, Indonesia, Ireland,
Italy, Hong Kong, Hungary, Israel, Japan, Korea, Malaysia, Mexico, New Zealand,
the Netherlands, Norway, Philippines, Pakistan, Peru, Poland, Portugal, Russia,
Singapore, South Africa, Spain, Sweden, Thailand, Turkey, the United Kingdom,
and the United States.3 The countries we include here cover much breadth across
regions, cultures, legal origins, and difference in national wealth. In selecting
independent variables we have tried and managed to avoid excess correlation
among them. We estimate models that focus on structural variables and then add
in turn a regional variable and then cultural, governance, and security protection
variables. All models have variance inflation factors (VIF) of less than 10 for all
regressors indicating that any multicollinearity is unlikely to be a problem. Never-
theless, in subsequent robustness checks, we also address other specific correlations
among particular pairs of independent variables.

12.4 Results

12.4.1 Descriptive Statistics

Table 12.3 presents the means, and standard deviations of the variables used in our
models. Together these independent variables reflect the factors we described
earlier that may affect cross-national differences in financing predilections. Column
4 shows the standard deviation divided by the mean. This column suggests that the
dummy variable for European region, and our Herfindahl for market concentration
are the most variable of the independent variables (1.06, 0.96), while FIN_ARCH,
and EQUITY PREMIUM are similarly variable (0.73, 0.40) (Table 12.3).

12.4.2 Results of Baltagi-Li EC2SLS Modeling

and Blundell-Bond Modeling

Table 12.4 shows the results of panel regressions using the system of Eqs. 12.3 and
12.4, with our estimate of financial architecture as a dependent variable and our
estimate of the equity premium as an instrumented variable. The estimation is carried
out using the random-effects EC2SLS estimator proposed by Baltagi and Li (1992,
1995), and Baltagi (2001). Baltagi (2001) suggests the EC2SLS estimator is more

3
We begin our period of study in 1996 in order to include our measure of market concentration
which is a Herfindahl index we construct using data from I/B/E/S. There is insufficient data from
I/B/E/S for many countries prior to 1996. Another important reason, however, is that the measures
we use for political stability, control of corruption, and regulatory from Kaufmann et al. (2008) are
only available from 1996. Generally, our sample is restricted to those country/years which have
sufficient data reported by I/B/E/S and are included by Kaufmann et al. (2008). We stop in 2006 to
avoid the effects of the global financial crises and recession that started in 2007.
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 329

Table 12.1 Mean financial architecture

Country Mean Country Mean
South Africa 2.47 Greece 0.69
The United States 2.35 Israel 0.64
Finland 2.00 Korea 0.60
Hong Kong 1.93 Indonesia 0.59
Sweden 1.72 Spain 0.58
Chile 1.52 Brazil 0.58
Singapore 1.47 Norway 0.57
Russia 1.40 Hungary 0.57
Switzerland 1.36 Ireland 0.57
Argentina 1.31 Belgium 0.56
Malaysia 1.20 New Zealand 0.51
The United Kingdom 1.15 Pakistan 0.49
Australia 1.11 Italy 0.48
Philippines 0.99 Poland 0.45
Canada 0.92 Thailand 0.41
India 0.87 Czech Republic 0.41
Mexico 0.87 Japan 0.40
The Netherlands 0.75 Germany 0.33
Denmark 0.72 Portugal 0.31
France 0.72 Austria 0.18
Turkey 0.71
This table lists the mean national financial architecture for 41 countries for 1996–2006. Financial
architecture is domestic stock market capitalization divided by domestic assets of deposit money
banks (this measure is constructed from the data from Beck et al. (2000))

efficient than the usual Balestra and Varadharajan-Krishnakumar (1987) G2SLS

estimator because of a broader set of transformations of the instruments. Alternatively
Table 12.5 shows, for the same respective sets of independent variables as Table 12.4,
the results of Blundell and Bond (1998) dynamic panel estimates.
Table 12.4 reports the results of EC2SLS random-effects estimates, while
Table 12.5 reports Blundell and Bond (1998) estimates. Model 1 restricts the
independent variables to just the equity premium. In this model EQ_PREM is
negatively significant at 5 % in both Tables 12.4 and 12.5. Both estimation pro-
cedures suggest in Model 1 an association of lower equity premia with more
market-oriented countries.
Model 2 adds to the independent variables that describe the volatility of the
financial and economic environment: INFLATION_VOLATILITY and
STOCK_VOLATILITY. The equity premium is again negatively significant, at 1 %
in Table 12.4 and 10 % in Table 12.5. Equity premium is again negatively significant
according to both estimation procedures. INFLATION_VOLATILITY is not signif-
icant with regard to either estimation procedure, while STOCK_VOLATILITY is
negatively significant in both Tables 12.4 and 12.5.
330 R. Aggarwal and J.W. Goodell

Table 12.2 Mean equity premia

Country Mean Country Mean
Brazil 4.24 Finland 7.24
Turkey 1.61 Australia 7.33
Argentina 1.71 Hong Kong 7.37
Hungary 3.08 Germany 7.51
Mexico 3.62 Austria 7.65
Poland 3.75 Canada 7.65
Norway 3.82 France 7.84
India 4.44 Japan 7.85
Indonesia 5.19 Chile 8.03
South Africa 5.54 Greece 8.35
Belgium 6.01 Switzerland 8.38
Denmark 6.08 Portugal 8.38
The United Kingdom 6.24 The Netherlands 8.41
Italy 6.27 Malaysia 8.62
Israel 6.38 Thailand 8.85
Philippines 6.42 Spain 9.07
New Zealand 6.57 Korea 9.27
Singapore 6.64 Pakistan 10.61
Czech Republic 6.84 Ireland 10.65
Sweden 6.84 Russia 39.16
The United States 7.02
This table lists the mean expected national mean equity premia for 41 countries for 1996–2006.
This chapter
 uses the following
 AEG model
d
gst þ keps1  glt
p0 ¼ epsk 1 k  glt
1

Where eps1 is the next-period forecasted earnings per share (when all firms aggregated, this
becomes next year’s forecasted earnings), p0 is the price which when all firms are aggregated
becomes aggregate market values, d1 are the dividends per share which becomes aggregate
dividends, glt is the long-term growth rate, and gst is short-term growth rate. We proxy the
short-term rate as the geometric average of growth in earnings forecasts up to the fifth forecast.
The long-term growth rate is proxied as rate of expected nominal GDP growth

Model 3 adds to the other three independent variables CONCENTRATION,

REGION_EUROPE, and LN_GDP. This results in CONCENTRATION being
positively significant with regard to both estimation procedures. Equity premium
is again negatively significant in both tables. A difference in the two tables is that in
Table 12.4 REGION_EUROPE is negatively significant, while in Table 12.5 this
variable is not significant. STOCK_VOLATILITY is negatively significant in both
Tables 12.4 and 12.5, although only at 10 % in Table 12.4.
Model 4 add to the independent variables the Hofstede cultural dimensions UAI,
PDI, MAS, and IDV. Model 4 also adds SPREAD, the spread between US
corporate bonds and US 10-year Treasuries. This results in differences between
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 331

Table 12.3 Summary of data sources used in this study

1 2 3 4 5
Standard Stdev/
Variable Mean deviation mean Source
FIN_ARCH 0.92 0.67 0.73 Beck et al. (2000)
EQUITY_PREMIUM 7.28 2.91 0.40 Equity premium estimated from the
data from I/B/E/S and treasury rate
data from International Financial
Statistics
INFLATION_VOLATILITY 8873.06 296.65 0.03 Variance of preceding 5 years of
inflation
STOCK_VOLATILITY 24.49 14.77 0.60 The annualized standard deviation of
monthly equity returns of respective
Morgan Stanley Country Index
CONCENTRATION 0.18 0.17 0.96 Herfindahl index created for this
chapter based on number of firms
REGION_EUROPE 0.47 0.50 1.06 A dummy variable that is assigned
“1” if the country is in Europe and
“0” otherwise
LN_GDP 9.12 1.24 0.14 PPP GDP per capita from World
Development Indicators
UAI 62.76 24.88 0.40 Uncertainty Avoidance, Hofstede
(2001)
PDI 53.74 21.72 0.40 Power Distance, Hofstede (2001)
MAS 51.22 19.44 0.38 Masculinity, Hofstede (2001)
IDV 51.52 23.83 0.46 Individualism, Hofstede (2001)
SPREAD 2.15 0.53 0.25 Difference from Moody’s Baa
corporate bond yield and 10-year
Treasury yield St Louis Federal
Reserve
ANTI_SELF_DEAL 0.52 0.24 0.47 Djankov et al. (2008)
Means, standard deviations, and sources of variables used in statistical estimates reported in
succeeding tables, 1996–2006

Tables 12.4 and 12.5. SPREAD is significantly negative in Table 12.5 while not
significant in Table 12.4. Three cultural variables, PDI, MAS, and IDV, are
significant in Table 12.4 while not significant in Table 12.5. In Table 12.4, UAI and
MAS are negatively significant at 1 % and 5 %, respectively. In Table 12.4,
IDV and PDI are both positively significant at 1 %. However, UAI, MAS, IDV,
and PDI are not significant in Table 12.5. In other results, both tables show
LN_GDP positively significant, suggesting equity premia are larger in wealthier
countries.
Model 5 adds to the independent variables of Model 4, the comprehensive
index of anti-self-dealing (ANTI_SELF_DEAL) of Djankov et al. (2008). These
result, in Table 12.4, with regulatory efficiency, as reflected in measures against
self-dealing being positively significant. In Table 12.5, however, this variable is
not significant.
Table 12.4 Cross-national determinants of financial architecture: static panel estimation
332

Dependent variable: FIN_ARCH Model

1 2 3 4 5
INTERCEPT 2.73*** (0.001) 2.87*** (0.000) 0.75*** (0.173) –0.68 (0.232) –0.73 (0.255)
EQUITY_PREMIUM –0.25** (0.023) –0.19*** (0.003) –0.03* (0.065) –0.01 (0.385) –0.03** (0.034)
INFLATION_VOLATILITY –0.00 (0.302) –0.00 (0.109) –0.00 (0.136) –0.00 (0.132)
STOCK_VOLATILITY –0.03*** (0.008) –0.01* (0.061) –0.00 (0.280) –0.00* (0.064)
CONCENTRATION 0.58*** (0.007) 0.80*** (0.000) 0.86*** (0.000)
REGION_EUROPE –0.30*** (0.007) –0.37*** (0.000) –0.20 (0.140)
LN_GDP 0.00 (0.204) 0.13*** (0.005) 0.10* (0.067)
UAI –0.01*** (0.001) –0.00 (0.110)
PDI 0.01*** (0.000) 0.01*** (0.000)
MAS –0.00** (0.039) –0.00** (0.049)
IDV 0.01*** (0.000) 0.01*** (0.004)
SPREAD –0.00 (0.968) 0.04 (0.465)
ANTI_SELF_DEAL 0.46* (0.071)
Observations/groups 397/41 397/41 396/41 395/41 395/41
Hausman 1.05 (0.305) 1.69 (0.429) 3.35 (0.646) 4.75 (0.447) 3.65 (0.723)
R-square (within, between, overall) (0.01,0.04,0.00) (0.01,0.03,0.00) (0.03,0.00,0.01) (0.06,0.23,0.20) (0.04, 0.15,0.12)
Wald chi square 5.18** (0.023) 9.69** (0.021) 24.30*** (0.000) 74.21*** (0.000) 62.87*** (0.000)
Baltagi Wu LBI 0.89 0.88 0.95 0.97 0.97
This table reports results of tobit regressions for 41 countries, for 1996–2006. FIN_ARCH is stock market capitalization divided by domestic assets of deposit money
banks from Beck, Demirguc-Kunt, Levine (2000); EQUITY PREMIUM is ex ante equity premium estimations from the data from I/B/E/S; INFLATION_VO-
LATILITY is variance of preceding 5 years of inflation International Financial Statistics; STOCK_VOLATILITY is the annualized standard deviation of monthly
equity returns from respective Morgan Stanley Country Index; CONCENTRATION is a Herfindahl index of equity market concentration created for this chapter;
REGION_EUROPE is a dummy variable that is assigned “1” if market is in Europe and “0” otherwise; LN_GDP is natural log of real GDP per capita; UAI, PDI MAS,
and IDV are cultural dimensions of Hofstede (2001); SREAD is the difference between Moody’s Baa corporate bond yield and 10-year Treasury yield from St Louis
Federal Reserve; ANTI_SELF_DEAL is the index of measures against self-dealing from Djankov et al. (2008)
Variance inflation factors are less than 10 for all variables and models. P values in parentheses. *** Significance at 1 % level, ** significance at 5 % level,
*
significance at 10 % level random-effects, EC2SLS estimates reported
R. Aggarwal and J.W. Goodell
 12

Table 12.5 Cross-national determinants of financial architecture: dynamic panel estimation

Dependent variable: financial architecture Model
1 2 3 4 5
INTERCEPT 0.17*** (0.000) 0.36*** (0.000) 1.11** (0.046) –5.77 (0.240) –6.48 (0.244)
LAG1_FIN_ARCH 0.85*** (0.000) 0.88*** (0.000) 0.85*** (0.000) 0.77*** (0.000) 0.75*** (0.000)
EQUITY_PREMIUM –0.00** (0.021) –0.00* (0.063) –0.00*** (0.010) –0.00** (0.024) –0.00** (0.036)
INFLATION_VOLATILITY –0.00 (0.822) –0.00 (0.922) –0.00 (0.593) –0.00 (0.692)
STOCK_VOLATILITY –0.01*** (0.000) –0.01*** (0.000) –0.01*** (0.000) –0.01*** (0.000)
CONCENTRATION 0.97*** (0.000) 0.96*** (0.000) 1.10*** (0.000)
REGION_EUROPE 0.11 (0.391) 0.33 (0.786) 0.12 (0.936)
LN_GDP –0.09 (0.158) 0.55** (0.025) 0.73** (0.028)
UAI 0.02 (0.405) 0.03 (0.380)
PDI 0.01 (0.658) 0.01 (0.793)
MAS 0.01 (0.688) 0.01 (0.661)
IDV –0.02 (0.510) –0.03 (0.464)
SPREAD –0.09** (0.013) –0.08** (0.025)
ANTI_SELF_DEAL –1.43 (0.629)
Observations/groups 366/41 366/41 365/41 365/41 365/41
Wald chi square 1453.14*** (0.000) 358.33*** (0.000) 268.08*** (0.000) 318.68*** (0.000) 244.79*** (0.000)
AR tests –0.04(0.965) 0.14(0.889) –0.31(0.757) 0.22(0.825) 0.23(0.820)
–2.11(0.035) –2.07(0.038) –2.26(0.024) 2.30(0.022) 2.31(0.021)
–0.80(0.422) –1.13(0.257) –1.22(0.222) 1.01(0.314) 1.02(0.308)
2.18(0.029) 2.24(0.025) 2.12(0.034) 2.36(0.018) 2.29(0.022)
This table reports results of tobit regressions for 41 countries, for 1996–2006. FIN_ARCH is stock market capitalization divided by domestic assets of deposit money
banks from Beck, Demirguc-Kunt, Levine (2000), EQUITY PREMIUM is ex ante equity premium estimations from the data from I/B/E/S;
INFLATION_VOLATILITY is variance of preceding 5 years of inflation from International Financial Statistics; STOCK_VOLATILITY is the annualized standard
deviation of monthly equity returns of respective Morgan Stanley Country Index; CONCENTRATION is a Herfindahl index of equity market concentration created for
this chapter; REGION_EUROPE is a dummy variable that is assigned “1” if market is in Europe and “0” otherwise; LN_GDP is natural log of real GDP per capita;
UAI, PDI MAS, and IDV are cultural dimensions of Hofstede (2001); SREAD is the difference between Moody’s Baa corporate bond yield and 10-year Treasury yield
from St Louis Federal Reserve; ANTI_SELF_DEAL is the index of measures against self-dealing from Djankov et al. (2008)
Variance inflation factors are less than 10 for all variables and models. P values in parentheses. *** Significance at 1 % level, ** significance at 5 % level,
*
Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 333

significance at 10 % level Blundell-Bond estimates reported

334 R. Aggarwal and J.W. Goodell

12.4.3 Additional Econometric Details

In our results for Table 12.4, we consistently report random-effects estimates.

We do this because the results of Hausman tests (Hausman 1978) suggest that
random-effects estimates may be used. Reading across the models in Table 12.4, we
see that Hausman tests for the models are consistently insignificant. Therefore we
do not reject the NULL hypothesis that the estimates are inconsistent (Hausman
1978). In the Appendix we also report the Pearson correlation coefficients
(Appendix 1) and the variance inflation factors (Appendix 2) for our five models
in Tables 12.4 and 12.5. Examining Appendix 1, we see that the largest correlation
among our independent variables is –0.67 between IDV and PDI. The next largest
correlation is between PDI and LN_GDP (–0.63). Since PDI is a partner in the two
largest correlations in our sample, we exclude this independent variable in robust-
ness tests described in Table 12.6.
In Table 12.4, we report the locally invariant test statistic of Baltagi and Wu
(1999) (LBI). The LBI for our models in Table 12.4 are all not close to 2. This may
suggest that serial correlation is a problem for these models. However, while
commonly used, the significances of LBI are difficult to interpret.
Because the dynamic panel estimator of Blundell and Bond (1998) assumes that
there is no autocorrelation in the idiosyncratic errors, we also report in Table 12.5
Arellano-Bond test statistics. The Arellano-Bond test for zero correlation in
first-differenced errors is significant for differences of one lag. However, because
the first difference of white noise is necessarily autocorrelated, it is only necessary to
be concerned with second and higher autocorrelations. Examining the Arellano-Bond
test across the models in Table 12.5, we see that the values for lags 2 and 4 are also
consistently significant. In subsequent robustness tests described below, we also
control for other lags in order to control for serial correlation.

12.4.4 Discussion of Initial Results

With the exception of Model 4 in Table 12.4, EQUITY_PREMIUM is negatively

significant in every model of both Tables 12.4 and 12.5. Both modelings point
toward a negative association of the size of the expected equity premium with
financial architecture (the ratio of stock market size to banking size). It is also
intuitively reasonable to expect a lower equity premium in more market-oriented
nations. This is consistent with the results of Aggarwal and Goodell (2011a, b).
With the exception of Model 4 in Table 12.4, STOCK_VOLATILITY
is negatively significant in every model of both Tables 12.4 and 12.5.
CONCENTRATION is positively significant in every model in which it is present
in both Tables 12.4 and 12.5. REGION_EUROPE is negatively significant in
Models 3 and 4 in Table 12.4, but is not significant in any models in Table 12.5.
LN_GPD is positively significant in Models 4 and 5 of both Tables 12.4 and 12.5.
SPREAD is negatively significant in Models 4 and 5 in Table 12.5, but is not
significant in any models in Table 12.4. ANTI_SELF_DEAL is positively
 12

Table 12.6 Cross-national determinants of financial architecture: dynamic panel estimation controlling for three lags
Dependent variable: financial architecture Model
1 2 3 4 5
INTERCEPT 0.23** (0.041) 0.36*** (0.001) 0.86* (0.093) 1.71 (0.946) –2.76 (0.655)
LAG1_FIN_ARCH 1.34*** (0.000) 1.36*** (0.000) 1.31*** (0.000) 1.20*** (0.000) 1.11*** (0.000)
LAG2_FIN_ARCH –0.80** (0.019) –0.76** (0.012) –0.70*** (0.000) –0.65*** (0.000) –0.62*** (0.000)
LAG3_FIN_ARCH 0.26 (0.121) 0.25* (0.069) 0.26*** (0.000) 0.23*** (0.002) 0.19** (0.030)
EQUITY_PREMIUM –0.00** (0.014) –0.00 (0.147) –0.00*** (0.002) –0.00*** (0.000) –0.00*** (0.000)
INFLATION_VOLATILITY 0.00 (0.684) 0.00 (0.708) 0.00 (0.437) 0.00 (0.723)
STOCK_VOLATILITY –0.01*** (0.000) –0.01*** (0.000) –0.01* (0.052) –0.01 (0.180)
CONCENTRATION 0.91*** (0.000) 0.97*** (0.001) 1.28*** (0.000)
REGION_EUROPE 0.08 (0.629) 0.33 (0.490) –0.31 (0.766)
LN_GDP –0.06 (0.248) 0.04 (0.834) 0.45 (0.366)
UAI 0.00 (0.989) –0.00 (0.935)
PDI 0.01 (0.637) 0.02 (0.454)
MAS –0.01 (0.300) –0.02 (0.348)
IDV –0.00 (0.905) –0.00 (0.983)
SPREAD –0.12** (0.013) –0.09** (0.025)
(continued)
Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 335
 336

Table 12.6 (continued)

Dependent variable: financial architecture Model
1 2 3 4 5
ANTI_SELF_DEAL –2.02 (0.227)
Observations/groups 291/41 291/41 290/41 290/41 290/41
Wald chi square 644.37*** (0.000) 1026.17*** (0.000) 1631.51*** (0.000) 569.15*** (0.000) 244.79*** (0.000)
AR tests –1.04(0.298) –1.09(0.277) –1.74(0.082)* –1.90(0.057)* –1.84(0.066)*
–0.57(0.566) –0.77(0.442) –0.71(0.475) –1.03(0.301) –0.80(0.422)
0.204(0.839) 0.42(0.674) 0.00(0.999) –0.30(0.766) 0.05(0.963)
1.37(0.170) 1.37(0.171) 1.06(0.289) 1.50(0.133) 1.28(0.201)
This table reports results of tobit regressions for 41 countries, for 1996–2006. FIN_ARCH is stock market capitalization divided by domestic assets of deposit
money banks from from Beck, Demirguc-Kunt, Levine (2000), EQUITY PREMIUM is ex ante equity premium estimations from the data from I/B/E/S;
INFLATION_VOLATILITY is variance of preceding 5 years of inflation from International Financial Statistics; STOCK_VOLATILITY is the annualized
standard deviation of monthly equity returns of respective Morgan Stanley Country Index; CONCENTRATION is a Herfindahl index of equity market
concentration created for this chapter REGION_EUROPE is a dummy variable that is assigned “1” if market is in Europe and “0” otherwise; LN_GDP is
natural log of real GDP per capita;; UAI, PDI MAS, and IDV are cultural dimensions of Hofstede (2001); SREAD is the difference between Moody’s Baa
corporate bond yield and 10-year Treasury yield from St Louis Federal Reserve; ANTI_SELF_DEAL is the index of measures against self-dealing from
Djankov et al. (2008)
Variance inflation factors are less than 10 for all variables and models. P values in parentheses. Random-effects, EC2SLS estimates reported for Model 1.
***
Significance at 1 % level, ** significance at 5 % level, * significance at 10 % level Blundell-Bond estimates reported, controlling for three lags
R. Aggarwal and J.W. Goodell
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 337

significant (at 10 %) in Model 5 in Table 12.4 but is not significant in Table 12.5.
Many of the cultural dimensions of Hofstede (2001) are significant in Table 12.4.
For instance, PDI, IDV, and MAS are significant in every model in which they are
present, while UAI is significant in Model 4. However, none of these cultural
variables are significant in Table 12.5.
Summarizing these differences, both static and dynamic estimation methods find
evidence of a negative association of financial architecture with the expected equity
premium. This suggests that investors demand less return for holding equity in more
market-oriented countries. Both static and dynamic estimation methods find evidence of
a negative association of financial architecture with stock volatility. This suggests that
less volatile markets are associated with being more market-oriented. Both static and
dynamic estimation methods find evidence of a positive association of financial archi-
tecture with equity market concentration. This suggests that, despite the widely dispersed,
market orientation of the United States, overall, more market-oriented countries have
greater concentration of equity ownership into fewer firms. Both static and dynamic
estimation methods also find evidence of a positive association of financial architecture
with national wealth, suggesting that wealthier countries are more market oriented.
While the two estimation procedures used in Tables 12.4 and 12.5 yield very
similar results with respect to a number of independent variables, some differences
also exist. For instance, SPREAD is significantly negative in Table 12.5 but not
significant in Table 12.4. SPREAD changes value each year but is the same across
countries. We consider that the difference between Tables 12.4 and 12.5 with
respect to SPREAD may be due to a difference in emphasis between these
tables with regard to cross-sectional versus time series effects, with the Baltagi-Li
estimator emphasizing cross-sectional effects and the Blundell-Bond estimator
emphasizing time series effects. Similarly the cultural dimensions of Hofstede
(2001) – as well as ANTI_SELF_DEALING – do not change over our time period.
However these variables vary widely across countries. And so the result that these
variables are significant in Table 12.4 but not significant in Table 12.5 is consistent
with the notion that Table 12.4 emphasizes cross-sectional differences, while
Table 12.5 emphasizes time series differences.

12.4.5 Robustness Checks

As noted above, because the dynamic panel estimator of Blundell and Bond (1998)
assumes that there is no autocorrelation in the idiosyncratic errors, we also report in
Table 12.5 Arellano-Bond test statistics. The Arellano-Bond test for zero correla-
tion in first-differenced errors is significant for differences of one lag. However,
because the first difference of white noise is necessarily autocorrelated, it is only
necessary to be concerned with second and higher autocorrelations that are higher
than 1 for the models in Table 12.5. However, examining the Arellano-Bond test
across the models in Table 12.5, the values for lags 2 and 4 are also consistently
significant. Therefore, in Table 12.6, we also control for the first three lags of our
dependent variable.
338 R. Aggarwal and J.W. Goodell

Additionally, in order to avoid multicollinearity, we also employ, in Table 12.7,

a two-stage model for some variables in order to lower multicollinearity. We note
that the Pearson correlation coefficients between IDV and PDI as reported in
Appendix 1 is –0.67. Further, the correlation between PDI and LN_GDP is –0.63.
We first regress the respective independent variable (either IDV or PDI) on our
measure of wealth along with some other cultural variables. Specifically, we
regress each variable on wealth (LN_GDP). We then use the residuals from these
regressions as our independent variables in Eq. 12.3:
X
IDVit ¼ ai þ b1  LN_GDPit þ eit (12.5)
X
PDIit ¼ ai þ b1  LN_GDPit þ eit (12.6)

We then include the residuals from Eqs. 12.5 and 12.6 as substitute independent
variables for IDV and PDI, respectively. We refer to these variables as RESID_IDV
and RESID_PDI, respectively. This procedure lowers the correlation between IDV
(RESID_IDV) and PDI (RESID_PDI) to zero. The correlation between
RESID_IDV and RESID_PDI is now lowered to –0.45.

12.4.6 Discussion of Robustness Tests

The models in Table 12.6 are the same as those in Tables 12.4 and 12.5, with the
exception of including three lags of the dependent variable. The results of
Table 12.6 substantially corroborate the results of Tables 12.5 and 12.4.
EQUITY__PREMIUM is negatively significant in every model. CONCENTRA-
TION is positively significant. STOCK_VOLATILITY is generally negatively
significant, as in Tables 12.4 and 12.5. However, unlike the previous tables this
variable is not significant in Model 5. Like Table 12.5, SPREAD is negatively
significant. Overall the results of Table 12.6 suggest that the results of Table 12.5
are not driven by serial correlation in higher lags and/or excessive multicollinearity.
Table 12.7 shows the results of the most comprehensive model (Model 5 in
Tables 12.4, 12.5, and 12.6) using three lags of the dependent variable and replacing
IDV and PDI with RESID_IDV and RESID_PDI. As seen in Table 12.7, controlling
for correlation between PDI and IDV by orthogonalizing both to LN_GDP results in
little change to the estimates for our most comprehensive model. Model 1 in
Table 12.7 is very similar to Model 5 in Table 12.4. Similarly, Model 2 in Table 12.7
is very similar to Model 5 in Table 12.5.

12.4.7 Discussion of Differences in the Results of the Two

Different Modelings

As noted above, while the two estimation procedures used in this study yield
very similar results with respect to a number of independent variables, some
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 339

Table 12.7 Cross-national determinants of financial architecture: comparison of static and

dynamic panel modeling
Dependent variable: financial architecture Model
1 2
INTERCEPT 0.88* (0.088) 0.86 (0.774)
LAG1_FIN_ARCH 1.11*** (0.000)
LAG2_FIN_ARCH –0.62*** (0.000)
LAG3_FIN_ARCH 0.19** (0.030)
EQUITY_PREMIUM –0.03** (0.021) –0.00***(0.000)
INFLATION_VOLATILITY –0.00 (0.133) 0.00 (0.723)
STOCK_VOLATILITY –0.01** (0.050) –0.01 (0.180)
CONCENTRATION 0.91*** (0.000) 1.28*** (0.000)
REGION_EUROPE –0.17 (0.215) –0.31 (0.766)
LN_GDP 0.04 (0.419) 0.18 (0.486)
UAI –0.00 (0.199) –0.00 (0.935)
RESID_PDI 0.01*** (0.000) 0.02 (0.454)
MAS –0.00* (0.054) –0.02 (0.348)
RESID_IDV 0.01*** (0.007) –0.00 (0.983)
SPREAD 0.04 (0.405) –0.09** (0.025)
ANTI_SELF_DEAL 0.46* (0.071) –2.02 (0.227)
Observations/groups 395/41 290/41
Wald chi square 1453.14*** (0.000) 850.11*** (0.000)
Hausman 3.47 (0.838)
LBI 0.97
R-square (0.04, 0.14,0.11)
AR tests –1.84(0.066)
–0.80(0.422)
0.05(0.963)
1.28(0.201)
This table reports results of tobit regressions for 41 countries, for 1996–2006. FIN_ARCH is
stock market capitalization divided by domestic assets of deposit money banks from Beck,
Demirguc-Kunt, Levine (2000), EQUITY PREMIUM is ex ante equity premium estimations
from the data from I/B/E/S; INFLATION_VOLATILITY is variance of preceding 5 years of
inflation from International Financial Statistics; STOCK_VOLATILITY is the annualized stan-
dard deviation of monthly equity returns of respective Morgan Stanley Country Index; CONCEN-
TRATION is a Herfindahl index of equity market concentration created for this chapter;
REGION_EUROPE is a dummy variable that is assigned “1” if market is in Europe and “0”
otherwise; LN_GDP is natural log of real GDP per capita; UAI, PDI MAS, and IDV are cultural
dimensions of Hofstede (2001); SREAD is difference between Moody’s Baa corporate bond yield
and 10 year Treasury yield from St Louis Federal Reserve; ANTI_SELF_DEAL is the index of
measures against self-dealing from Djankov et al. (2008)
Variance inflation factors are less than 10 for all variables and models. P values in parentheses.
Random-effects, EC2SLS estimates reported for Model 1. Blundell-Bond dynamic panel
estimates (Modified to control for three lags) reported in Model 2. *** Significance at 1 % level,
**
significance at 5 % level, * significance at 10 % level
340 R. Aggarwal and J.W. Goodell

differences in the results using Blundell-Bond dynamic panel modeling and

Baltagi-Li static panel modeling persist. For instance, the variable SPREAD
which has a time series aspect but no cross-sectional aspect is significantly negative
in the estimation of Blundell-Bond but not significant using Baltagi-Li EC2SLS
estimator. We consider that the difference between these two estimators with
respect to SPREAD may be due to a difference in emphasis between these estima-
tors with regard to cross-sectional versus time series effects, with the Baltagi-Li
estimator emphasizing cross-sectional effects and the Blundell-Bond estimator
emphasizing time series effects. Similarly the cultural dimensions of Hofstede
(2001) and our index of measures against self-dealing have no time series aspect
but these variables vary widely across countries. These variables are significant
using Baltagi-Li estimation but not significant using Blundell-Bond estimation.
These differences in the results using differing estimation procedures are consistent
with the Baltagi-Li estimator emphasizing cross-sectional differences, while
Blundell-Bond estimator places greater emphasis on time series differences,
due to the fact that the Blundell-Bond modeling includes as an independent variable
the one-period lag in the dependent variable and so controls for fixed effects across
one-period lags.

12.5 Conclusions

There has been a veritable explosion of studies in a wide variety of areas

that employ panel data econometric methodology in recent years. Analysis of
panel data allows independent variables to vary both cross-sectionally and across
time, while panel data econometrics correct for cross-correlations between time
series and cross-sectional error terms. In addition, there has now been the intro-
duction of a number of new refinements in analyzing panel data (Petersen 2009;
Wooldridge 2010), all maintaining the goal of accounting for cross-correlation
between time series and cross-sectional error terms when assessing coefficient
significance.
However, in the study of economic and financial panel data, it is often important
to assess the differential impact of time series versus cross-sectional effects; but
panel data techniques are unclear how this may be accomplished. In other words,
panel data methodologies typically do not inform us fully regarding which
effect (time series or cross-sectional) is more important or more dominant within
particular data sets or contexts. In this chapter we employ two contrasting
estimation procedures, which, respectively, emphasize cross-sectional versus time
series differences, to clarify the impacts of these two influences. We undertake this
comparison of econometric methods within a finance-related context which takes
into account possible endogeneity.
In this chapter we show that a number of independent variables are significant in
partially determining financial architecture using either or both of the estimation
procedures of Baltagi and Li (1992) and Blundell and Bond (1998). Overall, we find
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 341

a negative association of the equity premium with financial architecture. A smaller

equity premium inclines societies toward being more market oriented. A Herfindahl
Index for equity market concentration is positively significant. Generally, when
market capitalization is concentrated in fewer firms, societies are more market
oriented. We also find that stock volatility is generally negatively significant.
This suggests that stock volatility inclines societies away from markets and
toward banks.
However, of greater interest for this study, we find that some of our independent
variables impact cross-national differences in financial architecture according to
Baltagi and Li (1992) estimation, while other independent variables impact finan-
cial architecture according to Blundell and Bond (1998) estimation. While the two
estimation procedures yield very similar results with respect to a number of
independent variables, some differences exist also.
US Corporate bond spreads negatively determine financial architecture
according to Blundell and Bond (1998) estimation but not according to Baltagi
and Li (1992) estimation. US Corporate bond spreads change value each year but
have the same value across countries. Similarly some measures that change across
countries but do not change across time, such as the cultural dimensions of Hofstede
(2001) as well as the index of measures against self-dealing, are significant deter-
minants of financial architecture according to Baltagi-Li estimation but not
according to Blundell-Bond estimation. This is consistent with different estimation
techniques placing differing emphasis on cross-sectional and time series effects,
with the Baltagi-Li estimator emphasizing cross-sectional effects and the Blundell-
Bond estimator emphasizing time series effects.
These are critical findings that have important implications for the use of panel
data estimation procedures, especially where it is important to differentiate between
time series and cross-sectional influences. Our results show that using the two panel
estimation procedures used here together can better differentiate between time
series and cross-sectional variations in panel data. Thus, our results should be of
much interest to scholars especially in finance and economics.

Appendix 1: Pearson Correlation Coefficients

1 2 3 4 5 6 7 8 9 10 11 12 13
1 FIN_ARCH 1
2 EQUITY_PREMIUM 0.03 1
3 INFLATION_VOLATILITY 0.07 0.16 1
4 MARKET_VOLATILITY 0.05 0.10 0.11 1
5 CONCENTRATION 0.03 0.00 0.04 0.30 1
6 REGION_EUROPE 0.19 0.12 0.10 0.14 0.18 1
7 LN_GDP 0.08 0.00 0.09 0.42 0.02 0.38 1
8 UAI 0.32 0.07 0.06 0.13 0.15 0.18 0.12 1
9 PDI 0.04 0.08 0.07 0.27 0.18 0.35 0.63 0.18 1
10 MAS 0.10 0.07 0.01 0.10 0.21 0.16 0.06 0.15 0.06 1

(continued)
342 R. Aggarwal and J.W. Goodell

1 2 3 4 5 6 7 8 9 10 11 12 13
11 IDV 0.10 0.05 0.07 0.37 0.00 0.44 0.51 0.17 0.67 0.12 1
12 SPREAD 0.05 0.14 0.08 0.06 0.10 0.02 0.00 0.02 0.01 0.01 0.01 1
13 ANTI_SELF_DEAL 0.29 0.05 0.09 0.09 0.20 0.48 0.05 0.49 0.02 0.00 0.07 0.01 1

Appendix 2: Estimation Procedure of Baltagi-Li EC2SLS

Following Baltagi (1981), consider the jth structural equation:

yj ¼ Y j aj þ Xj bj þ uj ¼ Z j dj þ uj , j ¼ 1, 2, . . . , M,
 
yj ¼ NT  1; Y j ¼ NT  Mj  1 ; Xj ¼ NT  K j ; (12.7)
  0  0 0
Zj ¼ Y j Xj ; dj ¼ aj ; bj ; nj ¼ Mj þ K j  1

The additive error components structure is given by

uj ¼ Z u uj þ Z l lj þ vj , j ¼ 1, 2, . . . , M, (12.8)
N N
where Zu ¼ IN eT, Zl ¼ eN IT

IN and IT are identifying matrices of order N and T, respectively. eN and eN are

vectors of ones of order N and T, respectively.
0 0
uj0 ¼ (u1j, u2j, . . . , uNj), lj ¼ (l1j,l2j, . . . , lTj), and vj ¼ (v1j, v2j, . . . , vNTj) are
random vectors with means of zero and the following covariance matrix:

0 1 2 2 3
uj   sujl I N 0 0
6 7
E@ l j A u 1 l 1 v 1 ¼ 4 0
0 0 0
s2ljl I T 0 5 (12.9)
vj 0 0 svjl I TN
2

for j and l ¼ 1,2,. . . .M. As noted by Baltagi (1981), this implies that the covariance
matrix between jth and lth structural equation is
 0
Sjl ¼ E uj ul ¼ s2ujl A þ s2ljl B þ s2vjl I NT , (12.10)

N N
where A ¼ IN eTe0 T, B ¼ eNe0 N IT
It follows that
   
J NT A J NT B J NT
Sij ¼ s23jl þ s1jl
2
 þ s2jl
2
 þ s2vjl Q (12.11)
NT T NT N NT
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 343

where
0
Q ¼ I NT  AT  NB þ JNT
NT
and JNT ¼ eNTeNT
also
s23jl ¼ s2vjl þ Ns2ljl þ Ts2ujl and s21jl ¼ s2vjl þ Ts2ujl as well as

s22jl ¼ s2vjl þ Ns2ljl

s2vjl , s21jl , s22jl , s21jl are the characteristic roots of ∑ij of multiplicity (N1)(T1),
N1, T1, and 1, respectively (Baltagi 1981; Nerlove 1971).
If the rank condition K  K1 + M1  1 holds for Eq. 12.7, then we can apply
transformation Qh to Eq. 12.7 to get

ðhÞ ðhÞ ðhÞ ðhÞ ðhÞ ðhÞ

y 1 ¼ Y 1 a1 þ X 1 b 1 þ u1 ¼ Z 1 d1 þ u1 (12.12)

1 ¼ Qhy1, . . .,; u1 ¼ Qhu1 for h ¼ 1,2,3.

y(h) (h)

As noted by Baltagi (1981), u(h)

1 has a covariance matrix which is a scalar times
an identity matrix.
However, Baltagi (1981) applies a 2SLS procedure to Eq. 12.12 in order to
correct for simultaneity:

0 0 0
Xh yh1 ¼ Xh Zh1 d1 þ Xh uh1 (12.13)

The resulting estimator of d1 is

h 0 i1 h 0 i
^d ðhÞ ðhÞ ðhÞ
1, 2SLS ¼ Z 1 PwðhÞ Z 1
h
Z 1 PwðhÞ yh1 (12.14)
where
 0
1 0
Pwh ¼ XðhÞ XðhÞ XðhÞ XðhÞ XðhÞ

The variance components are then estimated by

 0  
ðhÞ ðhÞ ðhÞ ðhÞ
y1  Z 1 ^d 1, 2SLS y1  Z 1 ^d 1, 2SLS =nðhÞ
ðhÞ2 ðhÞ ðhÞ
s11 ¼ (12.15)

As noted by Baltagi (1981), since d1 is common to all three transformations of

Eq. 12.13, we can apply an Aitken estimation to the following sets of equations
(Baltagi 1981):
0 0 1 0 ð1Þ0 ð1Þ 1 0 ð1Þ0 ð1Þ 1
Xð1Þ y1ð1Þ X Z1 X u1
B ð2Þ0 ð2Þ C B ð2Þ0 ð2Þ C B ð2Þ0 ð2Þ C
¼ d
@ X y 1 A @ X Z 1 A 1 @ X u1 Aþ (12.16)
0 0 0
Xð3Þ y1ð3Þ Xð3Þ Z1ð3Þ Xð3Þ u1ð3Þ
344 R. Aggarwal and J.W. Goodell

This results in the following estimator:

 h i1  h i
3 3
^ ðhÞ0 ðhÞ ðhÞ2 ðhÞ0 ðhÞ ðhÞ2
d 1, GLS ¼ S Z1 Pwh Z1 =s11 S Z1 Pwh y1 =s11 (12.17)
h¼1 h¼1

ðhÞ2
^ 11 from Eq. 12.15 into Eq. 12.17 yields
Substituting the estimate of s
  1   
3
ðhÞ0 ðhÞ2 3
ðhÞ0 ðhÞ2
^
d 1, EC2SLS ¼ S
ðhÞ
Z 1 Pwh Z 1 =^s 11 S
ðhÞ
Z 1 Pwh y1 =^s 11
h¼1 h¼1

Appendix 3: Estimation Procedure of Blundell and

Bond (1998) System GMM Estimation

We consider the first-order autoregressive panel data model:

yit ¼ ayi, t1 þ uit (12.18)

In this case, uit ¼ i + vit

where i ¼ 1, . . . N and t ¼ 2, . . . . . . . T
As described by Bun and Windmeijer (2010), following Blundell and Bond
(1998), we assume that i and vit have the error components structure

Eði Þ ¼ 0; Eðvit Þ ¼ 0; Eðvit i Þ ¼ 0 (12.19)

Additionally

Eðvit vis Þ ¼ 0 for t 6¼ s (12.20)

Additionally the initial condition satisfies

Eðyi1 vit Þ ¼ 0: (12.21)

It then follows that the following (T – 1)(T – 2)/2 linear moments conditions are
valid:
 
E yit2 Duit t ¼ 3, . . . : , T, (12.22)

where

yit2 ¼ ðyi1 ; yi2 ; . . . ; yit2 Þ0 and Duit ¼ uit  uit1 ¼ Dyit  aDyit1
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 345

Defining
2 3
2 3 ui3
yi1 0 0 ... 0 ... 0 6 ui4 7
6 7
6 0 yi1 yi2 . . . 0 . . . 0 7 7 6 : 7
Z di ¼ 6
4: : ; Du ¼ 6 7
: ... : ... : 5 i 6 : 7
6 7
0 0 0 . . . yi1 . . . yiT2 4 : 5
DuiT

This leads to a more efficient representing of the moment conditions:

 0 
E Z di Dui ¼ 0 (12.23)

Following Arellano and Bover (1995), the GMM estimator of a is given by

0 0
Dy1 Z d W 1
N Z d Dy
ad ¼ 0 1 0
Dy1 Zd W N Z d Dy1
0 0
where Dy ¼ (Dy1, Dy12 . . . DyN)0 , Dyi ¼ (Dyi3Dyi4, . . . , DyiT)0 , Dy1 the lagged
version of Dy, and Zd ¼ (Z’d1, Z’d2, . . . , Z’dN) ’, and WN is a weight determining the
efficiency of the GMM estimator.
ad is a GMM estimator in the differenced model. As noted by Bun and
Windmeijer (2010), this estimator is referred to as the difference GMM estimator.
However, following Blundell and Bond (1998), who follow on Arellano and
Bover (1995), we also assume additional conditions:

Eði Dyi2 Þ ¼ 0 (12.24)

This condition holds when the process is mean stationary:

i
yi1 ¼ þ ei (12.25)
1a

Additionally, E(ei) ¼ E(eii) ¼ 0.

Further, if Eqs. 12.19, 12.20, 12.21, and 12.24 hold, then further (T – 1)(T – 2)/2
linear moments conditions are valid:
 
E uit Dyit1 ¼ 0 for t ¼ 3, . . . , T, (12.26)

where

yit1 ¼ ðDyi2 , Dyi3 , . . . , Dyit1 Þ0

346 R. Aggarwal and J.W. Goodell

Defining
2 3
2 3 ui3
Dyi2 0 0 ... 0 ... 0 6 ui4 7
6 7
6 0 Dyi2 Dyi3 ... 0 ... 0 7 6 7
Z di ¼ 6 7; Dui ¼ 6 : 7
4: : : ... : ... : 5 6 : 7
6 7
0 0 0 . . . Dyi2 . . . DyiT1 4 : 5
DuiT
The above moment conditions can be written as
 0 
E Z li ui ¼ 0 (12.27)

The GMM estimator based on these estimates is then given by

0 0
y1 Zl W 1
N Zl y
al ¼ 0 1 0
y1 Zl W N Zl y1

al is a GMM estimator in the levels model. This is referred to as the levels’ GMM
estimator (Bun and Windmeijer 2010).
The full set of linear moments under assumptions (Eqs. 12.19, 12.20, 12.21, and
12.24) is
 
E yit2 Duit ¼ 0 (12.28)
 
E uit Dyi, t1 ¼ 0
for t ¼ 3, . . ., T
Following Bun and Windmeijer (2010), this can be resolved as
 0 
E Z si pi ¼ 0, (12.29)
where
2 3
Zdi 0 ... 0 
60 Dyi2 ... 0 7
Z si ¼ 6 7; pi ¼ Dui
4: : : : 5 ui
0 0 . . . DyiT

As noted by Wind, the GMM estimator based on these conditions is

0 0
q1 Z s W 1
N Zs q
as ¼ 0 1 0
q1 Zs W N Z s q1
where
 0 0 0
qi ¼ Dyi yi
as is the system GMM estimator of Blundell and Bond (1998).
12 Assessing Importance of Time-Series Versus Cross-Sectional Changes in Panel Data 347

Acknowledgments Authors are grateful to John Hunter, Maurice Peat, Min Qi, an anonymous
reviewer, and others for their kind advice on earlier versions but remain responsible for the
contents including any remaining errors.

References
Aggarwal, R., & Goodell, J. W. (2009). Markets and institutions in financial intermediation:
National characteristics as determinants. Journal of Banking and Finance, 33, 1770–1780.
Aggarwal, R., & Goodell, J. W. (2011a). International variations in expected equity premia: Role
of financial architecture and governance. Journal of Banking and Finance, 35, 3090–3100.
Aggarwal, R., & Goodell, J. W. (2011b). Variations in emerging market equity premia: Impact of
financial architecture. Journal of the Asia Pacific Economy, 16, 313–330.
Aggarwal, R., & Zhao, S. (2009). The diversification discount puzzle: Empirical evidence for
a transactions cost resolution. Financial Review, 44, 113–135.
Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence
and an application to employment equations. Review of Economic Studies, 58, 277–297.
Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-
components models. Journal of Econometrics, 68, 29–51.
Balestra, P., & Varadharajan-Krishnakumar, J. (1987). Full information estimations of a system of
simultaneous equations with error components structure. Econometric Theory, 3, 223–246.
Baltagi, B. H. (1981). Simultaneous equations with error components. Journal of Econometrics,
17, 189–200.
Baltagi, B. H. (2001). Econometric analysis of panel data. New York: Wiley.
Baltagi, B. H., & Li, Q. (1992). A note on the estimation of simultaneous equations with error
components. Econometric Theory, 8, 113–119.
Baltagi, B. H., & Li, Q. (1994). Estimating error component models with general MA
(q) disturbances. Econometric Theory, 10, 396–408.
Baltagi, B. H., & Li, Q. (1995). Testing AR(1) against MA(1) disturbances in an error component
model. Journal of Econometrics, 68, 133–151.
Baltagi, B. H., & Wu, P. X. (1999). Unequally spaced panel data regressions with Ar(1) distur-
bances. Economic Theory, 15, 814–823.
Beck, T., Demirguc-Kunt, A., & Levine, R. (2000). A new database on financial development and
structure. World Bank Economic Review, 14, 597–605.
Bhattacharya, S., & Thakor, A. V. (1993). Contemporary banking theory. Journal of Financial
Intermediation, 3, 2–50.
Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data
models. Journal of Econometrics, 87, 115–143.
Bun, M. J. G., & Windmeijer, F. (2010). The weak instrument problem of the system GMM
estimator in dynamic panel data models. Econometrics Journal, 13, 95–126.
Coase, R. H. (1960). The problem of social cost. Journal of Law and Economics, 3, 1–44.
Diamond, D. W. (1984). Financial intermediation and delegated monitoring. Review of Economic
Studies, 51, 393–414.
Djankov, S., La Porta, R., Lopez-de-Silvanes, F., & Shleifer, A. (2008). The law and economics of
self-dealing. Journal of Financial Economics, 88, 430–465.
Ergungor, O. E. (2004). Market- vs. bank-based financial systems: Do rights and regulations really
matter? Journal of Banking and Finance, 28, 2869–2887.
Fligstein, N. (2001). The architecture of markets: An economic sociology of twenty-first-century
capitalist societies. Princeton: Princeton University Press.
Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica, 50, 1029–1054.
Hart, O. (1995). Firms, contracts, and financial structure. New York: Oxford University Press.
Hart, O. (2001). Financial contracting. Journal of Economic Literature, 39, 1079–1100.
348 R. Aggarwal and J.W. Goodell

Hausman, J. A. (1978). Specification tests in econometrics. Econometrica, 6, 1251–1271.

Hofstede, G. (2001). Culture’s consequences. London: Sage.
Ibbotson, R. G., Diermeier, J. J., & Siegel, L. B. (2006). The demand for capital market returns:
A new equilibrium theory. In W. N. Goetzmann & R. G. Ibbotson (Eds.), The equity risk
premium. Oxford: Oxford University Press.
Kaufmann, D., Kraay, A., & Mastruzzi, M. (2008). Governance Matters Vii: Aggregate and
Individual Governance Indicators, 1996-2007. The World Bank.
Kwok, C. C., & Tadesse, S. (2006). National culture and financial systems. Journal of Interna-
tional Business Studies, 37, 227–247.
Modigliani, F., & Perotti, E. (2000). Security markets versus bank finance: Legal enforcement and
investor protection. International Review of Finance, 1, 81–96.
Moerman, G. A., & van Dijk, M. A. (2010). Inflation risk and international asset returns. Journal of
Banking and Finance, 34, 840–855.
Morck, R., Yeung, B., & Yu, W. (2000). The information content of stock markets: Why do
emerging markets have synchronous stock price movements? Journal of Financial Economics,
58, 215–260.
Myers, S. (1977). The determinants of corporate borrowing. Journal of Financial Economics, 5,
147–175.
Nerlove, M. (1971). A note on error components model. Econometrica, 39, 383–396.
North, D. C. (1990). Institutions, institutional change, and economic performance. Cambridge,
UK: Cambridge University Press.
Penman, S. H. (2005). Discussion of “on Accounting-Based Valuation Formulae” and “Expected
Eps and Eps Growth as Determinants of Value”. Review of Accounting Studies, 10(2–3):
367–378.
Petersen, M. A. (2009). Estimating standard errors in finance panel data sets: Comparing
approaches. Review of Financial Studies, 22, 435–480.
Rajan, R. G. (1992). Insiders and outsiders: The choice between informed and arm’s length debt.
Journal of Finance, 47, 1367–1400.
Rajan, R. G., & Zingales, L. (1998). Which capitalism? Lessons from the East Asian crisis.
Journal of Applied Corporate Finance, 11, 40–48.
Rajan, R. G., & Zingales, L. (2003). Saving Capitalism from the Capitalists. New York: Crown
Business.
Shirai, S. (2004). Assessing the impact of financial and capital market reforms on firms’ corporate
financing patterns in India. South Asia Economic Journal, 5, 189–208.
Turner, P. (2003). Bond market development: What are the policy issues? In R. E. Litan,
M. Pomerleano, & V. Sundararajan (Eds.), The future of domestic capital markets in develop-
ing countries. Washington: Brookings Institute Press.
Williamson, O. E. (1988). Corporate finance and corporate governance. Journal of Finance, 43,
567–592.
Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step
GMM estimators. Journal of Econometrics, 126, 25–51.
Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data. Cambridge, MA:
MIT Press.
Does Banking Capital Reduce Risk?
An Application of Stochastic Frontier 13
Analysis and GMM Approach

Wan-Jiun Paul Chiou and Robert L. Porter

Contents
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
13.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
13.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
13.3.1 Instrumental Variable for Capital . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
13.3.2 Generalized Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
13.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
13.4.1 Overall Observations of BHC Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
13.4.2 Instrumental Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
13.4.3 Measures of Bank Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
13.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
13.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Appendix 1: Stochastic Frontier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Appendix 2: Generalized Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380

Abstract
In this chapter, we thoroughly analyze the relationship between capital and bank
risk-taking. We collect cross section of bank holding company data from 1993 to
2008. To deal with the endogeneity between risk and capital, we employ
stochastic frontier analysis to create a new type of instrumental variable. The
unrestricted frontier model determines the highest possible profitability based
solely on the book value of assets employed. We develop a second frontier based

W.-J.P. Chiou (*)

Department of Finance and Law College of Business Administration, Central Michigan
University, Mount Pleasant, MI, USA
e-mail: Chiou1P@cmich.edu
R.L. Porter
Department of Finance School of Business, Quinnipiac University, Hamden, CT, USA
e-mail: RLPorter@quinnipiac.edu; robert.porter1@quinnipiac.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 349
DOI 10.1007/978-1-4614-7750-1_13,
# Springer Science+Business Media New York 2015
350 W.-J.P. Chiou and R.L. Porter

on the level of bank holding company capital as well as the amount of assets. The
implication of using the unrestricted model is that we are measuring the uncon-
ditional inefficiency of the banking organization.
We further apply generalized method of moments (GMM) regression to avoid
the problem caused by departure from normality. To control for the impact of
size on a bank’s risk-taking behavior, the book value of assets is considered in
the model. The relationship between the variables specifying bank behavior and
the use of equity is analyzed by GMM regression. Our results support the theory
that banks respond to higher capital ratios by increasing the risk in their earning
asset portfolios and off-balance-sheet activity. This perverse result suggests that
bank regulation should be thoroughly reexamined and alternative tools devel-
oped to ensure a stable financial system.

Keywords
Bank capital • Generalized method of moments • Stochastic frontier analysis •
Bank risks • Bank holding companies • Endogeneity of variables

13.1 Introduction

Bank capital management has become an important issue to both commercial

bankers and central bankers after the recent financial crisis. As the subprime
mortgage debacle spread, the balkanized regulatory system designed over half
a century ago appeared totally inadequate for today’s complex financial system.
In this study, we evaluate the role of capital in regulatory risk management by
examining a wide-range bank holding company data. Historically, both theoretical
and empirical papers on the relationship between capital and risk have produced
mixed results.1 Yet a new look at the role of bank capital in risk management is now
critical if we are to protect the financial system of the twenty-first century.
This chapter fits into a long history of literature dealing in general with bank risk
management and more specifically with the question of what constitutes an ade-
quate level of bank capital. Our contribution consists of the analysis of a large cross
section of bank holding companies over the years that the Basle Accords have been
implemented. In addition, we introduce a unique, to our knowledge, method to
exogenously model an instrumental variable for capital in a regression with risk.
One of the primary goals of bank regulators is to minimize the risk held on and
off their balance sheets by financial institutions. In this way, the negative external-
ities of bank failures and the risk to taxpayers from losses from the federal bank
safety net are avoided or reduced. A mandatory bank capital requirement is one of
the most important tools historically used by regulators to stabilize the financial
industry. The recent financial crisis, however, challenges the effectiveness of these

1
For detailed discussion, please see Berger et al. (1995); Gatev et al. (2009); Hovakimian and Kane
(2000); Shrieves and Dahl (1992); and VanHoose (2007).
13 Does Banking Capital Reduce Risk? 351

mandatory capital requirements. The inherent characteristics of today’s banking

industry such as rapid financial innovation, high financial leverage, information
asymmetry, liquidity creation, and the federal bank safety net all distort incentives
and reward risk-taking. If maintaining a certain level of capital is viewed by bank
managers only as a necessary evil, then critical questions emerge: How is capital
related to specific measures of risk including credit risk, liquidity risk, interest rate
risk, off-balance-sheet risk, market risk, and overall bank risk? Do higher levels of
capital improve or lower the efficiency of banking? Answers to these questions will
help to establish the role of capital in regulatory risk management.
Empirical studies of bank capital and bank risk, however, face an inherent
problem. In order to measure the effect of the level of capital on bank risk-taking,
it would be useful to regress risk, as the dependent variable, on capital, as the
independent variable. However, there is an obvious endogeneity problem.
The amount of risk a bank can undertake is dependent on its amount of capital,
and the amount of capital needed is dependent on the amount of risk that a bank
wants to undertake. In other words, they are jointly determined, much like price
and quantity in a basic microeconomic analysis. The solution to this problem is
normally either to use a simultaneous equation model or to use instrumental
variables. However, a simultaneous equation model must be properly identified,
and no one has yet been able to accomplish this in regard to risk and bank capital.
Likewise, no one, to our knowledge, has yet found a true instrument for capital that
is independent of risk.
We present a methodology for the development of an exogenous instrument for
capital in a regression with risk by using stochastic frontier analysis. First, we
determine the maximum possible income that can be achieved from a given level of
assets. This is referred to as fitting an upper envelope. Such a frontier is obviously
exogenous to any specific bank because it is determined by the data from all banks
in the sample. The distance from the frontier to any specific bank’s actual income
can be considered a measure of bank inefficiency. Next, to develop the instrument
for capital, we create a second frontier conditioned on bank capital as well as the
amount of assets employed. The incremental inefficiency from the second frontier
is a function of the bank’s capital but independent of the bank’s risk, and it is this
incremental inefficiency that we propose to use as an instrument for capital.
Our analysis adds to the existing literature in several ways. First, we employ
a large panel data set to consider the capital-risk relationship for a wider range of
bank holding companies than typically reviewed. Previous empirical studies have
commonly used market measures of risk. However, this approach necessarily limits
the sample to publically owned banks or bank holding companies. In this study, we
acknowledge the importance of small banks and bank holding companies, as well as
the largest bank holding companies. This concern is significant since public policy
related to the banking industry must consider a broad sample of banks and not only
the largest organizations. As a result, we turn to the typical accounting measures of
a bank’s risk and utilize a large panel data set. In a second contribution, stochastic
frontier analysis is applied to exogenously generate the effect of the use of capital in
banking.
352 W.-J.P. Chiou and R.L. Porter

Finally, the results provide evidence of bank holding companies reacting to

higher mandatory capital requirements by increasing the amount of risk the bank
holding company accepts. We use the generalized method of moments and look at
seven different measures of risk: credit risk, liquidity risk, interest rate risk,
off-balance-sheet risk, market risk, overall risk, and leverage risk. In general,
many results support the proposition that increased capital requirements reduce
risk in BHCs. There are, however, some results that suggest the opposite – that
BHCs increase risk as their capital ratios increase. This is obviously an important
finding with major public policy implications. If the primary tool used by regulators
to ensure a stable financial system is creating perverse results, then alternative tools
must be developed.
The rest of the chapter is organized as follows. Section 13.2 summarizes the
literature that deals with bank capital regulation. Section 13.3 presents our meth-
odology, and Sect. 13.4 reports the data along with its univariate analysis. In
Sect. 13.5, we present our empirical results. Section 13.6 concludes.

13.2 Literature Review

It has been argued that excessively high capital requirements can produce social
costs through lower levels of intermediation. In addition, there can be unintended
consequences of high capital requirements such as risk arbitrage (increasing risk to
offset the increase in capital and thereby maintain the same return on capital),
increased securitization, and increased off-balance-sheet activity, all of which
could mitigate the benefits of increased capital standards. See Berger et al. (1995)
and Santos (2001). The extent to which these unintended consequences played
a role in our recent crisis is yet to be determined.
Moral hazard is high on the list of problems receiving attention in this post-
financial crisis environment. The presence of a federal safety net creates moral
hazard because bank management does not have to worry about monitoring by
depositors (see Merton 1977; Buser et al. 1981; Laeven and Levine 2009). Absent
depositor monitoring, banks are free to increase risk. If, however, deposit insurance
and other elements of a federal safety net are reasons for increases in bank risk, why
do they continue to exist? The answer lies in the contemporary theory of financial
intermediation. It has been well established in the literature that there is need for
both demand deposit contracts and the possibility of bank runs (Diamond and
Dybvig 1983; Calomiris and Kahn 1991; Diamond and Rajan 2000; and Santos
2001). If the possibility of bank runs is needed, and bank runs are harmful, then
government deposit insurance is an optimal solution. There is a related issue. Banks
have a unique ability to resolve information asymmetries associated with risky
loans. As a result, bank failures can produce a serious contraction in credit avail-
ability, especially among borrowers without access to public capital markets. The
federal safety net is needed to avoid this credit contraction. Likewise, if a bank is
considered “too big to fail,” then the government will always bail the bank out, and
there is no reason for bank management to limit risk.
13 Does Banking Capital Reduce Risk? 353

It needs to be noted that not everyone is in agreement that the use of capital
requirements is the best way to reduce risk in banking. Marcus (1984) and Keeley
(1990) argue that a bank’s charter value mitigates against increased risk. Banks
operate in a regulated environment, and therefore, a charter to operate contains market
power. Excessive risk increases the cost of financial distress, and this can cause a loss
of charter value. Kim and Santomero (1988) argue that a simple capital ratio cannot be
effective, and any ratio would need to have exactly correct risk weights in a risk-based
system. Gorton and Pennacchi (1992) discuss “narrow banking” and propose splitting
the deposit services of banks from the credit services. In other words, the financial
system would include money market accounts and finance companies. The money
market accounts would only invest in short-term high-quality assets and leave the
lending to the finance companies that would not take in any deposits.
In Prescott (1997), he reviews the precommitment approach to risk management.
Briefly, banks commit to a level of capital, and if that level proves to be insufficient,
the bank is fined. This is used currently in the area of capital in support of a trading
portfolio but cannot be used for overall capital ratios since a fine against a failed
bank is not effective. Esty (1998) studies the impact of contingent liability of
stockholders on risk. In the late nineteenth and early twentieth century, bank
stockholders were subject to a call or an assessment for more money if needed to
meet the claims on a bank. There was a negative relation between increases in risk
and the possible call on bank stockholders. Calomiris (1999) makes a strong case
for requiring the use of subordinated debt in bank capital structures. The need to
issue unguaranteed debt and the associated market discipline would act as an
effective limit to the amount of risk a bank would be able to assume. John
et al. (2000) argue that a regulatory emphasis on capital ratios may not be effective
in controlling risk. Since all banks will have a different investment opportunity set,
an efficient allocation of funds must incorporate different risk-taking for different
investment schedules. These authors go on to argue that senior bank management
compensation contracts may be a more promising avenue to control risk using
incentive-compatible contracts to achieve the optimal level of risk.
Marcus and Shaked (1984) show how Merton’s (1977) put option pricing formula
can be made operational and then used the results to estimate appropriate deposit
insurance premium rates. The results of their empirical analysis indicated that the then
current FDIC premiums were higher than was warranted by the ex ante default risk of
the sample banks. This implies that banks are not transferring excessive risk to the
deposit insurance safety net and capital regulation is effectively working.
Duan et al. (1992) address the question of the impact of fixed-rate versus risk-
based deposit insurance premiums directly. The authors tested for specific risk-
shifting behavior by banks. If banks were able to increase the risk-adjusted value of
the deposit insurance premiums, then they had appropriated wealth from the
FDIC. This is because the FDIC, at the time, could not increase the insurance
premium even though risk had increased. Their empirical findings were that only
20 % of their sample banks were successful in risk-shifting behavior and therefore
the problem was not widespread. This also implies that capital management has
been effective.
354 W.-J.P. Chiou and R.L. Porter

Keeley (1992) empirically studied the impact of the establishment of objective

capital-to-assets ratio requirements in the early 1980s. His evidence documents an
increase in the book value capital-to-assets ratio of previously undercapitalized
banks, and this, of course, was the goal of the new capital regulations. His study,
however, is unable to confirm the same result when looking at the market value
capital ratios. While the market value capital-to-assets ratios also increased, there
was no significant difference between the undercapitalized banks and the ade-
quately capitalized banks. Nevertheless, this was more evidence that capital regu-
lation was working.
Hovakimian and Kane (2000) use the same empirical design as Duan
et al. (1992) but for a more recent time period, and they obtain opposite results.
They also start with the argument of Merton (1977) that the value of deposit
insurance increases in asset return variance and leverage. They regress the change
in leverage on the change in risk and find a positive rather than a negative coeffi-
cient. The coefficient must be negative if capital regulation forces banks to decrease
leverage with increases in risk. In a second test, they regress the change in the value
of the deposit insurance premium on the change in the asset return variance. Here
again the coefficient must be negative (or zero) if there is any restraint. In this
equation, the coefficient measures how much the bank can benefit from increasing
the volatility of its asset returns. The option-model evidence presented shows that
capital regulation has not prevented risk-shifting by banks and that it was possible
for banks to extract a deposit insurance subsidy.
In Hughes et al. (2001), the authors study the joint impact of two functions of
bank capital. First is the capital’s influence on market value conditioned on risk, and
second is its impact on production decisions incorporating endogenous risk. Effi-
cient BHCs are determined according to frontier analysis, and then these BHCs are
assumed to be value-maximizing firms. The conclusion is that these value-
maximizing firms do achieve economies of scale, but the analysis of production
must include capital structure and risk-taking.
Berger et al. (2008) note that US banks hold significantly more equity capital
than the minimum amount required by regulators. Their evidence documents the
active management of capital levels by BHCs including setting target levels of
capital above regulatory minimums and moving quickly to achieve their targets.
Over the 15-year period of their study, BHCs regularly used new issues of shares
and share repurchase programs to actively manage their capital levels. Several
reasons for differing capital ratios among BHCs are given by the authors. Banks
with high earnings volatility would likely hold more capital. Banks whose cus-
tomers are more sensitive to default risk via counterparty exposure may be forced to
hold more capital. Firms with high charter values will want to minimize their costs
of financial distress by maintaining high capital ratios. On the other hand, larger
banks by asset size tend to be more diversified, enjoy scale economies in risk
management, have ready access to capital markets, and are possibly viewed as “too
big to fail” with attendant implicit government guarantees.
Flannery and Rangan (2008) also document a large increase in bank capital
during the 1990s. The authors note the timing correlation with deregulation of the
13 Does Banking Capital Reduce Risk? 355

banking industry and the related increase in risk exposure. They suggest that
increased diversification may have been offset by the increased risk of the newly
permissible activities. As a result, it was counterparty risk that was the driving force
for higher capital ratios.

13.3 Methodology

13.3.1 Instrumental Variable for Capital

We differ from previous studies that deal with the endogeneity between risk and
capital using traditional methods such as a simultaneous equation approach or two-
or three-stage regression analysis.2 In this study, we follow the method and concept
of Hughes et al. (2001, 2003); and others and use stochastic frontier analysis to
estimate the inefficiency of our sample of bank holding companies. See Jondrow
et al. (1982) for a discussion of fitting production frontier models. We then create
a unique instrumental variable for bank capital to be used in regressions of capital
and risk. The question we ask is: “How efficient is a bank holding company in
converting the resources with which it has to work into profit?” The frontier
developed is exogenous to any specific bank since it is based on the results of all
banks in the sample. From this frontier, we measure the inefficiency of each bank as
the distance between the frontier and that specific bank’s pretax income. This
measure, however, must be adjusted for those elements that are beyond the control
of the bank.
Our unrestricted frontier model determines the highest possible profitability
based solely on the book value of assets employed. The unrestricted model is
specified as

PTI ðBVA; sBANK Þ ¼ a þ b1 BVA þ b2 ðBVAÞ2 þ e

e ¼ x  B

 (13.1)
x  iid N 0; s2x , Bð 0Þ  iid N 0; s2B

where PTI is pretax income, BVA is of book value of assets, x is statistical noise, B is
systematic shortfall (under management control), and B  0. A quadratic specifi-
cation is used to allow for a nonlinear relation between the pretax income and the
book value of assets.
Our next step is to develop a second frontier based on the level of bank holding
company capital as well as the amount of assets. The implication of using the
unrestricted model is that we are measuring the unconditional inefficiency of the
banking organization. By also conditioning the model on capital, we can develop
a measure of the incremental efficiency or inefficiency of an organization due to its

2
In Appendix 1, we explain the execution of stochastic frontier in this chapter.
356 W.-J.P. Chiou and R.L. Porter

capital level. It is this incremental inefficiency due to a bank’s capital level that we
propose to use as an instrument for capital in a regression of risk on capital.
Specifically, our restricted model, again in a quadratic form, is as follows:

PTI ðBVA; BVC; sBANK Þ ¼ a þ b1 BVA þ b2 ðBVAÞ2 þ b3 BVC þ e

e ¼ v  u    (13.2)
v  iid N 0; s2v uð 0Þ  iid N 0; s2u

where BVC is the book value of capital, v is statistical noise, and u denotes the
inefficiency of a bank considering its use of both assets and capital.
The two assessments of inefficiency allow us to measure the difference in
profitability due to the use of capital by calculating the difference in the inefficiency
between the restricted and unrestricted model. Specifically,

d¼uB (13.3)

This becomes our instrumental variable for capital. While any measure of
profitability endogenously includes risk, our instrument, the difference between
two measures of profitability conditioned only on capital, is related to capital but
not to risk which is included in both models.

13.3.2 Generalized Method of Moments

We first apply a generalized method of moments (GMM) regression in this study.3

There are several reasons why we need to consider the infeasibility of the OLS
regression. First, the departure from normality of the variable d due to the combined
error terms should be taken into account in the analysis. There is no theory to
support a Gaussian distribution of these variables. Furthermore, in practice, the
ranges of the independent and dependent variables are bounded within certain
intervals. Unlike other estimators, GMM is robust and does not require information
on the exact distribution of the disturbances. We follow Hamilton (1994) to
construct our GMM estimation. To control for the impact of size on a bank’s
risk-taking behavior, the book value of assets is considered in the model (Gatev
et al. 2009). The relationship between the variables specifying bank behavior and
the use of equity is analyzed by GMM regression. Specifically,

yk, t ¼ ct þ bk, t di, t þ gk, t lnðBVAÞ þ k, t , (13.4)

where yk,t is one of the measures of risk or behavior (e.g., total equity/total asset)
for bank i in year t; c is a constant; bk,t is the coefficient of instrumental

3
In Appendix 2, we provide the derivation of the GMM model.
13 Does Banking Capital Reduce Risk? 357

variable of capital, di,k, for k’s regression in year t; gk,t is the coefficient of natural
logarithm of bank’s book value; and k,t is the error term.

13.4 Data

We obtain our data on bank holding companies from Federal Reserve reports FR
Y-9C for the years 1993–2008. Data on risk-weighted assets, tier 1 capital, and tier
2 capital were not included with the FR Y-9C reports from 1993 to 1996. We were
graciously provided this missing information by the authors of Berger et al. (2008).
Table 13.1 displays the descriptive statistics of the sample BHCs in our analysis.
The total of 24,973 bank-year observations ranges from 2,256 in 2005 to 678 in
2008. From 2005 to 2006, there is an especially large drop in the number of BHCs
included in our data. This is primarily due to a change in the reporting criteria for
the FR Y-9C report. Starting in 2006, the threshold for required reporting by a BHC
was increased from BHCs with $150 million in total assets to BHCs with $500
million in total assets. Note that in spite of the 57 % drop in the number of BHCs
reporting in 2006 compared with 2005, the total assets represented in the sample for
these 2 years decreased by only 14 %.
Our data start in 1993 because 1992 was the final year in which capital ratios were
still adjusting in order to conform to the Basle I Capital Accord. As a result, 1993
represents the first year that does not include any mandated changes in the capital ratios.
The entire period of 1993–2008 contains a number of significant events affecting the
banking industry. For instance, the Riegle-Neal Interstate Banking and Branching Act
was passed in 1994 eliminating geographic restrictions on bank expansion. In 1999 the
Gramm-Leach-Bliley Financial Services Modernization Act was passed effectively
repealing the Glass-Steagall Act. Together these two acts overturned 65 years of
legislation and regulation intended to keep banks financially sound.
From an economic point of view, the early portion of our time period represented
a time of recovery from recession. The economy then moved from recovery to
growth, and the decade ended in a tech-stock boom followed by a bursting of the
tech-stock price bubble and an attendant recession. The new decade brought
traditional financial policies intended to stimulate the economy which, in hindsight,
probably helped to lay the foundation for the housing price bubble which precip-
itated the 2007–2009 financial crisis. The time period from 1993 to 2008 seems to
be a very appropriate period in which to analyze bank capital ratios.
Previous empirical studies have used market measures of risk and various risk
measures derived from a market model based on return data. However, this
approach necessarily limits the sample to publically owned banks or bank holding
companies. In this study, we wish to determine the impact of capital on various
measures of risk and acknowledge the importance of small banks and bank holding
companies, as well as the largest bank holding companies. This concern is signif-
icant since public policy related to the banking industry must consider the broadest
sample and not only the largest organizations. As a result, we utilize a large panel
data set and turn to the typical accounting measures of a bank’s risk.
 358

Table 13.1 Statistical summary

BVA (US$ million)
Year N Mean Max Min SD BE PTI ROE OHE OBS RA Cap E/A NPA Gap
1993 1,525 2,630 216,574 22 12,260 207 43 0.135 4.10 0.120 0.438 0.161 0.085 0.069 0.160
1994 1,306 3,442 250,489 8 15,334 265 57 0.125 4.45 0.139 0.470 0.164 0.087 0.060 0.166
1995 1,355 3,526 256,853 27 16,410 283 62 0.124 4.18 0.240 0.493 0.172 0.093 0.052 0.160
1996 1,405 3,516 336,099 28 18,238 283 63 0.127 4.27 0.220 0.585 0.148 0.093 0.047 0.121
1997 1,493 3,614 365,521 30 20,027 281 65 0.127 4.04 0.222 0.607 0.146 0.094 0.043 0.086
1998 1,563 3,742 668,641 32 28,629 292 62 0.124 4.04 0.218 0.621 0.157 0.094 0.039 0.073
1999 1,658 3,845 716,937 38 29,651 292 75 0.130 4.18 0.215 0.652 0.151 0.089 0.041 0.032
2000 1,726 3,674 715,348 38 29,664 293 66 0.122 4.11 0.190 0.672 0.145 0.091 0.045 0.042
2001 1,818 3,869 693,575 38 31,651 319 56 0.117 3.77 0.224 0.682 0.145 0.091 0.055 0.026
2002 1,968 3,911 758,800 40 32,953 330 66 0.124 3.72 0.215 0.681 0.149 0.093 0.056 0.066
2003 2,129 3,960 820,103 41 34,712 334 75 0.127 3.71 0.224 0.688 0.152 0.093 0.056 0.079
2004 2,240 4,558 1,157,248 40 45,012 421 77 0.123 3.99 0.225 0.709 0.150 0.092 0.044 0.115
2005 2,252 5,407 1,494,037 40 55,416 466 91 0.130 4.15 0.242 0.725 0.147 0.090 0.041 0.115
2006 969 10,842 1,463,685 43 76,903 942 179 0.124 3.49 0.338 0.764 0.137 0.091 0.046 0.086
2007 888 10,801 1,720,688 72 88,290 1,010 152 0.110 3.82 0.322 0.778 0.133 0.092 0.085 0.076
2008 678 13,884 2,175,052 79 123,635 1,194 70 0.084 3.21 0.315 0.760 0.145 0.092 0.085 0.041
The numbers of bank holding companies (BHCs) and statistics of their book value of asset (BVA) over sample years are reported. The means of other
descriptive statistics are listed: BE, book value of equity; PTI, pretax income; ROE, return of equity; OHE (overhead efficiency), noninterest expenses/
noninterest income; OBS (off-balance-sheet activities), all OBS activities/total assets; RA (risk-based asset ratio), total risk-based assets/total assets; Cap
(total risk-based capital ratio), (tier 1 capital + tier 2 capital)/total risk-based asset; E/A, total equity/total asset; NPA (nonperforming assets ratio),
nonperforming assets/total equity capital; Gap, (interest-sensitive gap) (IS assets – IS liabilities)/total asset. The BVA, BE, and PTI are in million US dollar
W.-J.P. Chiou and R.L. Porter
13 Does Banking Capital Reduce Risk? 359

13.4.1 Overall Observations of BHC Data

We see the significant events and the economic activity listed above in the statistics
in Table 13.1. First, the size of BHCs measured by either their asset values or equity
has increased, while the number of banks has decreased. This trend is still evident
after adjusting for changes in the reporting criteria for the FR Y-9C report. The
government deregulation noted above has resulted in increased concentration in the
banking industry. We note also the significant cross-sectional variation in scale of
BHCs that suggests the utilization and operation of their resources vary
considerably.
When we look in Table 13.1 at the basic leverage ratio of equity to assets (E/A),
we see a generally rising ratio. In 1993, the ratio was 8.5 %, while in 2008 it was
9.2 %. These ratios appear to be in line with mandatory capital requirements. We
also see variation in this trend consistent with prevailing economic activity. For
example, the decline from 9.4 % in 1998 to 8.9 % in 1999 reflects the tech-stock
problems of that time period. In Table 13.1, we also see a rising trend in RA, the
ratio of risk-based assets to total assets. Here, however, the trend is far more
pronounced, rising from 43.80 % in 1993 to 76.00 % in 2008. Confirmation of
these two trends comes from the trend in CAP, the ratio of tier 1 plus tier 2 capital to
risk-based assets. This ratio declines from 16.10 % in 1993 to 14.50 % in 2008.
While these ratios are substantially above the Basle Capital Accord standards, the
trend is clearly down.
Another dramatic trend over this time period is the increase in off-balance-sheet
activity. In Table 13.1, the off-balance-sheet activities-to-total assets ratio (OBS)
has increased from 12.00 % in 1993 to 31.50 % in 2008. While this trend is not
a surprise, we need to ask if there is capital to support this expansion and consider
the makeup of the components of off-balance-sheet activities. It is unclear whether
BHCs use off-balance-sheet activities to decrease or increase risk.
The time-varying overall performance measures of our sample of BHCs such as
pretax income (PTI), return on equity (ROE), nonperforming assets ratio (NPA),
and the interest-sensitive gap (Gap) are shaped by major economic occurrences and
policies. Return on equity has varied in a relatively narrow band over this time
period. With the exception of 2007 and 2008, the return on equity ranged from
12.20 % to 13.50 %. In line with the financial crisis that started in 2007, ROE
declined to 11.00 % in 2007 and to 8.40 % in 2008. It is also noteworthy that the
highest return on equity was in the first year of our sample period, 1993. Nonper-
forming assets appear to move in concert with economic activity. The recovery and
expansion period of 1993–1998 is marked by a steady decrease in the ratio of
nonperforming assets to equity. This is followed by an increase in this ratio during
the tech-stock bubble and recession after which we see another decline until the
crisis of 2007 and 2008.
Since the industrial structure of financial services changes intertemporally, we
analyze the risk and use of capital by BHCs year by year. The analysis suggests banks
progressively depend more on aggressive funding sources and new product lines over
our sample period. Given that financial leverage (e.g., equity/asset ratio) must remain
360 W.-J.P. Chiou and R.L. Porter

approximately stable due to regulatory requirements, bankers may try to improve their
ROE by (1) enhancing overhead efficiency (OHE), (2) engaging in more off-balance-
sheet activities (OBS), and (3) using interest-sensitive gap management in an attempt
to decrease their total risk-based capital ratio (Cap) while maintaining an attractive
ROE. The above developments in the banking industry generate potential improve-
ment in performance but also intensify uncertainties and complexities of bank man-
agement. Therefore, a study to investigate the impact of the use of capital on the
riskiness of banks is an indispensible element in bank management.

13.4.2 Instrumental Variable

The statistical summary of our instrumental variable, d, for each year is shown in
Table 13.2. Consistent with the findings documented by Hughes et al. (2001), John
et al. (2000), Keeley (1990), and Kim and Santomero (1988), the use of equity
capital by banks, on average, triggers a loss in efficiency. The dispersion of d is
substantial both cross-sectionally and intertemporally. For our sample, the distri-
bution of d in the same year tends to be skewed to the left-hand side and leptokurtic
(i.e., has positive excess kurtosis). Therefore, we look at nonparametric statistics
and use a normality-free regression model in our analysis to avoid the possible
errors in estimation.

Table 13.2 Distribution of instrumental variables

Mean SD Skewness Kurtosis Max Min
1993 0.038 0.106 0.167 1.936 0.470 0.494
1994 0.040 0.112 0.386 3.394 0.511 0.805
1995 0.036 0.114 0.196 3.791 0.551 0.753
1996 0.041 0.111 0.037 1.852 0.506 0.557
1997 0.223 0.162 2.848 16.357 0.302 1.253
1998 0.032 0.112 0.018 3.332 0.544 0.744
1999 0.053 0.144 0.266 1.274 0.502 0.714
2000 0.054 0.144 0.101 1.358 0.541 0.761
2001 0.049 0.144 0.146 1.286 0.558 0.624
2002 0.042 0.153 0.160 1.707 0.726 0.739
2003 0.021 0.151 0.078 0.573 0.540 0.559
2004 0.021 0.121 0.070 1.425 0.531 0.634
2005 0.024 0.122 0.044 2.229 0.555 0.841
2006 0.016 0.137 0.506 4.024 0.519 1.006
2007 0.015 0.190 0.438 1.536 0.542 0.616
2008 0.012 0.049 2.758 27.015 0.129 0.581
Descriptive statistics of the instrumental variable for capital over sample years are presented.
The instrumental variable d ¼ u  B is a measure of incremental bank inefficiency due to
capital level, where the stochastic frontiers are PTI ¼ a + b1BVA + b2(BVA)2 + e, e ¼ x  B,
and PTI ¼ a + b1BVA + b2(BVA)2 + b3BVC + e, e ¼ v  u
13 Does Banking Capital Reduce Risk? 361

13.4.3 Measures of Bank Risk

We investigate the risks faced by banks from various aspects. Table 13.3 displays
the measures of risk used in this study: credit risk, liquidity risk, interest rate risk,
off-balance-sheet (OBS) risk, market risk, and finally leverage risk. Credit risk is
concerned with the quality of a bank’s assets. Historically this has focused on
a bank’s loan portfolio, but recent events have shown the importance of looking at
all bank assets in light of potential default risk. Liquidity risk measures the ability of
a bank to meet all cash needs at a reasonable cost whenever they arise. Interest rate
risk is the extent to which banks have protected themselves from market-driven
changes in the level of interest rates. Banks have the opportunity to use asset/
liability management tools to mitigate the impact of changes in interest rates on
both bank earnings and bank equity. We also collect data on off-balance-sheet
activities and investigate their relationship with bank capital. Market risk is the risk
of changes in asset prices that are beyond the control of bank management. Finally,
leverage risk is the risk arising from the capital structure decisions of the BHC. The
first five measures of risk relate to the various elements of business risk confronting
bank management. Leverage risk, on the other hand, relates directly to the financial
decisions taken in terms of the amount of capital employed. From another perspec-
tive, it can be said that minimum capital requirements (i.e., maximum leverage
standards) are mandated by regulators to mitigate the various elements of business
risk that the BHC accepts.
Table 13.4 displays the Spearman correlation coefficients between bank size
and our instrumental variable over the sample period. We look at this nonpara-
metric test due to the non-normal distribution of the instrument and variables. We
believe that the generally insignificant correlation between our instrument and the
book value of assets in combination with the generally significant correlation of
our instrument and the book value of equity justifies the use of delta as an
instrument for capital. In addition, Table 13.4 shows that, measured by book
value of equity and pretax income, large BHCs tend to suffer a greater loss in
efficiency than their smaller counterparts at a statistically significant level. On the
other hand, the value of assets does not necessarily demonstrate a negative relation
with bank efficiency. These findings suggest that the inefficiency of BHCs comes
from the use of equity capital but is not directly led by the expansion of business
scale and/or scope. Therefore, a careful investigation of the impact of capital on
banking risks is appropriate.

13.5 Empirical Results

We look at seven different measures of risk: credit risk, liquidity risk, interest rate
risk, off-balance-sheet risk, market risk, composite risk, and leverage risk. While
many of the results support the proposition that increased capital requirements
reduce risk in BHCs, there are some very significant results that suggest the
opposite – that BHCs increase risk as their capital ratios increase.
362 W.-J.P. Chiou and R.L. Porter

Table 13.3 Variables

Symbol Definition
Overall risk
Eq/A Total equity/total asset
RA/A Total risk-based assets/total assets
RC/RA Capital requirement ratio (total risk-based capital/total risk-based assets)
Tier 1/RA Tier 1 capital/total risk-based assets
Tier 2/tier 1 Tier 2 capital/tier 1 capital
Credit risk
NPL/LL Nonperforming assets/total loans and leases
NPL/E Nonperforming assets/total equity capital
Charge-offs/L Net loan charge-offs/total loans and leases
Provision/L Annual provision for loan losses/total loans and leases
Provision/E Annual provision for loan losses/total equity capital
Allowance/L Allowance for loan losses/total loans and leases
Allowance/E Allowance for loan losses/total equity capital
Liquidity risk
STPF/A Short-term purchased funds (Eurodollars, federal funds, security RPs, large CDs,
and commercial paper)/total assets
Cash/A Cash and due from other banks/total assets
HLA/A Cash assets and government securities/total assets
FFS/A (Federal funds sold + reverse RPs – sum of federal funds purchased – RPs)/total
assets
FFP/A (Federal funds purchased + RPs)/total assets
Cash/STPF Cash and due from other banks/short-term purchased funds (Eurodollars, federal
funds, security RPs, large CDs, and commercial papers)
Interest rate risk
Gap Interest-sensitive gap (IS assets – IS liabilities)/total assets
Off-balance-sheet risk
OBS/A Off-balance-sheet assets/total assets
Der/A Credit equivalent amount of off-balance-sheet derivative contracts/total assets
Der/RA Credit equivalent amount of off-balance-sheet derivative contracts/total risk-based
assets
IR Der Notional amount of interest rate derivatives held for trading/notional amount of
interest rate derivatives held for other purposes
FX Der Notional amount of foreign exchange derivatives held for trading/notional amount
of foreign exchange derivatives held for other purposes
Eq Der Notional amount of equity derivatives held for trading/notional amount of equity
derivatives held for other purposes
Cmd Der Notional amount of commodity derivatives held for trading/notional amount of
commodity derivatives held for other purposes
OBS/E Total OBS LC, commitments, credit card lines of credit, and loan/total equity
Der/A Total derivatives/total assets
(continued)
13 Does Banking Capital Reduce Risk? 363

Table 13.3 (continued)

Symbol Definition
Market risk
Trading Trading account assets/total assets
assets
Trading A/L Trading account assets/trading account liabilities
Investment Market value of investment portfolio/book value of investment portfolio
M/B
Performance
PTI/A Pretax income/asset
ROE Return on equity
ROA Return on asset
ATR Average tax rate (taxes/pretax income)
Spread Earning spread (interest income/(loan + investment)(interest expenses/deposits))
OHE Overhead efficiency (noninterest expenses/noninterest income)

Our results are displayed in Tables 13.5, 13.6, 13.7, 13.8, 13.9, and 13.10 and
provide a number of interesting insights. To enhance robustness, we present both
Spearman’s rank correlation coefficient between the tested variable and the instru-
mental variable for capital, d, and the coefficient of d in GMM regressions. To
control for the size of the bank holding companies, each coefficient of d is generated
by GMM regression with a constant and the natural logarithm of the book value of
assets. While the coefficient of the control variable and the constant term are omitted
from the tables, they are available upon request. In Table 13.5, we find a positive
relationship between ratio of total equity to total assets and our instrument for capital
(see Eq/A). The coefficient on our instrument is strictly positive and statistically
significant. This is clearly what we would expect. As leverage decreases, so does
risk; therefore, higher capital should be associated with higher levels of this risk
measure. In other words, it should be a positive relationship, and it is. However, for
the ratio of risky assets to total assets, risk increases as the ratio increases. Therefore,
higher capital should be associated with lower levels of this risk measure (a negative
relationship), and again that is what we find. When we look at just tier 1 capital to
total risky assets, we find the expected positive relationship, and when we look at the
ratio of tier 2 to tier 1 capital, we find the expected negative relationship.
In Table 13.6, we look at some traditional measures of credit risk. As the ratio of
nonperforming loans to total loans increases, so does risk. Therefore, the coefficient
on capital should be negative, and they are with several exceptions over the years. Our
second measure of credit risk is the ratio of nonperforming loans to total equity. Here
again higher levels of the ratio imply higher risk, so we expect to find a negative
relationship and we do, and this time without exception and at high levels of
significance. When we look at the ratio of loan charge-offs to loans outstanding, we
have more exceptions, but in general we find an expected negative relationship.
 364

Table 13.4 Explanatory variables and instrument for capital of BHCs: Spearman’s rank correlation coefficient
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
r(BVA, d) 0.27 0.02 0.01 0.01 0.01 0.07 0.01 0.03 0.03 0.01 0.03 0.25 0.07 0.07 0.00 0.01
p-value 0.001 0.409 0.422 0.429 0.412 0.101 0.413 0.323 0.327 0.450 0.337 0.000 0.145 0.136 0.495 0.467
r(BE, d) 0.51 0.27 0.23 0.32 0.34 0.26 0.32 0.31 0.30 0.33 0.34 0.51 0.34 0.33 0.26 0.20
p-value 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
r(PTI, d) 0.38 0.09 0.04 0.13 0.13 0.04 0.13 0.16 0.20 0.20 0.17 0.35 0.23 0.23 0.16 0.17
p-value 0.000 0.123 0.304 0.005 0.006 0.247 0.007 0.002 0.000 0.000 0.003 0.000 0.000 0.000 0.008 0.002
Spearman’s rank correlation coefficients between book value of assets (BVA), book value of equity (BE), and pretax income (PTI) with the instrumental
variable for capital (d) in each year are reported
W.-J.P. Chiou and R.L. Porter
 13

Table 13.5 Leverage risk

Panel A: Spearman’s rank correlation coefficient between variable and d
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
Eq/A 0.71 0.82 0.83 0.90 0.89 0.83 0.89 0.89 0.93 0.91 0.92 0.72 0.90 0.90 0.91 0.92
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
RA/A 0.10 0.18 0.15 0.21 0.22 0.15 0.20 0.20 0.17 0.12 0.16 0.10 0.11 0.16 0.15 0.20
p-value 0.144 0.012 0.022 0.000 0.000 0.001 0.000 0.000 0.002 0.020 0.004 0.047 0.047 0.008 0.012 0.000
RC/RA 0.41 0.42 0.46 0.43 0.44 0.44 0.44 0.44 0.39 0.42 0.42 0.41 0.42 0.42 0.36 0.40
Does Banking Capital Reduce Risk?

p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Tier 0.38 0.45 0.50 0.44 0.42 0.46 0.44 0.44 0.45 0.44 0.44 0.44 0.44 0.44 0.44 0.44
1/RA
p-value 0.000 0.000 0.000 0.000 0.009 0.003 0.004 0.005 0.004 0.004 0.005 0.004 0.004 0.004 0.004 0.004
Tier 0.23 0.37 0.38 0.33 0.36 0.35 0.35 0.35 0.35 0.35 0.35 0.35 0.35 0.35 0.35 0.35
2/tier 1
p-value 0.006 0.000 0.000 0.002 0.001 0.001 0.006 0.002 0.003 0.004 0.003 0.003 0.003 0.003 0.003 0.003
Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
Eq/A 0.312 0.102 0.144 0.206 0.202 0.161 0.161 0.165 0.174 0.165 0.222 0.098 0.221 0.215 0.202 0.204
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
RA/A 0.035 0.105 0.045 0.199 0.242 0.152 0.155 0.152 0.154 0.114 0.165 0.049 0.066 0.021 0.070 0.118
p-value 0.375 0.000 0.150 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.106 0.116 0.372 0.172 0.225
(continued)
365
 366

Table 13.5 (continued)

Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
RC/RA 0.192 0.068 0.102 0.249 0.246 0.200 0.216 0.240 0.256 0.263 0.381 0.165 0.399 0.259 0.253 0.284
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Tier 0.212 0.080 0.114 0.267 0.266 0.212 0.229 0.250 0.261 0.266 0.380 0.167 0.406 0.421 0.411 0.427
1/RA
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Tier 0.441 0.216 0.229 0.379 0.393 0.276 0.292 0.280 0.222 0.196 0.231 0.055 0.306 0.196 0.255 0.338
2/tier 1
pvalue 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.081 0.000 0.000 0.000 0.000
Spearman’s rank correlation coefficients between each risk measure variable with d, the instrument for capital, are reported for each year in Panel A. In
Panel B, the GMM regression coefficients of bk,t for each dependent variable k in each year t are reported. Specifically, yk,t ¼ ct + bk,tdi,t + gk,t ln(BVA) + k,t,
where yk,t is one of the measures of risk or performance (e.g., total equity/total asset) for bank i in year t; c is a constant; bk,t is the coefficient of instrumental
variable of capital, di,k, for k’s regression in year t; gk,t is the coefficient of natural logarithm of bank’s book value; an k,t is the error term
W.-J.P. Chiou and R.L. Porter
 13

Table 13.6 Credit risk

Panel A: Spearman’s rank correlation coefficient between variable and d
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
NPL/LL 0.02 0.03 0.02 0.02 0.03 0.02 0.01 0.03 0.05 0.05 0.02 0.01 0.04 0.08 0.05 0.07
p-value 0.412 0.340 0.418 0.346 0.285 0.290 0.407 0.291 0.209 0.189 0.381 0.453 0.246 0.118 0.215 0.123
NPL/E 0.23 0.22 0.17 0.18 0.14 0.16 0.21 0.25 0.26 0.26 0.22 0.16 0.23 0.25 0.22 0.27
p-value 0.005 0.003 0.015 0.000 0.002 0.001 0.000 0.000 0.000 0.000 0.000 0.005 0.000 0.000 0.000 0.000
Charge-offs/L 0.01 0.02 0.05 0.03 0.02 0.02 0.06 0.03 0.02 0.00 0.00 0.10 0.01 0.04 0.07 0.08
p-value 0.452 0.415 0.270 0.328 0.341 0.360 0.142 0.290 0.396 0.489 0.498 0.048 0.410 0.282 0.148 0.106
Provision/L 0.01 0.13 0.18 0.11 0.14 0.14 0.08 0.14 0.13 0.15 0.15 0.06 0.15 0.17 0.13 0.07
Does Banking Capital Reduce Risk?

p-value 0.442 0.052 0.010 0.083 0.052 0.052 0.065 0.005 0.010 0.005 0.006 0.160 0.008 0.004 0.022 0.142
Provision/E 0.24 0.33 0.39 0.41 0.38 0.35 0.37 0.41 0.40 0.41 0.38 0.24 0.34 0.33 0.27 0.26
p-value 0.004 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Allowance/L 0.18 0.08 0.08 0.13 0.12 0.13 0.17 0.13 0.15 0.14 0.14 0.17 0.11 0.06 0.05 0.02
p-value 0.021 0.144 0.135 0.045 0.069 0.068 0.001 0.012 0.005 0.009 0.011 0.003 0.038 0.168 0.204 0.350
Allowance/E 0.40 0.54 0.57 0.60 0.57 0.50 0.56 0.57 0.55 0.52 0.50 0.33 0.43 0.44 0.43 0.49
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
NPL/LL 0.074 0.006 0.002 0.002 0.004 0.007 0.001 0.001 0.001 0.001 0.000 0.004 0.004 0.009 0.001 0.003
p-value 0.022 0.017 0.219 0.134 0.007 0.000 0.258 0.402 0.373 0.246 0.415 0.012 0.028 0.003 0.437 0.204
NPL/E 9.510 0.159 0.103 0.088 0.087 0.004 0.122 0.118 0.087 0.088 0.090 0.064 0.116 0.173 0.191 0.195
p-value 0.096 0.000 0.000 0.000 0.000 0.463 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Charge-offs/L 0.007 0.001 0.000 0.001 0.002 0.005 0.000 0.003 0.000 0.000 0.001 0.003 0.002 0.003 0.002 0.004
p-value 0.074 0.138 0.431 0.134 0.037 0.000 0.358 0.142 0.435 0.313 0.199 0.010 0.045 0.044 0.113 0.021
(continued)
367
 368

Table 13.6 (continued)

Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
Provision/L 0.009 0.003 0.003 0.002 0.001 0.001 0.002 0.001 0.001 0.001 0.003 0.004 0.006 0.016 0.013 0.006
p-value 0.074 0.000 0.000 0.008 0.047 0.044 0.024 0.419 0.182 0.089 0.000 0.003 0.008 0.105 0.090 0.029
Provision/E 15.83 0.064 0.055 0.074 0.070 0.037 0.089 0.065 0.061 0.063 0.087 0.051 0.076 0.064 0.062 0.075
p-value 0.096 0.000 0.000 0.000 0.000 0.000 0.000 0.005 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Allowance/L 0.009 0.002 0.002 0.006 0.006 0.005 0.004 0.006 0.004 0.005 0.006 0.001 0.004 0.001 0.008 0.026
p-value 0.013 0.049 0.015 0.000 0.000 0.000 0.000 0.005 0.003 0.000 0.000 0.332 0.043 0.306 0.086 0.145
Allowance/E 6.659 0.118 0.153 0.201 0.201 0.133 0.183 0.163 0.163 0.165 0.202 0.095 0.199 0.237 0.265 0.300
p-value 0.096 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Please see the description of Table 13.5
W.-J.P. Chiou and R.L. Porter
 13

Table 13.7 Liquidity and interest rate risk

Panel A: Spearman’s rank correlation coefficient between variable and d
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
STPF/A 0.05 0.10 0.11 0.30 0.34 0.38 0.38 0.38 0.36 0.34 0.41 0.16 0.31 0.40 0.39 0.45
p-value 0.287 0.104 0.080 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.005 0.000 0.000 0.000 0.000
Cash/A 0.01 0.01 0.04 0.03 0.04 0.06 0.00 0.04 0.00 0.01 0.02 0.06 0.09 0.13 0.01 0.07
p-value 0.440 0.474 0.290 0.252 0.209 0.133 0.487 0.260 0.490 0.446 0.379 0.176 0.073 0.025 0.430 0.142
HLA/A 0.01 0.01 0.02 0.05 0.07 0.09 0.03 0.07 0.08 0.10 0.09 0.04 0.07 0.04 0.12 0.07
p-value 0.476 0.443 0.382 0.168 0.089 0.041 0.274 0.096 0.088 0.042 0.059 0.250 0.130 0.291 0.032 0.142
Does Banking Capital Reduce Risk?

FFS/A 0.16 0.04 0.01 0.05 0.05 0.17 0.06 0.08 0.08 0.08 0.03 0.10 0.01 0.04 0.09 0.00
p-value 0.037 0.294 0.471 0.177 0.148 0.000 0.127 0.081 0.070 0.077 0.287 0.048 0.461 0.279 0.091 0.485
FFP/A 0.04 0.04 0.15 0.32 0.33 0.43 0.39 0.35 0.32 0.29 0.34 0.12 0.65 1.00 0.99 1.00
p-value 0.346 0.289 0.024 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.023 0.000 0.000 0.000 0.000
Cash/STPF 0.04 0.03 0.04 0.10 0.10 0.10 0.13 0.11 0.10 0.09 0.13 0.08 0.20 0.31 0.24 0.31
p-value 0.431 0.434 0.374 0.266 0.241 0.196 0.315 0.240 0.275 0.261 0.288 0.250 0.277 0.266 0.259 0.295
Gap 0.07 0.07 0.08 0.05 0.08 0.03 0.11 0.05 0.06 0.02 0.07 0.03 0.05 0.06 0.09 0.10
p-value 0.207 0.019 0.015 0.014 0.016 0.255 0.021 0.163 0.141 0.352 0.117 0.326 0.203 0.185 0.092 0.054
Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
STPF/A 0.091 0.015 0.038 0.045 0.044 0.024 0.028 0.031 0.040 0.046 0.029 0.019 0.026 0.011 0.021 0.011
p-value 0.004 0.016 0.018 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.002 0.003 0.060 0.016 0.073
Cash/A 0.001 0.001 0.012 0.004 0.005 0.007 0.005 0.006 0.002 0.004 0.005 0.018 0.023 0.031 0.004 0.002
p-value 0.487 0.397 0.046 0.409 0.224 0.131 0.272 0.121 0.405 0.218 0.252 0.015 0.001 0.000 0.300 0.441
(continued)
369
 370

Table 13.7 (continued)

Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
HLA/A 0.011 0.000 0.020 0.014 0.029 0.024 0.020 0.026 0.035 0.039 0.060 0.027 0.069 0.043 0.127 0.002
p-value 0.346 0.484 0.013 0.234 0.002 0.003 0.021 0.001 0.006 0.000 0.000 0.024 0.002 0.031 0.000 0.441
FFS/A 0.029 0.001 0.011 0.018 0.028 0.054 0.020 0.026 0.036 0.045 0.040 0.004 0.021 0.001 0.053 1.324
p-value 0.323 0.461 0.204 0.027 0.002 0.000 0.020 0.009 0.001 0.000 0.006 0.333 0.047 0.464 0.000 0.154
FFP/A 0.105 0.005 0.055 0.016 0.021 0.020 0.017 0.027 0.029 0.024 0.034 0.013 0.000 0.000 0.000 0.000
p-value 0.060 0.353 0.019 0.039 0.011 0.004 0.004 0.007 0.001 0.002 0.003 0.014 0.241 0.430 0.092 0.261
Cash/STPF 172.87 21.99 904.47 276.21 193.34 72.96 58.75 88.84 4.764 39.15 195.89 0.744 5434.9 356.92 2.226 1990.9
p-value 0.112 0.347 0.180 0.007 0.137 0.143 0.035 0.248 0.470 0.197 0.235 0.489 0.157 0.063 0.493 0.088
Gap 0.082 0.072 0.123 0.050 0.110 0.012 0.133 0.071 0.077 0.020 0.080 0.068 0.112 0.116 0.103 0.126
p-value 0.258 0.026 0.004 0.084 0.001 0.340 0.000 0.015 0.037 0.261 0.026 0.012 0.007 0.002 0.009 0.001
Please see the description of Table 13.5
W.-J.P. Chiou and R.L. Porter
 13

Table 13.8 Off-balance-sheet risk

Panel A: Spearman’s rank correlation coefficient between variable and d
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
OBS/A 0.10 0.16 0.19 0.29 0.28 0.26 0.29 0.26 0.28 0.29 0.28 0.18 0.29 0.28 0.30 0.30
p-value 0.328 0.247 0.128 0.070 0.064 0.113 0.084 0.101 0.104 0.153 0.101 0.150 0.120 0.113 0.117 0.116
Der/A 0.93 0.99 0.96 0.96 0.94 0.92 1.00 0.92 0.92 0.90 0.89 0.96 1.02 1.00 1.00 1.00
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Der/RA 0.21 0.20 0.29 0.28 0.26 0.29 0.26 0.28 0.22 0.27 0.27 0.27 0.26 0.26 0.26 0.24
Does Banking Capital Reduce Risk?

p-value 0.319 0.227 0.195 0.140 0.129 0.225 0.168 0.201 0.208 0.306 0.202 0.288 0.240 0.225 0.176 0.175
IR Der 0.14 0.16 0.20 0.25 0.27 0.27 0.28 0.27 0.26 0.26 0.27 0.25 0.25 0.27 na na
p-value 0.256 0.276 0.179 0.119 0.087 0.103 0.115 0.111 0.121 0.149 0.152 0.149 0.158 0.143 na na
FX Der 0.16 0.48 0.56 0.59 0.60 0.60 0.61 0.62 0.60 0.59 0.58 0.59 0.62 0.63 na na
p-value 0.085 0.103 0.093 0.065 0.045 0.039 0.043 0.045 0.046 0.053 0.059 0.060 0.061 0.060 na na
Eq Der 0.71 0.35 0.38 0.42 0.43 0.44 0.44 0.44 0.43 0.42 0.43 0.43 na na na na
p-value 0.000 0.122 0.123 0.103 0.080 0.086 0.094 0.092 0.099 0.119 0.123 0.122 na na na na
Cmd Der 0.89 0.45 0.31 0.31 0.34 0.35 0.36 0.36 0.35 0.35 0.35 0.35 na na na na
p-value 0.000 0.109 0.135 0.110 0.083 0.073 0.079 0.082 0.084 0.095 0.106 0.109 na na na na
OBS/E 0.26 0.43 0.49 0.49 0.48 0.44 0.47 0.48 0.47 0.40 0.42 0.21 0.33 0.30 0.31 0.33
p-value 0.019 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Der/A 0.39 0.54 0.44 0.44 0.45 0.45 0.45 0.46 0.46 0.44 0.42 0.38 0.26 0.12 0.07 0.06
p-value 0.132 0.167 0.145 0.107 0.085 0.105 0.100 0.106 0.111 0.144 0.129 0.146 0.123 0.126 0.130 0.257
(continued)
371
 372

Table 13.8 (continued)

Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
OBS/A 0.002 0.001 0.000 0.001 0.002 0.001 0.001 0.001 0.003 0.003 0.005 0.002 0.005 0.004 0.000 0.000
p-value 0.377 0.092 0.124 0.005 0.014 0.065 0.081 0.144 0.030 0.074 0.088 0.062 0.052 0.065 0.500 0.500
Der/A 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
p-value 0.410 0.080 0.277 0.038 0.050 0.093 0.394 0.427 0.058 0.083 0.069 0.118 0.064 0.500 0.500 0.500
Der/ 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000
RA
p-value 0.474 0.083 0.285 0.054 0.063 0.223 0.338 0.462 0.061 0.097 0.071 0.118 0.075 0.500 0.500 0.500
IR Der 0.002 0.304 0.173 0.104 0.291 0.157 0.121 0.019 4.707 1.127 3.398 1.459 1.383 2.868 na na
p-value 0.397 0.146 0.070 0.006 0.071 0.065 0.084 0.440 0.097 0.228 0.080 0.210 0.334 0.010 na na
FX Der 0.106 5.001 8.842 102.09 85.289 1.508 0.716 0.478 0.964 0.078 1.691 1.829 4.028 9.633 na na
p-value 0.236 0.184 0.149 0.147 0.144 0.006 0.190 0.280 0.292 0.483 0.210 0.321 0.024 0.036 na na
Eq Der 0.003 0.304 0.415 1.004 0.876 0.219 0.804 0.692 2.029 0.903 3.828 2.313 na na na na
p-value 0.412 0.200 0.178 0.271 0.201 0.391 0.215 0.367 0.468 0.259 0.112 0.043 na na na na
Cmd 0.005 0.020 0.035 0.265 0.228 0.117 0.074 0.094 0.171 0.122 0.331 0.004 0.895 0.367 na na
Der
p-value 0.390 0.298 0.403 0.442 0.171 0.081 0.483 0.452 0.408 0.381 0.156 0.476 0.220 0.322 na na
OBS/E 1.910 2.453 4.405 4.672 4.639 2.778 2.912 3.953 3.203 2.939 4.320 1.913 4.142 3.237 3.536 3.104
p-value 0.096 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Der/A 0.406 0.396 0.128 0.382 0.560 0.413 0.378 0.242 0.478 0.493 0.693 0.165 0.724 0.450 0.000 0.000
p-value 0.312 0.095 0.189 0.004 0.060 0.071 0.103 0.163 0.097 0.102 0.089 0.050 0.048 0.057 0.500 0.500
Please see the description of Table 13.5
W.-J.P. Chiou and R.L. Porter
 13

Table 13.9 Market risk

Panel A: Spearman’s rank correlation coefficient between variable and d
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
Trading assets 0.75 0.92 0.96 1.00 0.99 0.97 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.98
p-value 0.000 0.000 0.000 0.000 0.000 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Trading A/L 0.57 0.13 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.11 0.10 0.09 n.a
0.044 0.153 0.207 0.204 0.176 0.163 0.164 0.165 0.170 0.190 0.206 0.214 0.214 0.206 0.206 n.a
Does Banking Capital Reduce Risk?

p-value
Investment M/B 0.20 0.21 0.16 0.15 0.16 0.14 0.14 0.17 0.13 0.16 0.14 0.13 0.16 0.13 0.13 n.a
p-value 0.059 0.070 0.118 0.070 0.059 0.052 0.050 0.050 0.051 0.041 0.057 0.063 0.041 0.038 0.059 n.a
Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
Trading assets 0.028 0.001 0.023 0.007 0.007 0.002 0.005 0.003 0.009 0.006 0.009 0.004 0.014 0.010 0.009 0.003
p-value 0.062 0.399 0.049 0.003 0.020 0.173 0.056 0.104 0.040 0.075 0.067 0.047 0.077 0.077 0.074 0.252
Trading A/L 16.484 0.227 0.970 1.365 1.467 3.105 0.992 0.925 0.752 1.869 0.169 2.241 1.526 1.025 0.366 na
p-value 0.000 0.497 0.388 0.183 0.188 0.148 0.230 0.146 0.285 0.167 0.324 0.147 0.128 0.057 0.391 na
Investment M/B 0.002 0.001 0.006 0.002 0.009 0.007 0.010 0.012 0.026 0.019 0.009 0.001 0.001 0.006 0.010 na
p-value 0.066 0.030 0.106 0.046 0.014 0.000 0.001 0.000 0.035 0.001 0.083 0.042 0.098 0.055 0.009 na
Please see the description of Table 13.5
373
Table 13.10 Performance
374

Panel A: Spearman’s rank correlation coefficient between variable and d

2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
PTI/A 0.12 0.06 0.06 0.22 0.16 0.32 0.26 0.34 0.42 0.39 0.32 0.19 0.39 0.35 0.36 0.44
p-value 0.100 0.215 0.235 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000
ROE 0.20 0.37 0.43 0.40 0.43 0.28 0.32 0.24 0.18 0.28 0.31 0.35 0.26 0.26 0.26 0.19
p-value 0.138 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.001
ROA 0.10 0.07 0.06 0.20 0.17 0.31 0.29 0.36 0.43 0.40 0.36 0.24 0.45 0.42 0.41 0.48
p-value 0.136 0.175 0.214 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
ATR 0.19 0.03 0.00 0.02 0.03 0.06 0.08 0.06 0.06 0.04 0.10 0.11 0.14 0.13 0.06 0.09
p-value 0.018 0.353 0.488 0.312 0.245 0.120 0.079 0.148 0.163 0.236 0.056 0.043 0.012 0.024 0.164 0.064
Spread 0.05 0.19 0.20 0.16 0.17 0.14 0.23 0.21 0.22 0.10 0.17 0.05 0.10 0.10 0.08 0.17
p-value 0.287 0.009 0.005 0.057 0.019 0.058 0.000 0.000 0.000 0.041 0.002 0.226 0.050 0.060 0.124 0.003
OHE 0.15 0.04 0.05 0.08 0.05 0.05 0.00 0.01 0.02 0.23 0.04 0.10 0.01 0.00 0.01 0.00
p-value 0.049 0.295 0.244 0.053 0.167 0.154 0.490 0.439 0.366 0.113 0.264 0.054 0.454 0.477 0.453 0.489
Panel B: GMM regressions
2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993
PTI/A 0.027 0.006 0.009 0.013 0.011 0.010 0.015 0.017 0.017 0.019 0.017 0.015 0.023 0.020 0.020 0.024
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
ROE 0.024 0.061 0.124 0.146 0.136 0.108 0.058 0.037 0.055 0.065 0.098 0.010 0.058 0.063 0.015 0.083
p-value 0.317 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.003 0.000 0.000 0.001 0.000
ROA 0.022 0.004 0.005 0.009 0.008 0.008 0.011 0.013 0.012 0.013 0.013 0.011 0.017 0.016 0.016 0.018
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
ATR 3.887 0.223 0.355 0.194 0.180 0.078 0.442 0.154 0.040 0.001 0.061 0.480 0.049 0.021 0.163 0.103
p-value 0.145 0.007 0.005 0.000 0.001 0.005 0.023 0.117 0.367 0.489 0.197 0.006 0.162 0.382 0.020 0.080
Spread 0.042 0.009 0.030 0.019 0.017 0.021 0.019 0.021 0.019 0.011 0.020 0.006 0.032 0.016 0.117 0.042
p-value 0.059 0.000 0.010 0.000 0.000 0.049 0.000 0.000 0.000 0.000 0.000 0.033 0.019 0.000 0.179 0.001
OHE 2.238 0.628 2.920 1.924 0.209 0.627 0.461 0.490 0.809 1.220 1.617 0.275 10.730 3.912 1.622 0.478
p-value 0.321 0.281 0.157 0.009 0.341 0.068 0.120 0.238 0.188 0.246 0.036 0.321 0.139 0.085 0.291 0.290
Please see the description of Table 13.5
W.-J.P. Chiou and R.L. Porter
13 Does Banking Capital Reduce Risk? 375

The ratio of the provision for loan loss to total loans is an ambiguous measure.
A high provision could indicate bank management is expecting high loan losses. On
the other hand, a high ratio could indicate conservative bank management that is taking
no chances on an underfunded allowance for loan loss. When we look at the provision
as a percentage of total equity, we again have the same ambiguous results possible. In
general, we find that both of these risk measures produce a negative coefficient on our
measure of capital. The ambiguity seems to be resolved in that as the provision for loan
losses increases, so does risk. The alternative explanation is that risk should decrease
with this ratio, but higher capital levels produce counterintuitive results. The ratio of
the allowance for loan loss to total loans moves inversely with capital while the ratio
of allowance for loan loss to total equity moves in the opposite direction.
We turn now to the allowance for loan losses. Like the provision for loan loss
ratios, we have the same ambiguous expectations, but now we find conflicting
results. There is no clear expectation for the impact of this ratio on risk. In other
words, both the ratio of the allowance for loan losses to total loans and the ratio of
the allowance for loan losses to total equity can be reflecting either high risk or low
risk. What we find is that the allowance for loan loss as a percentage of loans
produces a positive sign for the coefficient on capital, while the allowance for loan
loss as a percentage of equity produces a negative sign on the coefficient.
We find even more consistent counterintuitive results when we look at the
relationship between capital and liquidity risk in Table 13.7. Lower capital ratios are
generally related to higher levels of short-term purchased liabilities (see STPF/A
and FFP/A). Since short-term purchased money is more volatile than core deposits,
for example, we would expect high levels of purchased money to be associated with
high levels of capital, yet this is not what we find. On the other hand, we do find that
higher capital ratios are related to more liquid assets (HLA/A) and better coverage
of short-term liabilities (FFS/A). Since both of these ratios imply higher levels of
liquidity, we expected them to be related to lower levels of capital. Apparently
liquidity risk is not reflected in a BHC’s capital level.
When we look at a BHC’s exposure to interest rate risk, we again find counter-
intuitive results. As noted above, Table 13.7 shows that low capital ratios are related
to high levels of short-term purchased funds. This can result in a fundamental
liquidity problem if some markets for short-term borrowing completely dry up as
we have observed in the recent financial crisis. Our more direct measure of interest
rate risk is the interest-sensitive gap (gap) which we define as interest-sensitive
assets minus interest-sensitive liabilities divided by total assets. Here we find
ambiguous results. It is obvious that wider gaps expose banks to more risk if
interest rates move against the bank. However, wide gaps can be held in both
a positive and negative direction. A high positive gap indicates a BHC has a large
amount of interest-sensitive assets in relation to interest-sensitive liabilities and will
be hurt by falling interest rates. A high negative gap indicates a BHC has a large
amount of interest-sensitive liabilities in relation to interest-sensitive assets and will
be hurt if interest rates rise. Our results in Table 13.7 indicate that wider gaps are
associated with lower levels of capital, but this is only a true measure if BHCs
typically held a positive gap.
376 W.-J.P. Chiou and R.L. Porter

In Table 13.8, we turn our attention to BHCs’ off-balance-sheet activity. Rather

surprisingly, off-balance-sheet exposures seem to be inversely related to capital levels
in spite of the Basle Capital Accords. Recall that the Basle Accords require the
maintenance of capital in support of off-balance-sheet activity. Yet all of our measures
of off-balance-sheet risk are associated with low capital levels with one exception. The
notional amount of commodity derivatives held for trading compared with the
notional amount of commodity derivatives held for other purposes is associated
with a higher level of capital. Our interpretation of the ratios that measure the amount
of derivatives held for trading compared with the derivatives held for other purposes is
that the derivatives not held for trading are held to hedge an existing position on the
books of the BHC. As a result, a high ratio implies more trading activity in relation to
hedging activity, and therefore, more capital should be required. However, we again
find high OBS ratios associated with low levels of capital.
In Table 13.9, we look at two measures of market risk: the size of the BHCs’
trading account and the amount of unrealized gains or losses on the BHCs’
investment portfolio. We again find what we believe are counterintuitive results.
First, larger trading portfolios inherently contain a larger amount of market risk. On
the other hand, a large amount of unrealized gains in the investment portfolio
mitigates market risk, at least to some extent. We find, however, low levels of
capital associated with higher trading portfolios, while high levels of capital are
associated with higher unrealized gains in the investment portfolio.
Our results concerning performance measures are shown in Table 13.10. Here
we find evidence of higher capital ratios being associated with higher return on
assets ratios and with higher net interest spreads. Since higher returns on earning
assets are logically associated with higher risk, it is appropriate that higher capital is
used in support of the additional risk. However, while the direction of the causality
is not clear, this could be interpreted as more evidence that BHCs increase risk to
maintain a target return on equity in the face of higher capital requirements.

13.6 Conclusions

In this study, we thoroughly analyze a large cross section of bank holding company
data from 1993 to 2008 to determine the relationship between capital and bank risk-
taking. Our sample includes a minimum of almost 700 BHCs in 2008 and
a maximum of about 1,500 BHCs in 1993. This produces nearly 25,000
company-year observations of BHCs starting with the year that risk-based capital
requirements were first in place. Our data cover a period containing significant
changes in the banking industry and varying levels of economic activity. The
Riegle-Neal and Gramm-Leach-Bliley acts were passed during this time period,
and the tech-stock and housing bubbles both burst with attendant recessions. By
including a larger size range of BHCs in our analysis over a long sample period, our
results are applicable to relatively small BHCs as well as to the largest 200 or so
BHCs traditionally included in empirical studies.
13 Does Banking Capital Reduce Risk? 377

We employ stochastic frontier analysis to create a new type of instrumental

variable for capital to be used in regressions of risk and capital, thereby mitigating
the obvious endogeneity problem. The instruments are validated to confirm their
high correlation with capital and limited correlation with risk. We conclude that
they are suitable for use in our models and employ a GMM estimator to acknowl-
edge the non-normal distribution of the instruments.
Our results are consistent with the theory that BHCs respond to higher capital
ratios by increasing the risk in the earning asset portfolios. We find an inverse
relationship between the proportion of risky assets held by a bank holding company
and the amount of capital they hold. We also find lower levels of capital associated
with measures of credit risk that indicate a riskier loan portfolio. For example, the
amount of nonperforming assets held by the bank holding company is inversely
related to the bank holding company’s capital.
Our findings also demonstrate a counterintuitive relationship between bank
capital and liquidity risk. Less liquid banks tend to have low capital ratios, while
more liquid banks tend to have high capital ratios. These same results suggest that
a higher level of interest rate risk is also related to lower levels of capital. Our
direct measure of the mismatch between interest-sensitive assets and liabilities
provides additional evidence, although somewhat ambiguously, of higher interest
rate risk being associated with lower capital. High levels of off-balance-sheet
activity and of market risk exposure are likewise surprisingly related to low capital
levels. Finally, we note the association of high levels of capital with high return on
asset ratios. This association at least suggests that bank holding companies do
increase the risk of their earning assets in order to provide an adequate return on
their capital.
Our analysis adds to the existing literature with three contributions. First, we
employ a large panel data set to consider the capital-risk relationship for a wider
range of bank holding companies than previously reviewed. Second, stochastic
frontier analysis is applied to exogenously generate the effect of the use of capital in
banking. Finally, our results provide what we believe are important findings with
potentially major public policy implications. If the primary tool used by bank
regulators to ensure a stable financial system is, instead, creating perverse results,
then alternative tools must be developed. Further exploring the relationship
between the efficiency of capital and the risk strategy adopted by a bank would
be a contribution to this literature.

Appendix 1: Stochastic Frontier Analysis

Stochastic frontier analysis (SFA) is an economic modeling method that is intro-

duced by Jondrow et al. (1982). The frontier without random component can be
written as the following general form:

PTI i ¼ TEi  f ðxi ; bÞ, (13.5)

378 W.-J.P. Chiou and R.L. Porter

where PTIi is the pretax income of the bank i, i ¼ 1,..N, TEi denotes the technical
efficiency defined as the ratio of observed output to maximum feasible output, xi is
a vector of J inputs used by the bank i, f(xi, b) is the frontier, and b is a vector of
technology parameters to be estimated. Since the frontier provides an estimate of
the maximum feasible output, we then can measure the shortfall of the observed
output from the maximum feasible output. Considering a stochastic component that
describes random shocks affecting the production process in this model, the sto-
chastic frontier becomes

PTI i ¼ evi  TEi  f ðxi ; bÞ: (13.6)

The shock, ℯvi , is not directly attributable to the bank or the technology but may
come from random white noises in the economy, which is considered as a two-sided
Gaussian distributed variable.
We further describe TEi as a stochastic variable with a specific distribution
function. Specifically,

TEi ¼ eui , (13.7)

where ui is the nonnegative technical inefficiency component, since it is required

that TEi  1. Thus, we obtain the following equation:

PTI i ¼ evi ui  f ðxi ; bÞ: (13.8)

We then can describe the frontier according to a specific production model.

In our case, we assume that bank’s profitability can be specified as the log-linear
Cobb-Douglas function:

X
H
PTI i ¼ a þ bh lnxi, h þ vi  ui : (13.9)
h¼1

Because both vi and ui constitute a compound error term with a specific distri-
bution to be determined, hence the SFA is often referred as composed error model.
In our study, we use the above stochastic frontier with different inputs to
generate the net effect of bank capital without mixing the impact of risk. The
unrestricted model (without including bank equity) is

PTI ðBVA; sBANK Þ ¼ a þ b1 BVA þ b2 ðBVAÞ2 þ e

e ¼ x  B

 , (13.10)
x  iid N 0; s2x , B  iid N 0; s2B

where BVA is the natural logarithm of book value of assets, x is statistical noise, B is
systematic shortfall (under management control), and B  0. Our restricted model is
as follows:
13 Does Banking Capital Reduce Risk? 379

PTI ðBVA; BVC; sBANK Þ ¼ a þ b1 BVA þ b2 ðBVAÞ2 þ b3 BVC þ e:

e ¼ v  u   , (13.11)
v  iid N 0; s2v u  iid N 0; s2u

where BVC is the natural logarithm of book value of capital, v is statistical noise,
and u denotes the inefficiency of a bank considering its use of both assets and
capital. The difference in the inefficiency between the restricted and unrestricted
model,

d ¼ u  B, (13.12)

is our instrumental variable. The instrumental variable for capital can be used in
regressions of various measures of risk, as the dependent variable, on our instru-
ment for capital, as the independent variable, while controlling for BHC size.

Appendix 2: Generalized Method of Moments

Hansen (1982) develops generalized method of moments (GMM) to estimate

parameters that its full shape of the distribution function is not known. The method
requires that a certain number of moment conditions were specified for the model.
These moment conditions are functions of the model parameters and the data, such
that their expectation is zero at the true values of the parameters. The GMM method
then minimizes a certain norm of the sample averages of the moment conditions.
Suppose the error term «t ¼ «(xt, y) is a (T  1) vector that contains
T observations of the error term «t, where xt includes the data relevant for the
model and y is a vector of Nb coefficients. Assume there are NH instrumental
variables in an (NH  1) column vector, ht and T observations of this vector form
a (T  NH) matrix H. We define

f t ðyÞ  ht «ðxt ; yÞ: (13.13)

The notation denotes the Kronecker product of the two vectors. Therefore,
ft (y) is a vector containing the cross product of each instrument in h with each
element of «. The expected value of this cross product is a vector with NeNΗ
elements of zeros at the parameter vector:

E½f t ðy0 Þ
¼ 0: (13.14)

Since we do not observe the true expected values of f, thus we must work instead
with the sample mean of f,

X
T X
T
0
gt ðyÞ  T 1 f t ðyÞ ¼ T 1 ht «t ðyÞ ¼ T 1 H «t ðyÞ: (13.15)
t¼1 t¼1
380 W.-J.P. Chiou and R.L. Porter

We can minimize the quadratic form

0
QT ðyÞ  gT ðyÞ WT gT ðyÞ, (13.16)

where WT is an (NH  NH) symmetric, positive definite weighting matrix. We then

find the first-order condition is

0

DT y^T WT gT y^T ¼ 0, (13.17)

where DT (yT) is a matrix of partial derivatives defined by

0
DT ðyT Þ ¼ ∂gT ðyT Þ=∂y :

Note the above problem is nonlinear; thus, the optimization must be solved
numerically.
Applying the asymptotic distribution theory, the coefficient estimate y^T is

pffiffiffi
 d
T y^T  y0 ! N ð0; OÞ, (13.18)

where O ¼ (D00 WD0)1 D00 WSWD0 (D00 WD0)1. D0 is a generalization of MHX in

those equations and is defined by D0  E[∂f(xt, y0)/∂y0]. S is defined as
" #
X
T h i
S  lim Var T 1=2
f t ðy0 Þ ¼ lim Var T 1=2 gT ðy0 Þ : (13.19)
T!1 T!1
t¼1

The GMM estimators are known to be consistent, asymptotically normal, and

efficient in the class of all estimators that do not use any extra information aside
from that contained in the moment conditions. For more discussion of the execution
of the GMM, please refer to Campbell et al. (1997) and Hamilton (1994).

References
Berger, A. N., Herring, R. J., & Szego, G. P. (1995). The role of capital in financial institutions.
Journal of Banking and Finance, 24, 1383–1398.
Berger, A. N., DeYoung, R., Flannery, M. J., Lee, D., & Oztekin, O. (2008). How do large
banking organizations manage their capital ratios? Journal of Financial Services Research,
343, 123–149.
13 Does Banking Capital Reduce Risk? 381

Buser, S. A., Chen, A. H., & Kane, E. J. (1981). Federal deposit insurance, regulatory policy, and
optimal bank capital. Journal of Finance, 36, 51–60.
Calomiris, C. W. (1999). Building an incentive-compatible safety net. Journal of Banking and
Finance, 23, 1499–1519.
Calomiris, C. W., & Kahn, C. M. (1991). The role of demandable debt in structuring optimal
banking arrangements. American Economic Review, 81, 497–513.
Campbell, J., Lo, A., & MacKinlay, A. C. (1997). The econometrics of financial markets.
Princeton: Princeton University Press.
Diamond, D. W., & Dybvig, P. H. (1983). Bank runs, deposit insurance, and liquidity. Journal of
Political Economy, 91, 401–419.
Diamond, D. W., & Rajan, R. (2000). A theory of bank capital. Journal of Finance, 55,
2431–2465.
Duan, J., Moreau, A. F., & Sealey, C. W. (1992). Fixed rate deposit insurance and risk-shifting
behavior at commercial banks. Journal of Banking and Finance, 16, 715–742.
Esty, B. C. (1998). The impact of contingent liability on commercial bank risk taking. Journal of
Financial Economics, 47, 189–218.
Flannery, M. J., & Rangan, K. P. (2008). What caused the bank capital build-up of the 1990s?
Review of Finance, 12, 391–430.
Gatev, E., Schuermann, T., & Strahan, P. E. (2009). Managing bank liquidity risk: How
deposit-loan synergies vary with market conditions. Review of Financial Studies, 22,
995–1020.
Gorton, G., & Pennacchi, G. (1992). Financial innovation and the provision of liquidity services.
In J. R. Barth & R. Dan Brumbaugh (Eds.), Reform of federal deposit insurance. New York:
Harper Collins.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Hansen, L. P. (1982). Large sample properties of Generalized Method of Moments estimators,
Econometrica, 50, 1029–1054.
Hovakimian, A., & Kane, E. J. (2000). Effectiveness of capital regulation at U.S. commercial
banks: 1985 to 1994. Journal of Finance, 55, 451–468.
Hughes, J. P., Mester, L. J., & Moon, C. (2001). Are scale economies in banking elusive or
illusive? Incorporating capital structure and risk into models of bank production. Journal of
Banking and Finance, 25, 2169–2208.
Hughes, J. P., Lang, W. W., Mester, L. J., Moon, C., & Pagano, M. (2003). Do bankers sacrifice
value to build empires? Managerial incentives, industry consolidation, and financial perfor-
mance. Journal of Banking and Finance, 27, 417–447.
John, K., Saunders, A., & Senbet, L. (2000). A theory of bank regulation and management
compensation. Review of Financial Studies, 13, 95–125.
Jondrow, J., Lovell, C. A., Materov, I. S., & Schmidt, P. (1982). On the estimation of technical
inefficiency in the stochastic frontier production function model. Journal of Econometrics, 19,
233–238.
Keeley, M. C. (1990). Deposit insurance, risk, and market power in banking. American Economic
Review, 80, 1183–1200.
Keeley, M. C. (1992). Bank capital regulation in the 1980s: Effective or ineffective? In
S. Anthony, G. F. Udell, & L. J. White (Eds.), Bank management and regulation: A book of
readings. Mountain View: Mayfield Publishing Company.
Kim, D., & Santomero, A. M. (1988). Risk in banking and capital regulation. Journal of Finance,
42, 1219–1233.
Laeven, L. A., & Levine, R. (2009). Bank governance, regulation and risk taking. Journal of
Financial Economics, 93, 259–275.
Marcus, A. J. (1984). Deregulation and bank financial policy. Journal of Banking and Finance, 8,
557–565.
382 W.-J.P. Chiou and R.L. Porter

Marcus, A. J., & Shaked, I. (1984). The valuation of FDIC deposit insurance using option-pricing
estimates. Journal of Money, Credit and Banking, 16, 446–460.
Merton, R. C. (1977). An analytic derivation of the cost of deposit insurance and loan guarantees.
Journal of Banking and Finance, 1, 3–11.
Prescott, E. S. (1997). The pre-commitment approach in a model of regulatory banking capital.
FRB of Richmond Economic Quarterly, 83, 23–50.
Santos, J. A. (2001). Bank capital regulation in contemporary banking theory: A review of the
literature. Financial Markets, Institutions and Instruments, 10, 41–84.
Shrieves, E., & Dahl, D. (1992). The relationship between risk and capital in commercial banks.
Journal of Banking and Finance, 16, 439–457.
VanHoose, D. (2007). Theories of bank behavior under capital regulation. Journal of Banking and
Finance, 31, 3680–3697.
Evaluating Long-Horizon Event Study
Methodology 14
James S. Ang and Shaojun Zhang

Contents
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
14.2 Fundamental Issues in Long-Horizon Event Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
14.2.1 The Buy-and-Hold Benchmark Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
14.2.2 The Calendar-Time Portfolio Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
14.3 A Review of Simulation Studies on Long-Horizon Event Study Methodology . . . . . . . 390
14.4 A Simulation Study of Large-Size Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
14.4.1 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
14.4.2 Simulation Results for the Buy-and-Hold Benchmark Approach . . . . . . . . . . . . 396
14.4.3 Simulation Results for the Calendar-Time Portfolio Approach . . . . . . . . . . . . . . 402
14.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409

Abstract
We describe the fundamental issues that long-horizon event studies face in
choosing the proper research methodology and summarize findings from
existing simulation studies about the performance of commonly used methods.
We document in details how to implement a simulation study and report our own
findings on large-size samples. The findings have important implications for
future research.
We examine the performance of more than 20 different testing procedures
that fall into two categories. First, the buy-and-hold benchmark approach uses

J.S. Ang (*)

Department of Finance, College of Business, Florida State University, Tallahassee, FL, USA
e-mail: jang@cob.fsu.edu; jang@garnet.acns.fsu.edu
S. Zhang
School of Accounting and Finance, Faculty of Business, Hong Kong Polytechnic University,
Hung Hom, Kowloon, Hong Kong
e-mail: afszhang@inet.polyu.edu.hk

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 383
DOI 10.1007/978-1-4614-7750-1_14,
# Springer Science+Business Media New York 2015
384 J.S. Ang and S. Zhang

a benchmark to measure the abnormal buy-and-hold return for every event firm
and tests the null hypothesis that the average abnormal return is zero. Second, the
calendar-time portfolio approach forms a portfolio in each calendar month
consisting of firms that have had an event within a certain time period prior to
the month and tests the null hypothesis that the intercept is zero in the regression
of monthly portfolio returns against the factors in an asset-pricing model. We
find that using the sign test and the single most correlated firm being the
benchmark provides the best overall performance for various sample sizes and
long horizons. In addition, the Fama-French three-factor model performs better
in our simulation study than the four-factor model, as the latter leads to serious
over-rejection of the null hypothesis.
We evaluate the performance of bootstrapped Johnson’s skewness-adjusted
t-test. This computation-intensive procedure is considered because the distribution
of long-horizon abnormal returns tends to be highly skewed to the right. The
bootstrapping method uses repeated random sampling to measure the significance
of relevant test statistics. Due to the nature of random sampling, the resultant
measurement of significance varies each time such a procedure is used. We also
evaluate simple nonparametric tests, such as the Wilcoxon signed-rank test or the
Fisher’s sign test, which are free from random sampling variation.

Keywords
Long-horizon event study • Johnson’s skewness-adjusted t-test • Weighted
least squares regression • Bootstrap test • Calendar-time portfolio approach •
Fama-French three-factor model • Johnson’s skewness-adjusted t-statistic •
Large-scale simulations

14.1 Introduction

A large number of papers in finance literature have documented evidence that firms
earn abnormal returns over a long time period (ranging from 1 to 5 years) after
certain corporate events. Kothari and Warner (2007) report that a total of 565 papers
reporting event study results were published between 1974 and 2000 in five leading
journals: the Journal of Business (JB), Journal of Finance (JF), Journal of Finan-
cial Economics (JFE), Journal of Financial and Quantitative Analysis (JFQA), and
the Review of Financial Studies (RFS). Approximately 200 of the 565 event studies
use a maximum window length of 12 months or more.
The evidence of long-horizon abnormal returns contradicts the efficient market
hypothesis that stock prices adjust to information fully within a narrow time
window (a few days). To reconcile the contradiction, Fama (1998) argues that
“Most important, consistent with the market efficiency prediction that apparent
anomalies can be due to methodology, most long-term return anomalies tend to
disappear with reasonable changes in technique.” Several simulation studies such as
Kothari and Warner (1997) and Barber and Lyon (1997) document evidence that
statistical inference in long-horizon event studies is sensitive to the choice of
14 Evaluating Long-Horizon Event Study Methodology 385

methodology. Therefore, it is crucial to gain an understanding of the properties and

limitations of the available approaches before choosing a methodology for a long-
horizon event study.
At the core of a long-horizon event study lie two tasks: the first is to measure the
event-related long-horizon abnormal returns, and the second is to test the null
hypothesis that the distribution of these long-horizon abnormal returns concentrates
around zero. A proper testing procedure for long-horizon event studies has to do
both tasks well. Otherwise, two types of error could arise and lead to incorrect
inference. The first error occurs when the null hypothesis is rejected, not because
the event has generated true abnormal returns, but because a biased benchmark has
been used to measure abnormal returns. A biased benchmark shifts the concentra-
tion of abnormal returns away from zero and leads to too many false rejections of
the null hypothesis. The second error occurs when the null hypothesis is accepted,
not because the event has no impact, but because the test itself does not have
enough power to statistically discriminate the mean abnormal return from zero.
A test with low power is undesirable, as it will lead researchers to reach the
incorrect inference that long-term effect is statistically insignificant. Thus, the
researchers would want a procedure that minimizes both sources of error or at
least choose a balance between them.
Two approaches have been followed in recent finance literature to measure and
test long-term abnormal returns. The first approach uses a benchmark to measure the
abnormal buy-and-hold return for every event firm in a sample and tests whether the
abnormal returns have a zero mean. The second approach forms a portfolio in each
calendar month consisting of firms that have had an event within a certain time period
prior to the month and tests the null hypothesis that the intercept is zero in the
regression of monthly calendar-time portfolio returns against the factors in an asset-
pricing model. To follow either approach, researchers need to make a few choices as
illustrated in Fig. 14.1. For the calendar-time portfolio approach, researchers choose
an asset-pricing model and an estimation technique to fit the model. Among the most
popular asset-pricing models are Fama and French’s (1993) three-factor model and
its four-factor extension proposed by Carhart (1997) that includes an additional
momentum-related factor. Two techniques are commonly used to fit the pricing
model: the ordinary least squares (OLS) technique and the weighted least squares
(WLS) technique. On the other hand, if adopting the buy-and-hold benchmark
approach, researchers choose either a reference portfolio or a single control firm as
the benchmark for measuring abnormal returns and select either parametric or
nonparametric statistic for testing the null hypothesis of zero abnormal return.
Permutations of these choices under both approaches generate a large number of
possible testing procedures that can be used in a long-horizon event study. It is
neither practical nor sensible to implement all the testing procedures in an empirical
study of a financial event. Therefore, it would be very useful to provide guidance on
the strength and weakness of the procedures based on simulation results. Simulation
study generates large number of repetitions under various circumstances for
each testing procedure, which allows the tabulations of these two types of
error for comparison.
386 J.S. Ang and S. Zhang

To Test the Significance of Long-Horizon Abnormal Returns

Buy-and-hold
Calendar-Time
Benchmark
Portfolio Approach
Approach

Asset Pricing Reference Portfolio Single Firm

Model Benchmarks, e.g., SZBM, Benchmarks, e.g.,
SZBMBT, MC10 SZBMI, MCI

Model Estimation
Specification, e.g., Technique, e.g., Parametric Nonparametric Parametric Nonparametric
Fama and French’s Tests, e.g., Tests, e.g., Tests, e.g., Tests, e.g.,
Three-Factor Model, OLS, t-test, Bootstrapped test, t-test, Bootstrapped test
Carhart’s Four- WLS Johnson’s test Sign test Johnson’s test Sign test
Factor Model

Fig 14.1 Overview of the two approaches to choose a methodology for long-horizon event study

We organize this chapter as follows. Section 14.2 discusses the fundamental

issues in long-horizon event studies that have been documented in the literature.
Section 14.3 reviews existing simulation studies. Section 14.4 reports results from
a simulation study of large-size samples. Section 14.5 contains some suggestions
for future research.

14.2 Fundamental Issues in Long-Horizon Event Studies

14.2.1 The Buy-and-Hold Benchmark Approach

The long-term buy-and-hold abnormal return of firm i, denoted as ARi, is calculated as

ARi ¼ Ri  BRi , (14.1)

where Ri is the long-term buy-and-hold return of firm i and BRi is the long-term
return on a particular benchmark of firm i. The buy-and-hold return of firm i over t
months is obtained by compounding monthly returns, that is,
Yt
Ri ¼ t¼1
ð1 þ rit Þ  1, (14.2)
where rit is firm i’s return in month t. Calculation of the benchmark return BRi is
given below. The benchmark return, BRi, estimates the return that an event firm
would have had if the event had not happened.
Several articles clearly show that long-term abnormal returns are very sensitive to
choice of benchmarks; see, e.g., Ikenberry et al. (1995), Kothari and Warner (1997),
Barber and Lyon (1997), and Lyon et al. (1999). If wrong benchmarks were used in
measuring long-term abnormal returns, inference on the significance of a certain event
14 Evaluating Long-Horizon Event Study Methodology 387

would be erroneous. Most existing studies use either a single matched firm or a matched
reference portfolio as the benchmark. Barber and Lyon (1997) point out that the control
firm approach eliminates the new listing bias, the rebalancing bias, and the skewness
problem. It also yields well-specified test statistics in virtually all the situations they
consider. Further, Lyon et al. (1999) advocate a reference portfolio of firms that match
on size and BE/ME. The issue on choice of the benchmark is practically unresolved.
Ang and Zhang (2004) additionally argue that the control firm method overcomes
another important problem that is associated with the event firm not being representative
in important aspects of the respective matched portfolio in the reference portfolio
approach. This leads to the matched portfolio return generating a biased estimate of
expected firm return. This problem is particularly severe with small firms.
A common practice in computing an event firm’s long-term abnormal return is to
utilize a benchmark that matches the event firm on size and BE/ME. The practice is
often justified by quoting the findings in Fama and French (1992) that size and
BE/ME combine to capture the cross-sectional variation in average monthly stock
returns and that market beta has no additional power in explaining cross-sectional
return differences. However, in a separate paper, Fama and French (1993)
demonstrate that expected monthly stock returns are related to three factors:
a market factor, a size-related factor, and a book-to-market equity ratio
(BE/ME)-related factor. To resolve this issue, Ang and Zhang (2004) show that
matching based on beta in addition to size and BE/ME does not improve the
performance of the approach.
A recent trend is to use computation-intensive bootstrapping-based tests, such as
the bootstrapped Johnson’s skewness-adjusted t-statistic (e.g., Sutton 1993 and Lyon
et al. 1999) and the simulated empirical p-values (e.g., Brock et al. 1992 and
Ikenberry et al. 1995). These procedures rely on repeated random sampling to
measure the significance of relevant test statistics. Due to the nature of random
sampling, the resultant measurement of significance varies every time such
a procedure is used. As a consequence, different researchers could reach contradic-
tory conclusions using the same procedure on the same sample of event firms. In
contrast, simple nonparametric tests, such as the Wilcoxon signed-rank test or the
Fisher’s sign test, are free from random sampling variation. Barber and Lyon (1997)
examined the performance of the Wilcoxon signed-rank test in a large-scale simula-
tion study. They show that the performance depends on choice of the benchmark. The
signed-rank test is well specified when the benchmark is a single size and BE/ME
matched firm and misspecified when the benchmark is a size and BE/ME matched
reference portfolio. However, Barber and Lyon (1997) present only simulation
results for 1-year horizon. No simulation study in the finance literature has examined
the performance of these simple nonparametric tests for 3- or 5-year horizons, which
are the common holding periods in long-horizon event studies.1

1
The sign test has an advantage over the signed-rank test in that it does not require a symmetric
underlying distribution, while the signed-rank test does.
388 J.S. Ang and S. Zhang

Power is an important consideration in statistical hypothesis testing. Lyon

et al. (1999) report that bootstrapping-based tests are more powerful than Student’s
t-test in testing 1-year abnormal returns in a large-scale simulation study. However,
they do not report evidence on the power of these tests for the longer 3- or 5-year
horizon. In statistics literature, bootstrapping is primarily for challenging situations
when the sampling distribution of the test statistic is either indeterminate or difficult
to obtain and that bootstrapping is less powerful in hypothesis testing than other
parametric or simple nonparametric methods when both bootstrapping and other
methods are applicable (see, e.g., Efron and Tibshirani 1993, Chap. 16 and Davison
and Hinkley 1997, Chap. 4). In a recent study on 5-year buy-and-hold abnormal
returns to holders of the seasoned equity offerings, Eckbo et al. (2000) note that
bootstrapping gives lower significance level relative to the Student’s t-test.
Ang and Zhang (2004) find that most testing procedures have very low power for
samples of medium size over long event horizons (3 or 5 years). This raises concern
about how to interpret long-horizon event studies that fail to reject the null
hypothesis. Failure to reject is often interpreted as evidence that supports the null
hypothesis. However, when power of the test is low, such interpretation may no
longer be warranted. This problem gets even worse when event firms are primarily
small firms. They observe that all tests, except the sign test, have much lower power
for samples of small firms.
More recently, Schultz (2003) argue via simulation that the long-run IPO
underperformance could be related to the endogeneity of the number of new issues.
Firms choose to go IPO at the time when they expect to obtain high valuation in the
stock market. Therefore, IPOs cluster after periods of high abnormal returns on new
issues. In such a case, even if the ex ante returns on IPO are normal, the ex post
measures of abnormal returns may be negative on average. Schultz suggests using
calendar-time returns to overcome the bias. However, Dahlquist and de Jong (2008)
find that it is unlikely that the endogeneity of the number of new issues explains the
long-run underperformance of IPOs. Viswanathan and Wei (2008) present
a theoretical analysis on event abnormal returns when returns predict events.
They show that, when the sample size is fixed, the expected abnormal return is
negative and becomes more negative as the holding period increases. This implies
that there is a small-sample bias in the use of long-run event returns. Asymptoti-
cally, abnormal returns converge to zero provided that the process of the number of
events is stationary. Nonstationarity in the process of the number of events is
needed to generate a large negative bias.
The issues discussed above are associated with the buy-and-hold approach to
testing long-term abnormal returns.2 In addition, this approach suffers from the
cross-correlation problem and the bad model problem (Fama 1998; Brav 1999;
Mitchell and Stafford 2000). The cross-correlation problem arises because matching
on firm-specific characteristics fails to completely remove the correlation between

2
Variations of this approach have been used extensively; see, e.g., Ritter (1991); Ikenberry
et al. (1995); Ikenberry et al. (1996); and Desai and Jain (1997), among many others.
14 Evaluating Long-Horizon Event Study Methodology 389

event firms’ returns. The bad model problem arises because no benchmark gives
perfect estimate of the counterfactual (i.e., what if there was no event) return of an
event firm and benchmark errors are multiplied in computing long-term buy-and-
hold returns. Therefore, Fama (1998) advocates a calendar-time portfolio approach.3

14.2.2 The Calendar-Time Portfolio Approach

In the calendar-time portfolio approach, for each calendar month, an event portfolio
is formed, consisting of all firms that have experienced the same event within the t
months prior to the given month. Monthly return of the event portfolio is computed
as the equally weighted average of monthly returns of all firms in the portfolio.
Excess returns of the event portfolio are regressed on the Fama-French three factors
as in the following model:
 
Rpt  Rft ¼ a þ b Rmt  Rft þ sSMBt þ hHMLt þ et , (14.3)

where Rpt is the event portfolio’s return in month t; Rft is the 1-month Treasury bill rate,
observed at the beginning of the month; Rmt is the monthly market return; SMBt is the
monthly return on the zero investment portfolio for the common size factor in stock
returns; and HMLt is the monthly return on the zero investment portfolio for the common
book-to-market equity factor in stock returns.4 Under the assumption that the Fama-
French three-factor model provides a complete description of expected stock returns, the
intercept, a, measures the average monthly abnormal return on the portfolio of event
firms and should be equal to zero under the null hypothesis of no abnormal performance.
A later modification that has gained popularity is the four-factor model that
added a momentum-related factor to the Fama-French three factors:
 
Rpt  Rft ¼ a þ b Rmt  Rft þ sSMBt þ hHMLt þ pPR12t þ et , (14.4)

where PR12t is the momentum-related factor advocated by Carhart (1997).

Typically, we compute PR12t by first ranking all firms by their previous
11-month stock return lagged 1 month and then taking the average return of the
top one third (i.e., high past return) stocks minus the average return of the bottom
one third (i.e., low past return) stocks.
Under the assumption that the asset-pricing model adequately explains variation
in expected stock returns, the intercept, a, measures the average monthly abnormal
return of the calendar-time portfolio of event firms and should be equal to zero
under the null hypothesis of no abnormal performance. If the test concludes that the

3
Loughran and Ritter (1995), Brav and Gompers (1996), and Brav et al. (2000), among others,
have used the calendar-time portfolio approach.
4
See Fama and French (1993) for details on construction of the mimicking portfolios for the
common size and book-to-market equity factors. We thank Eugene Fama for providing us with
returns on Rft, Rmt, SMBt, and HMLt.
390 J.S. Ang and S. Zhang

time series conforms to the asset-pricing model, the event is said to have had no
significant long-term effect; otherwise, the event has produced significant
long-term abnormal returns. Lyon et al. (1999) report that the calendar-time
portfolio approach together with the Fama-French three-factor model, which shall
be referred to as the Fama-French calendar-time approach later, is well specified for
random samples in their simulation study.
However, we do not know how much power the Fama-French calendar-time
approach has. Loughran and Ritter (1999) criticize the approach as having very low
power. They argue that reduction in power is caused by using returns on contam-
inated portfolios as factors in the regression, by weighting each month equally and
by using value-weighted returns of the calendar-time portfolios. However, their
empirical evidence is based only on one carefully constructed sample of firms and is
hardly conclusive. No large-scale simulation study has been done to examine power
of the Fama-French calendar-time approach, which we will remedy in this paper.
The Fama-French calendar-time approach, estimated with the ordinary least
squares (OLS) technique, could suffer from a potential heteroskedasticity problem
due to unequal and changing number of firms in the calendar-time portfolios. The
weighted least squares (WLS) technique, which is helpful in addressing the
heteroskedasticity problem, has been suggested as a way to deal with the changing
size of calendar-time portfolios. When applying WLS, we use the monthly number
of firms in the event portfolio as weights.

14.3 A Review of Simulation Studies on Long-Horizon Event

Study Methodology

Several papers have documented performance of testing procedures in large-scale

simulations. Table 14.1 surveys these papers with reference to testing procedures
under their investigation and their simulation settings. The simulation technique
was pioneered by Brown and Warner (1980, 1985) to evaluate size and power of
testing procedures. In this section, we review these simulation studies.
As shown in Fig. 14.1, there are two approaches for a long-term event study: the
calendar-time portfolio approach versus the buy-and-hold benchmark approach.
There has been a debate on which approach prescribes the best procedure for long-
term event studies. Both approaches have been under criticisms. The buy-and-hold
benchmark approach is susceptible to biases associated with cross-sectional corre-
lation, insufficient matching criteria, new equity issues, periodic balancing, and
skewed distribution of long-term abnormal returns, while the calendar-time
portfolio approach may suffer from an improper asset-pricing model and heteroske-
dasticity in portfolio returns. See Kothari and Warner (1997), Barber and Lyon
(1997), Fama (1998), Loughran and Ritter (1999), Lyon et al. (1999); and others
for more detailed discussions. Kothari and Warner (1997) argue that the combined
effect of these issues is difficult to specify a priori and, thus, “a simulation study with
actual security return data is a direct way to study the joint impact, and is helpful in
identifying the potential problems that are empirically most relevant.”
 14

Table 14.1 Summary of existing simulation studies

Procedures under investigation Validation settings
Calendar-time portfolio
approach Buy-and-hold benchmark approach Evidence on Evidence on
Estimation specification at power at
Authors (year) Pricing model method Matching criteria Test statistics Simulation design horizon of horizon of
Kothari and This paper examines procedures that are based on cumulating monthly 250 simulated samples of 1 month, 1, 2, 1 year
Warner (1997) abnormal returns. Such procedures are severely misspecified in most cases 200 firms each and 3 years
and are not recommended
Barber and Size, BE/ME t-test, Wilcoxon test 1,000 simulated samples 1, 3, and 1 year
Lyon (1997) of 200 firms each 5 years
Lyon Three-factor OLS Size, BE/ME t-test, Johnson’s test, 1,000 simulated samples 1, 3, and 1 year, only for
et al. (1999) model bootstrapped test of 200 firms each 5 years the buy-and-hold
approach
Mitchell and Three-factor OLS 1,000 simulated samples 3 years 3 years
Stafford model of 2,000 firms each
Evaluating Long-Horizon Event Study Methodology

(2000)
Cowan and Size, BE/ME t-test, two-group test with 1,000 simulated samples 1, 3, and 3 years
Sergeant winsorized data with sample size of 5 years
(2001) 50, 200, and 1,000
Ang and Three-factor OLS, Size, BE/ME, t-test, 250 simulated samples 1, 3, and 1, 3, and 5 years
Zhang (2004) model, four- WLS beta, and Johnson’s test, with sample size of 5 years
factor model correlation bootstrapped test, sign 200 and 1,000
coefficient test
Jegadeesh and Size, BE/ME t-test, t-test adjusted for 1,000 simulated samples 1, 3, and 1, 3, and 5 years
Karceski heteroskedasticity and of 200 firms each 5 years
(2009) serial correlation
391
392 J.S. Ang and S. Zhang

In their simulation study, Kothari and Warner (1997) measure the long-term
(up to 3 years) impact of an event by cumulative monthly abnormal returns, where
monthly abnormal returns are computed against four common models: the market-
adjusted model, the market model, the capital asset-pricing model, and the Fama-
French three-factor model. They find that tests for cumulative abnormal returns are
severely misspecified. They identify sample selection, survival bias, and bias in
variance estimation as potential sources of the misspecification and suggest that
nonparametric and bootstrap tests are likely to reduce misspecification.
Barber and Lyon (1997) address two main issues in their simulation study. First,
they argue that buy-and-hold return is a better measure of investors’ actual experience
over a long horizon and should be used in long-term event study (up to 5 years). They
show simulation evidence that approaches using cumulative abnormal returns cause
severe misspecification, which is consistent with the observation in Kothari and
Warner (1997). Second, they use simulations to measure both size and power of
testing procedures that follow the buy-and-hold benchmark approach. An important
finding is that using a single control firm as benchmark yields well-specified tests,
whereas using reference portfolio causes substantial over-rejection.
In a later paper, Lyon et al. (1999) report another simulation study (for up to the
5-year horizon) that investigates the performance of both buy-and-hold benchmark
approach and calendar-time portfolio approach. They find that using the
Fama-French three-factor model yields a well-specified test. However, they advo-
cate a test that uses carefully constructed reference portfolio as benchmark and the
bootstrapped Johnson’s statistic for testing abnormal returns. They present
evidence that this test is well specified and has high power at the 1-year horizon.
Two questions remain unanswered in Lyon et al. (1999). First, how much power
does the bootstrap test have for event horizons longer than 1 year (e.g., 3 or 5 years
that is common in long-horizon studies)? It is known in statistics literature that
a bootstrap test is not as powerful as simple nonparametric tests in many occasions
(see Efron and Tibshirani 1993, Chap. 16 and Davison and Hinckley 1997,
Chap. 4). It is necessary to know the actual power of such test for event horizons
beyond 1 year. Second, is the calendar-time portfolio approach as powerful as the
buy-and-hold benchmark approach? Loughran and Ritter (2000) argue that the
calendar-time portfolio approach has low power, using simulations and empirical
evidence from a sample of new equity issuers. However, they do not measure how
much power the approach actually has, which makes it impossible to compare the
two approaches directly in more general settings.
Mitchell and Stafford (2000) is the only study that empirically measures power
of the calendar-time portfolio approach using simulations. Their main focus is to
assess performance of several testing procedures in three large samples of major
managerial decisions, i.e., mergers, seasoned equity offerings, and share
repurchases (up to 3 years). They find that different procedures lead to contradicting
conclusions and argue that the calendar-time portfolio approach is preferred. To
resolve Loughran and Ritter’s (2000) critique that the calendar-time portfolio
approach has low power, they conduct simulations to measure the empirical
power and find that the power is actually very high with an empirical rejection
14 Evaluating Long-Horizon Event Study Methodology 393

rate of 99 % for induced abnormal returns of 15 % over a 3-year horizon. Since they
have a large sample size, this finding is actually consistent with what we document in
Table 14.5. However, their simulations focus on only samples of 2,000 firms. Many
event studies have much smaller sample sizes, especially after researchers slice and
dice a whole sample into subsamples. More evidence is needed in order to have great
confidence in applying the calendar-time portfolio approach in such studies.
Cowan and Sergeant (2001) focus on the buy-and-hold benchmark approach in their
simulations. They find that using the reference portfolio approach cannot overcome the
skewness bias discussed in Barber and Lyon (1997) and that the larger the sample size,
the smaller the magnitude of the skewness bias. They also argue that cross-sectional
dependence among event firms’ abnormal returns increases in event horizon due to
partially contemporaneous holding periods, which may cause the overlapping horizon
bias. They propose a two-group test using abnormal returns winsorized at three
standard deviations to deal with these two biases and report evidence that this test
yields correct specifications and considerable power in many situations.
All previous simulation studies use only size and BE/ME to construct
benchmarks, which is often justified by the findings in Fama and French (1992)
that size and BE/ME together adequately capture the cross-sectional variations in
average monthly stock returns. Ang and Zhang (2004) use two other matching
criteria to explore whether better benchmarks could be used for future studies. The
two criteria are market beta and pre-event correlation coefficient. Using market beta
is motivated by the fact that Fama and French’s (1993) three-factor model has
a market factor, a size-related factor, and a BE/ME-related factor. Matching on the
basis of size and BE/ME does not account for the influence of the market factor. The
rationale for using pre-event correlation coefficient is that matching on size and
BE/ME may fail to control for other factors that could influence stock returns, such as
industry factor, seasonal factor, momentum factor, and other factors shared by only
firms of the same characteristics, such as geographical location, ownership, and
governance structures. Matching on the basis of pre-event correlation coefficient
helps remove the effect of these factors on the event firm’s long-term return.
The main findings in Ang and Zhang (2004) include the following. First, the four-
factor model is inferior to the well-specified three-factor model in the calendar-time
portfolio approach in that the former causes too many rejections of the null hypothesis
relative to the specified significance level. Second, WLS improves the performance of
the calendar-time portfolio approach over OLS, especially for long event horizons.
Third, the Fama-French three-factor model has relatively high power in detecting
abnormal returns, although power decreases sharply as event horizon increases. Fourth,
the simple sign test is well specified when it is applied with a single firm benchmark,
but misspecified when used with reference portfolio benchmarks. More importantly,
the combination of the sign test and the benchmark with the single most correlated firm
consistently has much higher power than any other test in our simulations and is the
only testing procedure that performs well in samples of small firms.
Jegadeesh and Karceski (2009) propose a new test of long-run performance that
allows for heteroskedasticity and autocorrelation. Previous tests used in Lyon
et al. (1999) implicitly assume that the observations are cross-sectionally uncorrelated.
394 J.S. Ang and S. Zhang

This assumption is frequently violated in nonrandom samples such as samples with

industry clustering or with overlapping returns. To overcome the cross-correlation bias
in event firms’ returns, they recommend a t-statistic that is computed using a generalized
version of the Hansen and Hodrick (1980) standard error. Their simulation studies show
that the new tests they propose are reasonably well specified in random samples, in
samples that are concentrated in particular industries, and also in samples where event
firms enter the sample on multiple occasions within the holding period.
In summary, these simulation studies show that testing procedures differ
dramatically in performance. Some procedures reject the null hypothesis at an
excessively high rate, while others have very low power. These findings confirm
the Fama (1998) statement that evidence for long-term return anomalies is
dependent upon methodology and suggest that caution must be exercised in
choosing the proper methodology for a long-term event study.

14.4 A Simulation Study of Large-Size Samples

A simulation study of large-size samples serves two purposes. First, it is well

documented that the distribution of buy-and-hold abnormal returns tends to be skewed
to the right. Kothari and Warner (2007) mention that the extent of skewness bias is
likely to decline with sample size. It is of interest to provide evidence on how much is
the level of right-skewness in the average abnormal returns of large-size samples.
Second, although it is expected that testing power increases with sample size, it is of
practical interest to know more precisely how much power a test can have in a sample
of 1,000 observations. Large sample simulation defines the limits of a procedure.

14.4.1 Research Design

In this simulation study, we construct 250 samples each consisting of 1,000 event
firms. To produce one sample, we randomly select, with replacement, 1,000 event
months between January 1980 and December 1992, inclusively.5,6 This allows us to
calculate 5-year abnormal returns until December 1997. For each selected event
month, we randomly select, without replacement, one firm from a list of qualified
firms. The qualified firms satisfy the following requirements: (i) They are publicly
traded firms, incorporated in the USA, and have ordinary common shares with
Center for Research in Security Prices (CRSP) share codes 10 and 11; (ii) they have
return data found in the CRSP monthly returns database for the 24-month period

5
We use a pseudorandom number generator developed by Matsumoto and Nishimura (1998) to
ensure high quality of random sampling.
6
Kothari and Warner (1997) use 250 samples, each of 200 event months between January 1980 and
December 1989 inclusively. Barber and Lyon (1997) use 1,000 samples, each of 200 event months
in a much longer period from July 1963 to December 1994. The period under our study, between
January 1980 and December 1992, is of similar length to Kothari and Warner’s.
14 Evaluating Long-Horizon Event Study Methodology 395

prior to the event month; (iii) they have nonnegative book values on COMPUSTAT
prior to the event month so that we can calculate their book-to-market equity ratios.
The 250 samples, each of 1,000 randomly selected firms, comprise the
simulation setting for comparing the performance of different testing procedures.7
We apply all testing procedures under our study to the same samples. Such
controlled comparison is more informative because it eliminates difference in
performance due to variation in the samples.
For the buy-and-hold approach, we compute the long-term buy-and-hold abnor-
mal return of firm i as the difference between the long-term buy-and-hold return of
firm i and the long-term return of a benchmark. The buy-and-hold return of firm
i over t months is obtained by compounding monthly returns. In case that firm
i does not have return data for all t months, we replace missing returns by the
same-month returns of a size and BE/ME matched reference portfolio.8 We evaluate
a total of five benchmarks and four test statistics in this study. We briefly describe
them in the following and give the details in the Appendix.
Three of the benchmarks are reference portfolios. The first reference portfolio
consists of firms that are similar to the event firm in both size and BE/ME. We
follow the same procedure as in Lyon et al. (1999) to construct the two-factor
reference portfolio. We use the label “SZBM” for this benchmark. The second
reference portfolio consists of firms that are similar to the event firm not only in size
and BE/ME but also in market beta. We use the label “SZBMBT” for this bench-
mark. The third reference portfolio consists of ten firms that are most correlated
with the event firm prior to the event. We use the label “MC10” for this benchmark.
The other two of the five benchmarks consist of a single firm. The first single
firm benchmark is the firm that matched the event firm in both size and BE/ME. To
find the two-factor single firm benchmark, we first identify all firms whose market
value is within 70–130 % of the event firm’s market value and then choose the firm
that has the BE/ME ratio closest to that of the event firm. We use the label
“SZBM1” for this benchmark. The second single firm benchmark is the firm that
has the highest correlation coefficient with the event firm prior to the event. We use
the label “MC1” for this benchmark.
We apply four test statistics to test the null hypothesis that the mean long-term
abnormal return is zero. They include Student’s t-test, Fisher’s sign test, Johnson’s
skewness-adjusted t-test, and the bootstrapped Johnson’s t-test. Fisher’s sign test is
a nonparametric test and is described in details in Hollander and Wolfe (1999, Chap. 3).
Johnson’s skewness-adjusted t-statistic was developed by Johnson (1978) to deal with
the skewness-related misspecification error in Student’s t-test. Sutton (1992) proposes
to apply Johnson’s t-test with a computationally intensive bootstrap resampling

7
Ang and Zhang (2004) examine two other simulation settings. Under one setting, they have another
250 samples of 200 event firms, a smaller sample size than the setting in this chapter. Under the other
setting, they have the sample size of 200 with the requirement that event firms belong to the smallest
quintile sorted by NYSE firm size. The second setting is used to examine the effect of small firms.
8
Filling in missing returns is a common practice in calculating long-term buy-and-hold returns;
e.g., see Barber and Lyon (1997), Lyon et al. (1999), and Mitchell and Stafford (2000).
396 J.S. Ang and S. Zhang

technique when the population skewness is severe and the sample size is small. Lyon
et al. (1999) advocate use of the bootstrapped Johnson’s t-test because long-term
buy-and-hold abnormal returns are highly skewed when buy-and-hold reference port-
folios are used as benchmarks. We follow Lyon et al. (1999) and set the resampling
size in the bootstrapped Johnson’s t-test to be one quarter of the sample size.
For the Fama-French calendar-time approach, we use both the Fama-French
three-factor model and the four-factor model. We apply both ordinary least
squares (OLS) and weighted least squares (WLS) techniques to estimate parameters
in the pricing model. The WLS is used to correct the heteroskedasticity problem due
to the monthly variation in the number of firms in the calendar-time portfolio. When
applying WLS, we use the number of event firms in the portfolio as weights.

14.4.2 Simulation Results for the Buy-and-Hold Benchmark

Approach

In this section, we examine the performance of testing procedures that follow

the buy-and-hold benchmark approach. Implementation of the buy-and-hold
benchmark approach involves choosing both benchmark and test statistic. For this
reason, rather than focusing on what is the best among all benchmarks, or focusing
on what is the best among all test statistics, we address the more practical question
of finding the best combination of benchmark and test statistic. Combination of the
five benchmarks and the four test statistics yields 20 testing procedures, out of
which we look for the best combination.
For each sample of 1,000 abnormal returns, we compute mean, median, standard
deviation, interquartile range, skewness coefficient, and kurtosis coefficient.
Table 14.2 reports the average of these statistics over 250 samples.
Since these event firms, being randomly selected, may not experience any event or
may experience events that have offsetting effects on averaged stock returns, we
expect their abnormal returns to concentrate around zero. In Table 14.2, means are
close to zero for all five benchmarks at all three holding periods, but medians differ
systematically according to the type of benchmark used. Medians are clearly negative
under the three reference portfolio benchmarks (i.e., SZBM, SZBMBT, and MC10),
but close to zero under the two single firm benchmarks (i.e., SZBM1 and MC1). The
evidence suggests that reference portfolio benchmarks overestimate holding period
returns of many event firms, resulting in far too many event firms having negative
abnormal returns under the portfolio-based benchmarks. The extent of the
overestimation bias by portfolio-based benchmarks is quite severe and gets worse
as the time horizon lengthens. The bias, as measured by the magnitude of median,
ranges from around 4 % at a 1-year horizon to 12 % at a 3-year horizon and to more
than 20 % at a 5-year horizon. Bias of this magnitude could cause too many events to
be falsely identified as having significant long-term impact.
Volatility of abnormal returns increases with the length of holding period under all
five benchmarks. For the same holding period, volatility is higher under the two single
firm benchmarks than under the three reference portfolio benchmarks. This is expected
14 Evaluating Long-Horizon Event Study Methodology 397

Table 14.2 Descriptive statistics of abnormal returns in samples of 1,000 firms

Descriptive statistics
Standard Interquartile Skewness Kurtosis
Benchmark Mean Median deviation range coefficient coefficient
Panel A: 1-year holding period
SZBM 0.009 0.032 0.574 0.453 4.332 60.763
SZBMBT 0.001 0.043 0.586 0.462 4.074 58.462
MC10 0.000 0.040 0.591 0.463 3.853 56.733
SZBM1 0.005 0.005 0.814 0.638 0.203 53.034
MC1 0.002 0.003 0.780 0.584 0.229 53.202
Panel B: 3-year holding period
SZBM 0.034 0.112 1.240 0.963 4.561 57.644
SZBMBT 0.001 0.139 1.264 0.982 4.258 54.616
MC10 0.000 0.126 1.286 0.982 3.996 53.153
SZBM1 0.023 0.022 1.746 1.305 0.137 51.176
MC1 0.016 0.006 1.658 1.200 0.736 43.430
Panel C: 5-year holding period
SZBM 0.068 0.209 2.034 1.490 5.287 67.521
SZBMBT 0.002 0.248 2.073 1.514 4.982 64.364
MC10 0.007 0.223 2.106 1.516 4.652 61.091
SZBM1 0.054 0.039 2.802 1.979 0.269 41.428
MC1 0.036 0.000 2.745 1.834 0.500 50.365
This table reports descriptive statistics that characterize the probability distribution of long-term
abnormal returns, in samples of 1,000 firms. Abnormal return is calculated as the difference in holding
period return between the event firm and its benchmark. We use five benchmarks: a reference portfolio
matched by size and BE/ME (SZBM); a reference portfolio matched by size, BE/ME, and beta
(SZBMBT); a reference portfolio consisting of ten firms, within the event firm’s size and BE/ME
matched portfolio, whose returns are most correlated with the event firm’s MC10; a single firm
matched by size and BE/ME (SZBM1); and a single firm, from the event firm’s size and BE/ME
matched portfolio, whose returns have the highest correlation with the event firm’s MC1. We compute
mean, median, standard deviation, interquartile range, skewness coefficient, and kurtosis coefficient
for abnormal returns in every sample. Since there are 250 samples in the simulation, entries in the table
are the average of these statistics over the 250 samples

because reference portfolios have lower volatility due to averaging. As for kurtosis, all
five benchmarks produce highly leptokurtic abnormal returns, with kurtosis coeffi-
cients ranging from 41.4 to 67.5, which are far greater than three, the kurtosis
coefficient of any normal distribution. At last, skewness coefficients for the two single
firm benchmarks are close to zero regardless of event horizons, while skewness
coefficients for the three portfolio benchmarks are excessively positive.
To sum up, probability distributions of long-term abnormal returns exhibit
different properties, depending on whether the benchmark is a reference portfolio
or a single firm. Under a reference portfolio benchmark, the distribution is highly
leptokurtic and positively skewed, with a close-to-zero mean but a highly negative
median. Under a single firm benchmark, the distribution is highly leptokurtic but
symmetric, with both mean and median close to zero. Statistical properties of
long-term abnormal returns have important bearings on performance of test
398 J.S. Ang and S. Zhang

statistics. Overall, it seems single firm benchmarks have more desirable properties.
Between the two single firm benchmarks, MC1 shows better performance than
SZBM1, because the abnormal returns based on MC1 have both mean and median
being closer to zero and smaller standard deviation.
A superior test should control for the probability of committing two errors. First,
it is important to control for the probability of misidentifying an insignificant event
as having statistical significance; in other words, the empirical size of the test,
which is computed from simulations, is close to the prespecified significance level
at which the test is conducted. When this happens, the test is well specified. Second,
power of the test should be large, that is, the probability of finding a statistically
significant event if one did exist.
Table 14.3 reports empirical size of all 20 tests for three holding periods.
Empirical size is calculated as the proportion of 250 samples that rejects the null
hypothesis at the 5 % nominal significance level. With only a few exceptions,
Student’s t-test is well specified against the two-sided alternative hypothesis.
Despite excessively high skewness in abnormal returns from reference portfolio
benchmarks, Student’s t-test is well specified against two-sided alternative
hypothesis because the effect of skewness at both tails cancels out (see, e.g.,
Pearson and Please 1975). When testing against the two-sided alternative
hypothesis, Johnson’s skewness-adjusted t-test is in general misspecified, but its
bootstrapped version is well specified in most situations. The sign test is
misspecified when applied to abnormal returns from reference portfolio
benchmarks, and the extent of misspecification is quite serious and increases in
the length of holding period. This is not surprising because abnormal returns from
reference portfolio benchmarks have highly negative medians.
Table 14.4 reports empirical power of testing the null hypothesis of zero abnormal
return against the two-sided alternative hypothesis. We follow Brown and Warner
(1980, 1985) to measure empirical power by intentionally forcing the mean
abnormal return away from zero with induced abnormal returns. We induce nine levels
of abnormal returns ranging from 20 % to 20 % at an increment of 5 %. To induce an
abnormal return of 20 %, for example, we add 20 % to the observed holding period
return of an event firm. Empirical power is calculated as the proportion of 250 samples
that rejects the null hypothesis at 5 % significance level.
With a large sample size of 1,000, the power of these tests remains reasonably
high at the longer holding period. Ang and Zhang (2004) report that, with the sample
size of 200, the power of all tests deteriorates sharply as holding period lengthens
from 1 to 3 and to 5 years and is alarmingly low at the 5-year horizon. For example,
when the induced abnormal return is 20 % over a 5-year horizon, the highest power
of the bootstrapped Johnson’s t-test is 13.6 % for a sample of 200 firms, whereas the
highest power is 62.8 % for a sample of 1,000 firms.
We compare the power of the three test statistics: Student’s t-test, the
bootstrapped Johnson’s skewness-adjusted t-test, and the sign test. All three
test statistics are applied together with the most correlated single firm benchmark.
The evidence shows that all three tests are well specified. However, the sign
test clearly has much higher power than the other two tests.
 14

Table 14.3 Specification of tests in samples of 1,000 firms

Two-tailed test Lower-tailed test Upper-tailed test
Benchmark t Jt BJt sign t Jt BJt sign t Jt BJt sign
Panel A: 1-year holding period
SZBM 4.0 7.2 6.4 75.6* 2.8 2.0* 1.6* 85.6* 8.4* 15.2* 13.6* 0.0*
SZBMBT 5.2 6.4 4.0 92.0* 9.6* 9.2* 7.6 96.0* 2.0* 4.8 4.0 0.0*
MC10 5.6 6.4 6.4 85.6* 10.0* 8.0* 6.8 92.8* 2.4* 5.6 4.8 0.0*
SZBM1 4.4 5.6 3.6 4.0 2.8 4.4 2.8 1.6* 6.0 8.0* 6.4 10.8*
MC1 3.6 5.2 3.2 9.6* 6.0 8.0* 4.8 12.8* 6.8 7.6 6.4 2.4
Panel B: 3-year holding period
Evaluating Long-Horizon Event Study Methodology

SZBM 11.2* 14.4* 12.8* 99.6* 1.2* 1.2* 0.8* 100.0* 17.6* 22.8* 21.6* 0.0*
SZBMBT 5.2 5.2 5.6 100.0* 7.6 7.6 5.6 100.0* 2.8 3.2 3.6 0.0*
MC10 4.8 6.8 5.6 100.0* 6.8 5.6 4.4 100.0* 3.2 6.0 5.6 0.0*
SZBM1 6.0 7.6 5.2 9.2* 2.0* 2.8 1.6* 0.4* 11.2* 13.6* 10.0* 15.6*
MC1 6.8 8.4* 6.4 6.4 2.8 2.8 2.8 7.6 7.6 8.0* 6.8 1.2*
(continued)
399
 400

Table 14.3 (continued)

Two-tailed test Lower-tailed test Upper-tailed test
Benchmark t Jt BJt sign t Jt BJt sign t Jt BJt sign
Panel C: 5-year holding period
SZBM 17.6* 20.4* 19.6* 100.0* 0.4* 0.4* 0.4* 100.0* 22.8* 28.8* 28.0* 0.0*
SZBMBT 3.6 4.4 2.8 100.0* 5.6 4.0 3.2 100.0* 2.8 5.6 3.6 0.0*
MC10 2.0* 4.4 2.8 100.0* 3.6 3.2 2.4 100.0* 3.6 6.4 5.6 0.0*
SZBM1 8.0* 10.4* 6.4 12.0* 1.2* 1.2* 1.2* 0.0* 13.6* 15.2* 11.2* 19.6*
MC1 6.0 8.4* 5.2 2.4 1.6* 2.0* 2.0* 4.0 10.8* 12.4* 9.2* 3.2
This table reports empirical size of testing the null hypothesis of zero abnormal return against two-tailed, lower-tailed, and upper-tailed alternative hypothesis,
in samples of 1,000 firms. Empirical size is calculated as the proportion of 250 samples that reject the null hypothesis at 5 % significance level. Abnormal
return is calculated as the difference in holding period return between the event firm and its benchmark. We use five benchmarks (a reference portfolio matched
by size and BE/ME (SZBM); a reference portfolio matched by size, BE/ME, and beta (SZBMBT); a reference portfolio consisting of ten firms, within the
event firm’s size and BE/ME matched portfolio, whose returns are most correlated with the event firm’s MC10; a single firm matched by size and BE/ME
(SZBM1); and a single firm, from the event firm’s size and BE/ME matched portfolio, whose returns have the highest correlation with the event firm’s MC1)
and four test statistics (the conventional t-test (t), Fisher’s sign test (sign), Johnson’s skewness-adjusted t-test (Jt), and the bootstrapped Johnson’s skewness-
adjusted t-test (BJt)). It is indicated by * that the empirical size is significantly different from the 5 % significance level. The significance is judged against the
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
critical values 0:05  1:96 0:05ð1  0:05Þ=250, where 0.05 is the theoretical size, 1.96 is the 97.5th percentile of the standard normal distribution, and 250 is
the sample size
J.S. Ang and S. Zhang
14 Evaluating Long-Horizon Event Study Methodology 401

Table 14.4 Power of tests in samples of 1,000 firms

Induced abnormal return over the holding period (%)
Test Benchmark 20 15 10 5 0 5 10 15 20
Panel A: 1-year holding period
t SZBM 100.0 99.6 98.4 62.0 4.0 92.8 100.0 100.0 100.0
SZBMBT 100.0 99.6 98.8 76.8 5.2 79.2 100.0 100.0 100.0
MC10 100.0 99.6 98.4 73.6 5.6 77.6 100.0 100.0 100.0
SZBM1 100.0 99.6 93.6 46.8 4.4 58.4 97.6 99.2 100.0
MC1 100.0 99.6 97.2 50.8 3.6 58.0 96.8 99.6 100.0
Jt SZBM 89.2 94.4 93.2 55.2 7.2 94.4 100.0 100.0 100.0
SZBMBT 89.6 94.4 95.2 69.6 6.4 83.2 100.0 100.0 100.0
MC10 91.2 95.6 95.2 66.4 6.4 80.0 100.0 100.0 100.0
SZBM1 98.4 97.6 92.0 47.6 5.6 58.8 94.8 98.0 98.0
MC1 98.0 98.4 95.6 50.0 5.2 59.2 95.6 98.0 98.0
BJt SZBM 80.8 86.0 85.2 47.6 6.4 93.2 100.0 100.0 100.0
SZBMBT 79.2 85.2 86.0 57.2 4.0 81.2 100.0 100.0 100.0
MC10 81.6 86.4 87.2 56.4 6.4 78.8 100.0 100.0 100.0
SZBM1 96.0 96.0 87.2 40.4 3.6 51.6 90.0 95.2 94.0
MC1 95.6 95.6 88.8 44.4 3.2 51.6 91.6 95.6 95.6
Sign SZBM 100.0 100.0 100.0 100.0 75.6 28.4 100.0 100.0 100.0
SZBMBT 100.0 100.0 100.0 100.0 92.0 10.4 99.2 100.0 100.0
MC10 100.0 100.0 100.0 100.0 85.6 17.2 100.0 100.0 100.0
SZBM1 100.0 100.0 100.0 72.0 4.0 92.0 100.0 100.0 100.0
MC1 100.0 100.0 100.0 93.6 9.6 90.4 100.0 100.0 100.0
Panel B: 3-year holding period
t SZBM 96.0 80.8 43.2 9.6 11.2 58.0 96.0 100.0 100.0
SZBMBT 98.4 93.2 70.8 30.8 5.2 19.2 73.2 98.4 100.0
MC10 98.4 92.4 70.0 26.8 4.8 19.2 72.4 98.8 100.0
SZBM1 88.4 63.6 30.8 10.0 6.0 27.6 64.4 85.6 96.4
MC1 92.4 74.0 36.4 10.4 6.8 22.4 64.0 91.2 97.6
Jt SZBM 91.2 74.8 38.4 9.6 14.4 66.4 96.4 100.0 100.0
SZBMBT 94.8 88.0 65.6 26.0 5.2 24.4 78.4 98.8 100.0
MC10 94.0 87.6 62.8 24.4 6.8 23.2 76.8 99.2 100.0
SZBM1 86.4 63.2 32.4 12.4 7.6 29.2 64.0 84.8 94.4
MC1 90.4 72.4 36.0 12.0 8.4 24.4 64.8 90.4 97.6
BJt SZBM 84.8 66.0 32.4 7.6 12.8 62.4 96.0 100.0 100.0
SZBMBT 88.8 82.0 58.4 21.6 5.6 21.6 74.8 98.4 100.0
MC10 90.4 79.6 54.8 19.6 5.6 21.6 73.6 98.0 100.0
SZBM1 81.6 56.4 27.6 8.0 5.2 24.4 56.4 79.6 88.8
MC1 86.0 65.6 29.6 8.8 6.4 20.4 55.2 86.8 94.4
Sign SZBM 100.0 100.0 100.0 100.0 99.6 63.6 6.0 27.2 88.8
SZBMBT 100.0 100.0 100.0 100.0 100.0 89.6 27.6 6.8 64.4
MC10 100.0 100.0 100.0 100.0 100.0 78.4 15.6 12.4 78.8
SZBM1 100.0 94.8 56.4 14.0 9.2 54.8 94.0 100.0 100.0
MC1 100.0 100.0 95.2 50.0 6.4 37.2 86.4 100.0 100.0
(continued)
402 J.S. Ang and S. Zhang

Table 14.4 (continued)

Induced abnormal return over the holding period (%)
Test Benchmark 20 15 10 5 0 5 10 15 20
Panel C: 5-year holding period
t SZBM 58.0 28.4 9.6 1.6 17.6 40.8 79.2 97.6 99.2
SZBMBT 84.8 63.2 37.6 14.8 3.6 8.4 32.8 66.0 92.0
MC10 80.4 61.2 32.8 11.6 2.0 10.4 32.4 69.6 92.0
SZBM1 38.0 18.4 7.2 4.0 8.0 21.6 41.6 64.8 82.4
MC1 50.4 23.6 10.4 4.0 6.0 17.2 38.0 61.2 81.2
Jt SZBM 44.4 23.6 7.2 5.6 20.4 52.0 85.2 98.8 99.6
SZBMBT 72.0 51.6 27.6 11.2 4.4 14.8 40.0 73.2 94.8
MC10 71.6 51.2 27.6 7.6 4.4 15.6 38.0 73.6 96.4
SZBM1 38.0 20.0 8.8 6.0 10.4 23.6 42.4 65.2 82.0
MC1 48.8 24.4 11.2 6.0 8.4 18.8 38.4 60.4 79.6
BJt SZBM 35.2 19.6 5.2 2.8 19.6 48.0 82.8 98.0 99.6
SZBMBT 62.8 43.2 21.6 8.4 2.8 12.8 36.8 70.0 94.0
MC10 60.0 42.4 20.0 6.4 2.8 14.4 36.0 72.0 94.8
SZBM1 30.0 16.0 5.6 4.0 6.4 16.4 33.6 54.4 70.4
MC1 38.4 17.6 9.6 3.2 5.2 15.2 30.8 50.4 70.8
Sign SZBM 100.0 100.0 100.0 100.0 100.0 95.6 72.0 22.8 3.2
SZBMBT 100.0 100.0 100.0 100.0 100.0 99.6 93.2 61.6 19.2
MC10 100.0 100.0 100.0 100.0 100.0 98.4 81.6 40.8 7.6
SZBM1 91.2 63.2 20.8 4.0 12.0 48.8 86.0 97.2 100.0
MC1 99.2 92.8 59.2 22.8 2.4 20.4 67.2 93.6 99.2
This table reports empirical power of testing the null hypothesis of zero abnormal return against the two-
sided alternative hypothesis, in samples of 1,000 firms. Empirical power is calculated as the proportion
of 250 samples that reject the null hypothesis at 5 % significance level. Abnormal return is calculated
as the difference in holding period return between the event firm and its benchmark. We use five
benchmarks (a reference portfolio matched by size and BE/ME (SZBM); a reference portfolio matched
by size, BE/ME, and beta (SZBMBT); a reference portfolio consisting of ten firms, within the event
firm’s size and BE/ME matched portfolio, whose returns are most correlated with the event firm’s
MC10; a single firm matched by size and BE/ME (SZBM1); and a single firm, from the event firm’s size
and BE/ME matched portfolio, whose returns have the highest correlation with the event firm’s MC1)
and four test statistics (the conventional t-test (t), Johnson’s skewness-adjusted t-test (Jt), the
bootstrapped Johnson’s skewness-adjusted t-test (BJt), and Fisher’s sign test (sign)). We study power
at nine levels of induced abnormal return, ranging from 20 % to 20 % at an increment of 5 %

14.4.3 Simulation Results for the Calendar-Time Portfolio Approach

Table 14.5 reports the rejection frequency of the calendar-time portfolio approach
in testing the null hypothesis that the intercept is zero in the regression of monthly
calendar-time portfolio returns, against the two-sided alternative hypothesis. Rejec-
tion frequency is measured as the proportion of the total 250 samples that reject the
null hypothesis. We compute rejection frequencies at nine nominal levels of
induced abnormal returns, ranging from 20 % to 20 % at an increment of 5 %.
Since monthly returns of the calendar-time portfolio are used in fitting the model, to
examine the power of testing the intercept, we need to induce abnormal returns by
14 Evaluating Long-Horizon Event Study Methodology 403

Table 14.5 Rejection frequency of calendar-time portfolio approach in samples of 1,000 firms
Panel A: 1-year holding period
Average effective induced holding period return (%)
20.4 15.7 10.7 5.5 0 5.7 11.7 17.9 24.4
Three factors OLS 100.0 100.0 99.2 53.2 2.4 78.8 100.0 100.0 100.0
WLS 100.0 100.0 99.6 74.4 2.0* 82.8 100.0 100.0 100.0
Four factors OLS 100.0 99.2 90.8 18.0 28.0* 97.6 100.0 100.0 100.0
WLS 100.0 99.6 93.2 20.8 25.2* 98.8 100.0 100.0 100.0
Panel B: 3-year holding period
Average effective induced holding period return (%)
25.2 19.3 13.2 6.8 0 7.1 14.5 22.3 30.4
Three factors OLS 98.0 86.8 38.0 3.6 2.4 32.0 84.8 99.6 99.6
WLS 100.0 97.2 65.2 10.0 1.2* 36.0 91.6 100.0 100.0
Four factors OLS 69.2 22.0 1.6 6.4 55.2* 94.0 99.6 100.0 100.0
WLS 92.0 38.0 4.0 10.4 75.6* 99.6 100.0 100.0 100.0
Panel C: 5-year holding period
Average effective induced holding period return (%)
31.1 23.9 16.3 8.3 0 8.7 17.9 27.4 37.5
Three factors OLS 64.8 31.2 10.0 0.8 4.0 27.6 62.4 90.8 99.6
WLS 94.4 58.4 14.8 0.4 4.0 36.0 81.2 99.2 100.0
Four factors OLS 12.4 1.6 5.2 32.8 70.8* 89.2 98.8 100.0 100.0
WLS 14.0 1.2 14.8 62.4 94.0* 100.0 100.0 100.0 100.0
This table reports rejection frequency in testing the null hypothesis that the intercept in the
regression of monthly calendar-time portfolio returns is zero, in samples of 1,000 firms. Both
the Fama-French three-factor model and the four-factor model are used in the regression. Model
parameters are estimated with both OLS and WLS estimation techniques. Rejection frequency is
equal to the proportion of 250 samples that reject the null hypothesis at 5 % significance level. We
measure rejection frequency at nine levels of induced abnormal returns. We induce abnormal
returns by adding an extra amount to monthly returns of every event firm before forming the
calendar-time portfolios. The effective induced holding period return of an event firm is equal to
the difference in the firm’s holding period return between before and after adding the monthly
extra amount. The average effective induced holding period return is computed over all event
firms in the 250 samples

adding an extra amount to actual monthly returns of every event firm before
forming the calendar-time portfolios. For example, in order to induce the 20 %
nominal level of abnormal holding period return, we add the extra amount
of 1.67 % (¼20 %/12) to an event firm’s 12 monthly returns for a 1-year
horizon, or add the abnormal amount of 0.56 % (¼ 20 %/36) to the firm’s
24 monthly returns for a 3-year horizon, or the abnormal amount of 0.33 %
(¼20 %/60) to the firm’s 60 monthly returns for a 5-year horizon.
Note that the nominal induced holding period return is different from the
effective induced abnormal holding period return, because adding the abnormal
amount each month does not guarantee that an event firm’s holding period return
will be increased or decreased by the exact nominal level. We measure the effective
induced holding period return of an event firm as the difference in the firm’s holding
404 J.S. Ang and S. Zhang

period return between before and after adding the monthly abnormal amount. The
average effective induced holding period return is computed over all event firms in
the 250 samples. The average induced holding period return allows us to compare
power of the buy-and-hold benchmark approach with that of the calendar-time
portfolio approach at the scale of holding period return.
We first examine empirical size of the calendar-time portfolio approach, which is
equal to the rejection frequency when no abnormal return is induced. In Table 14.5,
the empirical size is in the column with zero induced return. It is very surprising that
when the four-factor model is used, the test has excessively high rejection frequency
at 3-year and 5-year horizons. The rejection frequency, for example, is 94.0 % at the
5-year horizon with the WLS estimation! In contrast, when the Fama-French three-
factor model is used, the empirical sizes are not significantly different from the 5 %
significance level. The evidence strongly suggests that the three-factor model is
preferred for the calendar-time portfolio approach, whereas the four-factor model
suffers from overfitting and should not be used.
Table 14.5 shows that, for a sample of 1,000 firms, the power of this approach
remains high as event horizon increases. WLS estimation does improve the power
of the procedure over the OLS, and the extent of improvement becomes greater as
holding period gets longer. By comparing Tables 14.4 and 14.5, we find that the
power of the Fama-French calendar-time approach implemented with WLS tech-
nique (i.e., FF, WLS) has almost the same power as the buy-and-hold benchmark
approach implemented with the most correlated single firm and the sign test (i.e.,
MC1, sign), at the 1-year horizon, but slightly less at the 3- and 5-year horizons.

14.5 Conclusion

Comparing the simulation results in Sect. 14.4 with those in Ang and Zhang (2004),
we find that sample size has a significant impact on the performance of tests in long-
horizon event studies. With a sample size of 1,000, a few tests perform reasonably
well, including the Fama-French calendar-time approach implemented with WLS
technique and the buy-and-hold benchmark approach implemented with the most
correlated single firm (MC1) and the sign test. In particular, they have reasonably
high power even for the long 5-year holding period. On the contrary, with a sample
size of 200, Ang and Zhang (2004) find that the power of most well-specified tests is
very low for the 5-year horizon, only in the range of 10–20 % against a high level of
induced abnormal returns, while the combination of the most correlated single firm
and the sign test stands out with a power of 41.2 %. Thus, the most correlated single
firm benchmark dominates for most practical sample sizes, and in addition, the
simplicity of the sign test is appealing.
The findings have important implications for future research. For long-horizon
event studies with a large sample, it is likely to be more fruitful to spend efforts on
understanding the characteristics of the sample firms, than on implementing
various sophisticated testing procedures. The simulation results here show that the
commonly used tests following both the Fama-French calendar-time approach and
14 Evaluating Long-Horizon Event Study Methodology 405

the buy-and-hold benchmark approach perform reasonably well. In a recent paper,

Butler and Wan (2010) reexamine the long-run underperformance of bond-issuing
firms and find that straight debt and convertible debt issuers appear to have system-
atically better liquidity than benchmark firms, and controlling for liquidity by having
an additional matching criterion eliminates the underperformance. This resonates
well with Barber and Lyon’s (1997) suggestion that “as future research in financial
economics discovers additional variables that explain the cross-sectional variation in
common stock returns, it will also be important to consider these additional variables
when matching sample firms to control firms” (pp. 370–71). One reason why the
benchmark with a single most correlated firm performs well in our simulations may
be that returns of highly correlated firms are likely to move in tandem in response to
changes in risk factors that are well known, such as the market, size, and book-to-
market ratio, but also changes in other factors, such as industry, liquidity, momen-
tum, and seasonality.
On the other hand, for long-horizon event studies with a small sample, it may be
necessary to use a wide range of tests and interpret their outcome with care. This
prompts researchers to continue searching for better test statistics. For example,
Kolari and Pynnonen (2010) find that even relatively low cross-correlation among
abnormal returns in a short event window causes serious over-rejection of the null
hypothesis. They propose both cross-correlation and volatility-adjusted as well as
cross-correlation-adjusted scaled test statistics and demonstrate that these statistics
perform well in samples of 50 firms. It is an open and interesting question
whether these statistics have high power in long-horizon event studies with
a small sample.

Appendix

This appendix includes the details on the benchmarks and the test statistics that are
used in our simulation studies. We use five benchmarks. The first benchmark is
a reference portfolio constructed on the basis of firm size and BE/ME. We follow
Lyon et al. (1999) to form 70 reference portfolios at the end of June in each
year from 1979 to 1997. At the end of June of year t, we calculate the size of
every qualified firm as price per share multiplied by shares outstanding. We sort all
NYSE firms by firm size into ten portfolios, each having the same number of firms,
and then place all AMEX/NASDAQ firms into the ten portfolios based on firm size.
Since a majority of NASDAQ firms are small, approximately 50 % of all firms fall
in the smallest size decile. To obtain portfolios with the same number of firms, we
further partition the smallest size decile into five subportfolios by firm size
without regard to listing exchange. We now have 14 size portfolios. Next, we
calculate each qualified firm’s BE/ME as the ratio of the book equity value
(COMPUSTAT data item 60) of the firm’s fiscal year ending in year t  1 to its
market equity value at the end of December of year t  1. We then divide each of
the 14 portfolios into five subportfolios by BE/ME and conclude the procedure
with 70 reference portfolios on the basis of size and BE/ME.
406 J.S. Ang and S. Zhang

The size and BE/ME matched reference portfolio of an event firm is taken to
be the one of the 70 reference portfolios constructed at the month of June prior to the
event month that matches the event firm in size and BE/ME. The return on a size and
BE/ME matched reference portfolio over t months is calculated as

2 Xnt 3
Y
t1 r
j¼1 jt
BRSZBM ¼ 41 þ 5  1, (14.5)
i
t¼0
nt

where month t ¼ 0 is the event month, nt is the number of firms in month t, and rjt is
the monthly return of firm j in month t. We use the label “SZBM” for the benchmark
that is based on firm size and BE/ME.
The second benchmark is a reference portfolio constructed on the basis of firm size,
BE/ME, and market beta. The Fama-French three-factor model suggests that expected
stock returns are related to three factors: a market factor, a size-related factor, and a BE/
ME-related factor. Reference portfolios constructed on the basis of size and BE/ME
account for the systematic portion of expected stock returns due to the size and BE/ME
factors, but not the portion due to the market factor. Our second benchmark is based on
firm size, BE/ME, and market beta to take all three factors into account.
To build a three-factor reference portfolio for a given event firm, we first
construct the 70 size and BE/ME reference portfolios as above and identify the
one that matches the event firm. Next, we pick firms within the matched portfolio
that have returns in CRSP monthly returns database for all 24 months prior to the
event month and compute their market beta by regressing the 24 monthly returns on
the value-weighted CRSP return index. Lastly, we divide these firms that have
market beta into three portfolios by their rankings in beta and pick the one that
matches the event firm in beta as the three-factor reference portfolio. The return on
a three-factor portfolio over t months is calculated as
2 Xlt 3
Y
t1 r
j¼1 jt
BRSZBMBT ¼ 41 þ 5  1, (14.6)
i
t¼0
nt

where month t ¼ 0 is the event month, nt is the number of firms in month t, and rjt is
the monthly return of firm j in month t. We use the label “SZBMBT” to indicate that
the benchmark is based on firm size, BE/ME, and market beta.
The third benchmark is a reference portfolio constructed on the basis of firm size,
BE/ME, and pre-event correlation coefficient. The rational for using pre-event
correlation coefficient as an additional dimension is that returns of highly correlated
firms are likely to move in tandem in response to not only changes in “global” risk
factors, such as the market factor, the size factor, and the BE/ME factor in the Fama-
French model, but also changes in other “local” factors, such as the industry factor,
the seasonal factor, liquidity factor, and the momentum factor. Over a long time
period following an event, both global and local factors experience changes that
affect stock returns. It is reasonable to expect more correlated stocks would be
14 Evaluating Long-Horizon Event Study Methodology 407

affected by these factors similarly and should have resulting stock return patterns
that are closer to each other. Therefore, returns of a reference portfolio on the basis
of pre-event size, BE/ME, and pre-event correlation coefficient are likely to
be better estimate of the status quo (i.e., what if there was no event) return of
an event firm.
To build a reference portfolio on the basis of size, BE/ME, and pre-event
correlation coefficient, we first construct the same 70 size and BE/ME reference
portfolios as above and identify the combination that matches the event firm. Next,
we pick firms within the matched size and BE/ME reference portfolio that have
returns in CRSP monthly returns database for all 24 months prior to the event month
and compute their correlation coefficients with the event firm over the pre-event
24 months. Lastly, we choose the ten firms that have the highest pre-event
correlation coefficient with the event firm to form the reference portfolio. Return
of the portfolio over t months is calculated as
Yt1  
X
10 1 þ rjt  1
BRMC10
i ¼ t¼0
, (14.7)
j¼1
10

where month t ¼ 0 is the event month and rjt is the monthly return of firm j in
month t. We use the label “MC10” to indicate that the benchmark consists of the
most correlated ten firms. The benchmark return is the return of investing equally in
the ten most correlated firms over the t months beginning with the event month.
The benchmark is to be considered as a hybrid between the reference portfolio
discussed above and the matching firm approach shown below.
The fourth benchmark is a single firm matched to the event firm in size and
BE/ME. Barber and Lyon (1997) report that using a size and BE/ME matched firm
as benchmark gives measurements of long-term abnormal return that is free of the
new listing bias, the rebalancing bias, and the skewness bias documented in Kothari
and Warner (1997) and Barber and Lyon (1997). To select the size and BE/ME
matched firm, we first identify all firms that have a market equity value between
70 % and 130 % of that of the event firm and then choose the firm with BE/ME
closest to that of the event firm. The buy-and-hold return of the matched firm
is computed as in Eq. 14.2. We use the label “SZBM1” to represent the single
size and BE/ME matched firm.
The fifth and last benchmark is a single firm that has the highest pre-event correlation
coefficient with the event firm. Specifically, to select the firm, we first construct the
70 size and BE/ME reference portfolios and identify the one that matches the event firm.
Next, we pick firms within the matched size and BE/ME reference portfolio that
have returns in CRSP monthly returns database for all 24 months prior to the event
month and compute their correlation coefficients with the event firm over the pre-event
24 months. We choose the firm with the highest pre-event correlation coefficient with the
event firm as the benchmark. The buy-and-hold return of the most correlated firm
is computed as in Eq. 14.2. We use the label “MC1” to represent the most
correlated single firm.
408 J.S. Ang and S. Zhang

We apply four test statistics to test the null hypothesis of no abnormal returns:
(a) Student’s t-test, (b) Fisher’s sign test, (c) Johnson’s skewness-adjusted t-test, and
(d) bootstrapped Johnson’s t-test.
(a) Student’s t-test
Given the long-term buy-and-hold abnormal returns for a sample of n event
firms, we compute Student’s t-statistic as follows:
AR
t¼ pffiffiffi , (14.8)
sðARÞ= n
where AR is the sample mean and s(AR) the sample standard deviation of the
given sample of abnormal returns. The Student’s t-statistic tests the null
hypothesis that the population mean of long-term buy-and-hold abnormal
returns is equal to zero. The usual assumption for applying the Student’s
t-statistic is that abnormal returns are mutually independent and follow the
same normal distribution.
(b) Fisher’s sign test
To test the null hypothesis that the population median of long-term buy-and-
hold abnormal returns is zero, we compute Fisher’s sign test statistic as follows:
X
n
B¼ IðARi > 0Þ, (14.9)
i¼1

where I(ARi > 0) equals 1 if the abnormal return on the ith firm is greater than
zero and 0 otherwise. At the chosen significance level of a, the null hypothesis
is rejected in favor of the alternative of nonzero median if B  b(a/2, n, 0.5) or
B < [n  b(a/2, n, 0.5)], or in favor of positive median if B  b(a, n, 0.5), or in
favor of negative median if B < [n  b(a, n, 0.5)]. The constant b(a, n, 0.5) is
the upper a percentile point of the binomial distribution with sample size n and
success probability of 0.5. The usual assumption for applying the sign test is
that abnormal returns are mutually independent and follow the same continuous
distribution. Note that application of the sign test does not require the popula-
tion distribution to be symmetric. When the population distribution is symmet-
ric, the population mean equals the population median, and the sign test then
indicates the significance of the population mean (see Hollander and Wolfe
2000, Chap. 3).
(c) Johnson’s skewness-adjusted t-test
Johnson (1978) developed the following skewness-adjusted t-test to correct the
misspecification of Student’s t-test caused by the skewness of the population
distribution. Johnson’s test statistic is computed as follows:
1 1
J ¼ t þ pffiffiffi t2 g þ pffiffiffi g, (14.10)
3 n 6 n
where t is Student’s t-statistic given in Eq. 14.8 and g is an estimate of the
Xn  3
coefficient of skewness given by g ¼ ARi  AR =sðARÞ3 n . Johnson’s
i¼1
14 Evaluating Long-Horizon Event Study Methodology 409

t-test is applied to test the null hypothesis of zero mean under the assumption
that abnormal returns are mutually independent and follow the same continuous
distribution. At the chosen significance level of a, the null hypothesis is rejected
in favor of the alternative of nonzero mean if J > t(a/2, u) or J <  t(a/2, u),
or in favor of positive mean if J > t(a, u), or in favor of negative mean if
J < t(a, u). The constant t(a, u) is the upper a percentile point of the Student’s
t distribution with the degrees of freedom u ¼ n  1.
(d) Bootstrapped Johnson’s skewness-adjusted t-test
Sutton (1992) proposes to apply Johnson’s t-test with a computer-intensive
bootstrap resampling technique when the population skewness is severe and the
sample size is small. He demonstrates it by an extensive Monte Carlo study that
the bootstrapped Johnson’s t-test reduces both type I and type II errors com-
pared to Johnson’s t-test. Lyon et al. (1999) advocate the bootstrapped
Johnson’s t-test in that long-term buy-and-hold abnormal returns are highly
skewed when buy-and-hold reference portfolios are used as benchmarks. They
report that the bootstrapped Johnson’s t-test is well specified and has consider-
able power in testing abnormal returns at the 1-year horizon. In this paper, we
document its power at 3- and 5-year horizons.
We apply the bootstrapped Johnson’s t-test as follows. From the given
sample of n event firms, we draw m firms randomly with replacement counted
as one resample until we have 250 resamples. We calculate Johnson’s test
statistic as in Eq. 14.10 for each resample and end up with 250 J values, labeled
as J1,   , J250. Let J0 denotes the J value of the original sample. To test the null
hypothesis of zero mean at the significance level of a, we first determine two
critical values, C1 and C2, such that the percentage of J values less than C1
equals a/2 and the percentage of J values greater than c2 equals a/2, and then
reject the null hypothesis if J0 < C1 or J0 > C2. We follow Lyon et al. (1999) to
apply the bootstrapped Johnson’s t-test with m ¼ 50.9

References
Ang, J. S., & Zhang, S. (2004). An evaluation of testing procedures for long horizon event studies.
Review of Quantitative Finance and Accounting, 23, 251–274.
Barber, B. M., & Lyon, J. D. (1997). Detecting long-run abnormal stock returns: The empirical
power and specification of test statistics. Journal of Financial Economics, 43, 341–372.
Brav, A., & Gompers, P. A. (1997). Myth or reality? The long-run underperformance of initial
public offerings: Evidence from venture and nonventure capital-backed companies. Journal of
Finance, 52, 1791–1821.
Brav, A., Geczy, C., & Gompers, P. A. (2000). Is the abnormal return following equity issuances
anomalous? Journal of Financial Economics, 56, 209–249.

9
Noreen (1989, ▶ Chap. 4) cautions that bootstrap hypothesis tests can be unreliable and that
extensive research is necessary to determine which one of many possible specifications can be
trusted in a particular hypothesis testing situation. We also apply the bootstrapped Johnson’s t-test
with m ¼ 100, 200. We find no significant difference in the test’s performance.
410 J.S. Ang and S. Zhang

Brown, S. J., & Warner, J. B. (1980). Measuring security price performance. Journal of Financial
Economics, 8, 205–258.
Brown, S. J., & Warner, J. B. (1985). Using daily stock returns: The case of event studies. Journal
of Financial Economics, 14, 3–31.
Butler, A. W., & Wan, H. (2010). Stock market liquidity and the long-run stock performance of
debt issuers. Review of Financial Studies, 23, 3966–3995.
Carhart, M. M. (1997). On persistence in mutual fund performance. Journal of Finance, 52, 57–82.
Cowan, A. R., & Sergeant, A. M. A. (2001). Interacting biases, non-normal return
distributions and the performance of tests for long-horizon event studies. Journal of Banking and
Finance, 25, 741–765.
Dahlquist, M., & de Jong, F. (2008). Pseudo market timing: A reappraisal. Journal of Financial
and Quantitative Analysis, 43, 547–580.
Davison, A. C., & Hinkley, D. V. (1997). Bootstrap methods and their application. Cambridge,
UK: Cambridge University Press.
Eckbo, B. E., Masulis, R. W., & Norli, O. (2000). Seasoned public offerings: Resolutions of the
‘new issues puzzle’. Journal of Financial Economics, 56, 251–291.
Eckbo, B. E., & Norli, O. (2005). Liquidity risk, leverage, and long-run IPO returns. Journal of
Corporate Finance, 11, 1–35.
Efron, B., & Tibshirani, R. J. (1993). An introduction to the bootstrap. New York: Chapman & Hall.
Fama, E. F. (1998). Market efficiency, long-term returns and behavioral finance. Journal of
Financial Economics, 49, 283–306.
Fama, E. F., & French, K. R. (1992). The cross-section of expected stock returns. Journal of
Finance, 47, 427–465.
Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics, 33, 3–56.
Hollander, M., & Wolfe, D. A. (1999). Nonparametric statistical methods. New York: Wiley.
Jegadeesh, N. (2000). Long-term performance of seasoned equity offerings: Benchmark errors
and biases in expectations. Financial Management, 29, 5–30.
Jegadeesh, N., & Karceski, J. (2009). Long-run performance evaluation: Correlation and
heteroskedasticity-consistent tests. Journal of Empirical Finance, 16, 101–111.
Jegadeesh, N., & Titman, S. (1993). Returns to buying winners and selling losers: Implications for
stock market efficiency. Journal of Finance, 48, 65–91.
Jegadeesh, N., & Titman, S. (2001). Profitability of momentum strategies: An evaluation of
alternative explanations. Journal of Finance, 56, 699–720.
Johnson, N. J. (1978). Modified t tests and confidence intervals for asymmetrical populations.
Journal of the American Statistical Association, 73, 536–544.
Kolari, J. W., & Pynnonen, S. (2010). Event study testing with cross-sectional correlation of
abnormal returns. Review of Financial Studies, 23, 3996–4025.
Kothari, S. P., & Warner, J. B. (1997). Measuring long-horizon security price performance.
Journal of Financial Economics, 43, 301–340.
Kothari, S. P., & Warner, J. B. (2007). Econometrics of event studies. In B. E. Eckbo (Ed.),
Handbooks of corporate finance: Empirical corporate finance. Amsterdam:
Elsevier/North-Holland.
Loughran, T., & Ritter, J. R. (1995). The new issues puzzle. Journal of Finance, 50, 23–52.
Loughran, T., & Ritter, J. R. (2000). Uniformly least powerful tests of the market efficiency.
Journal of Financial Economics, 55, 361–389.
Lyon, J. D., Barber, B. M., & Tsai, C.-L. (1999). Improved methods for tests of long-run abnormal
stock returns. Journal of Finance, 54, 165–201.
Matsumoto, M., & Nishimura, T. (1998). Mersenne twister: A 623-dimensionally equidistributed
uniform pseudorandom number generator. ACM Transactions on Modeling and Computer
Simulation, 8, 3–30.
Mitchell, M. L., & Stafford, E. (2000). Managerial decisions and long-term stock price
performance. Journal of Business, 73, 287–329.
14 Evaluating Long-Horizon Event Study Methodology 411

Noreen, E. W. (1989). Computer intensive methods for testing hypotheses: An introduction.

New York: Wiley.
Pearson, E. S., & Please, N. W. (1975). Relation between the shape of population distribution and
the robustness of four simple test statistics. Biometrika, 62, 223–241.
Schultz, P. (2003). Pseudo market timing and the long-run underperformance of IPOs. Journal of
Finance, 58, 483–517.
Sutton, C. D. (1993). Computer-intensive methods for tests about the mean of an asymmetrical
distribution. Journal of the American Statistical Association, 88, 802–808.
Viswanathan, S., & Wei, B. (2008). Endogenous events and long-run returns. Review of Financial
Studies, 21, 855–888.
The Effect of Unexpected Volatility Shocks
on Intertemporal Risk-Return Relation 15
Kiseok Nam, Joshua Krausz, and Augustine C. Arize

Contents
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
15.2 Theoretical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
15.3 Empirical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
15.3.1 Asymmetric Nonlinear Smooth Transition GARCH Model . . . . . . . . . . . . . . . . . . 418
15.3.2 Empirical Models for the Intertemporal Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
15.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
15.4.1 The Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
15.4.2 Estimation Results, Interpretations, and Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . 420
15.4.3 GJR Sample Period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
15.4.4 The Link Between Asymmetric Mean Reversion and
Intertemporal Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
15.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
Appendix 1: Method to Derive the Variance of the Parameters with Restrictions . . . . . . . . . . . 433
Appendix 2: News Impact Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
Appendix 3: Sign Bias Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435

Abstract
We suggest that an unexpected volatility shock is an important risk factor to
induce the intertemporal relation, and the conflicting findings on the relation
could be attributable to an omitting variable bias resulting from ignoring the
effect of an unexpected volatility shock on the relation. With the effect of an
unexpected volatility shock incorporated in estimation, we find a strong positive

K. Nam (*) • J. Krausz

Yeshiva University, New York, NY, USA
e-mail: knam@yu.edu; krausz@yu.edu
A.C. Arize
Texas A & M University-Commerce, Commerce, TX, USA
e-mail: chuck.arize@tamuc.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 413
DOI 10.1007/978-1-4614-7750-1_15,
# Springer Science+Business Media New York 2015
414 K. Nam et al.

intertemporal relation for the US monthly excess returns for 1926:12–2008:12.

We reexamine the relation for the sample period studied by Glosten, Jagannathan,
and Runkle (Journal of Finance 48, 1779–1801, 1993) and find that their sample
period is indeed characterized by a positive (negative) relation under a positive
(negative) volatility shock with the effect of a volatility shock incorporated in
estimation. We also find a significant link between the asymmetric mean reversion
and the intertemporal relation in that the quicker reversion of negative returns is
attributed to the negative intertemporal relation under a prior negative return shock.
For estimations we employ the ANST-GARCH model that is capable of
capturing the asymmetric volatility effect of a positive and negative return
shock. The key feature of the model is the regime-shift mechanism that allows
a smooth, flexible transition of the conditional volatility between different states
of volatility persistence. The regime-switching mechanism is governed by
a logistic transition function that changes values depending on the level of the
previous return shock. With a negative (positive) return shock, the conditional
variance process is described as a high (low)-persistence-in-volatility regime.
The ANST-GARCH model describes the heteroskedastic return dynamics more
accurately and generates better volatility forecasts.

Keywords
Intertemporal risk-return relation • Unexpected volatility shocks • Time-varying
rational expectation hypothesis • Stock market overreaction • Expected market
risk premium • Volatility feedback effect • Asymmetric mean reversion •
Asymmetric volatility response • Time-varying volatility • Volatility regime
switching • ANST-GARCH model

15.1 Introduction

The trade-off between risk and return is a core tenet in financial economics.
In particular, the intertemporal risk-return relation is a key element to explain the
predictable variation of expected asset returns.1 Despite its importance in asset
pricing, there has been a long-standing debate on the empirical sign of the
intertemporal relation, with findings that are mixed and inconclusive.
Criticisms of the mixed results often refer to a lack of conditional information. If
the predetermined conditional information set does not contain an important vari-
able that affects the risk-return trade-off, the econometric modeling of market
expectations suffers from the model misspecification problem and leads to
a wrong conclusion on the empirical nature of the intertemporal risk-return relation.

1
Fama and French (1989) argue that systematic patterns in the predictable variations of expected
returns are consistent with the intertemporal asset pricing model by Lucas (1978) and Breeden
(1979) and the consumption smoothing idea by Modigliani and Brumberg (1955) and Friedman
(1957). Ferson and Harvey (1991) and Evans (1994) also document the relative importance of the
time-varying risk premia to the conditional betas to explain predictable variations in expected
returns.
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 415

In this chapter, we suggest that an unexpected volatility shock is an important

risk factor to induce the intertemporal risk-return trade-off, and the conflicting
findings on the relation could be attributable to an omitting variable bias resulting
from ignoring the effect of an unexpected volatility shock on the relation. Incor-
porating the effect of an unexpected volatility shock in estimation of the relation,
we find a strong positive intertemporal relation between the expected market risk
premium and the predictable volatility for the US monthly excess returns of market
indices for the period of 1926:01–2008:12.
Conventional belief about the intertemporal relation is that a positive risk-return
relation is consistent with the time-varying rational expectation hypothesis in the
sense that a substantial amount of predictable variations of the expected risk premium
is induced by the risk-averse investors’ revision of their expectations in responding to
changing volatility. For example, Pindyck (1984) empirically shows that much of the
decline in stock prices during the 1970s in the US stock market is attributable to the
upward shift in risk premium arising from high stock market volatility.2 Studies that
support a positive relation include French et al. (1987), Fama and French (1988),
Ball and Kothari (1989), Turner et al. (1989), Harvey (1989), Cecchetti et al. (1990),
Haugen et al. (1991), Campbell and Hentschel (1992), Scruggs (1998), Kim
et al. (2001), Ghysel et al. (2005), Ludvigson and Ng (2007), Pastor et al. (2008),
Bali (2008), Darrat et al. (2011), and Nyberga (2012).
Although a positive intertemporal relation is consistent with Merton’s (1980)
dynamic CAPM, there is another side to the argument that the equilibrium asset
pricing does not necessarily imply a positive relation. Abel (1988) suggests that
a positive risk-return relation is consistent with the general equilibrium model only
when the coefficient of relative risk aversion is less than one. Barky (1989) suggests
that the directional effect of an increase in riskiness on stock prices depends on the
curvature of the utility function. Showing evidence of a strong negative relation for
their sample period, Glosten et al. (hereinafter GJR) (1993) suggest that both
positive and negative intertemporal relations are consistent with the equilibrium
asset pricing theory.3 Among others, Campbell (1987), Pagan and Hong (1989),
Breen et al. (1989), Nelson (1991), Backus and Gregory (1993), Harvey (2001), and
Ang et al. (2006) support a negative intertemporal relation.4

2
He suggested that a substantial portion of time variation in the expected risk premium is
associated with time-varying risk factors in investment opportunities.
3
They argue that investors may not require a large premium for bearing risk, but rather may reduce
the risk premium when they perceive exceptionally optimistic expectations on the future perfor-
mance of stock prices.
4
Brandt and Kang (2004) find that the conditional mean and volatility are negatively correlated
contemporaneously but positively correlated unconditionally due to the positive lead-lag relation
between the two moments of stock returns. Poterba and Summers (1986) suggested that, due to the
low level of volatility persistence, the volatility effect on the expected risk premium dissipates so
quickly that it cannot have a major effect on stock price movements. Note that there are some
studies that report weak evidence of the intertemporal relation. See Baillie and DeGennaro (1990),
Whitelaw (1994, 2000), Boudoukh et al. (1997), Yu and Yuan (2011), and M€ uller et al. (2011).
416 K. Nam et al.

A price shock generates two sources of forecasting errors: a return forecasting

error and a volatility forecasting error. Unlike a return forecasting error, the
unanticipated volatility shock has not been paid much attention by the literature.
We refer the volatility forecasting error to as an unexpected volatility shock. An
unexpected volatility shock is indeed an important risk factor to affect investors’
pricing behavior in that rational risk-averse investors revise their expectations
in responding not only to an underlying volatility but also to an unexpected
volatility shock. This implies that the intertemporal behavior of the expected
market risk premium is driven by both the predictable volatility and the unexpected
volatility change.
While the underlying return volatility has been widely examined by many
studies, the impact of an unexpected volatility shock has not been given much
attention by the literature. Almost all of the previous empirical studies on this
topic completely ignore the effect of an unexpected volatility shock on the
relation, so that their empirical results reflect only a partial intertemporal
risk-return relation. Thus, considering an unexpected volatility shock as the
conditional information in estimation, we reexamine the nature of the
intertemporal relation.
Especially, we examine an asymmetrical effect of a positive and
negative unexpected volatility shock on the relation. We conjecture that
a higher volatility level than predicted would increase the expected risk premium
and induce a stronger positive intertemporal relation. We define a positive
(negative) unexpected volatility shock as the case where actual market volatility
is higher (lower) than expected. Our estimation results show that a positive
unexpected volatility shock (the case in which the actual volatility is higher than
expected) causes a stronger positive intertemporal relation than does a negative
unexpected volatility shock (the case in which the actual volatility is lower than
expected).
The effect of an unexpected volatility shock considered in this chapter
is different from the volatility feedback effect proposed by Campbell and
Hentschel (1992). While the volatility feedback effect focuses on the contempo-
raneous effect of concurrent volatility shocks on the expected returns,
our volatility shock effect implies the consequence of a prior unexpected
volatility shock on the intertemporal relation. However, the estimation
of their contemporaneous volatility feedback effect is subject to the endogeneity
problem, as it is theoretically impossible to obtain the concurrent
volatility forecasting error due to an unavailability of the current volatility
information.5

5
To avoid the endogeneity problem, French et al. (1987) examined the volatility feedback effect
using ex post unexpected volatility changes. They found a strong negative relation between
unexpected returns and ex post unexpected volatility changes and interpreted it as evidence
supporting a positive intertemporal relation.
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 417

15.2 Theoretical Models

Merton (1973) suggests the intertemporal risk-return relation as a function of stock

market volatility, which can be specified in the following general form:



E Rmt
rf t jOt
1 Þ ¼ f spmt , p ¼ 1, 2, (15.1)

Where E() is the expectation operator, E(Rmt
rft|Ot
1) is the

time-varying expected market risk premium conditional on the information set
Ot
1, Rmt is the return on a stock market index portfolio, and rft is the risk-free
interest rate. The market volatility is represented by either smt or s2mt, where
s2mt ¼ Et
1[(Rmt
rft)
Εt
1(Rmt
rft)]2. Although @EðRmt@s

rf t jOt
1 Þ
2 > 0 is consistent
mt
with the equilibrium asset pricing theory, there has been a long-standing
controversy in the sign of the relation.
The intertemporal relation is driven not only by predictable volatility but also by
an unexpected volatility change. We thus consider the intertemporal relation in the
following form:

2 

EðRmt
rf t jOt
1 Þ ¼ f 1 s ^ 2mt
1 ,
^ mt þ f 2 e2mt
1
s (15.2)

where e 2mt
1 and s ^ 2mt
1 are the actual realized volatility series and the predicted
volatility series, respectively, such that e2mt
1
s ^ 2mt
1 represents a prior unexpected
volatility shock. Many studies employ the GARCH models to conditionally
estimate the predictable volatility series s ^ 2mt . Equation 15.2 implies that
@EðRmt
rf t jOt
1 Þ @f 1
consists of two components: @^ and @ e2 @f
^ 2
. While the first
@s2mt s 2mt ð mt
1 s 2mt
1 Þ
@f 1
term @^s 2 measures the effect of predictable market volatility on the relation, the
second term @ e2 @f
^
mt
2
captures the effects of an unexpected volatility shock on
ð mt
1 s 2mt
1 Þ
the intertemporal relation. However, most of the previous empirical studies
on this topic completely ignore the second term and focus only on the first
term. Consequently, their empirical results show only a partial intertemporal
relation.
Thus, we examine the empirical nature of the full intertemporal relation by
considering not only the predictable conditional volatility but also the effect of an
unexpected volatility shock on the relation. Especially, we define an unexpected
volatility shock in two separate cases of a positive and negative unexpected
volatility shock to examine an asymmetrical effect of a positive and negative
unexpected shock on the relation, if any. A positive (negative) unexpected volatility
shock is denoted as e2mt
1 > s ^ 2mt
1 (e2mt
1 < s
^ 2mt
1 ), implying that the actual market
volatility is higher (lower) than the predicted conditional market volatility. We then
examine the sign of intertemporal relation under each case of e2mt
1 > s ^ 2mt
1 and
emt
1 < s
2
^ mt
1 , respectively.
2
418 K. Nam et al.

We specify the linear form of intertemporal relation with a dummy variable to

capture the asymmetric effect of a positive and negative unexpected volatility shock
on the relation

^ mt þ l2 s
Rmt
rf t ¼ a þ l1 s ^ mt  dt þ emt , (15.3)

where emt is a series of white noise innovations, and s ^mt is the conditional standard
deviation of market portfolio. Note that we use s ^mt instead of s^mt
2
as the conditional
forecasts of stock market volatility. The use of s ^mt is suggested as the slope of the
capital market line by Merton (1980). Also, in estimation, using s ^mt is expected to
yield an improvement in the statistical efficiency, mainly due to a reduction in the
mean square error of the regression. dt is the dummy variable for a positive or
negative unexpected volatility shock. It takes the value 1 with a prior unexpected
^ 2mt
1 ) and 0 otherwise. The intertemporal
positive volatility shock (i.e., e2mt
1 > s
relation is thus measured by l1 + l2 when e2mt
1 > s ^ 2mt
1 with dt ¼ 1 or by l1
otherwise with dt ¼ 0. The asymmetrical effect of a positive and negative unex-
pected volatility shock on the intertemporal relation, if any, is captured by l2.
Specifically, l2 > 0 (or l1 + l2 > l1) implies that a positive volatility shock has
a positive impact on the intertemporal relation.

15.3 Empirical Models

15.3.1 Asymmetric Nonlinear Smooth Transition GARCH Model

To generate the forecast of time-varying market volatility, we employ the asym-

metric nonlinear smooth transition (ANST) GARCH model that is capable of
capturing the asymmetric volatility effect of a positive and negative return shock.
The key feature of the model is the regime-shift mechanism that allows a smooth,
flexible transition of volatility between different states of volatility persistence. For
monthly excess return series rt, we specify

   
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ, (15.4)

where F(et
1) ¼ {1 + exp[
g(et
1)]}
1 and et ¼ rt
E(rt|Ot
1) with et|Ot
1  N(0, ht).

pffiffiffiffi iid
Given et ¼ vt  ht , the normalized residuals are distributed as vt N ð0; 1Þ . The
logistic transition function F(et
1) is a smooth and continuous function of et–1 and
the speed parameter g and takes a value between 0 and 1: 0 < F(et
1) < 0.5 for
et
1 < 0, 0.5 < F(et
1) < 1 for et
1 > 0, and F(et
1) ¼ 0.5 for et
1 ¼ 0. The volatility
persistence is measured by (a1 + a2) + (b1 + b2)F, and the condition b1 + b2 < 0 captures
the excess volatility of a negative return shock. For any negative return shock that
causes 0 < F(et
1) < 0.5, the current volatility is described as a high-persistence-in-
volatility regime. For any positive return shock causing 0.5 < F(et
1) < 1, the current
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 419

volatility is described as a low-persistence-in-volatility regime. When et
1 ¼ 0,

F(et
1) ¼ 0.5, which implies that the current volatility ht is halfway between the upper
and lower volatility regimes. g governs the speed of transition between volatility
regimes. When g approaches 1, our model with b0 ¼ b2 ¼ 0 degenerates into the
GJR model.
Note that the volatility transition mechanism in GARCH models has
been applied in the several models, such as the modified GARCH model by
Glosten et al. (1993), the smooth transition GARCH model by Gonzalez-Rivera
(1998), the SVSARCH (sign- and volatility-switching ARCH) model by Fornari
and Mele (1997), and the MSVARCH (Markov switching volatility ARCH) model
by Turner et al. (1989) and Hamilton and Susmel (1994). For more details, see
Harvey (1993), Lutkepol (1993), Hamilton (1994), Campbell et al. (1997),
Gourieroux and Monfort (1997), and Rothman (1999).

15.3.2 Empirical Models for the Intertemporal Relation

We examine a simple linear form of the intertemporal relation in the following

model for monthly excess return series rt:
Model 1:
pffiffiffiffi
r t ¼ m þ fr t
1 þ d ht þ et
    (15.5)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,

where F(et
1) ¼ {1 + exp[
g(et
1)]}
1. We include the first-order autoregressive

term in the mean equation to capture the serial dependence in returns. The
intertemporal relation is measured by the coefficient d.
Model 1, however, ignores the asymmetric effect of an unexpected volatility shock
on the intertemporal relation; hence, it suffers from the omitting variable problem. The
estimate of d in Model 1 measures only a partial intertemporal relation. To measure
the full intertemporal relation, we present Model 2 in the following specification:
Model 2:
pffiffiffiffi
r t ¼ m þ fr t
1 þ ½d þ tMt Þ ht þ et
    (15.6)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,

where Mt is a dummy variable to capture the asymmetrical effect of a positive

and negative unexpected volatility shock on the intertemporal relation. We define


^e 2t
1
ht
1 > 0  ^e 2t
1
ht
1 < 0 as a positive (negative) unexpected volatility
shock, which implies that actual volatility is greater (less) than expected. It thus
takes a value 1 if ^e 2t
1 > ht
1 or 0 otherwise.
The sign of the intertemporal relation is measured by the sign of d + t under
a positive volatility shock (^e 2t
1 > ht
1), while it is measured by the sign of d under a
420 K. Nam et al.

negative volatility shock (^e 2t
1 < ht
1 ). Thus, t captures an asymmetrical effect of
a positive and negative volatility shock on the relation. We put an empirical focus
on t > 0 (or d + t > d), which indicates a positive impact of a positive unexpected
volatility shock on the intertemporal relation.6
We specify Model 3 to examine whether allowing the asymmetry for the
constant term affects the estimation of the intertemporal relation. A possibility is
that the asymmetrical effect of a positive and negative volatility shock on the
relation might disappear.
Model 3:
pffiffiffiffi
r t ¼ ðm1 þ m2 Mt Þ þ fr t
1 þ ðd þ tMt Þ ht þ et
    (15.7)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,
where m2Mt captures the asymmetric effect of a prior positive and negative
volatility shock on the level of the conditional mean return. In Model 3, we focus
on whether the estimated value of t is still statistically significant even with
a presence of m2Mt.

15.4 Empirical Results

15.4.1 The Data

We employ the excess market returns as the expected market risk premiums.
To generate the excess returns, we use the monthly nominal returns of the value-
and equal-weighted market portfolio index of the NYSE, AMEX, and NASDAQ
from the CRSP data files from 1926:01 to 2008:12. The monthly excess return series
is constructed by subtracting the 1-month US T-bill returns reported by Ibbotson
Associates from the monthly nominal index returns. The excess return series is
computed as percentage returns. We employ three sample periods: the full period
(1926:01–2008:12), the pre-87 Crash period (1926:01–1987:09), and the GJR period
(1951:04–1989:12). Table 15.1 reports the summary statistics for the data. The
descriptive statistics indicate that both the nominal and the excess returns series of
the two market indexes exhibit significant excess kurtosis and positive first-order
autocorrelation, characterizing the nonnormality of the short-horizon stock returns.

15.4.2 Estimation Results, Interpretations, and Diagnostics

We employ the maximum likelihood method with the analytical derivatives of each
parameter provided in the Gauss code. All the statistical inferences are based on
the Bollerslev-Wooldrige (1992) robust standard errors. Estimation results of

Note that t ¼ 0 supports the volatility irrelevance argument by Poterba and Summers (1986).
6
 15

Table 15.1 Summary statistics for monthly nominal and excess returns
Nominal value-weighted returns 1-monthT-bill rates Excess value-weighted returns
Statistics Full period Subperiod Full period Subperiod Full period Subperiod
Observations 996 741 996 741 996 741
Mean (100) 0.895 0.961 0.303 0.287 0.592 0.674
Std. dev. (100) 5.352 5.706 0.253 0.276 5.363 5.722
Skewness 0.18 0.368 1.104 1.261 0.227 0.406
Kurtosis 11.221 10.888 4.684 4.624 11.246 10.875
1st order Autocorrelation 0.112 (0.000) 0.107 (0.003) 0.966 (0.000) 0.967 (0.000) 0.115 (0.000) 0.111 (0.002)
The nominal return series are the monthly value-weighted market index returns for NYSE, AMEX, and NASDAQ stocks and were retrieved from the CRSP
tapes for the period from 1926:01 to 2008:12. The monthly nominal return series are the value-weighted market indexes retrieved from the CRSP tapes for the
same. The monthly excess return series is computed by subtracting 1-month treasury bill returns as reported by Ibbotson Associates from the nominal returns.
All returns are computed as percentage value. The analysis employs the full period (1926:01–2008:12) and the pre-87 Crash period (1926:01–1987:09). The
value in the parentheses is the p-value for the Ljung-Box Q test
The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation
421
422 K. Nam et al.

Model 1 for the full period (1926:01–2008:12) and the pre-87 Crash period
(1926:01–1987:09) are reported in Table 15.2. Model 1 examines the sign of the
simple linear relation between the expected risk premium and the predictable
market volatility. Estimation results for Model 1 show that the intertemporal
relation in a simple linear form is significantly positive for both the full sample
and subsample periods. The estimated value of d is 0.077 (0.080 for the pre-87
Crash period) and statistically significant at the 1 % level for both sample periods.
With regard to the conditional variance equation, the asymmetric
volatility response of a positive and negative return shock is well captured by
b1 + b2 < 0 with a statistical significance. The average estimated value of ht for
the full sample period under a lower volatility regime is 36.37, while it is 24.00
under a higher volatility regime. Also, the estimation results show a high estimated
value of the transition parameter g, which indicates that the transition between
volatility regimes occurs very quickly. This implies that the volatility regime is
divided into only two extreme regimes: the upper and lower volatility regimes. The
upper (lower) regime is induced by any negative (positive) return shock and
exhibits a high volatility persistence (a low volatility persistence) in the conditional
volatility process.
Estimation results of Model 2 are reported in Table 15.2. There are several
notable findings. First, a positive volatility shock has a positive impact on the
intertemporal relation. The results show that the estimated value of t is significantly
positive (0.052 for the full period and 0.041 the pre-87 Crash period) and statisti-
cally significant at the 1 % level for both sample periods. This implies that an
unexpected volatility shock is priced such that, for a positive volatility shock, the
expected risk premium for the full sample period increases p byffiffiffiffi5.2 % (4.1 % for the
pre-87 Crash period) of the predicted conditional volatility ht. Second, the results
show that the intertemporal coefficients are all positive for both sample periods.
The estimated value for the full period d + t ¼ 0.0878 is (0.093 for the pre-87 Crash
period) under prior positive volatility shock, while it is d ¼ 0.035 for the full period
and 0.052 for the pre-87 Crash period under prior negative volatility shock. The
results indicate that the sign of the intertemporal relation is indeed positive when
the effect of an unexpected volatility shock is incorporated in estimation. Third, the
estimation result of t > 0 indicates that the slope of the capital market line is
relatively steeper under a positive volatility shock than under a negative volatility
shock. This implies that a positive volatility shock increases the degree of risk
aversion.
Estimation results of Model 3 are also reported in Table 15.2. In Model 3, we
examine the possibility that the asymmetrical effect of a positive and
negative volatility shock may disappear under the presence of asymmetry in the
constant term allowed in the conditional mean equation. The estimation
results show a positive value of m2 for both periods, indicating that a prior
positive volatility shock raises the conditional mean returns. This asymmetrical
effect of a prior volatility shock on the constant term is more profound for the
pre-87 Crash period (m2 is 0.351 and statistically significant at the 1 % level).
With respect to the intertemporal relation, t is still positive and statistically
 15

Table 15.2 Estimation results of models 1–3

Model 1 Model 2 Model 3
Full period Subperiod Full period Subperiod Full period Subperiod
m1 0.420 (3.315) 0.366 (1.485) 0.150 (2.233) 0.251 (2.356) 0.116 (1.224) 0.168 (1.152)
m2 0.109 (1.163) 0.351 (3.250)
f 0.077 (6.006) 0.080 (3.732) 0.041 (3.015) 0.072 (8.384) 0.030 (3.013) 0.051 (2.893)
d 0.073 (4.749) 0.080 (3.126) 0.035 (2.426) 0.052 (2.610) 0.033 (1.915) 0.048 (2.115)
t 0.052 (5.756) 0.041 (4.015) 0.054 (6.913) 0.042 (2.533)
a0 0.005 (0.779) 0.001 (1.813) 0.030 (0.300) 0.007 (0.186) 0.022 (0.567) 0.022 (1.306)
a1 0.131 (3.324) 0.142 (2.689) 0.118 (3.352) 0.141 (3.586) 0.119 (3.438) 0.141 (3.491)
a2 1.028 (25.817) 1.028 (25.009) 1.063 (26.978) 1.025 (24.875) 1.057 (26.108) 1.027 (25.225)
b0 2.555 (2.306) 2.855 (2.457) 2.696 (2.816) 2.782 (3.233) 2.652 (2.776) 2.707 (3.367)
b1
0.055 (
0.820)
0.087 (
0.951)
0.038 (
0.682)
0.085 (
1.487)
0.038 (
0.695)
0.081 (
0.834)
b2
0.358 (
3.539)
0.368 (
3.846)
0.408 (
4.643)
0.354 (
3.916)
0.397 (
4.329)
0.364 (
3.142)
g 137.382 (2.057) 97.516 (1.090) 123.324 (3.044) 85.377 (1.577) 107.355 (2.380) 99.011 (1.330)
This table presents the maximum likelihood estimates for models 1
3 for monthly excess returns of the value-weighted index for the NYSE, AMEX, and
NASDAQ stocks. The estimation employs the full period (1926:01–2008:12) and the pre-87 Crash period (1926:01–1987:09). For the monthly excess return
series rt, each model is specified
pffiffiffiffias follows
Model 1: r t ¼ m1 þ fr t
1 þ d ht þ et pffiffiffiffi
Model 2: r t ¼ m1 þ fr t
1 þ ðd þ tMt Þ ht þ etpffiffiffiffi
Model 3: r t ¼ m1 þ m2 Mt þ fr t
1 þ ðd þ tMt Þ ht þ et ,
where the indicator function Mt is specified to capture the asymmetric effect of an unexpected volatility shock on the intertemporal relation, such that Mt ¼ 1 if
2
^e 2t
1 > ht
1 or Mt ¼ 0 otherwise. The conditional variance equation for the models is specified as ht ¼ [a0 + a1et
1 + a2ht
1] + [b0 + b1e2t
1 + b2ht
1]F(et
1),
where the transition function F(et
1) is defined as F(et
1)¼{1 + exp[
g(et
1)]}
1. The values in parentheses are the Bollerslev-Wooldridge robust t-statistics
The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation
423
424 K. Nam et al.

significant at the 1 % level for both periods (t1 is 0.054 for the full period and 0.048
for the pre-87 Crash period). This implies that, even with the asymmetry allowed in
the constant term, there still exists a significant asymmetrical effect of
a positive and negative volatility shock on the intertemporal relation. Also, d is
still significantly positive for both periods, confirming the positive full
intertemporal relation.
Table 15.3 reports the summary of diagnostics for the estimation results, such as
skewness, kurtosis, the Jarque-Bera normality test, and the Ljung-Box Q test on
the normalized and the squared normalized residuals. The Ljung-Box Q statistics
on the normalized residuals checks serial correlation in the residuals. Rejection
of the null of no autocorrelation up to a certain lag length indicates that either
the dynamics of the conditional mean or the lag structure of the conditional
variance process is not well specified or that both equations are not well specified
by the model. The Ljung-Box statistics on the squared normalized residuals
ascertains if the serial dependence in the conditional variance is well captured by
the equation.
We also perform the negative sign bias test (NSBT) suggested by Engle and Ng
(1993) to examine the ability of the model to capture the so-called leverage effect
of a negative return shock on the conditional variance process. The negative sign
bias test shows insignificant t-values for all estimations and indicates that the
asymmetric volatility response to a positive and negative return shock is well
captured by the ANST-GARCH model.7 The Ljung-Box Q(10) test indicates that
the serial dependence of the conditional mean and variance is well captured by
models 1–3.

15.4.3 GJR Sample Period

Using modified (E)GARCH-M models, Glosten et al. (1993) report a strong

negative intertemporal relation for their (GJR) sample period of
1951:04–1989:12. However, they do not consider the effect of an unanticipated
volatility shock on the estimation of the relation, so that their results reflect only
a partial intertemporal relation. We thus evaluate the full intertemporal relation for
their sample period by incorporating the effect of an unexpected volatility shock in
the estimation of the relation. Allowing a time dummy in the conditional mean
equation to capture the characteristic of the GJR sample period, we specify Model
5 as follows:

7
The negative sign

bias
pffiffiffiffitest is performed with the regression equation v2t ¼ a + bS
0 *
t
1et
1 + p zt +
et, where v2t ¼ et = ht . S

2
t
1 ¼ 1 if et
1 < 0, and S

t
1 ¼ 0 otherwise. Also, z 
¼ e
h ð C Þ=h t,
t
where eh ðCÞ ¼ @ht =@C evaluated at the values of maximum likelihood estimates of parameter C.
The test statistic of the NSBT is defined as the t-ratio of the coefficient b in the regression.
A statistically significant t-value implies the failure of the model to absorb the effect of sign bias
and indicates that the volatility model considered is misspecified.
 15

Table 15.3 Diagnostics of models 1–3

Model 1 Model 2 Model 3
Full period Subperiod Full period Subperiod Full period Subperiod
Skewness of vt
0.610
0.372
0.631
0.375
0.628
0.366
Kurtosis of vt 4.961 3.909 5.013 3.877 4.983 3.850
JB-Normality 220.845 (0.000) 42.531 (0.000) 233.794 (0.000) 41.003 (0.000) 228.080 (0.000) 38.743 (0.000)
Q(10) on vt 11.209 (0.341) 17.801 (0.058) 13.368 (0.204) 18.010 (0.055) 13.139 (0.216) 17.663 (0.061)
Q(10) on v2t 7.156 (0.711) 12.905 (0.229) 9.660 (0.471) 13.148 (0.215) 9.516 (0.484) 13.639 (0.190)
NSBT on ht
1.328 (0.184)
0.579 (0.563)
0.526 (0.599)
1.275 (0.203)
0.688 (0.492)
1.365 (0.173)
This table presents a summary pffiffiffiffi of diagnostics on the normalized residuals and the squared normalized residuals from the estimations. The normalized residual
series is defined as vt ¼ et = ht. JB-Normality refers to the Jarque-Bera normality test statistic, which is distributed as w2 with two degrees of freedom under the
null hypothesis of normally distributed residuals. Q(10) is the Ljung-Box Q(10) test statistic for checking serial dependence in the normalized residuals and the
squared normalized residuals from the estimations. NSBT refers to the negative sign bias test suggested by Engle and Ng (1993). It is a diagnostic test that
examines the ability of the specified model to capture the so-called leverage effect
of apnegative volatility process. The test is

ffiffiffiffi2
return shock on the conditional


performed with the regression equation v2t ¼ a + bSt
1 et
1 + p0 z*t + et, where v2t ¼ et = ht . St
1 ¼1 if et
1 < 0, and St
1 h ðCÞ=ht,
¼ 0 otherwise. Also, zt ¼ e
where e h ðCÞ ¼ @ht =@C, is evaluated at the values of the maximum likelihood estimates of parameter C. The test statistic of the NSBT is defined as the t-ratio
of the coefficient b in the regression. The value in the parentheses is the p-value of the individual test statistics considered
The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation
425
426 K. Nam et al.

Model 4 (Modified Model 1 for GJR Period):

pffiffiffiffi
r t ¼ m þ fr t
1 þ d þ dG Gt ht þ et
 
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ yrf t (15.8)
 
þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,

where Gt is a dummy variable that takes a value 1 for the GJR sample period or
0 otherwise. The coefficient dG captures the differential effect of the GJR period, if
any, on the intertemporal relation, while d measures the relation for the full period.
The negative intertemporal relation reported by Glosten et al. (1993) can be
confirmed by d + dG < 0. One of the important features of the above modified
Model 1 is its capacity to distinguish the relation for the GJR sample period from
that for the entire sample period. We also estimate the same model with and without
the 1-month T-bill returns rft included in the conditional variance equation.8
Estimation results of Model 4 are reported in Table 15.4. A notable finding is
that the estimated value of dG is strongly negative (–0.103 with rft and –0.092
without rft) and highly significant with d + dG < 0. This result implies that,
comparing to the full period, the GJR sample period is especially characterized
by a strong negative intertemporal relation, and this result is consistent with that of
Glosten et al. (1993).
As mentioned earlier, however, this result does not consider the effect of an
unexpected volatility shock on the relation, reflecting only a simple partial
intertemporal relation. In order to examine the full relation for the GJR sample
period, we specify Model 5 as follows:
Model 5 (Modified Model 2 for GJR Period):


 pffiffiffiffi
r t ¼ m þ fr t
1 þ ðd þ tMt Þ þ dG þ tG Mt Gt ht þ et
    (15.9)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ yrf t þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,

where Gt is a time dummy variable that takes the value 1 for the GJR period
or 0 otherwise. The full intertemporal relation for the GJR period is measured by
d + dG + t + tG under a positive volatility shock and by d + dG under a negative
volatility shock, such that the differential effect of the GJR period on the full
relation is captured by dG + tG. We also estimate the same model with and without
rft included in the conditional variance equation.
Estimation results of Model 5 for the GJR sample period are reported
in Table 15.4. It shows that the estimated value of all four important parameters
(d, t, dG, and tG) capturing the intertemporal relation are statistically significant
at the 1 % level. There are several notable findings. First, the estimation result of

8
Several studies show that the estimation results are sensitive to the inclusion of 1-month T-bill
return in the conditional variance equation. See Campbell (1987), Glosten et al. (1993), and
Scruggs (1998).
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 427

Table 15.4 Estimation results of models 4 and 5 for the GJR sample period
Model 4 for GJR sample period Model 5 for GJR sample period
Coeff. With rf Without rf With rf Without rf
m 0.546 (9.944) 0.443 (9.288) 0.531 (6.136) 0.543 (5.166)
f 0.056 (5.930) 0.072 (7.909) 0.051 (5.143) 0.051 (5.394)
d 0.059 (5.925) 0.080 (6.392) 0.103 (4.164) 0.100 (3.678)
t 0.046 (4.312) 0.046 (3.778)
dG
0.103 (
12.369)
0.092 (
8.882)
0.155 (
12.540)
0.156 (
11.720)
tG 0.067 (6.560) 0.091 (6.903)
a0 0.001 (0.766) 0.001 (0.581) 0.007 (0.238) 0.014 (0.202)
a1 0.110 (3.199) 0.110 (3.291) 0.104 (3.359) 0.102 (3.303)
a2 1.060 (25.142) 1.072 (28.727) 1.061 (26.070) 1.071 (29.469)
b0 2.482 (2.564) 3.007 (3.464) 2.270 (2.538) 2.689 (2.971)
b1
0.029 (
0.457)
0.042 (
0.795)
0.019 (
0.337)
0.021 (
0.399)
b2
0.412 (
4.672)
0.428 (
4.883)
0.404 (
4.420)
0.416 (
4.840)
y 0.885 (1.065) 0.901 (1.128)
g 121.735 (1.944) 217.354 (2.441) 287.546 (2.128) 254.235 (1.963)
The GJR sample period (1951:04–1989:12) is evaluated by incorporating the period as a dummy
variable in the mean equation, with the 1-month T-bill rates included or excluded in the condi-
tional variance equation. The modified models to estimate the GJR

samplep period
ffiffiffiffi are as follows
Model 4 (modified Model 1 for GJR period): r t ¼ m þ fr t
1 þ d þ dG Gt
ht þ et  pffiffiffiffi
Model 5 (modified Model 2 for GJR period): r t ¼ m þ fr t
1 þ ðd þ tMt Þ þ dG þ tG Mt Gt ht þ et
where the indicator function Gt is a dummy variable for the GJR sample period, and rft is
the yield on 1-month T-bill from Ibbotson Associates. The function Mt is specified to capture
the asymmetric effect of an unexpected volatility shock on the relation, such that Mt ¼ 1 if ^e 2t
1
> ht
1 ; otherwise, Mt ¼ 0. The conditional variance equation for the two estimation
models is specified as ht ¼ [a0 + a1e2t
1 + a2ht
1] + [b0 + b1e2t
1 + b2ht
1]F(et
1), where
F(et
1) ¼ {1 + exp[
g(et
1)]}
1. The values in parentheses are the Bollerslev-Wooldridge robust
t-statistics

dG + tG ¼
0.088 (–0.065 without rft) confirms that the GJR period is

indeed characterized by a significant negative relation. Second, the estimation
result of t + tG ¼ 0.113 (0.137 without rft) implies that, even for the GJR
sample period, there still exists a significant positive effect of a positive
volatility shock on the full intertemporal relation. Third, the estimation results of
d + dG + t + tG ¼ 0.061 (0.081 without rft) and d + dG ¼
0.052 (–0.056
without rft) imply that the GJR sample period exhibits a positive intertemporal
relation under a positive volatility shock and a negative relation under a negative
volatility shock. Note the above results are not sensitive to the inclusion of rft in
estimation.
Diagnostic tests in Table 15.5 indicate that all the estimations pass the
Ljung-Box Q(10) test on the normalized and squared normalized residuals. This
result implies that there is no serial dependence remaining in the conditional mean
and variance processes. The negative sign bias test shows insignificant t-values for
all estimations, indicating that the estimated conditional variance process well
captures the excess volatility response caused by a negative return shock.
428 K. Nam et al.

Table 15.5 Diagnostics of the models 4 and 5 for the GJR sample period
Model 4 for GJR sample period Model 5 for GJR sample period
Without rf With rf Without rf Without rf
Skewness of vt
0.636
0.623
0.646
0.626
Kurtosis of vt 5.251 5.083 5.216 5.018
JB-Normality 246.664 (0.000) 217.536 (0.000) 242.89 (0.000) 208.29 (0.000)
Q(10) on vt 10.190 (0.424) 10.397 (0.406) 10.111 (0.431) 10.846 (0.370)
Q(10) on v2t 10.514 (0.397) 10.418 (0.405) 9.7167 (0.466) 10.122 (0.430)
NSBT on ht
1.198 (0.231)
1.162 (0.246)
1.205 (0.228)
1.105 (0.269)
This table presents a summary of diagnostics on the normalized residuals and the
squared normalized residuals from the estimations. The normalized residual series is defined as
pffiffiffiffi
vt ¼ et = ht. JB-Normality refers to the Jarque-Bera normality test statistic, which is distributed as
w with two degrees of freedom under the null hypothesis of normally distributed residuals.
2

Q(10) is the Ljung-Box Q(10) test statistic for checking serial dependence in the normalized
residuals and the squared normalized residuals from the estimations. NSBT refers to the negative
sign bias test suggested by Engle and Ng (1993). It is a diagnostic test that examines the ability of
the specified model to capture the so-called leverage effect of a negative return shock on the
conditional volatility process. The test is performed with the regression equation v2t ¼ a +



pffiffiffiffi2


bSt
1 et
1 + p0 z*t + et, where v2t ¼ et = ht . St
1 ¼ 1 if et
1 < 0, and St
1 ¼ 0 otherwise. Also,
 e e
zt ¼ h ðCÞ=ht , where h ðCÞ ¼ @ht =@C , is evaluated at the values of the maximum likelihood
estimates of parameter C. The test statistic of the NSBT is defined as the t-ratio of the coefficient
b in the regression. The value in the parentheses is the p-value of the individual test statistics
considered

15.4.4 The Link Between Asymmetric Mean Reversion and

Intertemporal Relation

It has been known that the expected market returns exhibit an asymmetric mean-
reverting pattern that negative returns are more likely to revert to positive returns
than positive returns reverting to negative returns. However, the quicker reversion
of negative returns is hardly justified under a positive intertemporal relation. Under
a positive intertemporal relation, a negative return shock raises the risk premium to
compensate for the excess volatility, and an increase in risk premium reduces the
current stock price, which in turn reduces the concurrent stock price. Thus, if
a positive intertemporal relation is correct, a negative return should be more likely
to be accompanied by another negative return for the subsequent periods.
Nam et al. (2001) suggest that the quicker reversion of a negative return is
attributed to a reduction in the expected risk. They show that the intertemporal
relation is significantly negative under a prior negative return shock. They argue
that a negative return shock generates an optimistic expectation by investors, of
the future performance of a stock experiencing a recent price drop, thereby reducing
the expected market risk premium. As a reduction in risk premium in turn raises the
current stock price, negative returns are more likely to revert to positive returns.
While successfully explaining the link between the asymmetric mean-reverting
property and the intertemporal relation, Nam et al. (2001) do not consider the effect
of an unexpected volatility shock on the link. In this section, we investigate the
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 429

impact of an unexpected volatility shock on the observed link. First, we specify the
following nonlinear autoregressive model to confirm the asymmetric mean rever-
sion of the expected market risk premium:
Model 6:

r t ¼ ½m1 þ m2 Fðet
1 Þ þ ½f1 þ f2 Fðet
1 Þr t
1 þ et

    (15.10)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,

where the asymmetry is allowed in both the conditional mean and variance, such
that the asymmetry in both processes is controlled by a prior return shock.
The return serial correlation varies between f1 and f1 + f2 depending on the
value of the transition function F(et
1). Under an extreme negative prior return
shock that causes F(et
1) ¼ 0, serial correlation is measured by f1, while it is
measured by f1 + f2 for an extreme positive return shock yielding F(et
1)¼ 1.9
A quicker reversion of a negative return is captured by f1 > 0 (or f1 + f2 > f1).
Note that the condition f1 < 0 (a negative serial correlation under et
1 < 0)
indicates a stronger reverting tendency of a negative return.
Estimation results of Model 6 are reported in Table 15.6. It shows that the
estimated value of f2 is positive and highly significant for both the full
period and the pre-87 Crash period. The measured serial correlation is negative
(f1 ¼
0.078 and –0.076, respectively, for the two periods) under a prior negative
return shock, while it is positive (f1 + f2 ¼ 0.056 and 0.094, respectively, for the
two periods) under a prior positive return shock. This result confirms the asymmet-
rical reverting pattern of the expected returns that a negative return reverts more
quickly, while a positive return tends to persist.
Secondly, we specify Model 7 to examine if there is a link between the
asymmetric mean reversion and the intertemporal relation.
Model 7:
pffiffiffiffi
r t ¼ ½m1 þ m2 Fðet
1 Þ þ f1 r t
1 þ ½d1 þ d2 Fðet
1 Þ ht þ et
    (15.11)
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ:

The partial intertemporal relation is measured by the estimated value of d1 under

a prior negative return shock causing F(et
1) ¼ 0, while it is measured by d1 + d2
under a prior positive return shock causing F(et
1) ¼ 1. Thus, d2 measures the
differential impact of a positive and negative return shock on the partial
intertemporal relation.
The estimation results presented in Table 15.6 show two notable findings. First,
d2 is positive and statistically significant at the 1 % level (d2 ¼ 0.258 for the full
period and 0.334 for the pre-87 Crash period). This result implies that there is

9
Stationarity condition of rt is satisfied with |f1 + f2F(et
1)| < 1, i.e., |f1| < 1 for et
1 < 0 or
|f1 + f2| < 1 for et
1 > 0.
 430

Table 15.6 Estimation results of models 6–8

Model 6 Model 7 Model 8
Coef. Full period Subperiod Full period Subperiod Full period Subperiod
m1 0.464 (13.257) 0.489 (9.673) 0.486 (2.569) 0.614 (4.531) 0.972 (6.584) 0.879 (2.429)
m2 0.074 (1.667)
0.169 (
3.267)
0.034 (
0.147)
0.229 (
0.946)
0.181 (
1.172) 0.092 (0.264)
f1
0.078 (
8.476)
0.076 (
11.278)
0.058 (
7.059)
0.070 (
5.170)
0.023 (
2.424)
0.058 (
5.008)
f2 0.134 (5.972) 0.170 (10.188)
d1
0.155 (
3.251)
0.182 (
5.013)
0.260 (
4.592)
0.244 (
2.610)
t1 0.079 (2.277) 0.024 (1.111)
d2 0.258 (4.070) 0.334 (5.502) 0.229 (3.489) 0.225 (2.329)
t2
0.028 (
0.659) 0.052 (1.983)
a0 0.001 (0.259) 0.000 (1.115) 0.002 (1.149) 0.001 (0.997) 0.009 (0.982) 0.012 (1.833)
a1 0.122 (3.656) 0.130 (3.581) 0.133 (3.674) 0.133 (3.774) 0.123 (3.762) 0.139 (3.776)
a2 1.012 (31.079) 1.011 (29.957) 1.063 (18.379) 1.081 (24.110) 1.060 (35.832) 1.066 (22.949)
b0 2.154 (2.598) 2.331 (2.998) 2.279 (3.102) 2.339 (3.943) 1.914 (3.139) 1.966 (3.483)
b1
0.030 (
0.526)
0.076 (
1.218)
0.054 (
1.036)
0.070 (
1.329)
0.028 (
0.503)
0.061 (
1.039)
b2
0.312 (
4.200)
0.306 (
4.391)
0.383 (
3.406)
0.405 (
4.893)
0.368 (
5.506)
0.379 (
4.390)
g 126.008 (3.864) 131.391 (2.818) 157.246 (3.609) 216.327 (3.387) 148.326 (2.095) 67.239 (3.027)
This table presents the maximum likelihood estimates of models 6–8 for the monthly excess returns of the value-weighted index for the NYSE, AMEX, and
NASDAQ stocks over the period of 1926:01–2008:12 and the period of 1926:01–1987:09. For the monthly excess return series rt, each model is specified as
follows
Model 6: rt ¼ [m1 + m2F(et
1)] + [f1 + f2F(et
1)]rt
1 + et pffiffiffiffi
Model 7: r t ¼ ½m1 þ m2 Fðet
1 Þ þ fr t
1 þ ½d1 þ d2 Fðet
1 Þ ht þ et pffiffiffiffi
Model 8: r t ¼ ½m1 þ m2 Fðet
1 Þ þ f1 r t
1 þ ½ðd1 þ t1 Mt Þ þ ðd2 þ t2 Mt ÞFðet
1 Þ ht þ et
where the indicator function Mt is specified to capture the asymmetric effect of an unexpected volatility shock on the intertemporal relation, such that Mt ¼ 1 if
^e 2t
1 > ht
1 or Mt ¼ 0 otherwise. The conditional variance equation for Models 6–8 is specified as ht ¼ [a0 + a1et2
1 + a2ht
1] + [b0 + b1et2
1 + b2ht
1]
F(et
1). The values in parentheses are the Bollerslev-Wooldridge robust t-statistics
K. Nam et al.
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 431

a strong asymmetry in the partial intertemporal relation under a prior positive and
negative return shock. Second, the partial intertemporal relation is indeed negative
under a prior negative return shock (d1 ¼
0.155 for the full period and –0.182
for the pre-87 Crash period), while it is positive under a prior positive return shock
(d1 + d2 ¼ 0.103 for the full period and 0.152 for the pre-87 Crash period). This result
implies that while a positive return shock complies with the conventional
positive intertemporal relation, a negative return shock indeed induces a negative
intertemporal behavior of the expected market returns. More importantly, the results
of Models 6 and 7 confirm the highly significant asymmetric link between the mean
reversion and the intertemporal relation. The quicker reversion of negative returns is
attributed to the negative intertemporal relation under a prior negative return shock.
Model 7 does not incorporate the asymmetrical impact of a positive and negative
volatility shock in the estimation of the intertemporal relation. To get empirically
more reliable results on the link between the mean reversion and the full
intertemporal relation, we propose Model 8 in the following specification.
Model 8:
pffiffiffiffi
r t ¼ ½m1 þ m2 Fðet
1 Þ þ f1 r t
1 þ ½ðd1 þ t1 Mt Þ þ ðd2 þ t2 Mt ÞFðet
1 Þ ht þ e t
   
ht ¼ a0 þ a1 e2t
1 þ a2 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1 Fðet
1 Þ,
(15.12)

where the full intertemporal relation is measured separately under a positive and
negative return shock. When a negative return shock causing F(et
1) ¼ 0 is realized,
the full intertemporal relation is measured by d1 + t1, for which t1 captures an
asymmetrical impact of a prior unexpected volatility shock on the relation. With
a prior positive return shock, the relation is measured by d1 + d2 + t1 + t2, for which
t1 + t2 measures an asymmetrical effect of a prior volatility shock on the relation, if any.
The estimation results of Model 8 are reported in Table 15.6. There are several
notable findings. First, the results show that three out of the four important parameters
(d1, t1, d2, and t2) to capture the full intertemporal relation are statistically significant
at the 5 % level. Second, the result of d1 + t1 ¼
0.181 (–0.220 for the pre-87 Crash
period) and d1 + d2 + t1 + t2 ¼ 0.020 (0.057 for the pre-87 Crash period) implies that
the full intertemporal relation is still negative (positive) under a prior negative
(positive) return shock. Third, more importantly, the results of t1 ¼ 0.079 and t1 +
t2 ¼ 0.051 (0.024 and 0.076 for the pre-87 Crash period) indicate that a positive
volatility shock has a positive impact on the intertemporal relation, regardless of the
sign of a prior return shock. This implies that a positive unexpected volatility shock
consistently induces a positive impact on the intertemporal risk-return relation.
Table 15.7 reports the results of diagnostic tests. The Ljung-Box Q(10) test
results on vt and v2t indicate that serial dependence is well captured by the
specified conditional mean and variance processes. The negative sign bias test
shows insignificant t-values for all estimations, confirming the capability of the
ANST-GARCH model to capture the excess volatility response caused by
a negative return shock.
 432

Table 15.7 Diagnostics of models 6–8

Model 6 Model 7 Model 8
Full period Subperiod Full period Subperiod Full period Subperiod
Skewness of vt
0.627
0.368
0.620
0.343
0.592
0.320
Kurtosis of vt 5.071 3.917 5.111 3.764 4.775 3.725
JB-Normality 242.843 (0.000) 42.563 (0.000) 248.414 (0.000) 32.511 (0.000) 188.671 (0.000) 28.843 (0.000)
Q(10) on vt 14.059 (0.170) 20.323 (0.026) 10.852 (0.369) 17.070 (0.073) 9.669 (0.470) 15.935 (0.101)
Q(10) on v2t 8.379 (0.592) 9.902 (0.449) 10.005 (0.440) 13.429 (0.201) 10.095 (0.432) 13.714 (0.186)
NSBT on ht
1.415 (0.157)
0.769 (0.442)
1.058 (0.290)
0.547 (0.585)
1.248 (0.212)
0.334 (0.738)
This table presents a summary pffiffiffiffi of diagnostics on the normalized residuals and the squared normalized residuals from the estimations. The normalized residual
series is defined as vt ¼ et = ht. JB-Normality refers to the Jarque-Bera normality test statistic, which is distributed as w2 with two degrees of freedom under the
null hypothesis of normally distributed residuals. Q(10) is the Ljung-Box Q(10) test statistic for checking serial dependence in the normalized residuals and the
squared normalized residuals from the estimations. NSBT refers to the negative sign bias test suggested by Engle and Ng (1993). It is a diagnostic test that
examines the ability of the specified model to capture the so-called leverage effect of a negative return shock on the conditional volatility process. The test is
2
0 *

pffiffiffiffi2


performed with the regression equation v2t ¼ a + bS
e
t
1et
1 + p zt + et, where vt ¼ et = ht . St
1 ¼ 1 if et
1 < 0, and St
1 ¼ 0 otherwise. Also, zt ¼ h ðCÞ=ht,
where e h ðCÞ ¼ @ht =@C, is evaluated at the values of the maximum likelihood estimates of parameter C. The test statistic of the NSBT is defined as the t-ratio
of the coefficient b in the regression. The value in the parentheses is the p-value of the individual test statistics considered
K. Nam et al.
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 433

15.5 Conclusions

We suggest that the intertemporal risk-return relation is driven not only by the
underlying market volatility but also by an unexpected volatility shock. Most of the
previous literature on this topic ignores the effect of an unexpected volatility shock
on the relation. Thus, their results reflect only a partial intertemporal relation. With
the effect of an unexpected volatility shock incorporated in the estimation of the
relation, we find a strong positive intertemporal relation for the US monthly excess
returns for the period of 1926:01–2008:12.
We also reexamine the relation for the GJR sample period with the effect of
a volatility shock incorporated in estimation. The estimation results show that the
GJR sample period is indeed characterized by a strong positive (negative) relation
under a positive (negative) volatility shock. This implies that the negative relation
reported by Glosten et al. (1993) is attributed to ignoring the effect of an unexpected
volatility shock on the relation.
Lastly, we examine the observed link between the mean reversion property and
the intertemporal relation under a consideration of the impact of an unexpected
volatility shock on the link. We confirm that the quicker reversion of negative
returns is attributed to the negative intertemporal relation under a prior negative
return shock. We interpret this negative intertemporal relation as reflective of
strong optimistic expectations as perceived by investors of the future performance
of a stock experiencing a recent price drop.

Appendix 1: Method to Derive the Variance of the Parameters

with Restrictions

In estimations, we employ the parameter restrictions for the positivity of the

conditional variance process. To get the true variance of the restricted parameter,
we apply the following functional transformations:
Case 1. Logistic Function Transformation
Suppose that a parameter b^ is restricted to a logistic function of ^b in order to
guarantee 0 < b^ < 1.
Hence,

 1
b^ ¼ f ^b ¼ , (15.13)
1 þ e
b^

where ^b is the actual coefficient estimated from computation, and b^ is the true
parameter estimate we want to transform from Eq. 15.13. Using Taylor expan-
sion, Eq. 15.13 can be expressed as



f ^b ¼ f ðbÞ þ f 0 ðbÞ ^b
b , (15.14)

or
434 K. Nam et al.



f b^
f ðbÞ ¼ f 0 ðbÞ ^b
b : (15.15)

Using f ðbÞ ¼ 1þe1
b ¼ b and Eq. 15.15 can be expressed as


b^
b ¼ f 0 ðbÞ  ^b
b : (15.16)
 
Then, from Eq. 15.16 we can get var b^ as follows:
   2
2
var b^ ¼ E b^
b ¼ ½f 0 ðbÞ  E ^b
b ,
2
(15.17)

where the true value of f 0 (b) is not known. We thus use the MLE of b, ^b, to get
the value of f 0 (b). Then the variance of b^ can be calculated as follows:
  
 2

var b^ ¼ f 0 ^b  var ^b , (15.18)




where f 0 ^
b ¼ f ^b 1
f ^b :

Case 2. Exponential Function Transformation

Suppose that a parameter b^ is restricted to an exponential function of ^b in order to
guarantee 0 < b^ < 1. Hence,

 ^
b^ ¼ f ^b ¼ e
b : (15.19)

Using Eq. 15.18,

  
 2

var b^ ¼ f 0 ^b  var ^b , (15.20)

 
2
where f 0 ^
b ¼ f ^b ¼ b^ :
2

Appendix 2: News Impact Curve

We derive the functions of the news impact curve (NIC) for the ANST-GARCH and
the GJR models as follows:
ANST-GARCH (1,1) Model:



Model : ht ¼ a0 þ a1 e2t
1 þ g1 ht
1 þ b0 þ b1 e2t
1 þ b2 ht
1  Fðet
1 Þ (15.21)

NIC : ht ¼ C þ ½a1 þ b1 Fðet
1 Þe2t
1

and
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 435

C ¼ a0 þ b0 Fðet
1 Þ þ ½g1 þ b2 Fðet
1 Þs2 , (15.22)

where Fðet
1 Þ ¼ f1 þ exp½
gðet
1 Þg
1 :

GJR (1, 1) Model:


Model : ht ¼ a0 þ a1 e2t
1 þ b1 ht
1 þ a2 e2t
1  I ðet
1 > 0Þ, (15.23)

1, if et
1 > 0
where I ðet
1 > 0Þ ¼ :
0, if et
1 < 0

ða0 þ b1 s2 Þ þ a1 e2t
1 , if et
1 > 0

NIC : ht ¼ : (15.24)
ða0 þ b1 s2 Þ þ ða1 þ a2 Þe2t
1 , if et
1 < 0

Appendix 3: Sign Bias Tests

Based on the news impact curve, we perform three diagnostic tests for examining
the ability of a model to capture asymmetric effect of news on conditional variance:
the sign bias test (SBT), the negative sign bias test (NSBT), and the positive sign
bias test (PSBT). These tests are performed by the t-statistic on the coefficient
b under the following regression equations:
0
v2t ¼ a þ bS

t
1 þ b zt þ et (15.25)
0
v2t ¼ a þ bS

t
1 et
1 þ b zt þ et (15.26)
0
v2t ¼ a þ bSþ 
t
1 et
1 þ b zt þ et , (15.27)

pffiffiffiffi2
where v2t ¼ et = ht , S


t
1 ¼ 1 if et
1 < 0 and St
1 ¼ 0 otherwise, and

 e  e
St
1 ¼ 1
St
1. zt ¼ h ðyÞ=ht , where h ðyÞ ¼ @ht =@y evaluated at the
+

values of maximum likelihood estimates of parameter y, and h*t is the estimated

conditional variance by a model considered.

References
Abel, A. B. (1988). Stock prices under time-varying dividend risk: An exact solution in an infinite-
horizon general equilibrium model. Journal of Monetary Economics, 22, 373–393.
Ang, A., Hodrick, R., Xing, Y., & Zhang, X. (2006). The cross-section of volatility and expected
returns. Journal of Finance, 61, 259–299.
Backus, D., & Gregory, A. W. (1993). Theoretical relations between risk premiums and condi-
tional variances. Journal of Business and Economic Statistics, 11, 177–185.
436 K. Nam et al.

Baillie, R., & DeGennaro, R. P. (1990). Stock returns and volatility. Journal of Financial and
Quantitative Analysis, 25, 203–214.
Bali, T. G. (2008). The intertemporal relation between expected returns and risk. Journal of
Financial Economics, 87, 101–131.
Ball, R., & Kothari, S. (1989). Nonstationary expected returns: Implications for tests of market
efficiency and serial correlation in returns. Journal of Financial Economics, 25, 51–74.
Barky, R. B. (1989). Why don’t the prices of stocks and bonds move together? American
Economic Review, 79, 1132–1145.
Bollerslev, T., & Wooldridge, J. M. (1992). Quasi-maximum likelihood estimation and inference
in dynamic models with time-varying covariances. Econometric Review, 11, 143–172.
Boudoukh, J., Richardson, M., & Whitelaw, R. F. (1997). Nonlinearities in the relation between
the equity risk premium and the term structure. Management Science, 43, 371–385.
Brandt, W., & Kang, Q. (2004). On the relationship between the conditional mean and volatility of
stock returns: A latent VAR approach. Journal of Financial Economics, 72, 217–257.
Breeden, D. T. (1979). An intertemporal asset pricing model with stochastic consumption and
investment opportunities. Journal of Financial Economics, 7, 265–296.
Breen, W., Glosten, L., & Jagannathan, R. (1989). Economic significance of predictable variations
in stock index returns. Journal of Finance, 44, 1177–1189.
Campbell, J. Y. (1987). Stock returns and the term structure. Journal Financial Economics, 18,
373–399.
Campbell, J. Y., & Hentschel, L. (1992). No news is good news: An asymmetric model of
changing volatility in stock returns. Journal Financial Economics, 31, 281–318.
Campbell, J. Y., Lo, A. W., & MacKinlay, A. C. (1997). The econometrics of financial markets.
Princeton: Princeton University Press.
Cecchetti, S. G., Lam, P., & Mark, N. C. (1990). Mean reversion in equilibrium asset prices.
American Economic Review, 80, 398–418.
Darrat, A. F., Gilley, O. W., Li, B., & Wu, Y. (2011). Revisiting the risk/return relations in the
Asian pacific markets: New evidence from alternative models. Journal of Business Research,
64, 199–206.
Engle, R. F., & Ng, V. K. (1993). Measuring and testing the impact of news on volatility. Journal
of Finance, 48, 1749–1778.
Evans, M. D. (1994). Expected returns, time-varying risk, and risk premia. Journal of Finance, 49,
655–679.
Fama, E. F., & French, K. R. (1988). Permanent and temporary components of stock prices.
Journal of Political Economy, 96, 246–273.
Fama, E. F., & French, K. R. (1989). Business conditions and expected returns on stocks and
bonds. Journal of Financial Economics, 25, 23–49.
Ferson, W., & Harvey, C. (1991). The variation of economic risk premiums. Journal of Political
Economy, 99, 385–415.
Fornari, F., & Mele, A. (1997). Sign- and volatility-switching ARCH models: Theory and
applications to international stock markets. Journal of Applied Econometrics, 12, 49–65.
French, K. R., Schwert, G. W., & Stambaugh, R. F. (1987). Expected stock returns and volatility.
Journal of Financial Economics, 19, 13–29.
Friedman, M. (1957). A theory of the consumption function. Princeton: Princeton University Press.
Ghysel, E., Santa-Clara, P., & Valkanov, R. (2005). There is a risk-return trade-off after all.
Journal of Financial Economics, 76, 509–548.
Glosten, L., Jagannathan, R., & Runkle, D. (1993). On the relation between the expected value
and the volatility of the nominal excess return on stocks. Journal of Finance, 48, 1779–1801.
Gonzalez-Rivera, G. (1998). Smooth transition GARCH models. Studies in Nonlinear Dynamics
and Econometrics, 3, 61–78.
Gourieroux, C., & Monfort, A. (1997). Time series and dynamic models. Cambridge, UK:
Cambridge University Press.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
15 The Effect of Unexpected Volatility Shocks on Intertemporal Risk-Return Relation 437

Hamilton, J. D., & Susmel, R. (1994). Autoregressive conditional heteroskedasticity and changes
in regime. Journal of Econometrics, 64, 307–333.
Harvey, C. R. (1989). Time-varying conditional covariances in tests of asset pricing models.
Journal of Financial Economics, 24, 289–317.
Harvey, A. C. (1993). The econometric analysis of time series. Cambridge, MA: The MIT Press.
Harvey, C. R. (2001). The specification of conditional expectations. Journal of Empirical Finance,
8, 573–638.
Haugen, R. A., Talmor, E., & Torous, W. N. (1991). The effect of volatility changes on the level of
stock prices and subsequent expected returns. Journal of Finance, 46, 985–1007.
Kim, C., Morley, J. C., & Nelson, C. R. (2001). Does an intertemporal tradeoff between risk and
return explain mean reversion in stock price? Journal of Empirical Finance, 8, 403–426.
Lucas, R. E. (1978). Asset prices in an exchange economy. Econometrica, 46, 1429–1446.
Ludvigson, S. C., & Ng, S. (2007). The empirical risk-return relation: A factor analysis approach.
Journal of Financial Economics, 83, 171–222.
utkepol, H. (1993). Introduction to multiple time series analysis. New York: Springer.
L€
Merton, R. C. (1973). An intertemporal asset pricing model. Econometrica, 41, 867–888.
Merton, R. C. (1980). On estimating the expected return on the market: An exploratory investi-
gation. Journal of Financial Economics, 8, 323–361.
Modigliani, F., & Brumberg, R. (1955). Utility analysis and the consumption function. In
K. Kurihara (Ed.), Post Keynesian economics. London: G. Allen.
M€uller, G., Durand, R. B., & Maller, R. A. (2011). The risk–return tradeoff: A COGARCH
analysis of Merton’s hypothesis. Journal of Empirical Finance, 18, 306–320.
Nam, K., Pyun, C., & Avard, S. (2001). Asymmetric reverting behavior of short-horizon stock
returns: An evidence of stock market overreaction. Journal of Banking and Finance, 25,
805–822.
Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.
Econometrica, 59, 347–370.
Nyberga, H. (2012). Risk-return tradeoff in U.S. stock returns over the business cycle. Journal of
Financial and Quantitative Analysis, 47, 137–158.
Pagan, A. R., & Hong, Y. S. (1989). Nonparametric estimation and the risk premium. In W.A. Barnett,
J. Powell, & G. E. Tauchen (Eds.), Semiparametric and nonparametric methods in econometrics
and statistics (pp. 51–75). Cambridge: Cambridge University Press.
Pastor, L., Sinha, M., & Swaminthan, B. (2008). Estimating the intertemporal risk-return tradeoff
using the implied cost of capital. Journal of Finance, 63, 2859–2897.
Pindyck, R. (1984). Risk, inflation, and the stock market. American Economic Review, 74,
335–351.
Poterba, J. M., & Summers, L. H. (1986). The persistence of volatility and stock market fluctu-
ations. American Economic Review, 76, 1142–1151.
Rothman, P. (1999). Nonlinear time series analysis of economic and financial data. Boston:
Kluwer.
Scruggs, J. T. (1998). Resolving the puzzling intertemporal relation between the market risk
premium and conditional market variance: A two-factor approach. Journal of Finance, 52,
575–603.
Turner, C. M., Startz, R., & Nelson, C. R. (1989). A Markov model of heteroskedasticity, risk, and
learning in the stock market. Journal of Financial Economics, 25, 3–22.
Whitelaw, R. F. (1994). Time variations and covariations in the expectation and volatility of stock
market returns. Journal of Finance, 49, 515–541.
Whitelaw, R. F. (2000). Stock market risk and return: An equilibrium approach. Review of
Financial Studies, 13, 521–547.
Yu, J., & Yuan, Y. (2011). Investor sentiment and the mean–variance relation. Journal of
Financial Economics, 100, 367–381.
Combinatorial Methods for Constructing
Credit Risk Ratings 16
Alexander Kogan and Miguel A. Lejeune

Contents
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
16.1.1 Importance of Credit Risk Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
16.1.2 Contribution and Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
16.2 Logical Analysis of Data: An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
16.3 Absolute Creditworthiness: Credit Risk Ratings of Financial Institutions . . . . . . . . . . . . 447
16.3.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
16.3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
16.3.3 An LAD Model for Bank Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
16.3.4 LAD Model Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
16.3.5 Remarks on Reverse Engineering Bank Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
16.4 Relative Creditworthiness: Country Risk Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
16.4.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
16.4.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
16.4.3 Rating Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
16.4.4 Evaluation of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
16.4.5 Importance of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
16.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481

A. Kogan (*)
Rutgers Business School, Rutgers, The State University of New Jersey, Newark–New Brunswick,
NJ, USA
Rutgers Center for Operations Research (RUTCOR), Piscataway, NJ, USA
e-mail: kogan@rutgers.edu
M.A. Lejeune
George Washington University, Washington, DC, USA
e-mail: mlejeune@gwu.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 439
DOI 10.1007/978-1-4614-7750-1_16,
# Springer Science+Business Media New York 2015
440 A. Kogan and M.A. Lejeune

Abstract
This study uses a novel method, the Logical Analysis of Data (LAD), to reverse
engineer and construct credit risk ratings which represent the creditworthiness of
financial institutions and countries. LAD is a data mining method based on
combinatorics, optimization, and Boolean logic that utilizes combinatorial
search techniques to discover various combinations of attribute values that are
characteristic of the positive or negative character of observations. The proposed
methodology is applicable in the general case of inferring an objective rating
system from archival data, given that the rated objects are characterized by
vectors of attributes taking numerical or ordinal values. The proposed
approaches are shown to generate transparent, consistent, self-contained, and
predictive credit risk rating models, closely approximating the risk ratings
provided by some of the major rating agencies. The scope of applicability of
the proposed method extends beyond the rating problems discussed in this study
and can be used in many other contexts where ratings are relevant.
We use multiple linear regression to derive the logical rating scores.

Keywords
Credit risk rating • Reverse engineering • Logical Analysis of Data • Combina-
torial optimization • Data mining • Creditworthiness • Financial strength •
Internal rating • Preorder • Logical rating score

16.1 Introduction

16.1.1 Importance of Credit Risk Ratings

Credit ratings published by such agencies as Moody’s, Standard & Poor’s, and Fitch
are considered important indicators for financial markets, providing critical infor-
mation about the likelihood of future default. The importance of credit ratings is
recognized by international bodies, as manifested by the Basel Capital Accord
(2001, 2006). The progressively increasing importance of credit risk ratings is
driven by the dramatic expansion of investment opportunities associated with the
globalization of the world economies. Since these opportunities are often risky, the
internationalized financial markets have to rely on agencies’ ratings for the assess-
ment of credit risk.
The importance of credit risk rating systems is manifold, as shown below:
• Credit approval: the credit risk rating plays a major role in the credit
preapproval decisions. Indeed, a binary decision model in which the credit risk
rating of the obligor is an explanatory variable is typically used to preapprove the
decision to grant or not the credit.
• Pricing: in case of loan preapproval, the credit risk rating impacts the conditions
(interest rate, covenants, collaterals, etc.) under which the final credit is granted
(Treacy and Carey 2000). The credit risk rating is also utilized in the operations
subsequent to the preapproval stage.
16 Combinatorial Methods for Constructing Credit Risk Ratings 441

• Provisioning: the credit risk rating is a variable used for calculating the expected
losses, which, in turn, are used to determine the amount of economic capital
that a bank must keep to hedge against possible defaults of its borrowers.
The expected loss of a credit facility is a function of the probability of default
of the borrower to which the credit line is granted, as well as the exposure at
default and the loss given default associated with the credit facility. Since there
is usually a mapping between credit risk rating and the borrower’s probability of
default, and provided that the rating of a borrower is a key predictor to assess the
recovery rate associated with a credit facility granted to this borrower, the
importance of the credit risk rating in calculating the amounts of economic
capital is evident.
• Moral hazard: the reliance upon an objective and accurate credit risk rating
system is a valuable tool against some moral hazard situations. Some financial
institutions do not use a credit risk rating model but instead let lending officers
assign credit ratings based on their judgment. The lending officers are in charge
of the marketing of banking services, and their performance and therefore
compensation are determined with respect to the “profitability” of the relation-
ships between the bank and its customers. Clearly, the credit risk ratings
assigned by the lending officers will affect the volume of the approved loans
and the compensation of the officer who, as a result, could have an incentive to
assign ratings in a way that is not consistent with the employer’s interests. Thus,
the use of a reliable credit risk rating system could lead to avoiding such perverse
incentive situations.
• Basel compliance: the New Basel Capital Accord (Basel II) requires banks to
implement a robust framework for the evaluation of credit risk exposures of
financial institutions and the capital requirements they must bear. This involves
the construction and the cross-validation of accurate and predictive credit risk
rating systems.
Financial institutions, while taking into account external ratings (e.g., those
provided by Fitch, Moody’s, S&P’s), have increasingly been developing efficient
and refined internal rating systems over the past years. The reasons for this trend are
multiple.
First, the work of rating agencies, providing external, public ratings, has recently
come under intense scrutiny and criticism, justified partly by the fact that some of
the largest financial collapses of the decade (Enron Corp, etc.) were not anticipated
by the ratings. At his March 2006 testimony before the Senate Banking Committee,
the president and CEO of the CFA Institute1 highlighted the conflicts of interest in
credit ratings agencies (Wall Street Letter 2006). He regretted that credit rating
agencies “have been reluctant to embrace any type of regulation over the services
they provide” and reported that credit rating agencies “should be held to the highest
standards of transparency, disclosure and professional conduct. Instead, there are no
standards.” The problem is reinforced by the fact that rating agencies, charging fees

1
Wall Street Letter. 2006. CFA To Senate: Follow Our Lead On Credit Rating.
442 A. Kogan and M.A. Lejeune

to rated countries, can be suspected of reluctance to downgrade them, because of the

possibility of jeopardizing their income sources. This is claimed, for example, by
Tom McGuire, an executive vice-president of Moody’s, who states that “the pressure
from fee-paying issuers for higher ratings must always be in a delicate balance with
the agencies’ need to retain credibility among investors.”2 The necessity to please the
payers of the ratings opens the door to many possible issues. Kunczik (2001) notes
that the IMF fears the danger that “issuers and intermediaries could be encouraged to
engage in rating shopping – a process in which the issuer searches for the least
expensive and/or least demanding rating.” The reader is referred to Hammer
et al. (2006) for a discussion of some other criticisms (lack of comprehensibility,
procyclicality, black box, lack of predictive and crisis-warning power, regional bias,
etc.) commonly addressed to rating agencies.
Second, an internal rating system provides autonomy to a bank’s management in
defining credit risk in line with that bank’s core business and best international
practices. Internal ratings are also used to report to senior management various key
metrics such as risk positions, loan loss reserves, economic capital allocation, and
employee compensation (Treacy and Carey 2000).
Third, while the Basel Committee on Banking Supervision of the Bank for
International Settlements favored originally (i.e., in the 1988 Capital Accord) the
ratings provided by external credit ratings agencies, it is now encouraging the
Internal Ratings-Based (IRB) approach under which banks use their own internal
rating estimates to define and calculate default risk, on the condition that the robust
regulatory standards are met and the internal rating system is validated by the
national supervisory authorities. The committee has defined strict rules for credit
risk models used by financial institutions and requires them to develop and cross-
validate these models in order to comply with the Basel II standard (Basel Com-
mittee on Banking Supervision 2001, 2006).
Fourth, aside from the Basel II requirements, banks are developing internal risk
models to make the evaluation of credit risk exposure more accurate and transpar-
ent. Credit policies and processes will be more efficient, and the quality of data will
be improved. This is expected to translate into substantial savings on capital
requirements. Today’s eight cents out of every dollar that banks hold in capital
reserves could be reduced in banks with conservative credit risk policies, resulting
in higher profitability.

16.1.2 Contribution and Structure

The above discussion outlines the impact of the credit risk ratings, the possible
problems of objectivity and transparency of external (i.e., provided by rating
agencies) credit risk ratings, and the importance and need for financial institutions
to develop their own, internal credit risk rating systems.

2
The Economist, July 15, 1995, 62.
16 Combinatorial Methods for Constructing Credit Risk Ratings 443

In this chapter, we use the novel combinatorial pattern extraction method called
the logical analysis of data (LAD) to learn a given credit risk rating system and to
develop on its basis a rating system having the following characteristics:
• Self-containment: the rating system does not use as predictor variables any other
credit risk rating information (non-recursiveness). Clearly, this requirement pre-
cludes the use of lagged ratings as independent variables. It is important to note that
this approach is in marked contrast with that of the current literature (see Hammer
et al. 2006, 2011 for a discussion). The significant advantage of the non-recursive
nature of the rating system is its applicability to not-yet-rated obligors.
• Objectivity: the rating system only relies on measurable characteristics of the
rated entities.
• Transparency: the rating system has formal explicit specification.
• Accuracy: it is in close agreement with the learned, opaque rating system.
• Consistency: the discrepancies between the learned rating system and the
constructed one are resolved by subsequent changes in the learned rating system.
• Generalizability: applicability of the rating system to evaluate the creditwor-
thiness of obligors at subsequent years or of obligors that were not previously
rated.
• Basel compliance: it satisfies the Basel II Accord requirements (cross-
validation, etc.).
In this study, we derive a rating system for two types of obligors:
• Countries: Eliasson (2002) defines country risk as the “risk of national gov-
ernments defaulting on their obligations,” while Afonso et al. (2007) state that
“sovereign credit ratings are a condensed assessment of a government’s ability
and willingness to repay its public debt both in principal and in interests on
time.” Haque et al. (1996) define country credit risk ratings compiled by
commercial sources as an attempt “to estimate country-specific risks, particu-
larly the probability that a country will default on its debt-servicing obligations.”
• Financial institutions: bank financial strength ratings represent the “bank’s
intrinsic safety and soundness” (Moody’s 2006).
Two different approaches developed in this chapter for reverse engineering and
constructing credit risk ratings will be based on the following:
• Absolute creditworthiness, which evaluates the riskiness of individual obligors.
• Relative creditworthiness, which first evaluates the comparative riskiness of
pairs of obligors; the absolute riskiness of entities is then derived from their
relative riskiness using the combinatorial techniques of partially ordered sets.
As was noted in the literature (de Servigny and Renault 2004), banks and other
financial service organizations are not fully utilizing the opportunities provided by
the significant increase in the availability of financial data. More specifically, the
area of credit rating and scoring is lacking in up-to-date methodological advances
(Galindo and Tamayo 2000; de Servigny and Renault 2004; Huang et al. 2004).
This provides a tremendous opportunity for the application of modern data mining
and machine learning techniques, based on statistics (Jain et al. 2000) and combi-
natorial pattern extraction (Hammer 1986). The above described approaches are
implemented using the LAD combinatorial pattern extraction method, and it turns
444 A. Kogan and M.A. Lejeune

out to be a very conclusive case for the application and the contribution of data
mining to the credit risk industry.
This chapter is structured as follows. Section 16.2 provides an overview of LAD
(Hammer 1986), which is used to develop a new methodology for reverse engineer-
ing and rating banks and countries with respect to their creditworthiness. Section 16.3
details the absolute creditworthiness model constructed for rating financial institu-
tions and describes the obtained results. Section 16.4 is devoted to the description of
two relative creditworthiness models for rating countries and contains an extensive
description of the results. Section 16.5 provides concluding remarks.

16.2 Logical Analysis of Data: An Overview

The logical analysis of data (LAD) is a modern data mining methodology based on
combinatorics, optimization, and Boolean logic3. LAD can be applied for the analysis
and classification of archives containing both binary (Hammer 1986; Crama
et al. 1988) and numerical (Boros et al. 1997) data. The novelty of LAD consists in
utilizing combinatorial search techniques to discover various combinations of attri-
bute values that are characteristic of the positive or negative character of observations
(such as whether a bank is solvent or not or whether a patient is healthy or sick). Then
LAD selects (usually small) subsets of such combinations (usually optimizing
a certain quality objective) to construct what is called a model (Boros et al. 2000).
We briefly describe below the basic concepts of LAD, referring the reader for a more
detailed description to Boros et al. (2000) and Alexe et al. (2007).
Observations in archives analyzed by LAD are represented by n-dimensional
real-valued vectors which are called positive or negative based on the value of the
additional binary (0,1) attribute called the outcome or the class of the observation.
Consider a dataset as a collection of M points (a j, z j) where the outcome z j of
observation j has value 1 for a positive outcome and 0 for a negative one and a j is an
n-dimensional vector. Figure 16.1 illustrates a dataset containing five observations
described by three variables. Each component a j[i] of the [5
3]-matrix in

Observation Variables Outcome

j a[1] a[2] a[3] zj

1 3.5 3.8 2.8 1
2 2.6 1.6 5 1
3 1 2.2 3.7 1
4 3.5 1.4 3.9 0
5 2.3 2.1 1 0
Fig. 16.1 Set of observations

3
The presentation in this section is partially based on Hammer et al. (2006).
16 Combinatorial Methods for Constructing Credit Risk Ratings 445

Observations
Variables a[1] a[2] a[3]

Outcome
c1,1 c1,2 c1,3 c2,1 c2,2 c3,1 c3,2 c3,3
Cutpoints
3 2.4 1.5 3 2 4 3 2
j
yi,k zj
1 1 1 1 1 1 0 0 1 1
Binary 2 0 1 1 0 0 1 1 1 1
Variables
3 0 0 0 0 1 0 1 1 1
4 1 1 1 0 0 0 1 1 0
5 0 0 1 0 1 0 0 0 0

Fig. 16.2 Binarized dataset

Fig. 16.1 gives the value taken by variable i in observation j. We will use a[i] to
denote the variables corresponding to the components of this dataset. The rightmost
column provides the outcome of the observation.
LAD discriminates positive and negative observations by constructing a binary-
valued function f depending on the n input variables, in such a way that it closely
approximates the unknown actual discriminator. LAD constructs this function f as
a weighed sum of combinatorial patterns.
In order to specify how such a function f is found, we first transform the original
dataset into a binarized dataset in which the variables can only take the values 0 and
1. We shall achieve this goal by using indicator variables which show whether the
values the variables take in a particular observation are “large” or “small”; more
precisely, each indicator variable shows whether the value of a numerical variable
does or does not exceed a specified level. This is achieved by defining, for each
variable a[i], a set of K(i) values {ci,k j k ¼ 1,..., K(i)}, called cut points to which
binary variables {yi,k j k ¼ 1,. . .,K(i)} are associated. The values of these binary
variables for each observation (z j, a j) are then defined as:


1 if a j ½i  ci, k
yij, k ¼
0 otherwise

Figure 16.2 provides the binarized dataset corresponding to the data displayed in
Fig. 16.1 and shows the values ci,k of the cut points k for each variable a[i] and those
j
of the binary variables yi,k associated with any cut point k of variable i in observa-
tion j. For example, y1,1 ¼ 1 since a1[1 ¼ 3.5 is greater than c1,1 ¼ 3.
1

Positive (negative) patterns are combinatorial rules obtained as conjunctions of

binary variables and their negations, which, when translated to the original vari-
ables, constrain a subset of input variables to take values between identified upper
and lower bounds, so that:
• All the pattern conditions are satisfied by a sufficiently high proportion of the
positive (negative) observations in the dataset.
446 A. Kogan and M.A. Lejeune

• At least one of the pattern conditions is violated by a sufficiently high proportion

of the negative (positive) observations.
The number of variables the values of which are constrained in the definition of a
pattern is called the degree of the pattern. The fraction of positive (negative)
observations covered by a positive (negative) pattern is called the prevalence of it.
The fraction of positive (negative) observations among those covered by a positive
(negative) pattern is called the homogeneity of it. Considering the above example,
the pattern

y1, 3 ¼ 0 and y3, 1 ¼ 1

is a positive one of degree 2 covering one positive observation and no negative

observation. Therefore, its prevalence is equal to 100/3 % and its homogeneity is
equal to 100 %. In terms of the original variables, it imposes a strict upper bound (3)
on the value of a[1] and a strict lower bound (4) on the value of a[3].
LAD starts its analysis of a dataset by generating the pandect, that is, the
collection of all patterns in a dataset. Note that the pandect of a dataset of typically
occurring dimension can contain exponentially large number of patterns, but many
of these patterns are either subsumed by other patterns or similar to them. It is
therefore important to impose a number of limitations on the set of patterns to be
generated, by restricting their degrees (to low values), their prevalence (to high
values), and their homogeneity (to high values); these bounds are known as LAD
control parameters. The quality of patterns satisfying these conditions is usually
much higher than that of patterns having high degrees, or low prevalence, or low
homogeneity. Several algorithms have been developed for the efficient generation
of large subsets of the pandect corresponding to reasonable values of the control
parameters (Boros et al. 2000). The collections of patterns sufficient for classifying
the observations in the dataset are called models. A model includes sufficiently
many positive (negative) patterns to guarantee that each of the positive (negative)
observations in the dataset is “covered” by (i.e., satisfies the conditions of) at least
one of the positive (negative) patterns in the model. Good models tend to minimize
the number of points in the dataset covered simultaneously by both positive and
negative patterns in the model. LAD models can be constructed using the
Datascope software (Alexe 2002).
LAD classifies observations on the basis of model’s evaluation of them in the
following way. An observation (contained in the given dataset or not), satisfying the
conditions of some of the positive (negative) patterns in the model and not satisfying
the conditions of any of the negative (positive) patterns in the model, is classified as
positive (negative). To classify an observation that satisfies both positive and negative
patterns in the model, LAD utilizes a discriminant that assigns specific weights to the
patterns in the model (Boros et al. 2000). To define the simplest discriminant that
assigns equal weights to all positive (negative) patterns, let p and q represent the
number of positive and negative patterns in a model, and let h and k represent the
numbers of those positive, respectively, negative patterns in the model which cover
a new observation o. Then the discriminant D(o) is calculated as
16 Combinatorial Methods for Constructing Credit Risk Ratings 447

Table 16.1 Classification Classification of observations

matrix
Observation classes Positive Negative Unclassified
Positive a c e
Negative b d f

DðoÞ ¼ h=p  k=q (16.1)

and the corresponding classification is determined by the sign of this expression. LAD
leaves unclassified any observation for which D(o) ¼ 0, since in this case either the
model does not provide sufficient evidence or the evidence it provides is contradictory.
Computational experience with real-life problems has shown that the number of
unclassified observations is usually small. The results of classifying the observations
in a dataset can be represented in the form of a classification matrix (Table 16.1).
Here, the percentage of positive (negative) observations that are correctly
classified is represented by a (respectively d). The percentage of positive
(negative) observations that are misclassified is represented by c (respectively b).
The percentage of positive (negative) observations that remain unclassified is
represented by e (respectively f). Clearly, a + c + e ¼ 100 % and b + d + f ¼
100 %. The quality of the classification is defined by

1 1
Q ¼ ða þ dÞ þ ðe þ f Þ (16.2)
2 4

16.3 Absolute Creditworthiness: Credit Risk Ratings

of Financial Institutions

16.3.1 Problem Description

The capability of evaluating the credit quality of banks has become extremely
important in the last 30 years given the increase in the number of bank failures:
during the period from 1950 to 1980, bank failures averaged less than seven per
year, whereas during the period from 1986 to 1991, they averaged 175 per year
(Barr and Siems 1994).4 Curry and Shibut (2000) report that the so-called savings
and loan crisis cost around $123.8 billion. Central banks are afraid of widespread
bank failures since they could exacerbate cyclical recessions and result in more
severe financial crises (Basel Committee on Banking Supervision 2004). More
accurate credit risk models for banks could enable the identification of problematic
banks early, which is seen as a necessary condition by the Bank for International
Settlements (2004) to avoid failure, and could serve the regulators in their efforts to
minimize bailout costs.

4
The presentation in this section is based on Hammer et al. (2012).
448 A. Kogan and M.A. Lejeune

The evaluation and rating of the creditworthiness of banks and other financial
organizations is particularly challenging, since banks and insurance companies
appear to be more opaque than firms operating in other industrial sectors. Morgan
(2002) attributes this to the fact that banks hold certain assets (loans, trading assets,
etc.), the risks of which change fast and are very difficult to assess, and it is further
compounded by banks’ high leverage. Therefore, it is not surprising that the main
rating agencies (Moody’s and S&P’s) disagree much more often about the ratings
given to banks than about those given to obligors in other sectors. The difficulty of
accurately rating those organizations is also due the fact that the rating migration
volatility of banks is historically significantly higher than it is for corporations and
countries and that banks tend to have higher default rates than corporations
(de Servigny and Renault 2004). Another distinguishing characteristic of the
banking sector is the external support (i.e., from governments) that banks receive
and other corporate sectors do not (Fitch Rating 2006). A thorough review of the
literature pertaining to the rating and evaluation of credit risk of financial institu-
tions can be found in Hammer et al. (2012).
In the next sections, we shall:
• Identify a set of variables that provides sufficient information to accurately
replicate the Fitch bank ratings.
• Construct LAD patterns to discriminate between banks with high and low ratings.
• Construct an optimized model utilizing (some of) the LAD patterns which is
capable of distinguishing between banks with high and low ratings.
• Define an accurate bank rating system on the basis of the discriminant values
provided by the constructed model.
• Cross-validate the proposed rating system.

16.3.2 Data

16.3.2.1 External Credit Risk Ratings of Financial Institutions

This section starts with a brief description of the Fitch individual bank rating
system. Long- and short-term credit ratings provided by Fitch constitute an opinion
on the ability of an entity to meet financial commitments (interest, preferred
dividends, or repayment of principal) on a timely basis (Fitch Ratings 2001).
These ratings are comparable worldwide and are assigned to countries and corpo-
rations, including banks.
Fitch bank ratings include individual and support ratings. Support ratings com-
prise five rating categories which reflect the likelihood that a banking institution
will receive support either from the owners or the governmental authorities if it runs
into difficulties. The availability of support, though critical, does not reflect
completely the likelihood that a bank will remain solvable in case of adverse
situations. To complement a support rating, Fitch also provides an individual
bank rating to evaluate credit quality separately from any consideration of outside
support. This rating is commonly viewed as assessing a bank were it entirely
independent and could not rely on external support. It supposedly takes into
16 Combinatorial Methods for Constructing Credit Risk Ratings 449

Table 16.2 Fitch individual rating system (Fitch Ratings 2001)

Category Numerical scale Description
A 9 A very strong bank. Characteristics may include outstanding
profitability and balance sheet integrity, management, or
operating environment
B 7 A strong bank. There are no major concerns regarding the
bank
C 5 An adequate bank which, however, possesses one or more
troublesome aspects. There may be some concerns
regarding its profitability, balance sheet integrity,
management, operating environment or prospects
D 3 A bank which has weaknesses of internal and/or external
origin. There are concerns regarding its profitability,
management, balance sheet integrity, franchise, operating
environment or prospects
E 1 A bank with very serious problems which either requires or
is likely to require external support

consideration such factors as profitability and balance sheet integrity, franchise,

management, operating environment, and prospects.
We present in Table 16.2 a detailed description of the nine rating categories
characterizing the Fitch individual bank credit rating system. Since individual bank
credit ratings are comparable across different countries, they will be used in the
remaining part of this chapter. In addition, Fitch uses gradations among these five
ratings: A/B, B/C, C/D, and D/E, the corresponding numerical values of which
being, respectively, 8, 6, 4, and 2. This conversion of the Fitch individual bank
ratings into a numerical scale is commonly used (see, e.g., Poon et al. 1999).

16.3.2.2 Variables and Observations

We use the following 14 financial variables (loans, other earning assets, total
earning assets, nonearning assets, net interest revenue, customer and short-term
funding, overheads, equity, net income, total liability and equity, operating income)
and nine representative financial ratios as predictors in our model. The variables are
as follows: ratio of equity to total assets (asset quality), net interest margin, ratio of
interest income to average assets, ratio of other operating income to average assets,
ratio of noninterest expenses to average assets, return on average assets (ROAA),
return on average equity (ROAE), cost-to-income ratio (operations), and ratio of net
loans to total assets (liquidity). The values of these variables were collected at the
end of 2000 and come from the database Bankscope.
As an additional variable, we use in this study the S&P’s risk rating of the country
where the bank is located. The S&P’s country risk rating scale comprises 22 different
categories (from AAA to D). We convert these categorical ratings into a numerical
scale, assigning the largest numerical value (21) to the countries with the highest
rating (AAA). Similar numerical conversions of country risk ratings are also used by
Ferri et al. (1999) and Sy (2004). Similarly, Bloomberg has also developed a standard
cardinal scale for comparing Moody’s, S&P’s, and Fitch-BCA ratings.
450 A. Kogan and M.A. Lejeune

Our dataset consists of 800 banks rated by Fitch and operating in 70 different
countries (247 in Western Europe; 51 in Eastern Europe; 198 in Canada and the
USA; 45 in developing Latin American countries; 47 in the Middle East; 6 in
Oceania; 6 in Africa; 145 in developing Asian countries; 55 in Hong Kong, Japan,
and Singapore).

16.3.3 An LAD Model for Bank Ratings

Our design of an objective and transparent bank rating system on the basis of the LAD
methodology is guided by the properties of LAD as a classification system. Therefore,
we define a classification problem associated to the bank rating problem, construct
an LAD model for it, and then define a bank rating system rooted in this LAD model.
We define as positive observations the banks which have been rated by Fitch
as A, A/B, or B and as negative observations those whose Fitch rating is D, D/E, or E.
In the binarization process, cut points were introduced for the 19 of the 24 numer-
ical variables shown in Table 16.3. Actually, the other five numerical variables
were also binarized, but since it turned out that they were redundant, only the
variables shown in Table 16.3 were retained for constructing the model. Table 16.3
provides all the cut points used in pattern and model construction. For example, two
cut points (24.8 and 111.97) are used to binarize the numerical variable “profit
before tax” (PbT), that is, two binary indicator variables replace PbT, one telling
whether PbT exceeds 24.8 and the other telling whether PbT exceeds 111.97.
The first step of applying the LAD technique to the problem binarized with the
help of these variable cut points was the identification of a collection of powerful
patterns. One example of such a powerful negative pattern is (i) the country risk
rating is strictly lower than A and (ii) the profits before tax are at most equal to
€111.96 millions. One can see that these conditions describe a negative pattern,
since none of the positive observations (i.e., banks rated A, A/B, or B) satisfy both
of them, while no less than 69.11 % of the negative observations (i.e., those banks
rated D, D/E, or E) do satisfy both conditions. This pattern has degree 2, prevalence
69.11 %, and homogeneity 100 %.
The model we have developed for bank ratings is very parsimonious, consisting
of only 11 positive and 11 negative patterns, and is built on a support set of only
19 out of the 24 original variables. All the patterns in the model are of degree at
most 3, have perfect homogeneity (100 %), and very substantial prevalence
(averaging 50.9 % for the positive and 37.7 % for the negative patterns).

16.3.4 LAD Model Evaluation

16.3.4.1 Accuracy and Robustness of the LAD Model

The classification of the banks whose ratings are A, A/B, B, D, D/E, or E with the
above LAD model is 100 % accurate. We use ten two-folding experiments to cross-
validate the model. In each of those experiments, we randomly assign the
 16

Table 16.3 Cut points

Numerical variables Cut points Numerical variables Cut points Numerical variables Cut points
Country risk rating 11, 16, 19.5, 20 Operating income (memo) 1,360.74 Noninterest expenses/ 2.00, 2.77, 3.71, 4.93
average assets
Other earning assets 3,809, 9,564.9 Profit before tax 24.8, 111.97 Return on average assets 0.30, 0.80, 1.52
Nonearning assets 364 Equity/total assets 4.90 Return on average equity 11.82, 15.85, 19.23
Total assets 5,735.8 Equity 370.45 Overhead 127, 324
Customer & short-term 4,151.3, 9,643.8 Net interest margin 1.87 Net loans/total assets 44.95, 53.90, 59.67, 66.50
funding
Cost-to-income ratio 50.76, 71.92 Net interest revenue/average 3.02
assets
Other operating income 46.8, 155.5, 470.5 Other operating income/ 0.83, 1.28, 1.86,
Combinatorial Methods for Constructing Credit Risk Ratings

average assets 3.10

451
452 A. Kogan and M.A. Lejeune

observations to a training and a testing sets of equal size. We use the training set to
build the model, and we apply it to classify the observations in the testing set. In the
second half of the experiment, we reverse the role of the two sets. The average
accuracy and the standard deviation of the accuracy are, respectively, equal to
95.12 % and 0.03. These results highlight the predictive capability and the robust-
ness of the derived LAD model.
We have also computed the correlation between the discriminant values of the
LAD model and the bank credit risk ratings (represented on their numerical scale).
Despite the fact that the correlation measure takes into account all the banks in the
dataset (i.e., not only those which were used in creating the LAD model but also
those rated B/C, C, or C/D, which were not used in the learning process), the
correlation is very high, equal to 80.70 %, attesting to the strong predictive power of
the LAD model. The ten two-folding experiments described above were then used
to verify the stability of the correlation between the LAD discriminant values and
the bank ratings. The average correlation is equal to 80.04 %, with a standard
deviation of 0.04, another testimony of the stability of the close positive association
between the LAD discriminant values and the bank ratings.

16.3.4.2 From LAD Discriminant Values to Ratings

The objective of this section is to map the numerical values of the LAD
discriminant to the nine bank rating categories (A, A/B, . . ., E). This will be
accomplished using a nonlinear optimization problem that partitions the interval of
the discriminant values into nine subintervals that we associate to the nine rating
categories. The partitioning is determined through cut points xi such that
1 ¼ x0  x1  x2  . . .. . .. . .  x8  x9 ¼ 1, where i indexes the rating categories
(with one corresponding to E and nine corresponding to A). A bank should be rated i if
its discriminant value is in the interval [xi, xi+1]. Since such a perfect partitioning may
not exist, we replace the LAD discriminant values di by an adjusted discriminant value
di and find values of di for which such a partitioning exists and which are “as close as
possible” to the values di. The expression “as close as possible” involves the minimi-
zation of the mean square approximation error. Referring to the rating category of bank
i by j(i) and to the set of banks by N, we solve the convex nonlinear problem
X
minimize ð di  d i Þ 2
i2N

subject to di  xjðiÞþ1 , i 2 N
(16.3)
xjðiÞ < di , i 2 N
 1 ¼ x0  x1  x2 , . . . ,  xj , . . . , x8  x9 ¼ 1
 1  di  1, i 2 N
on the NEOS server (Czyzyk et al. 1998) with the solver Lancelot to determine the cut
points xj and the adjusted discriminant values di that will be used for rating the banks
not only in the training sample but also those which are not. In this case, the bank rating
is defined by the particular subinterval containing the LAD discriminant value.
16 Combinatorial Methods for Constructing Credit Risk Ratings 453

Table 16.4 Discrepancy analysis

qk
N ¼ {A, A/B, B, D, N ¼ {A, A/B, B, B/C,
k D/E, E} N ¼ {B/C, C, C/D} C, C/D, D, D/E, E}
0 29.60 % 29.36 % 29.50 %
1 53.28 % 47.71 % 51.00 %
2 12.05 % 18.96 % 14.88 %
3 4.23 % 3.36 % 3.88 %
4 0.85 % 0.61 % 0.75 %
5–6–7–8 0.00 % 0.00 % 0.00 %

As compared to ordered logistic regression, the LAD-based approach has the

additional advantage that it can be used for generating any desired number of rating
categories. The LAD model can take the form of a binary classification model and
can be used to preapprove or not a loan to a bank. The LAD rating approach can
also be used to derive models with higher granularity (i.e., more than nine rating
categories), which are used by banks to further differentiate their customers and to
tailor accordingly their credit pricing policies.

16.3.4.3 Conformity of Fitch and LAD Bank Ratings

In this section, the goal is to investigate to goodness of fit of our rating system and
observe to which extent the original LAD discriminant values fit in the identified
rating subintervals. We denote by qk (k ¼ 0, . . ., 8) the fraction of banks whose
rating category determined in this way differs from its Fitch rating by exactly
k categories. It is worth recalling that the rating cut points were derived using all
the banks in the sample but that the LAD discriminant values were obtained by only
taking into account the banks rated A, A/B, B, D, D/E, and E. This is why we
calculate separately the discrepancy counts for the banks rated A, A/B, B, D, D/E,
and E and for these rated B/C, C, and C/D.
Table 16.4 highlights the high goodness of fit of the proposed LAD model. More
than 95 % of the banks are rated within two categories of their Fitch rating, with
about 30 % of the banks receiving exactly the same rating as in the Fitch rating
system and another 51 % being off by exactly one rating category. The very high
concordance between the LAD and the Fitch ratings is illustrated by the weighted
average distances between the two ratings equal to (i) 0.93 for the categories
A, A/B, B, D, D/E, and E; (ii) 0.98 for the categories B/C, C, and C/D; and (iii)
0.95 for all banks in the dataset. The stability of the proposed rating system and its
suitability to evaluate the creditworthiness of “new” banks, that is, banks which are
not rated by agencies or banks the rater has not dealt with before, are magnified by
the fact that the goodness of fit of the ratings of the banks not used in deriving the
LAD model is very close to the goodness of fit for the banks used for deriving the
LAD model (i.e., those rated A, A/B, B, D, D/E, and E).
In order to appraise the robustness of the proposed rating system, we apply ten
times the two-folding procedure described above to derive the average (qk )
454 A. Kogan and M.A. Lejeune

Table 16.5 Cross-validated discrepancy analysis

k
0 1 2 3 4 5 6 7 8
qk 29.01 % 51.18 % 14.10 % 4.36 % 1.14 % 0.20 % 0.01 % 0.00 % 0.00 %

discrepancy counts (Table 16.5) of the bank ratings predicted for the testing sets.
The fact that on the average the difference between the Fitch and the LAD ratings is
only 0.98 is a very strong indicator of the LAD model’s stability and the absence of
overfitting.

16.3.4.4 Importance of Variables

While the focus in LAD is on discovering how the interactions between the values
of small groups of variables (as expressed in patterns) affect the outcome (i.e., the
bank ratings), one can also use the LAD model to learn about the importance of
individual variables. A natural measure of importance of a variable in an LAD
model is the frequency of its appearance in the model’s patterns. The three most
important variables in the 22 patterns constituting the LAD model are the credit risk
rating of the country where the bank is located, the return on average total assets,
and the return on average equity. The importance of the country risk rating variable,
which appears in 18 of the 22 patterns, can be explained by the fact that credit rating
agencies are reluctant to give an entity a better credit risk rating than that of the
country where it is located. Both the return on average assets and the return on
average equity variables appear in six patterns. These two ratios, respectively,
representing the efficiency of assets in generating profits and that of shareholders’
equity in generating profits, are critical indicators of a company’s prosperity and are
presented by Sarkar and Sriram (2001) as key predictors to assess the wealth of
a bank. The return on average equity is also found significant for predicting the
rating of US banks (Huang et al. 2004).

16.3.5 Remarks on Reverse Engineering Bank Ratings

The results presented above demonstrate that the LAD approach can be used to
derive rating models with varying granularity levels:
• A binary classification model to be used for the preapproval operations
• A model with the same granularity as the benchmarked rating model
• A model with higher discrimination power, that is, with higher granularity than
that of the benchmarked rating model, to allow the bank to refine its pricing
policies and the allocation of regulatory capital
We show that the LAD model cross-validates extremely well and therefore is
highly generalizable and could therefore be used by financial institutions to develop
internal, Basel-compliant rating models.
The availability and processing of data have been major obstacles in the way of
using credit risk rating models. Until recently, many banks did not maintain such
16 Combinatorial Methods for Constructing Credit Risk Ratings 455

datasets and were heavily dependent on qualitative judgments. It is only after the
currency crises of the 1990s and the requirements imposed by the Basel Accord that
financial institutions have seen an incentive in collecting the necessary data and
maintaining the databases.
The move towards a heavier reliance on rating models is based on the assump-
tion that models produce more consistent ratings and that, over the long haul,
operating costs will diminish since less labor will be required to produce ratings.
The model proposed in this chapter will reinforce the incentives to develop and rely
upon credit risk models in bank operations due to the following:
• The accuracy and predictive ability of the proposed model will guarantee
dependable ratings.
• Its parsimony will alleviate the costs of extracting and maintaining large
datasets.
• It will result in leaner loan approval operations and faster decisions and could
thus reduce the overall operating costs.

16.4 Relative Creditworthiness: Country Risk Ratings

16.4.1 Problem Description

Country risk ratings have critical importance in the international financial markets
since they are the primary determinants of the interest rates at which countries can
obtain credit.5 There are numerous examples of countries having to pay higher rates
on their borrowing following their rating downgrade, an often cited example being
Japan. As mentioned above, another critical aspect of country risk ratings concerns
their influence on the ratings of national banks and companies that would make
them more or less attractive to foreign investors. That is why the extant literature
calls country risk ratings the “pivot of all other country’s ratings” (Ferri et al. 1999)
and considers them the credit risk ceiling for all obligors located in a country
(Eliasson 2002; Mora 2006). Historical record shows the reluctance of raters to give
a company a higher credit rating than that of the sovereign where the company
operates. Contractual provisions sometimes prohibit institutional investors from
investing in debt rated below a prescribed level. It has been demonstrated (Ferri
et al. 1999) that variations in sovereign ratings drastically affect the ratings of banks
operating in low-income countries, while the ratings of banks operating in high-
income countries (Kaminsky and Schmukler 2002; Larrain et al. 1997) do not
depend that much on country ratings. Banks, insurance companies, and public
institutions frequently assess the amount of exposure they have in each country
and establish lending limits taking into account the estimated level of country risk.
Credit managers in multinational corporations have to assess evolving conditions in
foreign countries in order to decide whether to request letters of credit for particular

5
The presentation in this section is based on Hammer et al. (2006, 2012).
456 A. Kogan and M.A. Lejeune

transactions. Country risk estimates and their updates are utilized on a real-time
basis by multinational corporations facing constant fluctuation in international
currency values and difficulties associated with moving capital and profits across
national boundaries. Financial institutions have to rely on accurate assessments of
credit risk to comply with the requirements of the Basel Bank for International
Settlements. Feeling the pressure to get higher ratings could lead to fraud attempts.
A notorious case was Ukraine’s attempt to obtain IMF credits through misleading
reporting of its reserve data on foreign exchanges.
The existing literature on country risk, defined by Bourke and Shanmugam
(1990) as “the risk that a country will be unable to service its external debt due to
an inability to generate sufficient foreign exchange,” recognizes both financial/
economic and political components of country risk. There are two basic approaches
to the interpretation of the reasons for defaulting. The first one is the debt-service
capacity approach which considers the deterioration of solvency of a country as
preventing it from fulfilling its commitments. This approach views country risk as
a function of various financial and economic country parameters. The second one is
the cost-benefit approach which considers a default on commitments or
a rescheduling of debt as a deliberate choice of the country. In this approach the
country accepts possible long-term negative effects (e.g., the country’s exclusion
from certain capital markets (Reinhart 2002) as preferable to repayment). Being
politically driven, this approach includes political country parameters in addition to
the financial and economic ones in country risk modeling (Brewer and Rivoli 1990,
1997; Citron and Neckelburg 1987).

16.4.2 Data

16.4.2.1 Ratings
We analyze in this chapter Standard & Poor’s foreign currency country ratings, as
opposed to the ratings for local currency debt. The former is the more important
problem, since the sovereign government has usually a lower capacity to repay
external (as opposed to domestic) debt, and as an implication, the international
bond market views foreign currency ratings as the decisive factor (Cantor and
Packer 1996). This is manifested by much higher likelihood for international
investors to acquire foreign currency obligations rather than domestic ones. In
evaluating foreign currency ratings, one has to take into account not only the
economic factors but also the country intervention risk, that is, the risk that
a country imposes, for example, exchange controls or a debt moratorium. The
evaluation of local currency ratings need not take into account this country
intervention risk.
Table 16.16 lists the different country risk levels used by S&P’s and also pro-
vides descriptions associated with these labels. A rating inferior to BB+ indicates
that a country is non-investment grade (speculative). A rating of CCC+ or lower
indicates that a country presents serious default risks. BB indicates the least degree
of speculation and CC the highest. The addition of a plus or minus sign modifies the
16 Combinatorial Methods for Constructing Credit Risk Ratings 457

rating (between AA and CCC) to indicate relative standing within the major rating
category. These subcategories are treated as separate ratings in our analysis.

16.4.2.2 Selected Variables

The selection of relevant variables is based on three criteria. The first criterion is the
variable’s significance (relevance) in assessing countries’ creditworthiness. The
review of the literature on predictors for country risk ratings (see Hammer
et al. 2011) played an important role in defining the set of candidate variables to
include in our model. The second criterion is the availability of complete and
reliable statistics, to allow us to maintain the significance and the scope of our
analysis. The third criterion is the uniformity of data across countries. This is why
we decided against incorporating the unemployment rate statistics provided by the
World Bank, since it is compiled according to different definitions: there are
significant differences between countries in the treatment of temporarily laid off
workers, those looking for their first job, and the criteria for being considered as
unemployed.
After applying the criteria of relevance, availability, and uniformity described
above, the following variables were incorporated in our model: gross domestic
product per capita (GDPc), inflation rate (IR), trade balance (TB), exports’ growth
rate (EGR), international reserves (RES), fiscal balance (FB), debt to GDP
(DGDP), political stability (PS), government effectiveness (GE), corruption
(COR), exchange rate (ER), and financial depth and efficiency (FDE).
In addition to eight economic variables that have already been used in the
country credit risk rating literature, a new one, called financial depth and efficiency
(FDE), was added. This variable is measured as the ratio of the domestic credit
provided by the banking sector to the GDP. It measures the growth of the banking
system since it reflects the extent to which savings are financial. The financial depth
and efficiency variable was first considered in Hammer et al. (2006, 2011) for the
evaluation of country risk ratings. The reason to include this variable is that it
captures some information that is relevant to the assessment of the creditworthiness
of a country and that is not accounted for by other variables and in particular by the
fiscal balance of a country.
Our analysis utilized the values of these nine economic/financial variables and
three political variables taken at the end of 1998. Our dataset included the values of
these 12 variables for the 69 countries considered: 24 industrialized countries,
11 Eastern European countries, eight Asian countries, ten Middle Eastern countries,
15 Latin American countries, and South Africa. The dependent variable is the
S&P’s country risk ratings for these countries at the end of 1998. The sources
utilized for compiling the values of the economic/financial variables include the
International Monetary Fund (World Economic Outlook Database), the World
Bank (World Development Indicators database), and, for the ratio of debt to gross
domestic product, Moody’s publications. Values of political variables are taken
from Kaufmann et al. (1999a, b), whose database is a joint product of the Macro-
economics and Growth, Development Research Group and Governance, Regulation
and Finance Institutes affiliated with the World Bank.
458 A. Kogan and M.A. Lejeune

16.4.3 Rating Methodologies

In the next subsections, we derive two new combinatorial models for country risk
rating by reverse engineering and “learning” from past S&P’s ratings. The models
are developed using the novel combinatorial-logical technique of logical analysis of
data which derives a new rating system only from the qualitative information
representing pairwise comparisons of country riskiness. The approach is based on
a relative creditworthiness concept, which posits that the knowledge of (pre)order
of obligors with respect to their creditworthiness should be the sole source for
creating a credit risk rating system. Stated differently, only the order relation
between countries should determine the ratings. Thus, the inference of a model
for the order relation between countries is the objective of this study. This is in
perfect accordance with the general view of the credit risk rating industry. Altman
and Rijken (2004) state that the objective of rating agencies is “to provide an
accurate relative (i.e., ordinal) ranking of credit risk,” which is confirmed by
Fitch ratings (2006) saying that “Credit ratings express risk in relative rank order,
which is to say they are ordinal measures of credit.” Bhatia (2002) adds that:
“Although ratings are measures of absolute creditworthiness, in practice, the ratings
exercise is highly comparative in nature. . . On one level, the ratings task is one of
continuously sorting the universe of rated sovereigns – assessed under one uniform
set of criteria – to ensure that the resulting list of sovereigns presents a meaningful
global order of credit standing. On another level, the sorting task is constrained by
a parallel need to respect each sovereign’s progression over time, such that shifting
peer comparisons is a necessary condition – but not a sufficient one – for upward or
downward ratings action.”
A self-contained model of country risk ratings can hardly be developed by
standard econometric methods since the dataset contains information only
about 69 countries, each described by 12 explanatory variables. An alternative at
hand possible with combinatorial techniques is to examine the relative riskiness of
one country compared to another one, rather than modeling the riskiness of each
individual country. This approach has the advantage of allowing the modeling to be
based on a much richer dataset (2,346 pairs of countries), which consists of the
comparative descriptions of all pairs of countries in the current dataset.
The models utilize the values of nine economic and three political variables
associated to a country, but do not use directly or indirectly previous years’ ratings.
This is a very important feature, since the inclusion of information from past ratings
(lagged ratings, rating history) does not allow the construction of a self-contained
rating system and does not make possible to rate the creditworthiness of not-yet-
rated countries. We refer to Hammer et al. (2006, 2011) for a more detailed
discussion of the advantages of building a non-recursive country risk rating system.
Moreover, the proposed LAD models completely eliminate the need to view the
ratings as numbers. Section 16.4.3.1 presents the common features of the two
developed LAD models, while Sects. 16.4.3.2 and 16.4.3.3 discuss the specifics
of the LAD-based Condorcet ratings and the logical rating scores, respectively.
16 Combinatorial Methods for Constructing Credit Risk Ratings 459

16.4.3.1 Commonalities
Pairwise Comparison of Countries: Pseudo-observations
Every country i 2 I ¼ {1, . . ., 69} in this study is described by the 13-dimensional
vector Ci, whose first component is the country risk rating given by Standard &
Poor’s, while the remaining 12 components specify the values of the nine eco-
nomic/financial and of the three political variables. A pseudo-observation Pij,
associated to every pair of countries i, j 2 I, provides in a way specified below
a comparative description of the two countries.
Every pseudo-observations is also described by a 13-dimensional vector. The
first component is an indicator which takes the value 1 if the country i in the pseudo-
observation Pij has a higher rating than the country j, –1 if the country j has a higher
rating than the country i, and 0 if the two countries have the same rating. The other
components k, k ¼ 2, . . ., 13 of the pseudo-observation Pij[k] are derived by taking
the differences of the corresponding components of Ci and Cj:

Pij ½k ¼ Ci ½k  Cj ½k, k ¼ 2, . . . , 13 (16.4)

Transformation (16.4) alleviates the problems related to the small size (j I j)

of the original dataset by constructing a substantially larger dataset containing
j I j*(j I j 1) pseudo-observations. While the set of pseudo-observations is not
independent, since Phi + Pij ¼ Phj, Hammer et al. (2006) show that this does not
create any problems for the LAD-based combinatorial data analysis techniques.
We illustrate the construction of pseudo-observations with Japan and Canada.
Rows 2 and 3 in Table 16.6 display the values of the 12 economic/financial and
political variables, as well as the S&P’s rating for Japan and Canada (at the end of
December 1998), while rows 4 and 5 report the pseudo-observations PJapan, Canada and
PCanada, Japan from the country observations CJapan and CCanada. Since Japan and
Canada are rated, respectively, AAA and AA + by S&P’s at the end of December
1998 and the rating AAA is better than rating AA+, the first component of the pseudo-
observation vector is equal to 1. The set of pseudo-observations is antisymmetric.
An advantage of this transformation is that it allows us to avoid the problems
posed by the fact that the original dataset contains only a small number (j I j)
of observations. The transformation (16.4) provides a larger dataset containing
j I j*(j I j 1) pseudo-observations.

Construction of Relative Preferences

The LAD-based reverse engineering of Standard & Poor’s rating system starts with
deriving an LAD model (constructed as a weighed sum of patterns) from the
archive of all those pseudo-observations Pij, corresponding to pairs of countries
i and j having different S&P’s ratings. A model resulting from applying LAD to the
1998 dataset consists of 320 patterns. As an example, let us describe two of these
patterns below. The positive pattern

FDE > 28:82; GDPc > 1539:135; GE > 0:553

 460

Table 16.6 Examples of country and pseudo-observations

S&P’s rating FDE RES IR TB EGR GDPc ER FB DGDP PS GE COR
CJapan AAA 138.44 5.168 .65 21.7471 2.54 24314.2 0.839 7.7 0.47 1.153 0.839 0.724
CCanada AA+ 94.69 1.01964 0.99 55.9177 8.79 24855.7 0.939 0.9 0.5 1.027 1.717 2.055
PJapan, Canada 1 43.75 4.15 0.34 34.17 11.33 541.5 0.1 8.6 0.03 0.126 0.878 1.331
PCanada, Japan 1 43.75 4.15 0.34 34.17 11.33 541.5 0.1 8.6 0.03 0.126 0.878 1.331
A. Kogan and M.A. Lejeune
16 Combinatorial Methods for Constructing Credit Risk Ratings 461

can be interpreted in the following way. If country i is characterized by

(i) A financial depth and efficiency (FDE) exceeding that of country j by at least
28.82, and
(ii) A gross domestic product per capita (GDPc) exceeding that of country j by at
least 1,539.135, and
(iii) A government efficiency (GE) exceeding that of country j by at least 0.553
then country i is perceived as more creditworthy than country j.
Similarly, the negative pattern

GDPc < 4, 886:96; ER < 0:195; COR < 0:213

can be interpreted in the following way. If country j is characterized by

(i) A gross domestic product per capita (GDPc) exceeding that of country i by
4,886.96, and
(ii) An exports growth rate (EGR) exceeding that of country i by 0.195, and
(iii) A level of incorruptibility (CR) exceeding that of country i by 0.213
then country i is perceived as less creditworthy than country j.
The constructed LAD model allows us to compute the discriminant D(Pij) for
each pseudo-observation Pij(i 6¼ j). We call the values D(Pij) of the discriminant the
relative preferences. They can be interpreted as measuring how “superior” the
country i’s rating over that of country j. We call the [69
69]-dimensional
antisymmetric matrix D, having the relative preferences as components, the relative
preference matrix. Its components are relative preferences D(Pij) (i 6¼ j) associated
with every pair of countries, including those that have the same S&P’s ratings, even
though only those pseudo-observation Pij for which i and j had different ratings
were used in deriving the LAD model.

Classification of Pseudo-observations and Cross-Validation

The dataset used to derive the LAD model contains 4,360 pseudo-observations,
with an equal number of positive and negative pseudo-observations, and is anti-
symmetric (Pij ¼ Pij). The results of applying the LAD model to classify the
observations in the dataset are presented in Table 16.7. The overall classification
quality of the LAD model (according to Eq. 16.2) is 95.425 %.
This very high classification accuracy may be misleading, since the constructed
LAD model can be overfitting the data, that is, it is adapted excessively well to
random noise in the training data and thus has an excellent accuracy on this training
data, but could perform very poorly on new observations. We therefore utilize
a statistical technique known as “jackknife” (Quenouille 1949) or “leave-one-out”

Table 16.7 Classification matrix

Classified as
Positive Negative Unclassified Total
Positive observations 93.90 % 3.72 % 2.38 % 100 %
Negative observations 3.72 % 93.90 % 2.38 % 100 %
462 A. Kogan and M.A. Lejeune

Table 16.8 Classification matrix for DJK

Classified as
Positive Negative Unclassified Total
Positive observations 93.49 % 3.67 % 2.84 % 100 %
Negative observations 3.67 % 93.49 % 2.84 % 100 %

to show that the achieved high classification accuracy is not due to overfitting. This
technique removes from the dataset one observation at a time, learns a model from
all the remaining observations, and evaluates the resulting model on the removed
observation; then it repeats all these steps for each observation in the dataset. If on
the average the predicted evaluations are “close to” the actual ones, then the model
is not affected by overfitting.
In the case at hand, it would not be statistically sound to implement the jackknife
technique in a straightforward way because of the dependencies among pseudo-
observations (Hammer et al. 2006). Indeed, even after explicitly eliminating a single
pseudo-observation Pij from the dataset, it would still remain in the dataset implicitly,
since Pih + Phj + Pij. This problem can be resolved by modifying the above described
procedure so that at each step, instead of just removing a single pseudo-observation Pij,
all the pseudo-observations which involve a particular country i are removed. Then, the
LAD discriminant is derived on the basis of the remaining pseudo-observations and
used to evaluate the relative preferences for every removed pseudo-observation,
resulting in a row of relative preferences of all pseudo-observations Pij which involve
the country i. This modified procedure is repeated for every country in the dataset, and
the obtained rows are combined into a matrix of relative preferences denoted by DJK.
The absence of overfitting is indicated by a very high correlation level of 96.48 %
between the matrix DJK and the original relative preference matrix D.
To further test for overfitting, we use the obtained matrix of relative preferences
DJK to classify the dataset of 4,360 pseudo-observations. The results of this classifi-
cation are presented in Table 16.8, and the overall classification quality of the LAD
model (according to formula (16.2)) is 95.10 %. The results in Table 16.8 are virtually
identical to those shown in Table 16.7, thus proving the absence of overfitting.
Since the derived LAD model does not suffer from overfitting, it is tempting to
interpret the signs of relative preferences as indicative of rating superiority and
infer that a positive value for D(Pi,j) indicates that country i is more creditworthy
than country j, while a negative one for D(Pi,j) justifies the opposite conclusion.
However, this naı̈ve approach towards relative rating superiority ignores the poten-
tial noise in the data and in relative preferences, which would make it difficult to
transform classifications of pseudo-observations to a consistent ordering of coun-
tries by their creditworthiness. It is shown in Hammer et al. (2006) that the
relationship based on the naı̈ve interpretation of the relative preferences can violate
the transitivity requirement of an order relation and therefore does not provide
a consistent partially ordered set of countries.
The following section is devoted to overcoming this issue by relaxing the overly
constrained search for (possibly nonexistent) country ratings whose pairwise
16 Combinatorial Methods for Constructing Credit Risk Ratings 463

orderings are in precise agreement with the signs of relative preferences. Instead,
we utilize a more flexible search for a partial order on the set of countries, which
satisfies the transitivity requirements and approximates well the set of relative
preferences.

16.4.3.2 Condorcet Ratings

The ultimate objective of this section, which is based on the results of Hammer
et al. (2006), is to derive a rating system from the LAD relative preferences. This is
accomplished in two stages. First, we use the LAD relative preferences to define
a partial order on the set of countries which represents their creditworthiness dominance
relationship. Then, we extend the derived dominance relationship to two rating systems,
respectively, based on the so-called weak Condorcet winners and losers. The number of
rating categories in these rating systems is the same and will be determined by the
structure of the partial order obtained in the first stage, that is, it is not a priory fixed.

From LAD Relative Preferences to a Partial Order on the Set of Countries

A partial order P(X) is a reflexive, antisymmetric, and transitive binary relation on
a set X. Any two distinct elements x and y of X such that (x, y) 2 P(X) are said to be
comparable and denoted by x  y. If neither (x, y) 2 P(X) nor (y, x) 2 P(X), then
x and y are called incomparable and denoted by x || y.
As was mentioned above, the naı̈ve rating superiority relation based on the LAD
model is not transitive. This difficulty can be overcome by defining a strengthened
version of the naı̈ve rating superiority relation, to be called the dominance rela-
tionship. While the former only relies on the sign of the relative preference D(Pij),
the definition of dominance of a country i over another country j takes into account
not only the sign of the relative preference D(Pij) but also the values of the relative
preferences of each of these two countries i and j over every other country k.
We will need the following notation to define the dominance relationship. Let
 
Sij ðkÞ ¼ DðPik Þ  D Pjk (16.5)

define the external preference of country i over country j with respect to country k, let
X
Sij ðkÞ
j2I
Sij ¼ (16.6)
jI j

define the average external preference of i over j, and let

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX    2
u DðPik Þ  D Pjk  Sij
u
t k2C
sij ¼ (16.7)
jI j

define the standard deviation of the external preference of i over j.

464 A. Kogan and M.A. Lejeune

The dominance relationship of a country i over another country j will be defined

by two conditions. The first one requires that D(Pij) > 0. It might also seem logical
to require that Sij(k) > 0 for every country k, k 6¼ i, j. However, this condition is so
difficult to satisfy that the resulting partially ordered set is extremely sparse, with
very few country pairs that are comparable. Thus, a more relaxed second condition
will require that, only at a certain confidence level, the external preference of i over
j should be positive. We parameterize the level of confidence by the multiplier  of
the standard deviation sij. More formally, for a given  > 0,
• A country i is said to dominate another country j if

 
D Pij > 0
(16.8)
Sij  sij > 0

• A country i is said to be dominated by another country j if

 
D Pij < 0
(16.9)
Sij þ sij < 0

• In all the other cases, countries i and j are said to be not comparable; this can be
due to the lack of evidence, or to conflicting evidence about the dominance of
i over j.
Note that the larger the value of  is, the stronger the conditions are, and the
fewer pairs of countries are comparable. If  is sufficiently large, then the domi-
nance relationship is transitive and is a partial order, since it becomes progressively
sparser until being reduced to empty. On the other hand, if  is small (and in
particular for  ¼ 0), then the dominance relationship becomes denser, but is not
necessarily transitive.
We are looking for a “rich” dominance relationship (applying to as many
country pairs as possible) which is transitive. Formally, our objective is to maxi-
mize the number of comparable country pairs subject to preserving the transitivity
of the dominance relationship.
The richest dominance relationship defined by the two conditions (16.8) and
(16.9) in the case of  ¼ 0 will be called the base dominance relationship. If i and
j are any two countries comparable in the base dominance relationship, then it is
possible to calculate the smallest value of the parameter ij ¼ j Sij/sij j such that the
i and j are not comparable in any dominance relationship defined by a parameter

Table 16.9 Correlation levels between LAD and canonical relative preferences
D dS&P dM dII dLRS
D 100 % 93.21 % 92.89 % 91.82 % 97.57 %
dS&P 93.21 % 100 % 98.01 % 96.18 % 95.54 %
dM 92.89 % 98.01 % 100 % 96.31 % 95.20 %
dII 91.82 % 96.18 % 96.31 % 100 % 94.11 %
dLRS 97.57 % 95.54 % 95.20 % 94.11 % 100 %
16 Combinatorial Methods for Constructing Credit Risk Ratings 465

value exceeding or equal to ij. This calculation is based on the fact that given
a value of , it is possible to check in polynomial time (Tarjan 1972) whether the
corresponding dominance relationship is transitively closed. Then the algorithm
that determines in polynomial time the minimum value * for which the
corresponding dominance relationship is still transitive, sorts at most j I j2 numbers
ij in ascending order, and then checks one by one the transitivity of the
corresponding dominance relationships. When the dominance relationship becomes
transitive for the first time, the algorithm stops and outputs * equal to the
corresponding value of the parameter ij. This study utilizes this value * and the
corresponding dominance relationship between countries, called here the logical
dominance relationship and denoted by the subscript of LAD (e.g.,  ).
LAD
The definition of dominance relationship between countries on the basis of
average external preferences bears some similarities to the so-called “column
sum methods” (Choo and Wendley 2004) utilized in reconciling inconsistencies
resulting from the application of pairwise comparison matrix methods.

Extending Partially Ordered Sets to “Extreme” Linear Preorders

The information about country preferences contained in the economic and political
attributes is represented most faithfully by the logical dominance relationship
defined above. However, it is impractical to use, since this partial order requires
a large amount of data to describe. On the other hand, country preferences can be
expressed very compactly by country ratings, since the latter are a very special type
of partial orders called linear preorders.
A partial order P(X) is called a linear preorder if there exists a mapping
M : X ! {0,1, . . ., k} such that x  y if and only if M(x) > M(y). Therefore,
a linear preorder is completely described by specifying its mapping M. Without loss
of generality, one can assume that for every i2 {0,1, . . ., k}, there exists x 2 X such
that M(x) ¼ i. Such a linear preorder is said to have k + 1 levels.
To make logical dominance of countries practically utilizable, this relationship
should be transformed into a linear preorder preserving all the order relations between
countries (i.e., is an extension of the partial order) and is as close as possible to it. The
logical dominance relationship can be extended in a multitude of ways to a variety of
linear preorders. In particular, two extreme linear preorders are constructed below that
we call the optimistic and the pessimistic extensions, denoted by OE and PE, respec-
tively. The names are justified since the former assigns to each country the highest level
it can expect, while the latter assigns to each country the lowest level it can expect:
• In the first step of OE construction, those countries that are not dominated by any
other country are assigned the highest level and are then removed from the set of
countries under consideration. Iteratively, non-dominated countries in the
remaining set of countries are assigned the highest remaining level until every
country is assigned a level denoted by OEi.
• In the first step of PE construction, those countries that do not dominate any
other country are assigned the lowest level and are then removed from the set of
countries under consideration. Iteratively, non-dominating countries in the
466 A. Kogan and M.A. Lejeune

remaining set of countries are assigned the lowest remaining level until every
country is assigned a level denoted by PEi.
The method utilized above to construct OE and PE is known as the Condorcet
method. It represents a specific type of voting system (Gehrlein and Lepelley
1998), and it is often used to determine the winner of an election. The Condorcet
winner(s) of an election is generally defined as the candidate(s) who, when
compared in turn with every other candidate, is preferred over each of them.
Given an election with preferential votes, one can determine weak Condorcet
winners (Ng et al. 1996) by constructing the Schwartz set as the union of all
possible candidates such that (i) every candidate inside the set is pairwise
unbeatable by any other candidate outside the set (ties are allowed) and (ii) no
proper subset of the set satisfies the first property.
The Schwartz set consists exactly of all weak Condorcet winners. The weak
Condorcet losers are the reverse of the weak Condorcet winners, that is, those
losing pairwise to every other candidate. OE assigns the highest level to those
countries that are the weak Condorcet winners, that is, better than or incomparable
with every other country. PE assigns the lowest level to those countries that are
the weak Condorcet losers, that is, worse than or are incomparable with every
other country. Note that both OE and PE have the minimum possible number of
levels. Indeed, in a directed graph whose vertices are the countries and whose arcs
represent comparable countries in the dominance relationship, the length of the
longest directed path bounds from below the number of levels in any linear
preorder extending the dominance relationship. This length equals the number
of levels in OE and PE.

16.4.3.3 Logical Rating Scores

The LAD relative preferences can be utilized in a completely different way
(compared to OE or PE) to construct country risk ratings. We describe here how
to derive new numerical ratings of all countries, called “logical rating scores”
(LRS), by applying multiple linear regression. LRS were defined by Hammer
et al. (2011) as numerical values whose pairwise differences approximate
optimally the relative preferences over countries as expressed in their risk
ratings. A common way to calculate the relative preferences is based on
interpreting sovereign ratings as cardinal values (see, e.g., Ferri et al. 1999;
Hu et al. 2002; Sy 2004). If the sovereign ratings b are viewed as cardinal values,
then one can view the relative preferences D as differences of the corresponding
ratings:
 
D Pij ¼ bi  bj , for all i, j 2 I, i 6¼ j (16.10)
Since the system (16.10) is not necessarily consistent, it should be relaxed in the
following way:
 
D Pij ¼ bi  bj þ eij , for all i, j 2 I, i 6¼ j (16.11)
16 Combinatorial Methods for Constructing Credit Risk Ratings 467

The values of the b’s providing the best L2 approximation of the D’s can be
found by solving the following multiple linear regression problem:
X
DðpÞ ¼ bk xk ðpÞ þ eðpÞ, (16.12)
k2I

8
< 1, for k ¼ i
where p ¼ {(i, j)ji, j 2 I, i 6¼ j} and xk ði; jÞ ¼ 1, for k ¼ j .
:
0, otherwise
The logical rating scores bk obtained by fitting the regression model are given in
Column 8 of Table 16.16.

16.4.4 Evaluation of the Results

In this section, we analyze the results obtained with the proposed rating systems.
The evaluation of the results involves the analysis of the following:
• The relative preferences
• The partially ordered set (i.e., the logical dominance relationship)
• The LRS scores and the Condorcet ratings (extremal linear extensions)
with respect to the rating system of S&P’s as well as that of Moody’s and The
Institutional Investor.

16.4.4.1 Canonical Relative Preferences

In this section, we evaluate the quality of the informational content in the matrix D
of relative preferences. To reach this goal, we first define, for any set of numerical
scores si representing sovereign ratings, the canonical relative preferences
dij ¼ sisj for each pair of countries. Second, we compare the LAD relative
preferences D(Pij) with the canonical relative preferences dS&Pij, dMij, dIIij, and
dLRSij associated, respectively, to the S&P’s ratings, Moody’s ratings, The Institu-
tional Investor’s scores, and the logical rating scores (Hammer et al. 2011). The
corresponding matrices of relative preferences are denoted dS&P, dM, dII, and dLRS,
respectively. We evaluate the proximity between the LAD relative preferences
(Hammer et al. 2006) and the canonical relative preferences based on their corre-
lation levels (Table 16.9).
The high correlation levels show that the LAD relative preferences are in strong
agreement with both the ratings of S&P’s and those of the other agencies, as well as
with the logical rating scores. The superiority of the logical rating scores compared
to the relative preferences associated with the LAD discriminant is explained by the
filtering of the noise done when deriving the LRS scores.

16.4.4.2 Preference Orders

The logical dominance relationship is compared to the preference orders derived
from the S&P’s and Moody’s ratings (viewed as partially ordered sets) and the
468 A. Kogan and M.A. Lejeune

partially ordered sets associated with The Institutional Investor’s scores and the
logical rating scores. The partially ordered sets corresponding to The Institutional
Investor’s scores and the logical rating scores are obtained as follows:

i  j if si  sj > y
i≺j if si  sj < y
ijjj otherwise

where y is the positive number chosen so as to obtain a partially ordered set of the
same density as the dominance relationship and si represents the numerical score
given to country i by the respective rating system. The incomparability between two
countries means, for S&P’s and Moody’s, that the two countries are equally
creditworthy, while for the logical dominance relationship, it means that the
evidence about the relative creditworthiness of the two countries is either missing
or conflicting. The concept of density of a partially ordered set is defined in
Hammer et al. (2006) and represents the extent to which a partial order on a set
of countries differentiates them by their creditworthiness.
To assess the extent to which the preference orders agree with each other, the
following concepts are introduced (Hammer et al. 2006). Given a pair of countries
(i, j), two partially ordered sets are:
• In concordance if one of the following relations i  j, i ≺ j, or i || j holds for both
partially ordered sets
• In discordance if i  j for one of the partially ordered sets and i ≺ j for the other one
• Incomparable otherwise, that is, if i || j for one of the partially ordered sets, and
either i  j or i ≺ j for the other partially ordered set.
We measure the levels of concordance, discordance, or incomparability between
two partially ordered sets by the fractions of pairs of countries for which the two
partially ordered sets are, respectively, in concordance, discordance, or incompa-
rable. Hammer et al. (2006) have shown that there is a very high level of agreement
between the following:
• The logical dominance relationship and the preference orders associated with
S&P’s and Moody’s ratings and The Institutional Investor scores
• The logical dominance relationship and the logical rating scores

16.4.4.3 Discrepancies with S&P’s

Logical Dominance Relationship
This section is devoted to the study of the discordance between the logical domi-
nance relationship and the preference order of S&P’s. We define as a discrepancy
(Hammer et al. 2006) a country pair for which the logical dominance relationship
and the preference order of S&P’s are in discordance. The 2.17 % discordance level
between the logical dominance relationship and the preference order of S&P’s
represents 51 discrepancies.
Next, in order to determine the minimum number of countries, for which the
S&P’s ratings must be changed so that the new adjusted S&P’s preference order has
16 Combinatorial Methods for Constructing Credit Risk Ratings 469

a 0 % discordance level with the dominance relationship, we solve the integer

program below:
X
min ai
i2I

subject to
 
 S  Si   M ai , for all i 2 I (16.13)
i

S i  S j for every pair ði; jÞ such that i  j

LAD

ai 2 f0; 1g, S i ef0; 1; . . . ; 21g for all i 2 I

where ai takes the value 1 if the S&P’s rating of country i must be modified and the
value 0 if otherwise, Si is the original S&P’s rating of country i, S*i is the adjusted
S&P’s rating of country i, and M is a sufficiently large positive number (e.g., M ¼ 22).
The optimal solution of Eq. 16.13 shows that the 0 % discordance level can be
achieved by adjusting the S&P’s ratings of nine countries: France, India, Japan,
Colombia, Latvia, Lithuania, Croatia, Iceland, and Romania. To check the rele-
vance of the proposed rating adjustments, we examine the S&P’s ratings posterior
to December 1998. We observe that Romania, Japan, and Columbia’s S&P’s
ratings have been modified in the direction suggested by our model. More precisely,
Columbia was downgraded by S&P’s twice, moving from BBB in December
1998 to BB+ in September 1999 and then to BB in March 2000. Japan was
downgraded to AA+ in February 2001 and AA in November 2001. Romania was
upgraded to B in June 2001. The S&P’s rating of the other countries (Iceland,
France, India, Croatia, Latvia, and Lithuania) has remained unchanged.

Logical Rating Scores

We have already shown that the LRS and the S&P’s ratings are in close agreement.
However, since LRS and the S&P’s ratings are not expressed on the same scale, the
comparison of the two scores of an individual country presents a challenge. In order
to bring the LRS and the S&P’s ratings to the same scale, we apply a linear
transformation a*bi + c to the logical rating scores bi in such a way that the mean
square difference between the transformed LRS and the S&P’s ratings is mini-
mized. This is obtained by solving a series of quadratic optimization problems in
which the decision variables are a and c. Clearly, the consistency of the LRS and
S&P’s ratings is not affected by this transformation.
In 1998, it appears that five countries (Columbia, Hong Kong, Malaysia,
Pakistan, and Russia) have an S&P’s rating that does not fall within the confidence
interval of the transformed LRS. The 1-year modification of the S&P’s ratings for
Columbia, Pakistan, and Russia is in agreement with the 1998 LRS of these countries and
highlights the prediction power of the LRS model. Moreover, the evolution of the S&P’s
ratings of Malaysia and Hong Kong is also in agreement with their 1998 LRS. Indeed,
both Malaysia and Hong Kong have been upgraded shortly thereafter, the former moving
from BBB- to BBB in November 1999 and the latter from A to A+ in February 2001.
470 A. Kogan and M.A. Lejeune

Table 16.10 Correlation analysis

S&P’s Moody’s II OE PE LRS
S&P’s 100 % 98.01 % 96.18 % 94.31 % 95.40 % 95.54 %
Moody 98.01 % 100 % 96.31 % 94.13 % 95.42 % 95.20 %
II 96.18 % 96.31 % 100 % 93.26 % 94.62 % 94.11 %
OE 94.31 % 94.13 % 93.26 % 100 % 99.15 % 99.24 %
PE 95.40 % 95.42 % 94.62 % 99.15 % 100 % 99.10 %
LRS 95.54 % 95.20 % 94.11 % 99.24 % 99.10 % 100 %

16.4.4.4 Optimistic and Pessimistic Extensions

The optimistic and pessimistic extensions of the logical dominance relationship
(Table 16.16) comprise 21 levels, while the S&P’s rating system contains 22 rating
categories. Table 16.10 provides the correlation levels between all the ratings
(scores). The analysis reconfirms the high level of agreement between the proposed
rating models and that of S&P’s. The high correlation levels attest that the LRS
approximate very well the S&P’s ratings, as well as those of the other rating
agencies.

16.4.4.5 Temporal Validity

In this section, we apply the “out-of-time” or “walk-forward” validation approach
(Sobehart et al. 2000; Stein 2002) to further verify the robustness and the relevance
of our rating models. This involves testing how well the LAD model derived from
the 1998 data performs when applied to the 1999 data. To evaluate the “temporal
validity” of the proposed models, we proceed as follows: (i) we derive the LAD
relative preferences, (ii) we build the logical dominance relationship and run the
regression model for the LRS scores, (iii) we calculate the weak Condorcet ratings
(pessimistic and optimistic extensions) and the logical rating scores, and (iv) we
compare these to the rating systems of the rating agencies (S&P’s, Moody’s, and
The Institutional Investor).

Relative Preferences
Table 16.11 shows that the LAD relative preferences are highly correlated with
those of the S&P’s rating system, as well as with the logical rating scores. The LAD
relative preferences and the LRS were obtained by applying to the 1999 data the
models derived from the 1998 data.
The high levels of pairwise correlations between the S&P’s 1999 ratings, the
relative preferences given by the LAD discriminant, and the canonical relative
preferences corresponding to LRS show that the LRS model has a very strong
temporal stability and indicate its high predictive power.

Preorders
The logical dominance relationship is compared to the 1999 preference order of the
S&P’s and the partially ordered set associated with the logical rating scores.
Table 16.12 displays the concordance, discordance, and incomparability levels
16 Combinatorial Methods for Constructing Credit Risk Ratings 471

Table 16.11 Correlation dS&P D dLRS

levels between relative
dS&P 100 % 91.70 % 94.12 %
preference matrices
D 91.70 % 100 % 96.98 %
dLRS 94.12 % 96.98 % 100 %

Table 16.12 Concordance, Logical dominance relationship

discordance, and
Concordance Incomparability Discordance
incomparability levels with
dominance relationship S&P’s 84.21 % 13.15 % 2.64 %
Logical rating 93.43 % 6.49 % 0.08 %
score
Logical rating score
Concordance Incomparability Discordance
S&P’s 83.46 % 12.69 % 3.85 %

between the logical dominance relationship, the preference order of S&P’s, and the
partial order associated with the logical rating scores and underlines their very
strong level of agreement.

Discrepancies with S&P’s

The discordance level between the logical dominance relationship obtained
using the 1999 data and the preference order of the 1999 S&P’s ratings is equal
to 2.64 %. The solution of the above integer programming problem (16.13) reveals
that the discrepancies would disappear if one modified the ratings of eight countries
(France, Japan, India, Colombia, Latvia, Croatia, Iceland, and Hong Kong). The
relevance of the ratings obtained with the logical dominance relationship is proven
by observing the rating changes published by S&P’s subsequent to December 1999.
The identification of the discrepancies between S&P’s ratings and the LRS
scores requires the derivation of the confidence intervals for the new, 1999 obser-
vations and therefore for the transformed LRS of each country (Hammer
et al. 2007). We denote by n and p the number of observations and predictors,
respectively. The expression t(1  a/2, n  p) refers to the Student test with (n  p)
degrees of freedom, and with upper and lower tail areas of a/2, Xj is the
p-dimensional vector of the values taken by the observation Yj on the
0
p predictors, Xp is the transposed of Xj, and (X0 X)1 is the variance-covariance
matrix, that is, the inverse of the [p
p]-dimensional matrix (X0 X). Denoting
by MSE the h mean square i of errors, the estimated variance
 
2 ^ 0  0 1

s Y j ¼ MSE 1 þ X X X Xj of the predicted rating Y ^ j and the (1a)
j
^ j, n are given by
confidence interval for Y

^ j, n  tð1  a=2, n  pÞ s½pred , Y

Y ^ j, n þ tð1  a=2, n  pÞ s½pred  (16.14)
472 A. Kogan and M.A. Lejeune

Hammer et al. (2007b) say that there is a discrepancy between S&P’s rating RSP j
and the logical
rating score if

^ j, n  tð1  a=2, n  pÞ s½pred , Y

^ j, n þ tð1  a=2, n  pÞ s½pred 
j 2
RSP = Y for
a ¼ 0.1.
Applying the 1998 LRS model to the 1999 data, only two countries (Russia and
Hong Kong) have S&P’s ratings that are outside the confidence intervals of the
corresponding transformed LRS. The creditworthiness of these two countries seems
to have been under-evaluated by S&P’s in 1998: the ratings of both were upgraded
in 1999.

Condorcet Ratings and LRS Scores

The optimistic and pessimistic extensions (Table 16.16) of the logical dominance
relationship obtained using the 1999 data both comprise 20 levels, while the S&P’s
rating system contains 22 rating categories. Table 16.13 provides the correlation
levels between the 1999 S&P’s ratings, the optimistic and pessimistic extension
levels, and the logical rating scores. The high levels of correlation and their
comparison with those presented in Table 16.10 provide further evidence of the
temporal validity of the proposed models.

16.4.4.6 Predicting Creditworthiness of Unrated Countries

Condorcet Approach
The application of the logical dominance relationship to predict the rating of
countries not included in the original dataset, and for years subsequent to 1998, is
an additional validation procedure, sometimes referred to as “out-of-universe”
cross-validation (Sobehart et al. 2000; Stein 2002). We use the 1998 LAD model
to calculate the relative preferences for all pseudo-observations involving one or
two of the four “new” countries (Ecuador, Guatemala, Jamaica, Papua New
Guinea), which allows us to derive the logical dominance relationship and the
computation of the optimistic and pessimistic extensions of previously unrated
countries. The levels assigned to these countries by the recalculated optimistic
and pessimistic extensions are shown in Table 16.14.
It appears that:
• Guatemala’s first S&P’s rating (in 2001) was BB. Guatemala’s OE/PE levels are
the same as Morocco’s (the only country with OE ¼ PE ¼ 5), and Morocco’s
S&P’s rating in 1999 was BB.

Table 16.13 Correlation S&P’s OE PE LRS

analysis
S&P’s 100.00 % 95.09 % 95.15 % 94.12 %
OE 95.09 % 100.00 % 99.59 % 98.43 %
PE 95.15 % 99.59 % 100.00 % 98.24 %
LRS 94.12 % 98.43 % 98.24 % 100.00 %
16 Combinatorial Methods for Constructing Credit Risk Ratings 473

Table 16.14 Out-of- S&P’s

universe validation Optimistic Pessimistic First S&P’s linear
extension extension rating extension
Ecuador 3 2 SD (07/2000) 0
Guatemala 5 5 BB (10/2001) 10
Jamaica 3 2 B (11/1999) 7
Papua 3 3 B+ (01/1999) 8
New
Guinea

• Jamaica’s first S&P’s rating (1999) was B. Its OE/PE levels (OE ¼ 3, PE ¼ 2)
are identical to these of Paraguay, Brazil, the Dominican Republic, and Bolivia,
which had 1999 S&P’s ratings of B, B+, B+, and BB-, respectively.
• The first S&P’s rating for Papua New Guinea was B+. Its OE/PE levels (OE ¼ 3,
PE ¼ 3) are the same as those of Peru and Mexico, which both had 1999 S&P’s
ratings of BB.
• Ecuador’s OE/PE levels (OE ¼ 3, PE ¼ 2) are the same as those of Paraguay, Brazil,
the Dominican Republic, and Bolivia, which had 1999 S&P’s ratings of B, B+, B+,
and BB-, respectively. Interestingly, while the initial S&P’s rating of Ecuador was
SD (in July 2000), it was upgraded in August 2000 (1 month later) to B.
The striking similarity between the initial S&P’s rating of each of the four
countries discussed above and the S&P’s ratings of those countries which have
the same OE/PE levels validates the proposed model, indicating its power to predict
the creditworthiness of previously unrated countries.

LRS Approach
The LAD discriminant, which does not involve in any way the previous years’
S&P’s ratings, allows the rating of previously unrated countries in the following
way. First, we construct all the pseudo-observations involving the new countries to
be evaluated. Second, we calculate the relative preferences for these pseudo-
observations, and we add the resulting columns and rows to the matrix of relative
preferences. Third, we determine the new LRS for all the countries (new and old) by
running the multiple linear regression model (16.12). Fourth, we apply the linear
transformation defined above to the LRS so that the transformed LRS and the
S&P’s ratings are on the same scale.
The evaluation of the ability of LRS to accurately predict S&P’s ratings is
carried out by comparing the predicted LRS (obtained as described above) and
the S&P’s ratings (when they first become available). We compute the confidence
intervals (16.14) for the transformed LRS of four countries never rated by S&P’s by
December 1998. The predictions for Guatemala, Jamaica, and Papua New Guinea
correspond perfectly to the first time (subsequent) S&P’s ratings. The comparison
between the LRS and the first (July 2000) S&P’s rating (SD) to Ecuador shows that
S&P’s rated it too harshly, since 1 month later S&P’s raised its rating to B-,
justifying the LRS prediction.
474 A. Kogan and M.A. Lejeune

16.4.5 Importance of Variables

The methodology developed in this chapter permits the assessment of the

importance of the variables in rating countries’ creditworthiness. In LAD,
the importance of variables is associated with their use in the patterns of the
LAD model and is usually measured by the proportion of patterns containing
a particular variable. The patterns of the 1998 LAD model show that the three
most frequently used variables are financial depth and efficiency, political stabil-
ity, and gross domestic product per capita (appearing in 47.5 %, 39.4 %, and
35.6 % of the LAD patterns, respectively).
The presence of political stability among the three most significant ones in the
selected set justifies the inclusion of political variables in country risk rating
models. This result is in agreement with the cost-benefit approach to country risk
(i.e., the risk of defaulting is heavily impacted by the political environment, see
Brewer and Rivoli 1990, 1997; Citron and Neckelburg 1987; Afonso 2003) which is
not a view shared by all (Haque et al. 1996, 1998).
The fact that the LAD approach identifies gross domestic product per capita as
significant was expected, for most studies on country risk ratings acknowledge its
crucial importance in evaluating the creditworthiness of a country. A key new result
is the identification of the financial depth and efficiency variable as a major factor
for the prediction of country risk ratings.

16.5 Conclusions

The central objective of this study is to develop transparent, consistent, self-

contained, and stable credit risk rating models, closely approximating the risk
ratings provided by some of the main rating agencies. We use the combinatorial
optimization method called LAD and develop a relative creditworthiness approach
for assessing the credit risk of countries, while an absolute creditworthiness
approach is used for financial institutions.
The evaluation of the creditworthiness of financial organizations is particularly
important due to the growing number of banks going bankrupt and the magnitude of
losses caused by such bankruptcies, and it is challenging due to the opaqueness of
the banking sector and the high variability of banks’ creditworthiness. We use the
logical analysis of data (LAD) to reverse engineer the Fitch bank credit ratings. The
LAD method identifies strong combinatorial patterns distinguishing banks with
high and low ratings. These patterns constitute the core of the rating model
developed here for assessing the credit risk of banks. The results show that the
LAD ratings are in very close agreement with the Fitch ratings. In that respect, it is
important to note that the critical component of the LAD rating system – the LAD
discriminant – is derived utilizing only information about whether a bank’s rating is
“high” or “low,” without the exact specification of the bank’s rating category.
Moreover, the LAD approach uses only a fraction of the observations in the dataset.
The higher classification accuracy of LAD appears even more clearly when
16 Combinatorial Methods for Constructing Credit Risk Ratings 475

performing cross-validation and applying the LAD model derived by using the
banks in the training set to those in the testing one.
This study also shows that the LAD-based approach to reverse engineering bank
ratings provides a model that is parsimonious and robust. This approach allows to
derive rating models with varying levels of granularity that can be used at different
stages in the credit granting decision process and can be employed to develop
internal rating systems that are Basel 2 compliant. Besides impacting the credit risk
of a financial institution, the use of the generalizable and accurate credit risk rating
system proposed here will also be critical in mitigating the financial institution’s
operational risk due to breakdowns in established processes and risk-management
operations or to inadequate process mapping within business lines. In particular, the
reliance on such risk rating system will reduce the losses due to mistakes made in
executing transactions, such as settlement failures, failures to meet capital regula-
tory requirements, or untimely debt collections, or the losses due to the offering of
inappropriate financial products or credit conditions, or giving incorrect advice to
counterparty.
The evaluation of the creditworthiness of countries is also of utmost importance,
since the country’s risk rating is generally viewed as the upper bound on the rating
that entities within a given country can be assigned. This study proposes an LAD
methodology for inducing a credit risk system from a set of country risk rating
evaluations. It uses nine economic and three political variables to construct the
relative preferences of countries on the basis of their creditworthiness. Two
methods are then developed to construct countries’ credit rating systems on the
basis of their relative creditworthiness. The first one is based on extending the
preorder of countries using the Condorcet voting technique and provides two rating
systems (weak Condorcet winners and losers), while the second one uses linear
regression to determine the logical rating scores.
The proposed rating systems correlate highly with those of the utilized rating
system (S&P’s) and those of other rating agencies (Moody’s and The Institutional
Investor) and are shown to be stable, having an excellent classification accuracy
when applied to the following years’ data or to the ratings of previously unrated
countries. Rating changes implemented by the S&P’s in subsequent years have
resolved most of the (few) discrepancies between the constructed partially ordered
set and S&P’s initial ratings. This study provides new insights on the importance of
variables by supporting the necessity of including in the analysis, in addition to
economic variables, also political variables (“political stability”), and by identify-
ing “financial depth and efficiency” as a new critical factor in assessing
country risk.
The rating systems proposed here for banks as well as countries are as follows:
• Avoid overfitting as attested by the back-testing analysis (i.e., extremely high
concordance between in- and out-of-sample rating predictions calculated using
the k-folding and jackknife cross-validation methods).
• Distinguish themselves from the rating models in the existing literature by their
self-contained nature, that is, by their non-reliance on any information derived
from lagged ratings. Therefore, the high level of correlation between predicted
476 A. Kogan and M.A. Lejeune

Table 16.15 Standard & Poor’s country rating system

Level Description
Investment AAA An obligor rated AAA has extremely strong capacity to meet its financial
rating obligations. AA is the highest issuer credit rating assigned by S&P’s
AA An obligor rated AA has very strong capacity to meet its financial
commitments. It differs from the highest rated obligors only in small
degree
A An obligor rated A has strong capacity to meet its financial commitments
but is somewhat more susceptible to the adverse effects of changes in
circumstances and economic conditions than obligors in higher-rated
categories
BBB An obligor rated BBB has adequate capacity to meet its financial
commitments. However, adverse economic conditions or changing
circumstances are more likely to lead to a weakened capacity of the obligor
to meet its financial commitments
Speculative BB An obligor rated BB is less vulnerable in the near term than other lower-
rating rated obligors. However, it faces major ongoing uncertainties and exposure
to adverse financial or economic conditions which could lead to its inability
to meet financial commitments
B An obligor rated B is more vulnerable than the obligors rated BB, but, at
the time of the rating, it has the capacity to meet financial commitments.
Adverse business, financial conditions could likely impair its capacity to
meet financial commitments
Default rating CCC An obligor rated CCC is vulnerable at the time of the rating and is
dependent upon favorable business, financial, and economic conditions to
meet financial commitments
CC An obligor rated CC is highly vulnerable at the time of the rating
C An obligor rated C is vulnerable to nonpayment at the time of the rating and
is dependent upon favorable business, financial, and economic conditions
to meet financial commitments
D An obligor rated D is predicted to default
SD An obligor rated SD (selected default) is presumed to be unwilling to repay

and actual ratings cannot be attributed to the reliance on lagged ratings and is
a reflection of the predictive power of the independent variables included in
these models. An important advantage of the non-recursive nature of the pro-
posed models is their applicability to not-yet-rated obligors.
The scope of the proposed methodology extends beyond the rating problems
discussed in this study and can be used in many other contexts where ratings are
relevant. The proposed methodology is applicable in the general case of inferring an
objective rating system from archival data, given that the rated objects are charac-
terized by vectors of attributes taking numerical or ordinal values.

Appendix

See Tables 16.15 and 16.16.

 16

Table 16.16 1998 Ratings

The
Institutional Optimistic Pessimistic Optimistic Pessimistic
S&P’s S&P’s Moody’s Investor extension extension LRS S&P’s S&P’s extension extension LRS
ratings preorder ratings ratings ratings ratings scores ratings preorder ratings ratings scores
Countries (1998) (1998) (1998) (1998) (1998) (1998) (1998) (1999) (1999) (1999) (1999) (1999)
Argentina BB 10 9 42.7 7 6 0.2768 BB 10 7 6 0.263
Australia AA 19 19 74.3 16 16 0.0289 AA+ 20 16 16 0.0128
Austria AAA 21 21 88.7 17 17 0.0094 AAA 21 17 16 0.0038
Belgium AA+ 20 20 83.5 15 14 0.0476 AA+ 20 15 15 0.0439
Bolivia BB 9 8 28 3 3 0.366 BB 9 3 2 0.3518
Brazil BB 9 7 37.4 2 2 0.3744 B+ 8 3 2 0.4016
Canada AA+ 20 20 83 16 16 0.0241 AA+ 20 16 16 0.0112
Chile A 15 14 61.8 11 11 0.191 A 15 11 11 0.1841
China BBB+ 14 15 57.2 10 9 0.2159 BBB 13 10 10 0.224
Colombia BBB 12 12 44.5 2 2 0.3854 BB+ 11 2 2 0.3964
Costa Rica BB 10 11 38.4 7 6 0.2748 BB 10 7 6 0.257
Croatia BBB 12 12 39.03 5 5 0.297 BBB 12 6 6 0.3202
Cyprus A+ 17 16 57.3 12 12 0.1081 A 16 12 12 0.1021
Czech A 15 14 59.7 10 10 0.2088 A 15 10 10 0.1904
Combinatorial Methods for Constructing Credit Risk Ratings

Republic
Denmark AA+ 20 20 84.7 15 15 0.048 AA+ 20 15 15 0.0492
Dominican B+ 8 10 28.1 3 2 0.3568 B+ 8 3 2 0.3431
Rep
Egypt BBB 12 11 44.4 6 5 0.2915 BBB 12 6 6 0.3067
El Salvador BB 10 12 31.2 4 4 0.3379 BB+ 11 5 4 0.3301
Estonia BBB+ 14 14 42.8 9 8 0.2518 BBB+ 14 8 8 0.245
Finland AA 19 21 82.2 14 14 0.064 AA+ 20 14 13 0.0458
477

(continued)
 478

Table 16.16 (continued)

The
Institutional Optimistic Pessimistic Optimistic Pessimistic
S&P’s S&P’s Moody’s Investor extension extension LRS S&P’s S&P’s extension extension LRS
ratings preorder ratings ratings ratings ratings scores ratings preorder ratings ratings scores
Countries (1998) (1998) (1998) (1998) (1998) (1998) (1998) (1999) (1999) (1999) (1999) (1999)
France AAA 21 21 90.8 13 13 0.0828 AAA 21 13 13 0.0614
Germany AAA 21 21 92.5 18 18 0.001 AAA 21 18 17 0.0126
Greece BBB 13 14 56.1 9 9 0.2255 A 15 9 9 0.1917
Hong Kong A 16 15 61.8 17 13 0.017 A 16 16 15 0.0213
Hungary BBB 13 13 55.9 9 8 0.2442 BBB 13 9 8 0.247
Iceland A+ 17 18 67 15 15 0.047 A+ 17 15 14 0.0378
India BB 10 10 44.5 1 1 0.4063 BB 10 2 1 0.3994
Indonesia CCC+ 5 6 27.9 0 0 0.4576 CCC+ 5 0 0 0.4316
Ireland AA+ 20 21 81.8 17 16 0.0179 AA+ 20 17 16 0.0249
Israel A 15 15 54.3 11 9 0.2215 A 15 10 9 0.2189
Italy AA 19 18 79.1 12 12 0.1064 AA 19 12 12 0.1122
Japan AAA 21 20 86.5 15 14 0.0604 AAA 21 15 14 0.0506
Jordan BB 9 9 37.3 4 3 0.323 BB 9 4 3 0.2818
Kazakhstan B+ 8 9 27.9 1 1 0.4095 B+ 8 1 1 0.4048
Korea Rep. BB+ 11 11 52.7 9 6 0.2649 BBB 13 10 8 0.2182
Latvia BBB 13 13 38 5 5 0.3026 BBB 13 5 5 0.3039
Lebanon BB 9 8 31.9 4 4 0.3223 BB 9 4 4 0.3121
Lithuania BBB 12 11 36.1 4 4 0.3247 BBB 12 4 4 0.3233
Malaysia BBB 12 12 51 11 9 0.1676 BBB 13 11 11 0.1712
A. Kogan and M.A. Lejeune
 16

Malta A+ 17 15 61.7 13 12 0.0999 A 16 12 12 0.2402

Mexico BB 10 10 46 3 3 0.3608 BB 10 3 3 0.3284
Morocco BB 10 11 43.2 6 5 0.2952 BB 10 5 5 0.2881
Netherlands AAA 21 21 91.7 19 19 0.0251 AAA 21 19 18 0.0337
New AA+ 20 19 73.1 18 17 0.0001 AA+ 20 17 16 0.0001
Zealand
Norway AAA 21 21 86.8 19 18 0.0125 AAA 21 18 18 0.0076
Pakistan CC 2 5 20.4 0 0 0.4563 B 6 0 0 0.4501
Panama BB+ 11 11 39.9 7 6 0.2712 BB+ 11 6 6 0.2487
Paraguay BB 9 7 31.3 2 2 0.3865 B 7 3 2 0.4066
Peru BB 10 9 35 3 3 0.3536 BB 10 3 3 0.3644
Philippines BB+ 11 11 41.3 4 4 0.3242 BB+ 11 4 3 0.349
Poland BBB 12 12 56.7 7 6 0.2772 BBB 13 7 7 0.2743
Portugal AA 19 19 76.1 14 13 0.0742 AA 19 14 13 0.0706
Romania B 6 6 31.2 1 1 0.3987 B 6 1 1 0.3942
Russia CCC 3 6 20 0 0 0.4428 SD 0 0 0 0.4197
Singapore AAA 21 20 81.3 19 18 0.0073 AAA 21 18 18 0.0225
Slovak BB+ 11 11 41.3 7 6 0.2814 BB+ 11 7 6 0.269
Republic
Combinatorial Methods for Constructing Credit Risk Ratings

Slovenia A 16 15 58.4 11 9 0.1922 A 16 10 9 0.1878

South BB+ 11 12 45.8 8 6 0.2523 BB+ 11 7 6 0.2386
Africa
Spain AA 19 19 80.3 13 12 0.0924 AA+ 20 13 13 0.0798
Sweden AA+ 20 19 79.7 18 17 0.0106 AA+ 20 17 17 0.0143
Switzerland AAA 21 21 92.7 20 20 0.071 AAA 21 19 19 0.0613
(continued)
479
 480

Table 16.16 (continued)

The
Institutional Optimistic Pessimistic Optimistic Pessimistic
S&P’s S&P’s Moody’s Investor extension extension LRS S&P’s S&P’s extension extension LRS
ratings preorder ratings ratings ratings ratings scores ratings preorder ratings ratings scores
Countries (1998) (1998) (1998) (1998) (1998) (1998) (1998) (1999) (1999) (1999) (1999) (1999)
Thailand BBB 12 11 46.9 8 8 0.2452 BBB 12 8 8 0.2383
Trinidad & BB+ 11 11 43.3 6 6 0.2824 BBB 12 6 6 0.248
Tob
Tunisia BBB 12 12 50.3 9 8 0.2488 BBB 12 8 8 0.242
Turkey B 7 8 36.9 0 0 0.4458 B 7 0 0 0.4177
UK AAA 21 21 90.2 18 18 0.0057 AAA 21 18 18 0.0062
United AAA 21 21 92.2 19 19 0.0205 AAA 21 19 18 0.0264
States
Uruguay BBB 12 12 46.5 7 7 0.2695 BBB 12 7 7 0.2409
Venezuela B+ 8 7 34.4 0 0 0.4444 B 7 1 0 0.3921
We have converted the Standard & Poor’s rating scale (columns 1 and 4) into a numerical scale (columns 2 and 5). Such a conversion is not specific to
us. Bouchet et al. (2003), Ferri et al. (1999), and Sy (2004) proceed similarly. Moreover, Bloomberg, a major provider of financial data services, developed
a standard cardinal scale for comparing Moody’s, S&P’s, and Fitch-BCA ratings (Kaminsky and Schmukler 2002). A higher numerical value denotes a higher
probability of default. The numerical scale is referred to in this chapter as Standard & Poor’s preorder
A. Kogan and M.A. Lejeune
16 Combinatorial Methods for Constructing Credit Risk Ratings 481

References
Afonso, A. (2003). Understanding the determinants of sovereign debt ratings: Evidence for the two
leading agencies. Journal of Economics and Finance, 27(1), 56–74.
Afonso, A., Gomes, P., & Rother, P. (2007). What “hides” behind sovereign debt ratings?
(European Central Bank Working Paper Series 711). https://www.repository.utl.pt/bitstream/
10400.5/2134/1/ecbwp711.pdf
Alexe, S. (2002). Datascope – A new tool for logical analysis of data. DIMACS Mixer Series.
DIMACS, Rutgers University.
Alexe, G., Alexe, S., Bonates, T. O., & Kogan, A. (2007). Logical analysis of data – The vision of
Peter L. Hammer. Annals of Mathematics and Artificial Intelligence, 49, 265–312.
Altman, E. I., & Rijken, H. A. (2004). How rating agencies achieve rating stability. Journal of
Banking and Finance, 28, 2679–2714.
Barr, R. S., & Siems, T. F. (1994). Predicting bank failure using DEA to quantify management
quality. Federal Reserve Bank of Dallas Financial Industry Studies, 1, 1–31.
Basel Committee on Banking Supervision. (2001). The internal ratings based approach. http://
www.bis.org/publ/bcbsca05.pdf
Basel Committee on Banking Supervision. (2004). Bank failures in mature economies (Working
Paper 13).
Basel Committee on Banking Supervision. (2006). International convergence of capital measure-
ment and capital standards: A revised framework (Basel II).
Bhatia, A.V. (2002). Sovereign credit risk ratings: An evaluation (IMF Working Paper WP/03/
170).
Boros, E., Hammer, P. L., Ibaraki, T., & Kogan, A. (1997). Logical analysis of numerical data.
Mathematical Programming, 79, 163–190.
Boros, E., Hammer, P. L., Ibaraki, T., Kogan, A., Mayoraz, E., & Muchnik, I. (2000). An
implementation of logical analysis of data. IEEE Transactions on Knowledge and Data
Engineering, 12(2), 292–306.
Bouchet, M. H., Clark, E., & Groslambert, B. (2003). Country risk assessment: A guide to global
investment strategy. Chichester: Wiley.
Bourke, P., & Shanmugam, B. (1990). An introduction to bank lending (Addison-Wesley Business
Series). Sydney: Addison-Wesley.
Brewer, T. L., & Rivoli, P. (1990). Politics and perceived country creditworthiness in international
banking. Journal of Money, Credit and Banking, 22, 357–369.
Brewer, T. L., & Rivoli, P. (1997). Political instability and country risk. Global Finance Journal,
8(2), 309–321.
Cantor, R., & Packer, F. (1996). Determinants and impact of sovereign credit ratings. FRBNY
Economic Policy Review, 2, 37–53.
Choo E.U., & Wedley W.C., (2004). A Common Framework for Deriving Preference Values from
Pairwise Comparison Matrices. Computers and Operational Research 31, 893–908.
Citron, J. T., & Neckelburg, G. (1987). Country risk and political instability. Journal of Develop-
ment Economics, 25, 385–395.
Crama, Y., Hammer, P. L., & Ibaraki, T. (1988). Cause-effect relationships and partially defined
boolean functions. Annals of Operations Research, 16, 299–326.
Curry, T., & Shibut, L. (2000). The Cost of the Savings and Loan Crisis: Truth and Consequences.
FDIC Banking Review, 2(2), 26–35.
Czyzyk, J., Mesnier, M. P., & Moré, J. J. (1998). The NEOS server. IEEE Computer Science
Engineering, 5(3), 68–75.
de Servigny, A., & Renault, O. (2004). Measuring and managing credit risk. New York:
McGraw-Hill.
Eliasson, A. (2002). Sovereign credit ratings (Working Papers 02-1). Deutsche Bank.
Ferri, G., Liu, L.-G., & Stiglitz, J. (1999). The procyclical role of rating agencies: Evidence from
the east asian crisis. Economic Notes, 3, 335–355.
482 A. Kogan and M.A. Lejeune

Fitch Ratings. (2001). Fitch simplifies bank rating scales. Technical Report.
Fitch Ratings. (2006). The role of support and joint probability analysis in bank ratings. Fitch
Special Report.
Galindo, J., & Tamayo, P. (2000). Credit risk assessment using statistical and machine
learning/basic methodology and risk modeling applications. Computational Economics,
15, 107–143.
Gehrlein, W. V., & Lepelley, D. (1998). The condorcet efficiency of approval voting and the
probability of electing the condorcet loser. Journal of Mathematical Economics, 29, 271–283.
Hammer, P. L. (1986). Partially defined boolean functions and cause-effect relationships. Inter-
national conference on multi-attribute decision making via OR-based expert systems. Passau:
University of Passau.
Hammer, P. L., Kogan, A., & Lejeune, M. A. (2006). Modeling country risk ratings using partial
orders. European Journal of Operational Research, 175(2), 836–859.
Hammer, P. L., Kogan, A., & Lejeune, M. A. (2011). Reverse-engineering country risk ratings:
Combinatorial non-recursive model. Annals of Operations Research, 188, 185–213.
Hammer, P. L., Kogan, A., & Lejeune, M. A. (2012). A logical analysis of banks’ financial
strength ratings. Expert Systems with Applications, 39(9), 7808–7821.
Haque, N. U., Kumar, M. S., Mark, N., & Mathieson, D. (1996). The economic content of
indicators of developing country creditworthiness. International Monetary Fund Working
Paper, 43(4), 688–724.
Haque, N. U., Kumar, M. S., Mark, N., & Mathieson, D. (1998). The relative importance of
political and economic variables in creditworthiness ratings. International Monetary Fund
Working Paper, 46, 1–13.
Hu, Y.-T., Kiesel, R., & Perraudin, W. (2002). The estimation of transition matrices for sovereign
credit ratings. Journal of Banking and Finance, 26(7), 1383–1406.
Huang, Z., Chen, H., Hsu, C.-J., Chen, W.-H., & Wu, S. (2004). Credit rating analysis with support
vector machines and neural networks: A market comparative study. Decision Support Systems,
37, 543–558.
Jain, K., Duin, R., & Mayo, J. (2000). Statistical pattern recognition: A review. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 22, 4–37.
Kaminsky, G., & Schmukler, S. L. (2002). Emerging market instability: Do sovereign ratings
affect country risk and stock returns? World Bank Economic Review, 16, 171–195.
Kaufmann, D., Kraay A., & Zoido-Lobaton, P. (1999a). Aggregating governance indicators.
(World Bank Policy Research Department Working Paper, 2195).
Kaufmann, D., Kraay, A., & Zoido-Lobaton, P. (1999b). Governance matters (World Bank Policy
Research Department Working Paper, 2196).
Kunczik, M. (2001). Globalization: News media, images of nations and the flow of international
capital with special reference to the role of rating agencies. Paper presented at the IAMCR
Conference, Singapore, pp. 1–49.
Larrain, G., Reisen, H., & von Maltzan, J. (1997). Emerging market risk and sovereign credit
ratings (pp. 1–30). OECD Development Centre Working Papers 124, OECD Publishing.
Moody’s. (2006). Bank financial strength ratings: Revised methodology. Moody’s Global Credit
Research Report.
Mora, N. (2006). Sovereign credit ratings: Guilty beyond reasonable doubt? Journal of Banking
and Finance, 30(7), 2041–2062.
Morgan, D. P. (2002). Rating banks: Risk and uncertainty in an opaque industry. The American
Economic Review, 92(4), 874–888.
Ng, W.-Y., Choi, K.-W., & Shum, K.-H. (1996). Arbitrated matching: Formulation and protocol.
European Journal of Operational Research, 88(2), 348–357.
Poon, W. P. H., Firth, M., & Fung, H.-G. (1999). A multivariate analysis of the determinants of
Moody’s bank financial strength ratings. Journal of International Financial Markets, Institu-
tions & Money, 9(3), 267–283.
16 Combinatorial Methods for Constructing Credit Risk Ratings 483

Quenouille, M. (1949). Approximate tests of correlation in time series. Journal of the Royal
Statistical Society, Series B, 11, 18–84.
Reinhart, C. M. (2002). Default, currency crises, and sovereign credit ratings. World Bank
Economic Review, 16, 151–170.
Sarkar, S., & Sriram, R. S. (2001). Bayesian models for early warning of bank failures. Manage-
ment Science, 47(11), 1457–1475.
Sobehart, J. R., Keenan, S. C., & Stein, R. M. (2000). Benchmarking quantitative default risk
models: A validation methodology. New York: Moody’s Investors Service.
Stein, R. M. (2002). Benchmarking default prediction models: Pitfalls and remedies in model
validation. New York: Moody’s KMV.
Sy, A. N. R. (2004). Rating the rating agencies: Anticipating currency crises or debt crises?
Journal of Banking and Finance, 28(11), 2845–2867.
Tarjan, R. E. (1972). Depth-first search and linear graph algorithms. SIAM Journal on Computing,
1, 146–160.
Treacy, W. F., & Carey, M. S. (2000). Credit risk rating systems at large US banks. Journal of
Banking & Finance, 24(1–2), 167–201.
Dynamic Interactions Between
Institutional Investors and the Taiwan 17
Stock Returns: One-Regime and Threshold
VAR Models

Bwo-Nung Huang, Ken Hung, Chien-Hui Lee, and Chin W. Yang

Contents
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
17.2 Sample Data and Basic Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
17.3 Lead-Lag Relation Among the Three Groups of Institutional Investors
in the TSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
17.3.1 The Unrestricted VAR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
17.3.2 The Structural VAR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
17.3.3 The Threshold VAR Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
17.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
The Unrestricted VAR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
The Structural VAR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
The Threshold VAR Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518

B.-N. Huang
National Chung-Cheng University, Minxueng Township, Chiayi County, Taiwan
e-mail: ecdbnh@ccu.edu.tw
K. Hung (*)
Division of International Banking & Finance Studies, Texas A&M International University,
Laredo, TX, USA
e-mail: ken.hung@tamiu.edu
C.-H. Lee
National Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan
e-mail: chlee@cc.kuas.edu.tw
C.W. Yang
Clarion University of Pennsylvania, Clarion, PA, USA
e-mail: yang@mail.clarion.edu; ecdycw@ccu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 485
DOI 10.1007/978-1-4614-7750-1_17,
# Springer Science+Business Media New York 2015
486 B.-N. Huang et al.

Abstract
This paper constructs a six-variable VAR model (including NASDAQ returns, TSE
returns, NT/USD returns, net foreign purchases, net domestic investment compa-
nies (dic) purchases, and net registered trading firms (rtf) purchases) to examine:
(i) the interaction among three types of institutional investors, particularly to test
whether net foreign purchases lead net domestic purchases by dic and rtf (the
so-called demonstration effect); (ii) whether net institutional purchases lead market
returns or vice versa; and (iii) whether the corresponding lead-lag relationship is
positive or negative? The results of unrestricted VAR, structural VAR, and multi-
variate threshold autoregression models show that net foreign purchases lead net
purchases by domestic institutions and the relation between them is not always
unidirectional. In certain regimes, depending on whether previous day’s TSE
returns are negative or previous day’s NASDAQ returns are positive, we find
ample evidence of a feedback relation between net foreign purchases and net
domestic institutional purchases. The evidence also supports a strong positive-
feedback trading by institutional investors in the TSE. In addition, it is found that
net dic purchases negatively lead market returns in Period 4. The MVTAR results
indicate that net foreign purchases lead market returns when previous day’s
NASDAQ returns are positive and have a positive influence on returns.
Readers are well advised to refer to chapter appendix for detailed discussion
of the unrestricted VAR model, the structural VAR model, and the threshold
VAR analysis.

Keywords
Demonstration effect • Multivariate threshold autoregression model • Foreign
investment • Lead-lag relationship • Structural VAR • Block Granger causality •
Institutional investors • Domestic investment companies • Registered trading
firms • Qualified foreign institutional investors

17.1 Introduction

As financial markets are gradually liberalized in emerging economies, capital invest-

ments have been flowing into these countries at increasing rates. Needless to say, such
capital movements in terms of bringing in direct investment along with technological
know-how can be instrumental in raising a nation’s productivity. On the downside,
capital inflows directed at investing in the host country’s security markets can be
disruptive. From a macroeconomics perspective, foreign capital inflows are beneficial
in that they provide much-needed capital. From a microeconomics perspective, they
also lower the cost of capital and enhance competitiveness. Nevertheless, capital
inflows can also be disruptive if arbitrage is their main purpose. When capital flight
occurs, as was the case in the 1997 Asian financial debacle and the 1994 Mexican Peso
crisis, movements in foreign capital could be extremely disruptive and damaging.
Given the potentially negative impact of foreign investments, Taiwan’s Ministry of
Finance exercised caution by implementing foreign capital policies in a three-stage
17 Dynamic Interactions 487

Table 17.1 Inflows and outflows of foreign capital investment in the TSE Unit: millions of
US dollars
Securities
investment Foreign institutional
companies investors Natural person Foreign total
Period Inflow Outflow Inflow Outflow Inflow Outflow Net inflow
1991 263 53 448 0 0 0 658
1992 57 61 447 17 0 0 426
1993 653 93 1,859 97 0 0 2,322
1994 451 207 2,279 634 0 0 1,889
1995 664 457 3,509 1,506 0 0 2,210
1996 565 477 6,213 3,881 334 8 2,747
1997/1
4 12 644 4,442 2,529 261 78 1,465
Cumulative amount 2,664 1,992 19,198 8,664 595 85 11,716
Source of data: Securities and Futures Commission, Ministry of Finance, Taiwan

process. First, it directed a number of local trust companies to issue investment funds
abroad for the local stock market. Second, qualified foreign investment companies
were allowed to invest in Taiwan’s stock market. Third, foreign individual as well as
institutional investors were permitted to directly participate in trading in the Taiwan
Stock Exchange Corporation (hereafter TSE). The first stage took effect from
September 1983 until December 29, 1990, when the second stage replaced the first
stage with a maximum investment limit of $2.5 billion. In response to the strong
demand to invest in the TSE, the ceiling was raised to $5 billion and to $7.5 billion in
August 1993 and March 1994, respectively. Foreign investments later increased
substantially (see Table 17.1), and as a result the ceiling was lifted entirely in 1996
except for a maximum investment limit to each individual stock.
The relatively slow pace in allowing foreign investment in Taiwan was the result
of ongoing debates between the Central Bank of China and the Securities and
Futures Commission of the Ministry of Finance regarding the stability of foreign
investment. The focus of the discussion was on the following three questions. First,
are there differences in the trading behaviors of different type of institutions? One
objective of opening up the domestic market to foreign investments is to utilize its
advantages in information acquisition, information processing, and trade execution
to improve the overall performance of local institutional investors.1 Second, will

1
This is because foreign institutions may well have better research teams and buy stocks according
to the fundamentals such as the firm’s future profitability. In contrast, local institutions and
individual investors usually choose stocks based on insider information or what the newspapers
write about. However, if stocks bought by foreign investors have a better performance than that of
local institutions and individual investors, the latter tend to buy the stocks bought by successful
foreign investors the previous day. Hence this gives rise to the so-called “demonstration effect.”
Such a concept is similar to the “herding” which Lakonishok et al. (1992) referred to as correlated
trading across institutional investors. Nonetheless, their definition is close to the contemporaneous
correlation rather than the lead-lag relation under study, and as such this leads to the demonstration
effect which we documented.
488 B.-N. Huang et al.

the trades of the three groups of classified institutional investors affect stock returns
of the TSE? Third, will the lifting of restrictions on foreign capital contribute to the
volatility of both the foreign exchange and stock markets in Taiwan?
The purpose of this paper is to provide answers to two of the aforementioned
issues: (1) Are there differences in the trading behaviors of different types of
institutions? Typically, we examine interaction among the three types of institu-
tions: qualified foreign institution investors (qfii), domestic investment companies
(dic), and registered trading firms (rtf). (2) Does institutional trading “cause” stock
returns or do institution investments follow movements in stock prices? The
distinction between this paper and previous literature lies in following aspects:
(i) Most previous studies uses low-frequency – yearly, quarterly, or at best weekly –
data (Nofsinger and Sias 1999; Cai and Zheng 2002; Karolyi 2002). In this paper we
employ daily data to help explore these issues in detail and to provide new evidence
on the short-term dynamics among institutional investors and stock returns.2
(ii) Unlike previous studies that used a bivariate VAR model (Froot et al. 2001),
we use a six-variable VAR model, which includes three types of institutional trades,
stock returns in the TSE, NT/USD exchange rate changes, and NASDAQ index
returns to test related relationships. (iii) Improving on the conventional linear VAR
analysis in the previous studies, we employ the threshold concept and split data into
two regimes based on whether previous trading day’s market returns are positive or
negative.
A number of studies have examined the relationship between investment flows
and stock returns. Brennan and Cao (1997) develop a theoretical model of interna-
tional equity flows that relies on the informational difference between foreign and
domestic investors. The model predicts that if foreign and domestic investors are
differently informed, portfolio flows between two countries will be a linear function
of the contemporaneous return on all national market indices. Moreover, if domes-
tic investors have a cumulative information advantage over foreign investors about
domestic securities, the coefficient of the host market return is expected to be
positive.
Nofsinger and Sias (1999) use US annual data to investigate the relationship
between stock returns and institutional and individual investors. They identify
a strong positive correlation between changes in institutional ownership and stock
returns measured over the same period. Their results suggest that (i) institutional
investors practiced positive-feedback trade more than individual investors and
(ii) institutional herding impacted price more than that by individual investors.
Nofsinger and Sias show that institutional herding is positively correlated with lag
returns and appears to be related to stock return momentum.
Choe et al. (1999) use order and trade data from November 30, 1996 to the end of
1997 to examine the impact of foreign investors on stock returns. They found strong

2
Froot et al. (2001) also use daily data, but they examine the behavior of capital flows across
countries. In addition, our models and approaches used in the estimation differ drastically from
theirs.
17 Dynamic Interactions 489

evidence of positive-feedback trading and herding by foreign investors before

South Korea’s economic crisis. During the crisis period, herding lessened and
positive- feedback trading by foreign investors mostly disappeared.3
Grinblatt and Keloharju (2000) use a Finland data set to analyze it to the extent
that past returns determine the propensity to buy and sell. They find that foreign
investors tend to be momentum investors, buying past winning stocks and selling past
losing ones. Domestic investors, particularly households, tend to be contrarians.
Froot et al. (2001) by making use of daily international portfolio flows (44 coun-
tries from 1994 to 1998) along with a bivariate unrestricted VAR model find that
lagged returns are statistically significant in predicting future flows. The evidence of
the predictability of returns by flows is, however, ambiguous. In developed markets,
there is no statistical evidence of stock predictability. For emerging markets, the
evidence for predictability is strong, although less so for the Emerging European
region. However, they estimate a restricted VAR model that assumes current inflows
will affect current prices, and the causality does not run from contemporaneous returns
to the flows. They provide the evidence of a positive contemporaneous correlation
between current inflows and returns in emerging markets.
Hamao and Mei (2001) investigate the impact of foreign investment on Japan’s
financial markets. Using monthly data from July 1974 to June 1992, they find that
(1) trades of foreign investors tend to increase market volatility more than that by
domestic investors; (2) foreign investors have more sophisticated investment technology
than do domestic investors; and (3) foreign investors seem to make investment decisions
on the basis of not only short-term gains but also long-term fundamentals.
Cai and Zheng (2002) use institutional holding data from the third quarter of 1981
to the last quarter of 1996 in order to examine the lead-lag relationship between
portfolio excess returns and the institutional trading. Beyond it, they compute institu-
tional trading as the change of institutional holdings from last quarter to the current
quarter. The unrestricted VAR analysis indicates that stock returns Granger-cause
institutional trading on quarterly basis, not vice versa. This implies the institutions
“herd” on past price behavior instead of being dominant price-setters in the market.
Using weekly data of Japan, Karolyi (2002) find consistent positive-feedback
trading among foreign investors before, during and after the Asian financial
debacle. Japanese banks, financial institutions, investment trusts and companies
are, on the other hand, aggressive contrarian investors. There is no evidence that the
trading activity by foreigners destabilized the markets during the crisis.4
Griffin et al. (2003) study the daily and intraday relationship between stock
returns and trading of institutional as well as individual investors for NASDAQ
100 securities. The daily unrestricted VAR results indicate that the institutional

3
Lakonishok et al. (1992) refer to the positive-feedback trading or trend chasing as buying winners
and selling losers and the negative feedback trading or contrarian as buying losers and selling
winners. Cai and Zheng (2002) point out that feedback trading occurs when lagged returns act as
the common signal that the investors follow.
4
Karolyi (2002) reaches such a conclusion because there is little evidence of any impact of foreign
net purchases on future Nikkei returns or currency returns.
490 B.-N. Huang et al.

buy-sell imbalances are positively related to previous day’s returns and the institu-
tional buy-sell imbalances (previous day) are not associated with current return.
The results are consistent with the finding by Sias and Starks (1997) using US data.
Griffin et al. (2003) estimate a structural VAR with the contemporaneous returns in
the institutional imbalance equation and discover a strong contemporaneous rela-
tionship between daily returns and institutional buy-sell imbalances.
Kamesaka et al. (2003) use Japanese weekly investment flow data over 18 years
to investigate the investment patterns and performance of foreign investors, indi-
vidual investors, and five types of institutional investors. Not surprisingly, they find
individual investors perform poorly, while securities firms, banks, and foreign
investors perform admirably over the sample period.
Several related studies focus mainly on Taiwan’s stock market. Huang and Hsu
(1999) detect decreased volatility in the weighted TSE using Levene’s F-statistic
following market liberalization. Lee and Oh (1995), implementing a vector
autoregression (VAR) model, find a reduction in the explanatory power of macro-
economics variables. Wang and Shen (1999) indicate that foreign investments exert
a positive impact on the exchange rate with only a limited effect on the TSE. In
addition, by using the turnover rate as a proxy for non-fundamental factors and
earnings per share for fundamental factors within the framework of a panel data
model, Wang and Shen are able to identify that (i) the non-fundamental factors
impacted the returns of the TSE before market liberalization and (ii) both the
fundamental and non-fundamental factors exerted an impact following market
liberalization.
Lee et al. (1999) investigate interdependence and purchasing patterns among
institutional investors, large, and small individual investors. Their results, based on
15-min intraday transaction data (3 months for 30 companies), highlight the
important role played by large individual investors, whose trading affects not
only stock returns but also small individual investors. However, net buys (i.e., the
difference between total buy and total sell) by institutional investors have no effect
on the TSE returns, and vice versa.
The previous literature is predominantly focused on the relationship between
institutional trading and stock returns, rarely on the interaction among institutional
investors. For example, the majority of prior studies find evidence of positive-
feedback trading by institutions, with the exception of Froot et al. (2001), who
discover that in Latin America and emerging East Asian markets, the trading by
institutions positively predicts future returns. Karolyi (2002) also detects that
foreign investors in Japan are positive-feedback traders, while Japanese financial
institution and companies are contrarian investors.
Most of the studies to date on these issues have been on the USA and Japanese
markets despite that some of the literature gives scant attention to Taiwan’s stock
market. When investigating the related issues in large countries such as the USA
and Japan, the influence of the foreign sector on the domestic market could be
neglected; however, for a small country such as Taiwan, it should not be ignored.
This is because the electronics industry in Taiwan is closely connected to the US
companies listed on the NASDAQ. Ultimately, after examining the interaction
17 Dynamic Interactions 491

among institutional investors and the dynamic relationship between stock returns
and institutional investors, the conclusion may also be affected by whether returns
of domestic market are positive or negative.
To circumvent the above problems, this paper employs a six-variable VAR
model, which takes into account trades of three types of institutional investors
(qfii, dic, and rtf), foreign returns, domestic returns, and changes in the NT/USD
exchange rate to jointly test hypotheses under different market conditions. Using
daily data, we find that net foreign purchases lead net domestic purchases. How-
ever, such a relation is not unidirectional. Under certain conditions (either when
previous day’s TSE returns are negative or previous day’s NASDAQ returns are
positive), we identify a feedback relation between net foreign purchases and net
domestic purchases. It highlights the well-known argument in Taiwan regarding
foreign investors: The demonstration effect on domestic institutional investors is
not entirely correct. As for the lead-lag relation between market returns and
institutional trading, we find that in most cases market returns at least lead both
net foreign and dic purchases; however, market returns also lead net rtf purchases if
the relationship between contemporaneous returns and institutional trading is
considered. On the other hand, our results also indicate that net dic purchases
lead market returns and are negatively associated with market returns in the fourth
period. The MVTAR analysis shows that when previous day’s NASDAQ returns
are positive, net foreign purchases positively lead stock returns.
The remainder of this paper is organized as follows. Section 17.2 describes the
sample data and the basic statistics. Section 17.3 investigates the lead-lag relation
for three groups of institutional investors in order to explore the issue of whether
foreign investments give rise to demonstration effects in Taiwan’s stock market and
examines the relationship between institutional trading activity and stock returns of
the TSE. To further explore the interaction among three types of market partici-
pants, the sample is divided into two regimes based on either the sign of the market
returns or that of the NASDAQ index returns of previous trading day, respectively.
The last section provides a conclusion.

17.2 Sample Data and Basic Statistics

This paper employs daily data from December 13, 1995 to May 13, 2004 for a large
sample analysis.5 The variables considered include purchases (qfiibuy) and sales
(qfiisell) by qfii, purchases (dicbuy) and sales (dicsell) by dic, purchases (rtfbuy) and
sales (rtfsell) by rtf, TSE daily weighted stock index (pt), NASDAQ stock index
(naspt), and the NT/USD exchange rate (et). The data are from the Taiwan Eco-
nomic Journal (TEJ). The changes in the exchange rate and the logarithmic returns
on TSE and NASDAQ indices are defined as

5
The data began on December 13, 1995, since the inception of the TEJ.
492 B.-N. Huang et al.

Table 17.2 Unit root tests for the six time series
Test nasrt rt Det qfiibst dicbst rtfbst
PP 44.04* 42.76* 43.32* 28.47* 31.12* 29.90*
The sample period starts from December 13, 1995, to May 13, 2004, a total of 1989 observations.
qfiibuyt and qfiisellt ¼ purchases and sales by qualified foreign institutional investors, and
qfiibst ¼ qfiibuyt-qfiisellt. dicbuyt and dicsellt ¼ purchases and sales by domestic investment
companies, and dicbst ¼ dicbuyt-dicsellt. rtfbuyt and rtfsellt ¼ purchases and sales by registered
trading firms, and rtfbst ¼ rtfbuyt-rtfsellt. nasrt are the NASDAQ index returns. rt is the TSE index
return. Det is changes in the NT/USD exchange rate. qfiibst is net purchases by qfii. dicbst is net
purchases by dic, and rtfbst is net purchases by rtf. y denotes the level of the variable. Dy denotes
the first difference of the variable. * denotes statistical significance at 1 % level

Table 17.3 Summary statistics for the net institutional purchases, stock returns and NT/USD
currency returns
Mean Median Maximum Minimum Std. dev.
Det 0.0105 0.0000 3.4014 2.9609 0.3331
rt 0.0083 0.0438 8.5198 12.6043 1.7839
nasrt 0.0302 0.1300 13.2546 10.4078 2.0210
qfiibuyt 5,386.30 4,130 31,415 151 4,513.78
qfiisellt 4,675.17 3,601 44,000 74 4,027.41
qfiibst 711.11 372 19,408 23772 3,171.87
dicbuyt 3,191.66 2,924 14,980 192 1,643.79
dicsellt 3,306.13 3,096 11,854 141 1,556.99
dicbst 114.49 120 10,070 8,876 1,277.43
rtfbuyt 1,814.00 1,418 10,972 49 1,433.95
rtfsellt 1,837.15 1,504 18,024 32 1,384.79
rtfbst 23.15 36 6,379 11,177 934.55
For variable definitions, see Table 17.2

Det ¼ ðlog et  log et1 Þ  100%,

r t ¼ ðlog pt  log pt1 Þ  100%

nasr t ¼ ðlog naspt  log naspt1 Þ  100%

Net foreign purchases (qfiibst) are computed as the daily purchases (qfiibuyt) less
sales (qfiisellt) of Taiwan stocks by foreigners. Similarly, net dic purchases (dicbst)
are computed as the daily purchases (dicbuyt) less sales (dicsellt) of Taiwan stocks
by dic, and net rtf purchases (rtfbst) are computed as the daily purchases (rtfbuyt)
less sales (rtfsellt) of Taiwan stocks by rtf. The VAR analysis used here depends on
whether the time series are stationary; hence, a unit root test is to be performed in
advance to avoid spurious regression. The Phillips and Perron test is applied and the
results are illustrated in Table 17.2.
The Phillips and Perron test results indicate that all time series are statistically
significant at 1 % level. There is no further differencing needed before applying VAR.
Table 17.3 presents the summary statistics for the time series used in this paper.
17 Dynamic Interactions 493

The average percentage change in the exchange rate on the daily basis is
0.011 %; the average daily TSE return is 0.0083 %; and the average daily NASDAQ
return equals 0.0302 %. Overall, qfii are net purchasers on average and two other
domestic institutional investors are net sellers of equity over the sample period,
reflecting different trading strategies adopted by foreign and domestic institutional
investors. Such distinct trading activities among institutional investors can also be
seen in Fig. 17.1.
Figure 17.1 presents the cumulative net purchases and daily net purchases by
qfii, dic, and rtf and how they are associated with the TSE returns, NASDAQ
returns, and NT/USD exchange rate. Over the entire period, the cumulative net
purchases by qfii suggest an upward trend in general, while those of dic and rtf
tend to present a downward trend. Overall, the NASDAQ index is more volatile
than the TSE index (1995/12/13 ¼ 100), and it seems that there exists some
correlation between the two indices. During the Asia financial crisis in 1997, the
NT/USD exchange rate suffered a great upward swing (depreciation of New
Taiwan Dollar) followed by a slight downward slide in 1999 and then rose
again from 2002 onwards. Over the sample period, the volatility of net purchases
by foreigners seemed to have increased since 2002. As for the relationship
between net purchases by institutions and stock returns, no clear correlation
could be detected as shown in Fig. 17.1. To grasp a better understanding on
their linkages, the contemporaneous correlation of net purchases by the three
types of institutional investors, stock returns, and currency returns are displayed
in Table 17.4.
An inspection of Table 17.4 points out that returns on the NT/USD
exchange rate (currency returns) are negatively correlated with both the
TSE returns and net purchases by the three types of institutional investors,
especially by foreign investors. Such relations are very much in line with the
expectation. When stock prices rise following the influx of foreign capital,
the local currency is expected to appreciate to a degree and as such negative
correlations among them is expected. In addition, we find that returns on the
TSE and NASDAQ are positively correlated. It is noteworthy that there exists
a positive contemporaneous correlation between net purchases by the three
types of institutional investors and the TSE returns with the correlation
coefficients ranging from about 0.3 (qfiibst and rt) to 0.4419 (rtfbst and rt).
In short, this finding largely echoes the previous results (e.g., Froot
et al. 2001; Karolyi 2002).
The greatest correlation between institutional trading and NASDAQ
returns is that of rtfbst and nasrt at 0.0938 followed by dicbst and qfiibst,
respectively. Owing to the time difference, the TSE returns may be
influenced by NASDAQ index returns. If nasrt1 is used instead, a higher
correlations between nasrt1 and net purchases by institutions are found:
0.3441 for qfiibst, 0.2369 for dicbst, and 0.1428 for rtfbst, respectively. It
implies that previous day’s NASDAQ returns exert a greater impact on net
purchases by each institutional investor in the TSE than do the current
NASDAQ returns.
494 B.-N. Huang et al.

Index
1000 (95/12/13=100)
1) Cumulative Net qfii Purchase
NT Dollars 500
a 97/07/02 Nasdaq
98/06/30 400
Asia Flu
300
200
1600000 100
Taiwan Stock Index
1200000 0
800000
400000 Cumulative Net QFII Purchase
0
−400000
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20
1000
NT Dollars NT/USD
2) Net qfii Purchase
40000 36
30000 a 35
20000 34
10000 33
0 32
−10000 31
−20000 Net QFII Purchase 30
−30000 29
NT/USD
−40000 28
−50000 27
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20
Index
1000 1) Cumulative Net dic Purchase (95/12/13=100)
NT Dollars 500
b 97/07/02 Nas daq
98/06/30 400
Asia Flu
300
200
50000
100
0 Taiwan Stock Index
0
−50000
−100000
Cumulative Net DIC Purchase
−150000
−200000
−250000
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20

Fig. 17.1 (continued)

17 Dynamic Interactions 495

1000 NT Dollars 2) Net dic Purchase NT/USD

20000 36
16000 b 35
12000 34
8000 33
4000 32
0 31
−4000 Net DIC Purchase 30
−8000 29
NT/USD
−12000 28
−16000 27
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20

Index
1000
NT Dollars 1) Cumulative Net rtf Purchase (95/12/13=100)
500
c 97/07/02 Nasdaq
98/06/30 400
Asia Flu
300

200
20000
Taiwan Stock Index 100
0
0
−20000
Cumulative Net RTF Purchase
−40000

−60000

−80000
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20

1000 NT Dollars 2) Net rtf Purchase NT/USD

16000 36
12000 c 35
8000 34
4000 33
0 32
−4000 31
Net RTF Purchase
−8000 30
−12000 29
−16000 NT/USD 28
−20000 27
97/01/07 98/02/05 99/03/03 00/03/24 01/04/02 02/05/09 03/05/20

Fig. 17.1 Trends of cumulative net purchase, net purchases by the three types of institutions,
Taiwan stock prices, NASDAQ index, and the NT/USD exchange rate. a-1 Cumulative net qfii
purchase, a-2 Net qfii purchase, b-1 Cumulative net dic purchase, b-2 Net dic purchase, c-1
Cumulative net rtf purchase, c-2 Net rtf purchase. Notes: qfii qualified foreign institutional
investors, dic domestic investment companies, rtf registered trading firms
496 B.-N. Huang et al.

Table 17.4 Correlation matrix of net purchases by institutions, stock returns, and exchange rate
changes
Det rt nasrt qfiibst dicbst rtfbst
Det 1.0000
rt 0.1338 1.0000
nasrt 0.0124 0.1322 1.0000
qfiibst 0.1270 0.2976 0.0482 1.0000
dicbst 0.0991 0.3750 0.0586 0.2400 1.0000
rtfbst 0.0767 0.4419 0.0938 0.3559 0.3779 1.0000
See also Table 17.3

17.3 Lead-Lag Relation Among the Three Groups

of Institutional Investors in the TSE

17.3.1 The Unrestricted VAR Model

Note that prices of many Taiwanese electronics securities are affected by the
NASDAQ returns and hence foreign portfolio inflows may induce fluctuations of
exchange rate. To investigate interactions emanated from the three types of insti-
tutional investors and the relationship between institutional trading activity and
stock returns in Taiwan, we employ a six-variable VAR model using the NASDAQ
returns (nasrt), currency returns (Det), TSE returns (rt), net foreign purchases
(qfiibst), net dic purchases (dicbst), and net rtf purchases (rtfbst) as the underlying
variables. We attempt to answer the issues pertaining to (i) the interaction among
trading activities of the three types of institutions and (ii) the relationship between
stock returns and institutional trading. First, we propose a six-variable unrestricted
VAR model shown below:
2 3 2 3
nasr t f11 ðLÞ f12 ðLÞ f13 ðLÞ f14 ðLÞ f15 ðLÞ f16 ðLÞ
6 De 7 6 7
6 t 7 6 f21 ðLÞ f22 ðLÞ f23 ðLÞ f24 ðLÞ f25 ðLÞ f26 ðLÞ 7
6 7 6 7
6 rt 7 6 f31 ðLÞ f32 ðLÞ f33 ðLÞ f34 ðLÞ f35 ðLÞ f36 ðLÞ 7
6 7¼6 7
6 qfiibs 7 6 f ðLÞ f ðLÞ f ðLÞ f ðLÞ f ðLÞ f ðLÞ 7
6 t7 6 41 42 43 44 45 46 7
6 7 6 7
4 dicbst 5 4 f51 ðLÞ f52 ðLÞ f53 ðLÞ f54 ðLÞ f55 ðLÞ f56 ðLÞ 5
rtfbst f61 ðLÞ f62 ðLÞ f63 ðLÞ f64 ðLÞ f65 ðLÞ f66 ðLÞ
2 3 2 3
nasr t1 e1t
6 De 7 6e 7
6 t1 7 6 2t 7
6 7 6 7
6 r t1 7 6 e3t 7
66 7þ6 7 (1)
7 6 7
6 qfiibst1 7 6 e4t 7
6 7 6 7
4 dicbst1 5 4 e5t 5
rtfbst1 e6t
17 Dynamic Interactions 497

where fij(L) is the polynomial lag of the jth variable in the ith equation.
To investigate the lead-lag relation among three types of institutional investors,
we
2 need to test the hypothesis3 that each off-diagonal element in the sub-matrix
f44 ðLÞ f45 ðLÞ f46 ðLÞ
4 f54 ðLÞ f55 ðLÞ f56 ðLÞ 5 is zero.
f64 ðLÞ f65 ðLÞ f66 ðLÞ
On the other hand, to determine whether the TSE returns of the previous day lead
net purchases by the three types of institutional investors,0 we test the hypothesis that
each polynomial lag in the vector ½ f43 ðLÞf53 ðLÞf63 ðLÞ  is zero. Conversely, if we
want to determine whether previous day’s net purchases by institutional investors
lead current market returns, we test the hypothesis that each element in the vector
[f34(L)f35(L)f36(L)] is zero. Before applying the VAR model, an appropriate lag
structure needs to be specified. A 3-day lag is selected based on the Akaike
information criterion (AIC). Table 17.5 presents the lead-lag relation among the
six time2series using block exogeneity
3 tests.
f44 ðLÞ f45 ðLÞ f46 ðLÞ
The 4 f54 ðLÞ f55 ðLÞ f56 ðLÞ 5 block represents the potential interaction among
f64 ðLÞ f65 ðLÞ f66 ðLÞ
three types of institutional investors. The results indicate that net foreign purchases lead
net dic purchases and the dynamic relationship between these two variables can be
provided by the impulse response function (IRF) in Fig. 17.2a. Clearly, a one-unit
standard error shock to net foreign purchases leads to an increase in net dic purchases,
but this effect dissipates quickly by Period 2. Figure 17.2b, c indicate a feedback
relation between net purchases by qfii and rtf. A one-unit standard error shock to net
foreign purchases results in a positive response to net rtf purchases over the next two
periods, which become negative in Period 3, followed by a positive response again after
Period 4. Furthermore, a one-unit standard error shock to net rtf purchases also gives rise
to an increase in net foreign purchases, which decays slowly over ten-period horizon.
Figure 17.2d shows that net purchases by dic lead net rtf purchases. A one-unit
standard error shock to net dic purchases produces an increase in net rtf purchases in the
first three periods and then declines thereafter. Overall, these impulse responses suggest
that previous day’s net foreign purchases exert a noticeable impact on net rtf purchases,
while previous day’s net rtf purchases also has an impact on net foreign purchases. It
implies that not only do foreign capital flows affect the trading activity of domestic
institutional investors but also the relation is not unidirectional. To be specific, there is
a feedback relation between net rtf purchases and net foreign purchases.
As for the effect of the three types of institutional trading activity on stock
returns in the TSE, Table 17.5 reveals that net dic purchases on previous day lead
the TSE returns. We can also see in Fig. 17.2e that after Period 3, net dic purchases
exert a negative (and thus destabilizing) effect on market returns, while the other
two institutional investors do not have such an effect over the sample period.
Examining the relationship between market returns and trading activity of the
three types of institutional investors, we find that either the net foreign purchases
or net dic purchases on previous trading day are affected by the previous day’s TSE
 498

Table 17.5 Results of Granger causality tests using the unrestricted VAR models
All nasrt Det rt qfiibst dicbst rtfbst
X3 5.23 (0.16) 103.73* (0.00) 329.58* (0.00) 103.46* (0.00) 40.92* (0.00)
nasr ti
i¼1
3
X 2.77 (0.43) 2.66 (0.45) 2.01 (0.57) 5.98 (0.11) 4.66 (0.20)
Deti
i¼1
3
X 0.04 (1.00) 17.43* (0.00) 14.22* (0.00) 97.30* (0.00) 0.27 (0.97)
r ti
i¼1
3
X 12.67* (0.01) 5.81 (0.12) 4.38 (0.22) 9.55** (0.02) 33.95* (0.00)
qfiibsti
i¼1
3
X 9.38** (0.02) 0.85 (0.84) 15.97* (0.00) 5.83 (0.12) 12.70* (0.01)
dicbsti
i¼1
3
X 0.66 (0.88) 0.81 (0.85) 2.33 (0.51) 52.84* (0.00) 0.85 (0.84)
rtfbsti
i¼1

*, **, and *** denote statistical significance at 1 %, 5 %, and 10 % levels, respectively. Values in parentheses are p values. The optimal lag length of three is
selected based on the Akaike information criterion
B.-N. Huang et al.
17 Dynamic Interactions 499

a) Response of dicbs b) Response of rtfbs c) Response of qfiibs

to qfiibs to qfiibs to rtfbs
1200 800 2500
1000 700 2000
600
800 1500
500
600 400 1000
400 300
200 200 500
100 0
0 0
−200 −100 −500
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

d) Response of rtfbs e) Response of r f) Response of qfiibs

to dicbs to dicbs to r
800 2.0 2500
700
600 1.6 2000
500 1.2 1500
400
300 0.8 1000
200 0.4 500
100
0 0.0 0
−100 −0.4 −500
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

g) Response of dicbs h) Response of r i) Response of qfiibs

to r to nasr j)to nasr
1200 2.0 2500
1000 1.6 2000
800 1.2 1500
600
0.8 1000
400
0.4 500
200
0 0.0 0
−200 −0.4 −500
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
j) Response of dicbs k) Response of rtfbs l) Response of nasr
to nasr to nasr to qfiibs
1200 800 2.4
1000 700 2.0
600
800 500 1.6
600 400 1.2
400 300 0.8
200 200
0.4
100
0 0 0.0
−200 −100 −0.4
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
m) Response of nasr n) Response of de
to dicbs to r
2.4 .4
2.0
.3
1.6
1.2 .2
0.8 .1
0.4
.0
0.0
−0.4 −.1
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Fig. 17.2 Impulse responses to innovations in the unrestricted VAR models (up to ten periods).
Notes: Solid lines represent response path and dotted lines are bands for the 95 % confidence
interval around response coefficients. r, TSE returns; nasr, NASDAQ returns; de, exchange rate
changes; qfiibs, net qfii purchases; dicbs, net dic purchases; rtfbs, net rtf purchases
500 B.-N. Huang et al.

returns. Moreover, the IRFs in Fig. 17.2f, g also reveal a significant positive relation
between TSE returns and net foreign purchases up to four periods and a significant
positive relation between TSE returns and net dic purchases for two periods, which
becomes negative after Period 3. In other words, foreign investors in the TSE
engage in positive-feedback trading, while those of the dic tend to change their
strategy and adopt negative-feedback trading after Period 3.
On the other hand, Table 17.5 indicates that previous day’s NASDAQ returns
significantly affect both current TSE returns and net purchases by the three types of
institutional investors. The impulse response in Fig. 17.2h–k also confirms that
previous day’s NASDAQ returns are positively related to both current returns and
net purchases by these institutional investors, with the exception that a negative
relation between net dic purchases and previous day’s NASDAQ returns is found
after Period 4. Such results are much in sync with the expectation since the largest
sector that comprises the TSE-weighted stock index is the electronics industry to
which many listed companies on NASDAQ have a strong connection. In addition,
although previous day’s net qfii and dic purchases also lead the NASDAQ returns,
we find that no significant relation exists except for Period 4 with a significantly
negative relation between them.
The liberalization of Taiwan’s stock market has ushered in significant amount of
short-term inflows and outflows of foreign capital, which have induced fluctuations in
the exchange rate. As is seen from Table 17.5, the TSE returns lead currency returns
and it appears that the initial significant effect of stock returns on currency returns is
negative for the first three periods and then turns to be significantly positive thereafter
(Fig. 17.2n). Given that foreign investors are positive-feedback traders, the capital
inflows is expected to grow in order to increase their stakes in TSE securities when
stock prices rise. Consequently, NT/USD is expected to appreciate.

17.3.2 The Structural VAR Model

The unrestricted VAR model does not consider the effect of current returns on net
purchases by institutions. The prior study by Griffin et al. (2003) includes current
returns in the institutional imbalance equation and finds a strong contemporaneous
positive relation between institutional trading and stock returns.6 Therefore, to
further examine the relationship between institutional trading and stock returns,
we introduce the current TSE returns (rt) in the net purchases equations of the three
types of institutional investors and reestimate the VAR model before conducting
the corresponding block exogeneity tests. Table 17.6 presents the estimation results.
As indicated in the last row of Table 17.6, we find evidence of a strong
contemporaneous correlation between current returns and net institutional

6
The main purpose of this paper is to improve our understanding on the interaction among
institutional investors and the relationship between institutional trading and stock returns. The
following discussion will therefore focus on these two issues.
 17

Table 17.6 Results of Granger causality tests using the structural VAR models
SVAR nasrt Det rt qfiibst dicbst rtfbst
X3 5.23 (0.16) 103.73* (0.00) 265.94* (0.00) 54.67* (0.00) 6.75*** (0.08)
Dynamic Interactions

nasr ti
i¼1
3
X 2.77 (0.43) 2.66 (0.45) 1.85 (0.60) 7.06*** (0.07) 4.45 (0.22)
Deti
i¼1
3
X 1.26 (0.74) 20.63* (0.00) 98.84* (0.00) 264.71* (0.00) 75.05* (0.00)
r ti
i¼1
3
X 12.67* (0.01) 5.81 (0.12) 4.38 (0.22) 14.05* (0.00) 36.64* (0.00)
qfiibsti
i¼1
3
X 9.38** (0.02) 0.85 (0.84) 15.97* (0.00) 5.09 (0.17) 11.09* (0.01)
dicbsti
i¼1
3
X 0.66 (0.88) 0.81 (0.85) 2.33 (0.51) 50.56* (0.00) 0.83 (0.84)
rtfbsti
i¼1
rt 332.55* [10.91] 247.47* [18.10] 213.66* [21.43]
The structural VAR model includes current TSE returns in the institutional trading equations to take the contemporaneous correlations into consideration.
See also Table 17.5 for definitions
501
502 B.-N. Huang et al.

a) Response of qfiibs to r b) Response of dicbs to r c) Response of rtfbs to r

2500 1200 800
1000 700
2000 600
1500 800 500
600 400
1000 300
400
500 200 200
100
0 0 0
−500 −200 −100
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Fig. 17.3 Selected impulse responses to innovations up to ten periods in the structural VAR
models. Notes: See also Fig. 17.2

purchases, which confirms the finding by the previous research. Moreover, as

shown in Table 17.5, we find no evidence that past returns lead net rtf purchases
when the unrestricted VAR model is used. In contrast to it, when the contempora-
neous impact of stock returns on net institutional purchases is considered, we find
that past returns also lead net rtf purchases as well as net purchases by qfii and dic
when the structural VAR model is used (Table 17.6). In other words, net purchases
by the three types of institutions are affected by past stock returns as was evidenced
by previous studies. The corresponding impulse response relations are presented
in Fig. 17.3.
Comparing the impulse response relations in Figs. 17.2 and 17.3, it is clear that
when the impact of current returns on net institutional purchases is considered,
a one-unit standard error shock from rt does not produce a positive impulse
response in institutional trading until Period 2.7 The responses of foreign investors
are rather distinct from those of domestic institutional investors after Period 3. In
general, a sustained positive response from foreign investors is observed, while
a negative response is witnessed for dic and sometimes, an insignificant response
for rtf manifests itself after Period 3.

17.3.3 The Threshold VAR Analysis

We pool all the data together when estimating either the unrestricted or restricted
VAR model; however, the trading activity of institutional investors may depend on
whether stock prices rise or fall.8 A small economy like Taiwan also depends to
a large degree on the sign of NASDAQ index returns. Consequently, to investigate
institutional trading under distinct regimes based on market returns, we use the
multivariate threshold autoregression (MVTAR) model proposed by Tsay (1998)

7
Figure 17.2f, g show significantly positive responses of qfiibst and dicbst to rt in Period 1 if the
impact of current returns is not considered.
8
Recall that both the positive-feedback and negative-feedback trading are associated with the sign
of market returns on the previous trading day.
17 Dynamic Interactions 503

Table 17.7 The C(d) Threshold variable Statistic

statistic
nasrt1 195.08 (0.00)
rt1 136.22 (0.08)
Values in parentheses are p values. The delay (d) is assumed to
be one

to test the relevant hypotheses. Let yt ¼ [nasrt, Det, rt, qfiibst, dicbst, rtfbst]0 be
a 6  1 vector and the MVTAR model can be described as
!
X
p
yt ¼ f0;1 þ fi, 1 yti  ½1  I ðztd > cÞ
i¼1
! (2)
X
p
þ f0;2 þ fi, 2 yti  I ðztd > cÞ þ et
i¼1

where E(e) ¼ 0, E(ee0 ) ¼ S, and I() is an index function, which equals 1 if the
relation in the bracket holds. It equals zero otherwise. ztd is the threshold variable
with a delay (lag) d.
In order to explore whether institutional trading activity would change during
different domestic and foreign market return scenarios, the potential threshold
variables used are rt1 and nasrt1.9 Before estimating Eq. 2, we need to test for
possible potential nonlinearity (threshold effect) in this equation. Tsay (1998)
suggests using the arranged regression concept to construct the C(d) statistic to
test the hypothesis Ho:fi,1 ¼ fi,2, i ¼ 0, . . . p. If H0 can be rejected, it implies that
there exists the nonlinearity in data with ztd as the threshold variable. Tsay (1998)
proves that C(d) is asymptotically a chi-square random variable with k(pk + 1)
degrees of freedom, where p is the lag length of the VAR model and k is the number
of endogenous variables yt.10 Table 17.7 presents the estimation results of the C(d)
statistic.
As shown in Table 17.7, the null hypothesis H0 is rejected using either past
returns on the TSE or NASDAQ, suggesting that our data exhibit nonlinear
threshold effect. Theoretically, one needs to rearrange the regression based on the
size of the threshold variable ztd before applying a grid search method to find the
optimal threshold value c*. Nonetheless, our goal is to know whether the institu-
tional trading behavior depends on the sign of market returns, as such the threshold
is set to zero in a rather arbitrary way.11 Table 17.8 lists the results of block

9
Here, we assume that net purchases by institutions are only affected by market returns on the
previous trading day.
10
For more details see Tsay (1998).
11
A previous study by Sadorsky (1999) also splits data into two regimes based on the sign of the
variable to discuss whether variables used would change their behaviors under different regimes.
 504

Table 17.8 Results of Granger causality tests using the MVTAR models with threshold variable rt1
rt1 < 0 nasrt Det rt qfiibst dicbst rtfbst
X 3 3.07 (0.38) 84.21* (0.00) 167.80* (0.00) 53.71* (0.00) 30.12* (0.00)
nasr ti
i¼1
3
X 0.51 (0.92) 2.01 (0.57) 1.00 (0.80) 3.11 (0.37) 0.65 (0.89)
Deti
i¼1
3
X 2.28 (0.52) 12.41* (0.01) 8.57** (0.04) 22.40* (0.00) 3.14 (0.37)
r ti
i¼1
3
X 7.75** (0.05) 3.04 (0.39) 4.33 (0.23) 9.79** (0.02) 32.73* (0.00)
qfiibsti
i¼1
3
X 3.83 (0.28) 0.46 (0.93) 4.92 (0.18) 2.38 (0.50) 4.62 (0.20)
dicbsti
i¼1
3
X 1.23 (0.75) 1.62 (0.65) 1.62 (0.66) 19.99* (0.00) 2.83 (0.42)
rtfbsti
i¼1
3
X 3.72 (0.29) 27.57* (0.00) 168.61* (0.00) 49.94* (0.00) 12.25* (0.01)
nasr ti
i¼1
B.-N. Huang et al.
 17

3
X 5.97 (0.11) 1.78 (0.62) 2.73 (0.44) 5.16 (0.16) 7.01*** (0.07)
Deti
i¼1
3
X 0.88 (0.83) 13.00* (0.00) 3.02 (0.39) 43.44* (0.00) 1.68 (0.64)
r ti
i¼1
3
X 4.83 (0.18) 5.87 (0.12) 1.85 (0.60) 5.72 (0.13) 5.16 (0.16)
qfiibsti
i¼1
Dynamic Interactions

3
X 8.99** (0.03) 2.04 (0.56) 13.17* (0.00) 3.90 (0.27) 12.54* (0.01)
dicbsti
i¼1
3
X 0.22 (0.97) 0.58 (0.90) 2.53 (0.47) 39.46* (0.00) 2.15 (0.54)
rtfbsti
i¼1

The results are premised on the condition that when previous day’s TSE returns are positive. See also Table 17.5
505
506 B.-N. Huang et al.

exogeneity tests for the lead-lag relation in the rt1 < 0 and rt1  0 regimes,
respectively.
The interaction among institutional investors is depicted in Table 17.8: Current
net purchases by foreign investors affect that by domestic institutions when previ-
ous day’s TSE returns are negative. Note that no such relation is evidenced when
previous day’s TSE returns are positive. A feedback relation between rtf and qfii is
observed when rt1 is positive or negative. However, dic is found to lead rtf only
when rt1 is positive. Such results reveal different institutional trading strategies
under distinct return regimes. The demonstration effect – previous day’s net foreign
purchases have on domestic institutions using the unrestricted VAR model – seems
to surface only when previous day’s market returns are negative. Therefore, it may
produce misleading results if we fail to consider the sign of previous returns.
The impulse responses in Fig. 17.4 illustrate that the responses of dic and rtf
from the qfii shock are quite similar to the ones in Fig. 17.2f, g.12 As for the impact
of previous day’s market returns on current net purchases by institutions, it can be
shown via the MVTAR model that market returns lead net purchases by qfii and dic
when previous day’s market returns are negative, which is consistent with the
finding using the one-regime VAR model. When previous day’s market returns
are positive, market returns lead net purchases by the dic only. Obviously, returns
have more influence on net institutional purchases when previous day’s
returns were negative. In addition, we find that net dic purchases on the previous
day may affect current returns when the one-regime VAR model is used. Actually,
the MVTAR analysis reveals that such a relation emerges only when previous day’s
market returns are positive. The impulse responses depicted in Fig. 17.4 (Panel B)
demonstrate that a one-unit standard error shock to net dic purchases produces an
increase in market returns in Period 2, and then they turns to be negative after
Period 4, a result similar to those using the one-regime VAR model.
Among the listed companies on the TSE, the electronics sector has the largest market
share, which accounts for more than 60 % of all trades. This being the case, Taiwan’s
stock market is closely related to the NASDAQ index as is evidenced using the
conventional VAR model. To further investigate whether the interaction among insti-
tutions and the relationship between institutional trading and stock returns are affected
by the sign of previous day’s NASDAQ index return, nasrt1 is used as the threshold
variable. That is, block Granger causality tests are performed by splitting our data into
two regimes based on the sign of the variable nasrt1. Table 17.9 reports the results.
The results indicate that net qfii purchases lead to that of two domestic institu-
tional investors regardless of the sign of previous day’s NASDAQ returns. More-
over, when nasrt1 < 0, net rtf purchases lead net qfii purchases, which is in line
with that using the one-regime VAR model. However, when nasrt1  0, the net
purchases by either dic or rtf lead net qfii purchases, and net qfii purchases lead net
purchases by either dic or rtf. In other words, we find strong evidence of a feedback

12
To economize space, only relevant impulse responses are presented here; the remaining are
available upon request.
17 Dynamic Interactions 507

rt-1< 0
a Response to Cholesky One S.D. Innovations ± 2 S.E.
a) Response of dicbs to qfiibs b) Response of rtfbs to qfiibs c) Response of qfiibs to rtfbs
1200 800 2400
600 2000
800
1600
400 1200
400
200 800
0 400
0
0
−400 −200 −400
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
d) Response of qfiibs to r e) Response of dicbs to r
2400 1200
2000
1600 800
1200
400
800
400 0
0
−400 −400
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
rt-1 ≥ 0
b Response to Cholesky One S.D. Innovations ± 2 S.E.
a) Response of qfiibs to rtfbs b) Response of rtfbs to dicbs
3000 800
2500
600
2000
1500 400
1000 200
500
0 0
−500 −200
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
c) Response of dicbs to r d) Response of r to dicbs
1200 2.0
1.6
800
1.2
400 0.8
0.4
0
0.0
−400 −0.4
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Fig. 17.4 Selected impulse responses to innovations up to ten periods in the MVTAR models.
(a) rt1 < 0, (b) rt1  0. Notes: See also Fig. 17.2

relation between net foreign purchases and two domestic net purchases when
previous day’s NASDAQ returns are positive. The results pertaining to the impact
of previous day’s returns on institutional trading parallel those using the
unrestricted VAR model: Previous day’s returns have an impact on the net pur-
chases by qfii and dic, but not on net purchases by rtf regardless of the sign of
previous day’s NASDAQ returns. As for the impact of net institutional purchases
 508

Table 17.9 Results of Granger causality tests using the MVTAR models with threshold variable nasrt1
nasrt  1  0 nasrt Det rt qfiibst dicbst rtfbst
X3 0.92 (0.82) 36.02* (0.00) 50.15* (0.00) 15.42* (0.00) 12.53* (0.01)
nasr ti
i¼1
3
X 0.50 (0.92) 4.95 (0.18) 0.29 (0.96) 3.81 (0.28) 3.08 (0.38)
Deti
i¼1
3
X 2.13 (0.55) 8.09** (0.04) 8.54** (0.04) 48.61* (0.00) 1.80 (0.62)
r ti
i¼1
3
X 4.62 (0.20) 9.55** (0.02) 3.25 (0.35) 7.43*** (0.06) 14.74* (0.00)
qfiibsti
i¼1
3
X 6.94*** (0.07) 1.05 (0.79) 7.97** (0.05) 2.77 (0.43) 9.64* (0.02)
dicbsti
i¼1
3
X 0.12 (0.99) 1.48 (0.69) 1.33 (0.72) 28.94* (0.00) 1.93 (0.59)
rtfbsti
i¼1
3
X 2.64 (0.45) 29.58* (0.00) 115.60* (0.00) 59.06* (0.00) 16.76* (0.00)
nasr ti
i¼1
B.-N. Huang et al.
 17

3
X 6.72*** (0.08) 5.21 (0.16) 3.85 (0.28) 5.63 (0.13) 5.73 (0.13)
Det-i
i¼1
3
X 1.49 (0.68) 27.44* (0.00) 14.75* (0.00) 55.66* (0.00) 0.27 (0.96)
r ti
i¼1
3
X 11.39* (0.01) 0.60 (0.90) 7.76** (0.05) 7.00*** (0.07) 33.48* (0.00)
qfiibsti
i¼1
Dynamic Interactions

3
X 3.48 (0.32) 3.25 (0.35) 11.08* (0.01) 6.74*** (0.08) 3.47 (0.32)
dicbsti
i¼1
3
X 1.82 (0.61) 2.59 (0.46) 4.12 (0.25) 25.09* (0.00) 4.65 (0.20)
rtfbsti
i¼1

The results are premised on the condition that when previous day’s NASDAQ returns are positive. See also Table 17.5
509
510 B.-N. Huang et al.

a nasrt−1 < 0
Response to Cholesky One S.D.Innovations ± 2 S.E.
a) Response of qfiibs b) Response of dicbs c) Response of rtfbs
to rtfbs to qfiibs to qfiibs
3000 1200 800
2500 600
2000 800
1500 400
400
1000 200
500 0 0
0
−500 −400 −200
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
d) Response of rtfbs e) Response of qfiibs f) Response of dicbs
to dicbs to r to r
800 3000 1200
600 2500
2000 800
400 1500
400
200 1000
500 0
0
0
−200 −500 −400
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
g) Response of r
to dicbs
2.0
1.6
1.2
0.8
0.4
0.0
−0.4
1 2 3 4 5 6 7 8 9 10

b nasrt-1 ≥ 0
Response to Cholesky One S.D. Innovations ± 2 S.E.
a) Response of qfiibs b) Response of qfiibs c) Response of dicbs
to dicbs to rtfbs to qfiibs
2500 2500 1200
2000 2000
800
1500 1500
1000 1000 400
500 500
0
0 0
−500 −500 −400
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Fig. 17.5 (continued)

17 Dynamic Interactions 511

d) Response of rtfbs to qfiibs e) Response of qfiibs to r f) Response of dicbs to r

800 2500 1200
600 2000
800
1500
400
1000 400
200
500
0 0
0
−200 −500 −400
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

g) Response of r to qfiibs h) Response of r to dicbs

2.0 2.0
1.6 1.6
1.2 1.2
0.8 0.8
0.4 0.4
0.0 0.0
−0.4 −0.4
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

Fig. 17.5 Impulse responses to innovations up to ten periods in the MVTAR models (portion).
(a) nasrt1 < 0, (b) nasrt1  0. Notes: See also Fig. 17.2

on previous day’s stock returns, only net dic purchases still lead stock returns when
nasrt1 < 0, as was the case in the one-regime model. However, we find that net qfii
purchases also lead stock returns when nasrt1  0.
The results from Panel B of Fig. 17.5g indicate that a one-unit standard error shock
from net qfii purchases made on previous days produces a positive response to stock
returns in Period 2, but no significantly negative responses are found during other
periods. The results of the MVTAR model also capture the phenomenon in the
one-regime model: Net dic purchases exert a negative impact on market returns.
However, such an effect is witnessed when previous day’s NASDAQ returns are
negative.

17.4 Conclusion

In this paper we investigate whether the trading behavior of foreign investors leads
that of Taiwanese institutional investors (i.e., the demonstration effect) and whether
institutional trading has a destabilizing effect on the stock market. The reason we
select Taiwan in our study is due to her unique role of being gradually opened up to
foreign investment and her high stock returns volatility. To provide more informa-
tion to these issues, this paper has constructed a six-variable VAR model including
trading activities of three types of institutional investors, the TSE returns,
NASDAQ returns, and currency returns so as to examine the interaction and the
dynamic relationship between institutional trading and stock returns using daily
data from December 13, 1995 to May 13, 2004.
512 B.-N. Huang et al.

The results from the conventional unrestricted VAR model indicate that net
purchases by foreign investors lead those by domestic institutions (and thus the
demonstration effect), while net purchases by rtf also lead those by foreign inves-
tors. That is, there exists a feedback relation between them. As for the relationship
between institutional trading and stock returns, we find that except for rtf, both
foreign and dic net purchases are positively affected by previous day’s TSE returns.
That is, both the qfii and dic engage in positive-feedback trading. We also find that
previous day’s net dic purchases first produce a positive and then a negative impact
on stock returns. Furthermore, we employ a structural VAR model with the
contemporaneous returns included in the three net institutional purchase equations.
A comparison of the structural and unrestricted VAR models suggests that the TSE
returns positively lead net rtf purchases using the structural VAR model, which
cannot be observed when the unrestricted VAR model is used. In other words, if the
contemporaneous relation between returns and net institutional purchases is taken
into account, we find that rtf are also positive-feedback traders.
On the other hand, the sign of market returns does affect trading activities of the
institutions. As a result, this paper makes use of the MVTAR model introduced by
Tsay (1998). By splitting data into two regimes based on the sign of both TSE and
NASDAQ returns on the previous trading day, we find that the demonstration effect
that foreign investors have on domestic institutions arises only when previous day’s
TSE returns are negative. In addition, when previous day’s TSE returns are nega-
tive, stock returns lead both net purchases by qfii and dic. However, stock returns
lead only net dic purchases when previous day’s TSE returns are positive. Finally,
we find the relation that net dic purchases lead market returns using the unrestricted
VAR model tends to emerge only when previous day’s TSE returns are positive.
As for the effect of NASDAQ returns on institutional trading, the results from
this paper suggest that when previous day’s NASDAQ returns are positive,
a feedback relation between net foreign purchases and net domestic purchases
prevails. Moreover, it is found that the net dic purchases lead the TSE returns, as
do the net foreign purchases. The latter, however, exert a positive influence on the
TSE returns.
In summary, our results suggest that net foreign purchases do lead net domestic
purchases, but more details manifest when the threshold model is applied. When
previous day’s TSE returns are negative (or previous day’s NASDAQ returns are
positive), a feedback relation between net foreign purchases and net domestic
purchases is observed. It implies that the widespread argument that foreign inves-
tors have a demonstration effect on domestic institutions in Taiwan is not entirely
correct. In examining the relation between market returns and institutional trading,
we find that market returns at least lead net purchases by both qfii and dic in most
cases. Market returns also lead net rtf purchases if the relationship of contempora-
neous returns and institutional trading is considered. Our analysis also indicates that
net dic purchases lead market returns and are negatively associated with market
returns in Period 4. The results of the MVTAR model suggest that when previous
day’s NASDAQ returns are positive, net foreign purchases positively lead stock
returns and thus will not exert a destabilizing influence on the market.
17 Dynamic Interactions 513

Appendix 1

The Unrestricted VAR Model

The prices of many Taiwanese electronics securities are affected by the NASDAQ
returns and hence foreign portfolio inflows may induce fluctuations of exchange
rate. To investigate interactions emanated from the three types of institutional
investors and the relationship between institutional trading activity and stock
returns in Taiwan, we employ a six-variable VAR model using the NASDAQ
returns (nasrt), currency returns (Det), TSE returns (rt), net foreign purchases
(qfiibst), net dic purchases (dicbst), and net rtf purchases (rtfbst) as the underlying
variables. We attempt to answer the issues pertaining to (i) the interaction among
trading activities of the three types of institutions and (ii) the relationship between
stock returns and institutional trading. First, we propose a six-variable unrestricted
VAR model shown below:

2 3 2 3
nasr t f11 ðLÞ f12 ðLÞ f13 ðLÞ f14 ðLÞ f15 ðLÞ f16 ðLÞ
6 De 7 6 f22 ðLÞ f23 ðLÞ f24 ðLÞ f25 ðLÞ f26 ðLÞ 7
6 t 7 6 f21 ðLÞ 7
6 7 6 7
6 rt 7 6 f31 ðLÞ f32 ðLÞ f33 ðLÞ f34 ðLÞ f35 ðLÞ f36 ðLÞ 7
6 7 6 7
6 qfiibs 7 ¼ 6 f ðLÞ f42 ðLÞ f43 ðLÞ f44 ðLÞ f45 ðLÞ f46 ðLÞ 7
6 t7 6 41 7
6 7 6 7
4 dicbst 5 4 f51 ðLÞ f52 ðLÞ f53 ðLÞ f54 ðLÞ f55 ðLÞ f56 ðLÞ 5
rtfbst f61 ðLÞ f62 ðLÞ f63 ðLÞ f64 ðLÞ f65 ðLÞ f66 ðLÞ
2 3 2 3
nasr t1 e1t
6 De 7 6e 7
6 t1 7 6 2t 7
6 7 6 7
6 r t1 7 6 e3t 7
66 7þ6 7
7 6 7
6 qfiibst1 7 6 e4t 7
6 7 6 7
4 dicbst1 5 4 e5t 5
rtfbst1 e6t

Where fij(L) is the polynomial lag of the jth variable in the ith equation. To
investigate the lead-lag relation among three types of institutional investors, we
need
2 to test the hypothesis
3 that each off-diagonal element in the sub-matrix
f44 ðLÞ f45 ðLÞ f46 ðLÞ
4 f54 ðLÞ f55 ðLÞ f56 ðLÞ 5 is zero.
f64 ðLÞ f65 ðLÞ f66 ðLÞ
On the other hand, to determine whether the TSE returns of the previous day lead
net purchases by the three types of institutional investors,0 we test the hypothesis that
each polynomial lag in the vector ½ f43 ðLÞf53 ðLÞf63 ðLÞ  is zero. Conversely, if we
want to determine whether previous day’s net purchases by institutional investors
lead current market returns, we test the hypothesis that each element in the vector
[f34(L)f35(L)f36(L)] is zero. Before applying the VAR model, an appropriate lag
structure needs to be specified. A 3-day lag is selected based on the Akaike
514 B.-N. Huang et al.

information criterion (AIC). Table 17.5 presents the lead-lag relation among the six
time series
2 using block exogeneity 3 tests.
f44 ðLÞ f45 ðLÞ f46 ðLÞ
The 4 f54 ðLÞ f55 ðLÞ f56 ðLÞ 5 block represents the potential interaction among
f64 ðLÞ f65 ðLÞ f66 ðLÞ
three types of institutional investors. The results indicate that net foreign purchases
lead net dic purchases and the dynamic relationship between these two variables
can be provided by the impulse response function (IRF) in Fig. 17.2a. Clearly,
a one-unit standard error shock to net foreign purchases leads to an increase in net
dic purchases, but this effect dissipates quickly by Period 2. Figure 17.2b, c indicate
a feedback relation between net purchases by qfii and rtf. A one-unit standard error
shock to net foreign purchases results in a positive response to net rtf purchases over
the next two periods, which become negative in Period 3, followed by a positive
response again after Period 4. Furthermore, a one-unit standard error shock to net rtf
purchases also gives rise to an increase in net foreign purchases, which decays
slowly over ten-period horizon.
Figure 17.2d shows that net purchases by dic lead net rtf purchases. A one-unit
standard error shock to net dic purchases produces an increase in net rtf purchases in the
first three periods and then declines thereafter. Overall, these impulse responses suggest
that previous day’s net foreign purchases exert a noticeable impact on net rtf purchases,
while previous day’s net rtf purchases also has an impact on net foreign purchases. It
implies that not only do foreign capital flows affect the trading activity of domestic
institutional investors but also the relation is not unidirectional. To be specific, there is
a feedback relation between net rtf purchases and net foreign purchases.
As for the effect of the three types of institutional trading activity on stock
returns in the TSE, Table 17.5 reveals that net dic purchases on previous day lead
the TSE returns. We can also see in Fig. 17.2e that after Period 3, net dic purchases
exert a negative (and thus destabilizing) effect on market returns, while the other
two institutional investors do not have such an effect over the sample period.
Examining the relationship between market returns and trading activity of the
three types of institutional investors, we find that either the net foreign purchases
or the net dic purchases on previous trading day are affected by the previous day’s
TSE returns. Moreover, the IRFs in Fig. 17.2f, g also reveal a significant positive
relation between TSE returns and net foreign purchases up to four periods and
a significant positive relation between TSE returns and net dic purchases for two
periods, which becomes negative after Period 3. In other words, foreign investors in
the TSE engage in positive-feedback trading, while those of the dic tend to change
their strategy and adopt negative-feedback trading after Period 3.
Table 17.5 indicates that previous day’s NASDAQ returns significantly affect
both current TSE returns and net purchases by the three types of institutional
investors. The impulse response in Fig. 17.2h–k also confirms that previous day’s
NASDAQ returns are positively related to both current returns and net purchases by
these institutional investors, with the exception that a negative relation between
net dic purchases and previous day’s NASDAQ returns is found after Period 4.
17 Dynamic Interactions 515

Such results are much in sync with the expectation since the largest sector that
comprises the TSE-weighted stock index is the electronics industry to which many
listed companies on NASDAQ have a strong connection. In addition, although
previous day’s net qfii and dic purchases also lead the NASDAQ returns, we find
that no significant relation exists except for Period 4 with a significantly negative
relation between them (Figs. 17.1 and 17.2).
The liberalization of Taiwan’s stock market has ushered in significant amount of
short-term inflows and outflows of foreign capital, which have induced fluctuations in
the exchange rate. As is seen from Table 17.5, the TSE returns lead currency returns
and it appears that the initial significant effect of stock returns on currency returns is
negative for the first three periods and then turns to be significantly positive thereafter
(Fig. 17.2n). Given that foreign investors are positive-feedback traders, the capital
inflows is expected to grow in order to increase their stakes in TSE securities when
stock prices rise. Consequently, NT/USD is expected to appreciate.

The Structural VAR Model

The unrestricted VAR model does not consider the effect of current returns on net
purchases by institutions. The prior study by Griffin et al. (2003) includes current
returns in the institutional imbalance equation and finds a strong contemporaneous
positive relation between institutional trading and stock returns. Therefore, to
further examine the relationship between institutional trading and stock returns,
we introduce the current TSE returns (rt) in the net purchases equations of the three
types of institutional investors and reestimate the VAR model before conducting
the corresponding block exogeneity tests. Table 17.6 presents the estimation results.
As indicated in the last row of Table 17.6, we find evidence of a strong contem-
poraneous correlation between current returns and net institutional purchases, which
confirms the finding by the previous research. As shown in Table 17.5, we find no
evidence that past returns lead net rtf purchases when the unrestricted VAR model is
used. When the contemporaneous impact of stock returns on net institutional pur-
chases is considered, we find that past returns also lead net rtf purchases as well as net
purchases by qfii and dic when the structural VAR model is used (Table 17.6). In other
words, net purchases by the three types of institutions are affected by past stock returns
as was evidenced by previous studies. The corresponding impulse response relations
are presented in Fig. 17.3.
Comparing the impulse response relations in Figs. 17.2 and 17.3, it is clear that
when the impact of current returns on net institutional purchases is considered,
a one-unit standard error shock from rt does not produce a positive impulse
response in institutional trading until Period 2. The responses of foreign investors
are rather distinct from those of domestic institutional investors after Period 3. In
general, a sustained positive response from foreign investors is observed, while
a negative response is witnessed for dic and sometimes, an insignificant response
for rtf manifests itself after Period 3.
516 B.-N. Huang et al.

The Threshold VAR Analysis

We pool all the data together when estimating either the unrestricted or restricted
VAR model; however, the trading activity of institutional investors may depend on
whether stock prices rise or fall. A small economy like Taiwan also depends to
a large degree on the sign of NASDAQ index returns. Consequently, to investigate
institutional trading under distinct regimes based on market returns, we use the
multivariate threshold autoregression (MVTAR) model proposed by Tsay (1998) to
test the relevant hypotheses. Let yt ¼ [nasrt, Det, rt, qfiibst, dicbst, rtfbst]0 be a 6  1
vector and the MVTAR model can be described as
!
X
p
yt ¼ f0;1 þ fi, 1 yti  ½1  I ðztd > cÞ
i¼1
!
X
p
þ f0;2 þ fi, 2 yti  I ðztd > cÞ þ et
i¼1

where E(e) ¼ 0, E(ee0 ) ¼ S, and I() is an index function, which equals 1 if the
relation in the bracket holds. It equals zero otherwise. ztd is the threshold variable
with a delay (lag) d.
In order to explore whether institutional trading activity would change during
different domestic and foreign market return scenarios, the potential threshold
variables used are rt1 and nasrt1. Before estimating Eq. 2, we need to test for
possible potential nonlinearity (threshold effect) in this equation. Tsay (1998)
suggests using the arranged regression concept to construct the C(d) statistic to
test the hypothesis Ho:fi,1 ¼ fi,2, i ¼ 0, . . .p. If H0 can be rejected, it implies that
there exists the nonlinearity in data with ztd as the threshold variable. Tsay (1998)
proves that C(d) is asymptotically a chi-square random variable with k(pk + 1)
degrees of freedom, where p is the lag length of the VAR model and k is the number
of endogenous variables yt. Table 17.7 presents the estimation results of the C(d)
statistic.
As shown in Table 17.7, the null hypothesis H0 is rejected using either past
returns on the TSE or NASDAQ, suggesting that our data exhibit nonlinear
threshold effect. Theoretically, one needs to rearrange the regression based on the
size of the threshold variable ztd before applying a grid search method to find the
optimal threshold value c*. Nonetheless, our goal is to know whether the institu-
tional trading behavior depends on the sign of market returns, as such the threshold
is set to zero in a rather arbitrary way. Table 17.8 lists the results of block
exogeneity tests for the lead-lag relation in the rt1 < 0 and rt1  0 regimes,
respectively.
The interaction among institutional investors is depicted in Table 17.8: Current
net purchases by foreign investors affect that by domestic institutions when previ-
ous day’s TSE returns are negative. Note that no such relation is evidenced when
17 Dynamic Interactions 517

previous day’s TSE returns are positive. A feedback relation between rtf and
qfii is observed when rt1 is positive or negative. However, dic is found to lead
rtf only when rt1 is positive. Such results reveal different institutional trading
strategies under distinct return regimes. The demonstration effect – previous day’s
net foreign purchases have on domestic institutions using the unrestricted VAR
model – seems to surface only when previous day’s market returns are negative.
Therefore, it may produce misleading results if we fail to consider the sign of
previous returns.
The impulse responses in Fig. 17.4 illustrate that the responses of dic and rtf
from the qfii shock are quite similar to the ones in Fig. 17.2f, g. As for the impact of
previous day’s market returns on current net purchases by institutions, it can be
shown via the MVTAR model that market returns lead net purchases by qfii and dic
when previous day’s market returns are negative, which is consistent with the
finding using the one-regime VAR model. When previous day’s market returns
are positive, market returns lead net purchases by the dic only. Obviously, returns
have more influence on net institutional purchases when previous day’s
returns were negative. In addition, we find that net dic purchases on the previous
day may affect current returns when the one-regime VAR model is used. Actually,
the MVTAR analysis reveals that such a relation emerges only when previous day’s
market returns are positive. The impulse responses depicted in Fig. 17.4 (Panel B)
demonstrate that a one-unit standard error shock to net dic purchases produces an
increase in market returns in Period 2, and then they turns to be negative after
Period 4, a result similar to those using the one-regime VAR model.
To further investigate whether the interaction among institutions and the
relationship between institutional trading and stock returns are affected by the
sign of previous day’s NASDAQ index return, nasrt1 is used as the threshold
variable. That is, block Granger causality tests are performed by splitting our
data into two regimes based on the sign of the variable nasrt1. Table 17.9 reports
the results.
The results indicate that net qfii purchases lead that of two domestic institutional
investors regardless of the sign of previous day’s NASDAQ returns. Moreover,
when nasrt1 < 0, net rtf purchases lead net qfii purchases, which is in line with that
using the one-regime VAR model. However, when nasrt1  0, the net purchases
by either dic or rtf lead net qfii purchases, and net qfii purchases lead net purchases
by either dic or rtf. In other words, we find strong evidence of a feedback relation
between net foreign purchases and two domestic net purchases when previous day’s
NASDAQ returns are positive. The results pertaining to the impact of previous
day’s returns on institutional trading parallel those using the unrestricted VAR
model: Previous day’s returns have an impact on the net purchases by qfii and dic,
but not on net purchases by rtf regardless of the sign of previous day’s NASDAQ
returns. As for the impact of net institutional purchases on previous day’s stock
returns, only net dic purchases still lead stock returns when nasrt1 < 0, as was the
case in the one-regime model. However, we find that net qfii purchases also lead
stock returns when nasrt1  0.
518 B.-N. Huang et al.

References
Brennan, M., & Cao, H. (1997). International portfolio investment flows. Journal of Finance, 52,
1851–1880.
Cai, F., & Zheng, L. (2002). Institutional trading and stock returns (Working paper). Ann Arbor:
University of Michigan Business School.
Choe, H., Kho, B. C., & Stulz, R. (1999). Do foreign investors destabilize stock markets? The
Korean experience in 1997. Journal of Financial Economics, 54, 227–264.
Froot, K. A., O’Connell, P. G. J., & Seasholes, M. S. (2001). The portfolio flows of international
investors. Journal of Financial Economics, 59, 151–193.
Griffin, J. M., Harris, J. H., & Topaloglu, S. (2003). The dynamics of institutional and individual
trading. Journal of Finance, 58, 2285–2320.
Grinblatt, M., & Keloharju, M. (2000). The investment behavior and performance of various
investor-types: A study of Finland’s unique data set. Journal of Financial Economics, 55,
43–67.
Hamao, Y., & Mei, J. (2001). Living with the “enemy”: An analysis of foreign investment in the
Japanese equity market. Journal of International Money and Finance, 20, 715–735.
Huang, C.-H., & Hsu, Y. Y. (1999). The impact of foreign investors on the Taiwan stock exchange.
Taipei Bank Monthly, 29(4), 58–77.
Kamesaka, A., Nofsinger, J. R., & Kawakita, H. (2003). Investment patterns and performance of
investor groups in Japan. Pacific-Basin Finance Journal, 11, 1–22.
Karolyi, G. A. (2002). Did the Asian financial crisis scare foreign investors out of Japan? Pacific-
Basin Finance Journal, 10, 411–442.
Lakonishok, J., Shleifer, A., & Vishny, R. (1992). The impact of institutional trading on stock
price. Journal of Financial Economics, 32, 23–44.
Lee, T. S., & Oh, Y. L. (1995). Foreign investment, stock market volatility and macro variables.
Fundamental Financial Institutions, 31, 45–47 (in Chinese).
Lee, Y.-T., Lin, J. C., & Liu, Y. J. (1999). Trading patterns of big versus small players in an
emerging market: An empirical analysis. Journal of Banking and Finance, 23, 701–725.
Nofsinger, J. R., & Sias, W. (1999). Herding and feedback trading by institutional and individual
investors. Journal of Finance, 54, 2263–2295.
Sadorsky, P. (1999). Oil price shocks and stock market activity. Energy Economics, 21, 449–469.
Sias, R. W., & Starks, L. T. (1997). Return autocorrelation and institutional investors. Journal of
Financial Economics, 46, 103–131.
Tsay, R. S. (1998). Testing and modeling multivariate threshold models. Journal of American
Statistical Association, 93(443), 1188–1202.
Wang, L.-R., & Shen, C. H. (1999). Do foreign investments affect foreign exchange and stock
markets? – The case of Taiwan. Applied Economics, 31(11), 1303–1314.
Methods of Denoising Financial Data
18
Thomas Meinl and Edward W. Sun

Contents
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520
18.2 Denoising (Trend Extraction) of Financial Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520
18.3 Linear Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
18.3.1 General Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
18.3.2 Transfer Functions: Time Versus Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . 525
18.4 Nonlinear Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
18.4.1 General Perception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
18.5 Specific Jump Detection Models and Related Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
18.5.1 Algorithms for Jump Detection and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
18.5.2 Further Related Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
18.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536

Abstract
Denoising analysis imposes new challenges for financial data mining due to the
irregularities and roughness observed in financial data, particularly, for instan-
taneously collected massive amounts of tick-by-tick data from financial markets
for information analysis and knowledge extraction. Inefficient decomposition of
the systematic pattern (the trend) and noises of financial data will lead to
erroneous conclusions since irregularities and roughness of the financial data
make the application of traditional methods difficult.
In this chapter, we provide a review to discuss some methods applied for
denoising analysis of financial data.

T. Meinl (*)
Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
e-mail: thomas.meinl@kit.edu
E.W. Sun
KEDGE Business School and BEM Management School, Bordeaux, France
e-mail: edward.sun@bem.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 519
DOI 10.1007/978-1-4614-7750-1_18,
# Springer Science+Business Media New York 2015
520 T. Meinl and E.W. Sun

Keywords
Jump detection • Linear filters • Nonlinear filters • Time series analysis • Trend
extraction • Wavelet

18.1 Introduction

Technical developments allow corporations and organizations to instantaneously

collect massive amounts of data, particularly the tick-by-tick data from financial
markets. Mining financial data turns to be the foundation for financial informatics
and stimulates the research interest in analyzing information conveyed at different
frequencies in this data. Financial data have the complex structure of irregularities
and roughness that are caused by a large number of instantaneous changes of the
markets and trading noises. The noises conveyed in the financial data usually
illustrate heavy tailedness, that is, the underlying time series data exhibits a large
number of occasional jumps. Ignoring these irregularities can easily lead to erro-
neous conclusions for data mining and statistical modeling. As consequence, the
statistical data mining methods (or models) require a denoising algorithm to clean
the data in order to obtain more significant results (see Sun and Meinl (2012)).
Most data cleaning methods focus only on a known type of irregularity.
In financial data, the irregularity is manifold, that is, the irregularity varies along
with time and changes with different measuring scales (see Fan and Yao (2003) and
Sun et al. (2008)). Finding an effective denoising algorithm turns out to be the
initial task of financial data mining (see Au et al. (2010) and Meinl and Sun (2012)).
In this chapter we will outline the classic and newly established methods for the
denoising analysis (trend extraction) of financial data analysis. We talk about
different approaches and a respective classification of them. Based on this classi-
fication we focus on nonparametric methods (i.e., linear and nonlinear filters) and
examine their pros and cons.

18.2 Denoising (Trend Extraction) of Financial Data

There is no universal definition of trend which applies to all application in different

fields, and it is generally accepted that a trend is a slowly evolving component that
is the main driving force for long-term development beneath the system. Pollock
(2001) characterizes trend as being limited to certain low frequencies of the data.
This notion excludes any noisy influences and fluctuations from higher frequency
levels. However, this notion of trends is not satisfactory for many financial data we
encounter in practice.
Some theoretical models (like the Black-Scholes model) do not incorporate
aspects like seasonalities or even jumps; they are still widely used today in practice,
assuming perfect division between the trend and stochastic fluctuations. It is
insufficient for many financial data we measure today, especially, when considering
trends over longer time horizon, we perceive significant jumps or steep slopes
18 Methods of Denoising Financial Data 521

which cannot be attributed to be part of the persistent stochastic noise. Another

example we can observe in electricity markets, where it has been found that jumps
occur on such a regular basis that is has been found reasonable to model these by
specific stochastic processes; see Seifert and Uhrig-Homburg (2007). These patterns
contradict the slow evolving characteristic and the low-frequency notion generally
associated with trends and have been considered as an inherent part of them.
A basic model (among others proposed by Fan and Yao (2003)) is stated as
follows. Given a time series Xt, t
0, this series can be decomposed into

X t ¼ # t þ st þ Y t , (18.1)

where #t represents a slowly varying function known as trend component, st a single

of combination of periodic functions (i.e., seasonal components), and Yt the
stochastic component, assumed to be stationary.
A trend is a mostly slow (in respect to the noise) evolving pattern in a time series,
with its driving force not being attributed to any noise present in the signal. Trends
may also exhibit edged frontiers (i.e., jumps and sudden regime changes) as well as
steep slopes, roofs, and valleys (see Joo and Qiu (2009)), as long as these patterns
can be contributed to the long-term dynamics of the time series and do not stem
from the noise component responsible for short-term variations.
We note that trend must always be interpreted in respect to the time series at
hand (i.e., the period coverage of the financial data) and the goal of the data
analysis, that is, on which scale the trend and the noise are relevant. We ignore
any distinction between long-, intermediate-, and short-term trends and/or
seasonalities, as these usually depend on the context of the financial data.
The question that may now arise is if we consider a trend # to be specified
according to Eq. 18.1, how does this agree with the notion that the trend may also
exhibit jumps, which clearly contradict this definition? Without going further into
this, we only note that this kind of trend would require a different (and more
accurate) definition of how trend is to be defined in a time series, than the generally
accepted notion in today’s literature. While we will not propose a new definition
either and leave this task to others, we can point out how jumps enter the trend.
As pointed out by, for example, Tsay (2005), jumps in financial time series data,
and particularly in high-frequency data, are attributed to external events, like the
increase or drop in interest rates by some governmental financial institution. These
events can be considered to happen only occasionally and are very sparse in relation
to the frequency the data is measured, that is, the amount of data exhibiting no
jumps at all. In the field of high-frequency financial data analysis, jumps are thus
assumed to be extreme events that happen with low probability but form neverthe-
less part of the stochastic distribution and must be considered to be modeled there.
Thus, Y will be modeled either by stochastic processes including jump components
(see, e.g., Seifert and Uhrig-Homburg (2007)) or by a distribution itself, depending
on the model. Such a distribution for high-frequency financial data has then found
to be heavy tailed, that is, jumps happen with enough regularity that they cannot
simply be discarded as nonrecurring events; see, for example, Rachev (2003).
522 T. Meinl and E.W. Sun

However, as these extreme events can have an enormous impact on the stochastic
variance analysis and its succeeding usage, and furthermore could lead to mislead-
ing results in the regions of the signal without any jumps, it is generally
preferred to include them into the trend component # rather than the stochastic
component Y.
In, for example, Wang (1995), a definition of an a-cusp in a continuous function
f at x0 is given, that is, if there exists an e > 0 so that

jf ðx0 þ hÞ  f ðx0 Þj
Kjhja (18.2)

holds for all h 2 [x0  e, x0 + e]. For a ¼ 0, f is said to have a jump at x0. It is
commonly agreed that the jump should significantly differ from the other fluctua-
tions (i.e., the noise) in the signal. As said above, jumps are just one particular
pattern of extreme events we are interested in. Others are steep slopes, roofs, and
valleys, which in Joo and Qiu (2009) are defined by at this point having a jump in
the first-order derivative of the regression curve.
Other extreme events frequently occurring in many practical applications are spikes
and outliers. However, these are usually undesirable features that should not be
included in the trend of affect it by any means. This is due to the following reasons.
First, in many cases, these outliers or spikes consist only of one or very few points often
caused by measurement errors, and it is obvious that they were caused by some factor
that plays no vital role in the ongoing time series analysis (unless the focus is on what
caused these outliers). Second, while jumps imply a permanent change in the
whole time series, outliers do not contribute to this. While we are aware that
the distinction between a few (adjacent) outliers and roofs/valleys is not precise,
from the context of the time series in most cases, it is evident whether an occurrence
should be considered as an outlier that is to be neglected or a significant feature to be
included in the trend.
In this work we only consider homogeneous financial time series data. However,
in many applications and particularly for high-frequency financial time series data,
this data is initially irregularly spaced. Homogeneity in this context means that for
a given series Xt, t 2 N, holds t + 1  t ¼ c, with a constant c > 0, that is, all time
steps are equally spaced. As we will see, this is not always the case for empirical
time series, especially in the area of financial high-frequency data. In this case, it is
necessary to preprocess the inhomogeneous (i.e., irregularly spaced) time series by
interpolation methods in order to regularize the raw data. Though there exist models
which can handle inhomogeneous time series directly (see Dacorogna (2001), but
they also remark that most today’s models are suited for regularly spaced time
series only), regarding the methods we discuss in the following sections, we restrict
ourselves to homogeneous ones.
In this chapter we focus on nonparametric methods for trend extraction. This is
due to the reason that in most time series we analyze in this work, we cannot
reasonably assume any model for the underlying trend. Yet, as noted in the
framework above, in case such assumptions hold, we can expect those models to
18 Methods of Denoising Financial Data 523

perform better than nonparametric models, since they exploit information that
nonparametric approaches cannot. Furthermore, the commitment to certain para-
metric time series or trend models can be seen as a restriction when considering the
general case, which may lead even to misleading results in case the trend does not
match model, as certain patterns might not be captured or considered by the
model itself. This can easily be seen at a most basic example, in case a linear trend
is expected, which in most cases will only be a poor estimator for any nonlinear
trend curve. In this case nonparametric approaches are less restrictive and can
more generally be applied, while of course not delivering the same accuracy
as parametric models which exactly match the underlying trend, with only
their parameters to be calibrated. If, e.g., the trend follows sinusoidal curve, a
sinusoidal curve with its parameters being estimated by, e.g., the least-squares
method will almost surely provide a better accuracy than any other nonparametric
approach. On the other hand, if the underlying trend is linear or even contains
only little deviations from a perfect sinusoidal curve, a sinusoidal fit to this trend,
it has been shown at simple examples that the parametric sinusoidal approach can
lead to confusing results and conclusions.
We further require that the method used for trend extraction be robust, that is,
the results are reliable and the error can be estimated or is at least bounded in some
way. In many cases (see, e.g., Donoho and Johnstone (1994)), the robustness
of a method is shown by proving its asymptotic consistency, that is, its conver-
gence towards a certain value for certain parameters tending towards infinity.
It should be remarked that the robustness should be independent of the
time series itself and/or any specific algorithm parameter sets, in order to be
applicable in practice. Of course this does not exclude specific assumptions on the
time series that must be met or parameter ranges for which the algorithm is
defined.
Therefore, as we cannot reasonably assume any model for the any time series
in general, in this work, we focus on nonparametric approaches only. Within these
approaches we focus on two main branches for trend extraction: linear
and nonlinear filters. This is for the reason because linear filters are known and
have proven to deliver a very smooth trend (given the filtering window size is
large enough), while nonlinear filters excel at preserving characteristic patterns
in a time series, i.e., especially jumps. Both methods in general require only
very few (or none at all) information about the underlying data, be-
sides their configuration of weights and calibration parameters, and are thus
applicable to a wide range of time series, independent of the field the data was
measured in.
Although there exists a variety of other nonparametric methods, most of these
already rely on specific assumptions or choices of parameters which in general
cannot easily be derived for any time series data or different analysis goals.
Nevertheless, for the sake of completeness, later in this chapter we list some
alternative methods, also including parametric approaches, which have been
applied in economic, financial, or related time series data.
524 T. Meinl and E.W. Sun

18.3 Linear Filters

Linear filters are probably the most common and well-known filters used for trend
extraction and additive noise removal. We first provide the most general notion of
this filter class and then depict the two viewpoints of linear filters and how they can
be characterized. While this characterization on the one hand is one of the most
distinguishable advantages of linear filters, on the other hand at the same time, it
leads to the exact problem we are facing in this work, that is, the representation of
sharp edges in otherwise smooth trends.

18.3.1 General Formulation

The filtered output depends linearly from the time series input. Using the notation of
Fan and Yao (2003), a linear filter of length 2h + 1 can be defined as
X
h
^t ¼
X wi Xtþi , (18.3)
i¼h

with X ^ t the filtered output and wi the filter weights. These kinds of filters are
also known as (discrete) convolution filters, as the outcome is the convolution of
the input signal with a discrete weight function, where often the following notation
is used:
X
h Xh
w Xt :¼ wi Xti ¼ wti Xi : (18.4)
i¼h i¼h

Thus, for every data point Xt, the filtered output X ^ t is the result of weighted
summation of data points around t. Applied for the whole time series, this results in
weighted average window of size L ¼ 2h + 1 which is moved throughout the whole
series. The size of this window is also called the bandwidth of the filter. It is also
common notation to set h  1 even if the window filter size should be finite, with
wi ¼ 0 for jij > h.
The probably best known linear filter is the mean filter, with wi ¼ 2h + 1, that is,
all filter weights are uniformly distributed. A more general viewpoint is given by
the notion of kernel filters. Given a kernel function w.l.o.g. with support [1, 1],
this function assigns the weights according to
K ði=hÞ
wi ¼ Xh : (18.5)
j¼h
K ðj=h Þ

Commonly used examples are the Epanechnikov kernel

3 
K E ð uÞ ¼ 1  u2 þ , (18.6)
4
18 Methods of Denoising Financial Data 525

the Gaussian kernel

 2
1 u
K ðuÞ ¼ pffiffiffiffiffiffi exp
G
, (18.7)
2p 2

and the symmetric beta family

1  g
K Bg ðuÞ ¼ 1  u2 I juj1 : (18.8)
Bð1=2, g þ 1Þ

For the values g 2 {0, 1, 2, 3}, the kernel KBg (u) corresponds to the uniform,
Epanechnikov, biweight, and triweight kernel functions, respectively.

18.3.2 Transfer Functions: Time Versus Frequency Domain

The previous section depicts the linear filtering method in the time domain, i.e., we
look at the time series Xt and its respective filtered output X ^ t and how they evolve
over time t. Another perception can be given by taking the frequency domain into
account. For all linear filters, we cannot only give its definition as depicted earlier,
but also in respect to the frequencies, the filters let pass. This notion can be derived
as follows.
While the sequence of filter weights wi, also called impulse response sequence,
determines the filtered output in the time domain (or equivalent are the linear filter’s
time domain representation), via the discrete Fourier transform (DFT), we can
derive the transfer function

X
1
Wð f Þ ¼ wj ei2pf j , (18.9)
j¼1

its counterpart in the frequency domain, also called frequency response function.
Alternatively, if this formulation is given in the beginning, we can also derive the
weights via the inverse transform:

Z 1=2
wj ¼ W ð f Þei2pf j df : (18.10)
1=2

Obviously, these two formulations are equivalent, as one can be derived from the
other, and vice versa. By considering the transfer function’s polar representation,

W ð f Þ ¼ jW ð f Þjeiyðf Þ , (18.11)
526 T. Meinl and E.W. Sun

with jW ð f Þj the gain function. The magnitude in gain jW ð f Þj describes the linear
filters behavior in the frequency domain, that is, what kind of frequencies and their
respective proportions will let passed or be blocked. For our needs, it
satisfies to distinguish between high – and low-pass filters, that is, filters that
let pass either the high frequencies and block the lower ones, or vice versa. Of
course, other filter types exist; for example, by combining high – and low-pass
filters (e.g., in a filter cascade), one can derive band-pass and stop filters,
so that the frequency domain output will be located only in a certain frequency
range. In this work we are specifically interested in low-pass filters, as they
block the high-frequency noise and the output consists of the generally
low-frequency trend.
As we assume that the weights wj are real valued, one can show
(see, e.g., Percival and Walden (2006)) that W ðf Þ ¼ W ð f Þ, and, with jW  ðf Þj ¼
jW ð f Þj , it follows that jW ðf Þj ¼ jW ð f Þj . Therefore, the transfer functions
are symmetric around zero. Because of its periodicity, we need to consider W ð f Þ
only an interval of unit length. For convenience, this interval is often taken
to be [1/2, 1/2], i.e., j fj  1/2. Therefore, with above-depicted
symmetry, it suffices to consider f 2 [0, 1/2] in order to fully specify the transfer
function.
While saying that certain frequencies are blocked and others are passed, this
holds only approximately true, since the design of such exact frequency filters is not
possible, but always a transition between the blocked and passing frequencies. The
goal of many linear filters is either to minimize these transitions (i.e., the range of
by this affected frequencies), which, on the other hand, inevitably causes ripples in
the other frequencies, that is, they are not any longer blocked, or to let them pass
completely (see Bianchi and Sorrentino (2007) and the references therein for
further details about this topic).
As we have seen at above examples, linear filters can be designed either from
a time or from a frequency perspective. The time domain usually focuses on putting
more weights on the surrounding events (i.e., events that recently before or
occurred shortly after) and, thus, gives an (economic) interpretation similar to
the, for example, ARMA and GARCH models. On the contrary, the frequency
domain is based on the point of view that certain disturbances (almost) exclusively
are located in a certain frequency range and are thus isolated from the rest of the
signal. Also, the slowly evolving trend can be seen to occupy only the lower
frequency ranges. Thus, the (economic) meaning lies here in the frequency of
events; see, for example, Dacorogna (2001).
Although linear filters can be designed to block or let pass certain frequencies
nearly optimal, at the same time this poses a serious problem when facing trends
that exhibit jumps or slopes. As these events are also located in the same (or in case
of slopes the adjacent) frequency range as the high-frequency noise, this has the
effect that jumps and edged frontiers are blurred out, while steep slopes mostly are
captured with poor precision only. Hence, from a frequency perspective, a smooth
trend and edge preservation are two conflicting goals. This is as the linear filters are
not capable to distinguish between the persistent noise and a single events that,
18 Methods of Denoising Financial Data 527

while located in the same frequency range, usually have a higher energy and, thus,
significantly larger amplitude. Thus, the same filtering rule is applied throughout
the whole signal, without any adaption. Note, however, that linear filters still give
some weight to undesirable events like outliers, due to their moving average
nature. Thus, while some significant features like jumps are not displayed in
enough detail, other unwanted patterns, like spikes, still partially carry over to
the filtered output. To overcome all these drawbacks for that kind of trends or
signals, besides the class of linear filters, the class of nonlinear filters has been
developed.

18.4 Nonlinear Filters

As we have seen, linear filters tend to blur out edges and other details though they
form an elementary part of the time series’ trend. In order to avoid this, a wide range
of nonlinear filters has been developed which on one hand preserve those details
while on the other try to smooth out as much of the noise as possible. We do not find
nonlinear filters only in time series, but they were in many cases developed
specifically for two-dimensional signals, specifically images, where the original
image, probably corrupted by noise during data transmission, consists mainly of
edges, which form the image.

18.4.1 General Perception

While linear filters generally provide a very smooth trend achieved through aver-
aging, two characteristics pose a problem for this class of filters:
• Outliers and spikes
Single, extreme outliers and spikes can cause the whole long-term trend to
deviate in the same direction, though they obviously do not play a part in it.
• Jumps, slopes, and regime changes
Whenever there occurs a sudden external event in the underlying main
driving force, it causes the trend to jump, that is, contrary to spikes it
changes permanently onto another plane. While slopes are not that extreme,
they also show a similar behavior as they decay or rise with the trend’s unusual
degree.
The reasons for the deviation sensitivity to these events are given by one of the
most favorable linear filters’ characteristics themselves: It follows directly from
them being characterizable in terms of frequency passbands we explained earlier
that all frequencies are treated the same (i.e., filtered according to the same rule)
throughout the whole signal. This means that no distinction is made (and even
cannot be made) between noise and those patterns, as they are located in approx-
imately the same frequency range. Technically, as long as an outlier or a jump is
contained in the weighted moving average filtering window, also a weight is
assigned to these outlier data points or the points on the other plane, i.e., before
528 T. Meinl and E.W. Sun

and after the jump. Nonlinear filtering procedures try to avoid this by using
a different approach, for example, by considering a single value only (instead of
multiple weighted ones) that was selected from an ordered or ranked permutation of
the original values in the filtering window.
Although we cannot characterize nonlinear filters in the same way as we do with
linear filters (i.e., by transfer functions) besides their characteristics not being
classifiable in that way, according to Peltonen et al. (2001), we can divide the
whole class of these filters into several subclasses that share the same or similar
approaches. Among them, there are stack filters, weighted median filters, polyno-
mial filters, and order statistic filters. Astola and Kuosmanen (1997) provide two
different taxonomies for further classification, though they remark that these divi-
sions are not unique. In their work, they extensively show how those different filters
behave (i.e., their characteristics) when applied onto different benchmark signals in
respect to the mean absolute error (MAE) and mean squared error (MSE) measures.
The behavior of nonlinear filters is generally characterized by their impulse and
step response, i.e., the filtered output when the input consists of a single impulse or
step only. These impulses generally are given by the sequence [. . . , 0, 0, 0, a, 0, 0,
0, . . .] and [. . . , 0, 0, 0, a, a, a, . . .], respectively, with a 6¼ 0. Though in most cases
no analytical result can be given, these characteristics help us to understand how the
nonlinear filter behaves in respect to those patterns, for which the linear filters
generally fail to deliver adequate results.
Despite their undoubtedly good ability to preserve outstanding features in a time
series while extracting the trend, nonlinear filters also suffer some drawbacks:
1. Insufficient smoothness
Though most nonlinear filters try to deliver a smooth trend and a good resolution
of edged frontiers, beyond the jumps’ surrounding regions, they usually fail to
deliver a smooth trend as accurate as even simple linear filter (e.g., the mean
filter) provides. Yet, by applying further smoothing procedures (e.g., by recur-
sive filters or some kind of linear filtering on the nonlinear output, with a smaller
bandwidth, as most of the noise is already smoothed out) comes at the price that
the prior preserved details of jumps or slopes tend to get lost.
This effect is even aggravated when the high-frequency noise present in the
signal is extremely high, i.e., the trend which evolves, besides the jumps, quite
slowly is dominated by the noise with extremely high amplitudes. Some filters
try to counter this effect, but they either:
• Provide only a tradeoff between an overall smooth signal and poor jump
resolution, or a trend still exhibiting ripples but preserved edges, or
• Rely on further information about the time series itself, that is, the noise
component and its structure, the jumps or the trend itself
This is due to the problem that most filtering rules are applied throughout the
whole signal, that is, they do not adapt themselves sufficiently when the filtering
window approaches a jump. An overview of these and other nonlinear filters’
performance is given by Astola and Kuosmanen (1997), proving again the
well-known insight that there cannot exist one solution that performs optimal
for all cases. Though the authors also report some approaches that try to
18 Methods of Denoising Financial Data 529

incorporate that behavior, these filters provide only dissatisfactory results (see
the examples below).
2. Lack of frequency control
Another feature nonlinear filters are lacking is the ability to regulate the filtered
output in terms of frequency passbands, as linear filters do. Since Pollock (2001)
defines the trend in terms of frequency bands and Ramsey (1999, 2002) points
out that frequency analysis is an important aspect in financial time series, so is
frequency control. Though not necessary for all applications, the ability to
a priori control and regulate the filters output (in contrast to an only a posteriori
frequency analysis of the filtered result) may come handy when one wants to ensure
that certain frequencies are not contained in the output, that is, when certain
information about the noise frequencies is at hand, the analyst can decide before
the actual filtering process (and thus, without any try and error procedures) what
frequency parts should be filtered out. Although a nonlinear filter can also provide
the same or a similar result, no theoretical results or statements are available before
the filtering procedure has been carried out completely. This incapacity of the
nonlinear filter follows directly from the fact that nonlinear filters do not rely on
frequency passbands, as they must be able to filter events, even though they occupy
nearly the same frequency range (i.e., noise and jumps), due to different rules.
As Astola and Kuosmanen (1997) classify linear filters to be a subclass of the
class of nonlinear filters, above statement does not exactly hold true, i.e., it would
mean that a subclass of nonlinear filters can be characterized by their transfer
function and frequency output. We, however, separate the classes of linear and
nonlinear filters by whether or not a filter can be described by a transfer function or
not, which marks a strict division of these two classes.

18.4.1.1 Examples
To illustrate the different courses of action of nonlinear filters and give the reader an
idea of their general procedure, in this section we outline several examples of above
named subclasses. As this list cannot be exhaustive by any means, of course, we
note that we do not take filters into account that already rely on specific assumptions
of the systems beneath the time series themselves.

Trimmed Mean Filter

This filter works essentially as the mean filter, with the difference that the extreme values
of the ordered series X(i) are trimmed. Index t is omitted here as the order is no longer in
concordance with time. Therefore, an (r, s)-fold trimmed mean filter is given by

1 X
N s
X ð iÞ : (18.12)
N  r  s i¼rþ1

A special case is the choice of r ¼ s. A further modification of the trimmed mean

filter is not to discard the ordered values beyond X(r) and X(s), but instead replace
them by X(r+1) and X(s+1) themselves. This is the Winsorized mean filter:
530 T. Meinl and E.W. Sun

!
1 X
N s
r  Xðrþ1Þ XðiÞ þ s  XðNsÞ : (18.13)
N i¼rþ1

In these methods, the (r, s) tuple is dependent of the data itself. Other filters
consider to make these values independent from the data or dependent from the
central sample itself, i.e., nearest neighbor techniques. All those filters have in
common that they discard all samples from the ordered series being too far away
(respectively) according to some measure.

L-Filters and Weighted Median

L-filters (also called order statistics filters) make a compromise between the
weighted moving averages of linear filters and the nonlinear ordering operation.
The idea is that the filtered output is generated by weighted averages over the
ordered samples, that is,

X
N
wi XðiÞ , (18.14)
i¼1

with wi the filter weights. A similar notion is given by weighted median filters, where
the weights are assigned to the time ordered sample Xt and where the weights denote
a duplication operation, i.e., wi ∘Xt = Xt, 1 , . . . , Xt, wi. The output is then given by

median fw1 ∘X1 , . . . , wN ∘XN g: (18.15)

Ranked and Weighted Order Statistic Filters

An rth-ranked-order statistic filter is simply given by taking X(r) as the filter output.
Examples are the median, the maximum (r ¼ N), and the minimum (r ¼ 1)
operation. This can also be combined with weights as depicted above, that is,

rth order statisticfw1 ∘X1 , . . . , wN ∘XN g: (18.16)

Hybrid Filters
Another approach is the design of nonlinear filters consisting of filter cascades,
that is, the repeated application of different filters on the respective
outputs. A general formulation is, for example, given by rth-order statistic
F1(X1, . . . , Xn), . . . , FM(X1, . . . , Xn), where F1, . . . , FM can denote any other
filtering procedure. A concrete example is the median hybrid filter that combines
prior linear filtering procedure with a succeeding median ordering operation, i.e.,
( )
Xk X N
median ð1=kÞ Xi , Xkþ1 , ð1=kÞ Xi : (18.17)
i¼1 i¼kþ2
18 Methods of Denoising Financial Data 531

Selective Filters
An interesting approach is given by the principle of switching between different
output rules depending on some selection rule, for example, based on the fact that
the mean filter delivers a larger (smaller) output than the median filter when the
filtering window approaches an upward (downward) jump. Thus, a selection rule
would be given by
   
mean X1 , ..., XN
median X1 , ..., XN : (18.18)

A certain drawback of this selection rule is that it is one-sided, that is, it considers
only the first half of the region around the jump. This is due to the fact that the mean
for the second half, after the jump has occurred, is generally smaller (larger) than the
median. Other rules can include thresholds and aim at deciding whether there has
actually happened a jump or there was an impulse in the signal, which is not caused
by the noise distribution, but due to some other explanatory effect.
Above-depicted examples of nonlinear filters should give the reader an overview
over the most common methods applied in practice. For detailed information about
each filter’s characteristics, advantages, and drawbacks, we refer to Astola and
Kuosmanen (1997), where there are also the references to the original works to be
found. Yet, we see that basically most of these filters rely on some ordered statistics,
with their input or output modified prior or afterwards, respectively. Since this basic
principle applies to most filters not directly dependent on some specific character-
istic or assuming a certain structure of the original series to be filtered, the different
methods pointed out above can be combined in numerous ways. In many cases,
however, even though we portrait only the very basic methods, we see that almost
all of them already incorporate an implicit or explicit choice of additional param-
eters besides the filter bandwidth, either by weights, rules, or thresholds. These
choices introduce further biases into the filtering process. Though some of these
parameters can be chosen to be optimal in some sense (i.e., minimize a certain
distance measure, e.g., the MSE for the L-filters), they lack the concrete meaning of
weights we have for linear filters.

18.5 Specific Jump Detection Models and Related Methods

In this section we list further related methods that are concerned with the estimation
of long-term trends exhibiting edged frontiers, i.e., jumps and/or steep slopes. We
first review methods being explicitly developed either only for the detection of
jumps in a signal corrupted by noise or approaches that also include capturing (i.e.,
modeling) those very jumps. We show the advantages and limits of applications of
these methods, highlighting in which aspects further research is still necessary. We
conclude this chapter by listing some of the most in practice well-known methods.
532 T. Meinl and E.W. Sun

18.5.1 Algorithms for Jump Detection and Modeling

The issue of detecting and modeling jumps in time series has been recognized as an
essential task in time series analysis and therefore has already been considered
extensively in the literature. Though we introduce wavelet methods in the next
chapter only, we list the works based on these methods here as well without going
into details. We only note that wavelets, based on their characteristics, make
excellent tools for jump and spike detection, as it is this what they were developed
for in the first place (see Morlet (1983)).
Generally, the most appreciated procedures in the recent literature can be seen as
two different general approaches. One is via wavelets, and the other uses local
(linear) estimators and derivatives.
One of the first approaches using wavelets for jump detection in time series,
besides the classical wavelet literature, for example, Mallat and Hwang (1992), was
given by Wang (1995, 1998). He uses wavelets together with certain data-
dependent thresholds in order to determine where in the signal jumps have hap-
pened and whether they are significantly different from short-varying fluctuations,
and provides several benchmark signals. Assumptions about the noise
structure were made according to Donoho and Johnstone (1994), that is, the
approach is applicable for white (i.e., uncorrelated) Gaussian noise only. This
work was extended by Raimondo (1998) to include even more general cusp
definitions. More recent contributions extend these works to stationary noise
(Yuan and Zhongjie (2000)) and other distributions (Raimondo and Tajvidi
(2004)) and also provide theoretical results about asymptotic consistency.
A further application specifically on high-frequency data is given by Fan and
Wang (2007).
The other line is given by Qiu and Yandell (1998) who estimate jumps using
local polynomial estimators. This work is continued in Qiu (2003) where jumps are
not only detected but also represented in the extracted time series. Gijbels
et al. (2007) further refine the results by establishing a compromise between
a smooth estimation for the continuous parts of a curve and a good resolution of
jumps. Again, this work is limited to pure jumps only and, since it uses local linear
estimators as the main method, has no frequency interpretation available. Sun and
Qiu (2007) and Joo and Qiu (2009) use derivatives to test the signal for jumps and to
represent them.
Finally, we note a completely different approach that is presented in Kim and
Marron (2006): a purely graphical tool for the recognition of probable jumps (and
also areas where almost certainly no jumps occurred). Yet, the authors confess that
their work is only to be seen as a complementary approach, and refer to Carlstein
et al. (1994) for further thorough investigation.
We note that there exists already a good body of work about how to detect (i.e.,
estimate their location), and even how to model jumps (i.e., estimate their height),
though in most works bounded to strict requirements on the noise or the trend
model. However, although most models will also automatically include the
detection of steep slopes, they fail at modeling the slope itself. While a jump
18 Methods of Denoising Financial Data 533

can easily be presented (either by indicator functions or any other methods used in
the cited works above), matters are different with slopes: Since there can be given
no general formulation or model of the exact shape of a slope, any parametric
approach will fail or deliver only a poor approximation if the model does not fit
the occurred slope. Examples of such different kinds of slopes are innumerous:
Sine, exponential, and logarithmic decays are only the most basic forms to
approximate such events, which in practical examples rarely follow such ideal-
ized curves. Naturally, only nonparametric approaches will adapt themselves to
the true shape of the underlying trend but generally suffer the same drawbacks as
all linear and nonlinear methods pointed out above, i.e., they always have to
bargain a tradeoff between bias and signal fidelity.

18.5.2 Further Related Methods

We conclude this section by outlining the most popular and established filtering
methods.

18.5.2.1 General Least-Squares Approaches

The most general approach can be seen by setting up a parameterized model and
calibrating the model afterwards, generally using some minimization procedure in
respect to some error measure. This can be seen as a straightforward approach,
requiring only the initialization of an appropriate model and choice of error
measure.
An example from the energy market is given by Seifert and Uhrig-Homburg
(2007) and Lucia and Schwartz (2002), who set up a trigonometric model in
conjunction with indicator functions and minimize the squared error for each
time step in order to estimate the deterministic trend as well as values for different
seasons and days. Though they find that their model works well in practice, for
general cases it might be difficult always finding an appropriate model, especially
when there is no information about the trend, its seasonal cycles, and other
(deterministic or stochastic) influences. This is especially the case when the data
set covers only a short period of time.
When analyzing financial time series data, one could suggest the trigonometric
model and the order of the sinusoidal functions; in general it may be difficult to set
up such a model or reason, why this model and its estimated trend are appropriate
for the respective time series. This requires either a rigorous a priori analysis of the
series itself or further information about the external factors (i.e., the system the
time series is derived from) and their interaction. In addition to this, the estimation
can never be better than the model and to which accuracy it approximates the true
trend.
We note another probable critical issue when using indicator functions in
combination with least-squares estimation. First, using indicator functions confines
the model to jumps only, that is, slopes or similar phenomena cannot be captured by
that approach, as the indicator functions automatically introduce jumps in the trend
534 T. Meinl and E.W. Sun

component. Thus, indicator functions excel at modeling jumps but perform poorly
with other types of sudden changes. Second, for this approach, it is extremely
important to determine the location of the jump as exact as possible, as otherwise
the estimated trend in this area may be highly inaccurate.
It is therefore dubious, whether such parametric approaches are appropriate to
deal with economic and financial time series, though they will perform very well, if
their requirements are met.

18.5.2.2 Smoothing Splines

Smoothing splines, though also utilizing the least-squares methods, on the contrary
do not rely on a specific model assumption. Instead, they penalize the regression
spline in respect to its roughness, i.e.,

X
N ð
ðXt  mðtÞÞ2 þ o ðm00 ðxÞÞ dx:
2
min (18.19)
m
t¼1

It follows directly from this definition that for o ¼ 0 this yields us an interpo-
lation while for o ! 1 m will approximate a linear regression.
In practice, there emerges another difficulty: In many cases, the (optimal) choice
of o remains unclear. Although there exist several works that have established some
data-dependent rules for this, in many cases, when the assumptions about the noise do
not hold or the time series incorporates additional deterministic (e.g., cycles) or
stochastic components (e.g., outliers that are part of the system and not due to
measurement or other errors), the choice of the penalizing smoothing parameter is
a difficult task that has been and is still undergoing extensive research (see Morton
et al. (2009), Lee (2003), Hurvich et al. (1998), Cantoni and Ronchetti (2001),
Irizarry (2004)). We only mention particularly the cross-validation method which is
used to determine the optimal smoothing parameters (see Green and Silverman
(1994)). Furthermore, though o is eventually responsible for the degree of smooth-
ness (i.e., on which scale or level the trend shall be estimated), one can hardly neither
impose nor derive any additional meaning on or from this parameter.

18.5.2.3 The Hodrick-Prescott Filter

The Hodrick-Prescott (HP) filter was first introduced by Leser (1961) and later
became popular due to the advanced works of Hodrick and Prescott (1997). In order
to extract the trend t ¼ [t0, . . . , tN + 1] from a given time series X, this trend is
derived by solving

X
N
min ðXt  tt Þ2 þ oððttþ1  tt Þ  ðtt  tt1 ÞÞ2 : (18.20)
t
t¼1

We note that (besides from not using a spline basis for approximation)
this approach can be seen as a discretized formulation of the smoothing spline.
18 Methods of Denoising Financial Data 535

The smoothing parameter o plays the same role, while the penalized smoothness
measure is the discretized version of the second derivative. Thus, though several
authors propose explicit rules (of thumb) for choosing o (see, e.g., Ravn and Uhlig
(2002), Dermoune et al. (2008), Schlicht (2005)), some researchers like Harvey and
Trimbur (2008) also recognize that in some way this choice still remains kind of
arbitrary or problematic for many time series which cannot be associated in the same
time terms.

18.5.2.4 The Kalman Filter

Another sophisticated filter was developed by Kalman (1960). It is a state-space
system specifically designed to handle unobserved components models and can be
used to either filter past events from noise or forecast. Another good introduction to
the Kalman filter can be found in Welch and Bishop (1995) and a thorough
discussion in Harvey (1989).
The basic Kalman filter assumes that the state x 2 ℝn of underlying process in
a time series can be described by a linear stochastic difference equation:

xt ¼ Axt1 þ But1 þ wt1 , (18.21)

with A the state transition model that relates in conjunction with the (optional)
control input B, the respective input ut, and the noise component
wt the previous state to the next. In the above equation, A and B are ass-
umed to be constant but may also change over time. Of this model (i.e., the true
state xt), only

zk ¼ Hxk þ vk (18.22)

can be observed. Both noise components are assumed to be independent of one

another and to be distributed according to

w  N ð0; QÞ and v  N ð0; RÞ: (18.23)

With A, B, Q, and R assumed to be known, the filter predicts the next state xt
based on xt1 and also provides an estimate of the accuracy of the actual prediction.
Since its first development, the Kalman filter has become popular in many areas;
see, for example, Grewal et al. (2001). However, a serious drawback of this Kalman
procedure is that many real-world models do not fit the assumptions of the model, for
example, above requirement of a linear underlying system is often not met. Though
there exist extensions for nonlinear systems (see e.g., Julier and Uhlmann (1997)),
there still exists the problem that one or more of the required parameters are unknown.
While for many technical systems (e.g., car or missile tracking systems) based on
physical laws the state transition model A is exactly known, this becomes a difficult
issue in many other application areas, including finance. Additionally, as, e.g., Mohr
(2005) notes, the performance of the Kalman filter can be very sensitive to initial
conditions of the unobserved components and their variances, while at the same time it
536 T. Meinl and E.W. Sun

requires an elaborate procedure of model selection. He also notes that in macroeco-

nomic time series, the Kalman filter does not work with annual data. Therefore, we see
that while the Kalman filter unquestionably delivers excellent results in many areas
(and has also applied for financial time series as well, though not without critique), its
usage is not convenient for general cases we treat in this work, requiring more
assumptions and knowledge about the underlying model.

18.6 Summary

In this chapter we provided an overview of the different tasks of time series analysis
in general and the specific challenges of time series trend extraction. We pointed
out several methods and outlined their advantages and disadvantages, together with
their requirements. The reader should keep in mind the following key points:
• Linear filters provide a very smooth trend but fail to capture sudden changes.
• Nonlinear filters capture jumps extremely well but are very sensitive to higher
levels of noise.
• Advanced methods may provide good results depending on the scenario but
often rely on very specific assumptions that can yield misleading results, in case
they are not completely met.
• The specific task of jump detection and modeling is well understood but lacks
appropriate methods for more general changes like steep slopes.
• And the most important: No methods universally perform best for all tasks!

References
Astola, J., & Kuosmanen, P. (1997). Fundamentals of nonlinear filtering. Boca Raton: CRC Press.
Au, S., Duan, R., Hesar, S., & Jiang, W. (2010). A framework of irregularity enlightenment for
data pre-processing in data mining. Annals of Operations Research, 174(1), 47–66.
Bianchi, G., & Sorrentino, R. (2007). Electronic filter simulation & design. New York: McGraw-
Hill. ISBN 0071494677.
Cantoni, E., & Ronchetti, E. (2001). Resistant selection of the smoothing parameter for smoothing
splines. Statistics and Computing, 11(2), 141–146.
Carlstein, E., Mueller, H. G., & Siegmund, D. (1994). Change-point problems. Hayward,
CA: Institute of Mathematical Statistics.
Dacorogna, M. M. (2001). An introduction to high-frequency finance. San Diego, CA: Academic Press.
Dermoune, A., Djehiche, B., & Rahmania, N. (2008). Consistent estimators of the smoothing
parameter in the hodrick-prescott filter. Journal of the Japan Statistical Society, 38, 225–241.
Donoho, D. L., & Johnstone, I. M. (1994). Ideal spatial adaptation by wavelet shrinkage.
Biometrika, 81(3), 425–455.
Fan, J., & Wang, Y. (2007). Multi-scale jump and volatility analysis for high-frequency financial
data. Journal of the American Statistical Association, 102(480), 1349–1362.
Fan, J., & Yao, Q. (2003). Nonlinear time series: Nonparametric and parametric methods.
New York: Springer.
Gijbels, I., Lambert, A., & Qiu, P. (2007). Jump-preserving regression and smoothing using local
linear fitting: A compromise. Annals of the Institute of Statistical Mathematics, 59(2), 235–272.
18 Methods of Denoising Financial Data 537

Green, P. J., & Silverman, B. W. (1994). Nonparametric regression and generalized linear models:
A roughness penalty approach. New York: Chapman & Hall/CRC. ISBN 0412300400.
Grewal, M. S., Andrews, A. P., & Ebooks Corporation. (2001). Kalman filtering theory and
practice using MATLAB. New York: Wiley Online Library. ISBN 0471392545.
Harvey, A. C. (1989). Structural time series models and the Kalman filter forecasting. Cambridge:
Cambridge University Press.
Harvey, A., & Trimbur, T. (2008). Trend estimation and the Hodrick-Prescott filter. Journal of the
Japan Statistical Society, 38(1), 41–49.
Hodrick, R. J., & Prescott, E. C. (1997). Postwar US business cycles: An empirical investigation.
Journal of Money, Credit and Banking, 29(1), 1.
Hurvich, C. M., Simonoff, J. S., & Tsai, C. L. (1998). Smoothing parameter selection in nonpara-
metric regression using an improved Akaike information criterion. Journal of the Royal
Statistical Society, Series B: Statistical Methodology, 60(2), 271–293.
Irizarry, R. A. (2004). Choosing smoothness parameters for smoothing splines by minimizing an
estimate of risk (p. 30). Johns Hopkins University, Department of Biostatistics Working Papers.
Joo, J. H., & Qiu, P. (2009). Jump detection in a regression curve and its derivative.
Technometrics, 51(3), 289–305.
Julier, S. J., & Uhlmann, J. K. (1997). A new extension of the Kalman filter to nonlinear systems.
In International symposium on aerospace/defense sensing, simulation and controls
(Vol. 3, p. 26). Citeseer.
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic
Engineering, 82(1), 35–45.
Kim, C. S., & Marron, J. S. (2006). Sizer for jump detection. Journal of Nonparametric Statistics,
18(1), 13–20.
Lee, T. (2003). Smoothing parameter selection for smoothing splines: A simulation study.
Computational Statistics & Data Analysis, 42(1–2), 139–148.
Leser, C. E. V. (1961). A simple method of trend construction. Journal of the Royal Statistical
Society: Series B: Methodological, 23(1), 91–107.
Lucia, J. J., & Schwartz, E. S. (2002). Electricity prices and power derivatives: Evidence from the
nordic power exchange. Review of Derivatives Research, 5(1), 5–50.
Mallat, S. G., & Hwang, W. L. (1992). Singularity detection and processing with wavelets. IEEE
Transactions on Information Theory, 38(2), 617–643.
Meinl, T., & Sun, E. W. (2012). A nonlinear filtering algorithm based on wavelet transforms for
high-frequency financial data analysis. Studies in Nonlinear Dynamics and Econometrics, 16(3),
1–24.
Mohr, M. F. (2005). A trend-cycle (-season) filter. Working Paper Series European Central Bank, (499).
Morlet, J. (1983). Sampling theory and wave propagation. Issues in Acoustic Signal/Image
Processing and Recognition, 1, 233–261.
Morton, R., Kang, E. L., & Henderson, B. L. (2009). Smoothing splines for trend estimation and
prediction in time series. Environmetrics, 20(3), 249–259.
Peltonen, S., Gabbouj, M., & Astola, J. (2001). Nonlinear filter design: Methodologies and
challenges. In Proceedings of the 2nd International Symposium on Image and Signal
Processing and Analysis, 2001. ISPA 2001, (pp. 102–107).
Percival, D. B., & Walden, A. T. (2006). Wavelet methods for time series analysis. Cambridge:
Cambridge University Press.
Pollock, D. S. G. (2001). Methodology for trend estimation. Economic Modelling, 18(1), 75–96.
Qiu, P. (2003). A jump-preserving curve fitting procedure based on local piecewise-linear kernel
estimation. Journal of Nonparametric Statistics, 15(4), 437–453.
Qiu, P., & Yandell, B. (1998). A local polynomial jump-detection algorithm in nonparametric
regression. Technometrics, 40(2), 141–152.
Rachev, S. T. (2003). Handbook of heavy tailed distributions in finance. North-Holland.
Raimondo, M. (1998). Minimax estimation of sharp change points. Annals of Statistics, 26(4),
1379–1397.
538 T. Meinl and E.W. Sun

Raimondo, M., & Tajvidi, N. (2004). A peaks over threshold model for change-point detection by
wavelets. Statistica Sinica, 14(2), 395–412.
Ramsey, J. B. (1999). The contribution of wavelets to the analysis of economic and financial data.
Philosophical Transactions: Mathematical, Physical and Engineering Sciences, 357(1760),
2593–2606.
Ramsey, J. B. (2002). Wavelets in economics and finance: Past and future. Studies in Nonlinear
Dynamics & Econometrics, 6(3), 1–27.
Ravn, M. O., & Uhlig, H. (2002). On adjusting the Hodrick-Prescott filter for the frequency of
observations. Review of Economics and Statistics, 84(2), 371–376.
Schlicht, E. (2005). Estimating the smoothing parameter in the so-called Hodrick-Prescott filter.
Journal of the Japan Statistical Society, 35(1), 99–119.
Seifert, J., & Uhrig-Homburg, M. (2007). Modelling jumps in electricity prices: Theory and
empirical evidence. Review of Derivatives Research, 10(1), 59–85.
Sun, E. W., & Meinl, T. (2012). A new wavelet-based denoising algorithm for high-frequency
financial data mining. European Journal of Operational Research, 217(3), 589–599.
Sun, J., & Qiu, P. (2007). Jump detection in regression surfaces using both first-order and second-
order derivatives. Journal of Computational and Graphical Statistics, 16(2), 289–311.
Sun, W., Rachev, S. T., Fabozzi, F., & Stoyan S. (2008). Multivariate skewed student’s t copula in
analysis of nonlinear and asymmetric dependence in German equity market. Studies in
Nonlinear Dynamics and Econometrics, 12(2).
Tsay, R. S. (2005). Analysis of financial time series. Hoboken: Wiley.
Wang, Y. (1995). Jump and sharp cusp detection by wavelets. Biometrika, 82(2), 385–397.
Wang, Y. (1998). Change curve estimation via wavelets. Journal of the American Statistical
Association, 93(441), 163.
Welch, G., & Bishop, G. (1995). An introduction to the Kalman filter. Chapel Hill: University of
North Carolina at Chapel Hill.
Yuan, L., & Zhongjie, X. (2000). The wavelet detection of the jump and cusp points of a regression
function. Acta Mathematicae Applicatae Sinica (English Series), 16(3), 283–291.
Analysis of Financial Time Series Using
Wavelet Methods 19
Philippe Masset

Contents
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
19.2 Frequency-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
19.2.1 Spectral Analysis: Some Basics and an Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
19.2.2 Filtering Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
19.3 Scale-by-Scale Decomposition with Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
19.3.1 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
19.3.2 Implementation and Practical Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
19.3.3 Illustration: Home Price Indices for Different US Cities . . . . . . . . . . . . . . . . . . . . . 563
19.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571

Abstract
This chapter presents a set of tools, which allow gathering information about the
frequency components of a time series. In a first step, we discuss spectral
analysis and filtering methods. Spectral analysis can be used to identify and to
quantify the different frequency components of a data series. Filters permit to
capture specific components (e.g., trends, cycles, seasonalities) of the original
time series. Both spectral analysis and standard filtering methods have two main
drawbacks: (i) they impose strong restrictions regarding the possible processes
underlying the dynamics of the series (e.g., stationarity) and (ii) they lead to
a pure frequency-domain representation of the data, i.e., all information from the
time-domain representation is lost in the operation.
In a second step, we introduce wavelets, which are relatively new tools in
economics and finance. They take their roots from filtering methods and Fourier
analysis, but overcome most of the limitations of these two methods. Their principal

P. Masset
Ecole Hôtelière de Lausanne, Le-Chalet-à-Gobet, Lausanne 25, Switzerland
e-mail: philippe.masset@ehl.ch

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 539
DOI 10.1007/978-1-4614-7750-1_19,
# Springer Science+Business Media New York 2015
540 P. Masset

advantages derive from (i) combined information from both time domain and
frequency domain and (ii) their flexibility as they do not make strong assumptions
concerning the data-generating process for the series under investigation.

Keywords
Filtering methods • Spectral analysis • Fourier transform • Wavelet filter •
Continuous wavelet transform • Discrete wavelet transform • Multiresolution
analysis • Scale-by-scale decomposition • Analysis of variance • Case-Shiller
home price indices

19.1 Introduction

The purpose of this chapter is to present a set of methods and tools belonging to the
so-called frequency-domain analysis and to explain why and how they can be used
to enhance the more conventional time-domain analysis. In essence, time-domain
analysis studies the evolution of an economic variable with respect to time, whereas
frequency-domain analysis shows at which frequencies the variable is active. We
focus on concepts rather than technicalities and illustrate each method with exam-
ples and applications using real datasets.
The usual time-domain approach aims at studying the temporal properties
of a financial or economic variable, whose realizations are recorded at
a predetermined frequency. This approach does not convey any information
regarding the frequency components of a variable. Thus, it makes the implicit
assumption that the relevant frequency to study the behavior of the variable
matches with its sampling frequency. An issue arises, however, if the variable
realizations depend (in a possibly complicate manner) on several frequency
components rather than just one. In such a case, the time-domain approach
will not be able to efficiently process the information contained in the original
data series.
In this chapter, we start by discussing methods belonging to the frequency-
domain analysis. These tools are very appealing to study economic variables that
exhibit a cyclical behavior and/or are affected by seasonal effects (e.g., GDP,
unemployment). Spectral analysis and Fourier transforms can be used to quantify
the importance of the various frequency components of the variable under investi-
gation. In particular, they allow inferring information about the length of a cycle
(e.g., business cycle) or a phase (e.g., expansion or recession). The presence of such
patterns also imposes the use of appropriate methods when it comes to model the
dynamics of the variable. Filtering methods have proven useful in this context.
Notably, filters may serve to remove specific frequency components from the
original data series.
In a second step, we introduce wavelets. During the last two decades, wavelets have
become increasingly popular in scientific applications such as signal processing and
functional analysis. More recently, these methods have also started to be applied to
financial datasets. They are very attractive as they possess the unique ability to provide
19 Analysis of Financial Time Series Using Wavelet Methods 541

a complete representation of a data series from both the time and frequency perspec-
tives simultaneously. Hence, they allow breaking down the activity on the market into
different frequency components and to study the dynamics of each of these compo-
nents separately. They do not suffer from some of the limitations of standard
frequency-domain methods and can be employed to study a financial variable, whose
evolution through time is dictated by the interaction of a variety of different frequency
components. These components may also behave according to nontrivial (noncyclical)
dynamics – e.g., regime shifts, jumps, and long-term trends.
For instance, the presence of heterogeneous agents with different trading
horizons may generate very complex patterns in the time series of stock prices
(see M€ uller et al. 1995; Lynch and Zumbach 2003). This heterogeneity may in
particular induce long memory in stock return volatility. In such a case, studying the
properties of a time series and trying to model it from the perspective of a single
frequency can be misleading. Much information will be lost because of the naive
and implicit aggregation of the different frequency components into a single com-
ponent. Furthermore, as these components may interact in a complicated manner
and may be time varying or even nonstationary, standard methods like Fourier
analysis are not appropriate. Therefore, one has to resort to more flexible filtering
methods like wavelets.
The remaining of this chapter is structured as follows. In Sect. 19.2, we
discuss spectral analysis and filtering methods. Section 19.3 is devoted to the
presentation of wavelets, Sect. 19.3.1 explains the relevant theoretical back-
ground, Sect. 19.3.2 discusses the implementation of these methods, and
Sect. 19.3.3 presents a complete case study. Section 19.4 offers a short
conclusion.

19.2 Frequency-Domain Analysis

19.2.1 Spectral Analysis: Some Basics and an Example

Studying the properties of an economic variable in the time domain is done

through time series analysis. Similarly, the purpose of spectral analysis is to
study the properties of an economic variable over the frequency spectrum, i.e.,
in the frequency domain. In particular, the estimation of the population spectrum
or the so-called power spectrum (also known as the energy-density spectrum)
aims at describing how the variance of the variable under investigation can be
split into a variety of frequency components. The subject has been extensively
researched during the previous 40 years (see Iacobucci (2003) for a short litera-
ture review). Our discussion is based primarily on Hamilton (1994) and Gençay
et al. (2002).

19.2.1.1 Fourier Transform

The basic idea of spectral analysis is to reexpress a covariance-stationary process
x(t) as a new sequence X(f), which determines the importance of each frequency
542 P. Masset

component f in the dynamics of the original series. This is achieved using the
discrete version of the Fourier transform1
X1
X ðf Þ ¼ t¼1
xðtÞei2pft , (19.1)

where f denotes the frequency at which X(f) is evaluated. In order to gain a deeper
insight into this decomposition, one may think about the De Moivre’s (Euler’s)
theorem, which allows to write e–i2pft as
ei2pft ¼ cosð2pf tÞ  i sinð2pftÞ: (19.2)

Hence application of formula (19.1) is similar to projecting the original signal

x(t) onto a set of sinusoidal functions, each corresponding to a particular frequency
component. Furthermore, one can use the inverse Fourier transform to recover the
original signal x(t) from X(f):
Z
1 p
xðtÞ ¼ Xðf Þei2pft df : (19.3)
2p p

Equation 19.3 shows that X(f) determines how much of each frequency compo-
nent is needed to synthesize the original signal x(t).

19.2.1.2 Population Spectrum and Sample Periodogram

Following Hamilton (1994), we define the population spectrum of x(t) as

1 X1
sx ðoÞ ¼ g eioj ,
j¼1 j
(19.4)
2p

where gj is the jth autocovariance of x(t)2; o ¼ 2pf is a real scalar, which is related to the
frequency f ¼ 1/t at which the spectrum is evaluated; and t is the period length of one
cycle at frequency f.3 One may notice that the right part of Eq. 19.4 is indeed the
discrete-time Fourier transform of the autocovariance series. There is also a close link
between this expression and the autocovariance generating function, which is defined by
X1
gx ð z Þ ¼ j¼1 j
g zj , (19.5)

where z denotes a complex scalar. This implies that one can easily recover the
autocovariance generating function from the spectrum. In the same spirit, an

1
The discrete version of the Fourier transform is used because the time series is recorded at
discrete-time intervals.
2
The jth autocovariance of x(t) is given by gj ¼ E[(x(t)  m)(x(t  j)  m)], where m denotes the
expected value of x(t).
3
As an example, let us consider an economic variable, whose evolution is fully determined by the
state of the economy. A complete business cycle lasts on average 36 months and therefore
f ¼ 1/36 months.
19 Analysis of Financial Time Series Using Wavelet Methods 543

application of the inverse discrete-time Fourier transform allows a direct estimation

of the autocovariances from the population spectrum.
In practice, the sample periodogram ^s x ðoÞ can be estimated with either nonpara-
metric or parametric approaches. We briefly describe these two approaches hereafter.
The first approach is nonparametric because it infers the spectrum from a sample of
realizations of the variable x without trying to assign an explicit structure to the data-
generating process underlying its evolution. The estimation of the sample
periodogram is straightforward as it is directly related to the squared magnitude of
the discrete-time Fourier transform │X(f)│ of the time series x(t)

1 1
^s x ðoÞ ¼ j X ðf Þj 2 , (19.6)
2p T

where T is the length of the time series x(t). │X(f)│2 is also known as the power
spectrum of x(t). This approach is usually called the “periodogram.” As noted in
Hamilton (1994), its accuracy seems questionable as the confidence interval for the
estimated spectrum is typically very broad. Furthermore, the variance of the
periodogram does not tend to zero as the length of the data series tends to infinity.
This implies that the periodogram is not a consistent estimator of the population
spectrum. Therefore, modified versions of the periodogram have been put forward.
For instance, smoothed periodogram estimates have been suggested as a way to
reduce the noise of the original estimator and to improve its accuracy. The idea
underlying this approach is that sx(o) will be close to sx(l) when o is close to l.
This suggests that sx(o) might be estimated with a weighted average of the values of
sx(o) for values of l in a neighborhood around o, where the weights depend on the
distance between o and l (Hamilton 1994). The weights are typically determined
by a kernel weighting function. Welch’s method (Welch 1967) and Childers (1978)
constitute another alternative based on a simple idea: instead of estimating a single
periodogram for the complete sample, one divides the original sample into sub-
samples, estimates the periodogram for each subsample, and computes the average
periodogram over all subsamples.
The second approach is based on some parameterization of the data-
generating process of x(t). Methods belonging to this category are close in spirit
to the population spectrum, i.e., to a direct application of Eq. 19.4. Typically
some specification based on an ARMA (autoregressive moving average) repre-
sentation is chosen to represent the temporal dynamics of the variable. The
model is then calibrated, i.e., the ARMA coefficients are estimated from the
realizations of the process x(t). These estimated coefficients are employed to
calculate the spectrum. As long as the autocovariances are reasonably well
estimated, the results would also be reasonably close to the true values.
A detailed discussion of the various parametric methods (e.g., the covariance,
Yule-Walker and Burg methods) is beyond the scope of this introduction, but
parametric methods are particularly effective when the length of the observed
sample is short. This is due to their ability to distinguish the noise from the
information contained in the data.
544 P. Masset

19.2.1.3 Example
We now turn to the discussion of a simple example. We consider a time series,
which has the following dynamic:
   
2pt 2pt
xðtÞ ¼ a  cos þ b  sin þ eðtÞ,
21 63

where e(t) is a random term that follows a normal distribution with mean zero and
unit variance. One may observe that the process is driven by two cyclical compo-
nents, which repeat themselves, respectively, each 21 and 63 units of time.
The full line in Fig. 19.1 shows the first 100 (simulated) realizations of x(t); the
dotted lines are for the cos and sin functions. At first glance, it seems difficult to
distinguish the realizations of x(t) from a purely random process. Figure 19.2
reports the autocorrelations (left panel) and partial autocorrelations (right panel)
of x(t) (upper panel) and of the cos and sin components (bottom panel). Again, it
remains difficult, when looking at this figure, to gather conclusive evidence
concerning the appropriate model specification for x(t).
On the other hand, results from the Fourier analysis, reported in Fig. 19.3, clearly
show that two cyclical components drive the evolution of x(t) and repeat themselves
around each 21 and 63 units of time. This demonstrates the effectiveness of Fourier
methods for the study of processes featuring cyclical components.

19.2.1.4 Illustration: Home Prices in New York City

We now illustrate how spectral methods can be applied to real economic data series.
We consider the Case-Shiller home price index for the city of New York. The
dataset covers the period January 1987 to December 2011 on a monthly basis. The
upper panel of Fig. 19.4 shows the evolution of the index level over this time period,
while the lower panel reports the time series of index returns. Results from the
Dickey-Fuller test cannot reject the null hypothesis that the index level series is
nonstationary. Application of the Fourier transform requires the series under study
to be stationary. We therefore study the spectral properties of the index using the
time series of returns rather than the levels themselves.
We also estimate the autocorrelations and partial autocorrelations of the index
returns up to 48 lags (i.e., 4 years of observations). These are reported in Fig. 19.5.
The structure of both the autocorrelations and the partial autocorrelations indicates
that the index returns are significantly autocorrelated and it also suggests a cyclical
(or seasonal) behavior of the returns. This observation is in line with previous
results from the literature (see Kuo 1996; Gu 2002). In order to gain more insight
into the presence of such patterns, we compute the power spectrum of the series
using parametric and nonparametric methods displayed in Fig. 19.6. The estimated
power spectra returned by the two nonparametric methods (periodogram and
Welch) are much noisier than the spectra obtained from the parametric methods
(Yule-Walker and Burg). The Welch method also seems to result in an
oversmoothed estimate of the power spectrum as compared to the other estimates
19 Analysis of Financial Time Series Using Wavelet Methods 545

Evolution of xt
4

−1

−2

−3 xt cos sin

−4
0 10 20 30 40 50 60 70 80 90 100

Fig. 19.1 Sample path of x(t)

ACF: xt PACF: xt

0.4 0.4

0.2 0.2

0 0

−0.2 −0.2
0 5 10 15 20 25 0 5 10 15 20 25

ACF: cos ACF: sin PACF: cos PACF: sin

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0

−0.2 −0.2 −0.2 −0.2

0 10 20 0 10 20 0 10 20 0 10 20

Fig. 19.2 Autocorrelations and partial autocorrelations of x(t). The upper panel reports the
autocorrelations (left panel) and partial autocorrelations (right panel) of x(t). The lower panel
shows similar statistics for the cosinus and sinus functions

(e.g., periodogram estimates). On the other hand, the difference between the
Yule-Walker and the Burg methods is minimal. Nevertheless, the key message
remains remarkably similar: strong seasonalities affect home prices with
a frequency of recurrence of 12 months.
546 P. Masset

Power Spectrum of xt

700 Periodogram
Population Spectrum
600

500

400

300

200

100

0
0 10 20 30 40 50 60 70 80 90 100
Periods

Fig. 19.3 Spectral analysis. The periodogram has been estimated directly from the realizations of
x(t). The population spectrum has been estimated on the basis of the theoretical autocovariances of x(t)

19.2.2 Filtering Methods

A filter is a mathematical operator that serves to convert an original time series x(t)
into another time series y(t). The filter is applied by convolution of the original
series x(t) with a coefficient vector w:
X1
yðtÞ ¼ ðw  xÞðtÞ ¼ k¼1
wðkÞxðt  kÞ, (19.7)

The purpose of this operation is to explicitly identify and to extract

certain components from x(t). In the present context, one may want to remove
from the original time series some particular features (e.g., trends, business
cycles, seasonalities, or noise) that are associated with specific frequency
components.

19.2.2.1 Frequency Response Function

Filters in the time domain can be characterized on the basis of their impulse-
response function, which traces the impact of a one-time unit impulse in x(t)
on subsequent values of y(t). Similarly, in the frequency domain, the analysis
of the frequency response function (or transfer function) of a filter tells us
which frequency components the filter captures from the original series. The
frequency response function is defined as the Fourier transform of the filter
coefficients
X1
H ðf Þ ¼ k¼1
wðkÞei2pfk : (19.8)
19 Analysis of Financial Time Series Using Wavelet Methods 547

Index level
200

150

100

1990 1995 2000 2005 2010

Returns [in %]
2

−2
1990 1995 2000 2005 2010

Fig. 19.4 Case-Shiller home price index for the city of New York. Price levels and returns are
reported in the upper and lower panel respectively.

ACF PACF
1 1

0.5 0.5

0 0

−0.5 −0.5
0 10 20 30 40 0 10 20 30 40

Fig. 19.5 Autocorrelations and partial autocorrelation of the returns on the New York home
price index

The frequency response, H(f), can be further split into two parts:

H ðf Þ ¼ Gðf Þeiyðf Þ , (19.9)

where G(f) is the gain function, eiy(f) the phase function, and y the phase angle
(or equivalently the argument of H( f )). The gain function is the magnitude of the
frequency response, i.e., G(f) ¼ jH( f )j. If the application of the filter on x(t) results in
a phase shift, i.e., if the peaks and lows of x(t) and y(t) have a different timing, the
phase angle y will be different from zero. The use of uncentered moving average
548 P. Masset

Power Spectrum
3

Periodogram
2.5
Welch
Yule-Walker
2 Burg

1.5

0.5

0
0 5 10 15 20

Fig. 19.6 Spectral analysis of the return series. Comparison between parametric (Burg and
Yule-Walker) and non-parametric (Periodogram and Welch) methods

filters leads to this (often) undesirable feature because turning points will be recorded
earlier in the original series than in the filtered series. On the other hand, centered
(symmetric) moving averages have y( f ) ¼ 0; hence, there is no phase shift for this
class of filters. For instance, the frequency response of a two-period uncentered
moving average filter with coefficients w(k) ¼ 0.5 for k ¼ 0, 1 is4
X1
H ðf Þ ¼ k¼0
0:5ei2pfk

¼ 0:5 þ 0:5ei2pf
 
¼ 0:5 eipf þ eipf eipf
¼ cosðpf Þeipf

This result shows the existence of a phase shift as the phase angle is y( f ) ¼ pf.5
Hence, the turning points from the original series will be shifted to the right in the
filtered series. On the other hand, a filter with coefficients w(k) ¼ 1/3 for k ¼ 1, 0, 1
has a zero phase angle.
Based on their gain functions, filters can be categorized as follows:
• High-pass filters should be able to capture the high-frequency components of
a signal, i.e., the value of their gain function G(f) should equal one for frequen-
cies f close or equal to 1/2.

4
See also Gençay et al. (2002) who uses a similar example.
 
Im½H ðf Þ
In full generality, the phase angle can be computed as yðf Þ = arctan
5
Re½Hðf Þ , where Im[H(f)] and
Re[H(f)] are, respectively, the imaginary part and the real part of H(F).
19 Analysis of Financial Time Series Using Wavelet Methods 549

• Low-pass filters should be able to capture the low-frequency components of

a signal, i.e., G(f) ¼ 1 for f close or equal to 0.
• Band-pass filters should be able to capture a range of frequency components of
a signal, i.e., G(f) ¼ 1 for flo < f < fhi.
• All-pass filters capture all the frequency components of a signal, i.e., G(f) ¼ 1 for
8f. Such filters leave the frequency components of the original signal unaltered.
Filters are commonly used in economics. The Hodrick and Prescott (1997) filter
is probably the best known. It has a structure and a gain function which enable it to
capture business cycle components.

19.2.2.2 Example
A basic example of a high-pass filter is a filter that takes the difference between two
adjacent values from the original series; its coefficients are whi ¼ [0.5,  0.5]. Similarly,
the most simple low-pass filter is p a ffiffiffi 2-period pmoving
ffiffiffi average; in this case
wlo ¼ [0.5, 0.5]. In wavelet theory, whi = 2 and wlo = 2 form the Haar wavelet family.
In this case, the low-pass filter wlo is basically an averaging filter, while the high-pass
filter whi is a differencing filter. The gain functions for these two filters are reported in
Fig. 19.7.

19.2.2.3 Illustration
The full line in the left panels of Fig. 19.8 shows the monthly (unadjusted) returns on the
Case-Shiller New York home price index from 1987 to 2011. We add to the top panel
the output series resulting from the application of both a centered and an uncentered
(causal) 3-period moving average on the original data. The three-filter coefficients have
a value of 1/3. This implies that the output series of the centered moving average at time
t is basically the average return from months t – 1 to t + 1. Similarly, the uncentered
moving returns the average return from t – 2 to t. It is apparent from the figure that the
uncentered moving average leads to a phase shift of 1 month. The bottom panel shows
the outputs of a 7-period moving average. The filter coefficients are equal to 1/7. Again,
we consider both centered and uncentered filters. The use of an uncentered moving
average leads to a phase shift of 3 months as compared to the centered moving average.
The right part of the figure reports the gain functions for the 3-period and the
7-period moving averages. The gain function is similar for both the uncentered and
the centered moving average. The only element that distinguishes these two filters
is indeed their phase function. One may notice that (i) both filters are low-pass
filters, and (ii) the longer filter captures more efficiently the low-frequency
components of the original signal than the short filter.

19.3 Scale-by-Scale Decomposition with Wavelets

To a large extent, wavelets can be seen as a natural extension to spectral and Fourier
analysis as (i) wavelets do not suffer the weaknesses of Fourier analysis, and
(ii) wavelets provide a more complete decomposition of the original time series
than Fourier analysis does.
550 P. Masset

Haar filter
1

whi
0.8 wlo

0.6
Gain

0.4

0.2

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency

Fig. 19.7 Gain functions of the high-pass (differencing) and low-pass (averaging) Haar filters

MA(3) G(f) - MA(3)

1
2
0.8
1
0.6
Gain

0
0.4
−1
0.2
−2
0
2000 0 0.1 0.2 0.3 0.4 0.5
Time Frequency

MA(7) G(f) - MA(7)

1
2
0.8
1
0.6
Gain

0
0.4
−1
0.2
−2
0
2000 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
Ch. CPI Centered Uncentered

Fig. 19.8 Two moving averages and their frequency responses. Left panels show the outputs from
a 3-period (upper panel) and a 7-period (lower panel) moving average filters applied on the returns on
the Case-Shiller New York home price index. Right panels report the corresponding gain functions
19 Analysis of Financial Time Series Using Wavelet Methods 551

There are some problems with spectral methods and Fourier transforms. Notably,
these methods require the data under investigation to be stationary. This is often not the
case in economics and finance. In particular, volatility is known to exhibit complicated
patterns like jumps, clustering, and long memory. Furthermore, the frequency decom-
position delivered by Fourier analysis only makes sense if the importance of the various
frequency components remains stable over the sample period. Ex ante, there is good
reason to expect this assumption not to hold for a variety of economic and financial
variables. For instance, volatility changes are likely to exhibit a different frequency
spectrum when trading activity is intense than when the market is quiet. The short-time
Fourier transform (which is also known as the Gabor or windowed Fourier transform)
has been suggested to overcome these difficulties. The idea is to split the sample into
subsamples and to compute the Fourier transform on these subsamples. Hence, this
extension achieves a better trade-off between the time and the frequency representation
of the original data. Nevertheless, this provides at best a partial solution to the
aforementioned issues as the strong restrictions regarding the possible data-generating
process over each subsample are maintained.
Wavelets do not make any of these assumptions. Furthermore, wavelets pro-
vide a complete decomposition of the original series, which is located both in time
and in frequency. From a mathematical viewpoint, a wavelet is a function, which
enables to split a given signal into several components, each reflecting the
evolution trough time of the signal at a particular frequency. Wavelet analysis
has originally been used in signal processing (e.g., image processing and data
compression). Its applications to economics and finance are relatively recent.
Nevertheless, the range of application of wavelets is potentially wide: denoising
and seasonality filtering, decorrelation and estimation of fractionally integrated
models, identification of regime shift and jumps, robust estimation of the covari-
ance and correlation between two variables at different time scales, etc.
From a physicist perspective, but with application to time series analysis,
Percival and Walden (2000) and Struzik (2001) provide a mathematically rigorous
and exhaustive introduction to wavelets. Struzik (2001) particularly emphasizes the
unique ability of nonparametric methods (like wavelets) to let the data speak by
themselves. Thus, such methods avoid making misleading interpretations of the
coefficients obtained from the calibration of misspecified models. Gençay
et al. (2002) discuss the use of wavelets for specific purposes in economics and
finance and adopt a more intuitive approach (with many illustrations and examples),
while Ramsey (2002) surveys the most important properties of wavelets and
discusses their fields of application in both economics and finance. Crowley
(2007) proposes a genuine guide to wavelets for economists. His article can be
considered as a complete and easily understandable toolkit, precisely explaining in
which circumstances to use wavelets and how to proceed. Schleicher (2002) is
a complementary reference to those already named; Schleicher focuses on some
mathematical concepts underlying the use of wavelets and discusses them in details
using examples.
This short literature review focuses only on textbook-style references. There are
however quite a large amount of economic and financial articles that have employed
552 P. Masset

wavelets for empirical purposes. Ramsey6 and Gençay7 and their respective coau-
thors can be considered as the pioneers of the use of wavelets in economics and
finance. Other recent contributions include Nielsen and Frederiksen (2005),
Vuorenmaa (2005), Oswiecimka et al. (2006), Elder and Jin (2007), Fan
et al. (2007), Fernandez and Lucey (2007), Rua and Nunes (2009), Manchaldore
et al. (2010), Curci and Corsi (2012), and Hafner (2012).

19.3.1 Theoretical Background

19.3.1.1 What Is a Wavelet?

As its name suggests, a wavelet is a small wave. In the present context, the term
“small” essentially means that the wave grows and decays in a limited time frame.
Figure 19.9 illustrates this notion by contrasting the values taken by a simple wavelet
function (the Morlet function)8 and the values of the sin function, which can be
considered as a sort of “large” wave. In order to clarify the notion of small wave, we
start by introducing a function, which is called the mother wavelet and is denoted by
c(t). This function is defined on the real axis and must satisfy two conditions:
Z 1
cðtÞdt ¼ 0: (19.10)
1
Z 1
jcðtÞj2 dt ¼ 1: (19.11)
1

Taken together, these conditions imply (i) that at least some coefficients of the
wavelet function must be different from zero and (ii) that these departures from zero
must cancel out. Clearly the sin function does not meet these two requirements.
A vast variety of functions meets conditions (19.10) and (19.11). Nevertheless, these
conditions are very general and not sufficient for many practical purposes. Therefore,
one has to impose additional conditions in order to run a specific analysis with
wavelets. One of these conditions is the so-called admissibility condition, which
states that a wavelet function is admissible if its Fourier transform
Z 1
Cðf Þ ¼ cðtÞei2pft dt, (19.12)
1
is such that
Z 1
jCðf Þj2
CC ¼ df satisfies 0 < CC < 1: (19.13)
0 f
These conditions allow reconstructing a function from its continuous wavelet
transform (see Percival and Walden (2000) for more details).

6
See Ramsey et al. (1995), Ramsey and Zhang (1997), and Ramsey (1999).
7
See Gençay et al. (2003), Gençay et al. (2010), Gençay and Fan (2010), and Gençay and
Gradojevic (2011).
8
The Morlet wavelet is actually similar to a sin curve modulated by a Gaussian envelope.
19 Analysis of Financial Time Series Using Wavelet Methods 553

Wave or wavelet?
1
Morlet Wavelet
0.8 sin(x)

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
−20 −15 −10 −5 0 5 10 15 20

Fig. 19.9 The Morlet wavelet and the sin function

19.3.1.2 The Continuous Wavelet Transform (CWT)

As a starting point, we discuss the CWT. The CWT primarily aims at quantifying
the change in a function at a particular frequency and at a particular point in time. In
order to be able to achieve this, the mother wavelet c(t) is dilated and translated:

1 t  u
cu,s ðtÞ ¼ pffiffi c , (19.14)
s s

where u and s are the location and scale parameters. The term p1ffis ensures that the
norm of cu,s(t) is equal to one. The CWT, W(u, s), which is a function of the two
parameters u and s, is then obtained by projecting the original function x(t) onto the
mother wavelet cu,s(t):
Z 1
W ðu; sÞ¼ xðtÞcu,s ðtÞdt: (19.15)
1

To assess the variations of the function on a large scale (i.e., at a low frequency),
a large value for s will be chosen and vice versa. By applying the CWT for
a continuum of location and scale parameters to a function, one is able to
554 P. Masset

decompose the function under study into elementary components. This is

particularly interesting for studying a function with a complicated structure,
because this procedure allows extracting a set of “basic” components that have
a simpler structure than the original function. By “synthesizing” W(u, s), it is also
possible to reconstruct the original function x(t) (see Gençay et al. (2002) for more
details).
In empirical applications, several difficulties with the CWT occur. First, it is
computationally impossible to analyze a signal using all wavelet coefficients. CWT
is thus more suitable for studying functions than signals or (economics) time series.
Second, as noted by Gençay et al. (2002), W(u, s) is a function of two parameters
and as such it contains a high amount of redundant information. We therefore turn
to the discussion of the discrete wavelet transform (DWT).

19.3.1.3 The Discrete Wavelet Transform (DWT)

The core difference between the CWT and the DWT is that the latter does not use
all translated and dilated versions of the mother wavelet to decompose the original
signal (Gençay et al. 2003). The idea is to select u and s such that the information
contained in the signal can be summarized in a minimum of wavelet coefficients.
This objective is achieved by setting

s ¼ 2j and u ¼ k2j ,

Where j and k are integers representing the set of discrete translations and discrete
dilatations. Gençay et al. (2002) refer to this procedure as the critical sampling of
the CWT. This implies that the wavelet transform of the original function or
signal is calculated only at dyadic scales, i.e., at scales 2j. A further implication is
that for a time series with T observations, the largest number of scales for the
DWT is equal to the integer J such that J ¼ blog2(T)c ¼ blog(T)/log(2)c. It is not
possible to directly apply the DWT if the length of the original series is not dyadic
(i.e., if J < log2(T) < J + 1). In such case, one has either to remove some
observations or to “complete” the original series in order to have a series of
dyadic length. Several methods exist to deal with this kind of boundary problems
(see Sect. 19.3.2).
The DWT is based on two discrete wavelet filters, which are called the mother
wavelet hl ¼ (hl, . . .,hL–1) and the father wavelet g1 ¼ (g1,. . .,gL–1). The mother
wavelet is characterized by three basic properties:

XL1 XL1 XL1

l¼0
hl ¼ 0, l¼0
hl 2 ¼ 1, and l¼0
hl hlþ2n ¼ 0 for all integers n 6¼ 0 (19.16)

These three properties ensure that (i) the mother wavelet is associated with
a difference operator, (ii) the wavelet transform preserve the variance of the original
data, and (iii) a multiresolution analysis can be performed on a finite variance data
series. The first property implies that the mother wavelet (also called “differencing
19 Analysis of Financial Time Series Using Wavelet Methods 555

function”) is a high-pass filter as it measures the deviations from the smooth

components. On the other hand, the father wavelet (“scaling function”) aims at
capturing long-scale (i.e., low-frequency) components of the series and generates
the so-called scaling coefficients.9
The mother and father wavelets must respect the following conditions:
XL1
l¼0
hl ¼ 0: (19.17)

XL1
l¼0
gl ¼ 1: (19.18)

The application of both the mother and the father wavelets allows separating the
low-frequency components of a time series from its high-frequency components.
Furthermore, a band-pass filter can be constructed by recursively applying
a succession of low-pass and high-pass filters.
Let’s assume that we have observed a sample of size T of some random variable
x(t), {x(1), x(2), . . ., x(T)}. The wavelet and scaling coefficients at the first level of
decomposition are obtained by convolution of the data series with the mother and
the father wavelets:
XL1 XL1
w1 ðtÞ ¼ l¼0
hl xðt0 Þ and v1 ðtÞ ¼ l¼0
gl xðt0 Þ (19.19)

where t ¼ 0,1, . . .,T/2  1 and t0 the time subscript of x is defined as t 0 ¼ 2t + 1 

l mod T. The modulus operator is employed to deal with boundary conditions.10
It ensures that the time subscript of x stays always positive. If, for some particular
values of t and l, the expression 2t + 1 – l becomes negative, the application of the
modulus operator returns t 0 ¼ 2t + 1  l + T. Thus, we are implicitly assuming that x
can be regarded as periodic. Alternative methods to deal with boundary conditions
are discussed thereafter. w1(t) and v1(t) are, respectively, the wavelet and the scaling
coefficients at the first scale. Hence, w1(t) corresponds to the vector containing the
components of x recorded at the highest frequency. One may notice that the operation
returns two series of coefficients that have length T/2. To continue the frequency-by-
frequency decomposition of the original signal, one typically resorts to what is known
as the pyramid algorithm.

19.3.1.4 Pyramid Algorithm

After having applied the mother and father wavelets on the original data series, one
has a series of high-frequency components and a series of lower-frequency

9
The low-pass filter can be directly obtained from the high-pass filter using the quadrature mirror
relationship; see Percival and Walden (2000, p. 75).
10
For two integer a and b, a modulus b is basically the remainder after dividing a by b, i.e., a mod
b ¼ a–c  b with c ¼ ba/bc.
556 P. Masset

j=1 j=2 j=3 j=4 […] j=J

x(t) v1(t) v2(t) v3(t) v4(t) vJ(t)

w1(t) w2(t) w3(t) w4(t) wJ(t)

Fig. 19.10 Flowchart of the pyramid algorithm

components. The idea of the pyramid algorithm is to further decompose the

(low-frequency) scaling coefficients v1(t) into high- and low-frequency components:
XL1 XL1
w2 ðtÞ ¼ l¼0
hl v1 ðt0 Þ and v2 ðtÞ ¼ l¼0
gl v1 ðt0 Þ, (19.20)

where t ¼ 0,1, . . .,T/4  1 and t0 ¼ 2t + 1  l mod T. After two steps, the

decomposition looks like w ¼ [w1 w2 v2]. One can then apply the
pyramid algorithm again and again up to scale J ¼ blog2(T)c to finally obtain
w ¼ [w1 w2 . . . wj vj]. Figure 19.10 summarizes these steps. One may also apply the
algorithm up to scale Jp < J only. This is known as the partial DWT.

19.3.1.5 The Maximal Overlap Discrete Wavelet Transform (MODWT)

The standard DWT suffers from three drawbacks. First, it requires a series with
a dyadic length. Second, DWT is not shift invariant, i.e., if one shifts the series one
period to the right, the multiresolution coefficients will be different. Third, it may
introduce phase shifts in the wavelet coefficients: peaks or troughs in the original
series may not be correctly aligned with similar events in the multiresolution
analysis. To overcome these problems, the MODWT has been proposed. This
wavelet transform can handle any sample size, it has an increased resolution at
coarser scales (as compared to the DWT), and it is invariant to translation. It also
delivers a more asymptotically efficient wavelet variance than the DWT.11
The main difference between the DWT and the MODWT lies in the fact that the
MODWT considers all integer translations, i.e., u ¼ k. This means that the
MODWT keeps at each frequency a complete resolution of the series. Whatever
the scale considered, the length of the wavelet and scaling coefficient vectors will
be equal to the length of the original series. The wavelet and scaling coefficients at
the first level of decomposition are obtained as follows:


XL1  
XL1 
0
w 1 ðt Þ ¼ l¼0
h l x ð t Þ and v 1 ð tÞ ¼ g xðt0 Þ,
l¼0 l
(19.21)

where t ¼ 0,1,. . .T and t0 ¼ t – l mod T. As for the DWT, the MODWT coefficients
 
for scales j > 1 can be obtained using the pyramid algorithm. For instance, w j and v j
are calculated as

11
See Crowley (2007) for more details about the properties of MODWT.
19 Analysis of Financial Time Series Using Wavelet Methods 557


XL1   
XL1  
w j ðt Þ ¼ h v ðt0 Þ and v j ðtÞ ¼
l¼0 l j1
g v ðt0 Þ,
l¼0 l j1
(19.22)
where t0 ¼ 2j–1 l mod T.
Using matrix notation, we can conveniently calculate

the wavelet and scaling
coefficients up to scale J. We first define a matrix W that is composed of J + 1
sub-matrices, each of them T x T:
2 3
W1
6 7
6 W 7
 6 27
W¼ 6 ⋮ 7 (19.23)
6 7
4 WJ 5

VJ

Each W j has the following structure:
2    3
h 0 =2j=2 0 0  0 h L1 =2j=2  h 1 =2j=2
6  j=2  7
6 h 1 =2 h 0 =2j=2 0  0 0  ⋮ 7
6    7
6 ⋮ h 1 =2j=2 h 0 =2j=2   h L1 =2 7j=2 7
6 0 0
6  7
6 h L1 =2j=2 ⋮ h 1 =2j=2  0 0  0 7
6  7
6 h L1 =2j=2 ⋮   7
 6 0 0 0 0 7
Wj ¼ 6  7
6 0 0 h L1 =2j=2  0 0  0 7
6 7
6 ⋮ ⋮ ⋮ ⋱ ⋮ ⋮  ⋮ 7
6  7
6 0 0 0  h 0 =2j=2 0  0 7
6   7
6  h 1 =2j=2 h 0 =2j=2  7
6 0 0 0 0 7
4 0 0 0  ⋮ ⋮  0 5
  
0 0 0  h L1 =2j=2 h L2 =2j=2  h 0 =2 j=2

(19.24)
 
V J has a similar structure as W J but it contains the coefficients associated to the
father wavelet instead of the mother wavelet.
We can now directly calculate all wavelet and scaling coefficients via
 
w ¼ W x, (19.25)

where w is a vector made up of J + 1 length T vectors of wavelet and scaling
       
T
coefficients, w, . . . , w j and vj ; i.e., w = w 1 w 2 w J    v J .

19.3.1.6 Multiresolution Analysis (MRA)

Multiresolution analysis can be used to reconstruct the original time series x from
 
the wavelet and scaling coefficients, vj and wj . In order to achieve this, one has
 
to apply the inverse MODWT on vj and wj , j ¼ 1,. . ., J.12

12
Our presentation of the multiresolution analysis is restricted to the case of the MODWT.
Nevertheless, a very similar procedure exists for the DWT; see Percival and Walden (2000).
558 P. Masset

  
W is an orthonormal matrix, as such W TW = 1. Hence, if we multiply both side

of Eq. 19.25 by W T , we get

   
W T w ¼ W T Wx ¼ x: (19.26)

h    i   
  
T
As W T = W 1 W 2    W J V J and w = w 1 w2 w J   v J , we can further
rearrange Eq. 19.26 and show that

XJ  T  T
x¼ j¼1
W j wj þ V J vJ : (19.27)

 T  T
Setting Dj = W j w j and Sj = V j v j , we can reconstruct the original time
series as

x ¼ D 1 þ . . . þ D J þ SJ : (19.28)

This “reconstruction” is known as multiresolution analysis (MRA). The ele-

ments of Sj are related to the scaling coefficients at the maximal scale and therefore
represent the smooth components of x. The elements of Dj are the detail (or rough)
coefficients of x at scale j.
On the basis of formula (19.28), one may also think of a way to compute an
approximation or a smooth representation of the original data. This can be achieved
by considering the scaling coefficients and the wavelet coefficients from scale Js
(Js < J) to J only, i.e.:

xS ¼ DS þ . . . þ DJ þ SJ : (19.29)

Equation 19.29 can be used, for instance, to filter out noise or seasonalities from
a time series. In image processing, Eq. 19.29 serves for data compression.
Equation 19.27 has been specifically derived for the MODWT but similar results
are available for the DWT (see Percival and Walden 2000).

19.3.1.7 Analysis of Variance

On the basis of the wavelet and scaling coefficients, it is also possible to decompose
the variance into different frequency components. There are some slight differences
between the variance decomposition for the DWT and for the MODWT. We will
therefore first present the main results for the DWT and then discuss their extension
to the MODWT.
19 Analysis of Financial Time Series Using Wavelet Methods 559

Using the wavelet and scaling coefficients of the discrete wavelet transform, it is
possible to decompose the energy of the original series on a scale-by-scale basis:
XT1 XJ XT=2J 1
kx k2 ¼ t¼0
xðtÞ2 ¼ j¼1 t¼0
wj ðtÞ2 þ vJ ðtÞ2 , (19.30)

where kxk2 denotes the energy of x. The wavelet coefficients capture the deviations
of x from its long-run mean at the different frequency resolutions. Therefore, at
scale j ¼ J ¼ log2(T), the last remaining scaling coefficient is equal to the sample
mean of x,
Eð x Þ ¼ v J : (19.31)

On this basis, we can express the variance of x as

  XJ   XJ  
V ðxÞ ¼ E x2  EðxÞ2 ¼ j¼1
E wj 2 ¼ j¼1
V wj , (19.32)

where V(wj) denotes the variance of the wavelet coefficients at scale j.

If we consider the wavelet and scaling coefficients obtained from a partial DWT,
the variance of x can be expressed as
XJ p    
V ðxÞ ¼ j¼1
V wj þ V vJ p : (19.33)

The variance of the scaling coefficients has to be taken into account because vJp
incorporates deviations of x from its mean at scales Jp < j < J.13
An alternative way to decompose the energy of x is based on the smooth and
detail coefficients of the MRA. As above, kxk2 can be computed as the sum of the
energy of the smooth and detail coefficients. This approach is, however, valid only
for the DWT (See Gençay et al. 2002).
It is important to note that some of the wavelet coefficients involved in Eq. 19.32
are affected by boundary conditions. One should remove the corresponding wavelet
coefficients in order to get an unbiased estimator of the wavelet variance:
" #
XJ 1 X 2Tj 1
^ ðxÞ ¼
V 0 wj ðtÞ
2
, (19.34)
j¼1 ^j
2lj T t¼Lj

where lj is the scale that is associated to the frequency interval [1/2j+1 1/2j].
0
Lj ¼ d(L  2) (1  2 j)e is the number of DWT coefficients computed using the
boundaries. Hence, T^ j = T=2j  L0 is the number of coefficients unaffected by the
j
boundary.

13
One may notice that the variance of the scaling coefficient at scale J is 0 as vJ is a scalar (the
sample mean of x).
560 P. Masset

We now turn to the analysis of variance in the context of the MODWT.

Equation 19.32 remains perfectly valid. The MODWT keeps the same number of
coefficients at each stage of the wavelet transform. The way of dealing with boundary
conditions must therefore be adapted. From the detail coefficients of a partial
MODWT of order Jp < log2(T), the wavelet variance can be estimated as follows:
" #
XJ p 1 XT1  2
V^ ðxÞ ¼ w ðtÞ (19.35)
j¼1 T ^j t¼Lj 1 j

where Lj ¼ (2j  1) (L  1) + 1 is the number of scale lj wavelet coefficients, which

are affected by boundary conditions. This number also corresponds to the length of
the wavelet filter at scale lj. T ^ j = T  Lj þ 1 is thus the number of wavelet
coefficients unaffected by the boundary.

19.3.2 Implementation and Practical Issues

19.3.2.1 Choice of a Wavelet Filter

Many different wavelet filters exist with each of them being particularly suitable
for specific purposes of analysis. Wavelet filters differ in their properties and in
their ability to match with the features of the time series under study. Further-
more, when it comes to implement a discrete wavelet transform, one also has to
decide about the filter length. Because of boundary conditions, longer filters are
well adapted for long time series. The simplest filter is the Haar wavelet, which is
basically a difference and average filter of length two. In finance, most researchers
have worked either with Daubechies (denoted as “D”) or with Least-Asymmetric
(“LA”) filters of length 4–8. Elder and Jin (2007) and Nielsen and Frederiksen
(2005) employ D(4) filters. Gençay et al. (2003, 2010) suggest that the LA
(8) wavelet (i.e., a Least-Asymmetric filter of length 8) is a good choice for
analyzing financial time series, while Subbotin (2008) uses a LA(4) wavelet.
Crowley (2007) argues that the impact of choosing another wavelet filter has
a rather limited impact on the distribution of the variance of the time series across
the scales.
Depending on the purpose of the analysis, it might be appealing to select
a wavelet filter which satisfies one or more of the following properties:
• Symmetry: symmetric filters are appealing as they ensure that there will be no
phase shift in the output series. Unfortunately, most wavelets are not
symmetric. An exception is the Haar wavelet. The requirement of a symmetric
wavelet is, however, less essential if a MODWT is used as it ensures that the
original series and its filter coefficients will be aligned.
• Orthogonality: this property refers to the fact that the wavelet and the scaling
coefficients contain different information. This is an important feature as it
allows for the wavelet decomposition to preserve the energy (variance) of the
original series (Crowley 2007). Daubechies and Least-Asymmetric wavelets
19 Analysis of Financial Time Series Using Wavelet Methods 561

meet this requirement, that is, their scaling and wavelet coefficients are orthog-
onal by construction.
• Smoothness: The degree of smoothness is measured by the number of continu-
ous derivatives of the basis function. As such, the Haar wavelet is the least
smooth wavelet. The choice of a more or less smooth filter depends essentially
on the data series to be represented. If the original time series is very smooth,
then one will opt for a smooth wavelet. For instance, the Haar wavelet is
appropriate for the analysis of a pure jump process.
• Number of vanishing moments: The number of vanishing moments of the
wavelet function has a direct implication on the ability of the wavelet
to account for the behavior of the signal. That is, if a signal has
a polynomial structure or if it can be approximated by a polynomial of
order q, then the wavelet transform will be able to properly capture this
polynomial structure only if it has q vanishing moments. For instance,
Daubechies wavelets have a number of vanishing moments which is half
the length of the filter. Thus, the Haar and D(8) have, respectively, 1 and
4 vanishing moments.
The last two properties depend not only on the wavelet filter but also on its
length. In fact, the most crucial point is probably not to choose the “right” filter but
to choose a filter with an appropriate length. Increasing filter length allows better
fitting the data. Unfortunately, this also renders the influence of boundary condi-
tions more severe. Hence, a trade-off has to be found.
Problems due to boundary conditions arise in two situations. The first case
concerns the DWT. To use the DWT, one requires a time series with a dyadic
length. If the series does not meet this requirement, i.e., if its length N is such that
2j < N < 2j+1, one has the choice between removing observations until N ¼ 2j and
completing the series such that N ¼ 2 j+1. Removing data might be the best
alternative but it leads to a loss of information.
The second case concerns both the DWT and the MODWT. The wavelet filter
has to be applied on all observations, including observations recorded at the
beginning (t ¼ 1). A problem arises because the convolution operator requires
that L – 1 observations are available before t. In this case, removing data is
useless. Therefore, one has to complete the data series. One solution is to pad
each end of the series with zeros; this technique is known as “zero padding.” An
alternative is to use the fit from a polynomial model to replace nonexisting data at
each end of the series (“polynomial approximation”). One may also complete
each end of the series either by mirroring the last observations (“mirror” or
“reflection”) or by taking the values observed at the beginning of the other end of
the series (“circular”). The choice depends on the data considered. For instance,
if working on stock returns, the use of a “mirror” seems to be the most suitable
approach as it accounts for the presence of volatility clustering. Moreover, after
the multiresolution decomposition of the original signal, one may obviously
discard the coefficients that are affected due to their proximity to the boundaries
of the series.
562 P. Masset

Instead of selecting a priori specific wavelet function, one may also use the
so-called optimal transforms. The idea is to choose the wavelet function that
minimizes a loss function (e.g., the entropy cost function; see Crowley 2007).

19.3.2.2 Examples of Wavelet Filters and Their Gain Functions

Figure 19.11 shows the coefficients of the Haar, D(4), D(8), and LA(8) wavelets for
level j ¼ 4. One may observe the very simple structure of the Haar wavelet. When
comparing the latter with the D(4) and D(8) filters, it becomes clear that the longer
the filter, the smoother it is. The LA(8) looks less asymmetric than the D(8).
Nevertheless it is still far from being symmetric.
Studying the frequency response of these filters permits to assess their ability to
capture the different frequency components. Figure 19.12 displays the gain function
for each wavelet filter at scales 1–4. It is evident from the figure that the longer
filters (D(8) and LA(8)) have better frequency localization. The gain functions of
the D(8) and LA(8) are similar. In order to make this statement clearer, we contrast
the gain functions of the Haar, D(4) and D(8) wavelets at scale j ¼ 5 in Fig. 19.13.
The D(8) captures much better the components corresponding to frequencies
between 1/2j and 1/2j+1.

19.3.2.3 Example
Let’s consider a variable x, whose dynamics is primarily driven by an AR(1) process
and three cyclical components:
X5  
2pt
yðtÞ ¼ 0:90yðt  1Þ þ s¼3
5 cos þ eðtÞ, (19.36)
s

e(t) is an i.i.d. Gaussian process with mean zero and unit variance and
t ¼ 1, . . . ,10,000. In the absence of seasonalities and noise, the autocorrelation
function of y should take value 0.90k at lag k. Wavelets can be used to remove the
impact of both noise and seasonalities. From Eq. 19.36, one may notice that the
cyclical components have a period length of 3–5 periods. Hence, they have an
impact on frequencies between 1/5 and 1/3.
At scale j, the wavelet detail Dj captures frequencies 1/2 j+1 f  1/2j and
the wavelet smooth Sj captures frequencies f < 1/2 j+1. If we use a level
2 multiresolution analysis, the wavelet smooth S2 will thus capture the compo-
nents of the time series, which have a frequency f < 1/8. This means that S2 will
take into account changes in y that are associated with a period length of at least
8 units of time. Therefore, S2 should keep the AR(1) dynamics of y, while
removing its cyclical behavior and noise.
Figure 19.14 reports the autocorrelation coefficients for the original time series y,
for the theoretical AR(1) process, and for the wavelet smooth S2. In order to assess the
impact of choosing a different wavelet filter, we use both the Haar and the LA
(8) wavelets. The results are very similar even if for short lags the LA(8) seems to
provide some improvements over the Haar. All in one, this example demonstrates the
ability of a wavelet filter to deal with a complex cyclical structure and with noise.
19 Analysis of Financial Time Series Using Wavelet Methods 563

Haar D4

0.1
0.05

0.05

0
0

−0.05 −0.05

−20 0 20 −60 −40 −20 0 20

D8 LA8
0.1
0.1

0.05 0.05

0 0

−0.05
−0.05

−100 −50 0 −100 −50 0

Fig. 19.11 Mother wavelet filters. The Haar, D(4), D(8), and LA(8) filters at scale j ¼ 4

19.3.3 Illustration: Home Price Indices for Different US Cities

19.3.3.1 Descriptive Statistics

In this section, we study the evolution of home prices in 12 US cities from January
1987 to February 2012 (302 months).14 We employ the Case-Shiller home price
indices. The data has been gathered from the Standard & Poors website. Pj(t)
denotes the price index level for the jth city in period t which has been standardized
so that it has a value of 100 in the first month of the sample, i.e., Pj(t ¼ 0) ¼ (8j). We
compute the index returns as

r j ðtÞ ¼ log Pj ðtÞ  log Pj ðt  1Þ :

Table 19.1 reports some descriptive statistics for the returns on each of the
12 home price indices. There are huge disparities in performance. The largest price
increases are to be found in Portland (+216 % since 1987 but 31 % since the

14
Los Angeles (LA), San Francisco (SF), Denver (De), Washington (Wa), Miami (Mi), Chicago
(Chi), Boston (Bos), Las Vegas (LV), New York (NY), Portland (Po), Charlotte (Cha), and
Cleveland (Cl).
564 P. Masset

G1(f) G2(f) G3(f) G4(f)

1 1 1 1
Haar

0.5 0.5 0.5 0.5

0 0 0 0
0 0.5 0 0.5 0 0.5 0 0.5
G1(f) G2(f) G3(f) G4(f)
1 1 1 1
D4

0.5 0.5 0.5 0.5

0 0 0 0
0 0.5 0 0.5 0 0.5 0 0.5
G1(f) G2(f) G3(f) G4(f)
1 1 1 1
D8

0.5 0.5 0.5 0.5

0 0 0 0
0 0.5 0 0.5 0 0.5 0 0.5
G1(f) G2(f) G3(f) G4(f)
1 1 1 1
LA8

0.5 0.5 0.5 0.5

0 0 0 0
0 0.5 0 0.5 0 0.5 0 0.5

Fig. 19.12 Gain function for the Haar, D(4), D(8), and LA(8) wavelets

beginning of the subprime/financial crisis), Washington (+174 % and 34 %

respectively), and Los Angeles (+169 % and 42 %). Las Vegas (+35 %
and 62 %), Charlotte (+71 % and 20 %), and Cleveland (+76 % and 24 %)
are the worst performers. The volatility of home price changes has been much larger
in cities like San Francisco, Los Angeles, Las Vegas, and Miami than in northeast
cities (plus Portland and Denver). This higher volatility is, to a large extent, the
result from the severe price downturn in these four cities during the last 6 years of
the sample. It is worth mentioning that some cities have been less affected by the
home price crash than others (e.g., Denver and Boston). The skewness is always
negative and the kurtosis is generally smaller than three.

19.3.3.2 Autocorrelations
Figure 19.15 shows the autocorrelations (up to 24 lags, i.e., 2 years) of the index
returns. The full lines are for the original series, while the dotted lines show the
autocorrelations computed from the smooth coefficients obtained using the MRA
from a partial MODWT. We employ a LA(8) filter with Jp ¼ 3. Hence, the wavelet
smooth (S3) should capture the frequency components that are associated with
a period length of at least 16 months and should therefore be free of seasonal
effects. In order to deal with the boundary conditions, we use the reflection method.
Some indices are very affected by seasonal effects (Portland, Boston, Denver,
Charlotte, and Cleveland) while other indices are much less affected
19 Analysis of Financial Time Series Using Wavelet Methods 565

G5(f)
1
Haar
0.9 D(4)
D(8)
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 19.13 Gain function for the Haar, D(4), and D(8) at scale j ¼ 5

1
0.8 AR(1) y S2 Haar S2 LA(8)

0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
0 5 10 15 20 25 30 35 40 45 50
Lags

Fig. 19.14 Autocorrelations estimated before and after having removed some specific frequency
components of the original time series
 566

Table 19.1 Descriptive statistics for the 12 home price indices

LA SF De Wa Mi Chi Bos LV NY Po Cha Cl
Averag 3.9 % 3.9 % 3.5 % 4.0 % 2.8 % 2.7 % 2.9 % 1.2 % 3.0 % 4.6 % 2.1 % 2.3 %
Std. 4.4 % 4.9 % 2.6 % 3.5 % 4.0 % 3.6 % 3.2 % 4.8 % 2.8 % 3.1 % 2.1 % 3.1 %
Skew. (0.4) (0.7) (0.6) (0.2) (1.2) (1.1) (0.1) (0.3) (0.2) (0.8) (0.7) (0.8)
Kurt. 1.0 1.8 1.1 0.7 2.9 3.1 (0.0) 4.9 (0.1) 2.0 2.1 6.9
Last 159.5 124.6 121.8 175.7 139.5 105.4 146.2 89.9 159.6 129.6 108.1 94.1
High 273.9 218.4 140.3 251.1 280.9 168.6 182.5 234.8 215.8 186.5 135.9 123.5
Bottom 159.2 117.7 120.2 165.9 137.0 105.4 145.8 89.9 159.6 129.6 108.1 94.1
We report the average return, the standard deviation (both annualized and in %), the skewness, and the kurtosis of home price index returns. We also report the
index level in the last month of the sample as well as the highest level reached before the crisis and the lowest level reached during the crisis
P. Masset
19 Analysis of Financial Time Series Using Wavelet Methods 567

LA SF Denver Washington
1 1 1 1

0 0 0 0

−1 −1 −1 −1
0 10 20 0 10 20 0 10 20 0 10 20
Miami Chicago Boston LV
1 1 1 1

0 0 0 0

−1 −1 −1 −1
0 10 20 0 10 20 0 10 20 0 10 20
NY Portland Charlotte Cleveland
1 1 1 1

0 0 0 0

−1 −1 −1 −1
0 10 20 0 10 20 0 10 20 0 10 20

Original series Wavelet smooth

Fig. 19.15 Autocorrelations of rj(t). This figure shows the autocorrelations for both the original
returns series and the wavelet smooth series for each home price index

(Miami, Las Vegas, and Los Angeles). In general the indices that display the least
significant seasonal patterns are also those that have been the most affected by the
recent crisis. This observation may suggest that the (quasi) absence of these patterns
is the result of the predominant role of speculation on price changes in these cities.
One may again observe the ability of wavelets to remove seasonal patterns. The
autocorrelations estimated from the wavelet smooth show the long-run temporal
dynamics of home prices. In contrast to financial markets, whose evolution is
almost unpredictable, home prices are strongly autocorrelated. The autocorrelation
remains positive even after 2 years.

19.3.3.3 Variance and Cross-Correlations

Next, we analyze the variance of each index returns series and the correlations
between the different indices at a variety of frequencies. We employ a partial
MODWT with Jp ¼ 5. Hence, at the largest scale, the wavelet coefficients contain
information regarding price changes over horizon of 32–64 months. Similarly, the
scaling coefficients capture the evolution for periods longer than 64 months.
Figure 19.16 shows the distribution of variance across the different frequencies.
From Fig. 19.16, one may notice that most of the variance is due to frequency
components associated with scales 3 and 6. On the other hand, variance at scales
1 and 2 is low and may be due to noise. Scales 4–5 are not related to important
frequency components, neither to seasonal patterns (scale j ¼ 3 nor to business
568 P. Masset

Distribution of variance across scales

0.7

0.6
Percentage of variance at scale j

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6
Scale j

Fig. 19.16 Distribution of the wavelet variance across scales. Each bar corresponds to the
variance recorded at a specific scale j and for a particular city. Cities are ordered as follows
(from left to right): Miami, Las Vegas, Los Angeles, Washington, New York, San Francisco,
Chicago, Portland, Boston, Denver, Charlotte, Cleveland

cycle components (scales j > 5). Thus, it comes as no surprise that they do not have
much information content.
Scale 3 corresponds to periods of 8–16 months, and as such the variance observed
at this scale reflects the importance of seasonal patterns in the dynamics of the various
home price indices. As before, we observe that seasonal effects have a very limited
impact on prices in Miami; they are also much less important in cities like Las Vegas,
Los Angeles, Washington, and New York than in Boston, Denver, Charlotte, and
Cleveland. Interestingly the ordering is reversed when considering the variance at
scale 6. That is, the index returns series on which seasonalities have a weak impact
exhibit the largest percentage of long-term volatility.
Figure 19.17 reports the correlation between the various city indices at different
scales. The largest correlations are observed at scales larger than 2. In particular at
scale 3, the correlations are very significant. This is because seasonalities affect
most indices simultaneously. One may notice that indices 5 and 8 (Miami and Las
Vegas) show less correlation with the other indices. At scales 4 and 5, the correla-
tions become highly significant. This demonstrates that the series tend to behave
very similarly in the long run.
An extension to this correlation analysis is to study how many (economic)
factors are important to explain the structure of the correlation matrix at each
19 Analysis of Financial Time Series Using Wavelet Methods 569

W1 W2 W3
1
2 2 2
4 4 4
0.9
6 6 6
8 8 8
0.8
10 10 10
12 12 12
0.7
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12

W4 W5 0.6
V6
2 2
2 0.5
4 4 4
6 6 6
0.4
8 8 8
10 10 10
0.3
12 12 12
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12

Fig. 19.17 Correlations between home price indices at different frequencies. Panels denoted by
W1 to W5 show the correlations between the wavelet coefficients at scales j ¼ 1, . . ., 5, while panel
denoted V5 show the correlations between the scaling coefficients at scale j ¼ 5. The 12 cities are
order as follows: Los Angeles, San Francisco, Denver, Washington, Miami, Chicago, Boston,
Las Vegas, New York, Portland, Charlotte, Cleveland

scale. In order to address this question, one may resort to random matrix theory
(RMT). A good introduction is provided by Bouchaud and Potters (2004) (See also
Sharifi et al. (2004) for a literature review). Here, we concentrate on the main
premises of RMT. That is, this approach should help to (i) assess if the correlation
coefficients have a genuine information content or if they are merely due to the
noise inherent in the data and (ii) estimate the number of factors that are necessary
to “explain” the correlation matrix. This is done by comparing the eigenvalues of
the empirical correlation matrix with those from a theoretical distribution.
Let’s consider a dataset with N time series of length T. We assume that the
theoretical random matrix belongs to the ensemble of Wishart matrices. On this
basis, we can derive the theoretical distribution of the eigenvalues of the correlation
matrix. Under the null of pure randomness, the eigenvalues must be confined within
the bounds15:
sffiffiffi! sffiffiffi!
1 1 1 1
lmin ¼ 1þ 2 and lmax ¼ 1þ þ2 , (19.37)
q q q q

where q = NT . If the correlation matrix is random, then the probability that any of
its eigenvalues lies outside the bounds defined by [lmin, lmax] is zero. Hence, the

15
See Sharifi et al. (2004) and Kwapien et al. (2007).
570 P. Masset

W1 W2
0.02 0.02

0.01 0.01

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2 2.5
-λ -
W3 W4
0.03
0.02
0.02
0.01
0.01

0 0
0 0.5 1 1.5 2 2.5 3 0 1 2 3 4

W5 V5
0.08
0.2
0.06
0.04 0.1
0.02
0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6

Empirical distribution of λ λmax 95%-percentile 99%-percentile

Fig. 19.18 Distribution of the empirical eigenvalues of the correlation matrix

presence of eigenvalues larger than the upper bound can be taken as an evidence
that the structure of the correlation matrix is not due to chance, i.e., that there are
deviations from RMT. Furthermore, the number of such eigenvalues can be
interpreted as corresponding to the number of factors underlying the evolution of
the N time series. A potential problem with the RMT approach is that it
requires N ! 1 and T ! 1. In our case, these conditions are evidently not fulfilled
as T ¼ 302 and N ¼ 12. To account for this, we also estimate lmax from the empirical
distribution of the correlation matrix eigenvalues. To this aim, we resample (without
replacement) the original time series of returns and then apply the MODWT on the
resampled data and calculate the correlation matrix and its eigenvalues at each scale
of the MODWT. This procedure is done 100,000 times. In Fig. 19.18, we report lmax
as computed on the basis of Eq. 19.37 and the empirical values of lmax obtained from
our simulations. In the latter case, we consider two different values for lmax, which
correspond to the 5 %- and, respectively, 1 %-percentile of the empirical cumulative
distribution of l. The theoretical (RMT) maximum eigenvalue is always much lower
than its empirical counterparts. This is probably due to the small sample size.
In Table 19.2, we report the eigenvalues of the correlation matrix for each
scale of the MODWT. Comparing the eigenvalues lk, k ¼ 1,. . .,12 with ^l max
demonstrates that there is a single factor underlying the evolution of home
19 Analysis of Financial Time Series Using Wavelet Methods 571

Table 19.2 Eigenvalues of the correlation matrix at each scale of the MODWT
w1 w2 w3 w4 w5 v5
l1 0.54 0.29 0.04 0.04 0.01 0.00
l2 0.60 0.30 0.07 0.07 0.02 0.00
l3 0.69 0.33 0.08 0.13 0.08 0.00
l4 0.78 0.46 0.12 0.15 0.13 0.01
l5 0.84 0.48 0.14 0.20 0.16 0.01
l6 0.86 0.61 0.18 0.23 0.18 0.05
l7 0.99 0.70 0.22 0.25 0.34 0.15
l8 1.02 0.77 0.29 0.44 0.50 0.16
l9 1.16 0.92 0.38 0.62 0.66 0.37
l10 1.38 1.06 0.77 0.82 0.69 1.10
l11 1.49 1.37 1.11 1.33 1.38 1.18
l12 1.64a 4.70a 8.60a 7.72a 7.84a 8.98a
lmax 1.30 1.31 1.31 1.32 1.34 1.34
^l max ð95 %Þ 1.43 1.53 1.77 2.16 2.79 3.25
^
l max ð99 %Þ 1.53 1.66 1.96 2.47 3.25 3.89
We report the 12 eigenvalues of the correlation matrix at each scale j as well as the theoretical
(RMT) maximum eigenvalue (lmax in the table) and the empirical 95 %- and 99 %-percentiles
of the empirical (bootstrapped) cumulative distribution of eigenvalues at scale j (^
l max ð95%Þ and
^ max ð99%Þ). a denotes significance at the 95 %-level
l

prices. At each scale, the eigenvalue that is attached to this factor is significant at
the 99 % level. This unique factor can be interpreted as a sort of national-wide
home price index.

19.4 Conclusion

This chapter discusses spectral and wavelet methods. It aims at being an easy-to-
follow introduction and it is structured around conceptual and practical explana-
tions. It also offers many supporting examples and illustrations. In particular, the
last section provides a detailed case study, which analyzes the evolution of home
prices in the USA over the last 20 years using wavelet methodology.

References
Bouchaud, J.-P., & Potters, M. (2004). Theory of financial risk and derivative pricing: From
statistical physics to risk management. Cambridge: Cambridge University Press.
Childers, D. (1978). Modern spectrum analysis. New York: IEEE Press.
Crowley, P. (2007). A guide to wavelets for economists. Journal of Economics Surveys,
21, 207–264.
Curci, G., & Corsi, F. (2012). Discrete sine transform for multi-scale realized volatility measures.
Quantitative Finance, 12, 263–279.
572 P. Masset

Elder, J., & Jin, H. (2007). Long memory in commodity futures volatility: A wavelet perspective.
The Journal of Futures Markets, 27, 411–437.
Fan, J., Fan, Y., & Yazhen, W. (2007). Multiscale jump and volatility analysis for high-frequency
financial data. Journal of American Statistical Association, 102, 618–631.
Fernandez, V., & Lucey, B. (2007). Portfolio management under sudden changes in volatility and
heterogeneous investment horizons. Physica A: Statistical Mechanics and its Applications,
375, 612–624.
Gençay, R., & Fan, Y. (2010). Unit root tests with wavelets. Econometric Theory, 26, 1305–1331.
Gençay, R., & Gradojevic, N. (2011). Errors-in-variables estimation with wavelets. Journal of
Statistical Computation and Simulation, 81, 1545–1564.
Gençay, R., Selçuk, F., & Whitcher, B. (2002). An introduction to wavelets and other filtering
methods in finance and economics. San Diego: Academic Press.
Gençay, R., Selçuk, F., & Whitcher, B. (2003). Systematic risk and timescales. Quantitative
Finance, 3, 108–116.
Gençay, R., Gradojevic, N., Selçuk, F., & Whitcher, B. (2010). Asymmetry of information flow
between volatilities across time scales. Quantitative Finance, 10, 895–915.
Gu, K. (2002). The predictability of house prices. Journal of Real Estate Research, 24, 213–234.
Hafner, C. (2012). Cross-correlating wavelet coefficients with applications to high-frequency
financial time series. Journal of Applied Statistics, 39, 1363–1379.
Hamilton, J. (1994). Time series analysis. Princeton.
Hodrick, R., & Prescott, E. (1997). Postwar U.S. business cycles: An empirical investigation.
Journal of Money, Credit, and Banking, 29, 1–16.
Iacobucci, A. (2003). Spectral analysis for economic time series. Documents de Travail de
l’OFCE.
Kuo, C. (1996). Autocorrelation and seasonality in the real estate market. Journal of Real Estate
Finance and Economics, 12, 139–162.
Kwapien, J., Oswiecimka, P., & Drozdz, S. (2007). Empirics versus RMT in financial cross-
correlations. Acta Physica Polonica B, 38, 4027–4039.
Lynch, P., & Zumbach, G. (2003). Market heterogeneities and the causal structure of volatility.
Quantitative Finance, 3, 320–331.
Manchaldore, J., Palit, I., & Soloviev, O. (2010). Wavelet decomposition for intra-day volume
dynamics. Quantitative Finance, 10, 917–930.
M€uller, U., Dacorogna, M., Davé, R., Pictet, O., Olsen, R., & Ward, J. (1995). Fractals and
intrinsic time – A challenge to econometricians (Working Papers). Olsen and Associates.
Z€urich
Nielsen, M., & Frederiksen, P. (2005). Finite sample comparison of parametric, semiparametric,
and wavelet estimators of fractional integration. Econometric Reviews, 24, 405–443.
Oswiecimka, P., Kwapien, J., & Drozdz, S. (2006). Wavelet versus detrended fluctuation analysis
of multifractal structures. Physical Review E, 74, 016103.
Percival, D., & Walden, A. (2000). Wavelet methods for time series analysis. Cambridge:
Cambridge University Press.
Ramsey, J. B. (1999). The contribution of wavelets to the analysis of economic and financial data.
Philosophical Transactions of the Royal Society of London A, 357, 2593–2606.
Ramsey, J. B. (2002). Wavelets in economics and finance: Past and future. Studies in Nonlinear
Dynamics & Econometrics, 3, 1–29.
Ramsey, J. B., & Zhang, Z. (1997). The analysis of foreign exchange data using waveform
dictionaries. Journal of Empirical Finance, 4, 341–372.
Ramsey, J. B., Zaslavsky, G., & Usikov, D. (1995). An analysis of U. S. stock price behavior using
wavelets. Fractals, 3, 377–389.
Rua, A., & Nunes, L. (2009). International comovement of stock market returns: A wavelet
analysis. Journal of Empirical Finance, 16, 632–639.
Schleicher, C. (2002). An introduction to wavelets for economists (Working Paper). Bank of
Canada.
19 Analysis of Financial Time Series Using Wavelet Methods 573

Sharifi, S., Crane, M., Shamaie, A., & Ruskin, H. (2004). Random matrix theory for portfolio
optimization: A stability approach. Physica A, 335, 629–643.
Struzik, Z. (2001). Wavelet methods in financial time-series processing. Physica A, 296, 307–319.
Subbotin, A. (2008). A multi-horizon scale for volatility. University of Paris-1: CES Working
Paper 2008.20.
Vuorenmaa, T. (2005). A wavelet analysis of scaling laws and long-memory in stock market
volatility, Bank of Finland Research Discussion Paper.
Welch, P. (1967). The use of fast Fourier transforms for the estimation of power spectra: A method
based on time averaging over short modified periodograms. IEEE Transactions on Audio and
Electroacoustics, 15, 70–73.
Composite Goodness-of-Fit Tests for
Left-Truncated Loss Samples 20
Anna Chernobai, Svetlozar T. Rachev, and Frank J. Fabozzi

Contents
20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
20.2 Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
20.3 EDF Statistics for Left-Truncated Loss Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580
20.3.1 Supremum Class Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
20.3.2 Quadratic Class Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
20.3.3 Cramér-von Mises Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
20.4 “Upper Tail” Anderson-Darling Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
20.4.1 Supremum Class “Upper Tail” Anderson-Darling Statistic . . . . . . . . . . . . . . . . . . 584
20.4.2 Quadratic Class “Upper Tail” Anderson-Darling Statistic . . . . . . . . . . . . . . . . . . . 585
20.5 Application to Loss Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
20.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
Derivation of AD2* Computing Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
Derivation of W2* Computing Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
Derivation of AD2up Computing Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594
Derivation of AD2* up Computing Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596

A. Chernobai (*)
Department of Finance, M.J. Whitman School of Management, Syracuse University, Syracuse,
NY, USA
e-mail: annac@syr.edu
S.T. Rachev
Department of Applied Mathematics and Statistics, College of Business, Stony Brook University,
SUNY, Stony Brook, NY, USA
FinAnalytica, Inc, New York, NY, USA
e-mail: rachev@pstat.ucsb.edu
F.J. Fabozzi
EDHEC Business School, EDHEC Risk Institute, Nice, France
e-mail: fabozzi321@aol.com; frank.fabozzi@edhec.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 575
DOI 10.1007/978-1-4614-7750-1_20,
# Springer Science+Business Media New York 2015
576 A. Chernobai et al.

Abstract
In many financial models, such as those addressing value at risk and ruin
probabilities, the accuracy of the fitted loss distribution in the upper tail of the
loss data is crucial. In such situations, it is important to test the fitted loss
distribution for the goodness of fit in the upper quantiles, while giving lesser
importance to the fit in the low quantiles and the center of the distribution of the
data. Additionally, in many loss models the recorded data are left truncated with
the number of missing data unknown. We address this gap in literature by
proposing appropriate goodness-of-fit tests.
We derive the exact formulae for several goodness-of-fit statistics that
should be applied to loss models with left-truncated data where the fit of
a distribution in the right tail of the distribution is of central importance.
We apply the proposed tests to real financial losses, using a variety of
distributions fitted to operational loss and the natural catastrophe insurance
claims data, which are subject to the recording thresholds of $1 and $25 million,
respectively.

Keywords
Goodness-of-fit tests • Left-truncated data • Minimum recording threshold •
Loss distribution • Heavy-tailed data • Operational risk • Insurance • Ruin
probability • Value at risk • Anderson-Darling statistic • Kolmogorov-Smirnov
statistic

20.1 Introduction

In most loss models, the central attention is devoted to studying the distribu-
tional properties of the loss data. The shape of the dispersion of the data
determines the vital statistics such as the expected loss, variance, and ruin
probability, value at risk, or conditional value at risk where the shape in the
right tail is crucial. Parametric procedures for testing the goodness of fit
(GOF) include the likelihood ratio test and chi-squared test. A standard
semi-parametric procedure to test how well a hypothesized distribution fits
the data involves applying the in-sample GOF tests that provide a comparison
of the fitted distribution to the empirical distribution. These tests, referred to
as empirical distribution function (EDF) tests, include the Kolmogorov-
Smirnov test, Anderson-Darling test, and the Cramér-von Mises tests. Related
works on the discussion of these widely used tests include Anderson and
Darling (1952, 1954), D’Agostino and Stephens (1986), and Shorack and
Wellner (1986).
In many applications, the data set analyzed is incomplete, in the sense that the
observations are present in the loss database only if they exceed a predetermined
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 577

threshold level. This problem is usually absent in risk models involving market
risk and credit risk. However, it is a common problem in operational risk or
insurance claims models. In operational risk, banks’ internal databases are subject
to a minimum recording threshold of roughly $6,000–$10,000, and external
databases usually collect operational losses starting from $1 million, BCBS
(2003). Similarly, in non-life insurance models, the thresholds are set at $5 million,
$25 million, or other levels. Consequently, in the analysis of operational losses,
recorded loss data are left truncated, and, as a result, it is inappropriate to employ
standard GOF tests.
GOF tests for truncated and censored data have been studied by Dufour and
Maag (1978), Gastaldi (1993), and Guilbaud (1998), among others. In this paper,
we derive the exact formulae for several GOF test statistics that should be applied
where there exist incomplete samples with an unknown number of missing data in
low quantiles and propose two new statistics to determine the goodness of fit in the
upper tail that can be used for loss models where the accuracy of the upper tail
estimate is of central concern.
The paper is organized as follows. In Sect. 20.2 we describe the problem of left-
truncated samples and explain the necessary adjustments that are required to the
GOF tests to make them applicable for the truncated samples. In Sect. 20.3 we
review the widely used test statistics for complete samples and derive the exact
formulae for the statistics to be used for left-truncated samples. We propose in
Sect. 20.4 two new EDF statistics to be used for the situations when the fit in the
upper tail is of the central concern. Application of the modified EDF tests to
operational loss data, obtained from Zurich IC2 FIRST Database, and the USA
natural catastrophe insurance claims data, obtained from Insurance Services Office
Inc. Property Claim Services, is presented in Sect. 20.5, with final remarks in
Sect. 20.6. Necessary derivations are provided in the Appendix.

20.2 Problem Setup

Suppose we have a left-truncated sample, with the data below a prespecified

threshold level H not recorded (not observable). The observable data sample
x ¼ {x1, x2, . . ., xn} has each data point at least as great as H and includes a total
of n observations. Let {X( j)}1jn be a vector of the order statistics, such
that X(1)  X(2)  . . .  X(n). The empirical distribution function of the sample is
defined as

8
>
> 0 x < xð 1Þ
# observations  xðjÞ <
j
Fn ðxÞ :¼ ¼ xðjÞ  x < xðjþ1Þ , j ¼ 1, 2, . . . , n  1 :
n >
> n
:
1 x  xn ,
(20.1)
578 A. Chernobai et al.

Graphically, the empirical distribution function of an observed data sample is

represented as a step function with a jump of size 1/n occurring at each recorded
sample value. On the other hand, with left-truncated data which is a part of a larger
complete data set, the true size of jumps of the EDF at each value of the complete
data sample would be of size 1/nc, nc ¼ n + m rather than 1/n, where nc is the total
number of points of the complete data set and m is the unknown number of missing
points. In the GOF tests the null hypothesis states that the observed loss sample
belongs to a family of truncated distributions, with the parameter specified (simple
test) or unspecified (composite test).
We fit a continuous truncated distribution F to the data, given that the data
exceed or equal to H, and estimate the conditional parameters y with the maximum
likelihood (or an alternative method) by
!
Yn
fy ðxk Þ
^
y MLE ¼ arg max log : (20.2)
y
k¼1
1  Fy ðH Þ

Assuming that F is the true distribution, the estimated number of missing points
^ and the number of observed points n are related as1
m

^
m zH
¼ (20.3)
n ð1  zH Þ

resulting in
zH n
n^c ¼ n þn¼ (20.4)
ð1  zH Þ ð1  zH Þ
where zH :¼ F^y ðHÞ is the estimated distribution evaluated at the truncation point.
Then, the estimated empirical distribution function F^nc ðxÞ of complete data
sample is

estimated # observations  xðjÞ

F^nc ðxÞ :¼ , (20.5)
n^c

where the numerator refers to the total number of observations of the complete data
sample, not exceeding in magnitude the jth-order statistic of the incomplete
(observed) data sample such that X(1)  X(2)  . . .  X(n). By Eqs. 20.3 and 20.4,
Eq. 20.5 becomes

m^ þj n zH þ jð1  zH Þ j
F^nc ðxÞ ¼ ¼ ¼ zH þ ð1  zH Þ, j ¼ 1, 2, . . . , n:
n=ð1  zH Þ n n

l m
1
More accurately, m ^ should be estimated as m
^ ¼ n ð1z
zH
H Þ , but this detail can be ignored for all
practical purposes.
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 579

Rearranging terms leads to the fitted distribution function of the observed sample
of the following form:

8
< F^y ðxÞ  F^y ðH Þ

F^y ðxÞ ¼ xH
^ (20.6)
: 1  F y ðH Þ
0 x < H,

so that Fy(X)  U[Fy(H), 1] and Fy ðXÞ  U ½0; 1 under the null hypothesis that the
fitted distribution function is true. Therefore, the estimated empirical distribution
function of the observed part of the data, using Eq. 20.1, is represented by
8
> ^ y ðH Þ
F x < xð1 Þ
>
<
   
^ y ðH Þ þ F
F n ðx Þ 1  F ^ y ðH Þ ¼ j
1F ^ y ðH Þ xðjÞ  x < xðjþ1Þ , j ¼ 1,2, . .. ,n  1 :
^ y ðH Þ þ F
>
>
:n
1 x  xn ,
(20.7)

Figure 20.1 gives a visual illustration of the idea we just described. With
these modifications, the in-sample GOF tests can be applied to the left-truncated
samples.
In this paper we consider tests of a composite hypothesis that the
empirical distribution function of an observed incomplete left-truncated loss data
sample belongs to a family of hypothesized distributions (with parameters not
specified), i.e.,
H0 : Fn ðxÞ 2 Fy ðxÞ vs: = Fy ðxÞ:
H A : Fn ð x Þ 2 (20.8)
Under the null Eq. 20.8, Fy ðXÞ
 U ½0; 1, and the null is rejected if the p-value is
lower than the level a, such as a from 5 % to 10 %. Letting D be the
observed value of a GOF statistic (such as Kolmogorov-Smirnov or Anderson-
Darling) and d the critical value for a given level a, the p-value is computed as
p-value ¼ P(D  d). Since the distribution of the statistic is not parameter-free, one
way to compute the p-values and the critical values is by means of Monte Carlo
simulation, for each hypothesized fitted distribution (Ross 2001). Under the
procedure, the observed value D is computed. Then, for a given level a, the
following algorithm is applied:
1. Generate a large number of samples I (such as I ¼ 1,000) from the fitted
truncated distribution of size n (the number of observed data points), such that
these random variates are above or equal to H.
2. Fit a truncated distribution and estimate conditional parameters y^ for each
sample i ¼ 1, 2, . . .I.
3. Estimate the GOF statistic value Di for each sample i ¼ 1, 2, . . .I.
4. Calculate p-value as the proportion of times the sample statistic values exceed
the observed value d of the original sample.
5. Reject H0 if the p-value is smaller than a.
580 A. Chernobai et al.

Fig. 20.1 Illustration of the

empirical distribution 1 fitted CDF = Fθ(x)
function for data with missing F(x)
observations below the
threshold H and the fitted EDF
cumulative distribution
function
jump size 1/(n+m)

Fθ(H)

EDF missing
EDF observed
Fitted CDF missing
H Fitted CDF observed
0
missing observed x

1
fitted CDF = (Fθ(x)−Fθ(H)) / (1−Fθ(H))
F(x)

EDF

jump size 1/n

EDF observed
Fitted CDF observed
H
x
observed

A p-value of, for example, 0.3, would mean that in 30 % of samples of the same
size simulated from the same distribution with the same parameter estimation
procedure applied, the test statistic value was higher than the one observed in the
original sample.

20.3 EDF Statistics for Left-Truncated Loss Samples

The EDF statistics are based on the vertical differences between the empirical
and fitted (truncated) distribution function. They are divided into two classes:
(1) the supremum class (such as Kolmogorov-Smirnov and Kuiper statistics) and
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 581

(2) the quadratic class (such as Anderson-Darling and Cramér-von Mises statistics).
In this section, we derive the exact computing formulae for a number of EDF
test statistics, modified so that they can be applied to left-truncated loss samples.
For the left-truncated samples, Fy(X) denotes the null distribution function for
left-truncated sample values. The variable Fy(X) is distributed uniformly over the
[0, 1] interval. The variable Fy(X) is distributed uniformly over the [Fy(H), 1]
interval. We reserve some other notations: zH :¼ F^y ðH Þ that was defined earlier and
 
zj :¼ F^y xðjÞ for truncated samples. In this section we discuss the Kolmogorov-
Smirnov, Kuiper, Anderson-Darling, and the Cramér-von Mises statistics and use
an asterisk (*) to denote their left-truncated sample analog.

20.3.1 Supremum Class Statistics

20.3.1.1 Kolmogorov-Smirnov Statistic

A widely used supremum class statistic, the Kolmogorov-Smirnov (KS) statistic,
measures the absolute value of the maximum distance between the empirical and fitted
distribution function and puts equal weight on each observation. Let {X( j)}1<j<n be
the vector of the order statistics and X(1) < X(2) < . . . < X(n), such that strict
inequalities hold. Usually, such distance is the greatest around the median of the
sample. For the left-truncated data samples, the KS statistic is expressed as
pffiffiffi  
  
KS ¼ n sup Fn ðxÞ  F^y ðxÞ: (20.9)
x
The KS* statistic can be computed from
n   pffiffiffi  
pffiffiffi  o n j
KSþ ¼ n sup Fn xðjÞ  F^ y xðjÞ ¼ sup zH þ ð1  zH Þ  zj ,
1  zH j n
j
n   o pffiffiffi   
p ffiffi
ffi   n j1
 ^
KS ¼ n sup F y xðjÞ  Fn xðjÞ ¼ sup zj  zH þ ð1  zH Þ ,
j 1  zH j n
and becomes
KS ¼ maxfKSþ , KS g: (20.10)

20.3.1.2 Kuiper Statistic

The KS statistic gives no indication of whether the maximum discrepancy between
  
Fn(x) and F^y ðxÞ occurs when Fn(x) is above F^y ðxÞ or when F^y ðxÞ is above Fn(x). The
Kuiper statistic (V) is closely related to the KS statistic. It measures the total sum of
the absolute values of the two largest vertical deviations of the fitted distribution
 
function from Fn(x), when Fn(x) is above F^y ðxÞ and when F^y ðxÞ is above Fn(x).
For left-truncated data samples, it is computed as
V  ¼ KSþ þ KS , (20.11)
+ 
with KS * and KS * defined in Sect. 20.1.1.
582 A. Chernobai et al.

20.3.1.3 Anderson-Darling Statistic

There are two variations of the Anderson-Darling (AD) statistic. Its
simplest, supremum class version is a variance-weighted KS statistic with

 


1=2
a weight of c F^y ðxÞ ¼ F^y ðxÞ 1  F^y ðxÞ attached to each observation
in Eq. 20.9. (The second version will be discussed in the next section.) Under this
specification, the observations in the lower and upper tails of the truncated sample are
assigned a higher weight. Let {X( j)}1 jn be the vector of the order statistics, such that
X(1)  X(2)  . . .  X(n). Then the AD statistic is defined for left-truncated samples as
 
 
  
p ffiffi
ffi  F ð x Þ  ^
F ð x Þ 
  n y
AD ¼ n sup rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : (20.12)
x 

 
^ 
ð Þ  ^ 
ð Þ 
 F y x 1 F y x 

For left-truncated samples, the computing formula is derived from

8 9 8 9
>
>     >
> > >
<  = <
þ pffiffiffi ^
F n x ð j Þ  F y xð j Þ pffiffiffi zH þ nð1  zH Þ  zj =
j
AD ¼ n sup rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ n sup qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
j >
>  
  > > z  z 1  z >
: F^ y xð jÞ 1  F ^  x ð jÞ > ; j :
j H j ;
y
8 9 8 9,
>
>     >
> > j1 >
pffiffiffi < F^ xð jÞ  Fn xð jÞ = pffiffiffi < z j  z H  ð 1  z H =
Þ
 y n
AD ¼ n sup rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ n sup qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  >
>  
  > >
j :
j >
: F ^  xð jÞ 1  F ^  xð jÞ , > ; zj  zH 1  zj ;
y y

and becomes

AD ¼ maxfADþ , AD g: (20.13)

20.3.2 Quadratic Class Statistics

The quadratic statistics for complete data samples are grouped under the Cramér-
von Mises family as
ð
1
 2  
Q¼n Fn ðxÞ  F^y ðxÞ c F^y ðxÞ dF^y ðxÞ, (20.14)
1
 
in which the weight function c F^y ðxÞ is assigned to give a certain weight
to different observations, depending on the purpose. For left-truncated
samples, we denote the Cramér-von Mises family as Q* and F^y ðxÞ is replaced by

F^y ðxÞ:
ð

1
2
  

Q ¼ n Fn ðxÞ  F^y ðxÞ c F^y ðxÞ dF^y ðxÞ:

(20.15)
H
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 583

Depending on the form of the weighting function, the sample observations are

 


1
given a different weight. c F^y ðxÞ ¼ F^y ðxÞ 1  F^y ðxÞ yields the quadratic

 
Anderson-Darling statistic and c F^y ðxÞ ¼ 1 yields the Cramér-von Mises statistic.
Derivation of the computing formulae makes use of Eq. 20.1 and involves the Proba-
bility Integral Transformation (PIT) technique. For left-truncated samples, this leads to
ð1 ð1  
   2 n   z  zH 2 z  zH
Q ¼ n ðFn ðz Þ  z Þ cðz Þdz ¼ Fn ð z Þ  c dz,
1  zH 1  zH 1  zH
0 zH
(20.16)

where Z ¼ F^y ðXÞ
¼ Zz  ^
1zH with Fy ðX Þ  U ½0; 1 under the null, and so Z ¼ F y ðX Þ
H

 
with Fy(X)  U[Fy(H), 1] under the null. Fn(Z ) ¼ Fn(X) ¼ Fy(Fn(X)) is the
 
empirical distribution function of Z*. zH ¼ F^y ðH Þ ¼ Fy F^y ðH Þ .

20.3.2.1 Anderson-Darling Statistic

The supremum version of the AD statistic was described in Sect. 20.1.3. Another,
more generally used, version of this statistic belongs to the quadratic class defined by
the Cramér-von Mises family (Eq. 20.15) with the weight function for left-truncated

 


1
samples of c F^y ðxÞ ¼ F^y ðxÞ 1  F^y ðxÞ . Again, with this specification,
most weight is being put on the outer left and right quantiles of the distribution,
proportional to the appropriate tails. Let {X( j)}1 jn be the vector of the order
statistics, such that X(1)  X(2)  . . .  X(n). The computing formula for the AD
statistic for left-truncated samples becomes (the derivation is given in Appendix)
1X n  
AD2 ¼  n þ 2n logð1  zH Þ  ð1 þ 2ðn  jÞÞlog 1  zj þ . . .
n j¼1
(20.17)
1X n  
þ ð1  2jÞlog zj  zH :
n j¼1

20.3.3 Cramér-von Mises Statistic

Cramér-von Mises (denoted as W2) statistic belongs

 tothe Cramér-von Mises
family (Eq. 20.15) with the weight function c F^ ðxÞ ¼ 1 . Let {X( j)}1 jn
y
be the vector of the order statistics, such that X(1)  X(2)  . . .  X(n).
The computing formula for the statistic for left-truncated samples becomes (the
derivation is given in Appendix)
n nzH 1 X n
1 X
n  2
W 2 ¼ þ þ ð1  2jÞzj þ zj  zH : (20.18)
3 1  zH nð1  zH Þ j¼1 2
ð1  zH Þ j¼1
584 A. Chernobai et al.

20.4 “Upper Tail” Anderson-Darling Statistic

In practice, there are often cases when it is necessary to test whether

a distribution fits the data well in the upper tail and the fit in the lower tail
or around the median is of little or less importance. Examples include opera-
tional risk and insurance claims modelling, in which goodness of the fit in the
tails determines the value at risk, conditional value at risk, and ruin probabil-
ities. Given the Basel II Capital Accord’s recommendations, under the loss
distribution approach the operational risk capital charge is derived from the
value-at-risk measure, which requires an accurate estimate of the upper tail of
the loss distribution. Similarly, in insurance, the upper tail of the claim size
distribution is central to obtaining accurate estimates of ruin probability. For
this purpose, we introduce a statistic, which we refer to as the upper tail
Anderson-Darling statistic and denote by ADup. We propose two different
versions of the statistic.

20.4.1 Supremum Class “Upper Tail” Anderson-Darling Statistic

The first version of ADup belongs to the supremum class EDF statistics. For
complete data samples, each observation of the KS statistic is assigned a weight
   1
of c F^y ðxÞ ¼ 1  F^y ðxÞ . Under this specification, the observations in the
upper tail are assigned a higher weight than those in the lower tail. Let {X( j)}1 jn
be the vector of the order statistics, such that X(1)  X(2)  . . .  X(n). Then we
define the ADup statistic for complete data samples as
 
pffiffiffi Fn ðxÞ  F^y ðxÞ
ADup 
¼ n sup  : (20.19)
x 1  F^y ðxÞ 
 
Denoting zj :¼ F^y xðjÞ , the computing formula is derived from
( )
pffiffiffi j
 zj
ADþ
up ¼ n sup n
,
j 1  zj
( )
pffiffiffi zj  j n 1
AD ¼ n sup ,
up
j 1  zj

and becomes
n o
ADup ¼ max ADþ
up , AD 
up : (20.20)

For left-truncated samples, the counterpart of the ADup statistic can be similarly
computed using
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 585

(    ) ( )
pffiffiffi Fn xðjÞ  F^y xðjÞ pffiffiffi zH þ njð1  zH Þ  zj
ADþ ¼ n sup   ¼ n sup ,
1  F^y xðjÞ 1  zj
up
j j
(    ) ( )
pffiffiffi F^y xðjÞ  Fn xðjÞ pffiffiffi zj  zH  j n 1ð1  zH Þ
AD ¼ n sup   ¼ n sup ,
1  F^ xðjÞ 1  zj
up
j y j

and becomes
n o
ADup ¼ max ADþ 
up , ADup : (20.21)

20.4.2 Quadratic Class “Upper Tail” Anderson-Darling Statistic

Another way to define the upper tail Anderson-Darling statistic is by an integral of

the Cramér-von Mises family (Eq. 20.14) with the weighting function2 of the form
   2
 

2
c F^y ðxÞ ¼ 1  F^y ðxÞ for complete samples and c F^y ðxÞ ¼ 1  F^y ðxÞ
for left-truncated samples. Such weighting function gives a higher weight
to the upper tail and a lower weight to the lower tail. We define this statistic
as AD2up.
Let {X( j)}1 jn be the vector of the order statistics, such that X(1)  X(2)  . . .  X(n).
The AD2up statistic’s general form for complete samples can be expressed as
ð 
1 2
Fn ðxÞ  F^y ðxÞ
ADup ¼ n  d F^y ðxÞ:
2
 (20.22)
1  ^y ðxÞ 2
F
H

For complete data samples Fy(X)  U[0, 1] under the null hypothesis. If we denote
 
zj :¼ F^y xðjÞ , straightforward calculations lead to the following computing formula:

X
n   1X n
1
AD2up ¼ 2 log 1  zj þ ð 1 þ 2ð n  j Þ Þ :
j¼1
n j¼1 1  zj

For incomplete left-truncated samples, F*y(X) is distributed U[Fy(H), 1]

under the null. Applying the PIT technique leads to the computing formula of the

2
It has been shown that the weighting function cðtÞ ¼ ð1  tÞb possesses nice asymptotic
properties only for b¼[0,2) (Deheuvels and Martynov 2003). For b¼2, which is the case consid-
ered in this chapter, the asymptotic distribution of the test statistic has infinite mean. This is indeed
a concern for very large samples (i.e., the asymptotic case n!1). Yet, because firms’ operational
loss data samples are typically relatively small, the asymptotic distribution of the test statistic
should not generate large concerns. Nevertheless, the results of the proposed quadratic class upper-
tail Anderson-Darling test should be treated with caution and with consideration of the properties
described above.
586 A. Chernobai et al.

Table 20.1 Description of EDF statistics for complete data samples

H0 : Fn(x) ∈ F(x) vs. HA : Fn(x) 2 = Fy(x)
 
Notations: zj :¼ F^ y xð jÞ , j ¼ 1, 2, . . . , n
Statistic Description and computing formula
pffiffiffi  
KS KS ¼ n supFn ðxÞ  F^ y ðxÞ
x
Computing formula:
( )
pffiffiffi n o n o
j j1
KS ¼ n max sup  zj , sup zj 
j n j n

V pffiffiffi  
V ¼ n sup Fn ðxÞ  F^ y ðxÞ þ sup F^ y ðxÞ  Fn ðxÞ
x x
Computing formula: !
pffiffiffi n o n o
j j1
V ¼ n sup  zj þ sup zj 
j n j n
 
AD  
pffiffiffi  F ðxÞ  F^ ðxÞ 
 n y 
AD ¼ n supqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
x  F
^ ^
 y ðxÞ 1  F y ðxÞ 

8 8
Computing formula: 9 8 99
< >
> < jz >
= >
< z  j1 > =>
=
pffiffiffi j j
AD ¼ n max sup qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
  , sup q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

n
 >
: j >
> : zj 1  zj > ; j > : zj 1  zj > ;;
 
ADup pffiffiffi Fn ðxÞ  F^y ðxÞ
ADup ¼ n sup 
x 1  F^y ðxÞ 
Computing formula:
(    )
pffiffiffi n  zj
j
zj  j1
ADup ¼ n max sup , sup n

j 1  zj j 1  zj
AD2 ð
1  
Fn ðxÞ  F^ y ðxÞ
2
AD2 ¼ n  d F^y ðxÞ
F^y ðxÞ 1  F^ y ðxÞ
1
Computing formula:
Xn Xn
1 1  
AD2 ¼ n þ ð1  2jÞlog zj  ð1 þ 2ðn  jÞÞ log 1  zj
j¼1 n j¼1 n
W2 ð
1
 2
W2 ¼ n Fn ðxÞ  F^ y ðxÞ dF^ y ðxÞ
1
Computing formula:
n 1 Xn Xn
W2 ¼ þ ð1  2jÞzj þ z2
j¼1 j
3 n j¼1
AD2up ð 
1 2
Fn ðxÞ  F^ y ðxÞ
AD2up ¼ n  2 dF^ y ðxÞ
1
1  F^ y ðxÞ
Computing formula:
1 Xn 1 Xn  
AD2up ¼ ð1 þ 2ðn  jÞÞ  þ2 log 1  zj
n j¼1 1  zj j¼1
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 587

Table 20.2 Description of EDF statistics for left-truncated (threshold ¼ H) data samples
ðxÞFy ðHÞ
H0 : Fn(x) ∈ F*y(x) vs. HA : Fn(x) 2 = F*y(x), Fy ðxÞ :¼ Fy1F y ðH Þ
 
^ ^
Notations: zj :¼ F y xðjÞ , zH ¼ F y ðH Þ, j ¼ 1, 2, . . . , n
Statistic Description and computing formula
pffiffiffi  
  
KS* KS ¼ n supFn ðxÞ  F^ y ðxÞ
x
Computing formula: (
pffiffiffi n o
n j
KS ¼ max sup zH þ ð1  zH Þ  zj ,
1  zH j n
  )
j1
sup zj  zH þ ð1  zH Þ
j n
 n o n  o
V* pffiffiffi 
V  ¼ n sup Fn ðxÞ  F^ y ðxÞ þ sup F^ y ðxÞ  Fn ðxÞ
x x
Computing formula:
pffiffiffi n o   !
 n j j1
V ¼ sup zH þ ð1  zH Þ  zj þ sup zj  zH þ ð1  zH Þ
1  zH j n j n
 
AD*  
 
pffiffiffi  Fn ðxÞ  F^ y ðxÞ 
 
AD ¼ n suprffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi 
x 


^ 
^  
 F y ð xÞ 1  F y ð xÞ 

8 8
Computing formula: 9 8 99
>
< > < z þ j ð1  z Þ  z > = <z  z  j1 ð1  z Þ>
> =>
=
p ffiffi
ffi
AD ¼ n max sup qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H H j j H H
n
   ffi , sup qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

n
  ffi
>
: j >: zj  zH 1  zj > ; j > : zj  zH 1  zj > ;>
;
 
AD*up pffiffiffi F ðxÞ  F^ ðxÞ

 n 
ADup ¼ n sup y

x  1F ^  ðxÞ 
y
Computing formula:
( ( ))
 
pffiffiffi zH þ nj ð1  zH Þ  zj zj  zH  j1
n ð1  zH Þ
ADup ¼ nmax sup , sup
j 1  zj j 1  zj

2
AD2* 
ð Fn ðxÞ  F^ ðxÞ
1
y 
AD2 ¼ n 


dF^ y ðxÞ
H
F^y ðxÞ 1  F^y ðxÞ
Computing formula: Xn  
AD2 ¼ n þ 2nlogð1  zH Þ  1n ð1 þ 2ðn  jÞÞlog 1  zj þ
Xn  
j¼1
1 ð1  2jÞlog z  z
n j¼1 j H

W2* ð
1

2
 
W 2 ¼ n Fn ðxÞ  F^ y ðxÞ dF^ y ðxÞ
H
Computing formula:
n n zH 1 X n
1 X
n  2
W 2 ¼ þ þ ð1  2jÞzj þ zj  zH
3 1  zH nð1  zH Þ j¼1 2
ð1  zH Þ j¼1
(continued)
588 A. Chernobai et al.

Table 20.2 (continued)

ðxÞFy ðHÞ
H0 : Fn(x) ∈ F*y(x) vs. HA : Fn(x) 2 = F*y(x), Fy ðxÞ :¼ Fy1F y ðH Þ
 
^ ^
Notations: zj :¼ F y xðjÞ , zH ¼ F y ðH Þ, j ¼ 1, 2, . . . , n
Statistic Description and computing formula

2
AD2up* ð Fn ðxÞ  F^  ðxÞ
1
y 
up ¼ n
AD2

2 dF^y ðxÞ
H 1  F^y ðxÞ
Computing formula:
X
n   1  zH X
n

up ¼ 2n logð1  zH Þ þ 2
AD2 log 1  zj þ ð1 þ 2ðn  jÞÞ 1z
1

j¼1
n j¼1 j

AD2*
up statistic for left-truncated samples of the following form (the derivation is
given in Appendix):
X
n   1  zH X
n
1
up ¼ 2n logð1  zH Þ þ 2
AD2 log 1  zj þ ð1 þ 2ðn  jÞÞ :
j¼1
n j¼1 1  zj

Tables 20.1 and 20.2 summarize the EDF statistics and their computing formulae
for complete and left-truncated samples.

20.5 Application to Loss Data

In this section we apply the GOF testing procedure to (1) operational loss data,
extracted from an external database, and (2) catastrophe insurance claims data. The
operational loss data set was obtained from Zurich IC Squared (IC2) FIRST Database
of Zurich IC Squared (IC2), an independent consulting subsidiary of Zurich Financial
Services Group. The external database is comprised of operational loss events
throughout the world. The original loss data cover losses in the period 1950–2002.
A few recorded data points were below $1 million in nominal value, so we excluded
them from the analysis, to make it more consistent with the conventional threshold for
external databases of $1 million. Furthermore, we excluded the observations before
1980 because of relatively few data points available (which is most likely due to poor
data recording practices). The final data set for the analysis covered losses for the time
period between 1980 and 2002. It consists of five types of losses: “relationship” (such
as events related to legal issues, negligence, and sales-related fraud), “human” (such
as events related to employee errors, physical injury, and internal fraud), “processes”
(such as events related to business errors, supervision, security, and transactions),
“technology” (such as events related to technology and computer failure and tele-
communications), and “external” (such as events related to natural and man-made
disasters and external fraud). The loss amounts have been adjusted for inflation using
the Consumer Price Index from the U.S. Department of Labor. The numbers of data
points of each type are n ¼ 849, 813, 325, 67, and 233, respectively.
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 589

Table 20.3 Goodness-of-fit tests for operational loss data

KS V AD ADup AD2 AD2up W2
Exponential
Relationship 11.0868 11.9973 1.3·107 1.2·1023 344.37 1.2·1014 50.5365
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
Human 14.0246 14.9145 2.4·106 1.1·1022 609.15 3.0·1012 80.3703
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
Processes 7.6043 8.4160 3.7·106 1.7·1022 167.61 6.6·105 22.5762
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
Technology 3.2160 3.7431 27.6434 1.4·106 27.8369 780.50 2.9487
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
External 6.5941 6.9881 4.4·106 2.0·1022 128.35 5.0·107 17.4226
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
Catastrophe 5.5543 5.9282 9.0·106 4.1·1022 72.2643 6.1·1013 13.1717
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
Lognormal
Relationship 0.8056 1.3341 2.6094 875.40 0.7554 4.6122 0.1012
[0.082] [0.138] [0.347] [0.593] [0.043] [0.401] [0.086]
Human 0.8758 1.5265 3.9829 1086.16 0.7505 4.5160 0.0804
[0.032] [0.039] [0.126] [0.462] [0.044] [0.408] [0.166]
Processes 0.6584 1.1262 2.0668 272.61 0.4624 4.0556 0.0603
[0.297] [0.345] [0.508] [0.768] [0.223] [0.367] [0.294]
Technology 1.1453 1.7896 2.8456 41.8359 1.3778 6.4213 0.2087
[<0.005] [0.005] [0.209] [0.994] [<0.005] [0.067] [<0.005]
External 0.6504 1.2144 2.1702 316.20 0.5816 2.5993 0.0745
[0.326] [0.266] [0.469] [0.459] [0.120] [0.589] [0.210]
Catastrophe 0.6854 1.1833 5.3860 1.1·104 0.7044 27.4651 0.0912
[0.243] [0.307] [0.064] [0.053] [0.068] [0.023] [0.111]
Weibull
Relationship 0.5553 1.0821 3.8703 2.7·104 0.7073 13.8191 0.0716
[0.625] [0.514] [0.138] [0.080] [0.072] [0.081] [0.249]
Human 0.8065 1.5439 4.3544 3.2·104 0.7908 8.6610 0.0823
[0.093] [0.051] [0.095] [0.068] [0.053] [0.112] [0.176]
Processes 0.6110 1.0620 1.7210 2200.75 0.2069 2.2340 0.0338
[0.455] [0.532] [0.766] [0.192] [0.875] [0.758] [0.755]
Technology 1.0922 1.9004 2.6821 52.5269 1.4536 4.8723 0.2281
[<0.005] [<0.005] [0.216] [0.944] [<0.005] [0.087] [<0.005]
External 0.4752 0.9498 2.4314 4382.68 0.3470 5.3662 0.0337
[0.852] [0.726] [0.384] [0.108] [0.519] [0.164] [0.431]
Catastrophe 0.8180 1.5438 5.6345 1.5·104 1.3975 15.8416 0.1965
[0.096] [0.041] [0.043] [0.028] [0.007] [0.025] [0.006]
Pareto (GPD)
Relationship 1.4797 2.6084 3.5954 374.68 3.7165 22.1277 0.5209
[<0.005] [<0.005] [0.172] [>0.995] [<0.005] [0.048] [<0.005]
(continued)
590 A. Chernobai et al.

Table 20.3 (continued)

KS V AD ADup AD2 AD2up W2
Human 1.4022 2.3920 3.6431 374.68 2.7839 23.7015 0.3669
[<0.005] [<0.005] [0.167] [>0.995] [<0.005] [0.051] [<0.005]
Processes 1.0042 1.9189 4.0380 148.24 2.6022 13.1082 0.3329
[<0.005] [<0.005] [0.104] [>0.995] [<0.005] [0.087] [<0.005]
Technology 1.2202 1.8390 3.0843 33.4298 1.6182 8.8484 0.2408
[<0.005] [<0.005] [0.177] [>0.995] [<0.005] [0.067] [<0.005]
External 0.9708 1.8814 2.7742 151.94 1.7091 8.6771 0.2431
[0.009] [0.005] [0.284] [0.949] [<0.005] [0.106] [<0.005]
Catastrophe 0.4841 0.8671 2.4299 1277.28 0.3528 4.3053 0.0390
[0.799] [0.837] [0.369] [0.239] [0.490] [0.235] [0.645]
This table reports goodness-of-fit test statistic values for operational loss data of various risk types.
p-values are reported in square brackets and were obtained via 1,000 Monte Carlo simulations

The insurance claims data set covers claims resulting from natural catastrophe
events occurred in the United States over the time period from 1990 to 1996. It was
obtained from Insurance Services Office Inc. Property Claim Services (PCS). The
data set includes 222 losses. The observations are greater than $25 million in
nominal value.
Left-truncated distributions of four types were fitted to each of the data set:
exponential, lognormal, Weibull, and Pareto (GPD). Table 20.3 presents the observed
statistic values and the p-values for the six data sets (five operational losses and
insurance claims), obtained with the testing procedure described in Sect. 20.2.
The results reported in Table 20.3 suggest that fitting heavier-tailed distribu-
tions, such as Pareto and Weibull, results in lower values of the GOF statistics,
which leads to acceptance of the null for practically all loss types and all
criteria of the goodness of fit, viewed from the high p-values. Since the analysis
of operational losses deals with estimating the operational value at risk
(VaR), it is reasonable to determine the ultimate best fit on the basis of the ADup
and AD2up statistics, introduced in this paper. As can be seen from Table 20.3, these
proposed measures suggest a much better fit of the heavier-tailed distributions.
Moreover, for the Pareto distribution, while the statistics that focus on the center of
the data (Kolmogorov-Smirnov, Kuiper, Cramér-von Mises) do not show a good
fit, the ADup and AD2up statistics indicate that the fit in the upper tail is very good.
It should be noted that the Pareto and Weibull distributions very often suggest
a superior fit in the upper tail to the lognormal distribution. Yet it is the lognormal
distribution that was suggested in 2001 by the Basel Committee (BCBS (2001)).

20.6 Conclusions

In this paper we present a technique for modifying the existing goodness-of-fit test
statistics so that they can be applied to loss models in which the available data set is
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 591

incomplete and is truncated from below. Such left truncation is often present in loss
data when the data are being recorded starting from a fixed amount and the data
below are not recorded at all. Exact computing formulae for the Kolmogorov-
Smirnov, Kuiper, Anderson-Darling, and Cramér-von Mises for the left-truncated
samples are presented.
In risk management, it is often vital to have a good fit of a hypothesized
distribution in the upper tail of the loss data. It is important in loss models that
deal with value at risk, conditional value at risk, and ruin probability. We suggest
using two other versions of the Anderson-Darling statistic (which we refer to as
the upper tail Anderson-Darling statistic) in which the weighting function is
proportional to the weight of only the upper tail of the distribution. Supremum
and quadratic versions of the statistic are proposed. Such statistic is convenient to
use when it is necessary to examine the goodness of fit of a distribution in the right
tail of the data, while the fit in the left tail is unimportant.
The technique is applied to check the goodness of fit of a number of
distributions using operational loss data and catastrophe insurance claims data
sets. From the empirical analysis we conclude that heavier-tailed distributions better
fit the data than Lognormal or thinner-tailed distributions in many instances. In
particular, the conclusion is strongly supported by the upper tail Anderson-Darling
tests.

Appendix

Derivation of AD2* Computing Formula

By the PIT technique


2
ð
þ1
Fn ðxÞ  F^y ðxÞ 
AD2 ¼ n 


 d F^y ðxÞ
^ ^
F y ðxÞ 1  F y ðxÞ
H
ð1
ðFn ðz Þð1  zH Þ þ zH  zÞ2
PIT n^c dz,
ðz  zH Þð1  zÞ
zH

where in the original integral Fn(Z) ¼ F

y(Fn(X)) ¼ Fn(X) is the empirical
 
distribution function of Z :¼ F^y ðXÞ ¼ Fy F^y ðXÞ so that Fy ðÞ  U ½0; 1 .
Changing variable and using Eq. 20.7, the integral becomes expressed in terms of
   
zH :¼ F^y ðHÞ ¼ Fy F^y ðH Þ and Z :¼ F^y ðXÞ ¼ Fy F^y ðXÞ so that Fy(·)  U[zH,1].
We estimate n^c as 1z
n
.  
Using Eq. 20.7, the computing formula is expressed in terms of zj :¼ F^y xðjÞ ¼
H

  
^
Fy F y xðjÞ , j ¼ 1, 2, . . ., n, as
592 A. Chernobai et al.

ð z1 n1 ð zjþ1 j
 2 ð1
1 ðz  zH Þ2 X nð1  zH Þ þ zH  z ð1  zÞ2
AD2 ¼ dz þ dz þ dz :
n^c zH ðz  zH Þð1  zÞ j¼1 zj
ðz  zH Þð1  zÞ zn ðz  zH Þð1  zÞ
|fflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl {zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl}
A B C

Separately solving for A, B, and C, we obtain

A ¼ zH  z1 þ ð1  zH Þðlogð1  zH Þ  logð1  z1 ÞÞ;
1  zH X
n1     
B ¼ z1  zn þ 2
ðn  jÞ2 log 1  zj  log 1  zjþ1  . . .
n j¼1

1  zH X
n1     
2 j2 log zj  zH  log zjþ1  zH
n2 j¼1

1  zH Xn  
¼ z1  zn þ ð1  zH Þlogð1  z1 Þ  2
ð1 þ 2ðn  jÞÞ log 1  zj þ . . .
n j¼1

1  zH Xn  
þ ð1  zH Þlogðzn  zH Þ þ 2
ð1  2jÞlog zj  zH ;
n j¼1

C ¼ zn  1 þ ð1  zH Þðlogð1  zH Þ  logðzn  zH ÞÞ:

Summing the terms A, B, and C, multiplying by n^c , and simplifying yields the
final computing formula:
1X n  
AD2 ¼  n þ 2n logð1  zH Þ  ð1 þ 2ðn  jÞÞ log 1  zj þ . . .
n j¼1

1X n  
þ ð1  2jÞlog zj  zH :
n j¼1

Derivation of W2* Computing Formula

By the PIT technique

ð
þ1

2 

W 2
¼n Fn ðxÞ  F^y ðxÞ d F^y ðxÞ
H
ð1
ð Fn ð z  Þ ð 1  z H Þ þ z H  z Þ 2
PIT n^c dz,
ð1  zH Þ2
zH
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 593

where in the original integral Fn(Z) ¼ F

y(Fn(X)) ¼ Fn(X) is the empirical
 
distribution function of Z :¼ F^y ðXÞ ¼ Fy F^y ðXÞ so that Fy ðÞ  U ½0; 1 .
Changing variable and using Eq. 20.7, the integral becomes  expressed
in terms of zH :¼ F^y ðHÞ ¼ Fy F^y ðH Þ and Z :¼ F^y ðXÞ ¼ Fy F^y ðXÞ so that
Fy(·)  U[Fy(H), 1]. We estimate n^c as 1z
n
.  
Using Eq. 20.7, the computing formula is expressed in terms of zj :¼ F^y xðjÞ ¼
H

  
^
Fy F y xðjÞ , j ¼ 1, 2, . . ., n, as

ð z1 n1 ð zjþ1 
X 2 ð1
ð1  zH Þ2 2 j
W ¼ ðzH  zÞ2 dz þ ð1  zH Þ þ zH  z dz þ ð1  zÞ2 dz :
n^c zH j¼1 zj n zn
|fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}
A B C

Separately solving for A, B, and C, we obtain

zH 3 z1 3
A¼ þ zH 2 z1  zH z1 2 þ ;
3 3
zn 3 z 1 3
B¼  þ zH z1 2  zH zn 2 þ z H 2 z1 þ zH 2 zn þ . . .
3 3
ð1  zH Þ2 X
n1   1  zH Xn1   1  zH X
n1  
þ 2
j2 zjþ1  zj þ j zj 2  zjþ1 2 þ 2zH j zjþ1  zj
n j¼1
n j¼1 n j¼1
!
zn 3 z 1 3 ð 1  zH Þ2 2 Xn
¼  þ zH z1  zH zn  z H z1 þ zH zn þ
2 2 2 2
n zn þ ð1  2jÞzj þ . . .
3 3 n2 j¼1
! !
1  zH X n
1  zH Xn
þ zj  nzn þ 2zH
2 2
nzn  zj
n j¼1
n j¼1

ð1  zH Þ2 Xn
¼ ð1  zH Þzn þ zH ð1  zH Þzn  ð1  zH Þzn 2 þ ð1  2jÞzj þ .. .
n2 j¼1

1  zH X
n
1  zH X
n
þ zj 2  2zH zj ;
n j¼1 n j¼1

1 zn 3
C ¼ þ zn 2  zn  :
3 3
n^c
Summing the terms A, B, and C, multiplying by ð1zH Þ2
, and simplifying yields
the final computing formula:
n nzH 1 X n
1 X
n  2
W 2 ¼ þ þ ð1  2jÞzj þ zj  zH :
3 1  zH nð1  zH Þ j¼1 2
ð1  zH Þ j¼1
594 A. Chernobai et al.

Derivation of AD2up Computing Formula

By the PIT technique

ð
þ1  2 ð1
Fn ðxÞ  F^y ðxÞ ðFn ðzÞ  zÞ2
AD2up ¼n  2 d ^
F y ð xÞ PIT n dz,
1
1  F^y ðxÞ 0
ð1  zÞ2

where Fn(Z) ¼ Fy(Fn(X)) ¼ Fn(X) is the empirical distribution function of Z ¼ F^y ðXÞ
 
¼ F F^y ðXÞ so that Fy(·)  U[0,1].  
Using Eq. 20.1, the computing formula is expressed in terms of zj :¼ F^y xðjÞ ¼
  
F F^y xðjÞ , j ¼ 1, 2, . . ., n, as

ð z1 n1 ð zjþ1
j 2 ð1
1 z2 X z ð1  zÞ2
AD2up ¼ 2
dz þ n
dz þ dz :
n 0 ð1  zÞ j¼1 zj ð1  zÞ2 zn ð1  zÞ2
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}
A B C

Separately solving for A, B, and C, we obtain

1
A ¼ z1  1 þ þ 2logð1  z1 Þ;
1  z1

1Xn1
2 1 1
B ¼ zn  z1  2 ðn  jÞ   ...
n j¼1 1  zj 1  zjþ1

1Xn1     
2 ðn  jÞ log 1  zj  log 1  zjþ1
n j¼1

1 1Xn
1 1X n  
¼ zn  z1  þ 2 ð1 þ 2ðn  jÞÞ  2logð1  z1 Þ þ 2 log 1  zj ;
1  z1 n j¼1 1  zj n j¼1

C ¼ 1  zn :

Summing the terms A, B, and C, multiplying by n, and simplifying yields the

final computing formula:

X
n   1X n
1
AD2up ¼ 2 log 1  zj þ ð 1 þ 2ð n  j Þ Þ :
j¼1
n j¼1 1  zj
20 Composite Goodness-of-Fit Tests for Left-Truncated Loss Samples 595

Derivation of AD2*
up Computing Formula

By the PIT technique


2
ð
þ1
Fn ðxÞ  F^y ðxÞ ð1
^  ðFn ðz Þð1  zH Þ þ zH  zÞ2
up ¼ n
AD2
2
d F y ð x Þ PIT ^
n c
dz,

1  F^y ðxÞ ð1  zÞ2
H zH

where in the original integral Fn(Z) ¼ F

y(Fn(X)) ¼ Fn(X) is the empirical
 
distribution function of Z :¼ F^y ðXÞ ¼ F F^y ðXÞ so that Fy ðÞ  U ½0; 1 .
Changing variable and using Eq. 20.7, the integral becomes expressed in terms of
   
zH :¼ F^y ðHÞ ¼ F F^y ðH Þ and Z :¼ F^y ðXÞ ¼ Fy F^y ðXÞ so that Fy(·)  U[zH,1].
We estimate n^c as 1z
n
.  
Using Eq. 20.7, the computing formula is expressed in terms of zj :¼ F^y xðjÞ ¼
H

  
F F^y xðjÞ , j ¼ 1, 2, . . ., n as

ð z1 n1 ð zjþ1
j 2 ð1
1 ðz  zH Þ2 X ð1  zH Þ þ zH  z ð1  zÞ2
AD2 ¼ dz þ n
dz þ dz :
n^c up zH ð1  zÞ 2
j¼1 zj ð1  zÞ 2
zn ð1  zÞ2
|fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}
A B C

Separately solving for A, B and C, we obtain

1
A ¼ z1  zH  ð1  zH Þ þ ð1  zH Þ2  2ð1  zH Þlogð1  zH Þ
1  z1
þ 2ð1  zH Þlogð1  z1 Þ;

ð1  z H Þ2 X
n1
2 1 1
B ¼ zn  z1  ð n  j Þ   ...
n2 j¼1
1  zj 1  zjþ1

1  zH X
n1     
2 ðn  jÞ log 1  zj  log 1  zjþ1
n j¼1

1 ð1  zH Þ2 Xn
1
¼ zn  z1  ð1  zH Þ2 þ ð1 þ 2ðn  jÞÞ  ...
1  z1 n 2
j¼1
1  zj

1  zH X
n  
 2ð1  zH Þlogð1  z1 Þ þ 2 log 1  zj ;
n j¼1

C ¼ 1  zn :
596 A. Chernobai et al.

Summing the terms A, B, and C, multiplying by n^c , and simplifying yields the
final computing formula:
Xn   1  zH Xn
1
AD2 ¼ 2nlogð 1  z H Þ þ 2 log 1  z j þ ð1 þ 2ðn  jÞÞ :
up
j¼1
n j¼1
1  zj

References
Anderson, T. W., & Darling, D. A. (1952). Asymptotic theory of certain ‘Goodness of Fit’ criteria
based on stochastic processes. The Annals of Mathematical Statistics, 23, 193–212.
Anderson, T. W., & Darling, D. A. (1954). A test of goodness of fit. Journal of the American
Statistical Association, 49, 765–769.
BCBS. (2001). Working paper on the regulatory treatment of operational risk. Basel, Switzerland:
Basel Committee on Banking Supervision. http://www.bis.org
BCBS. (2003). The 2002 loss data collection exercise for operational risk: summary of the data
collected. Basel, Switzerland: Basel Committee on Banking Supervision. http://www.bis.org
D’Agostino, R., & Stephens, M. (1986). Goodness-of-fit techniques. New York/Basel: Marcel
Dekker.
Deheuvels, P., & Martynov, G. (2003). Karhunen-Loève expansions for weighted Wiener pro-
cesses and Brownian bridges via Bessel functions. Progress in Probability, 55, 57–93.
Dufour, R., & Maag, U. R. (1978). Distribution results for modified kolmogorov-smirnov statistics
for truncated or censored samples. Technometrics, 20, 29–32.
Gastaldi, T. (1993). A kolmogorov-smirnov test procedure involving a possibly censored or
truncated sample. Communications in Statistics: Theory and Method, 22, 31–39.
Guilbaud, O. (1998). Exact kolmogorov-type test for left-truncated and/or right-censored data.
Journal of American Statistical Association, 83, 213–221.
Ross, S. (2001). Simulation (3rd ed.). Boston: Academic.
Shorack, G., & Wellner, J. (1986). Empirical processes with applications to statistics.
New York: Wiley.
Effect of Merger on the Credit Rating and
Performance of Taiwan Security Firms 21
Suresh Srivastava and Ken Hung

Contents
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
21.2 Merger Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
21.3 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
21.3.1 Comprehensive Performance Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
21.3.2 Test of Merger Synergy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
21.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
Appendix 1: Variables for Merger Synergy Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Merger Synergy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
Principal Component Factor Analysis of Merger Synergies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
Variability Percentage Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Factor Loading Adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Performance Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Appendix 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
Wilcoxon Sign Rank Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
Test Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614

S. Srivastava (*)
University of Alaska Anchorage, Anchorage, AK, USA
e-mail: afscs@uaa.alaska.edu
K. Hung
Division of International Banking & Finance Studies, Texas A&M International University,
Laredo, TX, USA
e-mail: ken.hung@tamiu.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 597
DOI 10.1007/978-1-4614-7750-1_21,
# Springer Science+Business Media New York 2015
598 S. Srivastava and K. Hung

Abstract
The effect of a merger on credit rating is investigated by testing the significance
of change in a firm’s rank based on comprehensive performance score and
synergistic gains. We extract principle component factors from a set of financial
ratios. Percentage of variability explained and factor loadings are adjusted to get
a modified average weight for each financial ratio. This weight is multiplied by
the standardized Z value of the variable, and summed a set of variables to get
a firm’s performance score. Performance scores are used to rank the firm.
Statistical significance of difference in pre- and post-merger rank is tested
using the Wilcoxon sign rank (double end).
We studied the merger of financial firms after the enactment of Taiwan’s
Merger Law for Financial Institution in November 2000 to examine synergies
produced by merger. Synergistic gains affect corporate credit ratings. After
taking into account the large Taiwan market decline from 1999 to 2000, test
results show there is no significant operating, market, and financial synergy
produced by the merger firms. Most likely explanations for the insignificant
rank changes are short observation period and the lack of an adequate sample in
this investigation.
We identify and define variables for merger synergy analysis followed
by principal component factor analysis, variability percentage adjustment,
and performance score calculation. Finally, Wilcoxon sign rank test is used
for hypothesis testing. Reader is well referred to the appendix for details.

Keywords
Corporate merger • Financial ratios • Synergy • Economies of scale • Credit
rating • Variability percentage adjustment • Principle component factors • Firm’s
performance score • Standardized Z • Wilcoxon rank test

21.1 Introduction

Corporate credit rating helps in determining the soundness of a financial system.

It gives lenders and venture capitalists confidence in making direct investment
in domestic and foreign countries. Credit rating determines the probability that
the corporation will be able to meet its obligations. There are a number of credit
rating agencies worldwide.1 The top three agencies that deal in credit ratings are
Moody’s, Standard & Poor’s, and Fitch IBCA. Independent objective assessments
of the credit worthiness of companies help investors decide the riskiness of
the security issued by the firm. Credit rating institutions base their subjective

1
World leading credit rating agencies are Moody’s Investors Service, Standard & Poor’s Corp,
Fitch Investors Service, Duff & Phelps, Japan Bond Research Institution, Nippon Investors
Service, Japan Credit Rating Agency, China Credit Rating Agency, and Taiwan New Economy
Newspaper (Weston and Mansinghka 1971; Williamson 1981).
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 599

rating judgments by the experiences and expertise of their analyst teams.

Their public confidences come from their reputation. Yet different analysts or
different credit rating institutions may give different meanings or powers to
the likelihood of the same incident. Credit rating agencies often revise their
ratings after a corporate event. This research deals with the event of corporate
mergers.
A merger is the quickest method of corporate growth. At the end of the
twentieth century, there was an upsurge of mergers in Europe, the USA, and
Asia. The same was true for Taiwan. Mergers occurred in various industries of
Taiwan, with financial industry the most dominant. For example, the merger of
Bank of Taiwan, Local Bank, and the Central Credit Bureau was the first large
merger initiated by the government. The institution resulting from the merger
ranked in the top 70 in the world. Yuan Dao, the number one corporation in the
security industry, captured 10 % of the Taiwanese brokerage business after its
acquisition of Jing Hua. The Merger Law for Financial Institutions passed in
November 2000 and permitted the merger of domestic financial institutions, such
as banks, insurance companies, and security companies, and allowed for the
establishment of assets Management Corporation. After a merger, operating
income, market share, revenues, and total asset increase. But does a merger help
to improve corporate quality and competence? When does merger synergy mani-
fest? What are its merits? There are a number of questions that need discussion and
analysis. The scope of this paper is limited. We study the effect of mergers on credit
rating by testing the significance of change in firm’s rank based on comprehensive
performance score and synergistic gains.

21.2 Merger Literature

Firms merge with the stated intent of shareholders’ wealth maximization. Value
maximization is achieved by increased profit, reduced risk, or both. Theories
that posit wealth maximization are efficiency theory (synergy), information and
signaling theory, and tax advantage.2 Non-value maximization concerns with
agency theory. In this research we focus primarily on efficiency theory, also called
synergy theory. Synergy theory implies three components: operation synergy,
market synergy, and financial synergy. Operation synergy is produced due to
economies of scale, operating reduction, and sharing of management expertise.
Under market synergy, a merger reduces market competitors and increases market
concentration leading to a monopolization or increased market share. The increased
market share produces superior profits via pricing strategy or corporate collusion.
The merging corporation increases its market share so as to influence product price

2
Theory of merger is discussed in Brigham and Daves (2007), Copeland and Weston (1979),
Gardner (1996), and Rhodes (1983). Tax incentive is discussed in Weston and Chung (1983).
600 S. Srivastava and K. Hung

and quantity and to achieve market synergy. However, corporations in pursuit of

market synergy must abide by the fair trade law. Seth (1990) said it was easy to
achieve market synergy via horizontal merger instead of pricing strategy or
corporate collusion. Financial synergy reduces the systematic risk and lowers the
capital cost. There is extensive literature of merger-related studies. The following is
a selected sample.
Ansoff et al. (1971) used a questionnaire to collect 12 financial variables:
sales, retained earnings, total asset, return on capital, etc. Their research
showed that sales management, technology, and R&D produced operating
synergy. Weston and Mansinghka (1971) estimated management energy with
corporate growth using variables: total asset, sales, net profit, retained earnings,
and stock price. Their research indicates that a conglomerate merger gives the
corporation a higher growth rate than the growth of other corporations within
the same industry. Beattie (1980) used nonsystem risk and stock return to
estimate merger synergy. His finding is that after a merger, nonsystem risk
declines, yet corporate stock returns do not increase. Hoshino (1982) estimated
corporate synergy with seven financial data: the ratio of net worth and total
liabilities, ratio between net worth and total assets, liquidity ratio, ratio of
interest rate to debt, turnover ratio, ratio of net profit and total liabilities, and
ratio of net profit to total assets. His finding is that a merged corporation has
increased liquidity, but lower profitability and stability. Muller (1985) used
market share percentage to estimate the effect of merger. His finding is that
there is no significant change in market share; some companies even lose
market share after merger. Sigh and Montgomery (1987) used stock returns
to discuss related and non-related merger synergy. Their finding is that
a related merger has a higher return. Healy et al. (1992) used the ratio between
before-tax capital flow and total assets to estimate merger synergy. Their
finding is that the return on operating cash flow improves significantly.
Williamson (1981) stated that vertical merger could reduce the communication
cost between upper and lower stream corporations, cost of product quality
check, storage cost, and delivery cost. Yet Scherer (1980) thought a different
market configuration could impact the gains due to economies of scale.
The more complete the market competition is, the less is the gain due to
economies of scale. On the other hand, Rhodes (1983) thought that the
resources configuration within a corporation is complicated and
bureaucratic. Hence, the internal capital market is less efficient than the
external capital market. In support of financial synergy, Fluck and Lynch
(1999) pointed out that after a merger, a corporation could have greater
investment opportunities at lower cost than before. Lewellen (1971) suggested
the merger could reduce the capital cost. Levy and Sarnat (1970) thought
stockholders could achieve risk reduction via portfolio diversification at
a lower cost than that of the merger. Higgins and Schall (1975) pointed out
the coinsurance effect and hence a reduced cost of bankruptcy. Hence, there is
a transfer of wealth from creditors to stockholders from creditors, as stock-
holders could reissue bonds with a lower interest rate.
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 601

Table 21.1 Merged security firms

Merged security firms Survived corporation Date of announcement Date of merger
YuanDa, JinHua, Dafa YuanDaJinHua Nov. 29, 1999 July 01, 2000
YuanFu, JiaHe, YongSheng YuanFu Jan. 28, 1999 July 24, 2000
JianHong, WanSheng JianHong Feb. 11, 2000 Aug. 28, 2000
BaoLai, DaShun, ShiTai, BaoLai Feb. 19, 2000 Sep. 9, 2000
HuaYu
FuBang, HuangQiu, FuBang Feb. 21, 2000 Sep. 9, 2000
ZhongRi, JinShan,
HuaXin, ShiLing, KuaiLe
13 non-merger firms used to test the model are as follows: TaiYu, DaHua, QunYi, ZhongXin,
YongChang, TaiZhen, JinDing, RiShen, DaXin, KangHe, YaZhou, XinBao, and TongYi

21.3 Empirical Analysis

The effect of a merger on credit rating is investigated by (1) testing the significance
of change in a firm’s rank based on a comprehensive performance score and
(2) examining post-merger synergy. Synergistic gains from a merger could be in
the form of operating (management) synergy, market synergy, and/or financial
synergy. Data for this research cover January 1999 to December 2000. Premerger
analysis covers year 1999, and post-merger analysis covers year 2000.3 Definitions
of financial ratios are presented in Appendix 1. The methodology of factor extrac-
tion and weight assignment is discussed in Appendix 2. Since the analyses are based
on ranks, the usual “t-test” method is unsuitable to examine the significance of
change in operating, financial, or market performance pre- and post-merger. Hence,
we use the Wilcoxon Sign Rank Test that checks whether two sets of ranks, pre- and
post-merger, come from the same sample or two samples (Appendix 3).4
In Table 21.1, we list five acquiring financial institutions and the acquired institu-
tions. The footnote of Table 21.1 lists 13 non-merger firms. These 13 firms are
included in the analysis for comparison.

21.3.1 Comprehensive Performance Score

Principle component factor analysis was used to extract common factors with
factor loading greater than 1. Using 12-month financial ratio covering 1999,
we produced five factors (Table 21.2, Panel A). The variability explained by five
extracted factors is 24.458 %, 23.764 %, 13.268 %, 12 %, and 11.031 %.

3
3-month and 6-month analyses were performed but are omitted as these did not add additional
foresight.
4
References for Statistical and Econometric issues are Johnson and Wichern (2007) and
Woolridge (2009).
602 S. Srivastava and K. Hung

Table 21.2 Extracted factors and explained variability

Explained variability
Factor Factor loading Percentage Cumulative percentage Adjusted percentage
Panel A: December 1999
1 4.647 24.458 24.458 28.937
2 4.515 23.764 48.222 28.116
3 2.521 13.268 61.49 15.698
4 2.28 12 73.49 14.198
5 2.096 11.031 84.521 13.051
Total 16.059 84.521 84.521 100
Panel B: December 2000
1 4.245 22.34 22.34 24.746
2 3.844 20.229 42.569 22.407
3 2.991 15.74 58.309 17.435
4 2.377 12.51 70.819 13.857
5 2.011 10.584 81.403 11.724
6 1.686 8.876 90.279 9.832
Total 17.154 90.279 90.279 100.000
Principle component factors. See Appendix 2

The percentage of variability explained was adjusted to make the total variability
explained equal to 100. The adjusted percentage of variability explained by five
extracted factors is 28.937 %, 28.116 %, 15.698 %, 14.198 %, and 13.051 %. Using
12-month financial ratio covering 2000, six extracted common factors with factor
loading greater than 1 are presented in Table 21.2 (Panel B). The adjusted percent-
age of variability explained by six extracted factors is 24.746 %, 22.407 %,
17.435 %, 13.857 %, 11.724 %, and 9.832 %. The adjusted variability percentage
is multiplied by factor loading and summed over all variables to get the total
loading for a variable. Then factor loadings are adjusted such that total factor
loadings add to 100 (Eq. 21.2). Initial and adjusted weights for each variable are
presented in Table 21.3. Finally, the values of adjusted weights for 1999 and 2000
are averaged and sign modified to reflect a positive or negative variable. It is
presented in the last column of Table 21.3 and used to calculate comprehensive
performance score. The modified average weight for each variable is multiplied by
the standardized Z value of the variable and summed over all the variables to get
a firm’s performance score and comprehensive rank (Table 21.4). The premerger
comprehensive ranks of the five merger firms and 13 non-merger firms are listed in
Table 21.4, Panel A. Panel B of Table 21.4 presents comprehensive post-merger
comprehensive ranks. Rank changes in Table 21.4 can be summarized as JianHong
dropped from rank 1 to 6, YuanDaJinHua rose from rank 3 to 2, YuanFu
rose from rank 7 to 1, FuBang dropped from rank 2 to 4, and BaoLai rose from
rank 16 to 8. The test of significance of the comprehensive rank differences, for
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 603

Table 21.3 Adjusted variable weights for comprehensive performance analysis

December 1999 December 2000
Initial Adjusted Initial Adjusted Average
Item Variable weight weight weight weight weighta
Financial structure Xi1
0.254
9.010
0.192
4.404 6.707
Xi2 0.263 9.329
0.025
0.583 4.373
Solvency Xi3 0.388 13.741 0.321 7.358 10.549
Xi4 0.408 14.455 0.323 7.405 10.930
Asset utilization Xi5 0.127 4.489 0.310 7.107 5.798
Xi6
0.214
7.575 0.171 3.925 1.825
Profitability Xi7 0.333 11.806 0.396 9.077 10.442
Xi8 0.324 11.472 0.380 8.704 10.088
Xi9 0.327 11.573 0.366 8.399 9.986
Xi10 0.302 10.706 0.372 8.525 9.616
Cash flow Xi11
0.021
0.732 0.325 7.459 3.363
Xi12
0.042
1.487 0.174 3.977 1.245
Xi13 0.035 1.232 0.207 4.743 2.988
Growth Xi14 0.194 6.891 0.282 6.456 6.674
Xi15 0.214 7.575 0.193 4.434 6.005
Size Xi16 0.185 6.559 0.255 5.833 6.196
Xi17 0.265 9.398 0.273 6.263 7.831
Industry specific Xi18 0.289 10.229 0.201 4.604 7.417
Xi19
0.301
10.654 0.031 0.717 4.969
Total 2.821 100.000 4.363 100.000
Variable definitions are in Appendix 1
a
Sign of average variable weights is changed to reflect a positive or negative variable

merger and non-merger firms, was conducted using the Wilcoxon sign rank (double
end) test discussed in Appendix 3. Panel A of Table 21.5 indicates that the merger
firm’s comprehensive rank difference is statistically insignificant. Panel B of
Table 21.5 indicates that the non-merger firm’s comprehensive rank difference is
also statistically insignificant. Plausible explanations for the insignificant rank
changes are a short observation period, the lack of adequate samples, and the
sharp decline in JianHong’s performance score. Another plausible reason may be
that some security firms (such as YuanDaJinHua and YuanFu) have already
achieved high premerger comprehensive performance scores, and any improvement
in post-merger performance did not make a significant difference.

21.3.2 Test of Merger Synergy

We use 12-month data to examine operating synergy, market synergy, and financial
synergy. If operating synergy exists, then operating cost ratio will be reduced
604 S. Srivastava and K. Hung

Table 21.4 Security firm’s comprehensive performance score and rank

Comprehensive Comprehensive
Firm score Rank Firm score Rank
Panel A: December 1999
JianHonga 112.441 1 YongChang 4.045 10
FuBanga 90.757 2 ZhongXin 2.177 11
YuanDaJinHuaa 88.033 3 TaiYu
3.648 12
TongYi 57.312 4 DaHua
4.140 13
QunYi 47.620 5 KangHe
47.811 14
RiShen 47.457 6 JinDing
81.604 15
YuanFua 40.041 7 BaoLaia
106.382 16
TaiZhen 21.349 8 YaZhou
116.505 17
XinBao 12.180 9 DaXin
163.323 18
Panel B: December 2000
YuanFua 162.4621 1 TaiZhen
10.0671 10
YuanDaJinHuaa 132.0642 2 DaHua
19.123 11
RiShen 118.3488 3 KangHe
29.1604 12
FuBanga 77.26606 4 YaZhou
75.6848 13
ZhongXin 50.47429 5 YongChang
77.5209 14
JianHonga 48.7586 6 TaiYu
96.5883 15
QunYu 32.33424 7 JinDing
98.3886 16
BaoLaia 31.40838 8 XinBao
101.256 17
TongYi 16.56453 9 DaXin
150.395 18
a
means Wilcoxon sign rank test for performance score difference between before after merger is
statistically significant

significantly; if market synergy exists, then two other ratios (ratio of operating
income to total assets and market share change rate) will increase significantly; and
existence of financial synergy is indicated by the decline in variability of operating
(business) risk.
Some financial variables of security firms are greatly influenced by market
conditions. The operating income, in particular, is greatly influenced by a bull or
bear market. In years 1999 and 2000, the Taiwan stock market fell from over 10,000
points to over 5,000 points. In order to reduce the market impact on variables,
we adjust the operating cost to operating income ratio and the operating income to
total asset ratio. The operating cost ratio and the operating return on assets are
normalized by the industry average of the ratio.
Table 21.6 presents a test of significance of operating synergy for merger and
non-merger firms. The test result shows there are no significant changes of operat-
ing costs ratio for merger firms. But for non-merger firms, there are significant
increases of operating cost ratio. Then operating income is adjusted for market drop
in 2000, and the significance of operating synergy is tested again. The test result is
presented in Table 21.7.
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 605

Table 21.5 Test of significance of merger and non-merger firms’ comprehensive rating
December December
Firm 1999 ranks 2000 ranks Difference Di Class W
Panel A: merger firms
JianHong 1 6
5 5 3 W (+) ¼ 33
YuanDaJinHua 3 2 1 1 1 W (
) ¼ 45
YuanFu 7 1 6 6 4 W ¼ 33
FuBang 2 4
2 2 2
BaoLai 16 8 8 8 5
Panel B: non-merger firms
TaiYu 12 15
3 3 6.5 W (+) ¼ 33
DaHua 13 11 2 2 3.5 W (
) ¼ 45
QunYi 5 7
2 2 3.5 W ¼ 33
ZhongXin 11 5 6 6 11
YongChang 10 14
4 4 8.5
TaiZhen 8 10
2 2 3.5
JinDing 15 16
1 1 1
RiShen 6 3 3 3 6.5
DaXin 18 18 0
KangHe 14 12 2 2 3.5
YaZhou 17 13 4 4 8.5
XinBao 9 17
8 8 12
TongYi 4 9
5 5 10
Wilcoxonsign rank (double end) test indicates that rank difference is statistically insignificant,
Eq. 21.6

The test of significance of merger synergy is performed by examining two

financial ratios: ratio of operating income to total assets (operating return ratio)
and market share. Table 21.8 presents the first test of significance of market synergy
using operating return ratio. Result shows there are significant increases in the
ratio of operating income to total assets for both merger and non-merger
security firms. This indicates a positive market synergy. We repeat the test of
significance after adjusting the operating income to account for the market decline.
Results presented in Table 21.9 show there are insignificant differences in the
ratio of operating income to total assets for both merger and non-merger
security firms. The second test of significance of market synergy using market
share change is presented in Table 21.10. Results in Table 21.10 show there are
significant increases of market share for merger security firms. This means
positive market synergy. For the non-merger firms, there are significant declines
of market share. Taking the two variables into consideration, we state that
positive market synergy is produced by the merger of security firms. We use the
variability of operating risk to assess financial synergy. The test results in
Table 21.11 show there is insignificant financial synergy for merger firms.
However, there is significant change in the variability of operating risk for
non-merger security firms.
606 S. Srivastava and K. Hung

Table 21.6 Test of significance of merger and non-merger firms’ operating cost ratioa
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merged firms
JianHong 71.634 67.332 4.302 4.302 2 W (+) ¼ 3
YuanDaJinHua 58.689 78.167 –19.478 19.478 5 W (
) ¼ 12
YuanFu 71.857 79.305 –7.448 7.448 4 W¼3
FuBang 70.040 75.378
5.338 5.338 3
BaoLai 92.007 91.503 0.503 0.503 1
Panel B: non-merger firms
TaiYu 76.548 122.296
45.748 45.748 12 W (+) ¼ 3
DaHua 84.481 83.181 1.299 1.299 2 W (
) ¼ 88
QunYi 71.105 73.431
2.325 2.325 4 W¼3
ZhongXin 79.395 84.665
5.269 5.269 5
YongChang 79.471 111.708
32.238 32.238 11
TaiZhen 81.915 95.024
13.109 13.109 8
JinDing 98.436 107.020
8.584 8.584 6
RiShen 66.200 66.167 0.034 0.034 1
DaXin 121.563 143.634
22.071 22.071 10
KangHe 81.531 93.207
11.676 11.676 7
YaZhou 104.699 106.014
1.316 1.316 3
XinBao 74.861 136.984
62.123 62.123 13
TongYi 63.744 85.364
21.621 21.621 9
a
Operating cost/operating income
Wilcoxon sign rank (double end) test indicates that rank difference is statistically
insignificant, Eq. 21.6

21.4 Conclusion

The effect of a merger on credit rating was examined by testing the significance
of change in firm’s rank based on a comprehensive performance score
and examining post-merger synergy. Synergistic gains from a merger could be
in the form of operating synergy, market synergy, and/or financial synergy.
Our test showed that merger and non-merger firms’ comprehensive performance
rank difference was statistically insignificant. Plausible explanations for the
insignificant rank changes are a short observation period and the lack of adequate
samples. Other plausible explanation may be that some security firms have already
achieved high premerger comprehensive performance scores, and any improvement
in post-merger performance did not make a significant difference.
We used a standardized score of operating cost ratio to rank firms for their
operating synergy. Test results show there is no significant operating synergy for
merger and non-merger security firms. We used standardized scores of ratio of
operating income to total assets and market share change to rank firms for their
market synergy. Test result shows there are significant increases in the ratio of
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 607

Table 21.7 Test of significance of merger and non-merger firms’ adjusted operating cost ratioa
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merger firms
JianHong 93.492 79.475 14.017 14.017 4 W (+) ¼ 10
YuanDaJinHua 76.597 92.263
15.667 15.667 5 W (
) ¼ 5
YuanFu 93.783 93.607 0.176 0.176 1 W¼5
FuBang 91.412 88.972 2.440 2.440 2
BaoLai 120.081 108.005 12.076 12.076 3
Panel B: non-merger firms
TaiYu 99.905 144.351
44.446 44.446 12 W (+) ¼ 32
DaHua 110.259 98.183 12.076 12.076 9 W (
) ¼ 59
QunYi 92.802 86.674 6.129 6.129 5 W ¼ 32
ZhongXin 103.622 99.933 3.688 3.688 3
YongChang 103.720 131.854
28.134 28.134 11
TaiZhen 106.911 112.161
5.250 5.250 4
JinDing 128.472 126.320 2.152 2.152 1
RiShen 86.400 78.099 8.301 8.301 6
DaXin 158.656 169.537
10.881 10.881 7
KangHe 106.409 110.016
3.608 3.608 2
YaZhou 136.646 125.133 11.513 11.513 8
XinBao 97.704 161.688
63.985 63.985 13
TongYi 83.194 100.759
17.565 17.565 10
a
Operating cost/operating income adjusted for the drop in the market
Wilcoxon sign rank (double end) test indicates that rank difference is statistically
insignificant, Eq. 21.6

operating income to total assets for both merger and non-merger security firms,
indicating a positive market synergy. However, after adjusting the operating
income market decline, the results indicate insignificant differences in the ratio of
operating income to total assets for both merger and non-merger security
firms. We used standardized score of variability of operating risk to rank firms
for their financial synergy. The test results show there is insignificant financial
synergy for merger firms; however, there is significant change in the variability of
operating risk for non-merger security firms.

Appendix 1: Variables for Merger Synergy Analysis

Post-merger credit rating of the firm depends on the extent of synergy produced by
the merger. Components of merger synergy are operating synergy, market synergy,
and financial synergy.
Each synergy component is determined by firm characteristics: financial
structure, solvency, asset utilization, profitability, cash flow, growth, scale, and
industry-specific ratio. In this study a number of financial ratios are used to assess
608 S. Srivastava and K. Hung

Table 21.8 Test of significance of merger and non-merger firms’ operating return on assetsa
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merger firms
JianHong 16.868 17.171
0.303 0.303 1 W (+) ¼ 0
YuanDaJinHua 13.591 22.133
8.542 8.542 3 W (
) ¼ 15
YuanFu 14.667 26.617
11.949 11.949 4 W¼0
FuBang 14.245 16.876
2.630 2.630 2
BaoLai 12.429 25.243
12.814 12.814 5
Panel B: non-merger firms
TaiYu 12.881 15.878
2.997 2.997 12 W (+) ¼ 10
DaHua 15.783 12.746 3.037 3.037 2 W (
) ¼ 81
QunYi 13.893 18.924
5.031 5.031 4 W ¼ 10
ZhongXin 13.214 26.546
13.332 13.332 5
YongChang 16.797 19.702
2.905 2.905 11
TaiZhen 12.783 14.188
1.405 1.405 8
JinDing 11.884 15.938
4.054 4.054 6
RiShen 18.111 24.489
6.379 6.379 1
DaXin 15.232 15.240
0.008 0.008 10
KangHe 14.146 16.561
2.414 2.414 7
YaZhou 16.653 16.268 0.385 0.385 3
XinBao 13.316 15.023
1.706 1.706 13
TongYi 13.732 21.170
7.438 7.438 9
a
Operating income/total assets
Wilcoxon sign rank (double end) test indicates that ratio difference is statistically significant,
Eq. 21.6

operating performance of the firm before and after the merger. The following
section outlines firm characteristics and variables to measure operating
performance.

Merger Synergy

Operating synergy – it refers to the improvement of operating efficiency achieved

via scale economy, transaction cost economy, and differential efficiency caused by
merger. Market synergy – increase in market share due to enhanced negotiating
power and dominant pricing strategy.
Financial synergy – diversification of financial risk and cost of capital reduction.

Operating Synergy

A. – Financial structure
Xi1 – Debt ratio
Xi2 – Ratio of long-term capital to fixed assets
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 609

Table 21.9 Test of significance of merger and non-merger firms’ adjusted operating return on
assetsa
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merger firms
JianHong 0.0190 0.0154 0.0035 0.0035 2 W (+) ¼ 3
YuanDaJinHua 0.0153 0.0199
0.0046 0.0046 3 W (
) ¼ 12
YuanFu 0.0165 0.0239
0.0074 0.0074 4 W¼3
FuBang 0.0160 0.0152 0.0008 0.0008 1
BaoLai 0.0140 0.0227
0.0087 0.0087 5
Panel B: non-merger firms
TaiYu 0.0145 0.0143 0.0002 0.0002 1 W (+) ¼ 53
DaHua 0.0177 0.0115 0.0063 0.0063 12 W (
) ¼ 38
QunYi 0.0156 0.0170
0.0014 0.0014 5 W ¼ 33
ZhongXin 0.0149 0.0239
0.0090 0.0090 13
YongChang 0.0189 0.0177 0.0012 0.0012 4
TaiZhen 0.0144 0.0128 0.0016 0.0016 7
JinDing 0.0134 0.0143
0.0010 0.0010 2
RiShen 0.0204 0.0220
0.0017 0.0017 8
DaXin 0.0171 0.0137 0.0034 0.0034 9
KangHe 0.0159 0.0149 0.0010 0.0010 3
YaZhou 0.0187 0.0146 0.0041 0.0041 11
XinBao 0.0150 0.0135 0.0015 0.0015 6
TongYi 0.0154 0.0190
0.0036 0.0036 10
a
(Operating income/total assets) adjusted for the drop in the market
Wilcoxon sign rank (double end) test indicates that ratio difference is statistically insignificant, Eq. 21.6

B. Solvency
Xi3– Current ratio
Xi4– Quick ratio
C. Asset utilization
Xi5 – Operating return on assets
Xi6 – Net worth turnover ratio
D. Profitability
Xi7 – Return on assets
Xi8 – Return on equity
Xi9 – Profit margin
Xi10 – Earnings per share
E. Cash flow
Xi11 – Cash flow to short-term liability
Xi12 – 5-year cash flow to debt obligations
Xi13 – Retention ratio
F. Growth
Xi14 – Growth in revenue
Xi15 – Growth in earnings
610 S. Srivastava and K. Hung

Table 21.10 Test of significance of merger and non-merger firms’ change in market share
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merger firms
JianHong 6.169 5.255 0.914 0.914 1 W (+) ¼ 1
YuanDaJinHua 11.463 17.270
5.807 5.807 5 W (
) ¼ 14
YuanFu 5.992 7.497
1.505 1.505 2 W¼1
FuBang 6.007 7.698
1.692 1.692 3
BaoLai 5.745 8.076
2.331 2.331 4
Panel B: non-merger firms
TaiYu 1.982 1.475 0.507 0.507 5 W (+) ¼ 80
DaHua 9.589 5.589 4.000 4.000 13 W (
) ¼ 11
QunYi 8.389 8.197 0.192 0.192 1 W ¼ 11
ZhongXin 4.932 6.402
1.470 1.470 11
YongChang 3.090 2.559 0.531 0.531 6
TaiZhen 5.462 4.929 0.532 0.532 7
JinDing 5.128 4.678 0.450 0.450 4
RiShen 9.547 7.487 2.061 2.061 12
DaXin 3.109 1.830 1.280 1.280 10
KangHe 2.580 2.209 0.371 0.371 3
YaZhou 2.300 1.462 0.838 0.838 8
XinBao 1.431 1.178 0.253 0.253 2
TongYi 7.085 6.210 0.876 0.876 9
Wilcoxon sign rank (double end) test indicates that ratio difference is statistically significant,
Eq. 21.6

G. Size
Xi16 – Total assets
Xi17 – Net worth
H. Industry-specific ratios
Xi18 – Consignment to current assets
Xi19 – Long-term financing to net worth

Market Synergy

Xi5 – Operating return on assets turnover

Xi20 – Market share variability

Financial Synergy

Xi21 – Operating risk variability

21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 611

Table 21.11 Test of significance of merger and non-merger firms’ operating risk
December December
Firm 1999 score 2000 score Difference Di Class W
Panel A: merger firms
JianHong 5250.000
8.178 5258.178 5258.178 5 W (+) ¼10
YuanDaJinHua 42.250
29.701 71.951 71.951 2 W (
) ¼ 5
YuanFu 10.324 14.973
4.649 4.649 1 W¼5
FuBang 1017.647
17.895 1035.542 1035.542 3
BaoLai
93.671 1010.000
1103.671 1103.671 4
Panel B: non-merge firms
TaiYu 1274.074
205.391 1479.465 1479.465 11 W (+) ¼ 70
DaHua 127.941
54.839 182.780 182.780 7 W (
) ¼ 21
QunYi 1657.895 1.198 1656.697 1656.697 12 W ¼ 21
ZhongXin 59.877
3.089 62.965 62.965 3
YongChang
249.479
216.376
33.103 33.103 1
TaiZhen
230.769
96.078
134.691 134.691 6
JinDing
101.587
7566.667 7465.079 7465.079 13
RiShen
29.612 67.241
96.853 96.853 4
DaXin
58.108 127.016
185.124 185.124 8
KangHe 228.378
72.840 301.218 301.218 9
YaZhou
50.000 0.000
50.000 50.000 2
XinBao 413.333
292.208 705.541 705.541 10
TongYi 64.151
37.701 101.852 101.852 5
Wilcoxon sign rank (double end) test indicates that ratio difference is statistically insignificant,
Eq. 21.6

Appendix 2

Principal Component Factor Analysis of Merger Synergies

Principal component factor analysis is used to examine merger synergies. It is

a multivariable statistic method focusing on the relationship between groups of
variables. Its purpose is to express the original data structure with fewer factors
while keeping most of the information provided by the original data structure.
Factor analysis is composed of two parts: one is common factor, and the other is
specific factor. Factor analysis intends to group the variables with same common
factors. In other words, it discusses how to break down every variable Xi of
P variables X1XP into the linear combination of q common factors fj(q, and
qp), j ¼ 1, 2 . . .q and specific factor ei. The model is as follows:
X1 ¼ m1 þ L11 F1 þ L12 F2 þ    þ L1q Fq þ e1
X2 ¼ m2 þ L21 F1 þ L22 F2 þ    þ L2q Fq þ e2
(21.1)
⋮
Xp ¼ mp þ Lp1 F1 þ Lp2 F2 þ    þ Lpq Fq þ ep
612 S. Srivastava and K. Hung

where F1 . . ., Fq are common factors, ei is specific factor for variable Xi, and Lij is
the factor loading of variable Xi and common factor Fj. Factors extracted are
independent and also analysis preserves the information in the original variables.
There is no overlap of information among principle components. The model should
be parsimonious, in the sense that a few principle components should be able to
replace the original set of variables.

Variability Percentage Adjustment

All principle components with factor loading greater than 1 are selected. Then
weights are assigned to different variables based on percentage of variability
explained. The variability percentage adjustment is adjusted as follows:

V adj
j ¼ Vj =Sj V j (21.2)

where Vj (j ¼ 1,2. . .q) is the percentage of variability explained by factor Fj.

Factor Loading Adjustment

The adjusted variability percentage Vadj j is multiplied by factor loading Lij and
summed over all variables (j ¼ 1,2. . .q) to get the total loading fo variable Xi. Then
adjust Li such that total factor loadings add to 100.
X
Li ¼ V adj
j Lij
 X  (21.3)
Ladj
i ¼ 100 Li = Li

Every variable has a weight, Ladji before and after the estimation period; we average
it to obtain weight Wi. For some variables, Xi, the greater the value of the variable,
the better it is for the operating, financial, or merger synergy. Those variables are
classified as positive variables. If opposite is true, then those variables are classified
as negative variable. Hence adjusted variable weights are redesigned to correctly
reflect operating synergy score: Wi* ¼ {Wi or –Wi} for positive and negative
variables, respectively.

Performance Scores

All variables are not measured in the same unit, so they are standardized as
Xi ¼ {Xi–Ave(Xi)}/s, where Ave(Xi) is the average and s is the standard
deviation of variable Xi (i ¼ 1,2. . .p). The adjusted variable weight Wi* multiplied
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 613

by standardized variable Zi gives the performance score of variable Xi. Appropriate

standardized performance score of the variable is used to rank the firm for its
operating or financial or market synergy. The sum of the standardized performance
scores over the set of variables gives the comprehensive performance score:

P
S¼ W i : Z i (21.4)

where the summation is over the first 19 variables listed in Appendix 1. Then firms
are ranked in terms of their respective comprehensive performance score. The
greater the total score is, the better is the comprehensive performance rating. On
the other hand, the smaller the total score, the worse is the performance rating and
lower is its rank.

Appendix 3

Wilcoxon Sign Rank Test

Usual “t-test” method is unsuitable to examine the significance of change in

operating, financial, or market performance before and after merger as we deal
with ranks. We use the Wilcoxon sign rank test which checks whether two sets of
ranks, pre- and post-merger, come from the same sample or two samples. Let D be
the difference between observed value from the sample and reference value. Then
we delete those observations whose value of D is zero and rank incrementally the
rest of observations in terms of the absolute value of D. If two or more absolute
values are the same, we give each value an appropriate rank, then average those
ranks, and use the averaged rank as the rank of the same absolute values.
The statistic analysis method of test is as follows:
The differences between the post-merger synergy ranks (Xi) and corresponding
premerger synergy ranks (Yi), Di is ranked in descending order. Let Ri be the serial
number of Di (if they have the same rank, then take their average value).

Di ¼ Xi
Yi i ¼ 1, 2,   n
Ri ¼ rankðjDi jÞ i ¼ 1, 2,   n
X
Wð þ Þ ¼ Ri Xi
Yi > 0 (21.5)
X
Wð
Þ ¼ Ri Xi
Yi < 0
W ¼ minðW þ , W
Þ

where W(+) is the total of absolute value of serial numbers of positive rank changes
and W(
) is the total of absolute value serial numbers of negative rank changes.
W is the test statistic.
614 S. Srivastava and K. Hung

Test Hypotheses

For comprehensive performance ranks:

H0: Performance rating after merger ¼ performance rating before merger
(Z1 ¼ Z2).
H1: Performance rating after merger 6¼ performance rating before merger
(Z1 6¼ Z2).
For operating, market, and financial synergy:
H0 ¼ No synergy occurred after merger (Z1 ¼ Z2).
H1 ¼ Synergy occurred after merger (Z1 6¼ Z2).
We undertake a double end test. Under the significant level of a, we find
the critical value W(a) from the appropriate statistical table. The null hypothesis
is rejected if

W  ð
Wða=2ÞÞ or
(21.6)
W  ðWða=2ÞÞ

This means that merger has produced significant synergy and performance
(and credit) rating has affected. The appropriate variables for operating, market,
and financial variables are operating cost ratio, ratio of operating income to total
assets and market share, and variability of operating risk, respectively.

References
Ansoff, H. I., Brandenburg, R. G., Portner, F. E., & Radosevich, R. (1971). Acquisition behavior of
U.S. manufacturing firms (pp. 1946–1965). Nashville: Vanderbilt University Press.
Beattie, D. L. (1980). Conglomerate diversification and performance: A survey and applied
economics. Applied Economics, 12, 251–273.
Brigham, E. F., & Daves, P. (2007). Intermediate financial management. Mason: Thomson South
West.
Copeland, T. E., & Weston, J. F. (1979). Financial theory and corporate policy. Reading:
Addison-Wesley.
Fluck, Z., & Lynch, A. W. (1999). Why do firms merge and then divest? A theory of financial
synergy. Journal of Business, 72, 319–346.
Gardner, D. C. (1996). Introduction to mergers and acquisitions. London: Pitman Publishing.
Healy, P. M., Palepu, K. G., & Ruback, R. S. (1992). Does corporate performance improved after
mergers? Journal of Financial Economics, 31, 135–175.
Higgins, R. C., & Schall, L. D. (1975). Corporate bankruptcy and conglomerate merger.
The Journal of Finance, 30, 93–119.
Hoshino, Y. (1982). The performance of corporate mergers in Japan. Journal of Business
Financial and Accounting, 9, 145–163.
Johnson, R., & Wichern, D. (2007). Applied multivariate statistical analysis (6th ed.). Upper
Saddle River: Pearson Education.
Levy, H., & Sarnat, M. (1970). Diversification, portfolio analysis and the uneasy case for
conglomerate mergers. The Journal of Finance, 25, 795–802.
Lewellen, W. G. (1971). A pure financial rationable for conglomerate merger. The Journal
of Finance, 26, 521–537.
21 Effect of Merger on the Credit Rating and Performance of Taiwan Security Firms 615

Muller, D. C. (1985). Mergers and market share. The Review of Economics and Statistics, 67,
66–74.
Rhodes, S. A. (1983). Power, empire building, and mergers. Lexington: Lexington Books.
Scherer, F. M. (1980). Industrial market structure and economic performance. Boston: Houghton
Mifflin.
Seth, A. (1990). Value creation in acquisitions: A re-examination of performance issues. Strategic
Management Journal, 11, 90–115.
Singh, P., & Montegomery, T. (1987). Using simulated mergers to evaluate corporate diversifica-
tion strategies. Strategic Management Journal, 8, 377–386.
Standard & Poor’s Corporation. (1996). Standard & Poor’s corporation ratings criteria.
New York: McGraw Hill.
Weston, J. F., & Chung, K. S. (1983). Do merger make money? A reach summary. Mergers &
Acquisitions, 18, 40–48.
Weston, J. F., & Mansinghka, S. K. (1971). Test of the efficiency: Performance of conglomerate
firms. The Journal of Finance, 26, 919–939.
Williamson, O. E. (1981). The modern corporation: Origins, evolution, attributes. Journal
of Economic Literature, 14, 1537–1568.
Woolridge, J. (2009). Introductory econometrics: A modern approach (4th ed.). Mason: Cengage
Learning.
On-/Off-the-Run Yield Spread Puzzle:
Evidence from Chinese Treasury Markets 22
Rong Chen, Hai Lin, and Qianni Yuan

Contents
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
22.2 Bond Pricing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
22.2.1 On-the-Run Bond Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
22.2.2 Off-the-Run Bond Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
22.3 Data and Empirical Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
22.3.1 Data Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
22.3.2 Empirical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
22.3.3 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
22.3.4 Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
22.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
Appendix 1: Nonlinear Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
Appendix 2: Matlab Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
Codes for the On-the-Run Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
Codes for the Off-the-Run Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637

Abstract
In this chapter, we document a negative on-/off-the-run yield spread in Chinese
Treasury markets. This is in contrast with a positive on-/off-the-run yield spread
in most other countries and could be called an “on-/off-the-run yield spread
puzzle.” To explain this puzzle, we introduce a latent factor in the pricing of
Chinese off-the-run government bonds and use this factor to model the yield

We thank the National Natural Science Foundation of China (Grant No. 71101121 and
No. 70971114) for financial support.
R. Chen (*) • Q. Yuan
Department of Finance, Xiamen University, Xiamen, China
e-mail: aronge@xmu.edu.cn; qnyuan@gmail.com
H. Lin
School of Economics and Finance, Victoria University of Wellington, Wellington, New Zealand
e-mail: hai.lin@vuw.ac.nz; cfc1080@gmail.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 617
DOI 10.1007/978-1-4614-7750-1_22,
# Springer Science+Business Media New York 2015
618 R. Chen et al.

difference between Chinese on-the-run and off-the-run issues. We use the

nonlinear Kalman filter approach to estimate the model. Regressions results
suggest that liquidity difference, market-wide liquidity condition, and disposi-
tion effect (unwillingness to sell old bonds) could help explain the dynamics of
a latent factor in Chinese Treasury markets. The empirical results of this chapter
show evidence of phenomena that are quite specific in emerging markets such as
China.
The Kalman filter is a mathematical method named after Rudolf E. Kalman.
It is a set of mathematical equations that provides an efficient computational
(recursive) means to estimate the state of a process, in a way that minimizes the
mean of the squared error. The nonlinear Kalman filter is the nonlinear version
of the Kalman filter which linearizes about the current mean and covariance.
The filter is very powerful in several aspects: it supports estimations of past,
present, and even future states, and it can do so even when the precise nature of
the modeled system is unknown.

Keywords
On-/off-the-run yield spread • Liquidity • Disposition effect • CIR model •
Nonlinear Kalman filter • Quasi-maximum likelihood

22.1 Introduction

It is well known that there exists an on-the-run phenomenon in worldwide Treasury

markets. This phenomenon refers to the fact that just-issued (on-the-run or new)
government bonds of a certain maturity are generally traded at a higher price or
lower yield than previously issued (off-the-run or old) government bonds maturing
on similar dates. For example, Amihud and Mendelson (1991), Warga (1992),
Kamara (1994), Furfine and Remolona (2002), Goldreich et al. (2005), and
Pasquariello and Vega (2009) report the existence of positive on-/off-the-run
yield spread in the US Treasury market with different frequency data. Mason
(1987) and Boudoukh and Whitelaw (1991, 1993) provide similar evidence in
Japan. In spite of different opinions on the information content of the yield spread,
there is no disagreement in the literature that the on-/off-the-run yield spread in
Treasury markets is significantly positive.
Academics have proposed many theories to explain the positive on-/off-the-run
yield spread. Early studies directly attribute this spread to the liquidity difference
between new bonds and old bonds (Amihud and Mendelson 1991; Warga 1992;
Kamara 1994). More recent work provides some other possible explanations, such
as different tax treatment (Strebulaev 2002), specialness in the repo market1
(Krishnamurthy 2002), the value of future liquidity (Goldreich et al. 2005), search

1
Specialness refers to the phenomenon that loans collateralized by on-the-run bonds offer lower
interest rates than their off-the-run counterparts in repo markets.
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 619

costs (Vayanos and Weill 2008), and market frictions of information heterogeneity
and imperfect competition among informed traders (Pasquariello and Vega 2009).
No matter what arguments are proposed, however, the important role that liquidity
plays in the positive on-/off-the-run yield spread and the liquidity premium
hypothesis (Amihud and Mendelson 1986) has never been denied. It is widely
accepted, by both practitioners and academics, that off-the-run bonds with a lower
liquidity level tend to have a higher yield than otherwise similar, yet more liquid,
on-the-run bonds.
In this chapter, we document a negative on-/off-the-run yield spread in Chinese
Treasury markets, which is contrary to the usual on-the-run phenomenon in other
countries and could be called the on-/off-the-run yield spread puzzle in China. Guo and
Wu (2006) and Li and He (2008) report that on-/off-the-run yield spread in Chinese
Treasury markets is significantly positive, but they did not match the on-the-run and
off-the-run bonds correctly. For example, they compare the yield of a just-issued
7-year government bond with that of a previously issued 7-year government bond
that has a different maturity. This is not consistent with the calculation of the usual on-/
off-the-run yield spread. In order to calculate the spread correctly, we need to match
the bonds in terms of maturity date. For example, we compare a just-issued 1-year
government bond and a previously issued government bond maturing on similar dates.
The maturities are both about 1 year and the durations are close to each other.
To explain on-/off-the-run yield spread puzzle, we introduce a latent factor in the
pricing of Chinese off-the-run bonds. This latent factor is used to model the yield
difference between on-the-run bonds and off-the-run bonds. We employ a nonlinear
Kalman filter to estimate the model and examine the temporal properties of the latent
factor. We find that the liquidity premium hypothesis still holds in Chinese Treasury
markets. In particular, the change of the latent factor is positively related to the
liquidity difference between off-the-run and on-the-run bonds and positively related
to the market-wide liquidity condition. Both findings are consistent with the liquidity
premium hypothesis. On the other hand, disposition effect (unwillingness to sell old
bonds in bear markets) dramatically changes the sign of the yield spread and causes
the puzzle. The change of latent factor in Chinese Treasury markets is negatively
related to 7-day repo rates. When interest rates go up and the returns of the bond
markets are negative, the holders of off-the-run bonds are reluctant to realize loss and
will not sell their bonds, which consequently leads to a relatively low yield level of
off-the-run bonds and a negative on-/off-the-run yield spread.
Our article makes several contributions to the literature. We document
a negative on-/off-the-run yield spread in China. We introduce a latent factor to
explain the yield difference between on-the-run bonds and off-the-run bonds and
employ the nonlinear Kalman filter in estimation. Our basic ideas are in line with
Longstaff et al. (2005) and Lin et al. (2011). We also provide evidence of irrational
investor behavior that is quite specific in emerging Treasury markets such as China.
In China, the liquidity premium hypothesis still holds, whereas the existence of
disposition effect causes the puzzle.
The chapter is organized as follows. In Sect. 22.2, we present a pricing model of
government bonds that introduces a latent factor for the off-the-run bonds.
620 R. Chen et al.

In Sect. 22.3, we describe the data, report the estimation results, and perform
a variety of regression analyses. We present our conclusions in Sect. 22.4.

22.2 Bond Pricing Models

22.2.1 On-the-Run Bond Pricing Model

We use the Cox et al. (1985, CIR) model to price the Chinese on-the-run govern-
ment bonds. The CIR model has been a benchmark interest rate model because of
its analytical tractability and other good properties. In this model, the risk-free short
rate, rt, is assumed to follow a square-root process as
pffiffiffiffi
dr t ¼ kðy  r t Þdt þ sr r t dW r, t , (22.1)

under the risk-neutral measure Q. k is the speed of mean reversion, y is the long-
term mean value, sr is the volatility parameter of rt, and Wr denotes a standard
Brownian motion under Q. Such specification allows for both mean reversion and
conditional heteroskedasticity and guarantees that interest rates are nonnegative.
At time t, the price of an on-the-run government bond maturing at tM could be
written as
"  ð tm #
X
M

t ¼ Et
Pon Cm exp  r s ds ,
Q
(22.2)
m¼1 t

where Ptonis the on-the-run bond price, Cm is the cash flow payments at time tm, and
M is the total number of cash flow payments. That is, the price of a government bond is
the expected present value of the cash flow payments under the risk-neutral measure.
Solving Eq. 22.2 gives

X
M  
t ¼
Pon Cm Am, t exp Bm, t r t , (22.3)
m¼1
where

 2ky=s2r
2h exp fðk þ hÞðtm  tÞ=2g
Am , t ¼ ,
2h þ ðk þ hÞðexp fðtm  tÞhg  1Þ

2ðexp fðtm  tÞhg  1Þ

Bm , t ¼ ,
2h þ ðk þ hÞðexp fðtm  tÞhg  1Þ

and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h¼ k2 þ 2s2r :
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 621

22.2.2 Off-the-Run Bond Pricing Model

In order to model the on-/off-the-run yield spread in Chinese Treasury markets,

we next incorporate a latent component, l, into the pricing model of Chinese
off-the-run government bonds and extend (22.2) to

"  ð tm #
X
M
Poff
t ¼ EQ
t Cm exp  ðr s þ ls Þds , (22.4)
m¼1 t

where Ptoff is the off-the-run bond price. Similar to Longstaff et al. (2005) and Lin
et al. (2011), we assume that under the risk-neutral measure Q,

dlt ¼ sl dW l, t , (22.5)

where Wl is a standard Brownian motion independent of Wr under Q and sl is the

volatility parameter.
Given the stochastic processes in (22.1) and (22.5), we can obtain the analytical
solution for the pricing formula of (22.4),

X
M  
Poff
t ¼ Cm Am, t exp Dm, t  Bm, t r t  ðtm  tÞlt , (22.6)
m¼1

s2l ðtm tÞ

where Dm, t = 6 and other notations are the same as in (22.3).

22.3 Data and Empirical Estimation

22.3.1 Data Summary

We use the price of Chinese government bonds in the interbank market to

estimate the pricing model. The data are from the RESSET dataset. There are
two main bond markets in China. One is the interbank bond market, while the
other is the exchange bond market. The interbank bond market is a quote-
driven over-the-counter market, and the participants are mainly institutional
investors. Its outstanding value and trading volume account for over 90 % of
Chinese bond markets. The number of bonds traded in the interbank market is
at least three times that traded in the exchange market. This is very important
for our empirical study, since we need enough bonds to match new and
old ones.
622 R. Chen et al.

Table 22.1 Summary statistics

Off-the-run On-the-run
Mean Std Mean Std Difference
One year
Yield (%) 2.24 0.92 2.31 0.72 0.07
Modified duration (years) 0.68 0.25 0.73 0.22 0.05
Age (years) 4.3 2.92 0.26 0.22 4.04a
Coupon (%) 2.95 2.73 1.39 1.46 1.63a
Three years
Yield (%) 2.72 0.78 2.83 0.79 0.11b
Modified duration (years) 2.47 0.33 2.48 0.31 0.01
Age (years) 3.05 1.45 0.38 0.32 2.67a
Coupon (%) 3.39 2.32 2.89 0.62 0.50
This table reports the summary statistics of the Chinese 1-year and 3-year on-the-run and off-the-
run government bonds between December 2003 and February 2009. The coupon rate and yield are
in percentages, while age and modified duration are in years. This table also reports the difference
between the off-the-run bonds and the on-the-run bonds
a
andb indicate statistical significance at the 5 % and 1 % level, respectively

Our sample period is from December 2003 to February 2009. We use monthly
data and choose the actively traded government bonds with 1 year and 3 years to
maturity. Thus, for each month in our sample period, a most recently issued 1-year
and 3-year government bonds are selected as the on-the-run bonds, and we get
another old bond maturing on similar dates to match each on-the-run bond.2
Altogether, 63 matched pairs of 1-year government bonds and 63 matched pairs
of 3-year government bonds are included in the final sample.
Table 22.1 reports the summary statistics for the sample bonds. As expected, the
modified durations of off-the-run and on-the-run bonds are quite close to each other.
This means if interest rate risk is the only risk factor, these bonds should be traded at
similar yields. To examine whether the same on-the-run phenomenon exists in
Chinese Treasury markets, we compute the on-/off-the-run yield spread as

DyM, t ¼ yoff
M, t  yM, t ,
on
(22.7)

where yoff on
M;t and yM;t are the time t yield of the off-the-run bond and the on-the-run
bond maturing at tM, respectively. In our sample, tM  t is equal to 1 year or 3 years.
Figure 22.1 plots the time series of DyM,t of the 1-year bond and 3-year bond.
Table 22.1 also reports the means of on-/off-the-run yield spreads and their
statistical significance. Both the 1-year on-/off-the-run yield spread and the 3-year
on-/off-the-run yield spread are negative, which is inconsistent with the findings in
other markets. Moreover, the 3-year on-/off-the-run yield spread is significantly

2
The main reason for the data selection comes from the concern of trading activity. Trading in
Chinese Treasury markets is not active, especially in the earlier period.
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 623

2.00%

1.50% One-year Three-year

1.00%

0.50%

0.00%

−0.50%

−1.00%

−1.50%

−2.00%

−2.50%

Fig. 22.1 Time series of Chinese on-/off-the-run yield spread. This figure plots the yield
difference between the off-the-run and on-the-run Chinese government bonds between December
2003 and February 2009

negative at the 10 % level, which suggests the existence of a significantly negative

on-/off-the-run yield spread and remains a puzzle. The 1-year on-/off-the-run yield
spread is negative but not significant. However, this spread has some noise,
since the difference of coupon rate between the on-the-run issues and off-the-run
issues is 1.63 % and significant at the 1 % level. This difference could affect
the significance of the on-/off-the-run spread at the 1-year level. Generally
speaking, we find evidence of a negative on-/off-the-run spread, which is contrary
to the established fact of a positive on-/off-the-run yield spread in most other
countries and could be called the on-/off-the-run yield spread puzzle in China.

22.3.2 Empirical Methodology

To explain the on-/off-the-run yield spread puzzle in China, we use the CIR
model to price the on-the-run issues and the CIR model with the latent factor to
price the off-the-run issues. We first estimate the parameters of the CIR model
using all on-the-run bonds. Given the parameters of the CIR process, we further
estimate the parameters of the latent factor using off-the-run bonds. Thus, the latent
factor represents the yield difference between on-the-run and off-the-run bonds.
In our empirical study, we employ the Kalman filter to estimate the parameters.3
The standard Kalman filter is not appropriate here because it requires linear state
functions and the measurement functions, while Eqs. 22.3 and 22.6 are nonlinear.

3
See Hamilton (1994) for an explanation of Kalman filter.
624 R. Chen et al.

Table 22.2 Estimates of pricing models

Off-the-run issues
pffiffiffiffi
On-the-run issues dr t = kðy  r t Þdt þ sr r t dW r, t
pffiffiffiffi
dr t = kðy  r t Þdt þ sr r t dW r, t dlt ¼ sldWl,t
The state function rt ¼ g + frt  1 + et r t = g þ fr t1 þ et lt = lt1 þ sl ðt  t1 Þ
The measurement yon t ¼ at + bt rt + vt t ¼ at + bt rt + jt lt + vt
on on on
yoff off off off off

function
k 0.089(6.403)a
y 0.023(6.486)a
sr 0.063(247.921)a
sl 0.009(8.530)a
var(et) 0.000008
var(t) 0.000007
One year Three years One year Three years
var(vont ) 0.000035 0.000020
var(voff
t ) 0.000011 0.0000006
RMSE 0.0045 0.0038 0.0062 0.0058
MAD 0.0596 0.0517 0.0705 0.0672
This table reports the estimate results of pricing models for on-the-run and off-the-run
Chinese government bonds. The parameters of the CIR model are estimated from monthly
on-the-run bond data. These parameters are then used to estimate the latent factor in the monthly
off-the-run bond " data. # We use a nonlinear Kalman" filter# approach to estimate the parameters.
 on  off  on  off
y1, t y 1, t o1, t o 1, t
yon off
t = yon , yt = t = oon , and vt =
, von off
, where subscript 1 and 3 denote 1-year
3, t yoff
3, t 3, t ooff
3, t
and 3-year bonds and superscript on and off denote on-the-run and off-the-run bonds, respectively.
The numbers in parentheses are t values
a
indicates statistical significance at the 1 % level respectively. RMSE and MAD are root mean
square error and mean absolute deviation, respectively

We therefore use the nonlinear Kalman filter for estimation. The details of the
nonlinear Kalman filter with its Matlab codes are reported in the Appendix.
After we estimate the parameters, we then study the dynamic of the latent
component and examine its temporal properties to explore the explanations of the
on-/off-the-run yield spread puzzle in China.

22.3.3 Estimation Results

22.3.3.1 Estimation Results of On-the-Run Issue

The left-hand column of Table 22.2 reports the estimation results of the CIR model
using on-the-run bonds. As shown, all the parameters are significant at the 1 %
level. The long-term mean value, the speed of mean reversion, and the volatility
parameter of r are 0.023, 0.089, and 0.063, respectively. These results are reason-
able and close to the results of other research on dynamic models in the Chinese
interest rate (Hong et al. 2010).
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 625

5.00%
r 7-day Repo Rate
4.50%
4.00%
3.50%
3.00%
2.50%
2.00%
1.50%
1.00%
0.50%
0.00%

Fig. 22.2 Time series of implied risk-free interest rate and 7-day repo rate in Chinese interbank
market. This figure plots the time series of implied risk-free interest rate estimated from the CIR
model using Chinese on-the-run government bonds, and the time series of 7-day repo rate in
Chinese interbank market

Figure 22.2 plots the time series of r estimated from the model. Most of the time,
the value of r is in the interval between 2 % and 4 %. We also plot the time series of
the 7-day repo rate in the Chinese interbank market. Similar trends in these two
curves suggest that r does capture the change of the market interest rate.

22.3.3.2 Estimation Results of Off-the-Run Issue

With all the parameters obtained for r, we next estimate the parameters of l using
the data of off-the-run bonds.4 The right-hand column of Table 22.2 reports the
estimation results. The parameter of sl is significant at the 1 % level.
Figure 22.3 plots the time series of l estimated from the data. As we can see from
the figure, most of l are negative. We conduct the t-test and find the average of l is
significantly negative at the 5 % level (the t-statistic is 017). Since l represents the
yield difference between off-the-run bonds and on-the-run bonds, negative
l provides further, strong evidence of the on-/off-the-run yield spread puzzle in
Chinese Treasury markets.

22.3.4 Regression Analysis

The analysis so far reveals that on average, the off-the-run bonds are traded at
a higher price or lower yield than the on-the-run bonds in Chinese Treasury
markets, which is hard to explain rationally. We next explore the information
contained in this negative yield spread by examining the temporal properties of
the latent component, l.

4
We also estimate the parameters of r and l jointly using the on-the-run and off-the-run data
together and find the results are quite similar.
626 R. Chen et al.

0.40%

0.20%

0.00%

−0.20%

−0.40%

−0.60%

−0.80%

Fig. 22.3 Time series of latent factor. This figure plots the time series of latent factor estimated
from the Chinese off-the-run government bonds

In order to explain the temporal properties of the latent component, we introduce

several variables. One is the turnover ratio difference between on-the-run issues and
off-the-run issues as a measure of liquidity difference, and it is used to examine
whether an on-/off-the-run yield spread in Chinese Treasury markets is related to
the difference in liquidity conditions. The turnover ratio difference between
on-the-run issues and off-the-run issues is defined as

TRoff
1, t  TR on
1, t þ TR off
3, t  TR on
3, t
TRt ¼  105 ,
2
where TR is the turnover ratio, the subscript 1 and 3 denote 1-year bonds and 3-year
bonds, and the superscript on and off denote on-the-run bonds and off-the-run bonds.
Pasquariello and Vega (2007, 2009) find that the release of macroeconomic news
changes liquidity, and hence the on-/off-the-run yield spread, in the US Treasury
market. For example, when macroeconomic news brings more funds into the bond
market, market-wide liquidity conditions will be better, investors might trade old
bonds more actively, and the yield difference between off-the-run bonds and
on-the-run bonds might decrease, and vice versa. Similarly, we introduce the
percentage change of a broad money supply measure, DM2, as a proxy of
market-wide liquidity conditions to examine whether there is covariation of the
latent component with changes in market-wide liquidity conditions. We use one
lagged DM2 to examine the impact of macroeconomic conditions on Dlt.
The last factor we investigate is investors’ behavior. It is observed that in
Chinese bond markets, there exists a “disposition effect.” Bond holders are
reluctant to realize loss and will not sell old bonds if they have a loss from the
investment. Consequently, old bonds might be traded at a lower yield than new
bonds. In China, the 7-day repo market is one of the most active bond markets, and
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 627

Table 22.3 Time regression results

(1) (2) (3) (4)
Dlt ¼ b0 + Dlt ¼ b0 + Dlt ¼ b0 + Dlt ¼ b0 + b1TRt + b2DM2t  1
b1TRt + et b1DM2t  1 + et b1Rt  1 + et + b3Rt  1 + et
Intercept 4.330(0.253) 0.001(2.349)a 0.001(2.348)a 0.0005(0.905)
TRt 0.904(1.625)b 0.231(0.390)
DM2t1 0.037(2.263)a 0.027(1.663)b
Rt1 0.057(2.703)c 0.045(1.930)a
Adj. R2 0.043 0.084 0.114 0.163
This table reports the results of regressing the change of latent component, Dlt, on the on-/off-the-
run turnover ratio difference (TRt), the lagged percentage change of M2 (DM2t1), and the lagged
7-day repo rate (Rt1). The numbers in parentheses are t values
ab
, , and c indicate statistical significance at the 10 %, 5 %, and 1 % level, respectively
Univariate regression of Dlt on the turnover ratio difference
Dlt ¼ b0 + b1TRt + et.
Univariate regression of Dlt on the lagged percentage change of the money supply:
Dlt ¼ b0 + b1DM2t  1 + et.
Univariate regression of Dlt on the lagged 7-day repo rate:
Dlt ¼ b0 + b1Rt  1 + et.
Multivariate regression of Dlt on the turnover ratio difference, lagged percentage change of the
money supply, and lagged 7-day repo rate:
Dlt ¼ b0 + b1TRt + b2DM2t  1 + b3Rt  1 + et

the change of the 7-day repo rate is a good measure of market conditions. When the
7-day repo rate goes up, the bond investment will generate a loss for the investors
and the disposition effect might occur. We use the interbank 7-day repo rates as the
proxy of the market interest rate to investigate whether the on-/off-the-run yield
spread puzzle in China is related to this irrational behavior. If the investors are
rational and the liquidity premium hypothesis holds, the liquidity of the whole
market will decrease in the bear market, and the old bonds will be traded at a higher
yield. Thus, detecting the response of the on-/off-the-run yield spread to the
change of the 7-day repo rate could help us distinguish whether the disposition
effect or the liquidity factor dominates. Similarly, we use one lagged 7-day repo
rate in the time series regression.
In what follows, we first run univariate time series regressions of Dlt against each
variable and then conduct multivariate regression analysis against all the three
factors. The regression models are specified as follows:
Univariate regression of Dlt on the turnover ratio difference, TRt:
Dlt ¼ b0 þ b1 TRt þ et :
Univariate regression of Dlt on the lagged percentage change of the money supply,
DM2t1:
Dlt ¼ b0 þ b1 DM2t1 þ et :
Univariate regression of Dlt on the lagged 7-day repo rate, Rt1:

Dlt ¼ b0 þ b1 Rt1 þ et :
628 R. Chen et al.

Multivariate regression of Dlt on the turnover ratio difference, lagged percentage

change of money supply, and lagged 7-day repo rate:
Dlt ¼ b0 þ b1 TRt þ b2 DM2t1 þ b3 Rt1 þ et :

22.3.4.1 Univariate Regression Analysis

Columns (1), (2), and (3) of Table 22.3 report the results of univariate time series
regressions of Dlt against the turnover ratio difference, the lagged percentage
change of the money supply, and the lagged 7-day repo rate, respectively.
As shown, all coefficients are significant, indicating that all these factors are useful
to explain the change of the on-/off-the-run yield spread in China. It is an interesting
finding that is worth further exploring.
The coefficients for TR and the lagged DM2 are significantly positive at the
10 % level and the 5 % level, respectively. That means the latent component
contains information about liquidity conditions. In particular, since lt is negative
most of the time, the positive sign of coefficients implies that when the explan-
atory variable increases, the on-/off-the-run yield spread will increase and move
close to zero. In other words, there will be less difference between off-the-run
bond yields and on-the-run bond yields. The significantly positive coefficient of
TR suggests that if the on-/off-the-run turnover ratio difference increases, that is,
the liquidity of old bonds becomes better, the yield difference between off-the-
run bonds and on-the-run bonds declines, and vice versa. This is consistent with
the liquidity premium hypothesis. Similarly, the significantly positive coefficient
of lagged DM2 reveals that if the money supply increases and market-wide
liquidity improves, the yield difference between old bonds and new bonds
declines, and vice versa. Both results provide evidence of a liquidity premium
in Chinese government bond yields. The liquidity premium hypothesis still holds
in Chinese Treasury markets despite the existence of the on-/off-the-run yield
spread puzzle.
The lagged 7-day repo rate is significantly negatively related to Dlt. That is,
when the market interest rate goes up and the bond investors have a loss, the yield
difference between old bonds and new bonds increases and the on-/off the run
spread becomes more negative. This indicates the disposition effect dominates the
effect of liquidity. When investors have a loss, the unwillingness to sell old bonds
leads to a lower yield for old bonds and a negative on-/off-the-run yield spread. The
regression against the lagged 7-day repo rate has the largest adjusted R square,
which implies the disposition effect could better explain the change in the on-/off-
the-run yield spread in China than the liquidity condition.

22.3.4.2 Multivariate Regression Analysis

Column (4) of Table 22.3 reports the results of multivariate regression. The
coefficient of TR is not significant any more, indicating that after controlling for
market-wide liquidity conditions and the disposition effect, the liquidity difference
has no influence on the on-/off-the-run yield spread. On the other hand, market-
wide liquidity and the disposition effect are still significant at the 10 % level.
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 629

The adjusted R square of the multivariate is about 16 %. This suggests that the
on-/off-the-run spread could be partly explained by the change in market-wide
liquidity conditions and the disposition effect.

22.4 Conclusion

In this chapter, we document a negative on-/off-the-run yield spread in Chinese

Treasury markets. This is contrary to the positive on-/off-the-run yield spread found
in most other countries and could be called the “on-/off-the-run yield spread
puzzle.” We introduce a latent factor into the pricing formula of off-the-run
bonds to capture the yield spread and estimate this factor by the nonlinear Kalman
filter. The result confirms the existence of the puzzle.
To reveal the information content of the negative on-/off-the-run yield spread, we
perform univariate and multivariate time series regressions of the change of the latent
factor against the turnover ratio difference between off-the-run issues and on-the-run
issues (a measure of the liquidity difference between off-the-run issues and on-the-run
issues), the lagged percentage change of M2 (a measure of market-wide liquidity
conditions), and the lagged 7-day repo rates (a measure of disposition effect). We find
that the liquidity premium hypothesis still holds in Chinese Treasury markets. The yield
spread, however, is dominated by the irrational disposition effect. When the investors
have a loss from the bond investment, they are more reluctant to sell old bonds, which
leads to a higher price and a lower yield for old bonds and hence causes the puzzle.
Our study is an attempt to explore the coexistence of a standard theoretical
hypothesis and irrational behavior in emerging Treasury markets such as China.
These markets have been a topic of interest increasingly, as the role of the emerging
markets in the global economy becomes more and more important.

Appendix 1: Nonlinear Kalman Filter

Let yon
M;t represent the time t yield of an on-the-run government bond maturing at tM.
Equation 22.3 could be written as
XM   X M

t ¼
Pon Cm Am, t exp Bm, t r t ¼ Cm exp yon
M, t ðtm  tÞ : (22.8)
m¼1 m¼1

As shown, yon
M;tis a nonlinear function of rt, which is inconsistent with the
requirements of the standard Kalman filter that state functions and measurement
functions should be linear. So we use the extended (nonlinear) Kalman filter to
linearize nonlinear functions. The idea is to employ the Taylor expansions around
the estimate at each step. That is, we express yon
M;t as
  @yon
M, t  
yon ðr
M, t t Þ  y on
^
M, t tjt1 þ
r jrt ¼^r tjt1  r t  r^tjt1 , (22.9)
@r t
where r^tjt1 is the estimate of rt at time t–1.
630 R. Chen et al.

@yon
M, t
To get @rt , we calculate the first-order derivative of Pon
t with respect to rt,

@Pon X M   @Pon @yonM, t

t
¼ Cm Am, t Bm, t exp Bm, t r t ¼ ont : (22.10)
@r t m¼1
@y M, t @r t

Thus, we have
X
M  
Cm Am, t Bm, t exp Bm, t r t
@yon
M, t m¼1
¼
: (22.11)
@r t X
M
Cm ðtm  tÞ exp yM, t  ðtm  tÞ
on

m¼1
 
Given r^tjt1, we can use Eq. 22.8 to calculate yon
M, t r^tjt1 and then use Eq. 22.11
@yon
to get M, t
@rt
Finally, the linearized measurement model for the on-the-run issues at time t is

t ¼at þbt r t þ vt
yon on on on
(22.12)

where
 
yon
1, t
t ¼
yon
yon
,
3, t
2 3
  @yon 1, t
on ^
6 y1, t ^
r tjt1  r tjt1  jr ¼^r
aon ¼6 @r t t tjt1 7 7,
t 4   @y3, t
on 5
y3, t r tjt1  r^tjt1 
on ^
jrt ¼^r tjt1
@r
2 on 3
@y1, t
j
6 @r rt ¼^r tjt1 7
bon 6 7,
t ¼ 4 @yon
t
3, t
5
jrt ¼^r tjt1
@r t

and von
t is the error term,
 
oon
1, t
von ¼ ,
t oon
3, t

where the subscript of 1 and 3 represent the 1-year bonds and 3-year bonds, while
the superscript on refers to on-the-run bonds.
After we get the measurement function, the third step is to rewrite (22.1) as
a discrete state function,

r t ¼ g þ fr t1 þ et , (22.13)
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 631

where
g ¼ yð1  expðk  DtÞÞ,
f ¼ expðk  DtÞ,

and et is the error term of rt and Dt is the size of the time interval in the discrete
sample. In our study, Dt ¼ 0.0833. The conditional mean and conditional variance
of rt are
r^tjt1 ¼ yð1  expðkDtÞÞ þ expðkDtÞ  r t1
  
  2 1  expðkDtÞ 1
Var r tjt1 ¼ sr yð1  expðkDtÞ þ expðkDtÞ  r t1 Þ :
k 2
(22.14)
Similarly, the state functions of the off-the-run issues are
r t ¼ g þ fr t1 þ et
(22.15)
lt ¼ lt1 þ sl et :
The conditional mean and conditional variance of lt are lt1 and s2l Dt,
respectively.
The corresponding measurement function is

t ¼ at þ bt r t þ jt lt þ vt ,
off
yoff off off off
(22.16)

where off refers to off-the-run bonds,

" #
yoff
1, t
yoff ¼ ,
t
yoff
3, t
2 3
off   @yoff
1, t @yoff
,
6 y1, t r^tjt1 ; l^tjt1 jr ¼^r  l^tjt1 
1 t
 r^tjt1  j ^ 7
aoff ¼6
6 @r t t tjt1 @lt lt ¼l tjt1 7,
7
t
4 off   off
@y3, t off
@y3, t 5
y3, t r^tjt1 ; l^tjt1  r^tjt1  jrt ¼^r tjt1  l^tjt1  jlt ¼l^tjt1
@r t @lt
2 3
@yoff
1, t
6 j 7
6 @r t rt ¼^r tjt1 7
boff ¼ 6 7,
t
4 @yoff
3, t
5
jrt ¼^r tjt1
@r t
2 off 3
@y1, t
6 j 7
6 @lt lt ¼l^tjt1 7
joff ¼ 6 7,
t
4 @yoff
3, t
5
jlt ¼l^tjt1
@lt
632 R. Chen et al.

" #
ooff
1, t
voff ¼ ,
t
ooff
3, t

X
M  
Cm Am, t Bm, t exp Dm, t  Bm, t r t  ðtm  tÞlt
@yoff
M, t
¼ m¼1
,
@r t X
M
Cm ðtm  tÞ exp yoff
M, t  ðtm  tÞ
m¼1

X
M  
Cm Am ðtm  tÞ exp Dm, t  Bm, t r t  ðtm  tÞlt
@yoff
M, t
¼ m¼1
,
@lt X
M
Cm ðtm  tÞ exp yoff
M, t  ðtm  tÞ
m¼1

and l^tjt1 is the estimate of lt at time t–1.

Once we get the state functions and the measurement functions, we employ the
regular iterative prediction-update procedure and the method of quasi-maximum
likelihood to estimate the parameters. When estimating the parameters of the
off-the-run issues, we use just the parameters g and f estimated from the on-the-run
issues to identify sl.

Appendix 2: Matlab Codes

Codes for the On-the-Run Bonds

%*************define the likelihood function***************%

function [logfun v1 zz QQ RR rr]¼kalfun(param)
k¼param(1);
theta¼param(2);
sigm¼param(3);
sigm2¼param(4);
v1¼zeros(63,2);
v¼zeros(2,1);
rr¼zeros(63,1);
zz¼zeros(63,2);
RR¼0;
QQ¼0;
load data.mat
z1¼data(:,1); % YTMs of one-year bonds
z3¼data(:,3); % YTMs of three-year bonds
c1¼data(:,2); % cash flows of one-year bonds
c3¼data(:,4:6); % cash flows of three-year bonds
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 633

couponrate¼data(:,7); % the coupon rate of three-year

bonds
z¼[z1,z3];
%gam¼sqrt(k^2+2*theta^2);
gam¼sqrt(k^2+2*sigm^2);
tao¼[1 2 3]’;
dt¼1/12;
a¼zeros(3,1);
b¼zeros(3,1);
for j¼1:3
a(j)¼log((2*gam*exp(k*j/2+gam*j/2))/((k+gam)*(exp
(gam*j)-1)+2*gam)^(2*k*theta/sigm^2));
b(j)¼(2*exp(gam*j)-2)/((k+gam)*(exp(gam*j)-1)
+2*gam);
end
r_(1)¼theta; %the initial value of r
A¼ exp(-k*dt);
P_¼(1/(1-A^2))*
(sigm^2*(1-exp(-k*dt))/k)*(theta*(1-exp(-k*dt))/2+r_
(1)*exp(-k*dt)); %the initial value of P
C¼theta*(1-exp(-k*dt)); %r(i)¼C+A*r(i-1)
zm¼zeros(63,2); %the prediction of YTM
R¼sigm2*[1 0;0 sqrt(1/3)]; %the covariance of measure-
ment functions
logfun¼0;
for i¼1:63
Q¼(sigm^2*(1-exp(-k*dt))/k)*(theta*(1-exp(-k*dt))/2
+r_(i)*exp(-k*dt)); %the conditional variance of state
functions
pz1(i)¼(c1(i)*b(1)*exp(a(1)-b(1)*r_(i)))/c1(i)*exp
(-z1(i)*1); %the partial derivative of one year z against r
pz3(i)¼sum(c3(i,:)’.*b.*exp(a-b*r_(i)))/sum(c3(i,:)’.
*tao.*exp(-z3(i)*tao)); %the partial derivative of three
year z against r
P1¼c1(i)*exp(a(1)-b(1)*r_(i)); %the prediction price
of one-year bond
P3¼sum(c3(i,:)’.*exp(a-b*r_(i))); % the prediction
price of three-year bond
%zm1¼bndyield(P1,c1(i),’20-Jan-1997’,’20-Jan-
1998’,1);
zm1¼-log(P1/c1(i)); %nonlinear measurement function
for one-year bonds
zm3¼bndyield(P3,couponrate(i),’20-Jan-1997’,’20-Jan-
2000’,1); %nonlinear measurement function for three-year
bonds
634 R. Chen et al.

H¼[pz1(i) pz3(i)]’;
%C1¼[zm1 zm3]’-H*r_(i);
%zm(i,:)¼ C1+H*r_(i);
%zm(i,:)¼ C1+H*r(i);
zm(i,:)¼[zm1 zm3]’; %the prediction of YTMs
v¼z(i,:)’-zm(i,:)’; %the error of measurement functions
v1(i,:)¼v’; % the error between the prediction and the
real value
F¼H*P_*H’+R; %the kalman gain
if det(F)<¼0
logfun¼0;
return
end
rr(i,:)¼r_(i);
zz(i,:)¼zm(i,:);
r(i)¼r_(i)+P_*H’*inv(F)*v; %update r
P¼P_-P_*H’*inv(F)*H*P_; %update P
ll¼-0.5*log(det(F))-0.5*v’*inv(F)*v; %likelihood
function
logfun¼logfun+ll;
r_(i+1)¼A*r(i)+C; %predict r
P_¼A*P*A’+Q; %predict P
end
QQ¼Q;
RR¼R;
logfun¼-logfun;
function covv¼covirance(param)
covv¼zeros(4,4);
for i¼1:4
for j¼1:4
parama¼param;
paramb¼param;
paramab¼param;
parama(i)¼param(i)*1.01;
paramb(j)¼param(j)*0.99;
paramab(i)¼param(i)*1.01;
paramab(j)¼paramab(j)*0.99;
ua¼kalfun(parama);
db¼kalfun(paramb);
udab¼kalfun(paramab);
kk¼kalfun(param);
covv(i,j)¼(ua+db-kk-udab)/((0.01*param(i))*
(0.01*param(j)));
end
end
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 635

Codes for the Off-the-Run Bonds

%************define the likelihood function *************%

function [logfun LL zz v1 QQ RR]¼kalfunL(paramL)
sigm3¼paramL(1);
sigm4¼paramL(2);
load dataL.mat
z1¼dataL(:,1); %YTMs of one-year bonds
z3¼dataL(:,3); % YTMs of three-year bonds
c1¼dataL(:,2); % cash flows of one-year bonds
c3¼dataL(:,4:6); % cash flows of three-year bonds
couponrate¼dataL(:,7); %the coupon rate of three-year
bonds
r¼dataL(:,8); %the estimated r in the CIR model
% the estimated parameters in the CIR model
k¼0.08899;
theta¼0.022659;
sigm¼0.063329;
gam¼sqrt(k^2+2*sigm^2);
z¼[z1,z3];
tao¼[1 2 3]’;
dt¼1/12;
a¼zeros(3,1);
b¼zeros(3,1);
e¼zeros(3,1);
zz¼zeros(63,2);
v1¼zeros(63,2);
v¼zeros(2,1);
RR¼0;
QQ¼0;
for j¼1:3
a(j)¼log((2*gam*exp(k*j/2+gam*j/2))/((k+gam)*(exp
(gam*j)-1)+2*gam)^(2*k*theta/sigm^2));
b(j)¼(2*exp(gam*j)-2)/((k+gam)*(exp(gam*j)-1)
+2*gam);
e(j)¼(sigm3^2*tao(j)^3)/6;
end
L_(1)¼0; %the initial value of L
v1¼zeros(63,2);
v¼zeros(2,1);
P_¼0;
zm¼zeros(63,2); %the prediction of YTM
R¼sigm4*[1 0;0 sqrt(1/3)]; %the covariance of measure-
ment functions
logfun¼0;
636 R. Chen et al.

for i¼1:63
Q¼sigm3^2*dt; %the conditional variance of state
functions
pz1(i)¼(c1(i)*exp(a(1)-b(1)*r(i)+e(1)-L_(i)*1))/c1(i)
*exp(-z1(i)*1); %the partial derivative of one year z against r
pz3(i)¼sum(c3(i,:)’.*tao.*exp(a-b*r(i)+e-L_(i)
*tao))/sum(c3(i,:)’.*tao.*exp(-z3(i)*tao)); %the par-
tial derivative of three year z against r
P1¼c1(i)*exp(a(1)-b(1)*r(i)+e(1)-L_(i)*1); %the pre-
diction price of one-year bond
P3¼sum(c3(i,:)’.*exp(a-b*r(i)+e-L_(i)*tao)); % the
prediction price of three-year bond
%zm1¼bndyield(P1,c1(i),’20-Jan-1997’,’20-Jan-
1998’,1);
zm1¼-log(P1/c1(i)); %nonlinear measurement function
for one-year bonds
zm3¼bndyield(P3,couponrate(i),’20-Jan-1997’,’20-Jan-
2000’,1); %nonlinear measurement function for three-year
bonds
H¼[pz1(i) pz3(i)]’;
%C1¼[zm1 zm3]’-H*r_(i);
%zm(i,:)¼ C1+H*r_(i);
%zm(i,:)¼ C1+H*r(i);
zm(i,:)¼[zm1 zm3]’; %the prediction of YTMs
v¼z(i,:)’-zm(i,:)’; %the error of measurement
functions
v1(i,:)¼v’; % the error between the prediction and
the real value
F¼H*P_*H’+R; %the kalman gain
if det(F)<¼0
logfun¼0;
return
end
LL(i,:)¼L_(i);
zz(i,:)¼zm(i,:);
L(i)¼L_(i)+P_*H’*inv(F)*v; %update r
P¼P_-P_*H’*inv(F)*H*P_; %update P
ll¼-0.5*log(det(F))-0.5*v’*inv(F)*v;
logfun¼logfun+ll;
L_(i+1)¼L(i); %predict r
P_¼P+Q; %predict P
end
QQ¼Q;
RR¼R;
logfun¼-logfun;
22 On-/Off-the-Run Yield Spread Puzzle: Evidence from Chinese Treasury Markets 637

function covvL¼coviranceL(paramL)
covvL¼zeros(2,2);
for i¼1:2
for j¼1:2
parama¼paramL;
paramb¼paramL;
paramab¼paramL;
parama(i)¼paramL(i)*1.0000001;
paramb(j)¼paramL(j)*0.9999999;
paramab(i)¼paramL(i)*1.0000001;
paramab(j)¼paramab(j)*0.9999999;
ua¼kalfunL(parama);
db¼kalfunL(paramb);
udab¼kalfunL(paramab);
kk¼kalfunL(paramL);
covvL(i,j)¼(ua+db-kk-udab)/((0.0000001*paramL(i))*
(0.0000001*paramL(j)));
end
end

References
Amihud, Y., & Mendelson, H. (1986). Asset pricing and the bid-ask spread. Journal of Financial
Economics, 17, 223–249.
Amihud, Y., & Mendelson, H. (1991). Liquidity, maturity, and the yields on US treasury
securities. The Journal of Finance, 46, 1411–1425.
Boudoukh, J., & Whitelaw, R. (1991). The benchmark effect in the Japanese government bond
market. The Journal of Fixed Income, 1, 52–59.
Boudoukh, J., & Whitelaw, R. (1993). Liquidity as a choice variable: A lesson from the Japanese
government bond market. Review of Financial Studies, 6, 265–292.
Cox, J., Ingersoll, J. E., & Ross, S. A. (1985). A theory of the term structure of interest rates.
Econometrica, 53, 385–407.
Furfine, C., & Remolona, E. (2002). What’s behind the liquidity spread? On-the-run and off-the-
run US treasuries in autumn 1998, BIS Quarterly Review, 6, 51–58.
Goldreich, D., Hanke, B., & Nath, P. (2005). The price of future liquidity: Time-varying liquidity
in the US treasury market. Review of Finance, 9, 1–32.
Guo, H., & Wu, K. (2006). Liquidity premium analysis of treasury bonds market in Shanghai stock
exchange. Finance and Economics (In Chinese), 217, 23–29.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Hong, Y., Lin, H., & Wang, S. (2010). Modeling the dynamics of Chinese spot interest rates.
Journal of Banking and Finance, 34, 1047–1061.
Kamara, A. (1994). Liquidity, taxes, and short-term treasury yields. Journal of Financial and
Quantitative Analysis, 29, 403–417.
Krishnamurthy, A. (2002). The bond/old-bond spread. Journal of Financial Economics, 66, 463–506.
Li, D., & He, Y. (2008). Study on the price gap between on-the-run and off-the-run treasury
securities of Shanghai stock exchange. Journal of Shanxi Finance and Economics University
(In Chinese), 30, 80–85.
638 R. Chen et al.

Lin, H., Liu, S., & Wu, C. (2011). Dissecting corporate bond and CDS spreads. The Journal of
Fixed Income, 20, 7–39.
Longstaff, F., Mithal, S., & Neis, E. (2005). Corporate yield spreads: Default risk or liquidity?
New evidence from the credit default swap market. The Journal of Finance, 60, 2213–2253.
Mason, R. (1987). The 10-year bond markets. Credit Suisse First Boston, CSFB Research.
Pasquariello, P., & Vega, C. (2007). Informed and strategic order flow in the bond markets. Review
of Financial Studies, 20, 1975–2019.
Pasquariello, P., & Vega, C. (2009). The on-the-run liquidity phenomenon. Journal of Financial
Economics, 92, 1–24.
Strebulaev, I. (2002). Liquidity and asset pricing: Evidence from the US Treasury securities
market (Working paper). London Business School, SSRN.
Vayanos, D., & Weill, P. (2008). A search-based theory of the on-the-run phenomenon. The
Journal of Finance, 63, 1361–1398.
Warga, A. (1992). Bond returns, liquidity, and missing data. Journal of Financial and Quantitative
Analysis, 27, 605–617.
Factor Copula for Defaultable Basket
Credit Derivatives 23
Po-Cheng Wu, Lie-Jane Kao, and Cheng-Few Lee

Contents
23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
23.2 Factor Copula with Issuer Default Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
23.3 Pricing a Defaultable Basket Credit Linked Note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
23.4 Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 646
23.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650
Appendix 1: Factor Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
One-Factor Gaussian Copula Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Law of Iterated Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652
Appendix 2: Cholesky Decomposition and Correlated Gaussian Random Numbers . . . . . . . . . 653
Cholesky Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
Correlated Gaussian Random Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
Appendix 3: Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
Weak Law of Large Numbers (WLLN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
Strong Law of Large Numbers (SLLN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
Uniform and Nonuniform Random Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655

Abstract
In this article, we consider a factor copula approach for evaluating basket credit
derivatives with issuer default risk and demonstrate its application in a basket
credit linked note (BCLN). We generate the correlated Gaussian random

P.-C. Wu (*) • L.-J. Kao

Department of Finance and Banking, Kainan University, Taoyuan, ROC, Taiwan
e-mail: pcwu@mail.knu.edu.tw; ljkao@mail.knu.edu.tw
C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 639
DOI 10.1007/978-1-4614-7750-1_23,
# Springer Science+Business Media New York 2015
640 P.-C. Wu et al.

numbers by using the Cholesky decomposition, and then the correlated default
times can be decided by these random numbers and the reduced-form model.
Finally, the fair BCLN coupon rate is obtained by the Monte Carlo simulation.
We also discuss the effect of issuer default risk on BCLN. We show that the
effect of issuer default risk cannot be accounted for thoroughly by considering
the issuer as a new reference entity in the widely used one-factor copula model,
in which constant default correlation is often assumed. A different default
correlation between the issuer and the reference entities affects the coupon rate
greatly and must be taken into account in the pricing model.

Keywords
Factor copula • Issuer default • Default correlation • Reduced-form model •
Basket credit derivatives • Cholesky decomposition • Monte Carlo simulation

23.1 Introduction

Structural and reduced-form models are the two main approaches for modeling
default risk. The structural model (Merton 1974) defines default as occurring when
a firm’s asset value falls below its debt. The reduced-form model (Jarrow and
Turnbull 1995), also known as the intensity model, views the default event as an
unexpected exogenous stochastic event. It estimates the intensity of the default
occurrence by using market data.
However, whether by the structural or reduced-form model, obtaining the joint
distribution of default times among a set of assets will be very complicated. Li (1999,
2000) first introduces the copula function (Sklar 1959) to simplify the estimation of
the joint distribution. Li assumes that the default events of reference entities follow
a Poisson process and sets the dependence structure as a Gaussian copula function.
Finally, he performs Monte Carlo simulation to obtain the default times. The copula
approach is the main approach for multi-name credit derivatives pricing in the last
decade. Mashal and Naldi (2003) use the copula approach to analyze how the default
probabilities of the protection sellers and buyers affect basket default swap (BDS)
spreads. While pricing the single-name credit default swap (CDS) with counterparty
risk based on the continuous-time Markov model, Walker (2006) indicates that using
a time-dependent correlation coefficient can improve the market-standard Gaussian
copula approach. By connecting defaults through a copula function, Brigo and
Chourdakis (2009) find that when the counterparty risk is involved, both the default
correlation and credit spread volatility impact the contingent CDS value.
In the implementation of the Gaussian copula using Monte Carlo simulation, the
computational complexity increases with the number of reference entities. Thus the
factor copula method, which makes the default event conditional on independent
state variables, is introduced to deal with these problems. Andersen et al. (2003)
find that one or two factors provide sufficient accuracy for the empirical correlation
matrices one encounters in credit basket applications. Hull and White (2004)
employ a multifactor copula model to price the kth-to-default swap and
23 Factor Copula for Defaultable Basket Credit Derivatives 641

collateralized debt obligation (CDO). Moreover, Laurent and Gregory (2005) use
one-factor Gaussian copula to simplify the dependence structure of reference
entities and apply this approach to price BDS and CDO. Wu (2010) develops
three alternative approaches to price the basket credit linked note (BCLN) with
issuer default risk using only one correlation parameter. Wu et al. (2011) analyze
how issuer default risk impacts the BCLN coupon rate by an implied default
correlation between the issuer and the reference entities.
On the other hand, acceleration techniques such as the importance sampling method
and others are used to improve the simulation efficiency. Chiang et al. (2007) and Chen
and Glasserman (2008) apply the Joshi-Kainth algorithm (Joshi and Kainth 2004), and
Bastide et al. (2007) use the Stein method (Stein 1972) for the multi-name credit
derivative pricing to reduce variance of the simulation results.
This article constructs a factor copula framework to evaluate defaultable basket
credit derivatives. The effect of the default correlation between the issuer and the
reference entities is considered in the proposed model. Its application in BCLN with
issuer default risk is also demonstrated. This study shows that the default correla-
tion between the issuer and the reference entities plays an important role in the
decision of fair BCLN coupon rate.
The remainder of this article is organized as follows. Section 23.2 reviews the
factor copula model and shows the proposed basket credit derivative pricing
model with issuer default event. Subsequently, Section 23.3 introduces the BCLN
and demonstrates how to price it when issuer default risk exists. Section 23.4 presents
the results of numerical analysis. Conclusions are finally drawn in Section 23.5.

23.2 Factor Copula with Issuer Default Risk

The most widely used copula function is the Gaussian copula and its definition is as
follows:

1 1 1

CGa ðu1 , u2 , . . . , uN Þ ¼ FR f ðu1 Þ, f ðu2 Þ, . . . , f ðuN Þ (23.1)

where FR denotes a multivariate cumulative normal (Gaussian) distribution,

R represents the correlation coefficient matrix, and f1 is the inverse function of
one dimensional cumulative Gaussian distribution.
Consider a credit portfolio which contains N reference entities, the default times
of each reference entity are t1,t2, . . . ,tN, respectively. According to the reduced-
form model, each reference entity default follows a Poisson process. The cumula-
tive default probability before time t is

Fi ðtÞ ¼ Pðti  tÞ ¼ 1  eli t , i ¼ 1, 2, . . . , N (23.2)

where li is the hazard rate of the reference entity i. Because Fi(t)  U(0,1), applying
the Gaussian copula obtains the multivariate joint distribution of default times, as
follows:
642 P.-C. Wu et al.


Fðt1 , t2 , . . . , tN Þ ¼ FR f1 ðF1 ðt1 ÞÞ, f1 ðF2 ðt2 ÞÞ, . . . , f1 ðFN ðtN ÞÞ
(23.3)

Let Xi represent the Gaussian random variable corresponding to the default time
of the reference entity i. In the one-factor copula model, the default time of
reference entity i depends on a common factor eY and a firm specific risk factor
exi. Both eY and eXi are independent standard Gaussian variables. The details of
one-factor copula model are given in Appendix 1. Thus Xi can be created via
Cholesky decomposition, as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X i ¼ r X i Y eY þ 1  rX i Y 2 e X i , i ¼ 1, 2, . . . , N (23.4)

where rXiY denotes the correlation coefficient between the reference entity Xi and
the common factor eY. How to apply the Cholesky decomposition for generating
correlated random variables is shown in Appendix 2.
One-factor Gaussian copula model with constant pairwise correlations has
become the standard market model. In the standard market model, all rXi Y in
Eq. 23.4 are equal to r, then the constant pairwise correlation rXi Xj (i 6¼ j) will
be r2. Let X1 ¼ f1(F1(t1)), X2 ¼ f1(F2(t2)), . . . , XN ¼ f1(FN(tN)), by mapping
ti and Xi, we can simulate the default time of the reference entity i using the
following equation:

lnð1  fðXi ÞÞ
ti ¼ F1
i ðfðXi ÞÞ ¼ , i ¼ 1, 2, . . . , N (23.5)
li

When issuer default risk is involved in the basket credit derivative, a natural
way is to view it as one reference entity of the derivative holder’s credit portfolio,
as discussed in Wu (2010). Wu (2010) assumes the issuer default time is deter-
mined by a Gaussian random variable Z. Like the reference entity variable Xi, Z is
decided by the common factor eY and the issuer’s specific risk factor eZ. Both eY
and eZ are independent standard Gaussian variables. Because the issuer is
viewed as one additional reference name in the portfolio, the correlation coeffi-
cient between Z and eY is also r. In this approach, Z and Xi are formulated as
follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ¼ reY þ 1  r2 e Z (23.6)

pffiffiffiffiffiffiffiffiffiffiffiffiffi
Xi ¼ reY þ 1  r2 e X i , i ¼ 1, 2, . . . , N (23.7)

In the above approach, the default correlation between the issuer and the reference
entities will be fixed to r2, which is always positive. Thus, it is not flexible enough
to deal with the default correlation. The default correlation between the issuer and
23 Factor Copula for Defaultable Basket Credit Derivatives 643

Fig. 23.1 The default

correlations between the
issuer, reference entities, and Common Factor
the common factor in the Y
proposed model
r r

Reference
Entities Issuer
Xi, i = 1,2,…,N Z
rXZ

the reference entities has a different impact on the fair coupon rate, and this needs to
be considered in credit derivative pricing.
Suppose that the default correlation between the issuer and the reference entities is
rXZ. The relationship between the issuer, reference entities, and the common factor in
the proposed model is shown in Fig. 23.1. Given that the three random variables eY,
eZ, and eXi are independent of each other, Xi, which is the Gaussian random variable
corresponding to the default time of the reference entity i, can be obtained by the
Cholesky decomposition. Thus, when the default correlation between the issuer and
the reference entities is incorporated into the pricing model, Z and Xi should be
formulated as follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ¼ reY þ 1  r2 e Z (23.8)

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2
u
rXZ  r 2 u r  r 2
t
Xi ¼ reY þ pffiffiffiffiffiffiffiffiffiffiffiffiffi eZ þ 1  r2  pffiffiffiffiffiffiffiffiffiffiffiffiffi eXi
XZ
(23.9)
1  r2 1  r2

To obtain a real number value of Xi, the following criteria must be satisfied.

!2
r  r2
r2 þ pXZffiffiffiffiffiffiffiffiffiffiffiffiffi 1 (23.10)
1  r2

By rearranging the above equation, the criteria can be written as follows:

2r2 þ r2XZ  2r2 rXZ  1 (23.11)

According to the above settings, the correlation coefficient between the reference
entity Xi and Xj will be
644 P.-C. Wu et al.


Cov Xi ; Xj
rXi Xj ¼
sXi sXj
0 1
rXZ  r2
B re Y þ p ffiffiffiffiffiffiffiffiffiffiffiffiffi eZ C
1  r2
B vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi C
B u !2 C
B u C
B t rXZ  r2 C
B þ 1  r  pffiffiffiffiffiffiffiffiffiffiffiffiffi eXi ,
2 C
B 1  r2 C
B C
¼ CovB C
B r r 2
C
B reY þ pXZffiffiffiffiffiffiffiffiffiffiffiffiffi eZ C (23.12)
B 1  r2 C
B vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi C
B u !2 C
B u rXZ  r2 C
@ t A
þ 1  r  pffiffiffiffiffiffiffiffiffiffiffiffiffi eXj
2
1  r2
2
ðrXZ  r2 Þ
¼ r2 þ
1  r2
r þ rXZ  2rXZ r2
2 2
¼
1  r2

23.3 Pricing a Defaultable Basket Credit Linked Note

Basket credit linked note (BCLN) is a kind of basket credit derivative

product. It is a note with a price or coupon linked to credit events of
reference entities (obligations). The conventional form of BCLN is the kth-
to-default BCLN. The BCLN holder (the protection seller) pays the notional
principal to the BCLN issuer (the protection buyer) at the start of the
contract and receives the coupon payments until either the kth default or
the contract maturity, whichever occurs earlier. If the kth default occurs
before contract maturity, the BCLN holder receives the recovered value of
the reference entity from the BCLN issuer. Otherwise, the BCLN holder
receives the notional principal back on contract maturity. In derivative
markets, the issuer default risk is attracting considerable attention because
of the recent financial turmoil and collapses of large financial institutions. If
the BCLN issuer defaults, the BCLN holder will not receive the recovered
value of the reference entity as the credit event happens nor the notional
amount at the contract maturity. The coupon payments also cease due to the
issuer default. Thus the issuer default results in a large loss. Therefore, it is
important to incorporate issuer default risk in BCLN pricing to obtain
a reasonable coupon rate.
If the issuer default risk is not considered, the value of a kth-to-default BCLN
with N reference entities, of which the notional principal is one dollar, can be
written as follows:
23 Factor Copula for Defaultable Basket Credit Derivatives 645

2 3
XT
rti
6c  e I ð t i < tk Þ 7
6 7
BCLN ¼ EQ 6 i¼1 7 (23.13)
4 þ dk  ertk  I ðtk  tT Þ 5
þ ertT  I ðtk > tT Þ

where the coupon rate is c and is paid annually. The coupon payment dates are
ti,i ¼ 1,2,  ,T. The maturity date is tT. Furthermore, tk is the kth default time, and
t1 < t2 <  < tN. dk is the recovery rate of the kth default reference entity. Thus dk
denotes the recovery value, which the issuer pays to the BCLN holder on the kth
default. The discount rate is r%. Finally, Q denotes the risk-neutral probability
measure, and I(.) is an indicator function.
Let the above equation equals one, the equation can be rewritten as:
" #
X
T
cE Q
erti I ðti < tk Þ
 i¼1 rtk
 (23.14)
Q 1  dk  e  I ð tk  t T Þ
¼E
ertT  I ðtk > tT Þ

Here we employ the Monte Carlo simulation, which is described in Appendix 3, to

obtain the initial fair BCLN coupon rate c as follows:

XW 

1  dsk  e
rtk  I tsk  tT
s

ertT  I tsk > tT

c¼
s¼1
" # (23.15)
XW X T

rti
e I t i < tk s

s¼1 i¼1

where W represents the number of simulation runs. tks represents the kth default
time, and dks denotes the recovery rate of the kth default reference entity at the sth
simulation, respectively.
When the issuer default risk is involved, whether the issuer default occurs before
or after the kth default must be taken into account. This article defines ^t as the issuer
default time and ^d as the issuer recovery rate. The BCLN holder gets back the
recovered value of the reference obligation if the kth default occurs before both the
issuer default time ^t and maturity date tT. If the issuer default occurs before the kth
default and maturity date, the issuer will not provide the BCLN holder with the
redemption proceeds and stop the coupon payments. In this situation, the notional
principal multiplied by the issuer recovery rate is returned to the BCLN holder. To
obtain all of the notional principal back, both the kth default time and the issuer
default time must be later than the contract maturity date. Thus, the value of a kth-
to-default BCLN with issuer default risk is modified as follows:
646 P.-C. Wu et al.

2 3
X T
rti
6 c  e I ð ti < min ð t k , ^
t Þ Þ 7
6 7
6 i¼1
7
6 7
BCLN ¼ EQ 6 þ d  ertk  I ðt < minð^t , t ÞÞ 7 (23.16)
6 k k T 7
6 7
4 þ ^d  er^t  I ð^t < minðtk , tT ÞÞ 5
þ ertT  I ðtT < minðtk , ^t ÞÞ

Therefore, the fair value of the coupon rate c with issuer default risk is
2
3
1  dsk  ertk  I tsk < minð^t s , tT Þ
s

X
W 6 7
6

 7
4 ^d  er^t s  I ^t s < min ts , tT 5
s¼1
rtT


s s k 
e  I tT < min tk , ^t
c¼ " # (23.17)
XW XT

 
erti I ti < min tsk , ^t s
s¼1 i¼1

where ^t s represents the issuer default time at the sth simulation.

23.4 Numerical Analysis

The numerical example presented here is a 5-year BCLN with three reference
entities; all of which are with notional principal one dollar, hazard rate 5 %, and
recovery rate 30 %. Furthermore, the coupon is paid annually; the hazard rate and
recovery rate of the issuer are 1 % and 30 %, respectively. Sixty thousand runs of
Monte Carlo simulation are executed to calculate the coupon rates, and the results
are shown in Tables 23.1, 23.2 and 23.3.
As we can see in Tables 23.1, 23.2 and 23.3, when issuer default risk
is considered by viewing it as one reference entity of the credit portfolio
(column II), the BCLN coupon rate increases compared to those without issuer
default risk (column I) for k ¼ 1–3. This is reasonable because the existence of
issuer default risk increases the risk of holding a BCLN; thus, the holder will
demand a higher coupon rate.
When the default correlations between the issuer and the reference entities are
considered as in the proposed model, the BCLN coupon rates with issuer default risk
(column III) are greater than those without issuer default risk (column I) for k ¼ 2 and 3
in Tables 23.2 and 23.3. However, when k ¼ 1 in Table 23.1, most of the BCLN
coupon rates with issuer default risk are lower than those without issuer default risk,
especially when the issuer and the reference entities are highly negatively or positively
correlated. This result shows that the BCLN coupon rates with issuer default risk are
not necessarily greater than those without issuer default risk. Moreover, from Figs. 23.2,
 23

Table 23.1 First-to-default BCLN coupon rates without and with issuer default risk. (I) Default free: Issuer default risk is not included in the pricing model.
(II) In credit portfolio: The issuer is viewed as one reference entity of the credit portfolio. The default correlations between the issuer and the reference entities
are fixed to r2, which is always positive. (III) The proposed model: The default correlation between the issuer and the reference entities is rXZ, which may be
positive or negative
(I) (II) (III) The proposed model
rXZ
r Default free In credit portfolio 0.9 0.6 0.3 0 0.3 0.6 0.9
0.9 8.0195 % 8.0346 % – – – – – – 7.6700 %
0.8 9.2625 % 9.3471 % – – – – 7.0292 % 9.3593 % 7.8852 %
0.7 10.2417 % 10.4279 % – – – 7.2895 % 10.2124 % 10.1562 % 7.9327 %
0.6 11.0469 % 11.3501 % – – – 10.5546 % 11.4061 % 10.5488 % 7.9452 %
0.5 11.7604 % 12.1928 % – – 10.2184 % 12.0213 % 12.0974 % 10.7840 % 7.9650 %
0.4 12.3336 % 12.8846 % – 8.8110 % 11.8856 % 12.9609 % 12.5139 % 10.9275 % 7.9665 %
0.3 12.7890 % 13.4509 % – 10.6576 % 12.9594 % 13.5479 % 12.7852 % 11.0273 % 7.9659 %
0.2 13.1304 % 13.8645 % 7.5062 % 11.7035 % 13.5910 % 13.9236 % 12.9839 % 11.0791 % 7.9894 %
0.1 13.3336 % 14.1181 % 8.7846 % 12.2398 % 13.9263 % 14.1327 % 13.1152 % 11.1038 % 7.9758 %
0 13.3875 % 14.1785 % 9.0738 % 12.3865 % 14.0210 % 14.1785 % 13.1205 % 11.1105 % 7.9933 %
Factor Copula for Defaultable Basket Credit Derivatives

0.1 13.3547 % 14.1215 % 8.7274 % 12.1743 % 13.8962 % 14.1373 % 13.0764 % 11.1036 % 8.0044 %
0.2 13.1236 % 13.8336 % 7.4400 % 11.5964 % 13.4923 % 13.8863 % 12.9678 % 11.0703 % 7.9941 %
0.3 12.7744 % 13.4073 % – 10.5344 % 12.8035 % 13.4871 % 12.7940 % 10.9995 % 7.9702 %
0.4 12.2933 % 12.8209 % – 8.6831 % 11.7668 % 12.8436 % 12.5034 % 10.9036 % 7.9704 %
0.5 11.7091 % 12.1252 % – – 10.1270 % 11.9387 % 12.0415 % 10.7603 % 7.9715 %
0.6 10.9854 % 11.2806 % – – – 10.4846 % 11.3343 % 10.5102 % 7.9448 %
0.7 10.1941 % 10.3750 % – – – 7.1962 % 10.1661 % 10.1106 % 7.9210 %
0.8 9.2061 % 9.2802 % – – – – 6.9644 % 9.2933 % 7.8398 %
0.9 7.9788 % 7.9892 % – – – – – – 7.6319 %
647
 648

Table 23.2 Second-to-default BCLN coupon rates without and with issuer default risk. (I) Default free: Issuer default risk is not included in the pricing
model. (II) In credit portfolio: The issuer is viewed as one reference entity of the credit portfolio. The default correlations between the issuer and the reference
entities are fixed to r2, which is always positive. (III) The proposed model: The default correlation between the issuer and the reference entities is rXZ, which
may be positive or negative
(I) (II) (III) The proposed model
rXZ
r Default free In credit portfolio 0.9 0.6 0.3 0 0.3 0.6 0.9
0.9 5.2169 % 5.2585 % – – – – – – 5.2870 %
0.8 4.9714 % 5.1490 % – – – – 5.9481 % 5.1998 % 5.2213 %
0.7 4.7203 % 5.0406 % – – – 6.1463 % 5.3390 % 4.9617 % 5.2041 %
0.6 4.4753 % 4.9077 % – – – 5.5463 % 4.9738 % 4.8270 % 5.2135 %
0.5 4.2476 % 4.7844 % – – 5.8264 % 5.1061 % 4.7372 % 4.7592 % 5.1736 %
0.4 4.0568 % 4.6587 % – 6.1656 % 5.3827 % 4.8129 % 4.6044 % 4.7055 % 5.1541 %
0.3 3.8888 % 4.5440 % – 5.7880 % 5.0664 % 4.6078 % 4.4986 % 4.6566 % 5.1422 %
0.2 3.7766 % 4.4490 % 6.3028 % 5.5060 % 4.8417 % 4.4584 % 4.4238 % 4.6153 % 5.1375 %
0.1 3.6917 % 4.3717 % 6.1079 % 5.3332 % 4.7028 % 4.3759 % 4.3711 % 4.6020 % 5.1434 %
0 3.6646 % 4.3412 % 6.0386 % 5.2612 % 4.6477 % 4.3412 % 4.3441 % 4.5867 % 5.1507 %
0.1 3.6833 % 4.3559 % 6.0959 % 5.3033 % 4.6836 % 4.3606 % 4.3478 % 4.5939 % 5.1636 %
0.2 3.7480 % 4.4025 % 6.2475 % 5.4584 % 4.7916 % 4.4244 % 4.3710 % 4.5969 % 5.1732 %
0.3 3.8817 % 4.5070 % – 5.7235 % 4.9942 % 4.5580 % 4.4430 % 4.6288 % 5.1900 %
0.4 4.0242 % 4.5995 % – 6.0943 % 5.3165 % 4.7513 % 4.5276 % 4.6679 % 5.1908 %
0.5 4.2155 % 4.7249 % – – 5.7533 % 5.0596 % 4.6854 % 4.7051 % 5.2016 %
0.6 4.4092 % 4.8338 % – – – 5.4952 % 4.9002 % 4.7565 % 5.2005 %
0.7 4.6447 % 4.9559 % – – – 6.0639 % 5.2704 % 4.8851 % 5.1929 %
0.8 4.9138 % 5.0873 % – – – – 5.8734 % 5.1404 % 5.2065 %
0.9 5.1952 % 5.2336 % – – – – – – 5.2602 %
P.-C. Wu et al.
 23

Table 23.3 Third-to-default BCLN coupon rates without and with issuer default risk. (I) Default free: Issuer default risk is not included in the pricing model.
(II) In credit portfolio: The issuer is viewed as one reference entity of the credit portfolio. The default correlations between the issuer and the reference entities
are fixed to r2, which is always positive. (III) The proposed model: The default correlation between the issuer and the reference entities is rXZ, which may be
positive or negative
(I) (II) (III) The proposed model
rXZ
r Default free In credit portfolio 0.9 0.6 0.3 0 0.3 0.6 0.9
0.9 3.5790 % 3.7167 % – – – – – – 3.8150 %
0.8 3.0115 % 3.3632 % – – – – 5.1041 % 3.4122 % 3.6937 %
0.7 2.6429 % 3.1441 % – – – 5.2536 % 3.3862 % 3.1083 % 3.6545 %
0.6 2.3993 % 2.9853 % – – – 3.4885 % 3.0143 % 2.9859 % 3.6223 %
0.5 2.2319 % 2.8735 % – – 3.7706 % 3.0416 % 2.8644 % 2.9228 % 3.6225 %
0.4 2.1107 % 2.7835 % – 4.5560 % 3.2013 % 2.8419 % 2.7752 % 2.8858 % 3.6088 %
0.3 2.0280 % 2.7177 % – 3.6782 % 2.9511 % 2.7344 % 2.7243 % 2.8614 % 3.5933 %
0.2 1.9809 % 2.6597 % 5.3797 % 3.3040 % 2.8003 % 2.6635 % 2.6743 % 2.8273 % 3.5886 %
0.1 1.9499 % 2.6251 % 4.5085 % 3.1384 % 2.7242 % 2.6247 % 2.6480 % 2.8155 % 3.5889 %
0 1.9359 % 2.6020 % 4.3246 % 3.0819 % 2.6916 % 2.6020 % 2.6360 % 2.8013 % 3.5745 %
Factor Copula for Defaultable Basket Credit Derivatives

0.1 1.9470 % 2.6101 % 4.5012 % 3.1331 % 2.7072 % 2.6095 % 2.6356 % 2.7996 % 3.5754 %
0.2 1.9719 % 2.6320 % 5.3262 % 3.2787 % 2.7677 % 2.6357 % 2.6482 % 2.8099 % 3.5761 %
0.3 2.0171 % 2.6668 % – 3.6289 % 2.8962 % 2.6780 % 2.6788 % 2.8205 % 3.5859 %
0.4 2.0883 % 2.7214 % – 4.4910 % 3.1651 % 2.7836 % 2.7186 % 2.8426 % 3.5879 %
0.5 2.2105 % 2.8231 % – – 3.7405 % 2.9960 % 2.8098 % 2.8894 % 3.6007 %
0.6 2.3793 % 2.9471 % – – – 3.4615 % 2.9825 % 2.9468 % 3.6257 %
0.7 2.6279 % 3.1099 % – – – 5.2006 % 3.3573 % 3.0784 % 3.6345 %
0.8 2.9783 % 3.3144 % – – – – 5.0265 % 3.3628 % 3.6818 %
0.9 3.5583 % 3.6858 % – – – – – – 3.7992 %
649
650 P.-C. Wu et al.

First to Default BCLN

15%
14%
rxz
13%
0.9
Coupon Rate

12% 0.5
11% 0
10% −0.5
−0.9
9%
ND
8%
7%
−1 −0.5 0 0.5 1
r

Fig. 23.2 First-to-default BCLN coupon rates under various default correlations between the
common factor and reference entities/issuer (r)

Second to Default BCLN

7%

rxz
6%
0.9
Coupon Rate

0.5
5% 0
−0.5

4% −0.9
ND

3%
−1 −0.5 0 0.5 1
r

Fig. 23.3 Second-to-default BCLN coupon rates under various default correlations between the
common factor and reference entities/issuer (r)

23.3 and 23.4, we find that when the correlation between the issuer and the reference
entities approaches a strongly positive correlation (rXZ ¼ 0.9), the BCLN coupon rate
curve becomes flatter and less sensitive to the common factor.

23.5 Conclusion

This article applies a factor copula approach for evaluating basket credit derivatives
with issuer default risk. The proposed model considers the different effects of
23 Factor Copula for Defaultable Basket Credit Derivatives 651

Third to Default BCLN

6%

5% rxz
0.9
Coupon Rate

4% 0.5
0
3%
−0.5
−0.9
2%
ND

1%
−1 −0.5 0 0.5 1
r

Fig. 23.4 Third-to-default BCLN coupon rates under various default correlations between the
common factor and reference entities/issuer (r)

the default correlation between the issuer and the reference entities.
A numerical example of the proposed model on BCLN is demon-
strated and discussed in the article. The example shows that viewing the issuer
default as a new reference entity cannot reflect the effect of issuer default risk
thoroughly. The different default correlation between the issuer and the reference
entities affects the coupon rate greatly and must be taken into account in credit
derivative pricing.

Appendix 1: Factor Copula

In a factor copula model, we assume that different variables depend on some

common factors. The most widely used model in finance is the one-factor Gaussian
copula model.

One-Factor Gaussian Copula Model

Let Si(t) ¼ P(ti > t) and Fi(t) ¼ P(ti  t) be the marginal survival and marginal
default distributions, respectively. Let Y be the common factor and f its density
function. Assume the default times are conditionally independent, given the
ijY ijY
common factor Y. qt ¼ P(ti > t | Y) and pt ¼ P(ti  t | Y) are the conditional
survival and conditional default distributions, respectively. According to the law
of iterated expectations, the joint survival and default distribution functions are
as follows:
652 P.-C. Wu et al.

Sðt1 ; t2 ; . . . ; tn Þ
¼ Pðti > t1 , t2 > t2 , . . . , tn > tn Þ
ðY n (23.18)
ijy
¼ qt f ðyÞdy
i¼1

Fðt1 ; t2 ; . . . ; tn Þ
¼ Pðti  t1 , t2  t2 , . . . , tn  tn Þ
ðY n (23.19)
ijy
¼ pt f ðyÞdy
i¼1

In the one-factor Gaussian copula, the credit variable Xi is Gaussian

distributed. Xi depends on the common factor Y and an individual factor eXi as
follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xi ¼ rXi Y Y þ 1  rXi Y 2 eXi , i ¼ 1, 2, . . . , N (23.20)

where Y and eXi are two independent standard Gaussian random variables.
We can get the default time ti ¼ F1
i (f(Xi)), where f() is the cumulative density
function of a standard Gaussian variable. Then the conditional distribution of ti,
given the common factor Y, is
0 1
1
ijY Bf ðFi ðtÞÞ  rXi Y Y C
pt ¼ f @ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi A, (23.21)
1  r2Xi Y

and the joint distribution function of t1,t2,  ,tn is as follows:

ðY
n
ijy
Fð t 1 ; t 2 ; . . . ; t n Þ ¼ pt f ðyÞdy
i¼1
2 0 13
ð Y 1
6
n
B f ð F i ð t ÞÞ  r Y C7
¼ 4 f@ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXi Y A5f ðyÞdy (23.22)
i¼1 1  rX i Y2

where f(y) is the standard Gaussian density. The copula of default times is
a Gaussian copula.

Law of Iterated Expectations

The law of iterated expectations states that E(Y) ¼ E(E(Y|X)). The proof is as follows:
23 Factor Copula for Defaultable Basket Credit Derivatives 653

ð
EðEðY jXÞÞ ¼ EðY jXÞf X ðxÞ dx
ð ð 
¼ y f XjY ðyjxÞ dy f X ðxÞ dx
ðð
¼ y f X, Y ðx; yÞ dxdy (23.23)
ð ð 
¼ y f X, Y ðx; yÞ dx dy
ð
¼ y f ðyÞ dy ¼ EðY Þ

Appendix 2: Cholesky Decomposition and Correlated Gaussian

Random Numbers

Cholesky Decomposition

A symmetric positive defined real number matrix A can be decomposed as

follows:

A ¼ LLT (23.24)

where L is a lower triangular matrix with strictly positive diagonal entries and LT is
the transpose of L. This format is the Cholesky decomposition and it is unique. The
entries of L are as follows:

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u Xj1
Lj, j ¼ tAj, j  L2j, k (23.25)
k¼1

!
1 X
j1
Li, j ¼ Ai , j  Li, k Lj, k , for i>j (23.26)
Lj , j k¼1

Correlated Gaussian Random Numbers

In financial applications, the Cholesky decomposition is usually used to create

correlated Gaussian random variables. Suppose we need to generate n correlated
random Gaussian numbers, X1, X2,   , Xn, given a positive defined correlation
coefficient matrix R. We first generate iid standard Gaussian random variables Z1,
Z2,  , Zn and let
654 P.-C. Wu et al.

2 3 2 3
X1 Z1
6 X2 7 6 Z2 7
X¼6 7 6 7
4 ⋮ 5, Z ¼ 4 ⋮ 5:
Xn Zn

Let X ¼ CZ, where C is an n  n matrix; then R ¼ Var(X) ¼ E(XXT) ¼ CCT and

C is the lower triangular matrix of the Cholesky decomposition of R. Therefore, we
can obtain the correlated Gaussian random variables X1, X2,   , Xn by letting
X ¼ CZ.

Appendix 3: Monte Carlo Simulation

Monte Carlo simulation is a computational algorithm based on random number

sampling. It generates n iid random samples X1, X2,   , Xn of the random
variable X and estimates E[X] by X = X1 þX2 nþþ Xn . X converges to E[X] as
n ! 1, according to the law of large numbers.

Weak Law of Large Numbers (WLLN)

Let X1, X2,    , Xn be an iid sequence of random variables for which E[X] < 1.
Then X1 þX2 þþX
P
n
n
! E½X as n ! 1. The WLLN implies that the Monte
Carlo method converges to E[X].

Strong Law of Large Numbers (SLLN)

Let X1, X2,    , Xn be an iid sequence of random variables for which E[X] < 1.
Then X1 þX2nþþXn ! E½X as n ! 1. The SLLN implies that the Monte
a:s:

Carlo method almost surely converges to E[X].

Uniform and Nonuniform Random Numbers

Usually, the computer has a uniform random number generator, which can generate
a sequence of iid uniform random numbers U1, U2,    on [0,1]. Then, how do we
get the nonuniform random numbers? Usually, these can be achieved by inversion.
Given a distribution function F(∙), there is a one-to-one mapping from U(∙) to F(∙).
Since F(∙) is nondecreasing, F-1(x) ¼ inf{y : F(y)  x}. If F is continuous and
strictly increasing, the inverse of the function is F(F1(x)) ¼ F1(F(x)) ¼ x. Thus,
the nonuniform random numbers with distribution F can be obtained by x ¼ F1(u),
where u is a uniform random number on [0,1].
23 Factor Copula for Defaultable Basket Credit Derivatives 655

References
Andersen, L., Sidenius, J., & Basu, S. (2003). All your hedges in one basket. Risk, 16(11), 67–72.
Bastide, D., Benhamou, E., & Ciuca, M. (2007). A comparative analysis of basket default swaps
pricing using the Stein method. Working paper, pricing partners.
Brigo, D., & Chourdakis, K. (2009). Counterparty risk for credit default swaps: Impact of spread
volatility and default correlation. International Journal of Theoretical and Applied Finance,
12(7), 1007–1026.
Chen, Z., & Glasserman, P. (2008). Fast pricing of basket default swaps. Operations Research,
56(2), 286–303.
Chiang, M. H., Yueh, M. L., & Hsieh, M. H. (2007). An efficient algorithm for basket default swap
valuation. Journal of Derivatives, 15(2), 1–22.
Hull, J., & White, A. (2004). Valuation of a CDO and an nth to default CDS without Monte Carlo
simulation. Journal of Derivatives, 12(2), 8–23.
Jarrow, R., & Turnbull, S. (1995). Pricing derivatives on financial securities subject to credit risk.
Journal of Finance, 1, 53–85.
Joshi, M. S., & Kainth, D. (2004). Rapid and accurate development of prices and Greeks for Nth to
default credit swaps in the Li model. Quantitative Finance, 4, 266–275.
Laurent, J. P., & Gregory, J. (2005). Basket default swaps, CDOs and factor copulas. Journal of
Risk, 7(4), 103–122.
Li, D. X. (1999). The valuation of basket credit derivatives. CreditMetrics Monitor, 34–50.
Li, D. X. (2000). On default correlation: A copula function approach. Journal of Fixed Income, 9,
43–54.
Mashal, R., & Naldi, M. (2003). Pricing multi-name default swaps with counterparty risk.
Quantitative Credit Research, 1–16.
Merton, R. C. (1974). On the pricing of corporate debt: The risk structure of interest rates. Journal
of Finance, 29, 449–470.
Sklar, A. (1959). Fonctions de repartition a n dimensions et leurs marges. Public Institute Statistic
University of Paris, 8, 229–231.
Stein, C. (1972). A bound for the error in the normal approximation to the distribution of a sum of
dependent random variables. In Proceedings of the sixth Berkeley symposium on mathematical
statistics and probability (pp. 583–602). Berkeley: University of California Press.
Walker, M. B. (2006). Credit default swaps with counterparty risk: A calibrated Markov model.
Journal of Credit Risk, 2(1), 31–49.
Wu, P. C. (2010). Applying a factor copula to value basket credit linked notes with issuer default
risk. Finance Research Letters, 7(3), 178–183.
Wu, P. C., Kao, L. J., & Lee, C. W. (2011). How issuer default risk affects basket credit linked note
coupon rate. International Journal of Information and Management Sciences, 22(1), 59–71.
Panel Data Analysis and Bootstrapping:
Application to China Mutual Funds 24
Win Lin Chou, Shou Zhong Ng, and Yating Yang

Contents
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
24.2 Wild-Cluster Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 660
24.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
24.3.1 Data and Definitions of the Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
24.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
24.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
Appendix 1: Wild-Cluster Bootstrap Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667

Abstract
Thompson (Journal of Financial Economics 99, 1–10, 2011) argues that double
clustering the standard errors of parameter estimators matters the most when the
number of firms and time periods are not too different. Using panel data of
similar number in firms and time periods on China’s mutual funds, we estimate
double- and single-clustered standard errors by wild-cluster bootstrap procedure.
To obtain the wild bootstrap samples in each cluster, we reuse the regressors
(X) but modify the residuals by transforming the OLS residuals with weights
which follow the popular two-point distribution suggested by Mammen (Annals
of Statistics 21, 255–285, 1993) and others. We then compare them with other

W.L. Chou (*)

Department of Economics and Finance, City University of Hong Kong, Hong Kong, China
e-mail: wlchou@e.cuhk.edu.hk
S.Z. Ng
Hong Kong Monetary Authority, Hong Kong, China
e-mail: bszng@hkma.gov.hk
Y. Yang
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: yatingyang.iof98g@nctu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 657
DOI 10.1007/978-1-4614-7750-1_24,
# Springer Science+Business Media New York 2015
658 W.L. Chou et al.

estimates in a set of asset pricing regressions. The comparison indicates that

bootstrapped standard errors from double clustering outperform those from
single clustering. Our findings support Thompson’s argument. They also suggest
that bootstrapped critical values are preferred to standard asymptotic t-test
critical values to avoid misleading test results.

Keywords
Asset-pricing regression • Bootstrapped critical values • Cluster standard
errors • Double clustering • Firm and time effects • Finance panel data • Single
clustering • Wild-cluster bootstrap

24.1 Introduction

Researchers using finance panel data have increasingly realized the need to account
for the residual correlation across both firms and/or time in estimating standard
errors of regression parameter estimates. Ignoring such clustering can result in
biased OLS standard errors. Two forms of residual dependence that are common in
finance applications are time-series dependence and cross-sectional dependence.
The former is called a firm effect, whereas the latter a time effect. The usual
solution to account for the residual dependence is to compute clustered standard
errors. The notable examples are Petersen (2009) and Thompson (2011).
Using the Monte Carlo-simulated panel data, Petersen (2009) compares the
performance of many different standard error estimation methods surveyed in the
literature. These methods include White’s heteroskedasticity-robust standard
errors, single clustering (by firm or by time), and double clustering (by both firm
and time). His findings suggest that the performance of different methods depends
on the forms of residual dependence. For example, in the presence of a firm effect,
the clustered standard errors are unbiased and can produce correctly sized
confidence intervals while those estimated by OLS, White, or Fama-MacBeth
method are biased.
Much of the analysis in Petersen (2009) is based on the simulated panel data set
whose data structure is certain. With the simulated panel data set, it is easier to
choose among the estimation methods. This paper chooses an alternative method,
namely, bootstrapping, to investigate the performance of standard errors estimated
by White’s OLS and single- and double-clustering methods with actually observed
data. The use of the bootstrap method is motivated by Kayhan and Titman (2007)
who show that bootstrapped standard errors are robust to heteroskedasticity, and
serial correlation problems in panel finance data applications. Moreover, despite
the wide use of the bootstrap in statistical and econometric applications, the
survey finding of Petersen (2009) found that the bootstrap applications are
relatively scarce in the finance literature. Hence, it may be of some interest to
investigate the bootstrapping application to a set of panel finance data on
Chinese mutual funds.
24 Panel Data Analysis and Bootstrapping: Application to China Mutual Funds 659

Table 24.1 Summary statistics for variables used in China’s mutual fund regressions
(Sept 2002–Aug 2006)
Std. Normality test stat.
Variable Mean deviation Skewness Kurtosis W-sq. A-sq.
Mutual funds 0.0006 0.0580 0.1313 0.1489 0.1809*** 0.2062***
returns
Market 0.0025 0.0465 0.1554
0.7801 2.4993*** 18.3650***
excess
returns
Notes: Sample size N ¼ 2,592. Individual mutual funds return is the dependent variable, while the
market excess return is the independent variable. W-sq or W2 ¼ Cramer-von Mises test statistic, and
A-sq or A2 ¼ Anderson-Darling test statistic. Both test statistics are empirical distribution function (EDF)
statistics. The computing formulas for W2 and A2 statistics are available in Stephens (1974), p. 731
***
denotes statistical significance at the 1 % level

The bootstrap method is applied to a panel data set on the monthly returns for
54 Chinese mutual funds over the period of September 2002–August 2006. The data
set is applied to a set of asset pricing regressions. Table 24.1 contains summary
statistics such as sample skewness, sample excess kurtosis, and two test statistics for
normality for the variables used in the asset pricing regressions. They suggest that
normality does not characterize the variables. Additionally, since the time-series
and/or cross-sectional independence assumption is most likely to be violated in panel
data sets, ignoring these dependence could result in biased estimates of the standard
errors. As evidenced in Kayhan and Titman (2007), bootstrapping is a possible
alternative to handle this dependence issue.
In this paper, we are particularly interested in the performance of the bootstrapped
double-clustered standard error estimates, because Thompson (2011) has argued that
double clustering matters most when the number of firms and time periods are not too
different. Given the panel data set we have collected which consists of 54 China mutual
fund returns for 48 months with data exhibiting firm and time effects, double clustering
is likely to show a significant difference. Our findings show that the bootstrapped
standard errors from double clustering leads to more significant test results. We also
demonstrate the importance of using bootstrapped critical values in hypothesis testing.
A number of bootstrap procedures are available in the literature. The
bootstrap procedure we consider in this paper is the wild-cluster bootstrap procedure,
which is an extended version of the wild bootstrap proposed by Cameron et al. (2008)
in a cluster setting. This procedure has been shown by Cameron et al. (2008) to perform
very well in practice, despite the fact that the pairs cluster bootstrap works well in
principle. In this paper, our comparison of the finite-sample size of the bootstrapped
t-statistics resulting from the pairs cluster bootstrap and wild-cluster bootstrap also
indicates that the wild-cluster bootstrap performs better.
The rest of the paper is organized as follows. Section 24.2 presents the wild-cluster
bootstrap procedure. Section 24.3 discusses the empirical results and the last section
gives conclusions.
660 W.L. Chou et al.

24.2 Wild-Cluster Bootstrap

The bootstrap we use in this paper is known as the “wild-cluster bootstrap”

which is based on the nonclustered wild bootstrap proposed by Wu (1986).
Proofs of the ability of the wild bootstrap to provide refinements in the
linear regression model for linear regression models with heteroskedastic errors
can be found in Liu (1988) and Mammen (1993). Cameron et al. (2008)
extended Wu’s (1986) wild bootstrap to the clustered setting.
The wild-cluster bootstrap procedure involves two stages. In the first stage,
we consider the asset pricing model with G clusters (subscripted by g), and
0
with Ng observations within each cluster, namely, yg ¼ Xg b + eg, g ¼ 1, . . . ,G,
b is k  1, Xg is Ng  k, and yg and eg are Ng  1 vectors. We fit the model to
the actually observed panel data set by OLS and estimate White’s
heteroskedasticity-robust standard errors, as well as standard errors,
clustered by firm, by time, and by both. We then save residuals and
denote them as êg.
The second stage is the 
resampling

procedure  which creates bootstrap
samples for each cluster, ^y 1 ; X1 ; . . . ; ^y G ; Xg
0
where ^y g = Xg b^ þ e . For
g
each bootstrap sample in a cluster, the explanatory variables are reused and
unchanged. The residuals eg are constructed according to eg = ag^e g , where the
weight ag serves as a transformation of the OLS residuals ^e g . A variety of
constructions of weights ag have proposed in the literature1. We use the
two-point distribution of the weight variable ag suggested in Mammen (1993),
Brownstone and Valletta (2001), and Davidson and
Flachaire pffiffiffi (2008), namely, ag
which takes on
one of the following
pffiffiffi
pffiffiffi values: (i) 1


p5ffiffiffi=2 
0:6180 with
probability 1
þ 5p=ffiffiffi2
5pffiffi ffi 0:7236 or (ii) 1 þ 5 =2  1:6180 with
probability 1
1 þ 5 = 2 5  0:2764 . Note that this random variable ag
has a mean zero with variance equal to one and the constraint E(a3g ) ¼ 1. We
perform 1,000 replications. On each replication, a new set of eg is generated and
a new set of bootstrap-data is created based on ^y G = Xg b
0
^ þ e, and therefore a new
 g
set of parameter estimates, denoted as b ^ , is obtained.
For the 1,000 starred estimates (eg ), we calculate their bootstrapped
standard errors using different estimation methods. The bootstrapped test
statistics are calculated by dividing the 1,000 parameter estimates by the
corresponding bootstrapped standard errors. The bootstrapped critical values
can be obtained from the bootstrapped distribution of these test
statistics. A detailed explanation of the procedure we follow is documented
in Appendix 1.

1
For example, in Cameron et al. (2008), ag takes the value +1 with probability 0.5, or the value
1
with probability 1–0.5.
24 Panel Data Analysis and Bootstrapping: Application to China Mutual Funds 661

24.3 Empirical Results

24.3.1 Data and Definitions of the Variables

The Data. The sample consists of the returns on 54 publicly traded closed-end
mutual funds that are gathered for 48 months from September 2002 to August 2006,
a total of 2,592 observations. The mutual fund data set is purchased from the GTA
Information Technology Company, Shenzhen, China. For simplicity, we divide the
mutual funds investment objectives into equity growth and nongrowth funds. Our
sample consists of 37 (68.5 %) growth funds and 17 (31.5 %) nongrowth funds.
Although the first closed-end fund in China was sold to the public in April 19982,
complete data for all 54 mutual funds are collected from September 2002.
The summary statistics for the variables used in China’s mutual fund return
regressions are displayed in Table 24.1, and the test statistics for normality for these
variables suggest that the variables are non-normal.
Definitions of the variables. The following are the definitions of the variables
used in the estimation procedures: Ri, t ¼ the return of a mutual fund i in excess of
the risk-free rate in month t. Savings deposit rate of the People’s Bank of China3 is
used as the proxy for the risk-free rate. Rm, t ¼ is the market return in excess of the
risk-free rate in month t and is calculated4 as follows:

Rm, t ¼ 0:4  R1, t þ 0:4  R2, t þ 0:2  0:0006, (24.1)

where R1 is the monthly return on Shanghai Stock Exchange index, R2 the monthly
return on Shenzhen Stock Exchange index, and 0.06 % is the monthly return on
savings deposits.

24.3.2 Results

Table 24.2 presents the results from a set of asset pricing regressions of China
mutual fund returns on its market returns. The firm and time effect in OLS residuals
and data can be seen graphically in Fig. 24.1. Figure 24.1, Panel A, shows the
within-firm autocorrelations in OLS residuals and independent variable, respec-
tively, for lags 1–12. Panel B of Fig. 24.1 displays the within-month autocorrela-
tions for residuals for lags 1–12. As the independent variable is a monthly series
without cross-sectional units, we cannot calculate its within-month autocorrela-
tions, thus no within-month plot is available for the independent variable.
Figure 24.1 suggests that the residuals exhibit both firm and time effects, whereas

2
Chen and Lin (2006), p. 384
3
Data are taken from the website of the People’s Bank of China: http://www.pbc.gov.cn/publish/
zhengcehuobisi/627/index.html.
4
The calculation follows that of Shen and Huang (2001), p. 24.
662 W.L. Chou et al.

Table 24.2 Application to asset pricing modeling

t-statistics
Clustered by
White Firm Time Firm and time
Regressor Estimate I II III IV
Rm, t 0.8399 46.935*** 65.492*** 9.017*** 9.099***
1 % critical value (CV) 2.576 2.576 2.576 2.576
Intercept
0.0016
1.858
2.222**
0.326
0.327
1 % critical value (CV)
2.576
2.576
2.576
2.576
Coefficient estimates OLS
R-squared 0.4539
Regressor Estimate V VI VII VIII
Rm, t 0.8401 47.215*** 66.808*** 50.886*** 73.672***
Bootstrapped 1 % CV 1.951 2.661 2.520 11.898
Intercept
0.0016
1.878**
2.262**
2.033**
2.659
Bootstrapped 1 % CV
2.300
2.717
2.371
12.288
Coefficient estimates Wild-cluster
bootstrap
R-squared 0.4579
Notes: The dependent variable is the monthly mutual fund return in excess of risk-free rate,
denoted as Ri,t, and the independent variable Rm,t is the market returns in excess of risk-free rate.
Both variables are monthly observations from September 2002 to August 2006
***
, and ** denote statistical significance at the 1 %, and 5 % levels, respectively

independent variable shows firm effects. Given Thompson’s (2011) argument that
double clustering matters most when the number of firms and time periods are not
too different, our data set which has the number of mutual funds (54) similar to
the number of months (48) is expected to imply that double clustering is important
in our analysis.
In Table 24.2, the second column presents the OLS parameter estimates, whereas
the remaining columns report their corresponding t-statistics by dividing each param-
eter estimate by its corresponding standard error. These t-statistics indicate all beta
coefficients are statistically significant at the 1 % level, whereas the intercept is only
significant in one case (under column II) when the standard error computed by single
clustering by firm is used. More importantly, the t-statistics in Table 24.2 enable us to
compare the clustered standard errors constructed from double and single clustering
with the OLS White estimate. Notice that the t-statistic for beta coefficient
obtained from double clustering (SÊboth) is 9.10 (column IV) which is much smaller
than 46.9 calculated using the White method (column I), indicating the presence of
firm and time effects. It also means that the double-clustering standard errors are
much larger. A comparison of t-statistics in columns III and IV implies the SÊboth of
the beta coefficient (9.10) is similar to the standard error clustered by time which is
9.01. This means that the firm effects do not matter much. The comparison reveals
that OLS White standard errors are underestimated when residuals exhibit both firm
and time effects.
24 Panel Data Analysis and Bootstrapping: Application to China Mutual Funds 663

PANEL A: Within Firm

0.3
residual Rm,t
0.2
Autocorrelation

0.1

−0.1 1 2 3 4 5 6 7 8 9 10 11 12

−0.2

−0.3
Lag

PANEL B: Within Month

0.25
0.2
0.15
0.1
Autocorrelation

0.05
0
1 2 3 4 5 6 7 8 9 10 11 12
−0.05
−0.1
−0.15
−0.2
−0.25
Lag

Fig. 24.1 The autocorrelations of residuals and the independent variable are plotted for 1–12
lags. The solid lines in Panel A and B show, respectively, within-firm and within-month autocor-
relations in residuals, whereas the dashed line in Panel A shows the within-firm autocorrelations in
the independent variable

Turning to the results obtained by the wild-cluster bootstrapping, the story changes.
The bootstrapped t-statistics of the beta coefficient estimates displayed in columns
V–VIII of Table 24.2 differ quite significantly from those in columns I–IV. The
t-statistic is 47.2 when the bootstrapped White standard error is used and 66.8 if the
bootstrapped standard error clustered by the firm is used. This means the firm effect is
significant in the data. By a similar comparison, the bootstrapped t-statistic is 73.7 when
the double-clustered standard error is used, meaning both the time and firm effects are
strong in the residuals. The bootstrapped t-statistic is 50.9 when the bootstrapped
standard error clustered by time is used implying the time effect exists in the data.
These comparisons suggest both the firm and time effects matter in the computation of
the bootstrapped standard errors using double as well as single clustering. Their
implication is that we might follow what Kayhan and Titman (2007) have done in
their study to simply compute the bootstrapped standard errors with our panel data set.
664 W.L. Chou et al.

Table 24.3 Bootstrapped critical values

Percentiles 0.005 0.025 0.05 0.95 0.975 0.995
OLS White
Rm,t
1.842
1.396
1.174 1.213 1.490 1.951
Intercept
2.300
1.656
1.396 1.385 1.587 2.091
Clustered by firm
Rm,t
2.621
1.984
1.698 1.766 2.007 2.661
Intercept
2.717
2.035
1.700 1.778 2.014 2.672
Clustered by time
Rm,t
2.513
1.795
1.510 1.517 1.821 2.520
Intercept
2.371
1.956
1.608 1.629 1.917 2.364
Clustered by firm and time
Rm,t
6.517
4.376
3.491 3.081 4.416 11.898
Intercept
12.288
4.053
2.792 3.136 4.384 8.168
Note: Critical values are obtained by implementing the bootstrap procedure presented in the
Appendix 1

Table 24.4 Rejection rates using wild-cluster bootstrapped critical values

OLS White Clustered by firm Clustered by time Clustered by firm and time
Rm,t 0.050 0.067 0.053 0.052
Intercept 0.042 0.034 0.037 0.036
Note: The number of replications is 1,000

The statistical significance of the bootstrapped t-statistics of the beta coefficient

estimates is determined by using the bootstrapped critical values reported
in Table 24.3. Compared with the bootstrapped critical values presented in
Table 24.3, we notice that all bootstrapped t-statistics constructed from bootstrapped
standard errors are statistically significant at the 1 % level. The intercept is now
significant in three cases (columns V–VII) when the standard errors were computed
by White and by single clustering. It is noteworthy that in Table 24.3 the
bootstrapped critical values on beta coefficient estimates when double clustering is
used are numerically larger than the corresponding asymptotic t-test critical values
of 2.58 (1 %), 1.96 (5 %), and 1.65 (10 %), indicating that the use of the large-sample
(normal approximation) critical values can lead to misleading test results when both
firm and time effects exist in the residuals. On the other hand, intercept coefficient
in column VIII was not significant when bootstrapped double clustering is
used. Interestingly, we observe from Table 24.3 that bootstrapped critical values
differ considerably depending on different standard error estimation methods.
We now examine the finite-sample size of the bootstrapped t-statistics assuming
the beta coefficient takes the OLS estimate under the null hypothesis. Table 24.4
shows that for the bootstrapped t-statistics, no serious size distortion is found as
24 Panel Data Analysis and Bootstrapping: Application to China Mutual Funds 665

Table 24.5 Rejection rates using conventional critical values

OLS White Clustered by firm Clustered by time Clustered by firm and time
Rm,t 0.004 0.069 0.037 0.206
Intercept 0.012 0.046 0.036 0.203
Note: The number of replications is 1,000

reflected by the close to 5 % size values. However, for the tests based on the
standard asymptotic t-test critical values, all tests suffer from serious size distortion.
For example, those based on the OLS White and clustered by time are undersized,
while others oversized (see Table 24.5).

24.4 Conclusion

In this paper, we examine the performance of single- and double-clustered standard

errors using the wild-cluster bootstrap method. The panel data set on the Chinese
mutual funds used in the analysis has similar number of firms (54) and time periods
(48); we are particularly interested in the performance of the bootstrapped
double-clustered standard errors. This is mainly due to the conclusion made in
Thompson (2011) that double clustering the standard errors matters the most when
the number of firms and time periods are not too different.
In the presence of firm and time effects, the standard OLS White standard errors
are found to be underestimated when compared to standard errors computed from
double clustering (columns I–IV, Table 24.2). Further, the wild-cluster
bootstrapped standard errors are found to account for the firm and time effects in
residuals, as evidenced in column VIII of Table 24.2. The bootstrapped t-statistic
computed by OLS White method is found to be much smaller than that calculated
from the double clustering, suggesting that the firm and time effects in
residuals are strong.
The size values for the test statistics of the beta coefficient estimates
in Table 24.4 suggest that the bootstrapped double clustering outperforms the
single clustering either by firm or by time. They support Thompson’s (2011)
argument that double clustering the standard errors of parameter estimators matters
the most when the number of firms and time periods are not too different.
Size distortions reported in Table 24.5 imply that it may not be appropriate to
compare the bootstrapped t-statistics with standard t-test critical values. These
findings also suggest that to avoid obtaining misleading test results with the
presence of either firm or time effects or both, the bootstrapped critical values are
preferred to conventional critical values. Additionally, a comparison of the sizes
displayed in Table 24.4 with those calculated using the pairs cluster bootstrap
method shown in Table 24.6 also suggests the wild-cluster bootstrap approach
performs better.
666 W.L. Chou et al.

Table 24.6 Rejection rates using pairs cluster bootstrapped critical values
OLS White Clustered by firm Clustered by time Clustered by firm and time
Rm,t 0.043 0.040 0.040 0.040
Intercept 0.062 0.048 0.059 0.060
Note: The number of replications is 1,000

Appendix 1: Wild-Cluster Bootstrap Procedure

The following steps are used to obtain the wild-clustered bootstrapped standard
errors and critical values:5
(i) Define firm effects, and time effects. The asset pricing model using individual
variables Ri,t and Rm,t is specified as

Rit ¼ b0 þ Rm, t b1 þ eit , i ¼ 1, . . . , N, t ¼ 1, . . . , T (24.2)

where the variables Ri,t and Rm,t are, respectively, the return of mutual fund
i and the market return in excess of risk-free rate in month t. b0 and b1
are unknown parameters. The construction of the variables is detailed in
Sect. 24.3.1. eit is the error term. It may be heteroskedastic but is assumed to
be independent of the explanatory variable E(eit | Rm,t) ¼ 0.
Following Thompson (2011), we make the following assumptions on the
correlations between errors, eit:
(a) Firm effects: The errors may be correlated across time for a particular firm,
that is, E(eit, eik | Rm,t,Rm,k) 6¼ 0 for all t 6¼ k.
(b) Time effects: The errors may be correlated across firms within the same time
period, that is,

 
E eit , ejt Rm, t 6¼ 0 for all i 6¼ j:

Let G be the number of clusters, and let Ng be the number of observations

within each cluster. The errors are assumed to be independent across clusters
but correlated within clusters. The asset pricing model can be written as

Rig ¼ b0 þ Rm b1 þ eig , i ¼ 1, . . . , Ng g ¼ 1, . . . , G,
(24.3)
Rg ¼ db0 þ Rmg b1 þ eg , g ¼ 1, . . . , G,

where Rig, Rm, and eig are scalars; Rg, Rmg, and eg are Ng  1 vectors; and
d is Ng  1 vector with all elements equal to 1.
(ii) Fit data to model. We fit model (Eq. 24.3) to the observed data using OLS
and obtain the parameter estimates b ^ and b ^ together with the OLS residuals
0 1
êg, g ¼ 1,. . ., G

5
See also Cameron et al. (2008) for details
24 Panel Data Analysis and Bootstrapping: Application to China Mutual Funds 667

(iii) Construct 1,000 bootstrap samples. The bootstrap-residuals are obtained

according to the following transformation

pffiffirelation:
ffi eg = ag^e g, where ag takes
on one of the following values: (i) 1
5 =2 
0:6180 with probability

pffiffiffi
pffiffiffi
pffiffiffi
1 þ 5 = 2 5  0:7236 or (ii) 1 þ 5 =2  1:6180 with probability

pffiffiffi
pffiffiffi
1
1 þ 5 = 2 5  0:2764:
Hence,
(a) For each cluster g = 1, . . . , G, set eg = 1:618^e g with probability 0.2764
or eg =
0:618^e g with probability 0.7236.
(b) Repeat (a) 1,000 times to obtain eg and then construct the bootstrap
samples Rg as follows:

Rg ¼ db ^ þ e :
^ þ Rmg b (24.4)
0 1 g

(iv) With each pseudo sample generated in step (iii), we estimate the parameters b ^
0
^  by OLS, White standard errors (SÊwhite)which are OLS standard errors
and b 1
robust to heteroskedasticity, as well as standard errors clustered by firm (SÊfirm), by
time (SÊtime), and by both firm and time (SÊboth). The simulations are performed
using GAUSS 9.0. The standard error formulas can be found in Thompson (2011).
(v) Construct bootstrapped test statistics by taking ratios of b ^  ði = 0, 1Þ obtained
i
by OLS to its corresponding SÊwhite, SÊfirm, SÊtime, and SÊboth obtained in
step (iv). More specifically, the bootstrapped test statistics are expressed as
follows:
^
b
b ^
wi ¼ i   i , i ¼ 0, 1, (24.5)
^ b
SE ^
i
 
where SE ^ b^ can either be SÊwhite, SÊfirm, SÊtime, or SÊboth.
i
(vi) Obtain the empirical distribution of the individual test statistics by sorting the
1,000 test statistics computed in step (v) in an ascending order. Bootstrapped
critical values are then obtained from this empirical distribution at the following
quantiles: 0.5 %, 2.5 %, 5 %, 95 %, 97.5 %, and 99.5 %, respectively.

References
Brownstone, D., & Valletta, R. (2001). The bootstrap and multiple imputations: Harnessing
increased computing power for improved statistical tests. Journal of Economic Perspectives,
15, 129–141.
Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). Bootstrap-based improvements for
inference with clustered errors. Review of Economics and Statistics, 90, 414–427.
Chen, Z., & Lin, R. (2006). Mutual fund performance evaluation using data envelopment analysis
with new risk measures. OR Spectrum, 28, 375–398.
Davidson, R., & Flachaire, E. (2008). The wild bootstrap, tamed at last. Journal of Econometrics,
146, 162–169.
Kayhan, A., & Titman, S. (2007). Firms’ histories and their capital structures. Journal of Financial
Economics, 83, 1–32.
668 W.L. Chou et al.

Liu, R. Y. (1988). Bootstrap procedures under some non-IID models. Annals of Statistics, 16,
1696–1708.
Mammen, E. (1993). Bootstrap and wild bootstrap for high dimensional linear models. Annals of
Statistics, 21, 255–285.
Petersen, M. A. (2009). Estimating standard errors in finance panel data sets: Comparing
approaches. The Review of Financial Studies, 22, 435–480.
Shen, W. T., & Huang, S. L. (2001). An empirical analysis of China’s mutual funds: Performance
and evaluation. Economic Research, 9, 22–30 (in Chinese).
Stephens, M. A. (1974). EDF statistics for goodness of fit and some comparisons. Journal of the
American Statistical Association, 69, 730–737.
Thompson, S. B. (2011). Simple formulas for standard errors that cluster by both firm and time.
Journal of Financial Economics, 99, 1–10.
Wu, C. F. G. (1986). Jackknife, bootstrap and other resampling methods in regression analysis.
Annals of Statistics, 14, 1261–1295.
Market Segmentation and Pricing of
Closed-End Country Funds: 25
An Empirical Analysis

Dilip K. Patro

Contents
25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
25.2 Theoretical Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
25.2.1 Pricing of Country Funds in the Context of International CAPM . . . . . . . . . . . . 673
25.2.2 Cross-Sectional Variations in Country Fund Premiums . . . . . . . . . . . . . . . . . . . . . . 674
25.2.3 Conditional Expected Returns and Pricing of Country Funds . . . . . . . . . . . . . . . . 676
25.2.4 Conditional Expected Returns and Time-Varying Country
Fund Premiums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
25.3 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
25.4 Empirical Results: Pricing of Country Funds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
25.4.1 Unconditional Risk Exposures of Country Funds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
25.4.2 Pricing of Country Funds in the Context of International CAPM . . . . . . . . . . . . 689
25.4.3 Cross-Sectional Variations in Country Fund Premiums . . . . . . . . . . . . . . . . . . . . . . 690
25.4.4 Predictability of Closed-End Country Fund Returns . . . . . . . . . . . . . . . . . . . . . . . . . 693
25.4.5 Conditional Expected Returns and Pricing of Country Funds . . . . . . . . . . . . . . . . 694
25.4.6 Conditional Expected Returns and Time-Varying Country
Fund Premiums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 700
25.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 701
Appendix 1: Generalized Method of Moments (GMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 701
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703

Abstract
This paper finds that for closed-end country funds, the international CAPM can
be rejected for the underlying securities (NAVs) but not for the share prices. This
finding indicates that country fund share prices are determined globally where as
the NAVs reflect both global and local prices of risk. Cross-sectional variations
in the discounts or premiums for country funds are explained by the differences

The views expressed in this paper are strictly that of the author and not of the OCC or the US
department of Treasury.
D.K. Patro
RAD, Office of the Comptroller of the Currency, Washington, DC, USA
e-mail: dilip.patro@occ.treas.gov

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 669
DOI 10.1007/978-1-4614-7750-1_25,
# Springer Science+Business Media New York 2015
670 D.K. Patro

in the risk exposures of the share prices and the NAVs. Finally, this paper shows
that the share price and NAV returns exhibit predictable variation and country
fund premiums vary over time due to time-varying risk premiums. The paper
employs generalized method of moments (GMM) to estimate stochastic
discount factors and examines if the price of risk of closed-end country fund
shares and NAVs is identical. GMM is an econometric method that was a
generalization of the method of moments developed by Hansen (Econometrica
50, 1029–1054, 1982). Essentially GMM finds the values of the parameters so
that the sample moment conditions are satisfied as closely as possible.

Keywords
Capital markets • Country funds • CAPM • Closed-end funds • Market
segmentation • GMM • Net asset value • Stochastic discount factors • Time-
varying risk • International asset pricing

25.1 Introduction

The purpose of this paper is to provide an empirical analysis of pricing of closed-

end country equity funds in the context of rational asset-pricing models which
account for the role of market segmentation and time-varying risk premiums.
Specifically, the paper addresses the following issues. How are country fund
share prices and net asset values (NAVs) determined? What are the implications
of differential pricing of closed-end fund shares and NAVs for cross-sectional and
time-series variations in the premiums of the funds? The answers to these questions
contribute to the burgeoning literature on country funds.
Closed-end country equity funds are a relatively recent innovation in international
capital markets. Whereas only four closed-end country equity funds traded in New
York at the end of 1984, currently there are 94 closed-end country equity funds targeting
over 31 countries. Past researchers have examined issues related to the benefits of
diversification from holding these funds (Bailey and Lim 1992; Chang et al. 1995;
Bekaert and Urias 1996) and how they should be designed and priced (Hardouvelis
et al. 1993; Diwan et al. 1995; Bodurtha et al. 1995). Relatively unexplored is how the
expected returns of the country fund share prices and NAVs are determined, in
a framework of market segmentation and time-varying risk premiums.
Country fund share prices are determined in the USA, but their NAVs are
determined in the country of origin of the fund. Models of international
asset pricing (e.g., Errunza and Losq 1985) and models of country fund pricing
(e.g., Errunza et al. 1998) suggest that the expected returns on country fund share
prices should be determined by their covariances with the world market portfolio.
However, if there are barriers to investment, such as limits on ownership, the capital
markets may be segmented. In such a segmented market, the expected returns of the
NAVs will be determined by their covariances with the world market portfolio as
well as the home market portfolio, to the extent that the home market is not spanned
by the world market.
25 Market Segmentation and Pricing of Closed-End Country Funds 671

The foundation of this paper is in such a framework of market segmentation and

its implications for international capital market equilibrium. In such a market
structure where the local investors have complete access to the world market but
the foreign investors (e.g., the US investors) have only partial access to the local
equity market in terms of the percentage of equity of a firm they can own, the prices
paid by foreigners relative to local investors can be higher due to the limited supply
of the local securities. To preclude arbitrage, it is assumed that the local investors
cannot buy the securities at a lower price and sell it to the foreign investor at
a higher price. Thus, foreign investors are willing to pay a premium due to the
diversification benefits from adding that security to their purely domestic portfolio.
The premium arises only if the cash flows of the local security are not spanned by
the purely domestic portfolio for the foreign investor. Since closed-end country
fund’s share prices and NAVs are determined in two separate markets, the expected
returns on the prices and NAVs can be different, leading to premiums, both positive
and negative. Thus, barriers to investments may be sufficient to generate
premiums.1
Even in the absence of institutional restrictions on foreign equity ownership or
even in a purely domestic context (see, e.g., Pontiff 1997), it is still possible for
observed price returns to be much more volatile than NAV returns. As shown in the
academic literature, closed-end country fund share prices and NAVs are not perfect
substitutes. The sources of these differences may be attributable to differences in
numeraires, information, and possibly noise trading causing excess volatility in the
share price returns (see Lee et al. 1991). Apart from investment restrictions, these
imperfections alone may be sufficient to generate premiums. However, Bonser-
Neal et al. (1990) document that relaxation of investment restrictions leads to
a decrease in the share price-NAV ratio for a sample of country equity funds.
Further, Bodurtha et al. (1995) document that the correlation of changes in country
fund premiums and domestic fund premiums is low and insignificant, indicating
that the structure of international capital markets is an important contributor in
determining premiums.
Thus, the objective of this paper is to provide an analysis of what explains the
expected returns on closed-end country fund share prices and NAVs in a segmented
market framework. The paper utilizes both unconditional and conditional tests
on mean-variance efficiency of the world market index (as proxied by the
Morgan Stanley Capital International world index) and provides results on the
cross-sectional and time-series variations of premiums across closed-end country
equity funds. In addition, the paper employs generalized method of moments
(GMM) as discussed in Appendix 1 to estimate stochastic discount factors
and examines if the price of risk of closed-end country fund shares and NAVs
is identical.

1
Hence forth both premiums and discounts are referred to as premiums. Therefore, a discount is
treated as a negative premium.
672 D.K. Patro

The sample consists of 40 closed-end country equity funds. Twenty of the funds
are from developed markets and 20 funds are from emerging markets.2 The main
empirical findings of the paper are as follows. For country fund share prices, the
hypothesis that the unconditional international CAPM is a valid model cannot be
rejected. However, for the NAVs the international CAPM can be rejected.
This finding suggests that country fund share prices and NAVs are not priced
identically. The share prices reflect the global price of risk only, but the NAVs
reflect both the global and the local prices of risk. It is shown that the differences in
risk exposure to the world market index of share prices and NAVs can explain up to
18.7 % of the cross-sectional variations in premiums for developed market funds,
but only 1.9 % of the variation for emerging market funds.
When conditioning on information is allowed, the international CAPM explains
country fund share returns and NAV returns for both developed and emerging
markets. However, the hypothesis that the price of risk is identical between
closed-end fund shares and NAVs can be rejected for alternate stochastic discount
factors for majority of the markets. This finding is consistent with market segmen-
tation. Therefore, differential pricing of the fund shares and the underlying portfolio
causes expected returns to be different and explains the existence of premiums and
their variation over time. Finally, it is shown that the country fund premiums vary
over time due to differential conditional risk exposures of the share prices
and NAVs.
The principal contributions of this paper are as follows. Existence of premiums
on closed-end funds has been a long-standing anomaly. In the domestic setting,
negative premiums have been the norm, which has been attributed to taxes,
management fees, illiquid stocks, or irrational factors (e.g., noise trading).
Although these factors may also be important in explaining country fund premiums,
unlike domestic closed-end funds, country fund share prices and NAVs are deter-
mined in two different market segments. This paper provides a rational explanation
for the premiums on country funds based on differential risk exposures of the share
prices and NAVs and market segmentation. Unlike the noise trading hypothesis
which assumes the presence of “sentiment” or an additional factor for the share
prices but not the NAVs, this paper shows that the same factor may be priced for the
share prices and the NAVs, but priced differently. The differential risk exposures
are shown to be among the significant determinants of cross-sectional variations in
the premiums. Further, this paper examines the role of time variations in expected
returns for country fund returns and attributes the time-varying country fund
premiums to time-varying country fund returns.
The paper is organized as follows. Section 25.2 presents the theoretical motiva-
tion and the hypotheses. Section 25.3 presents the data and the descriptive statistics.
Section 25.4 provides the empirical results for pricing of country funds. Concluding
remarks are presented in the last section.

2
In 1993, the World Bank defined an emerging market as a stock market in a developing country
with a GNP per capita of $8,625 or less. This is the definition of an emerging market in this paper.
25 Market Segmentation and Pricing of Closed-End Country Funds 673

25.2 Theoretical Motivation

This section first focuses on the theoretical motivation for the empirical testing of the
unconditional mean-variance efficiency of the world market index in the context of
country funds. The following subsections present tests for cross-sectional variations in
country fund premiums and the methodology for pricing country funds using stochas-
tic discount factors. Finally, this section outlines the methodology for explaining the
time variations in country fund premiums attributable to time-varying risk premium.

25.2.1 Pricing of Country Funds in the Context of

International CAPM

In a global pricing environment, the world market portfolio surrogates the market
portfolio. If purchasing power parity is not violated and there are no barriers to
investment, the world market portfolio is mean-variance efficient (see, e.g., Solnik
1996; Stulz 1995) and expected returns are determined by the international CAPM.
The international CAPM implies that the expected return on an asset is proportional
to the expected return on the world market portfolio:

E½ri ¼ bi E½rw  (25.1)

where, for any asset i, E[ri] is the expected excess return on the asset and E[rw] is the
expected excess return on the world market portfolio. The excess returns are
computed in excess of the return on a risk-free asset. Following a standard practice
(see Roll 1977), mean-variance efficiency of a benchmark portfolio can be
ascertained by estimating a regression of the form

ri, t ¼ ai þ bi, w rw, t þ ei, t (25.2)

where ri,t is the excess return on a test asset and rw,t is the return on a benchmark
portfolio.
The GMM-based methodology as outlined in Mackinlay and Richardson (1991) is
employed to test mean-variance efficiency of the world market index in the context of
country funds, using just a constant and the excess return on the world market index as
the instruments.3 The hypothesis that ai ¼ 0, for all i ¼ 1, . . ., N, where N is the
number of assets, is tested. This is the restriction implied by the international CAPM.
The joint hypothesis is tested by a Wald test for country fund share prices and NAVs
returns. Since country fund’s share prices are determined with complete access to
their markets of origin, their expected returns are expected to be determined via the

3
Although the evidence in favor of the domestic CAPM is ambiguous, many studies such as
Cumby and Glen (1990) and Chang et al. (1995) do not reject the mean-variance efficiency of the
world market index. Interestingly, although Cumby and Glen (1990) do not reject mean-variance
efficiency of the world market index, they reject mean-variance efficiency of the US market index.
674 D.K. Patro

international CAPM (see, e.g., Errunza and Losq 1985). However, if the country from
which the fund originates has restrictions on foreign equity investments, the expected
returns of the NAVs will be determined via their covariances with both the world
market portfolio and the part of the corresponding local market portfolio which is not
spanned by the world market portfolio (see Errunza et al. 1998).

Letting ei, t ¼ ri, t  ai  bi, w rw, t (25.3)

for the N assets, the residuals from the above equation can be stacked into
a vector et+1. The model implies that E[ei,t+1jZt] ¼ 0, for 8 i and 8 t. Therefore, E
[et+1  Zt] ¼ 0 for 8 t, where Zt is a set of predetermined instruments. GMM
estimation is based on minimizing the quadratic form f0 Of, where the sample
counterpart of E[et+1  Zt] is given by f ¼ {1/T[∑E[et+1  Zt]} and O is an optimal
weighting matrix in the sense of Hansen (1982).

25.2.2 Cross-Sectional Variations in Country Fund Premiums

If the mean-variance efficiency of the world market portfolio is not rejected, then
expected returns of the test assets are proportional to their “betas” with respect to
the world market portfolio. Therefore,

Eðra Þ ¼ lw ba, w where a ¼ p or n (25.4)

where E(rp) and E(rn) are the expected excess returns on the share prices and NAVs,
respectively, and lw is the risk premium on the world market index. Stulz and
Wasserfallen (1995) demonstrate that the logarithm of two share prices can be
written as a linear function of the differences in their expected returns.4 Therefore,


Prem ¼ logðP=NAVÞ ¼ Eðrn Þ  E rp (25.5)

where Prem is the premium on a fund calculated as the logarithm of the price-NAV
ratio. Combining Eqs. 25.4 and 25.5 leads to the testable implication


Prem ¼ lw bn, w  bp, w (25.6)

The above equation assumes that the international CAPM is valid and world
capital markets are integrated. In reality, the world markets may be segmented.
Therefore, the effect of market segmentation is captured by introducing additional

4
Stulz and Wasserfallen (1995) use the log-linear approximation of Campbell and Ammer to write
the logarithm of the price of a stock as ln(P) ¼ ESZi[(1  Z)dt+j+1  rt+j+1] + y, where Z is
a log-linear approximation parameter and rt+j+1 is the return from t+j to t+j+1. Assuming that Z is
the same for prices and NAVs, and the relation holds period by period, the premium can be written
as above.
25 Market Segmentation and Pricing of Closed-End Country Funds 675

variables based on prior research of Errunza et al. (1998) who document that measures
of access, spanning, and substitutability of the prices and NAVs have explanatory
power for the cross-sectional variation of premiums. The spanning measure (SPN) is
the conditional variance of the NAV return of a fund, unspanned by the US market
return and the fund’s share price returns, with specification as follows:
rn, t ¼ ai þ bi rUS, t þ bi rp, t þ ei, t (25.7)

where ei,t has a GARCH (1, 1) specification. The conditional volatility of ei,t is the
measure of spanning. The measure of substitution (SUB) is the ratio of conditional
volatilities of the share price and NAV returns of a fund not spanned by the US
market. The specifications for the expected return equations are

rn, t ¼ ai þ bi rUS, t þ en, t (25.8)

rp, t ¼ ai þ bi rUS, t þ ep, t (25.9)

where the error terms en,t and ep,t have GARCH (1, 1) specifications.5 The ratio of
the conditional volatilities of the residuals in Eqs. 25.8 and 25.9 is used as the
measure of substitutability of the share prices and NAVs.
Since it is difficult to systematically classify countries in terms of degree of access,
a dummy variable is used to differentiate developed and emerging markets. The
dummy variable takes value of one for developed markets. Another measure of
access, which is the total purchase of securities from a country by US residents as
a proportion of total global purchase, is also used as a measure of access to
a particular market. It is expected that the premiums are lower for countries with
easy access to their capital markets. Therefore, the coefficient on the access
variable is expected to be negative. The measure of spanning is interpreted as the
degree of ease with which investors could obtain substitute assets for the NAVs. As
noted earlier, even if there are barriers to foreign equity ownership, availability of
substitute assets can overcome the barriers. Since the spanning variable is
the volatility of the residual as specified in Eq. 25.7, it is expected to be positive.
The measure of imperfect substitutability of the share prices and NAVs is the ratio of
the variances of the share price and NAV returns. It is expected that the premium is
inversely related to this measure. Therefore, the extended cross-sectional equation is

Premi ¼ d0 þ d1 DIFi þ d2 SPNi þ d3 SUBi þ d4 ACCi þ ei (25.10)

where:
DIFi ¼ the difference of the betas on the world market index for the NAVs and the
share prices.
SPNi ¼ the conditional residual volatility from Eq. 25.7 used as measures of spanning.

5
Note that unlike here, Errunza et al. (1998) use returns on industry portfolios to proxy for the US
market.
676 D.K. Patro

SUBi ¼ the imperfect substitutability of the share prices and the NAVs proxied by
the ratio of conditional volatilities in Eqs. 25.8 and 25.9.
ACC ¼ a measure of access to a market proxied by a dummy variable which is
1 for developed markets [ACC(d)] or the total purchase of securities from the US
residents from that country as a fraction of global security purchases [ACC(cf)].
The empirical specification in Eq. 25.10 or its variations is used in the empirical
analysis for explaining cross-sectional variations in premiums. The higher the
difference of risk exposures, the higher is the premium. Therefore, the sign of d1
is expected to be positive. However, as the level of unspanned component is higher,
the level of premiums should be higher. Therefore the sign of d2 is expected to be
positive. Also, the higher the substitutability of the share prices and the NAVs, the
lower should be the premium. Therefore, d3 is expected to be negative. Since higher
access is associated with lower premiums, d4 is also expected to be negative.

25.2.3 Conditional Expected Returns and Pricing of Country Funds

If investors use information to determine expected returns, expected returns could

vary over time because of rational variations in risk premium. In such a scenario,
the premium on a country fund can be time varying as a result of the differential
pricing of the share prices and the NAVs.
Let Rtþ1 ¼ Ptþ1 þD
Pt
tþ1
, where Pt+1 is the price at time t+1 and Dt+1 is the dividend
at time t+1. Asset-pricing models imply that E[Mt+1Rt+1jZt] ¼ 1, where Zt is
a subset of the investor’s information set at time t (see, e.g., Ferson 1995). M is
interpreted as a stochastic discount factor (SDF). The existence of the above
equation derives from the law of one price. Given a particular form of an SDF, it
is estimated by using GMM as outlined earlier.6
The SDFs examined in this paper are the SDFs implied by the international
CAPM and its extension under market segmentation. The international CAPM is
obtained by assuming that the SDF is a linear function of the return on the world
market index (Rw, t+1). Specifically, the SDF is:

International CAPM: M ¼ l0 þ l1 Rw, tþ1 (25.11)

A two-factor SDF, where the second factor is the return on a regional index
(specifically for Asia or Latin America), is also estimated. Such a model is implied
by the models of Errunza and Losq (1985).7 Specifically, the SDFs is:
Two-factor model: M ¼ l0 þ l1 Rw, tþ1 þ l2 Rh, tþ1 (25.12)

6
Ferson and Foerster (1994) show that an iterated GMM approach has superior finite sample
properties. Therefore the iterated GMM approach is used in the estimations.
7
For such conditional representation, see, for example, Ferson and Schadt (1996).
25 Market Segmentation and Pricing of Closed-End Country Funds 677

Once an SDF is estimated for a group of share prices and NAVs, the coefficients
of the estimated SDFs are compared to test the hypothesis that for a given SDF, the
share price and NAV are priced identically.8

25.2.4 Conditional Expected Returns and Time-Varying Country

Fund Premiums

To examine the role of time-varying risk premiums in explaining the time variation
in country fund premiums, the following procedure is used. Similar to Ferson and
Schadt (1996), a conditional international CAPM in which the betas vary over time
as a linear function of the lagged instrumental variables is used. The following
equations are estimated via GMM:

rp, tþ1 ¼ ap þ bp, w ðZt Þrw, tþ1 þ ep, tþ1 (25.13)

rn, tþ1 ¼ an þ bn, w ðZt Þrw, tþ1 þ en, tþ1 (25.14)

where the excess return on the world market index is scaled by a set of instrumental
variables. Using Eq. 25.6 and the above time-varying betas, the premium can be
written as
 
Premtþ1 ¼ g0 þ g1 bn, w ðZt Þ  bp, w ðZt Þ þ eprem, tþ1 (25.15)

The empirical specification in Eq. 25.15 is used for explaining the time-varying
premiums. If the time-varying betas are significantly different and their difference
can explain the time variation of country fund premiums, the coefficient g1 is
expected to be positive and significant.

25.3 Data and Descriptive Statistics

The sample includes all single-country closed-end country equity funds publicly
trading in New York as of 31 August 1995. The test sample is limited to country
funds with at least 100 weeks of weekly NAV observations.9 An overview of the
sample is presented in Table 25.1. The funds are classified as developed market or

8
See Cochrane (1996) for such specifications. Cochrane (1996) calls such models as “scaled factor
models.” See Bansal et al. (1993) for a nonlinear specification of stochastic discount factors.
9
Unlike industrial initial public offerings (IPOs), closed-end fund IPOs are “overpriced” (Weiss
1989). Peavy (1990) finds that new funds show significant negative returns in the aftermarket.
Hanley et al. (1996) argue that closed-end funds are marketed to a poorly informed public and
document the presence of flippers – who sell them in the immediate aftermarket. They also
document evidence in support of price stabilization in the first few days of trading. Therefore,
the first 6 months (24 weeks) of data for each fund is excluded in the empirical analysis.
678 D.K. Patro

Table 25.1 Closed-end country funds: sample overview

No of shares Net assets
Fund name IPO date Symbol (millions) ($ millions) Listing
Panel A: Developed market funds
First Australia 16 December AUS 15.9 163.4 AMEX
fund 1985
Italy fund 27 February 1986 ITA 9.5 89.4 NYSE
Germany fund 18 July 1986 GER 13.5 173.9 NYSE
UK fund 6 August 1987 GBR 4.0 51.2 NYSE
Swiss Helvetia 19 August 1987 SHEL 9.2 181.8 NYSE
fund
Spain fund 21 June 1988 SPN 10.0 94.2 NYSE
Austria fund 22 September AUT 11.7 107.9 NYSE
1989
New Germany 24 January 1990 GERN 32.5 373.8 NYSE
fund
Growth fund of 14 February 1990 GSPN 17.3 164.0 NYSE
Spain
Future Germany 27 February 1990 GERF 11.9 171.5 NYSE
fund
Japan OTC 14 March 1990 JPNO 11.4 115.4 NYSE
equity fund
Emerging 29 March 1990 GERE 14.0 130.6 NYSE
Germany fund
Irish investment 30 March 1990 IRL 5.0 51.4 NYSE
fund
France growth 11 May 1990 FRA 15.3 168.4 NYSE
fund
Singapore fund 24 July 1990 SGP 6.9 97.4 NYSE
China fund 7 July 1992 CHN 10.8 136.3 NYSE
Jardine Fleming 17 July 1992 JCHN 9.1 114.5 NYSE
China fund
Greater China 17 July 1992 GCHN 9.6 116.2 NYSE
fund
Japan equity fund 14 August 1992 JPNE 8.1 102.8 NYSE
First Israel fund 23 October 1992 FISR 5.0 53.7 NYSE
Total 230.7 2,657.8
Panel B: Emerging market funds
Mexico fund 3 June 1981 MEX 37.3 765.6 NYSE
Korea fund 22 August 1984 KOR 29.5 610.0 NYSE
Taiwan fund 16 December TWN 11.3 289.5 NYSE
1986
Malaysia fund 8 May 1987 MYS 9.7 180.6 NYSE
Thai fund 17 February 1988 THA 12.2 343.8 NYSE
Brazil fund 31 March 1988 BRA 12.1 376.4 NYSE
India growth 12 August 1988 INDG 5.0 146.7 NYSE
fund
(continued)
25 Market Segmentation and Pricing of Closed-End Country Funds 679

Table 25.1 (continued)

No of shares Net assets
Fund name IPO date Symbol (millions) ($ millions) Listing
ROC Taiwan 12 May 1989 RTWN 27.9 365.7 NYSE
fund
Chile fund 26 September CHL 7.0 367.0 NYSE
1989
Portugal fund 1 November PRT 5.3 75.9 NYSE
1989
First Philippine 8 November FPHI 9.0 212.0 NYSE
fund 1989
Turkish 5 December 1989 TUR 7.0 33.5 NYSE
investment fund
Indonesia fund 1 March 1990 INDO 4.6 42.3 NYSE
Jakarta growth 10 April 1990 JAKG 5.0 43.2 NYSE
fund
Thai capital fund 22 May 1990 THAC 6.2 124.8 NYSE
Mexico equity 14 Aug 1990 MEXE 8.6 103.0 NYSE
and income fund
Emerging 2 October 1990 EMEX 9.0 117.6 NYSE
Mexico fund
Argentina fund 10 October 1991 ARG 9.2 108.1 NYSE
Korea investment 14 February 1992 KORI 6.0 87.4 NYSE
fund
Brazilian equity 9 April 1992 BRAE 4.6 91.4 NYSE
fund
Total 226.5 4,484.5
This table presents the sample of closed-end country funds. The funds are classified as from
developed markets or emerging markets using the World Bank’s definition of emerging markets.
The China funds are classified as developed market funds because the majority of their invest-
ments are in Hong Kong. The net assets reported is as of 31 December 1994. The listing of the
funds is denoted as NYSE, New York Stock Exchange and AMEX, American Stock Exchange.
The Germany fund, Korea fund, and the Thai capital fund also trade on the Osaka Stock Exchange.
The future Germany fund is now called Central European equity fund

emerging market funds using the World Bank’s definition of an emerging market.
The three China funds are classified as developed market funds because the majority
of their assets are in Hong Kong. The weekly (Friday closing) prices, NAVs, and
corresponding dividends for each fund are obtained from Barrons and Bloomberg. 10
The returns are adjusted for stock splits and dividends. The 7-day Eurodollar deposit
rate, provided by the Federal Reserve, is used as the risk-free benchmark.

10
For some of the funds, such as the India growth fund, the prices and net asset values are as of
Wednesday closing. This may lead to nonsynchronous prices and NAVs. However, as Bodurtha
et al. (1995) and Hardouvelis et al. (1993) show, the effects of nonsynchronous trading are not
pervasive and do not affect the analysis.
680 D.K. Patro

Table 25.2 Closed-end country funds: descriptive statistics

Panel A: Developed market funds
Price returns (%) NAV returns (%) Premium (%)
Fund Mean Std Mean Std Mean Std Volatility ratio
AUS 0.092 3.051 0.106 2.438 10.182 5.172 1.565*
ITA 0.009 4.126 0.029 2.943 5.985 9.311 1.965*
GER 0.106 3.943 0.130 2.372 1.120 11.316 2.762*
GBR 0.140 3.363 0.148 2.285 11.873 5.728 2.165*
SHEL 0.330 4.482 0.259 2.017 3.497 4.989 4.938*
SPN 0.007 4.128 0.033 2.505 0.076 11.923 2.716*
AUT 0.005 3.874 0.048 2.334 9.860 7.709 2.754*
GERN 0.113 3.787 0.113 2.146 15.003 6.020 3.113*
GSPN 0.239 5.159 0.143 2.675 14.522 8.655 3.717*
GERF 0.194 3.288 0.179 2.392 14.246 5.942 1.889*
JPNO 0.171 5.004 0.014 3.474 6.522 10.494 2.074*
GERE 0.051 3.548 0.014 2.200 14.750 6.200 2.601*
IRL 0.261 3.220 0.189 2.226 15.473 5.816 2.091*
FRA 0.133 3.456 0.090 2.133 13.266 8.020 2.625*
SGP 0.261 4.238 0.134 2.599 1.136 10.606 2.658*
Average 0.138 2.284 0.092 1.541 8.310 4.775 1.481*
Panel B: Emerging market funds
Price returns (%) NAV returns (%) Premium (%)
Symbol Mean Std Mean Std Mean Std Volatility ratio
MEX 0.841 12.670 0.897 14.695 5.783 8.646 0.743
KOR 0.277 4.778 0.295 3.428 18.625 11.146 1.943*
TWN 0.177 5.636 0.123 3.562 8.477 12.806 2.503*
MYS 0.308 4.581 0.251 3.233 3.636 7.599 2.007*
THA 0.232 4.359 0.358 3.458 3.264 10.251 1.588*
BRA 0.786 7.048 0.831 6.607 1.464 8.481 1.137*
INDG 0.297 5.518 0.130 4.144 0.833 15.397 1.773*
RTWN 0.235 5.026 0.030 2.998 0.443 8.892 2.809*
CHL 0.354 5.629 0.315 4.454 8.589 8.363 1.597*
PRT 0.209 4.639 0.129 2.284 6.313 7.311 4.125*
FPHI 0.489 4.490 0.359 2.589 21.127 5.345 3.005*
TUR 0.112 6.078 0.130 7.164 13.052 16.077 0.720
INDO 0.199 5.387 0.000 2.638 15.190 10.105 4.168*
JAKG 0.249 4.987 0.045 2.308 5.221 7.947 4.666*
THAC 0.449 4.921 0.377 3.510 8.516 6.409 1.964*
Average 0.347 2.478 0.284 1.939 0.447 4.093 1.277*
This table presents the descriptive statistics of the sample of closed-end country funds. The sample
covers the period January 1991–August 1995 (244 weekly observations). All the returns are
weekly returns in US dollars. The premium on a fund is calculated as the logarithm of the price-
net asset value ratio. The “volatility ratio” is the ratio of variance of price returns over the variance of
NAV returns for that fund. An asterisk (*) denotes that the variance of price returns is significantly
higher than the variance of NAV returns, from an F-test at the 5 % level of significance
25 Market Segmentation and Pricing of Closed-End Country Funds 681

Table 25.2 reports the descriptive statistics for the sample of funds. The IFC
global indices are available weekly from December 1988. Therefore, for emerging
market funds which were launched before December 1988, the analysis begins in
December 1988. To enable comparison across funds, a common time frame of
January 1991–August 1995 is chosen. Since, the data for the first 6 months are not
used in the analysis, only 30 funds listed on or before July 1990 are considered.
The mean premium on the index of developed market funds is 8.31 %, whereas
the mean premium on the emerging market funds is 0.44 %. A t-test of the
difference of means indicates that the mean premium of the index of the emerging
market funds is significantly higher than the mean premiums on the index of
developed market funds at the 1 % level of significance.
Figure 25.1 plots the time series of the premiums for the index of developed and
emerging market funds. The figure clearly indicates that the premiums on emerging
market funds are usually positive and higher than the premiums on developed market
funds, highlighting the effects of market segmentation. Table 25.2 also presents the
ratio of the variance of the excess price returns over the variance of the excess NAV
returns. For 27 funds out of the 30 funds, the ratios are significantly higher than one.11
A set of predetermined instrumental variables similar to instruments used in
studies of predictability of equity returns for developed markets (Fama and French
1989; Ferson and Harvey 1993) and emerging markets (Bekaert 1995) are used in
the empirical analysis when testing conditional asset-pricing models. The instru-
ments are the dividend yield on the Standard and Poor’s 500 Index, calculated as the
last quarter’s dividend annualized and divided by the current market price (DIVY);
the spread between 90-day Eurodollar deposit rate and the yield on 90-day US
treasury bill (TED); and the premium on an index of equally weighted country
funds (FFD). Only three instruments are used in order to be parsimonious repre-
sentation of the conditional asset-pricing models.
TED is calculated using data from the Federal Reserve of Chicago and DIVY is
obtained from Barrons. FFD is constructed using the sample of 30 funds listed in
Table 25.2. Table 25.3 reports the sample characteristics of these instruments.
As Fama and French (1989) show, the dividend yields and interest rates track the
business cycle, and expected returns vary as a function of these instruments.
Table 25.4 reports the descriptive statistics for the returns on the market indices
and the correlations of returns on the market indices of various countries with the
returns on the US index. The developed market indices are from Morgan Stanley
Capital International and the emerging market indices are from the International
Finance Corporation. All the developed market indices have significantly higher
correlations with the US market index, compared to the emerging markets, which
implies that there are potential diversification benefits from investing in emerging
markets.

11
Assuming that the price returns and NAV returns are from two normal populations, the test
statistic which is the ratio of the variances has an F distribution (if the two variances are estimated
using the same sample size). The null hypothesis that the variance of price returns is greater than
the variance of NAV returns is tested using this statistic at the 5 % level of significance.
682 D.K. Patro

20

Developed market funds

15 Emerging market funds

10

5
Premium (%)

0
910104
910315
910517
910726
911004
911213
920221
920417
920626
920904
921113
930108
930319
930521
930730
931008
931217
940225
940506
940715
940923
941202
950210
950421
950630
−5

−10

−15

−20
Date

Fig. 25.1 Country fund premiums

Table 25.3 Summary statistics for the instrumental variables

Panel A: Means and standard deviations (%)
Mean Std. dev
TED 0.34 0.18
DIVY 2.91 0.23
FFD 3.87 3.93
Panel B: Correlations of the instrumental variables
TED DIVY FFD
TED 1.00
DIVY 0.50 1.00
FFD 0.55 0.25 1.00
The statistics is based on weekly data from January 1991–August 1995 (244 weekly observations).
The spread between 90-day Eurodollar deposits and 90-day US treasury yields (TED), the
dividend yield on the Standard and Poor’s 500 index (DIVY), and an equally weighted index of
the premiums for the sample of funds in Table 25.2 (FFD)
25 Market Segmentation and Pricing of Closed-End Country Funds 683

Table 25.4 Market indices: descriptive statistics

Country Mean (%) Std. (%) Correlation with USA
Panel A: Developed markets
Austria 0.002 2.675 0.178
Hong Kong 0.485 3.366 0.174
Australia 0.227 2.212 0.292
Israel 0.000 3.351 0.130
France 0.167 2.272 0.334
Germany 0.180 2.416 0.238
Spain 0.090 2.627 0.329
Ireland 0.209 2.675 0.299
Italy 0.085 3.558 0.146
Japan 0.117 3.006 0.180
Singapore 0.336 2.018 0.231
Switzerland 0.375 2.186 0.285
USA 0.235 1.371 1.000
UK 0.143 2.186 0.335
Europe 0.169 1.808 0.377
Panel B: Emerging markets
Argentina 0.841 6.562 0.136
Brazil 1.145 7.577 0.221
Chile 0.687 3.180 0.125
India 0.259 4.534 0.021
Indonesia 0.114 3.302 0.031
Korea 0.130 3.432 0.040
Malaysia 0.410 2.802 0.163
Mexico 0.351 4.303 0.195
Philippines 0.636 3.618 0.141
Portugal 0.147 2.560 0.155
Taiwan 0.132 4.483 0.153
Thailand 0.520 3.609 0.189
Turkey 0.121 7.660 0.061
Asia 0.211 2.137 0.188
Latin America 0.554 3.314 0.640
World 0.172 1.410 0.277
This table presents the descriptive statistics for the returns on market indices of developed and
emerging markets for the time period January 1991–August 1995. The table also presents the
correlation of the returns, with the returns on the US market. All the developed market indices are
weekly indices from Morgan Stanley Capital International. The emerging market indices are from
the International Finance Corporation
684 D.K. Patro

25.4 Empirical Results: Pricing of Country Funds

This section presents the empirical results for pricing of country funds. The results
from the unconditional tests indicate that, in general, developed market closed-end
fund returns have significant risk exposures to the world market index, while
emerging market closed-end fund returns have significant risk exposures to both
the world market index and the corresponding local market index. Second, the
hypothesis of unconditional mean-variance efficiency of the world market
index cannot be rejected using the share price returns of either developed or
emerging market funds and NAV returns of some of the developed market funds.
However, the hypothesis of unconditional mean variance of the world market index
can be rejected for emerging market NAVs and some developed market NAVs.
This finding indicates that while the share prices reflect the global price of risk, the
NAVs may reflect both the global and the respective local prices of risk. Tests of
predictability using global instruments indicate that country fund share price and
NAVs exhibit significant predictable variation. When conditional asset-pricing
restrictions are examined using alternate stochastic discount factors, the results
indicate that the share prices and NAVs of the closed-end country funds are not
priced identically for some developed and all the Asian market funds. This finding
is consistent with market segmentation. Finally, it is shown that the time-varying
premiums for country funds are attributable to time-varying risk premiums.
The detailed results are discussed below.

25.4.1 Unconditional Risk Exposures of Country Funds

To ascertain the risk exposures of the share price and NAV returns of the sample of
country funds, the following econometric specification is employed:

rp, t ¼ ap þ bp, w rw, t þ bp, h rh, t þ ep, t (25.16)

rn, t ¼ an þ bn, w rw, t þ bn, h rh, t þ en, t (25.17)

where, for any country fund “i” (the subscript “i” has been dropped for
convenience):
rp,t ¼ the excess total return on the share price (including dividends) of a country
fund between t1 and t.
rn,t ¼ the excess return on the NAV of a country fund between t1 and t.
rw,t ¼ the excess return on the Morgan Stanley Capital International (MSCI) world
market index.
rh,t ¼ the excess return on the MSCI or International Finance Corporation (IFC)
global index, corresponding to the country of origin of the fund.
The coefficients bp,w, bp,h are the risk exposures on the world and home market
portfolios for the price returns, and bn,w,bn,h are the risk exposures on the world and
25 Market Segmentation and Pricing of Closed-End Country Funds 685

home market portfolios for the NAV returns. Two hypotheses are tested. The
hypotheses are share price returns have significant risk exposures to the world market
index and the NAV returns have significant risk exposures to both the world market
index and the corresponding local market index. These hypotheses are implied by the
international CAPM and its extension under market segmentation (see Errunza and
Losq 1985 and Diwan et al. 1995). If assets are priced in a global environment,
the country fund share price as well as NAV returns should have significant
risk exposure to the world market index only. However, when there are barriers
to investment, the local market factor may be a source of systematic risk for
the NAVs.
The risk exposures on the prices and NAVs could be different when the prices
and NAVs are imperfect substitutes. By jointly estimating Eqs. 25.16 and 25.17,
the hypothesis that the risk exposures on the world market and the home market
indices are identical for the price and NAV returns is tested via a Wald
test which is w2 distributed, with degrees of freedom equal to the number of
restrictions.12 Since the local indices are significantly correlated with the world
index, for ease of interpretation, they are made orthogonal to each other by
regressing the local index returns on a constant and the world index return and
using the residuals as the local index return. Therefore, the risk exposure on
the local or regional index is the marginal exposure in the presence of the world
market index.
The results from the regressions to estimate risk exposures are reported in
Table 25.5. Panel A of Table 25.5 presents the results for developed market funds
and panel B presents the results for the emerging market funds. The results
presented in Table 25.5 indicate that 15 of the 15 developed market funds and
12 of the 15 emerging market funds price returns have significant exposure to the
world index at the 5 % level of significance. Also, 12 of the developed market funds
and 14 of the emerging market funds price returns have significant exposure to their
corresponding local market index, in the presence of the orthogonal world
market index. For the NAVs, all 15 of the developed funds and ten of the emerging
market funds returns have significant exposure to the world market index. More-
over, 14 of the developed and 13 of the emerging market NAV returns have
significant exposure to their local market index, in the presence of the world market
index. The adjusted R-squares across all the regressions vary from a low of zero to
a high of 92 %.13

12
For a set of linear restrictions, the Wald test statistic is given by W ¼ ½Rb-r0 ½R VarðbÞ R0 ½Rb-r,
where b is the vector of estimated parameters and Rb ¼ r is the set of linear restrictions.
13
Risk exposures of the price and NAV returns were also estimated using regional indices in place
of the local market indices for the funds from Europe, Latin America, and Asia. These results
indicate that, out of the 12 European funds, only two funds’ price returns have significant risk
exposure to the MSCI Europe index in the presence of the MSCI world index. Also, out of the
seven Latin American funds, all seven price returns and six NAV returns have significant exposure
to the Latin American index. Also, out of the 11 Asian funds, nine price returns and eight NAV
returns have significant risk exposures to the Asian index.
Table 25.5 Risk exposures of country funds
686

Panel A: Developed market funds

Coefficients on price returns (t-stats) Coefficients on NAV returns (t-stats) Chi-squares for Wald tests [p-values]
Fund ap bp,h bp,w Adj. R2 an bn,h bn,w Adj. R2 w2local w2world
AUS 0.00 0.47 0.71 0.22 0.00 0.92 0.39 0.67 6.61 24.00
(0.66) (5.03)* (5.99)* (1.26) (26.79)* (8.05)* [0.01] [0.00]
ITA 0.00 0.44 0.60 0.17 0.00 0.76 0.70 0.87 0.21 11.36
(0.73) (5.18)* (3.15)* (3.06)* (33.69)* (14.30)* [0.64] [0.00]
GER 0.00 0.49 1.05 0.19 0.00 0.91 0.95 0.89 0.28 11.68
(0.15) (3.89)* (5.38)* (0.95) (35.56)* (30.87)* [0.59] [0.00]
GBR 0.00 0.32 0.87 0.17 0.00 0.92 0.74 0.73 0.64 18.77
(0.11) (2.78)* (6.33)* (1.42) (21.17)* (11.44)* [0.42] [0.00]
SHEL 0.00 0.08 0.63 0.03 0.00 0.11 0.32 0.05 1.09 0.03
(1.07) (0.62) (2.24)* (1.41) (1.11) (3.07)* [0.29] [0.84]
SPN 0.00 0.34 0.89 0.11 0.00 0.80 0.82 0.68 0.06 19.63
(0.80) (3.85)* (3.27)* (1.79) (20.66)* (21.50)* [0.80] [0.00]
AUT 0.00 0.30 0.97 0.14 0.00 0.72 0.77 0.71 1.21 19.06
(0.79) (3.46)* (6.57)* (1.37) (17.57)* (9.59)* [0.26] [0.00]
GERN 0.00 0.50 1.10 0.23 0.00 0.79 0.85 0.87 1.73 4.70
(0.37) (4.12)* (6.27)* (1.02) (27.79)* (25.90)* [0.18] [0.03]
GSPN 0.00 0.35 0.77 0.06 0.00 0.75 0.89 0.59 0.15 22.75
(0.44) (3.85)* (3.22)* (0.11) (16.15)* (10.63)* [0.69] [0.00]
GERF 0.00 0.45 0.91 0.23 0.00 0.93 0.89 0.92 0.01 23.94
(0.29) (4.24)* (6.00)* (0.90) (39.49)* (30.39)* [0.88] [0.00]
JPNO 0.00 0.23 1.95 0.31 0.00 0.74 1.16 0.41 6.86 10.25
(0.31) (1.82) (6.35)* (0.71) (7.71)* (8.18)* [0.00] [0.00]
GERE 0.00 0.51 0.80 0.18 0.00 0.84 0.76 0.85 0.04 12.03
(0.90) (5.87)* (4.34)* (3.46)* (34.82)* (23.53)* [0.83] [0.00]
D.K. Patro
 25

IRL 0.00 0.08 0.81 0.13 0.00 0.75 0.65 0.78 1.04 48.12
(0.75) (0.95) (5.73)* (0.33) (21.43)* (12.18)* [0.30] [0.00]
FRA 0.00 0.56 0.79 0.21 0.00 0.84 0.68 0.79 0.63 6.13
(0.20) (5.46)* (6.02)* (1.47) (26.34)* (17.95)* [0.42] [0.01]
SGP 0.00 0.63 0.93 0.15 0.00 0.65 0.43 0.25 11.27 0.01
(0.26) (4.18)* (5.65)* (0.14) (6.35)* (4.61)* [0.00] [0.90]
Panel B: Emerging market funds
Chi-squares for Wald tests
Coefficients on price returns (t-stats) Coefficients on NAV returns (t-stats) [p-values]
Symbol ap bp,h bp,w Adj. R2 an bn,h bn,w Adj. R2 w2local w2world
MEX 0.00 0.42 0.84 0.02 0.00 0.39 0.96 0.01 0.07 0.16
(1.36) (2.86)* (2.60)* (1.12) (2.01)* (2.35)* [0.78] [0.68]
KOR 0.00 0.20 0.09 0.01 0.00 0.00 0.01 0.00 3.09 0.14
(0.71) (1.77) (0.58) (0.73) (0.05) (0.09) [0.07] [0.70]
TWN 0.00 0.49 0.95 0.19 0.00 0.38 0.30 0.22 1.49 8.57
(0.26) (4.85)* (4.34)* (0.33) (6.91)* (2.26)* [0.22] [0.00]
MYS 0.00 0.62 1.01 0.20 0.00 0.92 0.74 0.63 8.35 0.83
(0.60) (5.98)* (3.74)* (0.89) (23.92)* (7.15)* [0.00] [0.36]
THA 0.00 0.56 1.02 0.27 0.00 0.87 0.72 0.82 18.04 3.07
(0.26) (7.80)* (5.71)* (2.27)* (23.26)* (13.28)* [0.00] [0.07]
Market Segmentation and Pricing of Closed-End Country Funds

BRA 0.00 0.50 0.90 0.31 0.00 0.64 1.07 0.58 7.38 0.29
(1.89) (10.84)* (2.98)* (2.41)* (17.65)* (5.64)* [0.00] [0.59]
INDG 0.00 0.34 0.06 0.07 0.00 0.35 0.06 0.14 0.00 0.20
(0.77) (2.15)* (0.27) (0.26) (5.04)* (0.45)* [0.98] [0.64]
RTWN 0.00 0.56 1.01 0.30 0.00 0.53 0.42 0.62 0.12 10.87
(0.10) (6.47)* (5.87)* (1.58) (12.03)* (6.17)* [0.71] [0.00]
(continued)
687
 688

Table 25.5 (continued)

Panel B: Emerging market funds
Chi-squares for Wald tests
Coefficients on price returns (t-stats) Coefficients on NAV returns (t-stats) [p-values]
Symbol ap bp,h bp,w Adj. R2 an bn,h bn,w Adj. R2 w2local w2world
CHL 0.00 0.77 0.46 0.21 0.00 0.89 0.04 0.40 1.24 5.13
(0.90) (7.57)* (2.54)* (1.19) (25.34)* (0.71) [0.26] [0.02]
PRT 0.00 0.31 0.75 0.07 0.00 0.73 0.58 0.70 7.67 0.76
(0.18) (2.40)* (4.29)* (1.12) (13.25)* (9.52)* [0.00] [0.38]
FPHI 0.00 0.64 0.64 0.30 0.00 0.56 0.24 0.60 1.77 6.07
(0.68) (8.72)* (3.94)* (0.55) (9.56)* (2.49)* [0.18] [0.01]
TUR 0.00 0.03 0.34 0.00 0.00 0.02 0.08 0.00 0.00 0.92
(0.05) (0.60) (1.28) (0.05) (0.39) (0.24) [0.96] [0.33]
INDO 0.00 0.48 0.91 0.14 0.00 0.64 0.06 0.66 2.74 15.83
(0.10) (4.94)* (4.62)* (1.11) (12.85)* (0.92) [0.09] [0.00]
JAKG 0.00 0.54 1.07 0.23 0.00 0.57 0.04 0.68 0.06 28.64
(0.26) (4.35)* (6.26)* (0.83) (13.64)* (0.73) [0.79] [0.00]
THAC 0.00 0.62 1.26 0.29 0.00 0.89 0.73 0.84 7.80 4.78
(1.23) (7.12)* (5.10)* (1.83) (19.05)* (15.04)* [0.00] [0.02]
This table presents the results from GMM estimation of risk exposures of the price (rp, t) and the NAV (rn.t) returns of country fund share and NAV excess
returns on the MSCI world index (rw, t)and the excess return on the corresponding local market index (rh, t). For developed market funds, the time series covers
the time from January 1991–August 1995. The price and the NAV risk exposures are estimated simultaneously using an exactly identified system of equations.
The t-stats robust to heteroskedasticity and serial correlation (six Newey-West lags) are presented below the coefficients. An asterisk (*) denotes significance
at the 5 % level of significance. The last two columns present the Chi-squares from a Wald test for the equality of the coefficients on the world market and local
market indices for the price and NAV returns. The models estimated are of the form
ra, t ¼ aa þ ba, h rh, :t þ ba, w rw, t þ ea, t where a ¼ p and n
D.K. Patro
25 Market Segmentation and Pricing of Closed-End Country Funds 689

These results confirm the hypothesis that returns on prices and NAVs are
generated as a linear function of the returns on the world market and possibly the
corresponding local market index. The fact that 27 out of the 30 funds have
significant risk exposure to the world market index indicates that the share prices
may be determined via the global price of risk. The higher adjusted R-squares for the
NAVs indicate that the world market index and the corresponding local market index
have higher explanatory power for the time-series variations of the returns. The
intercepts in the univariate regressions are not significantly different from zero,
which suggests that the return on the world market index and an orthogonal return
on the local market index are mean-variance efficient for the price and NAV returns.
When Wald tests are performed to test the hypothesis that the risk exposures on
the prices and NAVs are identical, the results presented in the last two columns of
Table 25.5 indicate that for 21 (14 from developed) out of the 30 funds, the null
hypothesis of the same betas on the world index is rejected at the 10 % level of
significance. Similarly, for eight funds the null hypothesis of the same betas on the
local index is rejected at the 5 % level. This is a very important finding since it
indicates that the systematic risks of the shares and NAVs are different. The fact
that closed-end funds may have different risk exposures for the share and NAVs has
been documented for domestic funds by Gruber (1996). Also different risk expo-
sures for restricted and unrestricted securities have been documented by Bailey and
Jagtiani (1994) for Thai securities and by Hietala (1989) for Finnish securities.
These results indicate that country fund share prices and NAVs may have
differential risk exposures. The different risk exposures will result in different
expected returns for the share prices and NAVs, which is one of the sources of
the premiums. This section clearly shows that country fund share price and NAV
returns have significant risk exposure to the world market index. The different risk
exposures will result in different expected returns for the share prices and NAVs,
which is one of the sources of the premiums. The issue of whether the country funds
are priced in equilibrium via the international CAPM is the focus of next section.

25.4.2 Pricing of Country Funds in the Context of

International CAPM

The results of the mean-variance efficiency of the MSCI world index, using
Eq. 25.2, are presented in Table 25.6. The table reports the w2 and the p-values
from a Wald test for the hypothesis that a set of intercepts are jointly zero.14 This
hypothesis is the restriction implied by unconditional international CAPM. Failure
to reject this hypothesis would imply that the world market index is mean-variance
efficient and expected returns are proportional to the expected returns on the world

14
Other studies such as Cumby and Glen (1990) fail to reject the mean-variance efficiency of the
MSCI world market portfolio for national indices of developed markets. Also, Chang et al. (1995)
fail to reject the mean-variance efficiency of the MSCI world market index for a sample of
developed as well as emerging market indices.
690 D.K. Patro

market index. The hypothesis is tested using GMM for an exactly identified system
as shown in section IA. To ensure a longer time series, the tests are conducted using
funds listed before December 1990.
The null hypothesis that the intercepts are jointly zero cannot be rejected for the
share prices of the 15 developed market funds as well as the 15 emerging market
funds. But, the null hypothesis is rejected at the 5 % level for the NAVs of both
developed and emerging market funds. The tests are also conducted on subsets of
funds to check the robustness of the results. Since the classification of developed
and emerging markets is based on national output, it may not always capture
whether a country’s capital market is well developed. The subset consists of
funds with zero intercepts for the NAV returns, on an individual basis. For this
subset of the developed market funds consisting of 11 funds, though not reported
here, the mean-variance efficiency of the world index cannot be rejected for both
the share prices and NAVs. For emerging market funds NAVs, even for subsets of
funds, the MSCI world market index is not mean-variance efficient. These results
indicate that the world market index is an appropriate benchmark for the share
prices but not necessarily for the NAVs.
These results have important implications for capital market integration. Since
for the NAVs, the mean-variance efficiency of the world market index is rejected, it
implies that the share prices and NAVs are priced differently. This differential
pricing is sufficient to generate premiums on the share prices. Also, the fact that
mean-variance efficiency of the world market index cannot be rejected for share
prices of both developed and emerging markets is consistent with the theoretical
prediction of Diwan et al. (1995). Since the country fund share prices are priced
with complete access to that market, in equilibrium, their expected returns should
be proportional to their covariances with the world market portfolio. If the NAVs
are priced with incomplete access, the world market portfolio is not mean-variance
efficient with respect to those returns. The results in this section are consistent
with this notion. As the earlier section shows, the risk exposures of the share prices
and NAVs may differ. If the international CAPM is a valid model for the share
prices and NAVs, as outlined in section IB, this differential risk exposure can
explain cross-sectional variations in premiums. The next section analyzes the effect
of the differential risk exposures of the share prices and NAVs on the country fund
premiums.

25.4.3 Cross-Sectional Variations in Country Fund Premiums

This section presents the results for tests of the hypothesis that cross-sectional
variation in country fund premiums is positively related to the differences in the risk
exposures of the share prices and NAVs. The theoretical motivation for this was
presented in section IC. Equation 25.10 is estimated for each week during the
last 75 weeks of the sample and the parameter estimates are averaged. The results
indicate that cross-sectional variation in country fund premiums is explained by the
25 Market Segmentation and Pricing of Closed-End Country Funds 691

Table 25.6 Pricing of country funds in the context of international CAPM

Chi-squares for price returns Chi-squares for NAV returns
Test assets [p-value] [p-value]
Developed market country 15.71 [0.40] 44.48 [0.00]
funds (15 funds)
Emerging market country 10.64 [0.77] 25.47 [0.04]
funds (15 funds)
This table presents the Chi-squares for the null hypothesis that the intercepts are jointly zero for
a set of assets when their excess returns are regressed on the excess return on the MSCI world
index. The regression estimated is
ri,t ¼ ai + birw,t + ei,t, i ¼ 1.... N
The null hypothesis Ho: ai ¼ 0 for all i ¼ 1. . .N is tested by a Wald test using GMM estimates from
an exactly identified system. The p-values are based on standard errors robust to heteroskedasticity
and serial correlation (13 Newey-West lags). The sample covers January 1991–August 1995. The
test assets are excess price/NAV returns of country funds
The composition of the test assets are as follows – the developed market funds are AUT, AUS,
FRA, GER, GERE, GERF, GERN, GSPN, IRL, ITA, JPNO, SGP, SPN, SHEL, and
GBR. The emerging market funds are BRA, CHL, FPHI, INDG, INDO, JAKG, KOR, MYS,
MEX, PRT, RTWN, TWN, THAC, THA, and TUR

differential risk exposures of a closed-end country fund share price and NAV returns.
The estimation proceeds are as follows: after the betas are estimated using 75 weeks
of data prior to the last 75 weeks of the sample, the difference in the betas on the world
index is used as an explanatory variable in regressions with the premiums as the
dependent variable.
The results are reported in Table 25.7. The results indicate that, for the full
sample, the difference in risk exposure is significant and the adjusted R-square is
9.5 %. For the developed market funds, the results in panel B indicate that the
differences in the betas on the world index have significant explanatory power for
the cross-section of premiums for the developed market funds, with an adjusted
R-square of 18.7 %. For the emerging market funds, however, as reported in
panel C, the differences in risk exposures are not significant, and the adjusted
R-square is only 1.9 %. This result is not surprising, given the result of the previous
section that the world market index is not mean-variance efficient for the emerging
market NAVs. When a dummy variable which takes value of one for the developed
markets is added, it is highly negatively significant, indicating that the premiums
for developed market funds are significantly lower than the emerging market fund
premiums. The differential risk exposure is however significant only at the 10 %
level, when the dummy variable is added.
As an extended specification, measures of spanning, integration, and substitution
(based on prior research of Errunza et al. 1998) are used as additional explanatory
variables. When measures of spanning, substitution, and access are used as addi-
tional explanatory variables, the adjusted R-squares go up to 39.5 % for the full
sample. The measure of spanning has a positive sign as expected. The measure of
spanning is also significant across the developed market funds. However, contrary
692 D.K. Patro

Table 25.7 Cross-sectional variations in country fund premiums

CONST DIF SPN SUB ACC(d) ACC (CF) Adj. R2
Panel A: Full sample
0.05 0.08 0.095
(1.60) (2.85)*
0.15 0.06 63.77 0.03 0.02 0.395
(0.07) (2.00)* (2.04)* (1.55) (1.40)
0.17 0.07 69.48 0.03 0.00 0.383
(2.79)* (2.00)* (2.28)* (1.60) (0.02)
0.16 72.59 0.03 0.300
(2.73)* (2.26)* (1.73)
0.00 0.05 0.08 0.184
(0.18) (2.03)* (2.85)*
Panel B: Developed market funds
0.09 0.11 0.187
(2.51)* (4.33)*
0.29 0.12 207.76 0.05 0.551
(4.23)* (4.03)* (2.05)* (2.27)*
0.29 0.12 210.39 0.08 0.00 0.565
(4.28)* (3.81)* (2.06)* (2.12)* (0.00)
0.28 234.11 0.04 0.370
(4.07)* (2.12)* (2.14)*
Panel C: Emerging market funds
0.00 0.01 0.019
(0.25) (0.40)
0.10 0.02 53.72 0.02 0.395
(1.21) (0.60) (1.75)** (0.86)
0.01 0.02 49.56 0.01 0.03 0.383
(0.95) (0.51) (1.59) (0.95) (0.15)
0.09 50.91 0.01 0.301
(1.11) (1.64) (0.86)
Results from an OLS regression of premiums on country funds on DIF (the difference between the
betas on the world market index, for the price returns and NAV returns). The coefficient ACC
(CF) is a measure of access proxied by the total purchase of securities by US residents (divided
by the global purchase) from the country of origin of the fund. The coefficient ACC(d) is
a dummy variable that takes value one for developed markets and zero for emerging markets.
The coefficients below are the averages from the regression of the premiums on the independent
variables for each week for the last 75 weeks of the sample (Mar 94–Aug 95). The t-statistics are
presented in the parenthesis * and ** denote significance at the 5 % and 10 % levels of significance,
respectively. The spanning measure (SPN) is the conditional variance of the NAV return of a fund,
unspanned by the US market return and the fund’s share price returns, with specification as
follows:
rn,t ¼ ai + birUS,t + + birp,t + ei,t
where ei,t has a GARCH(1, 1) specification. The measure of substitution (SUB) is the ratio of
conditional volatilities of the share price and NAV returns of a fund not spanned by the US market.
The specification are
rn,t ¼ ai + birUS,t + en,t, rp,t ¼ ai + birUS,t + ep,t
The error terms en,t and ep,t have GARCH (1, 1) specifications as follows (ht is variance of the error
term):
ht ¼ y0 + y1e2i;t1 + y2ht1
The equation for the cross-sectional regressions is
Premi ¼ d0 + d1DIFi + d2SPNi + d3SUBi + d4ACCi + ei
25 Market Segmentation and Pricing of Closed-End Country Funds 693

to expectation, the measure of spanning is not significant across emerging market

funds. This could be due to very low variability of the variables within emerging
markets. The measure of substitutability, which is the ratio of conditional
variances of the price returns and NAV returns, is not significant. The measure of
substitutability persists to be insignificant for the subsets of emerging market funds.
The measure of access proxied by purchase of securities by US residents from that
market is not significant.
These results indicate that differential risk exposures and measures of access and
spanning are significant determinants of cross-sectional variation in closed-end
country fund premiums. Also, the measures of segmentation explain premiums
better across developed and emerging markets compared to only within emerging
markets. The phenomenon that differential risk exposures can explain cross-
sectional variations in premiums on unrestricted securities relative to securities
restricted to only local investors has been previously documented for Thai
equities by Bailey and Jagtiani (1994) and for Swiss equities by Stulz and
Wasserfallen (1995). Therefore, both market segmentation and other factors
which make the closed-end country fund shares and the underlying securities
imperfect substitutes – which is the source of the differences in the risk exposures
and the excess volatility of the share prices – account for the cross-sectional
variations in the premiums. The significance of the difference in risk exposures
indicates that the greater the difference in risk exposures across a sample of country
funds, the higher the premium. Second, the premium is higher for country funds
originating from countries whose capital markets are less integrated with the US
capital market.
Although an important finding, the differential risk exposures cannot explain
the time variation in premiums. The academic literature has documented that the
country fund premiums vary over time. Figure 25.1 illustrates the time variation of
country fund premiums. Section IIIB reported the results for pricing of country funds
in the context of the unconditional international CAPM which indicated that country
fund share prices are priced consistent with the international CAPM, whereas the
NAVs may or may not be priced via the international CAPM. Unconditional mean-
variance efficiency of a benchmark portfolio implies conditional mean-variance
efficiency, but not vice versa (Ferson 1995). If expected returns conditional on an
information set are also different for the share prices and the NAVs, it could explain
not only the premiums but also their variation over time, which is further explored in
section IIIF. To analyze the time variability of expected returns, the predictability of
the funds share price and NAV returns is examined in the next section.

25.4.4 Predictability of Closed-End Country Fund Returns

To assess the predictability of price and NAV returns, 4-week cumulative returns
(the sum of returns over a 4-week period beginning the next week) are used, since
most studies of predictability have used a monthly horizon. Predictability of the
returns would imply that expected returns vary over time due to rational variation in
694 D.K. Patro

risk premiums. To test the hypothesis of predictability, the returns for both
share prices and NAVs are regressed on a set of global and fund-specific
instruments, to ascertain the predictive power of these instruments. The global
instruments are TED and DIVY. The fund-specific instrument is the lagged
premium on an equally weighted index of all the funds in the sample (FFD).
The lagged premium was found to have predictive power for price returns by
Bodurtha et al. (1995). Also, Errunza et al. (1998) find that when time series of
premiums is regressed on a global fund premium, it is highly significant. However,
unlike Errunza et al. (1998) here the lagged premium is used. Significance of this
variable would indicate that investors use information about past premiums to form
expectations about future prices. The regression is of the form

X
t¼ tþ4
ra, t ¼ Z0 þ Z1 FFDt þ Z2 TEDt þ Z3 DIVYt þ ea, t (25.18)
t ¼ tþ1

where ra, t is the excess return on the price or NAV of an equally weighted portfolio
of developed or emerging market funds.
The results from the regressions are presented in Table 25.8. Wald tests
of no predictability, which test the hypothesis that a set of coefficients are jointly
zero, indicate that the fund factor predicts price returns for 4 funds only and it
also predicts 15 funds’ NAV returns. The predictive power of the lagged premiums
for price returns has been attributed to noise trading (see Bodurtha et al. 1995).
The fact that it predicts NAV returns may indicate that noise trading is not an
important determinant of expected returns of share prices of closed-end funds.
When global instruments are used, the Wald test of no predictability can be rejected
at the 5 % level for only five funds’ price returns and 15 funds’ NAV returns.
Overall, using all the instruments, the null hypothesis of no predictability can be
rejected for six share prices and 20 NAVs. The poor predictability of the share
prices is puzzling, since the share prices are determined in the USA and the
instruments used have been shown to predict returns in the USA.
In summary, closed-end country fund price and NAV returns exhibit
considerable predictability. The existence of predictable variation in returns is
interpreted as evidence of time-varying expected returns. The differences
between the predictability of prices and NAVs also imply that they are not priced
identically, which is further examined in the next section using time-varying
expected returns.

25.4.5 Conditional Expected Returns and Pricing of Country Funds

Table 25.9 reports the results of estimating the SDFs specified in Eqs. 25.11
and 25.12. The models in panels A and B correspond to the international
CAPM and a two-factor model, with the world market return and the regional
returns for Latin America or Asia as the factors. For both the models,
25 Market Segmentation and Pricing of Closed-End Country Funds 695

Table 25.8 Predictability of country fund returns

Panel A: Developed market funds
Chi-squares [p-values] Chi-squares [p-values]
Global All Global All
Fund Fund factor instruments instruments Fund factor instruments instruments
AUS 0.11 1.93 1.95 8.23 3.44 9.24
[0.73] [0.38] [0.58] [0.00] [0.17] [0.02]
ITA 0.71 1.30 1.36 5.86 7.27 9.19
[0.39] [0.51] [0.71] [0.01] [0.02] [0.02]
GER 4.91 2.51 8.67 5.18 1.98 8.04
[0.02] [0.28] [0.03] [0.02] [0.37] [0.04]
GBR 0.07 1.23 2.58 3.33 5.64 5.98
[0.78] [0.53] [0.45] [0.06] [0.05] [0.11]
SHEL 1.81 1.43 2.51 0.05 1.86 1.86
[0.17] [0.48] [0.47] [0.81] [0.39] [0.60]
SPN 4.96 6.19 7.06 10.01 16.95 18.88
[0.02] [0.04] [0.06] [0.00] [0.00] [0.00]
AUT 0.02 0.01 0.03 1.86 3.25 6.03
[0.86] [0.99] [0.99] [0.17] [0.19] [0.10]
GERN 1.67 0.10 2.62 3.33 2.13 6.26
[0.19] [0.95] [0.45] [0.06] [0.34] [0.09]
GSPN 1.36 2.53 2.61 4.58 11.22 11.29
[0.24] [0.28] [0.45] [0.03] [0.00] [0.01]
GERF 0.20 0.29 1.42 3.58 2.91 5.87
[0.64] [0.86] [0.69] [0.05] [0.23] [0.11]
JPNO 7.12 3.18 8.73 13.96 5.20 19.60
[0.00] [0.20] [0.03] [0.00] [0.07] [0.00]
GERE 0.14 0.03 0.44 4.69 1.32 8.12
[0.70] [0.98] [0.93] [0.03] [0.51] [0.04]
IRL 0.00 1.93 2.15 1.80 14.56 14.58
[0.95] [0.38] [0.54] [0.17] [0.00] [0.00]
FRA 3.56 1.52 5.02 0.00 0.23 0.25
[0.05] [0.46] [0.17] [0.94] [0.89] [0.96]
SGP 0.00 2.55 2.55 11.05 1.37 11.28
[0.96] [0.27] [0.46] [0.00] [0.50] [0.01]
Panel B: Emerging market funds
Chi-squares [p-values] Chi-squares [p-values]
Global All Global All
Symbol Fund factor instruments instruments Fund factor instruments instruments
MEX 0.85 5.65 5.82 0.53 6.62 7.24
[0.35] [0.05] [0.12] [0.46] [0.03] [0.06]
KOR 0.74 1.04 2.00 4.20 8.87 13.46
[0.38] [0.59] [0.57] [0.04] [0.01] [0.00]
TWN 0.06 3.15 3.67 19.77 16.71 25.93
[0.79] [0.20] [0.29] [0.00] [0.00] [0.00]
(continued)
696 D.K. Patro

Table 25.8 (continued)

Panel B: Emerging market funds
Chi-squares [p-values] Chi-squares [p-values]
Global All Global All
Symbol Fund factor instruments instruments Fund factor instruments instruments
MYS 0.20 1.89 2.83 1.91 2.21 3.47
[0.64] [0.38] [0.41] [0.16] [0.33] [0.32]
THA 0.03 1.67 1.85 4.07 4.92 7.24
[0.85] [0.43] [0.60] [0.04] [0.08] [0.06]
BRA 0.03 6.34 6.99 0.32 4.90 5.39
[0.84] [0.04] [0.07] [0.56] [0.08] [0.14]
INDG 0.47 2.03 2.03 7.50 3.79 24.21
[0.49] [0.36] [0.56] [0.00] [0.15] [0.00]
RTWN 1.72 15.45 15.58 19.35 15.42 24.72
[0.18] [0.00] [0.00] [0.00] [0.00] [0.00]
CHL 0.05 4.94 6.97 1.78 7.32 7.34
[0.80] [0.08] [0.07] [0.18] [0.02] [0.06]
PRT 1.17 0.45 2.41 6.56 0.01 9.51
[0.27] [0.79] [0.49] [0.01] [0.99] [0.02]
FPHI 0.00 3.30 3.43 1.03 2.91 5.61
[0.97] [0.19] [0.32] [0.31] [0.23] [0.13]
TUR 1.37 1.11 3.45 0.35 0.34 1.42
[0.24] [0.57] [0.32] [0.55] [0.83] [0.70]
INDO 2.06 1.09 2.26 2.70 7.81 7.90
[0.15] [0.57] [0.51] [0.09] [0.02] [0.04]
JAKG 0.28 0.87 1.35 1.75 19.36 19.93
[0.59] [0.64] [0.71] [0.18] [0.00] [0.00]
THAC 0.03 4.22 4.79 1.98 2.33 3.53
[0.84] [0.12] [0.18] [0.15] [0.31] [0.31]
Table presents the Chi-squares from a Wald test to ascertain importance of the fund factor
and global instruments in predicting the 4-week ahead cumulative returns of the prices and
NAVs. The fund factor FFD is the lagged premium on an equally weighted index of all country
funds in the sample. The global instruments include the lagged spread on 90-day Eurodollar
deposits and 90-day US treasury yields (TED) and the lagged dividend yield on the S&P 500 index
(DIVY). The estimates are obtained via GMM using an exactly identified set of moment condi-
tions. The table below reports Chi-squares for a Wald test that a set of one or more instruments is
zero. The p-values are robust to heteroskedasticity. The model estimated is as follows, where
tX
¼ tþ4
ra, t ¼ Z0 þ Z1 FFDt þ Z2 TEDt þ Z3 DIVYt þ ea, t
t ¼ tþ1

the conditional restrictions are tested. As noted earlier, if expected returns vary over
time due to rational variation in risk premiums, conditional expected returns should
be used. The conditional restrictions are tested by using a set of lagged instrumental
variables TED and DIVY as predictors of excess returns. The tests use sets of
assets – two sets of four developed market funds, a set of three Latin American
funds, and two sets of three Asian funds. To ensure an adequate time series,
25 Market Segmentation and Pricing of Closed-End Country Funds 697

the funds selected are the ones listed prior to December 1989. Also, using too many
assets would result in too many moment conditions.15
The funds selected are listed in Table 25.9. After estimating a model using
GMM, its goodness-of-fit is ascertained by the J-statistic. The J-statistic is a test of
the overidentifying restrictions of the model. The p-values given below the
J-statistic indicate that none of the models are rejected for any test assets, at
conventional significance levels. This finding is consistent with the idea that for
a group of assets, there may be a number of SDFs that satisfy the pricing relation.
Also, failure to reject the conditional international CAPM even for the NAVs
indicates that while the unconditional international CAPM is not a valid model for
emerging market funds and some developed market funds, the conditional interna-
tional CAPM may be a valid model. This finding is consistent with the finding of
Buckberg (1995) who fails to reject the conditional international CAPM for a set of
emerging market indices. It must be noted that, unlike traditional conditional models,
using stochastic discount factors implies a nonlinear relation between asset returns
and the returns on the market index. Failure to reject such specifications may indicate
that nonlinear specifications perform better in explaining equity returns.
Table 25.9 also reports the results for tests of hypothesis that the coefficients of
the SDFs of the price returns and the NAV returns are identical. The p-values for the
Wald tests indicate that the null hypothesis can be rejected at all conventional levels
for a subset of the developed market funds and all the Asian funds. This is a striking
result, since it indicates that, although the same factors may be priced for the share
prices and NAVs, the factors are priced differently. In traditional asset-pricing
framework, it is equivalent to saying that the risk premiums are different.
This result is consistent with the previous finding of Harvey (1991) who reports
that for a sample of industrial markets, although the conditional international
CAPM cannot be rejected, the price of risk varies across countries. Also, De Santis
(1995) finds that stochastic discount factors that can price developed market indices
cannot price emerging market indices. The difference in the price of risk is one of the
sources of the premiums for the country funds. When the two-factor model is used,
again the tests reject the hypothesis that the coefficients are identical for the Asian
funds. However, the hypothesis of identical coefficients for the price and NAV returns
cannot be rejected for a set of developed market funds and the Latin American funds.
The above results imply that a subset of the developed markets and the Asian
market are segmented from the US market. However, a subset of the developed
markets and the Latin American markets are integrated with the US market. The
results for Latin America are not affected, when tests are conducted excluding the
time period surrounding the Mexican currency crisis. Therefore, differential risk
exposures are more important in explaining the premiums for the Latin American
markets and the developed markets that are integrated with the world market.
The results in this section clearly show that while some markets may be integrated

15
Cochrane (1996) shows that iterated GMM estimates behave badly when there are too many
moment conditions (37 in his case).
 698

Table 25.9 GMM estimation of stochastic discount factors

Price returns NAV returns
Test assets l0 l1 l2 l0 l1 l2 DF J [p-value] J1 [p-value] J2 [p-value]
Developed International CAPM
market funds
AUT, GBR, 2.14 (4.02)* 1.14 2.19 (5.51)* 1.29 20 11.98 [0.91] 0.53 [0.76] 0.07 [0.79]
GER, ITA (2.15)* (3.11)*
SPN, SHEL, 2.81 (4.46)* 1.81 0.81 (1.64) 0.18 (0.37) 20 11.23 [0.94] 8.38 [0.01] 7.72 [0.00]
AUS, AUT (2.88)*
Asian funds
KOR, MYS, 4.27 (4.08)* 3.27 2.95 (3.56)* 1.95 14 8.11 [0.88] 5.25 [0.07] 6.38 [0.01]
TWN (3.13)* (2.36)*
THA, INDG, 0.99 1.99 (1.50) 3.42 4.42 (2.84)* 14 9.28 [0.81] 11.21 [0.00] 10.57 [0.00]
FPHI (0.74) (2.20)*
Latin American funds
BRA, CHL, 4.46 (2.45)* 3.46 3.82 (2.37)* 2.82 14 9.80 [0.77] 0.42 [0.80] 0.36 [0.54]
MEX (1.90)* (1.75)
D.K. Patro
 25

Twofactor model
Asian funds
KOR, MYS, 2.35 3.54 (4.06)* 1.16 3.10 4.10 (4.80)* 1.19 18 11.51 [0.87] 3.12 [0.26] 2.94 [0.08]
TWN (2.92)* (1.48) (3.63)* (1.74)
THA, INDG, 3.00 (5.93)* 2.00 1.35 1.88 (3.86)* 0.88 1.50 18 10.66 [0.90] 8.16 [0.04] 7.65 [0.00]
FPHI (3.95)* (1.63) (1.81) (1.84)
Latin American funds
BRA, CHL, 0.70 (0.57) 0.29 (0.24) 1.09 0.83 (0.98) 0.16 (0.19) 0.76 18 12.64 [0.81] 2.27 [0.51] 0.02 [0.88]
MEX (3.27)* (3.97)*
Estimation of stochastic discount factors, based on the equation E[M(t,1) R(t,1)jZt] ¼1 via GMM. The sample covers January 1990–August 1995.
The instrument set Zt includes a constant, the lagged spread on 90-day Eurodollar deposits and 90-day US Treasury yields (LTED), and the dividend yield
on Standard and Poor’s 500 index (LDIVY). The estimations with the returns on regional indices (Rh) also use the lagged regional return (for Asia or Latin
America) as an instrument
The stochastic discount factors estimated are
ICAPM: M ¼ l0 + l1Rw
Two-factor: M ¼ l0 + l1Rw + l2Rh
DF is the degrees of freedom for Hansen’s test of the overidentifying restrictions (the J-statistic), which has an w2 distribution. The t-stats for the coefficients
are given in parenthesis (an asterisk indicates significance at the 5 % level) and the p-values for the J-stat is given in brackets. The last two columns report
Chi-squares and p-values for tests of hypotheses that the coefficients of the price returns and NAV returns are identical. J1 is for a joint test for all the
coefficients and J2 is a test for the coefficients on the world market index
Market Segmentation and Pricing of Closed-End Country Funds
699
700 D.K. Patro

with the world capital market, others are not. If capital markets are segmented, that
is sufficient to generate the premiums on the share prices. However, if markets are
integrated, differential risk exposures may explain the premiums. This finding is
different from the existing literature on country fund pricing, which has attributed
the existence of premiums on irrational factors such as noise trading. The preceding
analysis clearly shows that differential pricing of the same factors or differential
risk exposures on the same factor may lead to premiums on the share prices.
The next section provides an analysis of the effect of time-varying expected returns
on country fund premiums.

25.4.6 Conditional Expected Returns and Time-Varying Country

Fund Premiums

Table 25.9 reports the results of regressing the premiums on the country funds on
the differences in conditional risk exposures of share price returns and NAV returns
estimated using Eqs. 25.13, 25.14, and 25.15. The conditional risk exposures are
estimated as a linear function of lagged dividend yields and the term premium. The
null hypothesis is that the coefficient of the difference in the conditional betas is
significant. The results indicate that the difference in conditional betas is highly
significant in explaining the time-varying premiums on country funds. For 24 out of
the 30 funds, the coefficient is significant. Also, the adjusted R-squares are high,
especially for the developed market funds. For many of the funds, however, the
coefficient is negative, implying that using the world market index alone is not
sufficient to explain the return-generating process and the returns on the NAVs may
reflect the local price of risk.
The adjusted R2 values are higher for the developed market funds. This result is
consistent with earlier results which show that the world market index is not an
appropriate benchmark for emerging market NAVs. Also, majority of the emerging
market funds are from Asia. The results in Table 25.9 indicated that these markets
are segmented from the world market. Therefore, different risk exposures to the
world market index do not explain much of the variation in the premiums for the
emerging markets.
This is an important finding since the existing literature on closed-end funds
has attributed the existence of premiums to irrational factors, such as noise trading.
The preceding analysis shows clearly that segmented capital markets in which risk
premiums vary over time are sufficient to generate two different prices for the same
set of cash flows. Also, the differences in prices may vary over time as expected
returns vary over time due to rational variations in expected returns. If the price of
risk is different across markets, the same security will have different expected
returns. The difference in these expected returns results in the premium on the share
price. If the expected returns vary over time because of rational variation in risk
premiums, the premiums will also vary over time as a function of the differential
expected returns (Table 25.10).
25 Market Segmentation and Pricing of Closed-End Country Funds 701

25.5 Conclusions

Closed-end country funds are becoming an increasingly attractive source of

capital in international equity markets. This paper provides an empirical analysis
of the pricing of these funds in the context of rational asset-pricing models. The
paper finds that differential risk exposures, market segmentation, and time-varying
risk premiums play important roles in the differential pricing of the share prices and
NAVs. Based on the unconditional international CAPM, the existence of premiums
can be attributed to different risk exposures, as is the case with developed market
funds and differential pricing of the shares and NAVs, as is the case with emerging
market funds and some developed market funds.
The paper also analyzes the pricing of country funds when conditioning infor-
mation is allowed. The results indicate that, for alternate stochastic discount factors,
for majority of the funds, the pricing of country fund shares and NAVs is consistent
with the conditional international CAPM. However, tests of the estimated stochas-
tic discount factors indicate that, for a subset of the developed market funds and
all the Asian funds, closed-end country fund share prices and NAVs are priced
differently. This result indicates that the international capital markets are not fully
integrated. Finally, this paper shows that the premiums on country funds vary over
time because of time variation in expected returns.
The findings in this paper have several possible extensions. The focus of the
paper has been on rational explanations for country fund premiums based on
differential risk exposures and market segmentation effects. It will be interesting
to extend this analysis and examine the effect of other factors such as taxes,
numeraires, liquidity, and bid-ask spreads.

Appendix 1: Generalized Method of Moments (GMM)

GMM is an econometric method that was a generalization of the method of

moments developed by Hansen (1982). The moment conditions are derived
from the model. Suppose Yt is a multivariate independently and identically distrib-
uted (i.i.d) random variable. The econometric model specifies the relationship
between Zt and the true parameters of the model (y0). To use GMM there must
exist a function g(Zt, y0) so that

mðy0 Þ  E½gðZt ; y0 Þ ¼ 0 (25.19)

In GMM, the theoretical expectations are replaced by sample analogs:
X
f ðy; Zt Þ ¼ 1=T gðZt ; yÞ: (25.20)

The law of large numbers ensures that the RHS of above equation is the same as
E½f ðZt ; y0 Þ: (25.21)
702 D.K. Patro

Table 25.10 Time-varying closed-end country fund premiums

Panel A: Developed market funds
Fund g0 g1 Adj. R2
AUS 0.07 (10.13)* 0.11 (5.04)* 0.08
ITA 0.07 (13.05)* 0.10 (5.86)* 0.19
GER 0.01 (1.55)* 0.00 (0.29) 0.00
GBR 0.11 (34.40)* 0.08 (8.53)* 0.22
SHEL 0.06 (13.38)* 0.10 (7.12)* 0.13
SPN 0.00 (0.14) 0.18 (4.89)* 0.19
AUT 0.03 (3.47)* 0.21 (7.47)* 0.13
GERN 0.14 (37.73)* 0.04 (3.45)* 0.13
GSPN 0.15 (35.72)* 0.13 (7.13)* 0.22
GERF 0.14 (40.05)* 0.03 (2.47)* 0.01
JPNO 0.09 (9.44)* 0.04 (3.79)* 0.05
GERE 0.15 (39.20)* 0.04 (4.54)* 0.06
IRL 0.16 (44.37)* 0.13 (5.72)* 0.12
FRA 0.15 (24.47)* 0.24 (5.22)* 0.21
SGP 0.37 (2.03)* 0.72 (1.96) 0.02
Panel B: Emerging market funds
Fund g0 g1 Adj. R2
MEX 0.07 (16.93)* 0.08 (6.66)* 0.24
KOR 0.21 (26.45)* 0.15 (8.42)* 0.20
TWN 0.07 (1.01) 0.26 (2.28)* 0.01
MYS 0.03 (6.67)* 0.02 (2.25)* 0.02
THA 0.00 (0.04) 0.08 (0.43) 0.00
BRA 0.01 (2.44)* 0.05 (4.81)* 0.09
INDG 0.00 (0.61) 0.01 (0.83) 0.00
RTWN 0.60 (0.76) 0.00 (1.46) 0.01
CHL 0.34 (6.76)* 0.62 (5.14)* 0.20
PRT 0.06 (13.60)* 0.03 (2.47)* 0.02
FPHI 0.20 (25.86)* 0.01 (0.83) 0.00
TUR 0.15 (11.34)* 0.10 (2.16)* 0.01
INDO 0.23 (7.85)* 0.10 (2.97)* 0.07
JAKG 0.38 (8.91)* 0.34 (7.76)* 0.27
THAC 0.22 (4.97)* 0.24 (3.11)* 0.03
Results from an OLS regression of premiums on country funds on the difference between
the time-varying betas on the world market index. The t-statistics robust to heteroskedasticity
are presented in the parenthesis. An asterisk (*) denotes significance at the 5 % level of signifi-
cance. The regression estimated is of the form
Premt+1 ¼ g0 + g1[bn,w(Zt)  bp,w(Zt)] + eprem,t+1
The time-varying betas are estimated using the conditional international CAPM for an index of
developed and emerging market X funds price and NAV returns, using the equations


ra, tþ1 ¼ aa þ ba, w rw, tþ1 þ ga, i rw, tþ1  zt þ ea, tþ1 a ¼ p and n
i
where rp,t+1 is the excess return on the share price and rn,t+1 is the excess NAV return and Prem t+1
is the premium for the time period January 1991–August 1995. rw,t+1 is the excess return on the
world market index and Zt is a set of instrumental variables DIVY and TED
25 Market Segmentation and Pricing of Closed-End Country Funds 703

The sample GMM estimator of the parameters may be written as

(see Hansen 1982)
h X i0 X i
Y ¼ arg min 1=T gðZt ; yÞ WT 1=T gðZt ; yÞ (25.22)

So essentially GMM finds the values of the parameters so that the

sample moment conditions are satisfied as closely as possible. In our case for the
regression model,

yt ¼ Xt 0 b þ et (25.23)

The moment conditions include

E½ðyt  Xt 0 bÞxt  ¼ E½et xt  ¼ 0 for all t (25.24)

So the sample moment condition is

X
1=T ðyt  Xt 0 bÞxt

and we want to select b so that this is as close to zero as possible. If we select b as

(X0 X)1(X0 y), which is the OLS estimator, the moment condition is exactly
satisfied. Thus, the GMM estimator reduces to the OLS estimator and this is
what we estimate. For our case the instruments used are the same as the
independent variables. If, however, there are more moment conditions than the
parameters, the GMM estimator above weighs them. These are discussed in detail
in Greene (2008, Chap. 15). The GMM estimator has the asymptotic variance
 1
1
X0 Z ðZ0 OZÞ Z0 X (25.25)

The White robust covariance matrix may be used for O as discussed in appendix
C when heteroskedasticity is present. Using this approach, we estimate GMM with
White heteroskedasticity consistent t-stats.

References
Bailey, W., & Jagtiani, J. (1994). Foreign ownership restrictions and stock prices in the Thai
Capital Market. Journal of Financial Economics, 36, 57–87.
Bailey, W., & Lim, J. (1992). Evaluating the diversification benefits of the new country funds.
Journal of Portfolio Management, 18, 74–80.
Bansal, R., Hsieh, D. A., & Viswanathan, S. (1993). A new approach to international arbitrage
pricing. Journal of Finance, 48, 1719–1747.
Bekaert, G. (1995). Market integration and investment barriers in equity markets. World Bank
Economic Review, 9, 75–107.
704 D.K. Patro

Bekaert, G., & Urias, M. (1996). Diversification, integration and emerging market closed-end
funds. Journal of Finance, 51, 835–869.
Bodurtha, J. N., Jr., Kim, D.-S., & Lee, C. M. C. (1995). Closed-end country funds and U.S. market
sentiment. Review of Financial Studies, 8, 881–919.
Bonser-Neal, C., Brauer, G., Neal, R., & Wheatly, S. (1990). International investment restrictions
and closed-end country fund prices. Journal of Finance, 45, 523–547.
Chang, E., Eun, C., & Kolodny, R. (1995). International diversification through closed-end
country funds. Journal of Banking and Finance, 19, 1237–1263.
Cochrane, J. H. (1996). A cross-sectional test of an investment-based asset pricing model. Journal
of Political Economy, 104(3), 572–621.
Cumby, R. E., & Glen, J. D. (1990). Evaluating the performance of international mutual funds.
Journal of Finance, 45(2), 497–521.
De Santis, G. (1995). Volatility bounds for stochastic discount factors: Tests and implications from
international financial markets. Working paper, University of Southern California, Jan.
Diwan, I., Vihang, E., & Senbet, L.W. (1995). Diversification benefits of country funds. In
M. Howell (Ed.), Investing in emerging markets. Washington, DC: Euromoney/World Bank.
Errunza, V., & Losq, E. (1985). International asset pricing under mild segmentation: Theory and
test. Journal of Finance, 40, 105–124.
Errunza, V., Senbet, L. W., & Hogan, K. (1998). The pricing of country funds from emerging
markets: Theory and evidence. International Journal of Theoretical and Applied Finance,
1, 111–143.
Fama, E. F., & French, K. R. (1989). Business conditions and expected returns on stocks and
bonds. Journal of Financial Economics, 25, 23–49.
Ferson, W. (1995). Theory and empirical testing of asset pricing models. In R. Jarrow,
V. Maksimovic, & W. T. Ziemba (Eds.), Handbooks in operation research and management
science (Vol. 9). Amsterdam: North-Holland.
Ferson, W., & Foerster, S. (1994). Finite sample properties of generalized methods of moments in
tests of conditional asset pricing models. The Journal of Financial Economics, 36, 29–55
Ferson, W. E., & Harvey, C. R. (1993). The risk and predictability of international equity returns.
Review of Financial Studies, 6(3), 527–566.
Ferson, W. E., & Schadt, R. (1996). Measuring fund strategy and performance in changing
economic conditions. Journal of Finance, 51(2), 425–461.
Greene, W. (2008). Econometric analysis. Washington, DC: Pearson.
Gruber, M. J. (1996). Presidential address: Another puzzle: The growth in actively managed
mutual funds. Journal of Finance, 51, 783–810.
Hanley, K. W., Lee, C., & Seguin, P. (1996). The marketing of closed-end fund IPOs. Journal of
Financial Intermediation, 5, 127–159.
Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica, 50, 1029–1054.
Hardouvellis, G., Wizman, T., & La Porta, R. (1993). What moves the discount on country equity
funds? Working paper 4571, NBER.
Harvey, C. (1991). The world price of covariance risk. Journal of Finance, 46, 111–159.
Hietala, P. T. (1989). Asset pricing in partially segmented markets: Evidence from the finnish
market. Journal of Finance, 44(3), 697–718.
Lee, C. M. C., Shleifer, A., & Thaler, R. (1991). Investor sentiment and the closed-end fund
puzzle. Journal of Finance, 46(1), 75–109.
Mackinlay, C., & Richardson, M. (1991). Using generalized method of moments to test mean-
variance efficiency. Journal of Finance, 46(2), 511–527.
Peavy, J. W., III. (1990). Returns on initial public offerings of closed-end funds. Review of
Financial Studies, 3, 695–708.
Pontiff, J. (1997). Excess volatility and closed-end funds. American Economic Review,
87(1), 155–169.
25 Market Segmentation and Pricing of Closed-End Country Funds 705

Roll, R. (1977). A critique of the asset pricing theory’s tests. Journal of Financial Economics,
4, 129–176.
Solnik, B. (1996). International investments (3rd ed.). New York: Addison-Wesley.
Stulz, R. (1995). International portfolio choice and asset pricing: An integrative survey.
In R. Jarrow, V. Maksimovic, & W. T. Ziemba (Eds.), Handbooks in operation research and
management science (Vol. 9). Amsterdam: North-Holland.
Stulz, R., & Wasserfallen, W. (1995). Foreign equity investment restrictions, capital flight, and
shareholder wealth maximization: Theory and evidence. Review of Financial Studies,
8, 1019–1057.
Weiss, K. (1989). The post-offering price performance of closed-end funds. Financial Manage-
ment, 18, 57–67.
A Comparison of Portfolios Using
Different Risk Measurements 26
Jing Rung Yu, Yu Chuan Hsu, and Si Rou Lim

Contents
26.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 708
26.2 Portfolio Selection Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
26.2.1 The Mean-Variance Model (MV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
26.2.2 The Mean Absolute Deviation Model (MAD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
26.2.3 The Downside Risk Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
26.3 The Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
26.3.1 The MV Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
26.3.2 The MAD Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
26.3.3 The DSR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
26.3.4 The MV_S Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
26.3.5 The MAD_S Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714
26.3.6 The DSR_S Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
26.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716
26.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
The Linearization of the MAD Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
The Linearization of the DSR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
Appendix 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Multiple Objective Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727

Abstract
In order to find out which risk measurement is the best indicator of efficiency in
a portfolio, this study considers three different risk measurements: the mean-
variance model, the mean absolute deviation model, and the downside risk
model. Meanwhile short selling is also taken into account since it is an important

J.R. Yu (*) • Y.C. Hsu • S.R. Lim

National Chi Nan University, Nantou, Taiwan
e-mail: jennifer@ncnu.edu.tw; s96213020@ncnu.edu.tw; s96213021@ncnu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 707
DOI 10.1007/978-1-4614-7750-1_26,
# Springer Science+Business Media New York 2015
708 J.R. Yu et al.

strategy that can bring a portfolio much closer to the efficient frontier by
improving a portfolio’s risk-return trade-off. Therefore, six portfolio rebalancing
models, including the MV model, MAD model, and the downside risk model,
with/without short selling, are compared to determine which is the most
efficient. All models simultaneously consider the criteria of return and risk
measurement. Meanwhile, when short selling is allowed, models also consider
minimizing the proportion of short selling. Therefore, multiple objective
programming is employed to transform multiple objectives into a single objec-
tive in order to obtain a compromising solution. An example is used to perform
simulation, and the results indicate that the MAD model, incorporated with
a short selling model, has the highest market value and lowest risk.

Keywords
Portfolio selection • Risk measurement • Short selling • MV model • MAD
model • Downside risk model • Multiple objective programming • Rebalancing
model • Value-at-risk • Conditional value-at-risk

26.1 Introduction

Determining how to maximize the profit and minimize the risk of a portfolio is an
important issue in portfolio selection. The mean-variance (MV) model of portfolio
selection is based on the assumptions that investors are risk averse and the return of
assets is normally distributed (Markowitz 1952). This model is regarded as the basis
of modern portfolio theory (Deng et al. 2005).
However, the MV model is limited in that it only leads to optimal decisions if the
investor’s utility functions are quadratic or if investment returns are jointly elliptically
distributed (Grootveld and Hallerbach 1999; Papahristoulou and Dotzauer 2004). Thus,
numerous researches have focused on risk, return, and diversification in the development
of investment strategies. Also, a large number of researches have been proposed to
improve the performance of investment portfolios (Deng et al. 2000; Yu and Lee 2011).
Konno and Yamazaki (1991) proposed a linear mean absolute deviation (MAD)
portfolio optimization model. The MAD model replaces the variance of objective
function in the MV model with the mean absolute deviation. The major advantage
of the MAD model is that the estimation of the covariance matrix of asset returns is
not needed. Also, it is much easier to solve large-scale problems with linear
programming than with quadratic approaches (Simaan 1997).
Portfolio selection under shortfall constraints originated from Roy’s (1952)
safety-first theory. Economists have found that investors care about downside
losses more than they care about upside gains. Therefore, Markowitz (1959)
suggested using semi-variance as a measure of risk, instead of variance, because
semi-variance measures downside losses rather than upside gains. The use of
downside risk (DSR) measures is proposed due to the problems encountered in
using a conventional mean-variance analysis approach in the presence of
non-normality in the emerging market data. Unlike the mean-variance framework,
26 A Comparison of Portfolios Using Different Risk Measurements 709

the downside risk measure does not assume that the return distributions of assets are
normal. In addition, the increasing emphasis of investors on limiting losses might
make the downside risk measure more intuitively appealing (Stevenson 2001; Ang
et al. 2006). The downside risk measure can help investors make proper decisions
when returns are non-normally distributed, especially for emerging market data or
for an international portfolio selection (Vercher et al. 2007).
Another type of shortfall constraint is value-at-risk (VaR), which is a percentile-
based metric system for risk measurement purposes (Jorion 1996). It is defined as the
maximum loss that a portfolio can suffer at a given level of confidence and at a given
horizon (Fusai and Luciano 2001). However, VaR is too weak to handle the situation
when losses are not “normally” distributed, as loss distribution tends to exhibit “fat tail”
or empirical discreteness. Conditional value-at-risk (CVaR) is an alternative measure
that quantifies losses that might be encountered in the tail of loss distribution
(Rockafellar and Uryasev 2002; Topaloglou et al. 2002). A confidence level is required
when employing the measurements of VaR and CVaR. The allocation of a portfolio
varies with the varying confidence level. Therefore, neither measure is included for
comparison in this chapter.
Generally, short selling has good potential to improve a portfolio’s risk-return
trade-off (White 1990; Kwan 1997) and is considered by most investors to obtain
interest arbitrage; however, it comes with high risk (Angel et al. 2003). Since high
risks should be avoided, the role of short selling is minimized in this chapter. Instead,
three kinds of portfolio models with and without short selling are compared.
Section 26.2 introduces the mean-variance, mean absolute, and downside risk
models. In Sect. 26.3, the rebalancing models with/without short selling are pro-
posed. In Sect. 26.4, the performances of three different risk measurements with/
without short selling are compared by using the historical data of 45 stocks listed in
the TSE50 index. Finally, Sect. 26.5 presents the conclusions along with sugges-
tions for future research.

26.2 Portfolio Selection Models

In this section, the mean-variance, the mean absolute deviation, and the downside
risk models are introduced separately. First, the notations are defined as follows:
n is the number of available securities.
wi is the investment portion in securities i for i ¼ 1, . . ., n.
ri is the return on securities i.
m is the expected portfolio return.
si2 is the variance of the return on securities i.
sij2 is the covariance between the returns of securities i and j.
rit is the return on securities i in period t for t ¼ 1, . . ., T, which is assumed to be
available through historical data.
XT
Ri is equal to T1 r it .
t¼1
dt is the deviation between the return and the average return.
710 J.R. Yu et al.

26.2.1 The Mean-Variance Model (MV)

The MV model uses the variance of the return as the measure of risk and formulates
the portfolio optimization problem as the following quadratic programming prob-
lem (Markowitz 1952):

X
n Xn X
n
Min sp ¼ w2i s2i þ sij wi wj
i¼1 i¼1 j¼1ði6¼jÞ
(26.1)
X
n
s:t: r i wi  m,
i¼1

X
n
wi ¼ 1, (26.2)
i¼1

wi  0, (26.3)

for i ¼ 1, . . ., n.
Constraint (26.1) expresses the requirements m of a portfolio return, and con-
straint (26.2) is the budget constraint. The model is known to be valid if an investor
is risk averse in the sense that he prefers less standard deviation of the portfolio
rather than more. Since wi  0, a short sale is not allowed here.

26.2.2 The Mean Absolute Deviation Model (MAD)

The mean-variance model is weak in constructing a large-scale portfolio due to the

computational difficulty associated with solving a large-scale quadratic program-
ming problem with a dense covariance matrix. The MAD model (Konno and
Yamazaki 1991) replaces the variance in the objective function of the MV model
with the mean absolute deviation as follows:
 
T X 
1X 
n

Min  ðr it  Ri Þwi 
T t¼1  i¼1 

s.t. Constraints (26.1)  (26.3).

Because of the absolute deviation, the MAD model can be linearized as
following (Chang 2005):

1X T
Min dt
T t¼1

s.t. Constraints (26.1)  (26.3),

26 A Comparison of Portfolios Using Different Risk Measurements 711

X
n
dt þ ðr it  Ri Þwi  0, t ¼ 1, . . . , T, (26.4)
i¼1

X
n
dt  ðr it  Ri Þwi  0, t ¼ 1, . . . , T: (26.5)
i¼1

If the return is lower than the average return, constraint (26.4) is a binding
Xn
constraint which means d t ¼  ðr it  Ri Þwi for t ¼ 1, . . .,T. Otherwise, constraint
i¼1
X
n
(26.5) is a binding constraint which means dt ¼ ðr it  Ri Þwi for t ¼ 1, . . .,T. For
i¼1
more details on the reformulation, please refer to Appendix 1.
Apparently, the MAD model does not require the covariance matrix of asset returns,
and consequently its estimation is not needed. Large-scale problems can be solved faster
and more efficiently because the MAD model has a linear rather than quadratic nature.

26.2.3 The Downside Risk Model

Vercher et al. (2007) consider the equivalent formulation of the portfolio selection
problem (Speranza 1993) and reformulate the following linear optimization model
by considering downside risk measurement:

1X T
Min d
T t¼1 t

s.t. Constraints (26.1)  (26.3),

X
n
d
t þ ðr it  Ri Þwi  0, t ¼ 1, . . . , T (26.6)
i¼1

where dt ¼ dt+ + dt, dt+, dt  0.

Downside risk measurement focuses on returns falling below some critical level
(Grootveld and Hallerbach 1999). Differing from the MAD model, the downside
risk model ignores constraint (26.5). If the return is lower than the average return,
constraint (26.6) is a binding constraint. Please refer to Appendix 2 for more details.

26.3 The Proposed Model

Multiple period portfolio selection models with rebalancing mechanisms have become
attractive in the financial field in order to get desired returns in situations that are
subject to future changes (Yu et al. 2010). To reflect a changing situation in the models,
the rebalancing mechanism is adopted for multiple periods (Yu and Lee 2011).
712 J.R. Yu et al.

Six rebalancing models are introduced. These are MV, MAD, DSR, MV_S,
MAD_S, and DSR_S. The first three models lack short selling, and the other three
have short selling. The model notations are denoted as follows: w+i,0 is the weight of
security i held in the previous period, w i,0 is the weight of securities i sold short in
the previous period, wþ i is the total weight of securities i bought after rebalancing,
and w i is the total weight of securities i sold short after rebalancing.
With each rebalancing, lþ 
i is the weight of securities i bought in this period, li is
þ
the weight of securities i sold in this period, si is the weight of securities i sold short
in this period, si is the weight of securities i repurchased in this period, ui is the
binary variable that indicates whether the securities i are selected for buying, vi is
the binary variable that indicates whether the securities i are selected for selling
short, and k is the initial margin requirement for short selling.

26.3.1 The MV Model

The conventional MV model can be regarded as a bi-objective model without short

selling, whose objective functions are the maximization of portfolio return and
minimization of portfolio risk, as measured by the portfolio variance:
X
n
Max Ri wi
i¼1

Min sp

s.t. Constraint (26.2),

þ 
wi ¼ wþ
i, 0 þ ‘i  ‘i , (26.7)

0:05ui  wi  0:2ui , (26.8)

for i ¼ 1, 2, . . ., n.
Constraint (26.7) is the rebalancing constraint; it shows the current weight for
the ith security according to the previous period. Constraint (26.8) is the required
range of weights for each security in buying. For simplexity, the upper and lower
bounds of each weight are set 0.2 and 0.05, respectively.

26.3.2 The MAD Model

The objectives of the MAD model include the maximization of return and minimi-
zation of the mean absolute deviation, which is transformed into a linear deviation
as follows:
Xn
Max Ri w i
i¼1
26 A Comparison of Portfolios Using Different Risk Measurements 713

1X T
Min dt
T t¼1

s.t. Constraints (26.2)  (26.5), (26.7), (26.8).

26.3.3 The DSR Model

The DSR model considers the objectives of maximizing the return and minimizing
the downside risk, which is transformed into a linear risk as follows:

X
n
Max Ri w i
i¼1

1X T
Min d
T t¼1 t

s.t. Constraints (26.2), (26.3), (26.6), (26.7), (26.8).

When short selling is allowed, the above three models are reformulated as
follows (Yu and Lee 2011):

26.3.4 The MV_S Model

X
n  
Max Ri w þ 
i  wi
i¼1

Min sp

X
n
Min w
i
i¼1
X
n  (26.9)

s:t: wþ 
i þ kwi ¼ 1,
i¼1

þ 
wþ þ
i ¼ wi, 0 þ ‘i  ‘i , (26.10)

w  þ 
i ¼ wi, 0 þ si  si , (26.11)

0:05ui  wþ
i  0:2ui , (26.12)
714 J.R. Yu et al.

0:05vi  w
i  0:2vi , (26.13)

ui þ v i ¼ y i , (26.14)

for i ¼ 1, 2, . . ., n.
Unlike the MV model, the MV_S model has an extra objective, namely, mini-
mizing the short selling proportion. The total budget is including the cost of buying
and short selling as shown in constraint (26.9). k is the initial margin requirement
for short selling. Constraint (26.10) indicates the current weight for the ith securities
based on the previous period. Constraint (26.11) indicates the current short selling
proportion of the ith security adjusted by the previous period. Constraints (26.12)
and (26.13) limit the upper and lower bounds, respectively, of the long and short
selling proportion for the ith security.

26.3.5 The MAD_S Model

Based on the MAD model (Konno and Yamazaki 1991), the MAD_S model
replaces the objective function of minimizing the variance in the MV_S model
(Yu and Lee 2011) with the objective of minimizing the absolute deviation of
average return, as follows:

X
n     
Min  wþ  w r it  wþ  w Ri 
i i i i
i¼1

The objective of the mean absolute deviation can be transformed into a linear
problem:

1X T
Min dt
T t¼1

X
n    þ 
s:t: dt þ wþ  
i  wi r it  wi  wi Ri  0, (26.15)
i¼1

X
n    þ 
dt  wþ  
i  wi r it  wi  wi Ri  0, (26.16)
i¼1

for t ¼ 1, . . ., T.
The following is the MAD_S model:

X
n  
Max Ri wþ 
i  wi
i¼1
26 A Comparison of Portfolios Using Different Risk Measurements 715

1X T
Min dt
T t¼1

X
n
Min w
i
i¼1

s.t. Constraints (26.9)  (26.16).

26.3.6 The DSR_S Model

The DSR_S model focuses on the deviation when the return falls below the average
return, as follows:

X
n  
Max Ri w þ 
i  wi
i¼1

1X T
Min d
T t¼1 t

X
n
Min w
i
i¼1

s.t. Constraints (26.9)  (26.14),

X
n    þ 
d
t þ wþ  
i  wi r it  wi  wi Ri  0, (26.17)
i¼1

for t ¼ 1, . . ., T.
Apparently, all six models have multiple objectives. Therefore, multiple
objective programming (Zimmermann 1978, and Lee and Li 1993) is adopted to
transform the multiple objectives into a single objective. For more details, please
refer to Appendix 3. Taking the MAD_S model as an example, we can reformulate
the multiple objectives as follows:

Max l
ðr   r l Þ (26.18)
s:t: l   ,
rg  rl

ð s   sl Þ
l , (26.19)
s g  sl
716 J.R. Yu et al.

  
w  w
l l
, (26.20)
w 
g  wl

X
n  
r ¼ Ri w þ 
i  wi , (26.21)
i¼1

1X T
s ¼ d , (26.22)
T t¼1 t
X
n
w ¼ w
i , (26.23)
i¼1
Constraints (26.9)  (26.16).
For multiple objective programming (Lee and Li 1993), r is the return of the
portfolio, rl is the anti-ideal return of the portfolio, rg is the ideal return of the
portfolio that maximizes the objective, s is the inherent risk of the portfolio, sl is
the anti-ideal risk of the portfolio, sg is the ideal risk of the portfolio, w is the
short selling proportion of the portfolio, wl is the anti-ideal short selling proportion
of the portfolio, and w g is the ideal short selling proportion of the portfolio.
The constraints (26.18–26.20) are the achievements for maximizing the return,
minimizing the absolute deviation and objectives minimizing the short selling
problem of the corresponding portfolio, which are less than or equal to the whole
achievement (l). The whole achievement (l) should be maximized.
In the same way, the other five multiple objective models can be reformulated in
turn as a single-objective model. The details of the transformation are introduced in
Appendix 3.

26.4 Experimental Results

Forty-five stocks listed on the Taiwan Stock Exchange were adopted and used to
compare the six models discussed above in order to determine which one is the best.
The benchmark is the Taiwan 50 Index (TSE50).
The exchange codes of the 45 stocks are listed in Table 26.1. The duration of the
analyzed data is from November 1, 2006, to November 24, 2009. The historical data
of the first 60 transaction days are used to build the initial models. For the monthly
updates, 20 transaction days are set as a sliding window. This study assumes
a budget of $1 million NTD is invested in the Taiwan Stock Market. For invest-
ments based on the weights generated by the initial models, the first transaction day
is January 25, 2007, and there are 34 rebalancing times in total. The models are
executed on an Intel Pentium Dual CPU E2200 2.20GHz and 2G RAM computer,
with Lingo11.0, an optimizing software.
From Table 26.2, it is apparent that the MAD and DSR models are more efficient
than the MV model because they both use linear transformation for problem
solving. This is much faster and more efficient when handling a large-scale
 26

Table 26.1 The exchange codes of 45 stocks

No 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Code 1101 1102 1216 1301 1303 1326 1402 1722 2002 2105 2308 2311 2317 2324 2325
No 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Code 2330 2347 2353 2354 2357 2382 2409 2454 2498 2603 2801 2880 2881 2882 2883
No 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
Code 2885 2886 2888 2890 2891 2892 2912 3009 3231 3474 3481 5854 6505 8046 9904
A Comparison of Portfolios Using Different Risk Measurements
717
718 J.R. Yu et al.

Table 26.2 The running time of six portfolios

Models without short selling MV MAD DSR
00:09:06 00:01:01 00:01:01
Models with short selling MV_S MAD_S DSR_S
00:10:01 00:01:15 00:01:14
(hh:mm:ss)

1.9

MV 14,000
1.7 MAD
DSR

1.5 TSE50 12,000

TAIEX
Market Value (million)

1.3
10,000

1.1
8,000

0.9

6,000
0.7

0.5 4,000
1/24/07

3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08

7/24/08

9/24/08

11/24/08

1/24/09

3/24/09

5/24/09

7/24/09

9/24/09

Fig. 26.1 The market value of three risk measurements without short selling, the TSE50, and
TAIEX

problem. As one can see in Table 26.2, the MV model takes 9 min and 6 sec to
compute the result. However, the MAD and DSR models only take 1 min or slightly
more to solve the same data. Moreover, it is not necessary to calculate the covari-
ance matrix to set up the MAD and DSR models; this makes it very easy to update
the models when new data are added (Konno and Yamazaki 1991).
Figures 26.1, 26.2 and 26.3 show the comparisons of the MV, MAD, and DSR
models, respectively. These are the models without short selling. Figures 26.4, 26.5
and 26.6 show the comparisons of the MV_S, MAD_S, and DSR_S models,
respectively. These are the models with short selling. The TSE50 is used as the
benchmark to these models. As shown in Fig. 26.1, the market value of the MAD
26 A Comparison of Portfolios Using Different Risk Measurements 719

0.016

MV
0.014
MAD
DSR
0.012

0.01
Expected Return

0.008

0.006

0.004

0.002

−1E-17
1/24/07
3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08
7/24/08

9/24/08

11/24/08

1/24/09
3/24/09

5/24/09

7/24/09

9/24/09
−0.002

Fig. 26.2 The expected return of three risk measurements without short selling

0.035

MV
MAD
0.03 DSR

0.025
Risk

0.02

0.015

0.01
1/24/07

3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08

7/24/08

9/24/08

11/24/08

1/24/09

3/24/09

5/24/09

7/24/09

9/24/09

Fig. 26.3 The risk of three risk measurements without short selling
720 J.R. Yu et al.

2.3 18,000
MV_S
2.1 MAD_S
16,000
DSR_S
1.9 TSE50
TAIEX 14,000
Market Value (million)

1.7

1.5 12,000

1.3 10,000

1.1
8,000
0.9

6,000
0.7

0.5 4,000
1/24/07

3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08

7/24/08

9/24/08

11/24/08

1/24/09

3/24/09

5/24/09

7/24/09

9/24/09
Fig. 26.4 The market value of three risk measurements with short selling, the TSE50, and TAIEX

model is always greater than that of the other models and the benchmark.
Figures 26.2 and 26.3 display the expected return and risk of the portfolios
constructed with three risk measurements without short selling. Figure 26.2
shows that the expected returns of these three portfolios are almost the same.
However, Fig. 26.3 shows that the MAD model has the lowest risk under a similar
expected return to that of the others. Especially, in January 24, 2009, the risk in the
MAD model was much lower than in the other two models.
Downside risk measurement is applied when investors only take the negative
return into consideration and focus on the loss of investment. In other words, their
concern is with the real loss, not with the positive deviation of the average return in
portfolios. Therefore, in Fig. 26.3, the risk generated by the DSR model is always
higher than other measurements.
In Fig. 26.4, the market value of the MV_S, MAD_S, and DSR_S models is
compared. Since these models take short selling into consideration, the portfolio
selection is more flexible, and the risk is much lower, as Fig. 26.6 shows. Under the
same expected return, the MAD_S model has the lowest risk among the three models,
using the mean absolute deviation risk measure. Figure 26.4 shows that it also has the
highest market value. Even though the market value of each model is increased after
short selling is allowed, the market value of the MAD_S model is always higher than
the other models and the benchmark. Evidently, the MAD model is suggested for use
as the best risk measurement tool with or without short sell.
26 A Comparison of Portfolios Using Different Risk Measurements 721

0.016
MV_S
0.014 MAD_S
DSR_S

0.012

0.01
Expected Return

0.008

0.006

0.004

0.002

0
1/24/07

3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08

7/24/08

9/24/08

11/24/08

1/24/09

3/24/09

5/24/09

7/24/09

9/24/09
−0.002

Fig. 26.5 The expected return of three risk measurements with short selling

0.035

MV_S
MAD_S
0.03 DSR_S

0.025
Risk

0.02

0.015

0.01
1/24/07

3/24/07

5/24/07

7/24/07

9/24/07

11/24/07

1/24/08

3/24/08

5/24/08

7/24/08

9/24/08

11/24/08

1/24/09

3/24/09

5/24/09

7/24/09

9/24/09

Fig. 26.6 The risk of three risk measurements with short selling
722 J.R. Yu et al.

26.5 Conclusion

This chapter compares six different rebalancing models, with/without short selling, in
order to determine which is more flexible for portfolio selection. One of the advan-
tages of the MAD and DSR models is that they can be linearized; thus, they are faster
and more efficient than the MV model, especially with large-scale problems.
The experimental results indicate that the MAD and MAD_S models are
efficient in handling data and show higher market value than the other models;
moreover, they have lower risks in situations with/without short selling.
However, there remain important risk measurements, such as VaR and CVaR, for
future research to investigate. Thus, future studies may focus on developing
rebalancing models with the measures of VaR and CVaR. Since the rebalancing period
is fixed, the dynamic rebalancing mechanism is required for the first change environ-
ment. In addition, a portfolio selection model able to predict future returns is required.

Appendix 1

The Linearization of the MAD Model

Konno and Yamazaki (1991) assume that rit is the realization of random variable ri
X
n
during period t (t ¼ 1, . . ., T), and Ri ¼ T1 r it :
i¼1
The MAD model is as follows:
 
T X 
1X 
n

Max  ðr it  Ri Þwi 

T t¼1 i¼1 
X
n
s:t: Ri xi  m,
i¼1
X
n
wi ¼ 1,
i¼1

wi  0,
for i ¼ 1, . . ., n.
 
X n 
 
Let  ðr it  Ri Þwi  ¼ dt ¼ dt þ þ d t  ,
 i¼1 

dt þ and dt   0,

X
n
then ðr it  Ri Þwi ¼ dt þ  d t  ,
i¼1
26 A Comparison of Portfolios Using Different Risk Measurements 723

X
n
dt  d
t ¼ ðr it  Ri Þwi þ d t  ,
i¼1

!
1 Xn

dt ¼ dt  ðr it  Ri Þwi  0:
2 i¼1

X
n
2dt  ¼ dt  ðr it  Ri Þwi :
i¼1

Similarity, perform the same process,

X
n
dt þ ¼ ðr it  Ri Þwi þ dt  ,
i¼1

X
n
dt  dt þ ¼  ðr it  Ri Þwi þ dt þ ,
i¼1

Xn
2dt þ ¼  ðr it  Ri Þwi þ d t ,
i¼1

!
þ 1 Xn
dt ¼ ðr it  Ri Þwi þ d t  0,
2 i¼1

dt þ , d
t  0:

Two constraints are added because dt+, dt  0.

Then the model can be transformed into a linear model as follows:

 
T X  1X
1X 
n

T
Min  ðr it  Ri Þwi  ¼ d þ þ dt 
T t¼1  i¼1  T t¼1 t

!
þ 1 Xn
dt ¼ ðr it  Ri Þwi þ dt  0,
2 i¼1

X
n
) ðr it  Ri Þwi þ dt  0:
i¼1
724 J.R. Yu et al.

!
1 X n

dt ¼  ðr it  Ri Þwi þ d t  0,
2 i¼1

X
n
) ðr it  Ri Þwi þ dt  0:
i¼1

where dt ¼ dt+ + dt.

Therefore, the MAD model can be linearized as the following linear model:

1X T
Min dt
T t¼1
Xn
s:t: d t þ ðr it  Ri Þwi  0, t ¼ 1, . . . , T,
i¼1

X
n
dt  ðr it  Ri Þwi  0, t ¼ 1, . . . , T,
i¼1

X
n
Ri xi  m,
i¼1

X
n
wi ¼ 1,
i¼1

wi  0,

for i ¼ 1, . . . , n.

Appendix 2

The Linearization of the DSR Model

The use of variance as a measure of risk makes no distinction between gains and
losses. The following mean semi-absolute deviation risk measurement proposed by
Speranza (1993) is used to find the portfolios with minimum semi-variance:

 
XT Xn  X n
 
1
 ðr it  Ri Þwi   ðr  Ri Þwi
T 
t¼1 i¼1
 i¼1 it
Min
2
26 A Comparison of Portfolios Using Different Risk Measurements 725

X
n
s:t: Ri xi  m,
i¼1

X
n
wi ¼ 1,
i¼1

wi  0,

for i ¼ 1, . . . n.  
X n 
 
Because of the absolute deviation,  ðr it  Ri Þwi  of the DSR model can be
 i¼1 
linearized in the same manner as the MAD model:
X
n
ðr it  Ri Þwi ¼ dt þ  dt 
i¼1
 
X n  X n
 
 ðr it  Ri Þwi   ðr  Ri Þwi
 i¼1  i¼1 it dt þ þ dt   dt þ þ dt 
¼ ¼ dt 
2 2
!
1 Xn
dt þ ¼ ðr it  Ri Þwi þ d t  0,
2 i¼1

d t   0:

Then the DSR model is reformulated as follows:

1X T
Min d
T t¼1 t

X
n
s:t: Ri xi  m,
i¼1

X
n
wi ¼ 1,
i¼1

wi  0,

X
n
d
t þ ðr it  Ri Þwi  0, t ¼ 1, . . . , T,
i¼1

dt   0, t ¼ 1, . . . , T,

for i ¼ 1, . . ., n.
726 J.R. Yu et al.

Appendix 3

Multiple Objective Programming

The aforementioned multiple objective models in Sect. 26.3 are solved by fuzzy
multiple objective programming (Zimmermann 1978; Lee and Li 1993) in order to
transform the multiple objective model into a single-objective model. Fuzzy mul-
tiple objective programming based on the concept of fuzzy set uses a min operator
to calculate the membership function value of the aspiration level, l, for all of the
objectives.
The following is a multiple objective programming problem (Lee and Li 1993):

Max Z ¼ ½Z 1 ; Z 2 ; . . . ; Z l T ¼ ½c1 x, c2 x, . . . , cl xT

Min W ¼ ½W 1 ; W 2 ; . . . ; W l T ¼ ½q1 x, q2 x, . . . , ql xT

s:t: Ax b,

x  0,

where Ck, k ¼ 1, 2, . . .,l, cs, s ¼ 1, 2, . . .,r, and x are n-dimensional vectors; b is an

m-dimensional vector; A is an m  n matrix; and * denotes the operators , ¼,
or  The program aimed to achieve its maximization of the achievement level for
each objective while also considering a trade-off among the conflicting objectives
or criteria. The ideal and anti-ideal solutions must be obtained in advance. This
ideal solution and anti- ideal solutions are given by the decision maker, respec-
tively, as follows:

 
I þ ¼ Z 1 ; Z 2 ; . . . ; Z l ; W 1 ; W 2 ; . . . ; W l ,
 
I ¼ Z     
1 ; Z2 ; . . . ; Zr ; W 1 ; W 2 ; . . . ; W r :

The membership (achievement) functions for the objectives are defined as

follows:

Z K ðxÞ  Z 
m k ðZ K Þ ¼ k
, k ¼ 1, 2, . . . , l,
Z k  Z 
k

W
s  W s ðxÞ
ms ðZ s Þ ¼  , s ¼ 1, 2, . . . , r:
Ws  Ws

Then the “min” operator is used; the multiple objective programming is

formulated as follows:
26 A Comparison of Portfolios Using Different Risk Measurements 727

Max l
   
s:t: l  ðZK ðxÞ  Z  
k = Z k  Z k , k ¼ 1, 2, . . . , l,

   
l  W 
s  W s ðxÞÞ= W s  W s , s ¼ 1, 2, . . . , r,

x 2 X,

where l is defined as l ¼ mini mðxÞ ¼ mink, s ðmk ðZÞ, ms ðW s ÞÞ.

References
Ang, A., Chen, J., & Xing, Y. (2006). Downside risk. The Review of Financial Studies, 19,
1191–1239.
Angel, J. J., Christophe, S. E., & Ferri, M. G. (2003). A close look at short selling on Nasdaq.
Financial Analysts Journal, 59, 66–74.
Chang, C. T. (2005). A modified goal programming approach for the mean-absolute deviation
portfolio optimization model. Applied Mathematics and Computation, 171, 567–572.
Deng, X. T., Wang, S. Y., & Xia, Y. S. (2000). Criteria models and strategies in portfolio selection.
Advanced Modeling and Optimization, 2, 79–104.
Deng, X. T., Li, Z. F., & Wang, S. Y. (2005). A minimax portfolio selection strategy with
equilibrium. European Journal of Operational Research, 166, 278–292.
Fusai, G., & Luciano, E. (2001). Dynamic value at risk under optimal and suboptimal portfolio
policies. European Journal of Operational Research, 135, 249–269.
Grootveld, H., & Hallerbach, W. (1999). Variance vs downside risk: Is there really that much
difference? European Journal of Operational Research, 114, 304–319.
Jorion, P. H. (1996). Value at risk: A new benchmark for measuring derivative risk. New York:
Irwin Professional Publishers.
Konno, H., & Yamazaki, H. (1991). Mean-absolute deviation portfolio optimization model and its
applications to Tokyo stock market. Management Science, 37, 519–531.
Kwan, C. C. Y. (1997). Portfolio selection under institutional procedures for short selling:
Normative and market equilibrium considerations. Journal of Banking & Finance, 21,
369–391.
Lee, E. S., & Li, R. J. (1993). Fuzzy multiple objective programming and compromise program-
ming with Pareto optimum. Fuzzy Sets and Systems, 53, 275–288.
Markowitz, H. M. (1952). Portfolio selection. The Journal of Finance, 7, 77–91.
Markowitz, H. M. (1959). Portfolio selection, efficient diversification of investments. New York:
Wiley.
Papahristoulou, C., & Dotzauer, E. (2004). Optimal portfolios using linear programming models.
Journal of the Operational Research Society, 55, 1169–1177.
Rockafellar, R. T., & Uryasev, S. (2002). Conditional value-at-risk for general loss distribution.
Journal of Banking and Finance, 26, 1443–1471.
Roy, A. D. (1952). Safety first and the holding of assets. Econometrica, 20, 431–449.
Simaan, Y. (1997). Estimation risk in portfolio selection: The mean variance model versus the
mean absolute deviation model. Management Science, 43, 1437–1446.
Speranza, M. G. (1993). Linear programming models for portfolio optimization. Finance, 14,
107–123.
Stevenson, S. (2001). Emerging markets, downside risk and the asset allocation decision. Emerg-
ing Markets Review, 2, 50–66.
728 J.R. Yu et al.

Topaloglou, N., Vladimirou, H., & Zenios, S. A. (2002). CVaR models with selective hedging for
international asset allocation. Journal of Banking and Finance, 26, 1535–1561.
Vercher, E., Bermudez, J. D., & Segura, J. V. (2007). Fuzzy portfolio optimization under downside
risk measures. Fuzzy Sets and Systems, 158, 769–782.
White, J. A. (1990). More institutional investors selling short: But tactic is part of a wider strategy.
Wall Street Journal, 7, 493–501.
Yu, J. R., & Lee, W. Y. (2011). Portfolio rebalancing model using multiple criteria. European
Journal of Operational Research, 209, 166–175.
Yu, M., Takahashi, S., Inoue, H., & Wang, S. (2010). Dynamic portfolio optimization with risk
control for absolute deviation model. Journal of Operational Research, 201, 349–364.
Zimmermann, H. J. (1978). Fuzzy programming and linear programming with several objective
functions. Fuzzy Sets and Systems, 1, 44–45.
Using Alternative Models and
a Combining Technique in Credit Rating 27
Forecasting: An Empirical Study

Cheng-Few Lee, Kehluh Wang, Yating Yang, and Chan-Chien Lien

Contents
27.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
27.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
27.2.1 Ordered Probit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
27.2.2 Ordered Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732
27.2.3 Combining Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732
27.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
27.3.1 Model Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
27.3.2 Credit Rating Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
27.3.3 Estimation Results Using the Combining Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
27.3.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741
27.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741
Appendix 1: Ordered Probit Procedure for Credit Rating Forecasting . . . . . . . . . . . . . . . . . . . . . . . . 744
Appendix 2: Ordered Logit Procedure for Credit Rating Forecasting . . . . . . . . . . . . . . . . . . . . . . . . 747
Appendix 3: Procedure for Combining Probability Forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 749

Abstract
Credit rating forecasting has long time been very important for bond classifica-
tion and loan analysis. In particular, under the Basel II environment, regulators

C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu
K. Wang • Y. Yang (*)
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: lkwang@mail.nctu.edu.tw; yatingyang.iof98g@nctu.edu.tw
C.-C. Lien
Treasury Division, E.SUN Commercial Bank, Taipei, Taiwan

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 729
DOI 10.1007/978-1-4614-7750-1_27,
# Springer Science+Business Media New York 2015
730 C.-F. Lee et al.

in Taiwan have requested the banks to estimate the default probability of the
loan based on its credit classification. A proper forecasting procedure for credit
rating of the loan is crucially important in abiding the rule.
Credit rating is an ordinal scale from which the credit category of a firm can
be ranked from high to low, but the scale of the difference between them is
unknown. To model the ordinal outcomes, this study first constitutes an attempt
utilizing the ordered logit and the ordered probit models, respectively. Then, we
use ordered logit combining method to weigh different techniques’ probability
measures as described in Kamstra and Kennedy (International Journal of
Forecasting 14, 83–93, 1998) to form the combining model.
The samples consist of firms in the TSE and the OTC market and are divided
into three industries for analysis. We consider financial variables, market vari-
ables, as well as macroeconomic variables and estimate their parameters for
out-of-sample tests. By means of cumulative accuracy profile, the receiver
operating characteristics, and McFadden R2, we measure the goodness-of-fit
and the accuracy of each prediction model. The performance evaluations are
conducted to compare the forecasting results, and we find that combining
technique does improve the predictive power.

Keywords
Bankruptcy prediction • Combining forecast • Credit rating • Credit risk • Credit
risk index • Forecasting models • Logit regression • Ordered logit • Ordered
probit • Probability density function

27.1 Introduction

This study explores the credit rating forecasting techniques for firms in Taiwan. We
employ the ordered logit and the ordered probit models for rating classification and
then a combining procedure to integrate both. We then examine empirically the
performance of these alternative methods, in particular, whether the combining
forecasting performs better than any individual method.
Credit rating forecasting has long time been very important for bond classifica-
tion and loan analysis. In particular, under the Basel II environment, regulators in
Taiwan have requested the banks to estimate the default probability of the loan
based on its credit classification. A proper forecasting procedure for credit rating of
the loan is crucially important in abiding the rule.
Different forecasting models and estimation procedures have various underlying
assumptions and computational complexities. They have been used extensively by
researchers in the literature. Review papers like Hand and Henley (1997), Altman
and Sounders (1997), and Crouhy et al. (2000) have traced the developments of the
credit classification and bankruptcy prediction models over the last two decades.
Since Beaver’s (1966) pioneered work, there have been considerable researches
on the subject of the credit risk. Many of them (Altman 1968; Pinches and Mingo
1973; Altman and Katz 1976; Altman et al. 1977; Pompe and Bilderbeek 2005)
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 731

use the multivariate discriminant analysis (MDA) which assumes normality for
the explanatory variables of the default class. Zmijewski (1984) utilizes the probit
model, and Ohlson (1980) applies the logit model in which discrete or continuous data
can be fitted.
Kaplan and Urwitz (1979), Ederington (1985), Lawrence and Arshadi (1995),
and Blume et al. (1998) show that it is a consistent structure considering credit
rating as ordinal scale instead of interval scale. That is, the different values of the
dependent variables as different classes represent an ordinal, but not necessarily
a linear scale. For instance, higher ratings are less risky than lower ratings, but we
don’t have a quantitative measure indicating how much less risky they are.
Kaplan and Urwitz (1979) conduct an extensive examination of alternative
prediction models including N-chotomous probit analysis which can explain the
ordinal nature of bond ratings. To test the prediction accuracy of various statistical
models, Ederington (1985) compares the linear regression, discriminant analysis,
ordered probit, and unordered logit under the same condition. He concludes that
the ordered probit can have the best prediction ability and the linear regression is
the worst.
In a survey paper on forecasting methods, Mahmoud (1984) concludes that
combining forecasts can improve accuracy. Granger (1989) summarizes the use-
fulness of combining forecasts. Clemen (1989) observes that combining forecasts
increase accuracy, whether the forecasts are subjective, statistical, econometric, or
by extrapolation. Kamstra and Kennedy (1998) integrate two approaches with logit-
based forecast-combining method which is applicable to dichotomous, polychoto-
mous, or ordered-polychotomous contexts.
In this paper, we apply the ordered logit and the ordered probit models in credit
rating classification for listed firms in Taiwan and then combine two rating models
with a logit regression technique. The performance of each model is then evaluated
and we find that combining technique does improve the predictive power.

27.2 Methodology

27.2.1 Ordered Probit Model

Credit rating is an ordinal scale from which the credit category of a firm can be
ranked from high to low, but the scale of the difference between them is unknown.
To model the ordinal outcomes, let the underlying response function be

Y
¼ Xb þ e (27.1)

where Y* is the latent variable, X is a set of explanatory variables, and e is the

residual. Y* is not observed, but from which we can classify the category j:

Yi ¼ j if tj1 < Y
i  tj ði ¼ 1, 2, . . . , n; j ¼ 1, 2, . . . , J Þ: (27.2)

732 C.-F. Lee et al.

Maximum likelihood estimation can be used to estimate the parameters given

a specific form of the residual distribution.
For the ordered probit model, e is normally distributed with mean 0 and variance
1. The probability density function is

 2
1 e
fðeÞ ¼ pffiffiffiffiffiffi exp  (27.3)
2p 2

and the cumulative density function is

ðe  2
1 t
F ðeÞ ¼ pffiffiffiffiffiffiexp  dt: (27.4)
1 2p 2

27.2.2 Ordered Logit Model

For the ordered logit model, e has a logistic distribution with mean 0 and variance
p2/3. The probability density function is

expðeÞ
lð e Þ ¼ (27.5)
½1 þ expðeÞ2

and the cumulative density function is

expðeÞ
LðeÞ ¼ : (27.6)
1 þ expðeÞ

27.2.3 Combining Method

To combine the ordered logit and the ordered probit models for credit forecasting,
the logit regression method as described in Kamstra and Kennedy (1998) is applied.
We first assume that firm’s credit classification is determined by an index y.
Suppose there are J rating classes, ordered from 1 to J. If y exceeds the threshold
value tj, j ¼ 1, . . ., j  1, credit classification changes from j rating to j + 1 rating.
The probability of company i being in rating j is given by the integral of a standard
logit from tj  1  yi to tj yi.
Each forecasting method is considered as producing J  1 measures
oji ¼ tj  yi, j ¼ 1, . . ., j  1 for each firm. These measures can be estimated as

 
P1i þ    þ Pji
oji ¼ ln (27.7)
1  P1i      Pji
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 733

Table 27.1 Sample numbers across the industry

Industry Non-bankruptcy observations Bankruptcy observations Total
Panel A: In-sample (2000.Q1  2004.Q4)
Traditional 3,993 509 4,502
Manufacturing 2,450 411 2,861
Electronics 8,854 191 9,045
Panel B: Out-of-sample (2005.Q1  2005.Q3)
Traditional 629 63 692
Manufacturing 432 38 470
Electronics 1,990 72 2,062

where Pji is a probability estimate for firm i in rating j. The combining method
proposed by Kamstra and Kennedy (1998) consists of finding, via MLE, an
appropriate weighted average of o’s in ordered logit and ordered probit
techniques.1
To validate the model, we use cumulative accuracy profile (CAP) and its
summary statistics, the accuracy ratio (AR). A concept similar to the CAP is the
receiver operating characteristic (ROC) and its summary statistics, the area under
the ROC curve (AUC). In addition, we also employ the McFadden’s R2 (pseudo R2)
to evaluate the performance of the credit rating model. McFadden’s R2 is defined as
1(unrestricted log-likelihood function/restricted log-likelihood function).

27.3 Empirical Results

Data are collected from the Taiwan Economic Journal (TEJ) database for the period
between the first quarter in 2000 and the third quarter in 2005, with the last three
quarters used for out-of-sample tests. The sample consists of firms traded in the
Taiwan Security Exchange (TSE) and the OTC market.
The credit rating of the sample firms is determined by the Taiwan Corporate
Credit Risk Index (the TCRI). Among ten credit ratings, 1–4 represent the invest-
ment grade levels, 5–6 represent the low-risk levels, and 7–9 represent the high-risk
or speculative grade levels. The final rating, 10, represents the bankruptcy level.
Table 27.1 exhibits the descriptive statistics for the samples which are divided
into three industry categories. Panel A contains the in-sample observations, while
panel B shows the out-of-sample observations. There are 509 bankruptcy cases in
the traditional industry, 411 in the manufacturing sector, and 191 in the electronics
industry for in-sample data. For out-of-sample data, there are 63 in the traditional,
38 in the manufacturing, and 72 in the electronics industries, respectively.
Table 27.2 displays the frequency distributions of the credit ratings for
in-samples in these three industries.

1
See Kamstra and Kennedy (1998) for the detail description.
734 C.-F. Lee et al.

Table 27.2 Frequency distributions of the credit ratings

Ratings Traditional Manufacturing Electronics
1 10 3 156
2 47 71 259
3 151 62 252
4 380 294 1,066
5 921 321 2,343
6 1,044 559 2,736
7 645 465 1,272
8 459 338 490
9 336 337 280
10 509 411 191
Subtotal 4,502 2,861 9,045
Note: Level 10 represents the bankruptcy class

Bonfim (2009) finds that not only the firms’ financial situation has a central role
in explaining default probabilities, but also macroeconomic conditions are very
important when assessing default probabilities over time. Based on previous studies
in the literature, 62 explanatory variables including financial ratios, market condi-
tions, and macroeconomic factors are considered. We use the hybrid stepwise
method to find the best predictors in the ordered probit and ordered logit models.
The combining technique using logit regression is then applied.

27.3.1 Model Estimates

27.3.1.1 Ordered Logit Model

Table 27.3 illustrates the in-sample estimation results under the ordered logit model
for each industry. From the likelihood ratio, score ratio, and Wald ratio, with
significant level at 1 %, we can determine the goodness-of-fit for each model.
For the traditional industry, the coefficients of Fixed Asset to Long Term Funds
Ratio (Fixed Asset to Equity and Long Term Liability Ratio), Interest Expense to Sales
Ratio, and Debt Ratio are positive. It shows that firms with higher ratios will get worse
credit ratings as well as higher default probabilities. On the other hand, the coefficients
of Accounts Receivable Turnover Ratio, Net Operating Profit Margin, Return on Total
Assets (Ex-Tax, Interest Expense), Depreciation to Sales Ratio, Free Cash Flow to
Total Debt Ratio, Capital Spending to Gross Fixed Assets Ratio, Retained Earning
to Total Assets Ratio, and Ln (Total Assets/GNP price-level index) are negative so
that firms tend to have good credit qualities as well as lower default probabilities
when these ratios become higher. All these signs meet our expectation.
For the manufacturing industry, the coefficient of the dummy variable for the
Negative Net Income for the last 2 years is positive, so the losses worsen the credit
rating. On the other hand, the coefficients of Equity to Total Asset Ratio, Total
Assets Turnover Ratio, Return on Total Assets (Ex-Tax, Interest Expense),
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 735

Table 27.3 Regression results estimated by the ordered logit. This table represents the
regression results estimated by the ordered logit model. Panel A shows the 11 explanatory vari-
ables fitted in the traditional industry. Panel B shows the nine explanatory variables fitted in the
manufacturing industry. Panel C shows the ten explanatory variables fitted in the electronics
industry
Parameters
Standard
Explanatory variables Estimates errors
Panel A: Traditional
X7 Fixed Asset to Long Term Funds Ratio 0.968*** (0.098)
X12 Accounts Receivable Turnover Ratio 0.054*** (0.0087)
X19 Net Operating Profit Margin 5.135*** (0.546)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 5.473*** (0.934)
X29 Depreciation to Sales Ratio 6.662*** (0.537)
X30 Interest Expense to Sales Ratio 22.057*** (1.862)
X35 Free Cash Flow to Total Debt Ratio 0.971*** (0.094)
X41 Capital Spending to Gross Fixed Assets Ratio 0.746*** (0.167)
X46 Debt Ratio 4.781*** (0.305)
X47 Retained Earning to Total Assets Ratio 5.872*** (0.336)
X50 Ln (Total Assets/GNP price-level index) 6.991*** (1.024)
Panel B: Manufacturing
X2 Equity to Total Asset Ratio 7.851*** (0.317)
X15 Total Assets Turnover Ratio 0.714*** (0.172)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 7.520*** (0.932)
X40 Accumulative Depreciation to Gross Fixed Assets 1.804*** (0.199)
X47 Retained Earning to Total Assets Ratio 3.138*** (0.340)
X50 Ln (Total Assets/GNP price-level index) 5.663*** (1.213)
X52 1:If Net Income was Negative for the Last Two Years 1.274*** (0.108)
0: Otherwise
X54 Ln (Age of the firm) 0.486*** (0.102)
X60 Ln (Net Sales) 1.063*** (0.065)
Panel C: Electronics
X2 Equity to Total Asset Ratio 3.239*** (0.207)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 6.929*** (0.349)
X30 Interest Expense to Sales Ratio 25.101*** (1.901)
X35 Free Cash Flow to Total Debt Ratio 0.464*** (0.031)
X40 Accumulative Depreciation to Gross Fixed Assets 0.781*** (0.127)
X44 Cash Reinvestment Ratio 0.880*** (0.167)
X45 Working Capital to Total Assets Ratio 3.906*** (0.179)
X47 Retained Earning to Total Assets Ratio 3.210*** (0.174)
X49 Market Value of Equity/Total Liability 0.048*** (0.004)
X50 Ln (Total Assets/GNP price-level index) 8.797*** (0.704)
***
Represents significantly different from zero at 1 % level
**
Represents significantly different from zero at 5 % level
*
Represents significantly different from zero at 10 % level
736 C.-F. Lee et al.

Accumulative Depreciation to Gross Fixed Assets, Retained Earning to Total

Assets Ratio, Ln (Total Assets/GNP price-level index), Ln (Age of the firm), and
Ln (Net Sales) are all negative. The ratios improve the credit standings.
For the electronics industry, the coefficient of Interest Expense to Sales Ratio is
positive, and the coefficients of Equity to Total Asset Ratio, Return on Total Assets
(Ex-Tax, Interest Expense), Free Cash Flow to Total Debt Ratio, Accumulative
Depreciation to Gross Fixed Assets, Cash Reinvestment Ratio, Working Capital to
Total Assets Ratio, Retained Earning to Total Assets Ratio, Market Value of
Equity/Total Liability, and Ln (Total Assets/GNP price-level index) are negative.
In general, the matured companies like those in the traditional and the manufactur-
ing industries should focus mainly on their capabilities in operation and in liquidity.
Also, the market factors are important for the manufacturing firms. For high-growth
industry like the electronics, we should pay more attention to their market factors and
the liquidity ratios. The common explanatory variables among three industries seem
related to the operating returns, the retained earnings, and the asset size.

27.3.1.2 Ordered Probit Model

Table 27.4 shows the in-sample estimation results using the ordered probit model
for each industry.
For the traditional industry, the coefficients of Fixed Asset to Long Term Funds
Ratio, Interest Expense to Sales Ratio, and Debt Ratio are positive in the ordered
probit model, which are similar to the results from the ordered logit model. On the
other hand, the coefficients of Accounts Receivable Turnover Ratio, Net Operating
Profit Margin, Return on Total Assets (Ex-Tax, Interest Expense), Depreciation to
Sales Ratio, Operating Cash Flow to Total Liability Ratio, Capital Spending to
Gross Fixed Assets Ratio, and Retained Earning to Total Assets Ratio are negative,
also showing no big difference with the results from the ordered logit model.
For the manufacturing industry, the coefficients of Accounts Payable Turnover
Ratio and the dummy variable for the negative Net Income for the last 2 years are
positive. On the other hand, the coefficients of Equity to Total Asset Ratio, Quick
Ratio, Total Assets Turnover Ratio, Return on Total Assets (Ex-Tax, Interest
Expense), Accumulative Depreciation to Gross Fixed Assets, Cash Flow Ratio,
Retained Earning to Total Assets Ratio, and Ln (Age of the firm) are negative. The
results are also similar to those of the ordered logit model, only that the ordered logit
seems focusing more on the size factors (sales, asset), while the ordered probit
concerns more on the liquidity (quick ratio, cash flow ratio, and payables) of the firm.
For the electronics industry, the coefficients of Interest Expense to Sales Ratio is
positive and the coefficients of Free Cash Flow to Total Debt Ratio, Working
Capital to Total Assets Ratio, Retained Earning to Total Assets Ratio, Market
Value of Equity/Total Liability, LN (Total Assets/GNP price-level index), and
LN (Age) are negative.
Table 27.5 shows the threshold values estimated by the two models. Threshold
values represent the cutting points for neighboring ratings.
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 737

Table 27.4 Regression results estimated by the ordered probit. This table represents the
regression results estimated by the ordered probit model. Panel A shows the ten explanatory
variables fitted in the traditional industry. Panel B shows the ten explanatory variables fitted in the
manufacturing industry. Panel C shows the nine explanatory variables fitted in the electronics
industry
Parameters
Explanatory variables Estimates Standard errors
Panel A: Traditional
X7 Fixed Asset to Long Term Funds Ratio 0.442*** (0.058)
X12 Accounts Receivable Turnover Ratio 0.027*** (0.005)
X19 Net Operating Profit Margin 2.539*** (0.311)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 3.335*** (0.534)
X29 Depreciation to Sales Ratio 3.204*** (0.287)
X30 Interest Expense to Sales Ratio 9.881*** (1.040)
X34 Operating Cash Flow to Total Liability Ratio 0.514*** (0.121)
X41 Capital Spending to Gross Fixed Assets Ratio 0.459*** (0.096)
X46 Debt Ratio 3.032*** (0.188)
X47 Retained Earning to Total Assets Ratio 3.317*** (0.192)
Panel B: Manufacturing
X2 Equity to Total Asset Ratio 4.648*** (0.183)
X10 Quick Ratio 0.036*** (0.009)
X11 Accounts Payable Turnover Ratio 0.021*** (0.004)
X15 Total Assets Turnover Ratio 0.742*** (0.072)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 4.175*** (0.522)
X40 Accumulative Depreciation to Gross Fixed Assets 0.989*** (0.112)
X43 Cash Flow Ratio 0.190*** (0.045)
X47 Retained Earning to Total Assets Ratio 1.804*** (0.188)
X52 1:If Net Income was Negative for the Last Two Years 0.753*** (0.061)
0: Otherwise
X54 Ln (Age of the firm) 0.289*** (0.056)
Panel C: Electronics
X2 Equity to Total Asset Ratio 1.851*** (0.117)
X27 Return on Total Assets (Ex-Tax, Interest Expense) 3.826*** (0.197)
X30 Interest Expense to Sales Ratio 11.187*** (1.052)
X35 Free Cash Flow to Total Debt Ratio 0.233*** (0.017)
X45 Working Capital to Total Assets Ratio 2.101*** (0.101)
X47 Retained Earning to Total Assets Ratio 1.715*** (0.098)
X49 Market Value of Equity/Total Liability 0.027*** (0.002)
X50 Ln (Total Assets/GNP price-level index) 4.954*** (0.399)
X54 Ln (Age of the firm) 0.247*** (0.028)
***
Represents significantly different from zero at 1 % level
**
Represents significantly different from zero at 5 % level
*
Represents significantly different from zero at 10 % level
738 C.-F. Lee et al.

Table 27.5 Threshold values estimated by the ordinal analysis. This table shows the threshold
values estimated by the ordered logit and the ordered probit models. There are nine threshold
parameters given ten credit ratings

Threshold Ordered logit model Ordered probit model

parameter Traditional Manufacturing Electronics Traditional Manufacturing Electronics
t1 23.361*** 28.610*** 31.451*** 12.272*** 16.891*** 17.306***
(0.633) (0.987) (0.478) (0.332) (0.447) (0.252)
t2 21.236*** 24.879*** 29.759*** 11.231*** 15.042*** 16.410***
(0.545) (0.766) (0.459) (0.294) (0.360) (0.243)
t3 19.441*** 23.848*** 28.855** 10.316*** 14.446*** 15.917***
(0.519) (0.749) (0.451) (0.282) (0.352) (0.240)
t4 17.639*** 21.657*** 26.846*** 9.339*** 13.175*** 14.797**
(0.502) (0.717) (0.436) (0.275) (0.337) (0.233)
t5 15.389*** 20.333*** 24.451*** 8.080*** 12.411*** 13.441***
(0.484) (0.703) (0.421) (0.268) (0.331) (0.227)
t6 13.317*** 18.309*** 21.777*** 6.918*** 11.251*** 11.937***
(0.473) (0.683) (0.407) (0.263) (0.322) (0.221)
t7 11.584*** 16.461*** 19.715*** 5.972*** 10.207*** 10.809***
(0.465) (0.670) (0.400) (0.261) (0.314) (0.218)
t8 9.825*** 14.738*** 18.019*** 5.049*** 9.246*** 9.929***
(0.459) (0.662) (0.397) (0.260) (0.308) (0.217)
t9 8.231*** 12.166*** 16.010*** 4.256*** 7.860*** 9.005***
(0.457) (0.655) (0.400) (0.260) (0.301) (0.219)
***
Represents significantly different from zero at 1 % level
**
Represents significantly different from zero at 5 % level
*
Represents significantly different from zero at 10 % level

27.3.2 Credit Rating Forecasting

Tables 27.6 and 27.7 illustrate the prediction results of the ordered logit and the
ordered probit models. Following Blume et al. (1998), we define the most probable
rating as the actual rating or its immediate adjacent ratings. The ratio of the number
of the predicted ratings as the most probable ratings to the total number of the
ratings being predicted can assess the goodness-of-fit for the model. For out-of-
sample firms, the predictive power of the ordered logit model for each industry is
86.85 %, 81.06 %, and 86.37 %, respectively; and the predictive power of the
ordered probit model for each industry is 86.42 %, 80.21 %, and 84.87 %,
respectively. The results from two models are quite similar.

27.3.3 Estimation Results Using the Combining Method

Table 27.8 depicts the regression results using the Kamstra-Kennedy combining
forecasting technique. The coefficients are the logit estimates on o’s. These o
values are all positive and strongly significant for each industry.
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 739

Table 27.6 Out-of-sample predictions by the ordered logit model

Panel A: Traditional
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 3 0 0 0 0 0 0 0
2 0 0 0 6 5 1 0 0 0 0
3 0 0 4 8 12 0 0 0 0 0
4 0 0 4 26 34 7 0 0 0 0
5 0 0 0 14 101 41 2 0 0 0
6 0 0 0 2 53 104 29 5 0 0
7 0 0 0 0 5 24 26 11 0 0
8 0 0 0 0 0 21 26 13 2 3
9 0 0 0 0 0 1 4 15 7 10
10 0 0 0 0 0 3 2 9 15 34
Prediction ratio 86.85%
Panel B: Manufacturing
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 0 0 0 0 0 0 0 0
2 0 3 0 4 3 2 0 0 0 0
3 0 2 2 4 2 5 0 0 0 0
4 0 0 0 7 21 24 0 1 0 0
5 0 0 0 5 29 51 5 3 0 0
6 0 0 0 8 20 68 18 3 0 0
7 0 0 1 1 0 18 38 9 3 1
8 0 0 0 0 0 4 16 5 9 1
9 0 0 0 0 0 0 9 12 10 5
10 0 0 0 0 0 0 4 5 8 21
Prediction ratio 81.06%
Panel C: Electronics
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 1 2 0 15 0 0 0 0 0
2 0 0 0 5 22 2 0 0 0 0
3 0 0 0 1 36 11 2 0 0 0
4 0 2 4 13 144 42 0 0 0 0
5 0 1 1 9 279 180 15 1 0 0
6 0 0 0 11 293 303 39 2 0 0
7 0 0 0 0 16 206 79 9 0 1
8 0 0 0 0 2 50 80 26 10 2
9 0 0 0 0 0 6 11 21 22 13
10 1 0 0 0 3 2 5 12 15 34
Prediction ratio 86.37%
740 C.-F. Lee et al.

Table 27.7 Out-of-sample predictions by the ordered probit model

Panel A: Traditional
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 2 1 0 0 0 0 0 0 0
2 0 0 0 4 6 2 0 0 0 0
3 0 0 3 5 16 0 0 0 0 0
4 0 0 6 22 36 7 0 0 0 0
5 0 0 0 10 97 48 3 0 0 0
6 0 0 0 2 45 106 36 4 0 0
7 0 0 0 0 4 28 21 13 0 0
8 0 0 0 0 1 19 26 14 2 3
9 0 0 0 0 1 0 5 10 9 12
10 0 0 0 0 0 2 3 11 15 32
Prediction ratio 86.42%
Panel B: Manufacturing
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 0 0 0 0 0 0 0 0
2 1 2 0 4 5 0 0 0 0 0
3 1 1 2 4 2 5 0 0 0 0
4 0 0 0 9 20 23 1 0 0 0
5 0 0 0 6 28 51 5 3 0 0
6 0 0 0 7 22 71 15 2 0 0
7 0 0 0 1 1 23 34 8 3 1
8 0 0 0 0 0 6 15 4 9 1
9 0 0 0 0 0 0 12 8 11 5
10 0 0 0 0 0 1 4 5 8 20
Prediction ratio 80.21%
Panel C: Electronics
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 0 3 15 0 0 0 0 0
2 0 0 0 1 26 2 0 0 0 0
3 0 0 0 0 40 8 2 0 0 0
4 0 0 0 5 155 45 0 0 0 0
5 0 0 0 7 295 164 20 0 0 0
6 0 0 0 8 313 286 40 1 0 0
7 0 0 0 0 11 207 87 5 0 1
8 0 0 0 0 2 54 96 14 2 2
9 0 0 0 0 0 5 26 30 6 6
10 0 0 0 0 3 4 12 21 8 24
Prediction ratio 84.87%
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 741

Table 27.8 Regression results estimated by the combining forecast

Threshold Combining forecasting model

parameter Traditional Manufacturing Electronics
t1 13.1875 (0.3033)*** 13.6217 (0.5475)*** 11.6241 (0.1591)***
t2 11.3484 (0.1854)*** 9.9864 (0.1918)*** 10.5484 (0.1483)***
t3 9.8997 (0.1524)*** 9.1712 (0.1704)*** 9.973 (0.1456)***
t4 8.3781 (0.1370)*** 7.5337 (0.1441)*** 8.7037 (0.143)***
t5 6.4390 (0.1257)*** 6.4459 (0.1315)*** 7.0973 (0.1414)***
t6 4.5627 (0.1137)*** 4.8259 (0.1142)*** 5.0093 (0.1364)***
t7 3.1037 (0.1016)*** 3.3724 (0.1003)*** 3.1906 (0.1261)***
t8 1.5619 (0.0885)*** 2.0849 (0.0913)*** 1.5682 (0.1145)***
t9 0.0816 (0.0861) 0.1127 (0.0929) 0.3735 (0.1169)**
oLogit 1.1363 (0.0571)*** 1.2321 (0.029)*** 0.6616 (0.0528)***
oProbit 0.0825 (0.0330)* 0.1437 (0.0085)*** 0.1867 (0.0272)***
Numbers in parentheses represent the standard errors
***
Represents significantly different from zero at 1 % level
**
Represents significantly different from zero at 5 % level
*
Represents significantly different from zero at 10 % level

Table 27.9 shows the prediction results of the combining forecasting model. For
out-of-sample test, the predictive power of the combining model for each industry
is 89.88 %, 82.77 %, and 88.02 %, respectively, which are higher than those of the
ordered logit or ordered probit models by 2–4 %.

27.3.4 Performance Evaluation

To evaluate the performance of each model, Fig. 27.1 illustrates the ROC curves
estimated by the three models, respectively.
From these ROC curves we can distinguish the performance of each rating
model. Furthermore, we can compare the AUC and AR calculated from the ROC
and CAP (See Table 27.10).
For the traditional industry, the AUCs from the ordered logit, the ordered probit, and
the combining model are 95.32 %, 95.15 %, and 95.32 %, respectively. For the
manufacturing industry, they are 94.73 %, 93.66 %, and 95.51 %, respectively. And
for the electronics industry, they are 92.43 %, 92.30 %, and 94.07 %, respectively. These
results apparently show that the combining forecasting model performs better than any
individual one.

27.4 Conclusion

This study constitutes an attempt to explore the credit rating forecasting

techniques. The samples consist of firms in the TSE and the OTC market and
742 C.-F. Lee et al.

Table 27.9 Out-of-sample credit rating prediction by the combining forecast

Panel A: Traditional
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 2 1 0 0 0 0 0 0
2 0 0 0 6 6 0 0 0 0 0
3 0 0 4 7 13 0 0 0 0 0
4 0 0 2 28 34 7 0 0 0 0
5 0 0 0 12 100 44 2 0 0 0
6 0 0 0 2 48 113 26 4 0 0
7 0 0 0 0 5 27 23 11 0 0
8 0 0 0 0 0 23 25 12 2 3
9 0 0 0 0 0 1 4 14 8 10
10 0 0 0 0 0 3 2 9 16 33
Prediction ratio 89.88%
Panel B: Manufacturing
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 0 0 0 0 0 0 0 0 0
2 0 2 1 3 6 0 0 0 0 0
3 0 1 2 5 2 5 0 0 0 0
4 0 0 0 6 24 22 0 1 0 0
5 0 0 0 5 30 50 5 3 0 0
6 0 0 0 6 23 70 17 1 0 0
7 0 0 1 1 0 27 32 7 2 1
8 0 0 0 0 1 3 16 6 8 1
9 0 0 0 0 0 1 8 12 11 4
10 0 0 0 0 0 0 4 4 7 23
Prediction ratio 82.77%
Panel C: Electronics
Predicted rating
1 2 3 4 5 6 7 8 9 10
Actual rating 1 0 3 0 0 15 0 0 0 0 0
2 0 0 0 4 18 7 0 0 0 0
3 0 0 0 1 34 13 2 0 0 0
4 0 4 4 13 128 56 0 0 0 0
5 0 1 1 11 256 193 23 1 0 0
6 0 0 0 12 248 338 47 3 0 0
7 0 0 0 0 14 190 91 15 0 1
8 0 0 0 0 1 44 82 32 10 1
9 0 0 0 0 0 5 10 23 24 11
10 0 0 0 0 3 1 6 11 17 34
Prediction ratio 88.02%
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 743

a 1.0 b 1.0

0.8 0.8
Hit Rate

Hit Rate
0.6 0.6

0.4 Source of the Curve 0.4 Source of the Curve

Ordered Logit Ordered Logit
Ordered Probit Ordered Probit
0.2 0.2
Combine Forecast Combine Forecast
Random Model Random Model
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
False Alarm Rate False Alarm Rate

c 1.0

0.8

0.6
Hit Rate

0.4
Source of the Curve
Ordered Logit
0.2 Ordered Probit
Combine Forecast
Random Model
0.0
0.0 0.2 0.4 0.6 0.8 1.0
False Alarm Rate

Fig. 27.1 ROC curves, (a) Traditional, (b) Manufacturing, (c) Electronics

Table 27.10 Performance evaluation for each model

Traditional Manufacturing Electronics
Panel A: Ordered logit
AUC 95.32 % 94.73 % 92.43 %
AR 90.63 % 89.46 % 84.86 %
McFadden’s R-square 35.63 % 38.25 % 39.63 %
Panel B: Ordered probit
AUC 95.15 % 93.66 % 92.30 %
AR 90.30 % 87.32 % 84.60 %
McFadden’s R-square 34.45 % 40.05 % 41.25 %
Panel C: Combining forecasting
AUC 95.32 % 95.51 % 94.07 %
AR 90.63 % 91.03 % 88.15 %
McFadden’s R-square 42.34 % 43.16 % 46.28 %
CAP represents the cumulative accuracy profile, AR represents accuracy ratio. McFadden’s R2 is
defined as 1(unrestricted log-likelihood function/restricted log-likelihood function)
744 C.-F. Lee et al.

are divided into three industries, i.e., traditional, manufacturing, and electronics,
for analysis. Sixty-two explanatory variables consisting of financial, market, and
macroeconomics factors are considered. We utilize the ordered logit, the ordered
probit, and the combining forecasting model to estimate the parameters
and conduct the out-of-sample tests. The main result is that the combining
forecasting method leads to a more accurate rating prediction than that of any
single use of the ordered logit or ordered probit analysis. By means of cumula-
tive accuracy profile, the receiver operating characteristics, and McFadden R2,
we can measure the goodness-of-fit and the accuracy of each prediction model.
These performance evaluations depict consistent results that the combining
forecast performs better.

Appendix 1: Ordered Probit Procedure for Credit Rating

Forecasting2

Credit rating is an ordinal scale from which the credit category of a firm can be
ranked from high to low, but the scale of the difference between them is unknown.
To model the ordinal outcomes, we follow Zavoina and McKelvey (1975) to begin
with a latent regression

Y
¼ Xb þ e (27.8)

where
2 3 2 3
Y
1 1 X11  XK1
Y
¼ 4 ⋮ 5, X ¼ 4 ⋮ ⋮ ⋮ 5
Y
N 1 X1N    XKN
2 3 2 3
b0 e0
b ¼ 4 ⋮ 5, e ¼ 4 ⋮ 5:
bK eN

Here, b is a vector of unknown parameters, X is a set of explanatory variables,

and e is a random disturbance term assumed to follow the multivariate normal
distribution with mean 0 and variance-covariance matrix s2I, that is,
 
e  N 0, s2 I : (27.9)
Y*, the dependent variable of theoretical interests, is unobserved, but from which
we can classify the category j:3

2
There are detailed discussion about ordered data in Ananth and Kleinbaum (1997), McCullagh
(1980), Wooldridge (2010), and Greene (2011).
3
In our case, J ¼ 10 and K ¼ 62.
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 745

Y i ¼ j if tj1 < Y
i  tj ði ¼ 1, 2, . . . , N; j ¼ 1, 2, . . . , J Þ (27.10)

where Y is the one we do observe, an ordinal version of Y*, and the t’s are
unknown parameters to be estimated with b. We assume
that  1 ¼ t0  t1      tJ ¼ + 1.
From Eqs. 27.8 and 27.10, we have
tj1 < Y
i  tj , tj1 < Xi b þ ei  tj
tj1  Xi b ei tj  Xi b
, <  (27.11)
s s s

where Xi is the ith row of X.

From Eqs. 27.9 and 27.11, the probability of Yi ¼ j can be written as
   
tj  X i b tj1  Xi b
PrðY i ¼ jÞ ¼ F F (27.12)
s s

where F(.) represents the cumulative density function of standard normal distri-
bution. The model (27.12) is under-identified since any linear transformation of
the underlying scale variable Y*, if applied to the parameters and t0,. . .,tJ as well,
would lead to the same model. We will assume without loss of generality
that,t1 ¼ 0 and s ¼ 1 in order to identify the model. The model we will estimate
turns out to be
   
PrðY i ¼ jÞ ¼ F tj  Xi b  F tj1  Xi b : (27.13)

Maximum likelihood estimation can be used to estimate the J + K  1 param-

eters, t2,. . .,tJ  1 and b0, b1,. . .,bK, in Eq. 27.13. To form the likelihood function,
first, define a dummy variable Yi,j:

1 if Y i ¼ j
Y i, j ¼ :
0 otherwise

Then, for simple notation, we set Zi,j ¼ tj  Xib and Fi,j ¼ F(Zi,j). The likelihood
function, L, is

Y i, j
n Y
Y J  
L ¼ Lðb0 , . . . , bK , t2 , . . . , tJ1 jY Þ ¼ Fi, j  Fi, j1 : (27.14)
i¼1 j¼1

So, the log-likelihood function, ln L, is

X
n X
J  
ln L ¼ Y i, j ln Fi, j  Fi, j1 : (27.15)
i¼1 j¼1
746 C.-F. Lee et al.

Now, we want to maximize ln L subject to t1  t2  . . .  tJ  1.

Z2
 i2, j 1 if l ¼ j
Let N i, j ¼ p1ffiffiffiffi e for 1  j  J and 1  i  N, and let dl, j ¼ ,
2p 0 if l ¼
6 j
it follows that

∂Fi, j ∂Zi, j
¼ N i, j ¼ N i, j Xu, i for 0  u  K
∂bu ∂bu (27.16)
∂Fi, j ∂Zi, j
¼ N i, j ¼ N i, j dl, j for 2  l  J  1
∂tl ∂tl

and

∂N i, j
¼ Z i, j N i, j Xu, i for 0  u  K
∂bu (27.17)
∂N i, j
¼ Zi, j N i, j dl, j for 2  l  J  1
∂tl

By using Eqs. 27.16 and 27.17, we can calculate the J + K  1 partial derivatives
of Eq. 27.15 with respect to the unknown parameters, b and t, respectively:

 
∂ln L X n X J
N i, j1  N i, j Xu, i
¼ Y i, j for 0  u  K
∂bu i¼1 j¼1
Fi, j  Fi, j1
  (27.18)
∂ln L Xn X J
N i, j dl, j  N i, j1 dl, j1 Xu, i
¼ Y i, j for 2  l  J  1:
∂tl i¼1 j¼1
Fi, j  Fi, j1

And the elements in the (J + K  1)  (J + K  1) matrix of second partials are

∂2 ln L Xn X J
¼ Y i, j
∂bu ∂bV i¼1 j¼1

    2
Fi, j  Fi, j1 Z i, j1 N i, j1  Z i, j N i, j  N i, j1  N i, j Xu, i Xv, i
 2 ,
Fi, j  Fi, j1
"  
∂2 ln L ∂2 ln L X n X J
Fi, j  Fi, j1 Zi, j N i, j dl, j  Zi, j1 N i, j1 dl, j1
¼ ¼ Y i, j  2
∂bu ∂tl ∂tl ∂bu i¼1 j¼1 Fi, j  Fi, j1
  #
N i, j1  N i, j N i, j dl, j  N i, j1 dl, j1
  2 Xu, i ,
Fi, j  Fi, j1
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 747

"  
∂2 ln L X n X J
Fi, j  Fi, j1 Zi, j1 N i, j1 dm, j1 dl, j1  Zi, j N i, j dm, j dl, j
¼ Y i, j  2
∂tl ∂tm i¼1 j¼1 Fi, j  Fi, j1
  #
N i, j dl, j  N i, j1 dl, j1 N i, j dm, j  N i, j1 dm, j1
  2 :
Fi, j  Fi, j1
(27.19)

We then set the J + K  1 equations in Eq. 27.18 to zero to get the MLE
of the unknown parameters. The matrix of second partials should be negative
definite to insure that the solution is a maximum. The computer program
NPROBIT, which uses the Newton-Raphson method, can solve the nonlinear
equations in Eq. 27.18.

Appendix 2: Ordered Logit Procedure for Credit Rating

Forecasting

Consider a latent variable model where Y* is the unobserved dependent variable, X a

set of explanatory variables, b an unknown parameter vector, and e a random
disturbance term:

Y
¼ Xb þ e (27.20)

e is assumed to follow a standard logistic distribution, so the probability density

function of e is

expðeÞ
lð e Þ ¼ (27.21)
½1 þ expðeÞ2

and the cumulative density function is

expðeÞ
LðeÞ ¼ : (27.22)
1 þ expðeÞ

Y*, the dependent variable of theoretical interests, is unobserved, but from which
we can classify the category j:

Y i ¼ j if tj1 < Y
i  tj ði ¼ 1, 2, . . . , N; j ¼ 1, 2, . . . , J Þ (27.23)

where Y is the one we do observe, an ordinal version of Y*, and the t’s are unknown
parameters satisfying t1  . . .  tj and to be estimated with b.
748 C.-F. Lee et al.

From Eqs. 27.20 and 27.22, we form the proportional odds model:
 
exp tj  Xi b
PrðY i  jjXÞ ¼   (27.24)
1 þ exp tj  Xi b

or equivalently
 
  Pij
log it Pij ¼ log
 1  Pij (27.25)
PrðY i  jjXi Þ
log ¼ tj  X i b
PrðY i > jjXi Þ

where Pij ¼ Pr(Yi  j). Notice that in the proportional odds model, b is assumed to
be constant and not depend on j. The validity of this assumption can be checked
based on a w2 score test. The model that relaxes the proportional odds assumption
can be represented as
 
log it Pij ¼ tj  Xi bj , (27.26)

where the regression parameter vector b is allowed to vary with j. Both models can
be fit through the procedure of maximum likelihood estimation.

Appendix 3: Procedure for Combining Probability Forecasts

To combine the ordered logit and the ordered probit models for credit forecasting,
the logit regression method as described in Kamstra and Kennedy (1998) is applied.
We first assume that firm’s credit classification is determined by an index y.
Suppose there are J rating classes, ordered from 1 to J. If y exceeds the threshold
value tj, j ¼ 1,. . . J  1, credit classification changes from j rating to j + 1 rating.
The probability of company i being in rating j is given by the integral of a standard
logit from tj  1  yi to tj  yi.
Each forecasting method is considered as producing J  1 measures,
oji ¼ tj  y, j ¼ 1, . . ., J  1 for each firm. These measures can be estimated as
 
P1i þ    þ Pji
oji ¼ ln (27.27)
1  P1i      Pji

where Pji is a probability estimate for firm i in rating j. For firm i we have
8 oji
>
>
e
for j ¼ 1
>
> oji
< 1 þoeji
> oj1i
e e
Pji ¼  for j ¼ 2  J  1 : (27.28)
>
> 1 þ e oji 1 þ eoj1i
>
>
>
:
1
for j ¼ J
1 þ eoj1i
27 Using Alternative Models and a Combining Technique in Credit Rating Forecasting 749

In our case, there are two forecasting techniques A and B. For firm i, the
combining probability estimate is
8 pj þpA ojiA þpB ojiB
>
>
e
for j ¼ 1
>
> pj þpA ojiA þpB ojiB
> þ
< pjeþpA ojiA þpB ojiB
1
e epj1 þpA oj1iA þpB oj1iB
Pji ¼  for j ¼ 2  J  1 :
>
> 1þe p þp o þp o 1 þ epj1 þpA oj1iA þpB oj1iB
>
j A jiA B jiB
>
> 1
: for j ¼ J
1 þ epj1 þpA oj1iA þpB oj1iB
(27.29)

This can be estimated using an ordered logit software package with the oA and
oB values as explanatory variables and the t1 and p2 parameters playing the role of
the unknown threshold values.

References
Altman, E. I. (1968). Financial ratios discriminant analysis and the prediction of corporate
bankruptcy. Journal of Finance, 23, 589–609.
Altman, E. I., & Katz, S. (1976). Statistical bond rating classification using financial and account-
ing data. In Proceedings of the conference on topical research in accounting, New York.
Altman, E. I., & Sounders, A. (1997). Credit risk measurement: Developments over the last
20 years. Journal of Banking and Finance, 21, 1721–1742.
Altman, E. I., Hadelman, R. G., & Narayanan, P. (1977). Zeta analysis, a new model to identify
bankruptcy risk of corporations. Journal of Banking and Finance, 10, 29–51.
Ananth, C. V., & Kleinbaum, D. G. (1997). Regression models for ordinal responses: A review of
methods and applications. International Journal of Epidemiology, 26, 1323–1333.
Beaver, W. (1966). Financial ratios as predictors of failure. Journal of Accounting Research, 4,
71–111.
Blume, M. E., Lim, F., & Mackinlay, A. C. (1998). The declining credit quality of U.S. corporate
debt: Myth or reality? Journal of Finance, 53, 1389–1413.
Bonfim, D. (2009). Credit risk drivers: Evaluating the contribution of firm level information and of
macroeconomic dynamics. Journal of Banking and Finance, 33, 281–299.
Clemen, R. T. (1989). Combining forecasts: A review and annotated bibliography. International
Journal of Forecasting, 5, 559–583.
Crouhy, M., Galai, D., & Mark, R. (2000). A comparative analysis of current credit risk models.
Journal of Banking and Finance, 24, 59–117.
Ederington, L. H. (1985). Classification models and bond ratings. The Financial Review, 20,
237–261.
Granger, C. W. (1989). Invited review: Combining forecasts-twenty years later. Journal of
Forecasting, 8, 167–173.
Greene, W. H. (2011). Econometric analysis. NJ: Prentice Hall.
Hand, D., & Henley, W. (1997). Statistical classification methods in consumer credit scoring:
A review. Journal of the Royal Statistical Society, Series A, 160, 523–541.
Kamstra, M., & Kennedy, P. (1998). Combining qualitative forecasts using logit. International
Journal of Forecasting, 14, 83–93.
Kaplan, R. S., & Urwitz, G. (1979). Statistical models of bond ratings: A methodological inquiry.
Journal of Business, 52, 231–261.
750 C.-F. Lee et al.

Lawrence, E. C., & Arshadi, N. (1995). A multinomial logit analysis of problem loan resolution
choices in banking. Journal of Money, Credit and Banking, 27, 202–216.
Mahmoud, E. (1984). Accuracy in forecasting: A survey. Journal of Forecasting, 3, 139–159.
McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society,
42, 109–142.
Ohlson, J. A. (1980). Financial ratios and the probabilistic prediction of bankruptcy. Journal of
Accounting Research, 18, 109–131.
Pinches, G. E., & Mingo, K. A. (1973). A multivariate analysis of industrial bond ratings. Journal
of Finance, 28, 1–18.
Pompe, P. M., & Bilderbeek, J. (2005). The prediction of bankruptcy of small- and medium-sized
industrial firms. Journal of Business Venturing, 20, 847–868.
Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.).
Cambridge, MA: MIT Press.
Zavoina, R., & McKelvey, W. (1975). A statistical model for the analysis of ordinal level
dependent variables. Journal of Mathematical Sociology, 4, 103–120.
Zmijewski, M. E. (1984). Methodological issues related to the estimation of financial distress
prediction models. Journal of Accounting Research, 22, 59–68.
Can We Use the CAPM as an Investment
Strategy?: An Intuitive CAPM and 28
Efficiency Test

Fernando Gómez-Bezares, Luis Ferruz, and Maria Vargas

Contents
28.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 752
28.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 756
28.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759
28.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 762
28.4.1 Strategy 1: The Investor Buys All the Undervalued Stocks . . . . . . . . . . . . . . . . . . 762
28.4.2 Strategy 2: The Investor Buys the Three Stock-Quartiles with the Highest
Degree of Undervaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
28.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 767
Strategy 1: Investor Buys All Undervalued Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
Strategy 2: The Investor Buys the Most Undervalued Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788

Abstract
The aim of this chapter is to check whether certain playing rules, based on the
undervaluation concept arising from the CAPM, could be useful as investment
strategies, and can therefore be used to beat the Market. If such strategies work,
we will be provided with a useful tool for investors, and, otherwise, we will

The authors wish to thank Professor José Vicente Ugarte for his assistance in the statistical analysis
of the data presented in this chapter.
F. Gómez-Bezares
Universidad de Deusto, Bilbao, Spain
e-mail: f.gomez-bezares@deusto.es
L. Ferruz (*)
Facultad de Economı́a y Empresa, Departamento de Contabilidad y Finanzas, Universidad de
Zaragoza, Zaragoza, Spain
e-mail: lferruz@unizar.es
M. Vargas
Universidad de Zaragoza, Zaragoza, Spain
e-mail: mvargas@unizar.es

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 751
DOI 10.1007/978-1-4614-7750-1_28,
# Springer Science+Business Media New York 2015
752 F. Gómez-Bezares et al.

obtain a test whose results will be connected with the Efficient Market Hypoth-
esis (EMH) and with the CAPM.
The basic strategies were set out in Gómez-Bezares, Madariaga, and
Santibáñez (Análisis Financiero 68:72–96, 1996). Our purpose now is to recon-
sider them, to improve the statistical analysis, and to examine a more recent
period for our study.
The methodology used is both intuitive and rigorous: analyzing how many
times we beat the Market with different strategies, in order to check whether
beating the Market happens by chance. Furthermore, we set out to study,
statistically, when and by how much we beat it, and to analyze whether this is
significant.

Keywords
ANOVA • Approximately normal distribution • Binomial distribution • CAPM •
Contingency tables • Market efficiency • Nonparametric tests • Performance
measures

28.1 Introduction

Finance, as it is currently taught in business schools and brought together in the

most prestigious text books, places great importance on the Capital Asset Pricing
Model (CAPM)1 and on the Efficient Market Hypothesis (EMH)2. For example,
Brealey, Myers, and Allen (2008) conclude their famous book by saying that
within what we know about Finance, there are seven key ideas: the first of which
is Net Present Value (NPV), the second CAPM, and the third EMH3;
leaving aside for the moment NPV,4 we have as the two clearly outstanding
concepts the CAPM and the EMH. In our opinion, asset pricing models
(above all the CAPM) and the efficiency of the Markets are among the founda-
tions of the current paradigm which has been in vogue since the 1970s.5 And this
is perfectly logical; the financial objective of a company is to maximize its
Market value6; to this end it must make investments with a positive NPV,
and to calculate the NPV financiers require an asset pricing model such as the
CAPM; ultimately the Markets need to be efficient so that they can notice
increases in value provided by investments. We can thus see by this simple
reasoning how the three concepts highlighted by Brealey, Myers, and Allen are
interrelated.

1
Sharpe (1964), Lintner (1965).
2
One of its main defenders has been Fama (1970, 1991, 1998).
3
There are also other four key ideas.
4
A fundamental concept, predating the other two and clearly related to the CAPM and the EMH.
5
See Gómez-Bezares and Gómez-Bezares (2006). Also of interest could be Dimson and
Mussavian (1998, 1999).
6
The interesting work by Danielson, Heck and Shaffer (2008) is also worth looking at.
28 Can We Use the CAPM as an Investment Strategy? 753

An alternative way of viewing this is to say that Finance, as it is currently

understood, is the science of pricing; we have to valuate so that we can know which
decisions will result in the greatest added value, and to valuate we need asset
pricing models (like the CAPM). Finally, the result of our labor will be recognized
by the Market, if the Market is efficient and values more highly those stocks with
a higher intrinsic value.
The Golden Age of these two principles was the decade of the 1970s, appearing
in works such as Fama (1970), Black, Jensen, and Scholes (1972) and Fama and
MacBeth (1973): the Market correctly values the assets (the EMH) and we
have a good model of asset pricing (the CAPM). However, since then there have
been many critiques of both of these principles: the efficient Market principle has
been criticized by psychologists and behavioral economists, who question the
rationality of human beings, as well as by econometricians, who claim that prices
are susceptible to prediction (Malkiel 2003). There is ample criticism in the
literature, but the EMH also has important apologists; among the classic defenders
are the works of Fama (1991, 1998) as well as the very noteworthy and recent study
by Fama and French (2010), in which they show that portfolio managers, on
average, do not beat the Market, thereby proving that the Market is indeed efficient.
The detractors of the efficient Market hypothesis (currently, above all, psychol-
ogists and behavioral economists) underline again and again the inefficiencies
they observe. These problems with the model, when they are seen to occur
repeatedly, are termed “anomalies” (for a summary, see Malkiel 2003). Fama and
French (2008) studied different anomalies, concluding that some are more
significant than others. However, they concluded that we cannot be sure whether
to attribute abnormal returns to inefficiencies of the Market or to rewards for
risk-taking.
This brings us back to the debate regarding asset pricing models. According to the
CAPM, the expected return on an asset should be a positive linear function of its
systematic risk measured by beta, which is the sole measurement of risk.
This statement has been questioned in many ways; one very important contribution
in this respect is that of Fama and French (1992) in which they comment that the beta
can contribute little to an explanation of expected returns and, in fact, that there are
other variables that can explain a lot more. Their work gave rise to the famous Fama
and French three-factor model. The existence of variables other than the beta which
can help to explain mean returns is not in itself an argument against the efficiency of
the Market if we consider them to be risk factors: as risk is our enemy we demand
higher returns for higher risk, and we can measure this risk in terms of a range of
factors. This approach may be compatible with Arbitrage Pricing Theory (APT), but
it conflicts with the CAPM.
But we could also consider that the problem is that the Markets are inefficient,
and hence expected returns respond to variables that are not risk factors, variables to
which they should not react. Let us take as example momentum, which tells us that
recent past high (or low) returns can help to predict high (or low) returns in the near
future. Some may think that this could be used as a risk factor to explain returns
whereas others would say that it is an inefficiency of the Market. The truth is that
754 F. Gómez-Bezares et al.

the asset pricing model and efficiency are tested jointly (Fama 1970, 1991, 1998), in
such a way that, when they work, they both work,7 but when they fail we cannot
know which of them has failed.8
The CAPM tests fall far short of giving definitive results. Although some have
given it up for dead and buried because it cannot fit the data or because of
theoretical problems, one recent magnificent study by Levy (2010) rejects
the theoretical problems posed by psychologists and behavioral economists and,
based on ex ante data, says that there is experimental support for the CAPM.
Brav, Lehavy, and Michaely (2005), in an approach related to the above, set out
to test the model, not by looking at past returns (as a proxy of expected returns) but
instead at the expectations of Market analysts, on the basis that these may be
assumed to be unbiased estimates of the Market expectations. From their analysis,
they found a positive relationship between expected returns and betas.
Asset pricing models (such as the CAPM) look at the relationship between return
and risk. Recently, there has been more focus on liquidity. Liu (2006) builds an
enhanced CAPM based on two factors: the Market and liquidity – he obtains their
corresponding factorial weighting which he uses to explain the expected returns.
The model works correctly and allows him to account several anomalies; according
to Liu, his enhanced CAPM lends new support to the risk-return paradigm.
The methods most commonly used to contrast the CAPM are time-series and
cross-sectional studies, which each present different statistical problems. What we
set out to achieve in this chapter is to replicate a simple and intuitive but also
rigorous methodology that we believe is much less problematic from a statistical
point of view. The procedure was first proposed by Gómez-Bezares, Madariaga,
and Santibáñez (1996), and in this chapter we attempt to replicate it with a more
updated sample and greater statistical rigor. The basic idea is simple: an individual
who knows the CAPM calculates at the beginning of each month the return that
each stock has rendered in the past and compares it with the return it ought to have
given according to the CAPM. The stocks which gave higher returns than they
ought to have are cheap (undervalued) while those which gave lower return are
expensive (overvalued). If we assume that return levels are going to persist, the
investor should buy the cheap stocks and sell off the expensive ones in order to beat
the Market. We illustrate this line of reasoning in Fig. 28.1.
In Fig. 28.1, the betas appear on the horizontal axis and the expected return on
the vertical axis (Rf is the riskless rate; Rm is the Market return). We also show
the Security Market Line (SML), with its formula above it, which is the formula
for the CAPM. Our investor decides to buy the stocks which are cheap and
refrain from selling short the expensive stocks (due to the limitations on short
selling). Based on the data at his disposal, the stocks which have gained a value
more than that the SML indicates are undervalued, that is to say, they are above

7
We refer to the situation we have described previously which occurred at the beginning of
the 1970s.
8
See also Copeland, Weston, and Shastri (2005, p. 244).
28 Can We Use the CAPM as an Investment Strategy? 755

Fig. 28.1 shows the Security E(Ri)

Market Line (SML). The E (Ri) = Rf + [E (Rm) – Rf] βi
betas appear on the horizontal SML
E(Rm)
axis, and the expected return
on the vertical axis

Rf

1 βi

the SML. Believing that this situation will continue, he will buy all of these
stocks, trusting that he will obtain an adjusted return higher than the Market
return. If this happens (which he can find out by using Jensen’s alpha), and this
strategy consistently proves to result in abnormal returns (positive Jensen alphas,
or to put it another way, the stocks stay above the SML in the next period), he
will have found a way to beat the Market; therefore it is not efficient. Moreover,
he shall be able to predict which stocks will, in the future, outperform the
SML and therefore do not comply with the CAPM, which goes against the
CAPM. There would then be one way left to save the EMH: risk should not
be measured using the beta (i.e., the CAPM is mistaken) and therefore a positive
value for alpha is not synonymous with beating the Market (which could not be
done consistently in an efficient Market); we could also save the CAPM as
follows: risk must be measured with the beta; however, the Market fails to value
stocks accurately and hence it can be beaten; what we cannot do is to save both
the EMH and the CAPM simultaneously.
On the other hand, if our investor cannot consistently beat the Market with his
strategy but can only achieve about a 50 % success rate, more or less, we must
conclude that the portfolio assembled with the previously undervalued stocks
sometimes performs better than the SML and other times not as well, purely by
chance, and then settles down to the value it ought to have according to the
CAPM. We will not have a Market-beating strategy, and the results are compatible
with the EMH; likewise, we will see that on a random basis the portfolios perform
better than the SML at times and other times worse, which is compatible with the
CAPM (every month there may be random oscillations around the SML); therefore
two key elements of the aforementioned paradigm would be rescued.
In our research, we use Jensen’s (1968, 1969) and Treynor’s (1965) indices,
since they consider systematic risk. These indices have been very widely used in the
literature.9

9
Recent studies that use these indices are, for example, Samarakoon and Hasan (2005), Abdel-
Kader and Qing (2007), Pasaribu (2009), Mazumder and Varela (2010), and Ferruz, Gómez-
Bezares, and Vargas (2010).
756 F. Gómez-Bezares et al.

One of the virtues of this method is that it is perfectly replicable, as we have

hypothesized an investor who makes decisions based on information in the public
domain whenever he makes them. Meanwhile, we feel, reasonably enough, that the
beta values and the risk premium may vary over time. We also use the most liquid
stocks (to avoid problems with lack of liquidity) and portfolios (which reduces the
measuring problems).
The results of our analysis clearly indicate that our strategy is not capable of
beating the Market consistently, and therefore it is compatible with both the CAPM
and the EMH. The result, to which we have come via different routes, which
supports the strategy’s robustness, supports the results reported by Fama and French
(2010) and Levy (2010) although they used very different methods. Doubtless, both
the CAPM and the EMH are simplifications of reality, but they help us to explain
that reality.
The rest of our chapter is organized as follows: in Sect. 28.2, we describe the
sample analyzed with information about the assets that it comprises, and we comment
on the method used to calculate the returns and betas; in Sect. 28.3, we comment on the
method employed, with reference to the proposed strategies and how these were
formulated, we then go on to analyze the scenarios in which the strategies succeed
in beating the Market and we describe the statistical tests used for the analyses; in
Sect. 28.4, we show the results obtained by the different strategies and we analyze the
number of occasions on which we were able to beat the Market and whether these
results occur by chance. Furthermore, we analyze when and in what magnitude we
beat the Market; in Sect. 28.5, we summarize the main conclusions drawn, highlight-
ing the implications of our results in relation to the possibilities of using strategies to
beat the Market according to the CAPM, and, therefore, we set out our conclusions as
to the efficiency of the Market and the validity of the CAPM. In addition, there is a list
of the references on which our research is based and an Appendix which contains, in
greater detail, the statistical evidence gathered during our study.

28.2 Data

Our analysis is based on the 35 securities that comprise the IBEX 35 (the official
index of the Spanish Stock Exchange Market) in each month from practically the
beginning of trading in Continuous Market in Spain (1989) up to March 2007.10 We
chose this index because it includes the 35 most liquid companies in the Spanish
Market and, taking into account that the CAPM in Sharpe-Lintner’s version is valid
for liquid companies, it would have been difficult to find such companies in Spain
outside the IBEX 35.

10
We have not extended our analysis beyond March 2007 in order to exclude the period of the
global financial crisis which had a very damaging effect on major companies listed on the
IBEX 35, such as banks, as this could have distorted our results.
28 Can We Use the CAPM as an Investment Strategy? 757

The IBEX 35 is a capitalization-weighted index comprising the 35 most liquid

companies which quote in the Continuous Market in Spain. It fulfills all of the
criteria required of an indicator which aspires to be a benchmark for trading: it
is representative (almost 90 % of the trading volume in the Continuous Market and
approximately 80 % of the Market capitalization is held by the 35 firms listed on the
IBEX 35), it can be replicated (ease of replication), its computation is guaranteed, it is
widely publicized and impartial (supervised by independent experts).
The selection criteria for securities listed on the IBEX 35 are as follows:
• Securities must be included in the SIBE trading system.
• Stocks must be representative with a big Market capitalization and trading
volume.
• There must be a number of “floating” stocks which is sufficient to ensure that the
index’s Market capitalization is sufficiently widespread and allows for hedge and
arbitrage strategies in the Market for derivatives on the IBEX 35.
• The average Market capitalization of a stock measurable on the index11 must be
greater than 0.30 % of the average capitalization of the index during the
monitoring period.12
• The stock must have been traded in at least 1/3 of the sessions during the
monitoring period or be among the top 15 stocks measured in terms of Market
capitalization.
The IBEX 35 is revised on a half-yearly basis in terms of its composition and the
number of stocks considered of each firm; nevertheless, if financial operations are
carried out which affect significantly any of the listed stocks, it can be adjusted
accordingly. In general, the index is adjusted when there are increases in capital with
preemptive rights, extraordinary dividend distributions, stocks integration as a result of
increases in capital excluding preemptive rights, reductions in capital due to stocks
redemption, capital reductions against own funds with distribution of the value to the
shareholders (this is not the payment of an ordinary dividend), as well as mergers,
takeovers, and demergers. Special adjustments of the index are often carried out.
In the 6-monthly selection of the 35 most liquid stocks, there is no minimum
or maximum number of adjustments made with regard to the previous period.
No changes at all may be required, or as many adjustments as necessary may
be made, based on the results obtained when measuring the liquidity.
Having described the context in which we intend to carry out our study, we will
propose a series of strategies that will enable us to test the validity of the CAPM and
the degree to which the Market is efficient. For this we will take a hypothetical
investor who, each month, examines the 35 stocks of the IBEX 35 listed at that
moment and for which there are at least 36 monthly data available prior to that
moment.

11
The average Market capitalization of a stock in the Index is the arithmetic average of the result
we obtain when multiplying the stocks allowed to be traded in each session during the monitoring
period by the closing price of the stock in each of those sessions.
12
The monitoring period is the 6-month period ending prior to each ordinary meeting of the
Commission.
758 F. Gómez-Bezares et al.

The data on returns for all the stocks were obtained from Bloomberg.
In Table 28.1, we show the annual descriptive statistics from our database.
There are two panels in Table 28.1: in panel A we show the descriptive statistics
for the monthly returns of the previous period (December 1989 to November 1992).
In this period, the index did not exist as such, so we have looked at the return
of those stocks which, being included in the IBEX 35 at the beginning of
the contrasting period, provide data in the corresponding month as they then were
quoted in the Continuous Market. In panel B, we bring together the descriptive

Table 28.1 Annual descriptive statistics for the monthly returns on the stocks
PANEL A: PRECEDING PERIOD

YEAR MEAN MEDIAN MAXIMUM MINIMUM St. Dev.

DEC 1989/NOV 1990 0.187118 −0.002995 11.17 − 0.359280 1,195

DEC 1990/NOV 1991 0.010761 0.0000 0.470758 −0.288640 0.095169
DEC 1991/NOV 1992 −0.013669 −0.011575 0.448602 −0.375127 0.113660
DEC 1989/NOV 1992 0.054038 -0.004545 11.17 -0.375127 0.661538

PANEL B: CONTRASTING PERIOD

YEAR MEAN MEDIAN MAXIMUM MINIMUM St. Dev.

DEC 1992/NOV 1993 0.035059 0.028576 0.343220 −0.237244 0.083789

DEC 1993/NOV 1994 0.009628 −0.000857 0.510076 −0.604553 0.101984
DEC 1994/NOV 1995 0.006480 0.002051 0.554368 −0.220966 0.077322
DEC 1995/NOV 1996 0.022723 0.022042 0.204344 −0.164185 0.062786
DEC 1996/NOV 1997 0.038109 0.034819 0.429381 −0.308638 0.097809
DEC 1997/NOV 1998 0.027631 0.028225 0.432660 −0.331141 0.120737
DEC 1998/NOV 1999 −0.001611 −0.003229 0.351075 −0.223282 0.080742
DEC 1999/NOV 2000 0.002103 −0.005529 0.556553 −0.357094 0.110788
DEC 2000/NOV 2001 0.003081 0.000000 0.468603 −0.282024 0.098028
DEC 2001/NOV 2002 −0.005393 −0.002908 0.369085 −0.403408 0.104471
DEC 2002/NOV 2003 0.013977 0.016353 0.551835 −0.347280 0.083299
DEC 2003/NOV 2004 0.018440 0.015852 0.297322 −0.147819 0.050610
DEC 2004/NOV 2005 0.023066 0.014685 0.257841 −0.110474 0.057863
DEC 2005/NOV 2006 0.029634 0.022335 0.372409 −0.169830 0.062640
DEC 2006/MAR 2007 0.021989 0.013010 0.264933 −0.285306 0.068472
DEC 1992/MAR 2007 0.016678 0.012686 0.556553 -0.604553 0.087537

Table 28.1 reports the annual descriptive statistics for our database. In Panel A, we show the
figures for the monthly returns for the previous period and in Panel B the figures for monthly
returns for the contrasting period
For Panel B, we consider only the returns on stocks which we have included in our study, that is
to say, those stocks which were part of the IBEX 35 at a given time and which had a track record of
monthly returns of at least 36 months. For panel A, as the index as such did not yet exist, we
include those stocks listed on the IBEX 35 at the beginning of the contrasting period, for which
there is data in the corresponding previous month; for example, the stocks we included for the
month of December 1989 are those which were quoted on the IBEX 35 in December 1992 but
which also were quoted in the Continuous Market back in December 1989. For the subsequent
months the number of stocks considered in the study rises consistently since the number of stocks
quoted continued to grow
28 Can We Use the CAPM as an Investment Strategy? 759

data for the monthly returns during the contrasting period, that is, it includes the
returns on the stocks which are included in our study (namely, the stocks which
comprised the IBEX 35 at each moment and for which there were at least 36 prior
quotations).
We built the Market portfolio,13 and as a risk-free asset we took the monthly
return on the 1-month Treasury Bills.

28.3 Methods

The aim of our study is to check whether it is possible to obtain abnormal returns
using the CAPM, that is to say, whether the returns derived from the use of the
model are greater than those which could be expected according to the degree of
systematic risk undertaken.
To this end, we analyzed the period December 1989 to March 2007. The study
focuses on the 35 securities comprising the IBEX 35 at any given moment during
the said period.
From this starting point, we propose two possible strategies for testing the efficient
Market hypothesis and the validity of the CAPM: the first strategy assumes an
individual who adjusts his portfolio at the end of each month, selling those stocks he
bought at the end of the previous month and buying those he considers to be
undervalued according to the CAPM; the second strategy is similar to the first, but
with the distinction that in this case the investor does not buy all of the undervalued
stocks but just 75 % of them, namely, those which are furthest removed from
the SML. We use two methods to calculate which securities are the most undervalued:
Jensen’s ratio and Treynor’s ratio. The reason for using a second method is in response
to the critique by Modigliani (1997) of Jensen’s alpha, in which she argues that
differences in return cannot be compared when the risks are significantly different.
To conduct our analysis, we begin by dividing the overall period analyzed
in our study (December 1989 to March 2007) into two subperiods: the first
(December 1989 to November 1992) allows us to calculate the betas of the
model at the beginning of the following (we carried out this calculation with the
36 previous month data). The second subperiod (December 199214 to March 2007)
allows us to carry out the contrast.

13
This has been constructed as a simple average of returns for the stocks comprising the
IBEX 35 each month, corrected for dividend distributions, splits, etc. Of course, in this case, it
is not necessary that there be at least 36 months of prior quotations for the stocks to be included in
the portfolio.
14
Initially, we intended to let the beginning of the second subperiod be the launch date of the
IBEX 35 (January 1992), however, given the requirement that we set ourselves of a prior 36-month
period to compute the betas of the model, we had to move this forward to December. Likewise, it
was originally our intention that the beginning of the first subperiod be the launch date of the
Continuous Market (April 1989) but most of the securities included in our study were listed for the
first time in December 1989, so data could only be obtained from that date onward.
760 F. Gómez-Bezares et al.

So our hypothetical investor will firstly, at the end of each month, observe the
stocks that comprise the IBEX 3515 and will buy those which are undervalued
(in our second analysis method he will buy just 75 % of the stocks which are the
most undervalued). A stock is undervalued if its return is higher than what is ought
to be based on the CAPM.
To obtain this we compute the monthly betas for each stock (bi) during the
contrasting period by regressing the monthly returns on each stock16 on the returns
on the Market portfolio17 in the 36 months immediately prior to this. Then, we
calculate the mean monthly returns on stock i (Ri ), on the Market portfolio (RM )
and on the risk-free asset (RF ) during the same period as a simple average of the
corresponding 36 monthly returns. From all of the foregoing we can compute the
mean monthly return (Ri ), which, according to the CAPM, the stock ought to have
gained during this 36-month period,


Ri ¼ RF þ RM  RF bi (28.1)

We then compare it with the actual return gained (Ri ) to determine whether the
stock is undervalued.
Once we have determined which stocks are undervalued, according to our first
strategy the investor buys all of the undervalued stocks each month,18 while with
the second strategy, he disregards the first quartile and just buys 75 % of the most
undervalued stocks. The quartiles are assembled based on either Jensen’s alphas or
Treynor’s ratio.19
The next step consists in evaluating how well the CAPM functions; if the
undervalued stocks continue to be undervalued indefinitely, we would expect any

15
And for which there is a minimum of 36-monthly data prior to the month analyzed so that the
beta can be calculated. In the case of companies which have recently merged, the beta is calculated
by looking at the weighted average monthly returns on the companies prior to the merger. The
weighting is proportional to the relative importance of each company involved in the merger.
16
We look at the return obtained by our investor from capital gains, dividends, and the sale of
preemptive rights. The returns are also adjusted to take splits into account.
17
The return on the Market portfolio is calculated based on a simple average of the returns gained
by the stocks which comprise the IBEX 35 in each month. The reason we decided to build our own
portfolio rather than relying on a stock Market index such as the IBEX 35 itself is that we wanted
to obtain an index corrected by capital gains, dividends, splits, and preemptive rights. Moreover,
an equally weighted portfolio is theoretically superior.
18
Overvalued stocks are not included in our analysis as we decided to exclude short selling.
19
In this second variant, although the quartiles are built based on Treynor’s ratio, we continue to
use Jensen’s index to work out whether or not a stock is undervalued. This is because of the
problems we have encountered when determining which stocks are undervalued using Treynor’s
ratio. These are related to the return premiums and the negative betas which appear in some cases
in our database. Naturally, we take on board the problems of using Treynor’s ratio to construct the
quartiles, as there could be stocks with positive alphas and low Treynor ratios which would
exclude such stocks from the analysis. This is due to the form of Treynor’s ratio as a quotient.
28 Can We Use the CAPM as an Investment Strategy? 761

portfolio assembled with them to beat the Market in a given month, but this will not
occur if in that month the CAPM operates perfectly.
Jensen’s index allows us to determine whether we have been able to beat the
Market, in which case it must yield a positive result:
 
ap ¼ Rp  RF  bp ðRM  RF Þ (28.2)

where,
bp is the beta for the portfolio, calculated as a simple average of the individual
betas20 of the stocks included in the portfolio (all of the undervalued stocks, or
alternatively the top 75 % thereof as we have outlined above).
Rp is the return on the portfolio and is calculated as the simple average of the
individual returns obtained each month for the stocks that comprise it.
RM and RF are the monthly returns on the Market portfolio and the risk-free asset
respectively, for the month in question.
ap is Jensen’s alpha for the portfolio p.
The Z-statistic, which follows an approximately normal distribution, (0.1),
allows us to determine whether the number of months in which our investor beats
the Market is due to chance:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ¼ ðY  npÞ= npð1  pÞ (28.3)

Y indicates the number of periods in which the portfolio comprising undervalued

stocks beats the Market, n represents the number of months analyzed and to p we
give a value of 0.5 as this is the probability of beating the Market if the CAPM and
the efficient Market hypothesis are fulfilled,21 and we want to find out if the
difference (Y – np) is due to chance.
A value of |Z| > 1.96 would lead us to reject the null hypothesis of a difference
due to chance, with a significance level of 5 %.
Moreover, in order to make our results more robust, we developed a series of
statistical tests which are presented in Appendix. We carried out a mean test for
each of the proposed strategies to ascertain whether we beat the Market or not, both
parametric – which means normality – as well as nonparametric (Wilcoxon’s test);
in addition we tested the hypothesis that we could beat the Market 50 % of the time,
and for this we used a binomial test. We also analyzed, on a monthly and on a yearly
basis, whether or not there were differences between the performance means
achieved in each month/year, using the ANOVA, a nonparametric test (Kruskal
and Wallis), and also the technique of contingency tables and the chi-square

20
Which have been calculated with the 36 previous months’ data.
21
If the CAPM and the efficient Market hypothesis are fulfilled exactly, the results of any portfolio
(adjusted for risk) should match, as an average, with those of the Market. Here we suppose that, by
pure chance, for any given month, they will end up 50 % above and 50 % below the Market
average.
762 F. Gómez-Bezares et al.

Pearson test together with the likelihood ratio. Finally, we conducted regression
analysis of the performance measurement as a dependent variable and the time
dimension (months) as an independent variable to see whether the performance
improved or worsened over time. These tests allow us to confirm our results, which
we shall see in the next section.

28.4 Results

28.4.1 Strategy 1: The Investor Buys All the Undervalued Stocks

In this first strategy, the investor observes, in a given month, the track record (for
the previous 36 months) of the 35 stocks which comprise the IBEX 35 at the
moment in question,22 and buys all those that are undervalued according to Jensen’s
index. The next step is to check whether he has managed to beat the Market with the
portfolio of undervalued stocks.
Table 28.2 shows the results of this strategy. We show for each of the 14 years in
our study and for the other 4 months the number of months (and the percentage) in
which the investor’s portfolio has beaten the Market. In addition, we show this
result for the whole period. Furthermore, and also with respect to the whole period,
we show the result for the Z-statistic.
As we can see in Table 28.2, only in 92 of the 172 months analyzed in our study
is the Market beaten, which is equal to 53.49 %23; moreover, the figure for
Z-statistic (which is below 1.96) confirms that the CAPM does not offer
a significantly better strategy than investing in the Market portfolio. In other
words, we could say that the success of the investor (in the months in which he
succeeds in beating the Market) could be due to chance. This conclusion is further
confirmed in Appendix by the mean tests which allow us to accept that the mean of
Jensen’s alphas is zero and by the binomial test which allows accepting a success
probability rate of 50 %.
If we now focus on the analysis of each year in our time frame, we can see that in
only two of the 14 years covered in our study, to be precise from December 2004 to
November 2006, the CAPM proves to be a useful tool for the investor allowing him
to use a strategy to beat the Market; in fact, in 75 % and 83.33 % of the months of
each of those 2 years our investor beats the Market, thus confirming the poor
performance of the model, or the inefficiency of the Market.

22
Which have a track record of at least 36 months in the Market. If, in a given month, there is
a stock which is listed on the IBEX35 but does not have at least 36 months of prior data, this will be
included in our study in the month in which it reaches this figure, unless by that stage it has already
been deleted from the IBEX35.
23
We have not included the transaction costs, which undoubtedly would be higher for the managed
portfolio and, therefore, would make that strategy less attractive.
28 Can We Use the CAPM as an Investment Strategy? 763

Table 28.2 Results of the Strategy 1 (buying all of the undervalued stocks)

YEAR No. of successful months % success

DEC 1992/NOV 1993 4 33.33%

DEC 1993/NOV 1994 4 33.33%
DEC 1994/NOV 1995 5 41.67%
DEC 1995/NOV 1996 7 58.33%
DEC 1996/NOV 1997 7 58.33%
DEC 1997/NOV 1998 6 50%
DEC 1998/NOV 1999 7 58.33%
DEC 1999/NOV 2000 6 50%
DEC 2000/NOV 2001 6 50%
DEC 2001/NOV 2002 5 41.67%
DEC 2002/NOV 2003 7 58.33%
DEC 2003/NOV 2004 7 58.33%
DEC 2004/NOV 2005 9 75%
DEC 2005/NOV 2006 10 83.33%
DEC 2006/MAR 2007 2 50% Z-statistic
DEC 1992/MAR 2007 92 53.49% 0.91
In Table 28.2, we bring together the results of a strategy in which our hypothetical investor buys all
the undervalued stocks on the Index. To be specific, we provide for each of the 14 years in our
study as well as for the other 4 months, the number of months (and the percentage) in which the
portfolio beats the Market. We also show this result for the period as a whole. Moreover, for the
period as a whole, we show the result for the Z-statistic

In the first 2 years (from December 1992 to November 1994), the opposite
happens: in just 33.33 % of the months the investor’s strategy beats the Market.
In those months it seems, therefore, that he could use the opposite strategy to beat it.
The statistical tests carried out and included in the Appendix do not provide us
with clear conclusions as to the basis for the differences in performance of the
investor’s strategy between the different months and years, because the different
tests used gave us different results. However, the regression analysis does confirm
that the performance of the portfolio improves over time, as we obtain a positive
and significant beta.

28.4.2 Strategy 2: The Investor Buys the Three Stock-Quartiles with

the Highest Degree of Undervaluation

With this strategy, which has two variants, what lets us find out whether a stock is or
is not undervalued is, just as with the previous strategy, the difference between the
mean monthly return actually achieved by the stock in a given month (using
information on the previous 36 months) and that which it ought to have achieved
according to the CAPM. However, with this second strategy, the investor disregards
the first quartile of undervalued stocks and buys the other 75 %, those that are the
most undervalued.
764 F. Gómez-Bezares et al.

In the 1st variant of this second strategy, the ranking of the stocks (in each
month) by quartiles is done using Jensen’s index and in the 2nd variant it is done
using Treynor’s ratio.

28.4.2.1 Strategy 2, 1st Variant: The Quartiles Are Built Based on

Jensen’s Index
In Table 28.3, we show the results of this strategy:
The results, although slightly better, are not very different from those
achieved by the previous strategy, as once again the CAPM proves not to be
a useful tool for beating the Market. In only 94 of the 172 months in our study
(54.65 %) does the strategy manage to beat the Market. The value for Z-statistic
again confirms this result. And again, the mean tests show that the mean for
the alphas for the investor’s portfolio can be zero; while the binomial test
shows that the probability of success in beating the Market can be 50 %
(see Appendix).
There is also a slight improvement in the results by years in comparison with the
results achieved by the previous strategy; the CAPM proves to be a useful Market-
beating tool in 5 of the 14 years analyzed (December 1995 to November 1996,
December 1998 to November 1999, December 2002 to November 2003, December
2004 to November 2005, and December 2005 to November 2006), which implies
that the model either works badly or that the Market is inefficient since it is possible
to beat it in most months.

Table 28.3 Results of Strategy 2, 1st Variant (buying the top three quartiles of the most
undervalued stocks, constructed based on Jensen’s Index)

YEAR No. of successful months % success

DEC 1992/NOV 1993 3 25%

DEC 1993/NOV 1994 4 33.33%
DEC 1994/NOV 1995 5 41.67%
DEC 1995/NOV 1996 8 66.67%
DEC 1996/NOV 1997 6 50%
DEC 1997/NOV 1998 7 58.33%
DEC 1998/NOV 1999 8 66.67%
DEC 1999/NOV 2000 6 50%
DEC 2000/NOV 2001 5 41.67%
DEC 2001/NOV 2002 5 41.67%
DEC 2002/NOV 2003 9 75%
DEC 2003/NOV 2004 6 50%
DEC 2004/NOV 2005 10 83.33%
DEC 2005/NOV 2006 10 83.33%
DEC 2006/MAR 2007 2 50% Z-statistic
DEC 1992/MAR 2007 94 54.65% 1.22
Table 28.3 shows the results of the strategy in which the investor disregards the first quartile of
undervalued stocks and buys the remaining 75 %, which are the stocks that are most undervalued.
The quartiles are built using Jensen’s Index
28 Can We Use the CAPM as an Investment Strategy? 765

The opposite results are obtained for the first 2 years (from December 1992 to
November 1994), in which the percentages of success (25 % and 33.33 %, respec-
tively) are the lowest.
The ANOVA, the Kruskal–Wallis test, and the contingency table with Pearson’s
chi-square-statistic and the likelihood ratio all lead us to the conclusion24 that there
are no differences between the performance means achieved by the portfolio in the
months in our analysis, so the small differences could be due to mere chance.
However, these tests do not allow us to offer conclusions when we look at the years
that comprise our sample, as different conclusions can be drawn from the different
tests used.
Moreover, the regression analysis once again allows us to confirm that, as with
the previous strategy, the performance of the portfolio improves over time.

28.4.2.2 Strategy 2, 2nd Variant: The Quartiles Are Constructed

Based on Treynor’s Ratio
In Table 28.4, we show the results of this strategy.
The conclusions are very similar to those which the two previous strategies lead
us to, thus we can see, for the whole period, a scant success rate for the strategy.

Table 28.4 Results of Strategy 2, 2nd Variant (buying the top three quartiles of the most
undervalued stocks, constructed based on Treynor’s Ratio)

YEAR No. of successful months % success

DEC 1992/NOV 1993 3 25%

DEC 1993/NOV 1994 4 33.33%
DEC 1994/NOV 1995 7 58.33%
DEC 1995/NOV 1996 7 58.33%
DEC 1996/NOV 1997 6 50%
DEC 1997/NOV 1998 7 58.33%
DEC 1998/NOV 1999 7 58.33%
DEC 1999/NOV 2000 7 58.33%
DEC 2000/NOV 2001 4 33.33%
DEC 2001/NOV 2002 5 41.67%
DEC 2002/NOV 2003 8 66.67%
DEC 2003/NOV 2004 6 50%
DEC 2004/NOV 2005 10 83.33%
DEC 2005/NOV 2006 10 83.33%
DEC 2006/MAR 2007 2 50% Z-statistic
DEC 1992/MAR 2007 93 54.07% 1.07
Table 28.4 shows the results of the strategy in which the investor disregards the first quartile of
undervalued stocks and buys the remaining 75 %, which are the stocks that are most undervalued.
The quartiles are built using Treynor’s Ratio

24
See Appendix.
766 F. Gómez-Bezares et al.

Moreover, the value for Z-statistic once again leads us to accept the hypothesis that
any possible success or failure is due to chance. Meanwhile, the tests we include
in Appendix confirm the previous results: the mean tests support the
interpretation that the mean performance of the portfolios is zero, therefore we do
not beat the Market, and the binomial test suggests that the strategy’s success rate
is 50 %.
Focusing now on the analysis by years, there are 3 years (December 2002 to
November 2003 and December 2004 to November 2006) in which we can confirm
that the Market is not efficient or the model does not work well, since in those years
the CAPM seems to be a useful Market-beating tool. However, in the first 2 years of
our study as well as in the ninth year, the lowest success rates are delivered, which
does not allow us to confirm the efficient Market hypothesis, as by using the
opposite strategy we could have beaten it.
With regard to the statistical tests to analyze the robustness of the above-
mentioned results for the various years and months in our database and those
shown in Appendix, we find that there are no significant differences between the
performance means achieved by the portfolio in the different months, so any
difference can be due to chance. Nonetheless, we cannot draw conclusions on
a yearly basis as the tests produce different results.
Finally, once again the result of the previous strategies is confirmed: the perfor-
mance of the investor’s portfolio improves over time. Overall, simply looking at the
graphics for the three strategies where we can see the Jensen alphas achieved by our
investor over time, if the reader disregards the initial data, he will see that the
adjusted line is flat.

28.5 Conclusion

The aim of our study was to analyze whether the design of strategies based on
the CAPM can enable an investor to obtain abnormal returns. We also set out
to explore this using a methodology that was both intuitive and scientifically
rigorous.
We also set out to determine whether the efficient Market hypothesis
was fulfilled and whether we could accept the validity of the CAPM, since if
strategies that can beat the Market with a degree of ease exist, we cannot
confirm either the efficiency of the Market or the workability of the CAPM, as
this defends that the only way to obtain extra returns is to accept a higher degree
of systematic risk. Conversely, we would then be in possession of a tool enabling
the investor to achieve higher returns than the Market for a given level of
systematic risk.
Analyzing the behavior of an investor who can buy the stocks comprising the
IBEX 35 at any given moment, that is, the benchmark for the Spanish stock
28 Can We Use the CAPM as an Investment Strategy? 767

Market, and who reconfigures his portfolio on a monthly basis, we find that,
regardless of the strategy used (buying all of the undervalued stocks or buying
the 75 % most undervalued stocks, either measured with Jensen’s alpha or with
Treynor’s ratio), the investor manages to beat the Market about 50 % of the time.
We opted to exclude from our calculations the transaction costs, so we in fact
overvalued the investor’s results. Therefore, we can conclude that the CAPM is
not an attractive tool for an investor who wishes to achieve abnormal returns.
It seems that undervalued stocks rapidly fit the SML, and so from another
perspective, we can confirm the efficient Market hypothesis and the workability
of the CAPM. These conclusions are backed up by a series of statistical tests
included in Appendix.
A positive aspect of our study from the point of view of its applicability is that
the behavior we have envisaged for our virtual investor is perfectly replicable as
he acts only with the information available in any given month. Meanwhile, by
focusing on the stocks that make up the IBEX 35, our study’s conclusions are
applicable to the most representative and consolidated stocks on the Spanish
Market.
Furthermore, our system possesses interesting statistical advantages: it allows
for variation of the beta and of the risk premium over time, it avoids the risk of
illiquidity, and it reduces errors in measurement. We have also considered its
robustness.
Finally, it is important to point out that at all times we have accepted implicitly
the logic of the CAPM, when measuring performance with Jensen’s index (which
implies basing ourselves on the CAPM). For this reason our results enable us to
confirm that it is the efficiency of the Market which prevents abnormally high
returns compared to those the Market itself is able to achieve, and that the CAPM
is a model that works reasonably well; hence it cannot be used as a Market-
beating tool.

Appendix

In this section, we report a series of statistical tests done using the “Stata” program,
which support the results obtained in our study, thereby, we believe, making it more
robust.25 These statistics are drawn up for each of the strategies we tried out: buying
all of the undervalued stocks, buying the top 75 % most undervalued stocks using
Jensen’s Index to select them, and buying the 75 % most undervalued stocks
according to Treynor’s Ratio.

25
See Agresti (2007), Anderson et al. (2011), Conover (1999) and Newbold et al. (2009) for
a widening of the statistical processing used in this Appendix.
768 F. Gómez-Bezares et al.

Strategy 1: Investor Buys All Undervalued Stocks

1. Summary statistics
We work here, as for all of the strategies, with the Jensen values recorded for each
of the portfolios (172).
Percentiles Smallest
1% −0.1064671 −0.1087249
5% −0.0279054 −0.1064671
10% −0.0180511 −0.0708467 Obs. 172
25% −0.0077398 −0.0540319 Sum of Wgt. 172
50% 0.0012046 Mean 0.001246
Largest Std. Dev. 0.0213603
75% 0.0123422 0.0389192
90% 0.0237416 0.0444089 Variance 0.0004563
95% 0.0309298 0.0509601 Skewness −1.397642
99% 0.0509601 0.0773151 Kurtosis 10.85719

We would point out that in this table the Jensen values are far from normality, as
can be seen in the readings for asymmetry and kurtosis.
2. Mean test

Variable Obs. Mean Std. Err. Std. Dev. [95% Conf. Interval]
jensen 172 0.001246 0.0016287 0.0213603 −0.001969 0.004461

mean = mean (jensen) t = 0.7650

Ho: mean = 0 degrees of freedom = 171

Ha: mean < 0 Ha: mean != 0 Ha: mean > 0

Pr(T < t) = 0.7773 Pr(|T| > |t|) = 0.4453 Pr(T > t) = 0.2227

From this mean test, we can accept that the mean for the Jensen alphas is zero; in fact,
if we focus on the two-sided test we obtain a probability of 0.4453, higher than 5 %.
This allows us to conclude that we are not able to beat the Market with this strategy.
3. Nonparametric mean test

sign Obs. sum ranks expected

positive 92 8451 7439
negative 80 6427 7439
zero 0 0 0
all 172 14878 14878
unadjusted variance 427742.50
adjustment for ties 0.00
adjustment for zeros 0.00
--------------
adjusted variance 427742.50
Ho: jensen = 0
z = 1.547
Prob > |z| = 0.1218
28 Can We Use the CAPM as an Investment Strategy? 769

We can see again that Wilcoxon’s nonparametric test leads to the same conclu-
sion: we cannot beat the Market with this strategy.
4. Frequencies
BEAT
BEAT
NO YES Total
80 0 80
NO
46.51 0 46.51
0 92 92
YES
0 53.49 53.49
80 92 172
Total
46.51 53.49 100

From this contingency table, we see that the proportion of months in which we
beat the Market is very similar to the proportion in which we fail to do so.
5. Probabilities test
Variable N Observed K Expected K Assumed p Observed p
Jensen 172 92 86 0.50000 0.53488

Pr(k >= 92) = 0.200842 (one-sided test)

Pr(k <= 92) = 0.839213 (one-sided test)
Pr(k <= 80 or k >= 92) = 0.401684 (two-sided test)

We use this binomial test to determine whether the probability of beating the Market
is similar to the probability of failing to beat it (50–50). If we look at the two-sided test,
we obtain a probability of 0.401684, above 5 %, and so we accept the success rate is
around 50 %, which supports the results obtained with the contingency table.
6. Fit of Jensen by month

.1

.05
jensen

−.05

−.1

1992m1 1994m1 1996m1 1998m1 2000m1 2002m1 2004m1 2006m1 2008m1

time
770 F. Gómez-Bezares et al.

7. Linear fit

Source SS df MS Number of obs. 172

Model 0.00459285 1 0.00459285 F( 1, 170) 10.63
Residual 0.07342831 170 0.00043193 Prob > F 0.0013
Total 0.07802116 171 0.00045626 R-squared 0.0589
Adj. R-squared 0.0533
Root MSE 0.02078

jensen Coef. Std. Err. t P>|t| [95% Conf. Interval]

time 0.0001041 0.0000319 3.26 0.001 0.0000411 0.0001671
_cons −0.048762 0.0154174 −3.16 0.002 −0.0791963 −0.0183277

From the above table we can observe that the regression slope which
connects Jensen’s index (the dependent variable) and time measured in months
(the independent variable) gives a positive and significant coefficient for
beta which allows us to conclude that this strategy gives results that improve
over time.
8. One-way analysis of Jensen by month

.1

.05

−.05

−.1
y

ry

ch

il

ay

er

r
us

be

be

be
r

l
ar

Ju
Ap
ua

ob
M
ar

Ju

g
nu

em

em

em
Au
M
br

ct
Ja

ov
pt

ec
Fe

Se

D
28 Can We Use the CAPM as an Investment Strategy? 771

9. One-way ANOVA (by month)

Source SS df MS Number of obs 172

Model 0.00607108 11 0.00055192 F( 11, 160) 1.23
Residual 0.07195009 160 0.00044969 Prob > F 0.2729
Total 0.07802116 171 0.00045626 R-squared 0.0778
Adj R-squared 0.0144
Root MSE 0.02121

jensen Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0084177 0.0056675 1.49 0.139 −0.0027751 0.0196104
1 −0.0094303 0.0080151 −1.18 0.241 −0.0252592 0.0063987
2 −0.0142236 0.0080151 −1.77 0.078 −0.0300526 0.0016053
3 0.0040468 0.0078803 0.51 0.608 −0.0115161 0.0196097
4 −0.0021065 0.0078803 −0.27 0.79 −0.0176694 0.0134564
5 −0.0167755 0.0078803 −2.13 0.035 −0.0323384 −0.0012126
6 −0.0049838 0.0080151 −0.62 0.535 −0.0208128 0.0108451
7 −0.0026922 0.0080151 −0.34 0.737 −0.0185211 0.0131368
8 −0.0082308 0.0078803 −1.04 0.298 −0.0237937 0.0073321
9 −0.0112643 0.0080151 −1.41 0.162 −0.0270932 0.0045647
10 −0.0092672 0.0080151 −1.16 0.249 −0.0250961 0.0065618
11 −0.0115344 0.0080151 −1.44 0.152 −0.0273633 0.0042946
12 (dropped)

We carry out an ANOVA to determine whether the means for Jensen’s indices
obtained in the different months are uniform. We can see that F has a value of 1.23
and a related probability of 0.2729, above 5 %, which leads us to accept that the
means for Jensen’s indices are uniform throughout the months included in our study.
10. Kruskal-Wallis tests (Rank sums), by month
Month Obs. Rank Sum
January 15 949
February 15 1551
March 15 1310
April 14 1217
May 14 1093
June 14 1386
July 14 1270
August 14 798
September 14 1382
October 14 1025
November 14 1032
December 15 1865
chi-squared = 22.716 with 11 d.f.
probability = 0.0194

We use the Kruskal–Wallis nonparametric test to measure the same phenome-

non as with the ANOVA, and we see that for a significance level of 1 % we would
come to the same conclusion as we did with the ANOVA, while for a significance
level of 5 %, we can rule out uniformity from month to month.
772 F. Gómez-Bezares et al.

11. Contingency table (by month)

BEAT
Month
NO YES Total
January 11 4 15
February 5 10 15
March 7 8 15
April 6 8 14
May 7 7 14
June 2 12 14
July 6 8 14
August 11 3 14
September 6 8 14
October 8 6 14
November 9 5 14
December 2 13 15
Total 80 92 172
Pearson chi2(11) = 26.3578 Pr = 0.006
likelihood-ratio chi2(11) = 28.4294 Pr = 0.003

We use this contingency table to test the same point as with the ANOVA and the
Kruskal–Wallis test above, to determine whether there are any differences between
the months, and we find that both with Pearson’s chi-square test and with the
likelihood ratio the probability is less than 1 %, which leads us to conclude that
the chances of beating the Market are not the same month on month, so that there
will be some months in which it is easier to do so than others.
12. One-way analysis of Jensen by year

.1

.05

−.05

−.1
93

94

95

96

97

98

99

00

01

02

03

04

05

06
19

19

19

19

19

19

19

20

20

20

20

20

20

20
28 Can We Use the CAPM as an Investment Strategy? 773

13. One-way ANOVA (by year)

Source SS df MS Number of obs 168
Model 0.01216506 13 0.00093577 F(13, 154) 2.2
Residual 0.06538048 154 0.00042455 Prob > F 0.0117
Total 0.07754554 167 0.00046435 R-squared 0.1569
Adj R-squared 0.0857
Root MSE 0.0206

jensen Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0113553 0.005948 1.91 0.058 −0.000395 0.0231055
1 −0.0368126 0.0084118 −4.38 0 −0.0534299 −0.0201952
2 −0.0159892 0.0084118 −1.9 0.059 −0.0326065 0.0006282
3 −0.0088588 0.0084118 −1.05 0.294 −0.0254762 0.0077586
4 −0.0077928 0.0084118 −0.93 0.356 −0.0244102 0.0088246
5 −0.0082156 0.0084118 −0.98 0.33 −0.024833 0.0084018
6 −0.0062178 0.0084118 −0.74 0.461 −0.0228352 0.0103996
7 −0.0102373 0.0084118 −1.22 0.225 −0.0268546 0.0063801
8 −0.0061693 0.0084118 −0.73 0.464 −0.0227866 0.0104481
9 −0.0108019 0.0084118 −1.28 0.201 −0.0274192 0.0058155
10 −0.0127709 0.0084118 −1.52 0.131 −0.0293883 0.0038465
11 −0.009494 0.0084118 −1.13 0.261 −0.0261114 0.0071234
12 −0.0078177 0.0084118 −0.93 0.354 −0.0244351 0.0087997
13 0.0003579 0.0084118 0.04 0.966 −0.0162595 0.0169752
14 (dropped)

We performed an ANOVA based on years, to see whether the strategy gives

similar Jensen values from year to year for those years included in our study
(we excluded 1992 and 2007 because they could provide only 1 and 3 months of
data, respectively. These years are also excluded from the other tests done based on
years). We obtained a figure F of 2.20 with a related probability of 0.0117, thus for
a significance level of 1 % we would accept that the means are uniform from one
year to another, however, for a significance level of 5 % we must rule out this
hypothesis and conclude that the Jensen values differ from one year to another.
14. Kruskal-Wallis tests (Rank sums), by year
Year Obs. Rank Sum
1993 12 722
1994 12 661
1995 12 1017
1996 12 1092
1997 12 1024
1998 12 1083
1999 12 993
2000 12 1002
2001 12 931
2002 12 906
2003 12 1019
2004 12 1058
2005 12 1275
2006 12 1413
chi-squared = 16.527 with 13 d.f.
probability = 0.2218
774 F. Gómez-Bezares et al.

We now use the Kruskal–Wallis test to find out the same point as we did with the
ANOVA. We see that we obtain a probability greater than 5 % which means we can
accept that the mean Jensen values are uniform from one year to another.
15. Contingency table (by year)

BEAT
Year
NO YES Total
1993 8 4 12
1994 9 3 12
1995 6 6 12
1996 5 7 12
1997 5 7 12
1998 6 6 12
1999 5 7 12
2000 6 6 12
2001 6 6 12
2002 7 5 12
2003 5 7 12
2004 5 7 12
2005 4 8 12
2006 1 11 12
Total 78 90 168
Pearson chi2(13) = 15.2205 Pr = 0.294
likelihood-ratio chi2(13) = 16.7608 Pr = 0.210

The contingency table confirms the result of the previous test, namely, that there
are no differences between the Jensen values for the different years so that any
small difference can be due to chance.

Strategy 2: The Investor Buys the Most Undervalued Stocks

Strategy 2, 1st Variant: Quartiles Constructed Using Jensen’s Index

1. Summary statistics

Percentiles Smallest
1% −0.1414719 −0.1486773
5% −0.0391241 −0.1414719
10% −0.0192669 −0.1117398 Obs. 172
25% −0.0103728 −0.0683232 Sum of Wgt. 172
50% 0.0017188 Mean 0.0011871
Largest Std. Dev. 0.0280576
75% 0.0154674 0.0545513
90% 0.0294085 0.0690987 Variance 0.0007872
95% 0.0390722 0.0696097 Skewness −1.842765
99% 0.0696097 0.0745625 Kurtosis 11.929
28 Can We Use the CAPM as an Investment Strategy? 775

We would point out that in this table the Jensen values are far from normality, as
can be seen in the readings for asymmetry and kurtosis.
2. Mean test

Variable Obs. Mean Std. Err. Std. Dev. [95% Conf. Interval]
jensen (Q) 172 0.001187 0.0021394 0.0280576 −0.0030359 0.0054101

mean = mean (jensen) t = 0.5549

Ho: mean = 0 degrees of freedom = 171

Ha: mean < 0 Ha: mean != 0 Ha: mean > 0

Pr(T < t) = 0.7101 Pr(|T| > |t|) = 0.5797 Pr( T > t) = 0.2899

From this mean test we assume that the mean for Jensen’s alphas is zero, in fact,
if we focus on the two-sided test we can detect a probability of 0.5797, higher than
5 %. This allows us to conclude that we are not able to beat the Market with this
strategy.
3. Nonparametric mean test

sign Obs. sum ranks expected

positive 94 8567 7439
negative 78 6311 7439
zero 0 0 0
all 172 14878 14878
unadjusted variance 427742.50
adjustment for ties 0.00
adjustment for zeros 0.00
--------------
adjusted variance 427742.50
Ho: jensen = 0
z = 1.725
Prob > |z| = 0.0846

We can see again, that Wilcoxon’s nonparametric test leads to the same conclu-
sion: we cannot beat the Market with this strategy.
4. Frequencies

BEAT
BEAT
NO YES Total
78 0 78
NO
45.35 0 45.35
0 94 94
YES
0 54.65 54.65
78 94 172
Total
45.35 54.65 100
776 F. Gómez-Bezares et al.

From this contingency table, we see that the proportion of months in which we
beat the Market is very similar to the proportion of months in which we do not, with
a slightly higher probability of beating the Market.
5. Probabilities test

Variable N Observed K Expected K Assumed p Observed p

Jensen (Q) 172 94 86 0.50000 0.54651
Pr(k >= 94) = 0.126331 (one-sided test)
Pr(k <= 94) = 0.902626 (one-sided test)
Pr(k <= 78 or k >= 94) = 0.252662 (two-sided test)

With this binomial test we set out to find whether the probability of beating the
Market is the same as the probability of not beating it; (50–50). If we look at the
two-sided test, we get a probability of 0.252662, above 5 %, hence we can accept
that the success rate is 50 %, which backs up the results obtained with the
contingency table.
6. Fit of Jensen (quartiles) by month

.1

.05

0
qjensen

−.05

−.1

−.15
1992m1 1994m1 1996m1 1998m1 2000m1 2002m1 2004m1 2006m1 2008m1
time
28 Can We Use the CAPM as an Investment Strategy? 777

7. Linear fit

Source SS df MS Number of obs 172

Model 0.00979599 1 0.00979599 F(1, 170) 13.34
Residual 0.1248204 170 0.00073424 Prob > F 0.0003
Total 0.13461639 171 0.00078723 R-squared 0.0728
Adj R-squared 0.0673
Root MSE 0.0271

jensen (Q) Coef. Std. Err. t P>|t| [95% Conf. Interval]

time 0.000152 0.0000416 3.65 0 0.0000699 0.0002341
_cons −0.0718465 0.0201013 −3.57 0 −0.1115268 −0.0321663

From the above table we can observe that the regression slope, which connects
Jensen’s index (the dependent variable) and time measured in months (the inde-
pendent variable), gives a positive and significant coefficient for beta which allows
us to conclude that this strategy gives results that improve over time.
8. One-way analysis of Jensen (quartiles) by month

.1

.05

−.05

−.1

−.15
y

ry

ch

ril

ay

er

r
us

be

be

be
l
ar

Ju
Ap
ua

ob
M
ar

Ju

g
nu

em

em
Au
M
br

ct
te
Ja

ov

ec
Fe

p
Se

D
778 F. Gómez-Bezares et al.

9. One-way ANOVA (by month)

Source SS df MS Number of Obs. 172
Model 0.00735542 11 0.00066868 F(11, 160) 0.84
Residual 0.12726097 160 0.00079538 Prob > F 0.5997
Total 0.13461639 171 0.00078723 R-squared 0.0546
Adj. R-squared −0.0104
Root MSE 0.0282

jensen (Q) Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0105061 0.0075374 1.39 0.165 −0.0043796 0.0253918
1 −0.0142298 0.0106595 −1.33 0.184 −0.0352813 0.0068218
2 −0.015749 0.0106595 −1.48 0.142 −0.0368006 0.0053025
3 0.0028747 0.0104804 0.27 0.784 −0.017823 0.0235724
4 −0.0060897 0.0104804 −0.58 0.562 −0.0267874 0.0146081
5 −0.0172233 0.0104804 −1.64 0.102 −0.037921 0.0034744
6 −0.0069927 0.0106595 −0.66 0.513 −0.0280443 0.0140588
7 −0.0014246 0.0106595 −0.13 0.894 −0.0224761 0.019627
8 −0.014007 0.0104804 −1.34 0.183 −0.0347047 0.0066907
9 −0.0143055 0.0106595 −1.34 0.181 −0.035357 0.0067461
10 −0.0125205 0.0106595 −1.17 0.242 −0.0335721 0.008531
11 −0.0123631 0.0106595 −1.16 0.248 −0.0334147 0.0086884
12 (dropped)

We perform an ANOVA to see whether the means for Jensen values obtained for
the different months are uniform. We find that F-statistic has a value of 0.84 with
a related probability of 0.5997, above 5 %, which leads us to accept that the means
for Jensen values are uniform among the months included in our sample.
10. Kruskal-Wallis tests (Rank sums), by month

Month Obs. Rank Sum

January 15 1167
February 15 1428
March 15 1228
April 14 1105
May 14 1077
June 14 1451
July 14 1216
August 14 885
September 14 1385
October 14 1156
November 14 1020
December 15 1760
chi-squared = 14.369 with 11 d.f.
probability = 0.2133

We now use the Kruskal–Wallis test to find out the same point as we did with the
ANOVA. We come to the same conclusion as we did with the ANOVA.
28 Can We Use the CAPM as an Investment Strategy? 779

11. Contingency table (by month)

BEAT
Month
NO YES Total
January 7 8 15
February 5 10 15
March 7 8 15
April 8 6 14
May 9 5 14
June 3 11 14
July 7 7 14
August 9 5 14
September 6 8 14
October 7 7 14
November 8 6 14
December 2 13 15
Total 78 94 172
Pearson chi2(11) = 16.2331 Pr = 0.133
likelihood-ratio chi2(11) = 17.3939 Pr = 0.097

The contingency table allows us to test the same point as above with the ANOVA
and the Kruskal–Wallis test, and leads to the same conclusions; hence, we can accept
that all months show a similar tendency in terms of beating the Market.
12. One-way analysis of Jensen (quartiles) by year

.1

.05

−.05

−.1

−.15
93

94

95

96

97

98

99

00

01

02

03

04

05

06
19

19

19

19

19

19

19

20

20

20

20

20

20

20
780 F. Gómez-Bezares et al.

13. One-way ANOVA (by year)

Source SS df MS Number of obs. 168

Model 0.02482403 13 0.00190954 F(13, 154) 2.69
Residual 0.10923702 154 0.00070933 Prob > F 0.0019
Total 0.13406105 167 0.00080276 R-squared 0.1852
Adj. R-squared 0.1164
Root MSE 0.02663

jensen (Q) Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0129934 0.0076884 1.69 0.093 −0.0021949 0.0281817
1 −0.0494527 0.010873 −4.55 0 −0.0709322 −0.0279733
2 −0.0191305 0.010873 −1.76 0.08 −0.04061 0.002349
3 −0.00969 0.010873 −0.89 0.374 −0.0311695 0.0117894
4 −0.0092178 0.010873 −0.85 0.398 −0.0306973 0.0122617
5 −0.0100839 0.010873 −0.93 0.355 −0.0315634 0.0113956
6 −0.0066183 0.010873 −0.61 0.544 −0.0280978 0.0148612
7 −0.0095188 0.010873 −0.88 0.383 −0.0309983 0.0119607
8 −0.0095735 0.010873 −0.88 0.38 −0.031053 0.0119059
9 −0.0142031 0.010873 −1.31 0.193 −0.0356825 0.0072764
10 −0.0158029 0.010873 −1.45 0.148 −0.0372824 0.0056765
11 −0.0087823 0.010873 −0.81 0.421 −0.0302617 0.0126972
12 −0.0075224 0.010873 −0.69 0.49 −0.0290019 0.013957
13 0.0071267 0.010873 0.66 0.513 −0.0143528 0.0286061
14 (dropped)

We now perform an ANOVA by years to see whether the strategy gives similar
Jensen values among the years in our sample. We obtain a value for F-statistic of
2.69 with a related probability of 0.0019; hence we conclude that the means for
Jensen values are not uniform from one year to another.
14. Kruskal-Wallis tests (Rank sums), by year
Year Obs. Rank Sum
1993 12 735
1994 12 665
1995 12 980
1996 12 1042
1997 12 1002
1998 12 1064
1999 12 1045
2000 12 972
2001 12 846
2002 12 913
2003 12 1136
2004 12 1070
2005 12 1353
2006 12 1373
chi-squared = 17.864 with 13 d.f.
probability = 0.1627
28 Can We Use the CAPM as an Investment Strategy? 781

We now use the Kruskal–Wallis nonparametric test to find out the same point as we
did with the ANOVA. We see that we obtain a probability greater than 5 % which
means we can accept that the mean Jensen values are uniform from one year to another.
15. Contingency table (by year)

BEAT
Year
NO YES Total
1993 8 4 12
1994 9 3 12
1995 6 6 12
1996 4 8 12
1997 6 6 12
1998 5 7 12
1999 4 8 12
2000 6 6 12
2001 7 5 12
2002 7 5 12
2003 3 9 12
2004 6 6 12
2005 2 10 12
2006 2 10 12
Total 75 93 168
Pearson chi2(13) = 19.9673 Pr = 0.096
likelihood-ratio chi2(13) = 21.0731 Pr = 0.071

The contingency table confirms the result of the previous test, namely, that there
are no differences between the Jensen values for the different years so that any
small difference can be due to chance.

Strategy 2, 2nd Variant: Quartiles Built Using Treynor’s Ratio

1. Summary statistics

Percentiles Smallest
1% −0.0968895 −0.104714
5% −0.0382262 −0.0968895
10% −0.0231222 −0.0940315 Obs. 172
25% −0.0106754 −0.0702515 Sum of Wgt. 172
50% 0.0023977 Mean 0.0014829
Largest Std. Dev. 0.0256145
75% 0.0157561 0.0484799
90% 0.0285186 0.0629996 Variance 0.0006561
95% 0.04065 0.0648316 Skewness −0.9274022
99% 0.0648316 0.0745625 Kurtosis 6.503481
782 F. Gómez-Bezares et al.

We note that in this table the Jensen values are far from normality, as can be seen in
the readings for asymmetry and kurtosis, although this is less obvious than for the
previous strategies.
2. Mean test

Variable Obs. Mean Std. Err. Std. Dev. [95% Conf. Interval]
Jensen (Q) 172 0.001483 0.0019531 0.0256145 −0.0023724 0.0053381
mean = mean (jensen) t = 0.7592
Ho: mean = 0 degrees of freedom = 171

Ha: mean < 0 Ha: mean != 0 Ha: mean > 0

Pr(T < t) = 0.7756 Pr(|T| > |t|) = 0.4487 Pr(T > t) = 0.2244

From this mean test, we can accept that the mean for the Jensen alphas is zero; in
fact, if we focus on the two-sided test, we obtain a probability of 0.4487, which is
higher than 5 %. This allows us to conclude that we cannot beat the Market with this
strategy.
3. Nonparametric mean test
sign Obs. sum ranks expected
positive 93 8460 7439
negative 79 6418 7439
zero 0 0 0
all 172 14878 14878
unadjusted variance 427742.50
adjustment for ties 0.00
adjustment for zeros 0.00
--------------
adjusted variance 427742.50
Ho: jensen = 0
z = 1.561
Prob > |z| = 0.1185

We can see again that Wilcoxon’s nonparametric test leads to the same conclu-
sion: we cannot beat the Market with this strategy.
4. Frequencies
BEAT
BEAT
NO YES Total
79 0 79
NO
45.93 0 45.93
0 93 93
YES
0 54.07 54.07
79 93 172
Total
45.93 54.07 100
28 Can We Use the CAPM as an Investment Strategy? 783

From this contingency table, we see that the proportion of months in

which we beat the Market is very similar to the proportion in which we do
not, with a slightly higher figure for those months in which we do beat the
Market.
5. Probabilities test

Variable N Observed K Expected K Assumed p Observed p

Jensen (Q) 172 93 86 0.50000 0.5407
Pr(k >= 93) = 0.160787 (one-sided test)
Pr(k <= 93) = 0.873669 (one-sided test)
Pr(k <= 79 or k >= 93) = 0.321574 (two-sided test)

We use this binomial test to determine whether the probability of beating the
Market is similar to the probability of failing to beat it (50–50). If we look at the
two-sided test, we obtain a probability of 0.321574, above 5 %, and so we accept
the success rate is around 50 %, which supports the results obtained with the
contingency table.
6. Fit of Jensen (quartiles) by month

.1

.05
qtreynor

−.05

−.1

1992m1 1994m1 1996m1 1998m1 2000m1 2002m1 2004m1 2006m1 2008m1

time
784 F. Gómez-Bezares et al.

7. Linear fit

Source SS df MS Number of obs. 172

Model 0.00721202 1 0.00721202 F(1, 170) 11.68
Residual 0.1049811 170 0.00061754 Prob > F 0.0008
Total 0.11219312 171 0.0006561 R-squared 0.0643
Adj R-squared 0.0588
Root MSE 0.02485

jensen (Q) Coef. Std. Err. t P>|t| [95% Conf. Interval]

time 0.0001304 0.0000382 3.42 0.001 0.0000551 0.0002058
_cons −0.0611824 0.0184347 −3.32 0.001 −0.0975728 −0.024792

From the above table we can observe that the regression slope which
connects Jensen’s index (the dependent variable) and time measured in months
(the independent variable) gives a positive and significant coefficient for
beta which allows us to conclude that this strategy gives results that improve
over time.
8. One-way analysis of Jensen (quartiles) by month

.1

.05

−.05

−.1
ry

ch

il

ay

er

r
us

be

be

be
r

l
ar

Ju
Ap
a

ob
M
ar

Ju

g
nu

ru

em

em

em
Au
M

ct
b
Ja

ov
pt

ec
Fe

Se

D
28 Can We Use the CAPM as an Investment Strategy? 785

9. One-way ANOVA (by month)

Source SS df MS Number of Obs. 172
Model 0.00656026 11 0.00059639 F(11, 160) 0.9
Residual 0.10563286 160 0.00066021 Prob > F 0.5387
Total 0.11219312 171 0.0006561 R-squared 0.0585
Adj. R-squared −0.0063
Root MSE 0.02569

jensen (Q) Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0092267 0.0068671 1.34 0.181 −0.0043352 0.0227886
1 −0.0125889 0.0097116 −1.3 0.197 −0.0317683 0.0065906
2 −0.0150679 0.0097116 −1.55 0.123 −0.0342474 0.0041115
3 0.003986 0.0095484 0.42 0.677 −0.0148711 0.0228431
4 −0.0040562 0.0095484 −0.42 0.672 −0.0229133 0.0148009
5 −0.0145235 0.0095484 −1.52 0.13 −0.0333805 0.0043336
6 −0.0055123 0.0097116 −0.57 0.571 −0.0246917 0.0136671
7 −0.0003726 0.0097116 −0.04 0.969 −0.0195521 0.0188068
8 −0.0092886 0.0095484 −0.97 0.332 −0.0281457 0.0095684
9 −0.013379 0.0097116 −1.38 0.17 −0.0325584 0.0058005
10 −0.0120712 0.0097116 −1.24 0.216 −0.0312506 0.0071083
11 −0.0105584 0.0097116 −1.09 0.279 −0.0297378 0.008621
12 (dropped)

We perform an ANOVA to see whether the means for the Jensen values obtained
for the different months are uniform. We find that F-statistic has a value of 0.90
with a related probability of 0.5387, above 5 %, which leads us to accept that the
means for the Jensen values are uniform among the months included in our sample.
10. Kruskal-Wallis tests (Rank sums), by month
Month Obs. Rank Sum
January 15 1153
February 15 1442
March 15 1297
April 14 1107
May 14 1046
June 14 1430
July 14 1220
August 14 899
September 14 1352
October 14 1160
November 14 1044
December 15 1728
chi-squared = 12.840 with 11 d.f.
probability = 0.3039
786 F. Gómez-Bezares et al.

We now use the Kruskal–Wallis nonparametric test to find out the same point as we
did with the ANOVA. We come to the same conclusion as we did with the ANOVA.
11. Contingency table (by month)
BEAT
Month
NO YES Total
January 7 8 15
February 5 10 15
March 6 9 15
April 9 5 14
May 8 6 14
June 3 11 14
July 7 7 14
August 10 4 14
September 7 7 14
October 6 8 14
November 8 6 14
December 3 12 15
Total 79 93 172
Pearson chi2 (11) = 15.8416 Pr=0.147
likelihood-ratio chi2(11) = 16.5468 Pr = 0.122

The contingency table allows us to test the same point as above with the ANOVA
and the Kruskal–Wallis test, and leads to the same conclusions; hence we can accept
that all months show a similar tendency in terms of beating the Market.
12. One-way analysis of Jensen (quartiles) by year

.1

.05

−.05

−.1
93

94

95

96

97

98

99

00

01

02

03

04

05

06
19

19

19

19

19

19

19

20

20

20

20

20

20

20
28 Can We Use the CAPM as an Investment Strategy? 787

13. One-way ANOVA (by year)

Source SS df MS Number of obs. 168
Model 0.0170417 13 0.0013109 F(13, 154) 2.14
Residual 0.09434205 154 0.00061261 Prob > F 0.0147
Total 0.11138374 167 0.00066697 R-squared 0.153
Adj. R-squared 0.0815
Root MSE 0.02475

jensen Coef. Std. Err. t P>|t| [95% Conf. Interval]

_cons 0.0135106 0.007145 1.89 0.061 −0.0006043 0.0276254
1 −0.041029 0.0101045 −4.06 0 −0.0609904 −0.0210676
2 −0.0190657 0.0101045 −1.89 0.061 −0.0390271 0.0008957
3 −0.0091611 0.0101045 −0.91 0.366 −0.0291225 0.0108003
4 −0.0100757 0.0101045 −1 0.32 −0.0300372 0.0098857
5 −0.0077821 0.0101045 −0.77 0.442 −0.0277435 0.0121793
6 −0.0098938 0.0101045 −0.98 0.329 −0.0298552 0.0100676
7 −0.0154658 0.0101045 −1.53 0.128 −0.0354272 0.0044956
8 −0.0103002 0.0101045 −1.02 0.31 −0.0302616 0.0096612
9 −0.0158053 0.0101045 −1.56 0.12 −0.0357667 0.0041561
10 −0.0165104 0.0101045 −1.63 0.104 −0.0364718 0.003451
11 −0.0080536 0.0101045 −0.8 0.427 −0.028015 0.0119078
12 −0.0072511 0.0101045 −0.72 0.474 −0.0272125 0.0127103
13 0.004184 0.0101045 0.41 0.679 −0.0157774 0.0241454
14 (dropped)

We now perform an ANOVA by years to see whether the strategy gives similar
Jensen values among the years in our sample. We obtain a value for F-statistic of
2.14 with a related probability of 0.0147; hence we conclude that for a significance level
of 1 % we can accept that the means for the Jensen values are the same from one year to
another, however for a significance level of 5 % we must reject this hypothesis and we
would conclude that there are differences between the Jensen values of different years.
14. Kruskal-Wallis tests (Rank sums), by year
Year Obs. Rank Sum
1993 12 740
1994 12 679
1995 12 1025
1996 12 1033
1997 12 1065
1998 12 1022
1999 12 877
2000 12 1026
2001 12 824
2002 12 920
2003 12 1166
2004 12 1110
2005 12 1322
2006 12 1387
chi-squared = 18.336 with 13 d.f.
probability = 0.1452
788 F. Gómez-Bezares et al.

We now use the Kruskal–Wallis test to find out the same point as we did with the
ANOVA. We see that we obtain a probability greater than 5 % which means we can
accept that the means for the Jensen values are uniform from one year to another.
15. Contingency table (by year)

BEAT
Year
NO YES Total
1993 8 4 12
1994 9 3 12
1995 4 8 12
1996 5 7 12
1997 6 6 12
1998 5 7 12
1999 6 6 12
2000 4 8 12
2001 8 4 12
2002 7 5 12
2003 4 8 12
2004 6 6 12
2005 2 10 12
2006 2 10 12
Total 76 92 168
Pearson chi2 (13) = 19.9908 Pr = 0.095
likelihood-ratio chi2(13) = 21.0581 Pr = 0.072

The contingency table confirms the result of the previous test, namely, that there
are no differences between the Jensen values for the different years and that any
small difference can be due to chance.

References
Abdel-Kader, M. G., & Qing, K. Y. (2007). Risk-adjusted performance, selectivity, timing ability
and performance persistence of Hong Kong mutual funds. Journal of Asia-Pacific Business,
8(2), 25–58.
Agresti, A. (2007). An introduction to categorical data analysis (2nd ed.). Hoboken: Wiley
Interscience.
Anderson, D. R., Sweeney, D. J., & Williams, T. A. (2011). Statistics for business and economics
(11th ed.). Mason: South-Western, Cengage Learning.
Black, F., Jensen, M. C., & Scholes, M. (1972). The capital asset pricing model: Some empirical tests.
In Jensen, M. C. (Ed.), Studies in the theory of capital Markets (pp. 79–121). New York: Praeger.
Brav, A., Lehavy, R., & Michaely, R. (2005). Using expectations to test asset pricing models.
Financial Management, 34(3), 31–64.
Brealey, R. A., Myers, S. C., & Allen, F. (2008). Principles of corporate finance (9th ed.). New
York: McGraw-Hill.
Conover, W. J. (1999). Practical nonparametric statistics (3rd ed.). New York: Wiley.
Copeland, T. E., Weston, J. F., & Shastri, K. (2005). Financial theory and corporate policy
(4th ed.). Boston: Pearson Addison Wesley.
28 Can We Use the CAPM as an Investment Strategy? 789

Danielson, M. G., Heck, J. L., & Shaffer, D. R. (2008). Shareholder theory – How opponents and
proponents both get it wrong. Journal of Applied Finance, 18(2), 62–66.
Dimson, E., & Mussavian, M. (1998). A brief history of Market efficiency. European Financial
Management, 4(1), 91–103.
Dimson, E., & Mussavian, M. (1999). Three centuries of asset pricing. Journal of Banking and
Finance, 23(12), 1745–1769.
Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. Journal of
Finance, 25(2), 383–417.
Fama, E. F. (1991). Efficient capital markets: II. Journal of Finance, 46, 1575–1617.
Fama, E. F. (1998). Market efficiency, long-term returns, and behavioral finance. Journal of
Financial Economics, 49, 283–306.
Fama, E. F., & French, K. R. (1992). The cross-section of expected stock returns. Journal of
Finance, 47, 427–465.
Fama, E. F., & French, K. R. (2008). Dissecting anomalies. Journal of Finance, 63(4), 1653–1678.
Fama, E. F., & French, K. R. (2010). Luck versus skill in the cross-section of mutual fund returns.
Journal of Finance, 65(5), 1915–1947.
Fama, E. F., & MacBeth, J. D. (1973). Risk, return and equilibrium: empirical tests (pp. 607–636).
May-June: Journal of Political Economy.
Ferruz, L., Gómez-Bezares, F., & Vargas, M. (2010). Portfolio theory, CAPM, and performance
measures. In Lee, C. F. et al. (Eds.), Handbook of quantitative finance and risk management
(pp. 267–283). New York: Springer.
Gómez-Bezares, F., & Gómez-Bezares, F. R. (2006). On the use of hypothesis testing and other
problems in financial research. The Journal of Investing, 15(4), 64–67.
Gómez-Bezares, F., Madariaga, J. A., & Santibáñez, J. (1996). Valuation and efficiency models:
Does the CAPM beat the Market? Análisis Financiero, 68, 72–96.
Jensen, M. C. (1968). The performance of mutual funds in the period 1945–1964 (pp. 389–416).
May: Journal of Finance.
Jensen, M. C. (1969). Risk, the pricing of capital assets, and the evaluation of investment
portfolios. Journal of Business, 42, 167–247.
Levy, H. (2010). The CAPM is alive and well: A review and synthesis. European Financial
Management, 16(1), 43–71.
Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets. Review of Economics and Statistics, 47, 13–37.
Liu, W. (2006). A liquidity-augmented capital asset pricing model. Journal of Financial Econom-
ics, 82, 631–671.
Malkiel, B. G. (2003). The efficient Market hypothesis and its critics. Journal of Economic
Perspectives, 17(1), 59–82.
Mazumder, M. I., & Varela, O. (2010). Market timing, the trading of international mutual funds:
weekend, weekday and serial correlation strategies. Journal of Business, Finance & Account-
ing, 37(7–8), 979–1007.
Modigliani, L. (1997). Don’t pick your managers by their alphas. U.S. Investment Research,
U.S. Strategy, Morgan Stanley Dean Witter, June.
Newbold, P., Carlson, W. L., & Thorne, B. M. (2009). Statistics for business and economics
(7th ed.). New Jersey: Pearson Education Canada.
Pasaribu, R. B. (2009). Relevance of accrual anomaly information in stock portfolio
formation – case study: Indonesia stock exchange. Journal of Accounting and Business,
10(2), 1–45.
Samarakoon, L. P., & Hasan, T. (2005). Portfolio performance evaluation. In Lee, C. F. (Ed.), The
encyclopedia of finance (pp. 617–622). New York: Springer.
Sharpe, W. F. (1964). Capital asset prices: A theory of Market equilibrium under conditions of
risk. Journal of Finance, 19, 425–442.
Treynor, J. L. (1965). How to rate management of investment funds. Harvard Business Review, 43,
63–75.
Group Decision-Making Tools for
Managerial Accounting and Finance 29
Applications

Wikil Kwak, Yong Shi, Cheng-Few Lee, and Heeseok Lee

Contents
29.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793
29.2 Designing a Comprehensive Performance Evaluation System: Using the Analytic
Hierarchy Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793
29.2.1 Hierarchical Schema for Performance Measurement . . . . . . . . . . . . . . . . . . . . . . . . . 795
29.2.2 Analytic Hierarchical Performance Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 796
29.2.3 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 798
29.2.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 800
29.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
29.3 Developing a Comprehensive Performance Measurement System in the Banking
Industry: An Analytic Hierarchy Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
29.3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 802
29.3.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
29.3.3 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
29.3.4 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807

W. Kwak (*)
University of Nebraska at Omaha, Omaha, NE, USA
e-mail: wkwak@unomaha.edu
Y. Shi
University of Nebraska at Omaha, Omaha, NE, USA
Chinese Academy of Sciences, Beijing, China
e-mail: yshi@unomaha.edu
C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu
H. Lee
Korea Advanced Institute of Science and Technology, Yuseong-gu, Daejeon, South Korea
e-mail: hsl@business.kaist.ac.kr

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 791
DOI 10.1007/978-1-4614-7750-1_29,
# Springer Science+Business Media New York 2015
792 W. Kwak et al.

29.4 Optimal Trade-offs of Multiple Factors in International Transfer Pricing

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807
29.4.1 Existing Transfer Pricing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
29.4.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 810
29.4.3 Model Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
29.4.4 Optimal Trade-offs and Their Accounting Implications . . . . . . . . . . . . . . . . . . . . . . 816
29.4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822
29.5 Capital Budgeting with Multiple Criteria and Multiple Decision Makers . . . . . . . . . . . . 822
29.5.1 AHP and MC2 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
29.5.2 Model Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 829
29.5.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 833
29.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 833
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 836

Abstract
To deal with today’s uncertain and dynamic business environments with differ-
ent background of decision makers in computing trade-offs among multiple
organizational goals, our series of papers adopts an analytic hierarchy process
(AHP) approach to solve various accounting or finance problems such as devel-
oping a business performance evaluation system and developing a banking
performance evaluation system. AHP uses hierarchical schema to incorporate
nonfinancial and external performance measures. Our model has a broader set of
measures that can examine external and nonfinancial performance as well as
internal and financial performance. While AHP is one of the most popular
multiple goals decision-making tools, multiple-criteria and multiple-constraint
(MC2) linear programming approach also can be used to solve group decision-
making problems such as transfer pricing and capital budgeting problems. This
model is rooted by two facts. First, from the linear system structure’s point of
view, the criteria and constraints may be “interchangeable.” Thus, like multiple
criteria, multiple-constraint (resource availability) levels can be considered.
Second, from the application’s point of view, it is more realistic to consider
multiple resource availability levels (discrete right-hand sides) than a single
resource availability level in isolation. The philosophy behind this perspective
is that the availability of resources can fluctuate depending on the decision
situation forces, such as the desirability levels believed by the different
managers. A solution procedure is provided to show step-by-step procedure to
get possible solutions that can reach the best compromise value for the multiple
goals and multiple-constraint levels.

Keywords
Analytic hierarchy process • Multiple-criteria and multiple-constraint linear
programming • Business performance evaluation • Activity-based costing sys-
tem • Group decision making • Optimal trade-offs • Balanced scorecard •
Transfer pricing • Capital budgeting
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 793

29.1 Introduction

In this chapter, we provide an up-to-date review on our past works in AHP

and MC2 linear programming models to solve real-world problems faced by
managers in accounting and finance. These applications include developing
a business performance evaluation system using an analytic hierarchical model,
a banking performance evaluation system, capital budgeting, and transfer pricing
problems. A good performance measurement system should incorporate strategic
success factors, especially to be successful in today’s competitive environment.
Balanced scorecard is a hot topic, but it lacks linkages among different basic
units of financial and nonfinancial measures or across different levels of managers.
The model proposed in this study uses a three-level hierarchical schema to
combine financial and nonfinancial performance measures systematically. Its
emphasis is on an external as well as an internal business performance measures
such as the balanced scorecard method. This method is more likely to cover
a broader set of measures that include operational control as well as strategic
control.
The purpose of this chapter is to provide additional insight for managers who
face group decision-making problems in accounting and finance and want to find
practical solutions.

29.2 Designing a Comprehensive Performance Evaluation

System: Using the Analytic Hierarchy Process

AHP is one of the most popular multiple goals decision-making tools (Ishizaka
et al. 2011). Designing a comprehensive performance measurement system has
frustrated many managers (Eccles 1991). The traditional performance measures
enterprises have used may not well fit in for the new business environment and
competitive realities. The figures that enterprises have traditionally used are not
very useful for the information-based society we are becoming. We suspect that
firms are much more productive than these out-of-date measures.
A broad range of firms is deeply engaged in redefining how to evaluate the
performance of their businesses. New measurements for quantification are needed
to perform business evaluation. Drucker (1993) put the ever-increasing measure-
ment dilemma this way:
Quantification has been the rage in business and economics these past 50 years. Accoun-
tants have proliferated as fast as lawyers. Yet we do not have the measurements we need.
Neither our concepts nor our tools are adequate for the control of operations, or for
managerial control. And, so far, there are neither the concepts nor the tools for business
control, i.e., for economic decision making. In the past few years, however, we have
become increasingly aware of the need for such measurements.

Drucker’s message is clear: a traditional measure is not adequate for business

evaluation. A primary reason why traditional measures fail to meet new business
needs is that most measures are lagging indicators (Eccles and Pyburn 1992).
794 W. Kwak et al.

The emphasis of accounting measures has been on historical statements of financial

performance. They are the result of management performance, not the cause of it;
i.e., they are better at measuring the consequence of yesterday’s decisions but
unlikely to provide useful indicators for future success. As a result, they easily
conflict with new strategies and current competitive business realities.
To ameliorate this accounting lag situation, researchers have frequently attempted
to provide new measuring procedures (Kaplan 1986; Wu et al. 2011). Yet most
measuring guidelines may not be well represented analytically. Managers keep
asking: What are the most important measures of performance? What are the associ-
ations among those measures? Unfortunately, we know little about how measures are
integrated into a performance measurement system regarding a particular business.
For example, is the customer service perceived as a more important measure than cost
of quality? Another question often raised is: What is the association between the
customer service and on-time delivery or manufacturing cycle time?
The current wave of dissatisfaction with traditional accounting systems has been
intensified partly because most measures have internal and financial focus. The new
measure should broaden the basis of nonfinancial performance measurement. Mea-
sures must truly predict long-term strategic success. External performance relative to
competitors such as market share is as of importance as internal measures. In addition,
the recent rise of global competitiveness reemphasizes the primacy of operational, i.e.,
nonfinancial, performance over financially oriented performance. Nonfinancial
measures reflect the actionable steps needed for surviving in today’s competitive
environment (Fisher 1992).
When a company uses an activity-based costing system (Campi 1992; Ayvaz and
Pehlivanli 2011) or just-in-time manufacturing system, using nonfinancial measures is
inevitable. Nonfinancial measures reduce communication gap between workers and
managers; i.e., workers can better understand what they are measured by, and
managers can get timely feedback and link them to strategic decision making.
The answer proposed in this study is to use hierarchical schema to incorporate
nonfinancial and external performance measures. The model has a broader set of
measures that can examine external and nonfinancial performance as well as internal
and financial performance. On the basis of the schema, this chapter demonstrates how
Saaty’s analytic hierarchy process (e.g., see Saaty (1980) and Harker and Vargas
(1987)) can be merged with the performance measurement. The analytic hierarchy
process is a theory of measurement that has been widely applied in modeling human
judgment process. In this sense, the performance measuring method proposed in this
study is referred to as the Analytic Hierarchical Performance Model (AHPM).
While the AHP has been applied in a number of cases of capital budgeting,
auditing, preference analysis, and balanced scorecard to product planning, enter-
prise risk management, and internal control structure study (see Arrington
et al. 1984; Boucher and MacStravic 1991; Liberatore et al. 1992; Hardy and
Reeve 2000; Huang et al. 2011; Li et al. 2011), little attention is devoted to the
problem of an analytical and comprehensive business performance model to cover
a broader base of measures in the currently changing environment in accounting
information systems. Although Chan and Lynn (1991) originally investigated the
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 795

application of the AHP in business performance evaluation, the problem of struc-

turing the decision hierarchy in an appropriate manner is yet to be explored. The
methodology proposed in this chapter will resolve this issue.

29.2.1 Hierarchical Schema for Performance Measurement

The AHP is a theory of measurement that has been extensively applied in modeling
the human judgment process (e.g., Lee 1993; Muralidhar et al. 1990). It decom-
poses a complex decision operation into a multilevel hierarchical structure. The
primary advantage of the AHP is its simplicity and the availability of the software.
Several desirable features of the AHP can help resolve issues in performance
evaluation. For example, nonfinancial and external effects can be identified and
integrated with financial and internal aspects of business performance through the
AHP. Furthermore, the AHP is a participation-oriented methodology that can aid
coordination and synthesis of multiple evaluators in the organizational hierarchy.
Participation makes a positive contribution to the quality of the performance
evaluation process. This point is further explored within the context of hierarchical
schema of performance evaluation as follows.
Performance measures have the relationship with management levels. They need
to be filtered at each superior/subordinate level in an organization; i.e., measures do
not need to be the same across management levels. Performance measures at each
level, however, should be linked to performance measures at the next level up.
Performance measurement information is tailored to match the responsibility of
each management level. For example, at the highest level, the CEO has responsi-
bility for performance of the total business. In contrast, the production manager’s
main interest may be in cost control because he or she is responsible for this. Taking
this idea into account leads to the notion that the performance measuring process
consists of different levels according to management levels.
Depending on organizational levels, therefore, a three management level model
will be suggested: top, middle, and operational. To recognize the ever-increasing
importance of nonfinancial measures, at the top management level, the hierarchy
consists of two criteria: nonfinancial and financial performance. One level lower
(middle-level management) includes performance criteria such as market share,
customer satisfaction, productivity, ROI, and profitability.
The lowest level (operational management level) includes measures that lead to
simplifications in the manufacturing process as related to high-level performance
measures. Examples are quality, delivery, cycle time, inventory turnover, asset turn-
over, and cost. Typically, the criteria at the lowest level may have several sub-criteria.
For example, quality measures may have four sub-criteria: voluntary, (appraisal and
prevention,) and failure, (internal and external,) costs. With relation to these criteria,
accounting information is controlled typically with respect to each product (or service)
or division (or department). As a result, product and division levels are added. One
should keep in mind that the above three-level hierarchical schema is dynamic over
time. As a company evolves, the hierarchy must be accordingly adjusted.
796 W. Kwak et al.

Another interesting aspect is that the hierarchy is not company invariant. The
hierarchy must be adjusted depending on unique situation faced by each individual
company or division. Basically, the hierarchal schema is designed to accommodate
any number of levels and alternatives. New level or alternative can be easily added or
deleted to the hierarchy once introduced. For example, another level regarding prod-
ucts may be added at the bottom in order to evaluate performance of different products.

29.2.2 Analytic Hierarchical Performance Model

Based on the hierarchical structure, the performance indices at each level of the
AHPM are derived by the analytic hierarchy process. The AHPM collects input
judgment in the form of matrix by pairwise comparisons of criteria. An eigenvalue
method is then used to scale weights of criteria at each level; i.e., the relative
importance of each criterion at each level is obtained. The relative importance is
defined as a performance index with respect to each alternative (i.e., criterion,
product, or division).
From now on, each step of the AHPM in obtaining the weights is explored. First,
a set of useful strategic criteria must be identified. Let nt be the total number
of criteria under consideration at the top management level. Typically,
nt ¼ 2 (nonfinancial and financial measures). The relative weight of each criteria
may be evaluated by pairwise comparison; i.e., two criteria are compared at one time
until all combinations of comparison are considered (only one pairwise comparison
is needed if nt ¼ 2). The experience of many users of this method and the experiments
reported (Saaty 1980) are likely to support that the 1–9 scale for pairwise compar-
isons captures human judgment fairly well while the scale can be altered to suit each
application. The result from all of pairwise comparisons is stored in an input matrix:


At ¼ atij ðan nt by nt matrixÞ:

The element atij states the importance of alternative i compared to alternative j.

For instance, if at at12 ¼ 2, then criterion 1 is twice as important as criterion 2.
Applying an eigenvalue method to At results in a vector Wt ¼ (wti) that has nt
elements.
In addition to the vector, the inconsistency ratio (g) is obtained to estimate the
degree of inconsistency in pairwise comparisons. The common guideline is that if
the ratio surpasses 0.1, a new input matrix must be generated. Generally speaking,
each element of the vector resulting from an eigenvalue method is the estimated
relative weight of the corresponding criterion of one level with respect to one level
higher; i.e., the element wti is the relative weight of the ith criteria at this level.
At the second level of hierarchy, consider the ith criterion of the top management
level. Then, we have one input matrix of pairwise comparisons of criteria (at middle
management level) that corresponds to the ith criterion of the top management
level. The result is stored in an nm by nm matrix Aim. Here, nm is the total number of
criteria at this level. Applying an eigenvalue method to Aim results in the relative
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 797

weight for each criterion with relation to the ith criteria at one level higher. This
“local relative weight” is stored in a local weighing vector Wim. We need nt local
weighing vectors at this level. The “global relative weight” at this level is computed
and then stored in a vector Wm ¼ (wmi) that has the nm elements as
 
W m ¼ W t  W 1m . . . W nt
m :

The similar computing process continues at the operational management level

until we have all global relative weights as Wo.
A prototype for the AHPM with three-level hierarchies was built via commercial
software for the analytic hierarchy process, called Expert Choice (Forman
et al. 1985).
Above relative weights can be utilized in a number of ways for performance
measurement. Clearly, it implies the relative importance among criteria at each
level. For example, consider an automobile company where market share and ROI
are two important criteria in performance evaluation. If the AHPM generates their
weights as 0.75 and 0.25, respectively, it is reasonable to conclude that market
share affects the company’s performance three times higher than ROI. The weights
can be used as a measure for allocating future resources in products or
divisions. Assume the automobile company produces two types of autos, sedan
and minivan. If the AHPM generates global relative weights as 0.8 and 0.2,
respectively, i.e., the performance of sedans is four times higher than minivans,
then this may provide a good reason for the CEO to invest in sedans four times
higher than minivans.
One important performance control measure is the rate of performance change
that can be computed at any level. The change of performance can be measured.
This measure is useful in estimating the elasticity of the performance of any
alternative. This elasticity can aid in resource allocation decisions; i.e., further
resources may be assigned to more elastic products of divisions.
For the example of computing this elasticity, at the middle management level,
the rate is then
Xnm
em ¼ i¼1
Wmi DCmi =Cmi :

Here, DCmi is the amount of change of the ith criteria. For example, if the ROI is
increased from 5 % to 7 % and the market share is changed from 20 % to 25 % in the
automobile company discussed above,

2 5
em ¼ 0:25  þ 0:75  ¼ 0:2875:
5 20

We may conclude that the overall business performance has been increased by
$28.75 % with respect to middle management level. Similarly, the performance
change rates at any level can be obtained. Typically, they vary depending on levels
of the AHPM; i.e., they have different implications.
798 W. Kwak et al.

29.2.3 An Example

The suitability of the AHPM is illustrated with a study on a hypothetical automobile

company. Nonfinancial criteria include market share, customer satisfaction, and
productivity. Financial criteria include ROI and profitability. At the operational
level, six criteria such as quality, delivery, cycle time, inventory turnover, asset
turnover, and cost are considered.
First, nonfinancial and financial criteria are compared and stored in a vector:

Wt ¼ ð0:4; 0:6Þ:

Next, the middle management level is considered. The relative weights of

middle-level performance criteria with relation to each top-level criterion are to
be computed. First, the local relative weights were computed.
For the nonfinancial criteria,
2 3
1 2 4
A1m ¼ 4 1=2 1 35
1=4 1=3 1

For the sake of convenience, ROI and profitability are not listed in this compar-
ison matrix.
The lower triangle is not listed because, in the eigenvalue method, the lower
triangle is simply the reciprocity of the upper triangle. As a result,

W 1m ¼ ð0:558, 0:320, 0:122Þ:

For each weighing computing, an inconsistency ratio was computed and checked
for acceptance; i.e., in this case, the ratio (g ¼ 0.017) was accepted because g  0.1.
For the financial criteria, ROI is estimated to be twice more important than
profitability, i.e.,

W 2m ¼ ð0:667, 0:333Þ:

Accordingly, the global relative weights of the managerial criteria (in the order
of market share, customer satisfaction, productivity, ROI, and profitability) are then
 
0:558 0:320 0:122 0 0
Wm ¼ ð0:4, 0:6Þ 
0 0 0 0:667 0:333
¼ ð0:223, 0:128, 0:049, 0:400, 0:200Þ:

Here, the global relative weights of market share, customer satisfaction, produc-
tivity, ROI, and profitability are 22.3 %, 12.8 %, 4.9 %, 40 %, and 20 %, respec-
tively. Note that these percentages are elements of the above Wm.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 799

Let us move down to the operational level. The following are local relative
weights of operational level criteria (in the order of quality, delivery, cycle time,
cost, inventory turnover, and asset turnover).
For the market share,
2 3
1 2 3 3 5 7
6 1=2 1 2 3 4 57
6 7
6 1=3 1=2 1 2 4 57
A0 ¼ 6
1 6 7
6 1=3 1=3 1=2 1 2 377
4 1=5 1=4 1=4 1=2 1 25
1=7 1=5 1=5 1=3 1=2 1

For convenience, the local weights are arranged in the order of quality, cycle
time, delivery, cost, inventory turnover, and asset turnover:

W 1o ¼ ð0:372, 0:104, 0:061, 0:250, 0:174, 0:039Þ:

Similarly,

W 2o ¼ ð0:423, 0:038, 0:270, 0:185, 0:055, 0:029Þ;

W 3o ¼ ð0:370, 0:164, 0:106, 0:260, 0:064, 0:037Þ;

W 4o ¼ ð0:220, 0:060, 0:030, 0:4150, 0:175, 0:101Þ;

W 5o ¼ ð0:246, 0:072, 0:065, 0:334, 0:160, 0:122Þ:

Consequently, we get the global relative weights of operational criteria as

Wo ¼ ð0:223; 0:128; 0:049; 0:400; 0:200Þ

2 3
0:372 0:104 0:061 0:250 0:174 0:039
6 0:423 0:038 0:270 0:185 0:055 0:029 7
6 7
6 7
 6 0:370 0:164 0:106 0:260 0:064 0:037 7
6 7
4 0:220 0:060 0:030 0:414 0:175 0:101 5
0:246 0:072 0:065 0:334 0:160 0:122
¼ ð0:292; 0:074; 0:078; 0:324; 0:151; 0:079Þ:

Finally, the relative importance of operational level performance measures are

29.2 %, 7.4 %, 7.8 %, 32.4 %, 15.1 %, and 7.9 %, respectively. One should note that
nonfinancial measures are integrated with financial measures in the scaling process.
Our automobile company can adopt these performance measures for further
analysis. The measures serve as the basis for the rate of performance change.
In addition, they can be further used for evaluating the performance of each product
if the product level is connected to the operational management level in the AHPM.
800 W. Kwak et al.

29.2.4 Discussions

It is too optimistic to argue that there can be one right way of measuring perfor-
mance. There are two factors to be considered. First, each business organization
requires its own unique set of measures depending on its environment. What is most
effective for a company depends upon its history, culture, and management style
(Eccles 1991). To be used for any business performance measurement, a well-
designed model must be flexible enough to incorporate a variety of measures while
retaining major aspects. Second, managers should change measures over time.
In the current competitive business environment, change is not incidental. It is
essential. Change must be managed. The AHPM is highly flexible so that managers
may adjust its structure to evolving business environments. This flexibility will
allow a company to improve its performance measurement system continuously.
Generally, a performance measurement system exists to monitor the implemen-
tation of planning of an organization and aid to motivate desirable individual
performance through a realistic communication of performance information in
related goals of business. This premise of performance measurement requires
a significant number of feedbacks and corrective actions in the practice of account-
ing information systems (Nanni et al. 1990), i.e., feedbacks between levels
and also within level. In the implementation of AHPM, these activities are
common procedures. A clean separation of levels in the hierarchy of the AHPM
is an idealization for simplifying the presentation. The flexibility of the AHPM,
however, accommodates as many feedbacks and corrective actions as possible.
Performance is monitored by group rather than by an individual because of
its ever-increasing importance of teamwork in the success of businesses. The
integrated structure of the AHPM facilitates group decision and thus increases the
chance that managers put trust in the resulting measures. This structure will
enhance more involvement of lower-level managers as well as workers
when a firm implements the AHPM. Furthermore, the hierarchy of the AHPM
corresponds to that of business organization. As a result, group decision at each
level is facilitated. For example, middle-level managers will be responsible for
determining weights for middle-level criteria while lower-level managers will be
responsible for operational level criteria.
The iterative process of weighing goals among managers, as the implementation
of AHPM progresses, will help them to understand which strategic factors are
important and how these factors are linked to other goals to be a successful
company as a group.
Information technology plays a critical role in designing a performance mea-
surement system to provide timely information to management. A computer-based
decision support system can be used to utilize this conceptual foundation in a real-
world situation. The AHPM can be easily stored in an enterprise database because
of the commercial software, Expert Choice. As a result, the AHPM can be readily
available to each management level via the network system of an enterprise.
The AHPM fits any corporate information architecture to pursue the company’s
long-term strategy.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 801

29.2.5 Conclusions

A well-designed performance model is a must for an enterprise to gain competitive

edges. The performance measurement model proposed in this study is the first kind
of analytical model to cover a wide variety of measures while providing operational
control as well as strategic control. With comparison to previous evaluation
methods, the model shows advantages such as flexibility, feedbacks, group evalu-
ation, and computing simplicity. A prototype was built via a personal computer so
that the model can be applied to any business situations.
The possible real-time control is of importance for the competitive business
environment we are facing today. In sum, the contribution of this study is of both
conceptual and practical importance.

29.3 Developing a Comprehensive Performance Measurement

System in the Banking Industry: An Analytic Hierarchy
Approach

The objective of this study is to design a practical model for a comprehensive

performance measurement system that incorporates strategic success factors in the
banking industry. A performance measurement system proposed in this study can
be used to evaluate top managers of a main bank or managers of a branch office.
Performance measurement should be closely tied to goal setting of an organiza-
tion because it feeds back information to the system on how well strategies are
being implemented (Chan and Lynn 1991; Eddie et al. 2001). A balanced scorecard
approach (Kaplan and Norton 1992; Tapinos et al. 2011) is a hot topic currently in
this area, but it does not provide systematic aggregation of each level as well as
different levels of managers’ performance for the overall company. In other words,
there is no systematic linkage between financial and nonfinancial measures across
different levels of management hierarchy. A traditional performance measurement
system, which focuses on financial measures such as return on assets (ROA),
however, may not serve this purpose well for middle- or lower-level managers in
the new competitive business environment either.
The model proposed in this study will have a broader set of measures that
incorporate traditional financial performance measures such as return on assets
and debt to equity ratio as well as nonfinancial performance measures such as the
quality of customer service and productivity. Using Saaty’s analytic hierarchy
process (AHP) (Saaty 1980; Harker and Vargas 1987), the model will demonstrate
how multiple performance criteria can be systematically incorporated into
a comprehensive performance measurement system. The AHP enables decision
makers to structure a problem in the form of a hierarchy of its elements according to
an organization’s structure or ranks of management levels and to capture manage-
rial decision preferences through a series of comparisons of relevant factors or
criteria. The AHP has been applied recently to several business problems (e.g.,
divisional performance evaluation (Chan and Lynn 1991), capital budgeting
802 W. Kwak et al.

(Liberatore et al. 1992), and real estate investment (Kamath and Khaksari 1991),
marketing applications (Dyer and Forman 1991), information system project
selection (Schniederjans and Wilson 1991), activity-based costing cost driver
selection (Schniederjans and Garvin 1997), developing lean performance measures
(DeWayne 2009), and customer’s choice analysis in retail banking (Natarajan
et al. 2010)). The AHP is relatively easy to use and its commercial software is
available. This study will be an analytical and comprehensive performance evaluation
model to cover a broader base of measures in the rapidly changing environment of
today’s banking industry. The following section presents background. The third
section discusses methodology. The fourth section presents a numerical example.
The last section summarizes and concludes this chapter.

29.3.1 Background

Recent research topic guide from Institute of Management Accountants lists “per-
formance measurement” as one of top priority research issues. Therefore, this
project will be interesting to bank administrators as well as managerial accountants.
With unprecedented competitive pressure from nonbanking institutions, dereg-
ulation, and the rapid acquisition of smaller banks by large national or regional
banks, the most successful banks in the new millennium will be the ones that adapt
strategically to a changing environment (Calvert 1990). Management accounting
and finance literature have emphasized using both financial and nonfinancial
measures as performance guidelines in the new environment (e.g., Chan and
Lynn 1991; Rotch 1990). However, most studies do not propose specifically how
we should incorporate these financial and nonfinancial factors into a formal model.
The performance measurement system proposed in this study is the formal
model applied using the AHP in the banking industry to cover a wide variety of
measures while providing operational control as well as strategic control. The AHP
can incorporate multiple-subjective goals into a formal model (Dyer and Forman
1991). Unless we design a systematic performance measurement system that
includes financial as well as nonfinancial control factors, there may be incorrect
behavior by employees because they misunderstand the organization’s goals and
how they relate to their individual performance.
Compared with previous evaluation methods, the model proposed in this study will
have advantages such as flexibility, continuous feedback, teamwork in goal setting,
and computational simplicity. To be used for any business performance measurement,
a well-designed model must be flexible enough to incorporate a variety of measures
while retaining major success factors. The AHP model is flexible enough for managers
to adjust its structure to a changing business environment through an iterative process
of weighing goals. This flexibility will allow a company to improve its performance
measurement system continuously. Through the iterative process of goal comparisons,
management could get continuous feedback for the priority of goals and work as
a team. The possible real-time control is of importance in the competitive business
environment we are facing today.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 803

29.3.2 Methodology

The analytic hierarchy process (AHP) collects input judgments in the form of a matrix
by pairwise comparisons; i.e., two criteria are compared at one time. The experience of
many users of this method supports use of a 1–9 scale for pairwise comparisons to capture
human judgment while the scale can be altered to fit each application (Saaty 1980).
A simple example will be provided to explain how AHP operates. Consider the
situation where a senior executive has to decide on which of three managers to
promote to a senior position in the firm. The candidate’s profiles have been studied
and rated on three criteria: leadership, human relations skills, and financial man-
agement ability. First, the decision maker compares each of the criteria in pairs to
develop a ranking of the criteria. In this case, the comparisons would be:
1. Is leadership more important than human relations skills for this job?
2. Is leadership more important than financial management ability for this job?
3. Are human relations skills more important than financial management ability for
this job?
The response to these questions would provide an ordinal ranking of the three
criteria. By adding a ratio scale of 1–9 for rating the relative importance of one criterion
over another, a decision maker could make statements such as “leadership is four times
as important as human relations skills for this job,” “financial management ability is
three times as important as leadership,” and “financial management ability is seven
times as important as human relations skills.” These statements of pairwise compar-
isons can be summarized in a square matrix. The preference vectors are then computed
to determine the relative rankings of the three criteria in selecting the best candidate.
For example, the preference vectors of the three criteria are 0.658 for financial
management ability, 0.263 for leadership, and 0.079 for human relations skills.
Once the preference vector of the criteria is determined, each of the candidates
can be compared on the basis of the criteria in the following manner:
1. Is candidate A superior to candidate B in leadership skills?
2. Is candidate A superior to candidate C in leadership skills?
3. Is candidate B superior to candidate C in leadership skills?
Again, rather than using an ordinal ranking, the degree of superiority of one
candidate over another can be assessed. The same procedures can be applied to
human relations skills and financial management ability. The responses to these
questions can be summarized in matrices where the preference vectors are again
computed to determine the relative ranking of the three candidates for each criterion.
Accordingly, the best candidate should be the one who ranks “high” on the “more
important” criteria. The matrix multiplication of preference vectors of candidates on
evaluation criteria and the preference vector of evaluation criteria will provide the
final ranking of the candidates. In this example, the candidates are ranked A, B,
and C. This example provides the usefulness of the AHP for setting priorities for
both qualitative and quantitative measures (Chan and Lynn 1991).
We could apply the same procedures to a bank. Based on the hierarchical
structure of a banking institution, the relative weights of criteria at each level of
managers are derived by AHP. Here the relative importance of performance
804 W. Kwak et al.

measures can be defined as a performance index with respect to each alternative

(i.e., criterion, service, or branch office). For example, derive the relative weights of
financial and nonfinancial criteria at the highest management level through the
pairwise comparisons of criteria. Next, the relative weights of middle-level perfor-
mance criteria with relation to each top-level criterion are to be computed then enter
these relative weights into an n x n matrix format. Finally, this matrix is multiplied
by the relative weights of criteria of the top management level. The same pro-
cedures can be applied to the next lower-level management. This AHP approach
could link systematically the hierarchical structure of business performance mea-
surement between different levels of organizational structures.
Consider a local bank where market share and return on assets are two important
criteria in performance evaluation. If the AHP generates their weights as 0.4 and
0.6, respectively, it is reasonable to conclude that return on assets affects the bank’s
performance one and a half times higher than market share. The weights can be
used as a measure for allocating future resources in products or branch offices.
Assume the bank has two types of services, commercial loans and residential
mortgage loans. If the AHP generates relative weights as 0.8 and 0.2 (i.e., the
performance of commercial loans is four times higher than mortgage loans), then
this may provide a good reason for top management to invest resources in the
commercial loan market four times higher than the residential loan market.
The use of the AHP for multiple-criteria situation is superior to ad hoc weighing
because it has the advantage of forcing the decision maker to focus exclusively on the
criteria at one time and the way in which they are related to each other (Saaty 1980).
A model could be built using a microcomputer program called Expert Choice so
that the model can be applied to any bank easily.

29.3.3 A Numerical Example

This section presents a numerical example for a commercial bank. The Commercial
Omaha Bank (COB) is a local bank that specializes in commercial loans. Their head-
quarters are located in Omaha, Nebraska, and they have several branch offices through-
out rural areas of Nebraska. The top management of COB realized that the current
measurement system is not adequate for their strategic performance management and
identified the following measures based on the hierarchy of the organization for their
new performance measurement using AHP. These measures are shown in Table 29.1.
The OCB uses financial criteria such as return on assets and debt to equity and
nonfinancial criteria such as market share, productivity, and quality of service. At
the lowest management level, income to interest expense, service charges, interest
revenue, growth of deposits, default ratio, and customer satisfaction can be used.
Each computing step of the AHP is discussed as follows.
First, nonfinancial and financial criteria are computed and the result is entered in
a vector:

Wt ¼ ð0:5; 0:5Þ:
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 805

Table 29.1 New performance measures

Level of organization High Middle Low
Financial measures Return on assets Income to interest expenses
Debt to equity Service charges
Interest revenue
Nonfinancial measures Market share Growth of deposits
Productivity Default ratio
Quality Customer satisfaction

Next, the mid-level management is considered. The relative weight of mid-level

performance criteria with relation to each top-level criterion is to be computed.
Here, the local relative weights are computed.
For the nonfinancial criteria,
2 3
1 3 4
A1m 4 1=3 1 3 5:
1=4 1=3 1

Here, the market share is estimated to be three times more important than produc-
tivity and four times more important than the quality of service. Productivity is
estimated to be three times more important than the quality of service. From this result,

W 1m ¼ ð0:608; 0:272; 0:120Þ:

For each weight computation, an inconsistency ratio (g) was computed and
checked for the acceptance level. If g  0.1, it is acceptable. For this example, it
is acceptable since g ¼ 0.065. If it is not acceptable, the input matrix should be
adjusted or recomputed.
For the financial criteria, ROA is estimated to be three times more important than
debt to equity ratio. Therefore,

W 2m ¼ ð0:75; 0:25Þ:

The global relative weights of the criteria are


0:608, 0:272, 0:120, 0, 0
Wm ¼ ð0:5; 0:5Þ 
0, 0, 0, 0:75, 0:25
¼ ð0:304; 0:136; 0:060; 0:375; 0:125Þ:

Here the global relative weights of market share, productivity, quality of service,
ROA, and debt to equity ratio are 30.4 %, 13.6 %, 6 %, 37.5 %, and 12.5 %,
respectively.
806 W. Kwak et al.

Let us move to the lower-level managers. Income to interest expense, service

charges, interest revenue, growth of deposits, default ratio, and customer satisfac-
tion are criteria at this level.
For market share,

2 3
1 2 3 4 5 6
6 1=2 1 2 3 4 57
6 7
6 1=3 1=2 1 3 4 57
A0 ¼ 6
1
6 1=4
7
6 1=3 1=3 1 3 477
4 1=5 1=4 1=4 1=3 1 35
1=6 1=5 1=5 1=4 1=3 1

For simplicity of presentation, the local weights are arranged in the order of
income to interest expense, service charges, interest revenue, growth of deposits,
default ratio, and customer satisfaction:

W 1o ¼ ð0:367; 0:238; 0:183; 0:109; 0:065; 0:038Þ:

Here the local weights with relation to market share are 36.7 %, 23.8 %, 18.3 %,
10.9 %, 6.5 %, and 3.8 %, respectively.
For other criteria at one level higher, the local weights can be calculated in the
same way. These are

W 2o ¼ ð0:206; 0:163; 0:179; 0:162; 0:143; 0:146Þ;

W 3o ¼ ð0:155; 0:231; 0:220; 0:103; 0:169; 0:122Þ;

W 4o ¼ ð0:307; 0:197; 0:167; 0:117; 0:117; 0:094Þ;

W 5o ¼ ð0:266; 0:133; 0:164; 0:159; 0:154; 0:124Þ:

For the next step, the global relative weights of lower-level management
criteria are

0:367 0:238 0:183 0:109 0:065 0:038

2 3
0:206 0:163 0:179 0:162 0:143 0:146
6 0:155 0:231 0:220 0:169 0:122 7
6 0:103 7
W0 ¼ ð0:304; 0:136; 0:060; 0:375; 0:125Þ  6 7
4 0:307 0:197 0:167 0:117 0:117 0:094 5
0:266 0:133 0:164 0:159 0:159 0:124
¼ ð0:297; 0:199; 0:176; 0:125; 0:112; 0:089Þ:
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 807

Finally, the relative importance of lower-level performance measures are 29.7 %,

19.9 %, 17.6 %, 12.54 %, 11.2 %, and 8.9 %, respectively. Note that financial
measures are integrated with nonfinancial measures in the scaling process. The OCB
can extend these performance measures into the next lower level of each product
using the same method.

29.3.4 Summary and Conclusions

A performance measurement system should incorporate nonfinancial as well as

financial measures to foster strategic success factors for a bank in the new envi-
ronment. Generally, a good performance measurement system should monitor
employees’ behavior in a positive way and be flexible enough to adapt to the
changing environment. To motivate employees, a bank should communicate per-
formance information of an individual employee in relation to overall business
goals. This characteristic of performance measurement requires a significant
amount of feedback both between and within levels and corrective actions in the
practice of accounting information (Nanni et al. 1990).
The AHP model proposed in this study is flexible enough to incorporate the
“continuous improvement” philosophy of today’s business environment by chang-
ing weighting values of measures. In addition, the integrated structure of AHP
allows group performance evaluation, which is a buzzword for “teamwork” in
today’s business world. The iterative process of getting input data in the AHP
procedure also helps each manager as well as employee to be aware of the
importance of strategic factors of each performance measure of the bank.

29.4 Optimal Trade-offs of Multiple Factors in International

Transfer Pricing Problems

Ever since DuPont and General Motors Corporation of the USA initiated transfer
pricing systems for the interdivisional transfer of resources among their divisions,
many large organizations, with the creation of profit centers, have used a transfer
pricing system in one way or the other. Recently, transfer pricing problems have
become more important because most corporations increase transfer of goods or
services dramatically among their divisions as a result of restructuring or
downsizing their organizations (Tang 1992). Therefore, designing a good transfer
pricing strategy should be a major concern for both top management and divisional
managers (Curtis 2010).
Transfer pricing problems have been extensively studied by a number of scholars.
Many of them have recognized that a successful transfer pricing strategy should
consider multiple criteria (objectives), such as overall profit, total market share,
divisional autonomy, performance evaluation, and utilized production capacity
(Abdel-khalik and Lusk 1974; Bailey and Boe 1976; Merville and Petty 1978;
Yunker 1983; Lecraw 1985; Lin et al. 1993). However, few developed methods
808 W. Kwak et al.

have a capability of dealing with all possible optimal trade-offs of multiple criteria in
optimal solutions of the models with involvement of multiple decision makers.
In this chapter, we propose a multiple factor model to provide managers from
different background, who are involved in transfer price decision making of
a multidivisional corporation, with a systematic and comprehensive scenario
about all possible optimal transfer prices depending on both multiple-criteria and
multiple-constraint levels (in short, multiple factors). The trade-offs of the optimal
transfer prices, which have rarely been considered in the literature, can be used as
a basis for managers of corporations to make a high-quality decision in selecting
their transfer pricing systems for business competition.
This chapter proceeds as follows. First, existing transfer pricing models will be
reviewed. Then, the methodology of formulating and solving a transfer pricing
model with multiple factors will be described. A prototype of a transfer pricing
problem in a corporation will be illustrated to explain the implications of the
multiple factor transfer pricing model. Finally, conclusions and remaining research
problems will be presented.

29.4.1 Existing Transfer Pricing Models

In the literature of transfer pricing problems, the various approaches can be

categorized into four groups: (i) market-based pricing, (ii) accounting-based pric-
ing, (iii) marginal cost pricing (or opportunity cost pricing), and (iv) negotiation-
based pricing. Market-based prices are ideal for transfer prices when external
market prices are available. Even though empirical research has found that some
corporations prefer cost-based prices to market-based prices, market-based pricing
method is recommended when the emphasis is on the motivation of divisional
managers (Borkowski 1990). Pricing intermediate goods based on the market price
will motivate the supplying division to reduce its costs to achieve efficiency and to
allow divisional autonomy for both the supplying division and the purchasing
division. Statistics show that almost a third of corporations actually use market-
based transfer pricing (Tang 1992).
However, if there is no outside market for intermediate goods or services, then
accounting-based pricing, marginal cost pricing (economic models and mathemat-
ical programming techniques), or negotiation-based pricing (behavioral approach)
is commonly recommended for finding a transfer pricing system.
In the accounting-based pricing approach, the divisional managers simply use
accounting measurements of the divisions, such as full costs or variable costs, as
their transfer prices. Thus, the transfer price of one division may differ from that of
another division. These transfer prices may not be globally optimal for the corpo-
ration as a whole (Abdel-khalik and Lusk 1974; Eccles 1983).
In marginal cost pricing approaches, Hirshleifer (1956) recommended use of an
economic model to set transfer pricing at a manufacturing division’s marginal cost
to achieve the global optimal output. A problem of this economic model is that it
can destroy the divisional manager’s autonomy, and the supplying division may not
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 809

get the benefit of efficiencies. Moreover, the manufacturing division manager in this
case should be evaluated based on cost, not on profit. Similarly, Gould (1964) and
Naert (1973) recommended economic models based on current entry and current exit
prices. Their models also focus on global profit maximization and have the same
problems as Hirshleifer’s. Note that when the transfer price is set based on marginal
costs, the division should be either controlled as a standard cost center or merged into
a larger profit center with a division that processes the bulk of its output.
Ronen and McKinney (1970) suggested dual prices in which a subsidy is given
to the manufacturing division by the central office in addition to marginal costs.
This subsidy would not be added to the price charged to the purchasing division.
They believed that autonomy is enhanced because the corporate office is only
a transmitter of information, not a price setter, and that the supplying division has
the same autonomy as an independent supplier. However, there might be a gaming
chance where all divisions are winners but the central office is a loser. There is also
a “marginal cost plus” approach that charges variable manufacturing costs (usually
standard variable costs) and is supplemented by periodic lump-sum charges to the
transferees to cover the transferor’s fixed costs, or their fixed costs plus profit, or
some kind of subsidy. It is difficult, however, to set a fixed fee that will satisfy both
the supplying division and the purchasing division. This method enables the
purchasing division to absorb all the uncertainties caused by fluctuations in the
marketplace. Moreover, the system begins to break down if the supplying division
is operating at or above the normal capacity since variable costs no longer represent
the opportunity costs of additional transfers of goods and services (Onsi 1970).
As an alternative method to identify marginal costs as the transfer prices,
a mathematical programming approach becomes more attractive for the transfer
pricing problem because it handles complex situations in a trade setting (Dopuch
and Drake 1964; Bapna et al. 2005). The application of linear programming to the
transfer pricing problem is based on the relationship between the primal and dual
solutions in the linear programming problem. Shadow prices, which reflect
the input values of scarce resources (or opportunity cost) implied in the primal
problem, can be used as the basis for a transfer price system. However, these
transfer prices have the following limitations for decision making: (i) those transfer
prices based on dual values of a solution tend to reward divisions with scarce
resources, (ii) the linear formulation requires a great deal of local information, and
(iii) transfer prices based on shadow prices do not provide a guide for performance
evaluation of divisional managers.
Based on the mathematical decomposition algorithm developed by Dantzig and
Wolfe (1960), Baumol and Fabian (1964) demonstrated how to reduce complex
optimization problems into sets of smaller problems solvable by divisions and the
central office (corporation). Although the final analysis of output decisions is made by
the central manager, the calculation process is sufficiently localized that central
management does not have to know anything about the internal technological arrange-
ments of the divisions. However, this approach does not permit divisional autonomy.
Ruefli (1971) proposed a generalized goal decomposition model for incorporating
multiple criteria (objectives) and some behavioral aspects within a three-level
810 W. Kwak et al.

hierarchical organization into the mathematical formulation of transfer pricing

problems. Bailey and Boe (1976) suggested another goal programming model as
a supplement of Ruefli’s model. Both models overcome the shortcoming of linear
programming in dealing with multiple criteria and the organizational hierarchy.
Merville and Petty (1978) used a dual formulation of goal programming to directly
find the shadow price as the optimal transfer price for multinational corporations.
However, the optimal solution of these models that results in the transfer prices is
determined by a particular distance (norm) function because of the mathematical
structure of goal programming. Thus, the optimal solution represents only a single
optimal trade-off of the multiple criteria, not all possible optimal trade-offs.
Watson and Bulmer (1975) and Thomas (1980) criticized the lack of behavioral
considerations in the mathematical programming approaches. They suggested
a negotiated transfer pricing to further the integration of differentiation within the
organization. The differentiation of the organization exists because, for example,
different managers can interpret the same organizational problem differently.
Generally, the negotiated transfer pricing will be successful under conditions
such as the existence of some form of outside markets for the intermediate product,
freedom to buy or sell outside, and sharing all market information among the
negotiators. Negotiated prices may require iterative exchanges of information
with the central office as a part of a mathematical programming algorithm.

29.4.2 Methodology

29.4.2.1 MC2 Linear Programming

Practically speaking, since linear programming has only a single criterion (objective)
and a single resource availability level (right-hand side), it has limitations in handling
real-world transfer pricing problems. For instance, linear programming cannot be
used to solve the problem in which a corporation tries to maximize overall profit and
the total market share simultaneously. This dilemma is overcome by a technique
called multiple-criteria (MC) linear programming (Zeleny 1974; Goicoechea
et al. 1982; Steuer 1986; Yu 1985). To extend the framework of MC linear program-
ming, Seiford and Yu (1979) and Yu (1985) formulated a model of multiple-criteria
and multiple-constraint level (MC2) linear programming. This model is rooted by two
facts. First, from the linear system structure’s point of view, the criteria and constraints
may be “interchangeable.” Thus, like multiple criteria, multiple-constraint (resource
availability) levels can be considered. Second, from the application’s point of view, it
is more realistic to consider multiple resource availability levels (discrete right-hand
sides) than a single resource availability level in isolation. The philosophy behind this
perspective is that the availability of resources can fluctuate depending on the decision
situation forces, such as the desirability levels believed by the different managers. For
example, if the differentiation of budget among managers in transfer pricing problems
(Watson and Baumler 1975) is represented by different levels of budget, then this
differentiation can be resolved by identifying some best compromise of budget levels
as the consensus budget.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 811

The theoretical connections between MC linear programming and MC2 linear

programming can be found in Gyetvan and Shi (1992). Decision problems related to
MC2 linear programming have been extensively studied in Lee et al. (1990), Shi
(1991), and Shi and Yu (1992). Key ideas of MC2 linear programming, a primary
theoretical foundation of this chapter, are outlined as follows.
An MC2 linear programming problem can be formulated as
Max lt Cx
s:t: Ax  Dg
x  0,

where C 2 Rqxn, A 2 Rmxn, and D 2 Rmxp are matrices of qxn, mxn, and mxp
dimensions, respectively; x 2 Rn are decision variables; l 2 Rq is called the criteria
parameter; and g 2 Rp is called the constraint level parameter. Both (g,l) are
assumed unknown.
The above MC2 problem has q criteria (objectives) and p constraint levels. If the
constraint level parameter g is known, then the MC2 problem reduces to an MC
linear programming problem (e.g., Yu and Zeleny 1975). In addition, if the criteria
parameter l is known, it reduces to a linear programming problem (e.g., Charnes
and Cooper 1961; Dantzig 1963).
Denote the index set of the basic variables {xji, . . ., xjm} for the MC2 problem by
J ¼ {j1, . . ., jm}. Note that the basic variables may contain some slack variables.
Without confusion, J is also called a basis for the MC2 problem. Since a basic
solution J depends on parameters (g, l), define that (i) a basic solution J is feasible
for the MC2 problem if and only if there exists a g0 > 0 such that J is a feasible
solution for the MC2 problem with respect to g0 and (ii) J is potentially optimal for
the MC2 problem if and only if there exist a g0 > 0 and a l0 > 0 such that J is an
optimal solution for the MC2 problem with respect to (g0, l0). Let G(J) be the
constraint level parameter set of all g such that the basis J is feasible and L(J) be the
criteria parameter set of all l that the basis J is dual feasible. Then, for a given basis
J of the MC2 problem, (i) J is a feasible solution if and only if the set G(J) is not
empty and (ii) J is potentially optimal if and only if both sets G(J) and L(J) are not
empty. For an MC2 problem, there may exist a number of potentially optimal
solutions {J} as parameters (g, l) vary depending on decision situations.
Seiford and Yu (1979) derived a simplex method to systematically locate the set
of all potentially optimal solutions {J}. The computer software of the simplex
method (called MC2 software) was developed by Chien et al. (1989). This software
consists of five subroutines in each iteration: (i) pivoting, (ii) determining primal
potential bases, (iii) determining dual potential bases, (iv) determining the effective
constraints for the primal weight set, and (v) determining the effective constraints
for the dual weight set (Chap. 8 of Yu 1985). It is written in PASCAL and operates
in a man-machine interactive fashion. The user cannot only view the tableau of each
iteration but also trace the past iterations. In the next section, the framework of
the MC2 problem, as well as its software, will be used to formulate and solve the
multiple factor transfer pricing problems.
812 W. Kwak et al.

29.4.2.2 Multiple Factor Transfer Pricing Model

The review of existing transfer pricing models shows that two major shortcomings,
from a management point of view, need to be overcome in the previous mathemat-
ical models of transfer pricing problems. First, neither the linear programming
approach nor the goal programming approach can provide a comprehensive sce-
nario of all possible optimal trade-offs between multiple objectives under consid-
eration for a given transfer pricing problem, such as maximizing the overall profits
for a corporation and minimizing the underutilization of production capacity (Tang
1992). Transfer pricing scheme by linear programming only reflects a single objec-
tive of a corporation. As a result, the linear programming approach cannot help the
corporation seek to simultaneously achieve several objectives, some in conflict, in
business competition. The transfer price determined by the goal programming
approach is an optimal compromise (i.e., trade-off) among several objectives of
the corporation. However, it misses other possible optimal compromises of the
objectives that result from some linear combinations of objective weights. These
compromises lead to different optimal transfer prices for different decision situa-
tions that the corporation may face. Second, none of the past mathematical models
can deal with the organizational differentiation problems, as Watson and Baumler
(1975) pointed out. In real-life cases, when a corporation designs its transfer prices
for the divisions, the involved decision makers (executives or the members of the
task force) can give different opinions on the same issue, such as production
capacity and customer’s demand. In mathematical models, these different interpre-
tations can be represented by different “constraint levels.” Because both linear
programming and goal programming presume a fixed single constraint level, they
fail to mathematically describe such an organizational differentiation problem.
The MC2 linear programming framework can resolve the above shortcomings
inherent in the previous transfer pricing models. Based on Yunker (1983) and Tang
(1992), the four important objectives of transfer pricing problems in most corpora-
tions are considered: (i) maximizing the overall profit, (ii) maximizing the total
market share, (iii) maximizing the subsidiary profit (note that the subsidiary profit
maximization is used to reflect the degree of the subsidiary autonomy in decision
making. It may differ from the overall profit), and (iv) maximizing the utilized
production capacity. Even though the following model contains only these four
specific objectives, the generality of the modeling process fits in all transfer pricing
problems with multiple-criteria and multiple-constraint levels.
Let k be the number of divisions in a corporation under consideration and t be the
index number of the products that each division of the corporation produces.
Define xij as the units of the jth product made by the ith division, i ¼ 1, . . . , k;
j ¼ 1, . . . , t. For the coefficients of the objectives, let pij be the unit overall profit
generated from the jth product made by the ith division, mij be the market share value
for the jth product made by the ith division in the market, sij be the unit subsidiary
profit generated from the jth product made by the ith division, and cij be the unit
utilized production capacity of the ith division to produce the jth product. For the
coefficients of the constraints, let bij be the budget allocation rate for producing
the jth product by the ith division. For the coefficients of the constraint levels, let bijs
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 813

be the budget availability level believed by the sth manager (or executive) for
producing the jth product by the ith division, s ¼ 1, . . . , h; dis be the production
capacity level believed by the sth manager for the ith division; dijs be the production
capacity level believed by the sth manager for the ith division to produce the
jth product; and eijs be the initial inventory level believed by the sth manager for
the ith division to hold the jth product. Then, the multiple factor transfer pricing
model is
Max Ski¼1 Slj¼1 pij xij
Max Ski¼1 Slj¼1 mij xij
Max Ski¼1 Slj¼1 sij xij
Max Ski¼1 Slj¼1 cij xij


Subject to Ski¼1 Slj¼1 bijxij  b1 ij ; . . . ; bh ij (29.1)


Skj¼1 xij  d1 ij , . . . dh i


xij  d1 ij ; . . . ; dh ij


xij þ xiþ1, j  e1 ij ; . . . ; eh ij
Xij  0, i ¼ 1, . . . , k, j ¼ 1, . . . , t:
In the next section, a prototype of the transfer pricing model in a corporation will
be illustrated to demonstrate the implications for decision makers.

29.4.3 Model Implications

29.4.3.1 Numerical Example

As an illustration of the multiple factor transfer pricing model, the United Chemical
Corporation has two divisions that process raw materials into intermediate or final
products. Division 1, which is located in Kansas City, manufactures two kinds of
chemicals, called Products 1 and 2. Product 1 in Division 1 is intermediate product
and cannot be sold externally. It can, however, be processed further by Division
2 into a final product. Division 2, which is located in Atlanta, manufactures Product
2 and finalizes Product 1 in Division 1. The executives (president, vice president for
production, and vice president for finance) and all divisional managers agree on the
following multiple objectives:
(i) Maximize the overall company’s profit.
(ii) Maximize the market share goal of Product 2 in Division 1 and Products 1 and
2 in Division 2.
(iii) Maximize the utilized production capacity of the company so that each
division manager can avoid any underutilization of normal production
capacity.
The data related to these objectives is given in Table 29.2.
In Table 29.2, Product 1 of Division 1 is a by-product that has no sale value
in Division 1 at all. The unit profit $4 of this product means its production cost.
814 W. Kwak et al.

Table 29.2 Data of the objectives in the United Chemical Corporation

Division 1 Product 1 Product 2 Division 2 Product 1 Product 2
Unit profit ($) 4 8 13 5
Market price 0 40 46.2 38
Unit utilizing production capacity (h) 4 4 3 2

All other products generate profits. We use a market price of each product to
maximize the market share goal. Note that the market prices of three products in
Table 29.2 are different.
In the company, besides the president, the vice presidents for production and for
finance are the main decision makers and may have different interpretations of the
same resource availability across the divisions. The vice president for production
views the constraint level based on the material, manpower, and equipment under
control, while the vice president for finance views the constraint level based on the
available cash flow. The president will make the final decision for the company’s
transfer price setting on the basis of the compromises of both vice presidents. All
divisional managers will carry out the president’s decision, although they have their
autonomy to provide the information about the divisional profits and costs for the
executives. The interpretations of the vice presidents for the resource constraint
levels are summarized in Table 29.3.
All executives and divisional managers agree on the units of resources consumed
to produce the products. The data is given in Table 29.4.
Let xij be the units of the jth product produced by the ith division, i ¼ 1, 2;
j ¼ 1, 2. Using the information in Tables 29.2, 29.3, and 29.4, the multiple factor
transfer pricing model is formulated as
Max  4x11 þ 8x12 þ 13x21 þ 5x22
Max 40x12 þ 46:2x21 þ 38x22
Max 4x11 þ 4x12 þ 3x21 þ 2x22
Subject to  x11 þ x21  ð0; 100Þ
0:4x12 þ 0:4x21 þ 0:4x22  ð45; 000; 40; 000Þ (29.2)
X12  ð38; 000; 12; 000Þ
X21  ð45; 000; 50; 000Þ
X22  ð36; 000; 10; 000Þ
xij  0, i ¼ 1, 2; j ¼ 1, 2:
Since this multiple factor transfer pricing problem is a typical MC2 problem, the
MC2 software of Chien et al. (1989) can be used to solve the problem. Let l ¼ (l1,
l2, l3) be the weight parameter for the objectives, where l1 + l2 + l3 ¼ 1 and l1, l2,
l3  0. Let g ¼ (g1, g2) be the weight parameter for the constraint levels, where g1 +
g2 ¼ 1 and g1, g2  0. Because both weight parameters (g, l) are unknown before
design time, the solution procedure of MC2 linear programming must be used to
locate all possible potentially optimal solutions as (g, l) vary. The implications of
the potentially optimal solutions for accounting decision makers will be explained in
the next subsection. After putting (g, l) into the above model, it becomes
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 815

Table 29.3 The constraint levels of the vice presidents

Vice president for production Vice president for finance
Transfer product constraint 0 100
Budget constraint ($) 45,000 40,000
Production capacity of 38,000 12,000
Product 2 in Division 1
Production capacity of 45,000 50,000
Product 1 in Division 2
Production capacity of 36,000 10,000
Product 2 in Division 2

Table 29.4 The unit consumptions of resources

Division 1 Product 1 Product 2 Division 2 Product 1 Product 2
Transfer product constraint 1.0 0 1.0 0
Budget constraint ($) 0 0.4 0.4 0.4
Production capacity of 1.0
Product 2 in Division 1
Production capacity of 1.0
Product 1 in Division 2
Production capacity of 1.0
Product 2 in Division 2

Max l1 ð4x11 þ 8x12 þ 13x21 þ 5x22 Þ

þ l2 ð40x12 þ 46:2x21 þ 38x22 Þ
þ l3 ð4x11 þ 4x12 þ 3x21 þ 2x22 Þ
Subject to
 x11 þ x21  100g2
(29.3)
0:4x12 þ 0:4x21 þ 0:4x22  45, 000g1 þ 40, 000g2
x12  38, 000g1 þ 12, 000g2
x21  45, 000g1 þ 50, 000g2
x22  36, 000g1 þ 10, 000g2
xij  0, i ¼ 1, 2; j ¼ 1, 2:
Let sq, q ¼ 1, . . . , 5, be the slack variables corresponding to the constraints. The
MC2 software yields two potentially optimal solutions {J1, J2} and their associated
values of (g, l) as shown in Table 29.5. Here, V(Ji) denotes the objective value of
potentially optimal solution Ji, which is a function of (g, l). When (g, l) are
specified, V(Ji) is the payoff of using Ji.
Table 29.5 means that if the g takes the value from G(J1) and the l takes the value
from L(J1), J1 is the optimal solution for the transfer pricing problem. In this case,
products {x11, x12, x21, x22} will be produced to achieve the objective payoff
V(J1) and the resource of x22 has the amount of s5 unused. Similarly, the
potentially optimal solution J2 can be interpreted. However, J2 is different from J1
816 W. Kwak et al.

Table 29.5 All potentially optimal solutions

Ji G(Ji) L(Ji) V(Ji)
J1 ¼ (x11, x12, x21, x22, s5) g1 + g2 ¼ 1, l1 + l2 + l3 ¼ 1, 856,500 736,400 g
65g1  280g2  0, l1  l3, 4,720,000 4,234,000
g1, g2  0 l1, l2, l3  0 526,000 47,444,000
J2 ¼ (x11, x12, x21, x22, s2) g1 + g2 ¼ 1, l1 + l2 + l3 ¼ 1, 889,000 596,400 g
65g1  280g2  0, l1  l3, 4,967,000 3,170,000
g1, g2  0 l1, l2, l3  0 539,000 41,184,000

Table 29.6 All optimal transfer prices

Ji pq(Ji)
J1 ¼ (x11, x12, x21, x22, s5) p1(J1) ¼ 4l1  4l3,
p2(J1) ¼ 12.5l1 + 95l2 + 5l3,
p3(J1) ¼ 3l1 + 2l2 + 2l3
p4(J1) ¼ 4l1 + 8.2l2 + 5l3
p5(J1) ¼ 0
J2 ¼ (x11, x12, x21, x22, s2) p1(J2) ¼ 4l1  4l3,
p2(J2) ¼ 0,
p3(J2) ¼ 8l1 + 40l2 + 4l3
p4(J2) ¼ 9l1 + 46.2l2 + 7l3
p5(J2) ¼ 5l1 + 38l2 + 2l3

because there is the unused budget s2 for J2 and the parameter set G(J1) _ G(J2).
Because Product 1 in Division 1 is a by-product and its unit profit is $4, whenever
l1 < l3, both J1 and J2 are not optimal.
Let pq(Ji), q ¼ 1, . . . , 5; i ¼ 1, 2, be the shadow price of Ji for the qth constraint.
According to the marginal cost pricing approach, the optimal transfer prices of
{J1, J2} are designated as the shadow prices of {J1, J2}. These optimal transfer
prices are found in Table 29.6. Table 29.6 shows that (i) the relative transfer price
between x11 and x21 is p1(J1) ¼ 4 l1  4 l3; (ii) the transfer price for budget across
the divisions is p2(J1) ¼ 12.5 l1 + 95 l2 + 5 l3; (iii) the transfer price for x12 is
p3(J1) ¼ 3 l1 + 2 l2 + 2 l3; (iv) the transfer price for x21 is p4(J1) ¼ 4 l1 + 8.2 l2 +
5 l3; and (v) the transfer price for x22 is p5(J1) ¼ 0.

29.4.4 Optimal Trade-offs and Their Accounting Implications

Optimal trade-offs related to transfer prices of the multiple factor model consist of
three components: (i) trade-offs among multiple objectives, (ii) trade-offs among
multiple-constraint levels, and (iii) trade-offs between multiple objectives and
multiple-constraint levels. The trade-offs among multiple objectives imply
that all possible optimal compromises of the multiple objectives are determined
by locating all possible weights of importance of these objectives. Similarly,
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 817

Fig. 29.1 Optimal trade-offs

between profit (l1) and λ3
production capacity (l3)
1
λ1 = λ3

J1

J2

0 1 λ1

the trade-offs among multiple-constraint levels imply that all possible

optimal compromises of the multiple-constraint levels that represent the
executives’ different opinions are determined by locating all possible weights of
importance of these opinions. The trade-offs between multiple objectives and
multiple-constraint levels measure the effects on the transfer pricing problem by
the interaction of the objectives and constraint levels. These three cases of the
optimal trade-offs can be analyzed through the potentially optimal solutions of
the multiple factor model. The optimal transfer prices also result from those
solutions. In the following, three trade-off cases and accounting implications of
the corresponding optimal transfer prices are explored in detail by using the above
numerical example.
Because there are two potentially optimal solutions {J1, J2} in the example, the
optimal trade-offs should be studied in terms of both J1 and J2. For the trade-offs
among three objectives that are the overall company’s profit, the market share of the
company, and the utilized production capacity of the company:
(i) If the overall company’s profit is not considered (this implies l1 ¼ 0 and l3 6¼ 0),
then either J1 or J2 is not optimal since l3 < 0 (see Table 29.5).
(ii) If the market share of the company is not considered (i.e., l2 ¼ 0), then both J1 and
J2 are optimal for which l1 + l3 ¼ 1, l1  l3, and l1, l3  0. The graphical
representation is shown in Fig. 29.1. From Fig. 29.1, any weighting values of l1
and l3 taken from the feasible (dark) segment guarantee that the utility values of
both the overall company’s profit and the utilized production capacity of
the company are maximized. The resulting optimal transfer prices associated
with J1 are p1(J1) ¼ 4l1  4l3, p2(J1) ¼ 12.5l1 + 5l3, p3(J1) ¼ 3l1 + 2l3,
p4(J1) ¼ 4l1 + 5l3, and p5(J1) ¼ 0, respectively. The resulting optimal transfer
818 W. Kwak et al.

Fig. 29.2 Optimal trade-offs

between profit (l1) and
λ2
market share (l2)
1
J1

J2

0 1 λ1

prices associated with J2 are p1(J2) ¼ 4l1  4l3, p2(J2) ¼ 0, p3(J2) ¼ 8l1 + 4l3,
p4(J2) ¼ 9l1 + 7l3, and p5(J2) ¼ 5l1 + 2l3, respectively. The weighting values of
l1 and l3 in the transfer prices for J1 and J2 are given in Fig. 29.1, because
L(J1) ¼ L(J2).
(iii) If the utilized production capacity of the company is not considered (i.e., l3 ¼ 0),
then both J1 and J2 are optimal for which l1 + l2 ¼ 1 and l1, l2 _ 0, where the
graphical representation is shown in Fig. 29.2. Figure 29.2 implies that any
weighted combination of l1 and l2 taken from the indicated feasible segment
maximizes the utility values of both the overall company’s profit and the market
share of the company. The resulting optimal transfer prices of J1 are
p1(J1) ¼ 4l1, p2(J1) ¼ 12.5l1 + 95l2, p3(J1) ¼ 3l1 + 2l2, p4(J1) ¼ 4l1 +
8.2l2, and p5(J1) ¼ 0, while the resulting optimal transfer prices of J2 are
p1(J2) ¼ 4l1, p2(J2) ¼ 0, p3(J2) ¼ 8l1 + 40l2, p4(J2) ¼ 9l1 + 46.2l2, and
p5(J2) ¼ 5l1 + 38l2, respectively.
For the trade-offs among two constraint levels (the different opinions of two vice
presidents), the weight of the vice president for production is g1 while that of the vice
president for finance is g2 such that g1 + g2 ¼ 1, g1, g2  0. The range of g1 and g2 are
decomposed into two subsets: G(J1) ¼ {g1, g2  0 | 65g1  280g2  0 and g1 + g2 ¼ 1}
and G(J2) ¼ {g1, g2  0 | 65g1 + 280g2  0 and g1 + g2 ¼ 1} (see Table 29.5). The
graphical representation of G(J1) and G(J2) is shown in Fig. 29.3. Whenever the
weighting values of g1 and g2 are taken from G(J1), the corresponding compromise
of two vice presidents will result in the optimal transfer prices of J1 in Table 29.6.
Similarly, the decision situation of constraint levels for J2 can be explained.
Finally, there are many optimal trade-off situations between three objectives and
two constraint levels involved with the transfer pricing problem. For example, two
cases are illustrated as follows:
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 819

Fig. 29.3 Optimal trade-offs

between V.P. for production γ2
(g1) and V.P. for finance (g2)
1
J2

65γ1 = 280γ2

J1

0 1 γ1

λ1

J2 J1

Fig. 29.4 Optimal trade-offs

V.P. for production (g1) and
0 .81 1 γ1
total profit (l1)

(i) In the case that the utilized production capacity of the company is not considered
(i.e., l3 ¼ 0) in contrast to the weighting value of the market share and that of the
vice president for finance (i.e., 0  l2  1 and 0  g2  1), the optimal trade-offs
between the overall company’s profit and the constraint level believed by the vice
president for production for J1 and J2 are shown in Fig. 29.4. Here, if the values of
(g1, l1) are taken from 0  l1  1 to .81  g1  1, then the optimal transfer prices
820 W. Kwak et al.

Fig. 29.5 Optimal trade-offs

between V.P. for production λ3
(g1) and production
capacity (l3) 1

.5

J2 J1

0 .81 1 γ1

associated with J1 will be chosen. They are p1(J1) ¼ 4l1, p2(J1) ¼ 12.5l1 + 95l2,
p3(J1) ¼ 3l1 + 2l2, p4(J1) ¼ 4l1 + 8.2l2, and p5(J1) ¼ 0. Otherwise, the values
of (g1, l1) are taken from 0  l1  1 to 0  g1  .81, and the optimal
transfer prices associated with J2, p1(J2) ¼ 4l1, p2(J2) ¼ 0, p3(J2) ¼ 8l1 + 40l2,
p4(J2) ¼ 9l1 + 46.2l2, and p5(J2) ¼ 5l1 + 38l2 will be chosen.
(ii) In the case that the weighting value of the overall company’s profit is fixed at .5
(i.e., l1 ¼ .5) and the weighting value of the market share and that of the vice
president for finance are any of 0  l2  1 and 0  g2  1, respectively, Fig. 29.5
shows the optimal trade-offs between the utilized production capacity of the
company and the constraint level believed by the vice president for production
for J1 and J2. In Fig. 29.5, if the values of (g1, l3) are taken from 0  l3  .5 to
.81  g1  1, then the optimal transfer prices associated with J1 are p1(J1) ¼ 2 
4l3, p2(J1) ¼ 6.25 + 95l2 + 5l3, p3(J1) ¼ 1.5 + 2l2 + 2l3, p4(J1) ¼ 2 + 8.2l2 +
5l3, and p5(J1) ¼ 0, respectively. If the values of (g1, l3) are taken from 0  l3 
.5 to 0  g1  .81, then the optimal transfer prices associated with J2 are
p1(J2) ¼ 2  4l3, p2(J2) ¼ 0, p3(J2) ¼ 4 + 40l2 + 4l3, p4(J2) ¼ 4.5 + 46.2l2 +
7l3, and p5(J2) ¼ 2.5 + 38l2 + 2l3, respectively. Note that when the weighting
values fall in the range of .5  l3  1 and 0  g1  1, there is not any optimal
trade-off because both J1 and J2 are not optimal solutions (recall that this is caused
by the by-product 1 in Division 1 that has -$4 as the unit profit).
It is worth noting some important implications for accounting decision makers
from the above trade-off analysis. First, the multiple factor transfer pricing model
has a capability of systematically locating all possible optimal transfer prices
through the optimal trade-offs of multiple objectives and multiple-constraint levels.
Since the set of all possible optimal transfer prices found by this model describes
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 821

every possible decision situation within the model framework, the optimal transfer
price obtained from either linear programming or goal programming models is
included in the subset as a special case.
Second, the multiple factor transfer pricing model can be applied to solve
transfer pricing problems not only with complex structure but also with some
organizational behavior contents, such as the organizational differentiation (see
Table 29.3 for different constraint levels). Consequently, this model can foster more
autonomous flexibility than any other mathematical programming model by
allowing central management or local managers to express their own preference
structures and weights.
Third, the proposed method facilitates decision makers’ participation that may
make a positive management achievement of organizational goals (Locke
et al. 1981). The model can aid coordination and synthesis of multiple conflicting
views. This may be quite effective in a transfer pricing situation in which many
objectives are contradictory to each other and these objectives are measured
differently by a number of decision participants.
Fourth, in most multiple-criteria solution techniques, including the goal pro-
gramming approach, if the decision makers are not satisfied with the optimal
solution obtained by using their preferred weights of importance for the multiple
objectives, then an iterative process has to be conducted for incorporating the
decision makers’ new preference on the weights and finding new solutions until
decision makers are satisfied. However, the proposed model eliminates such a time-
and cost-consuming iterative process since it already considers all possible optimal
solutions with respect to the changes of parameter (g, l). Whenever decision
makers want to change their preference on the weights, the corresponding optimal
transfer prices can be immediately identified from the results like in Table 29.6.
This model, in turn, allows the performance evaluation of optimal transfer prices.
Finally, the optimal transfer prices obtained by the multiple factor model have
a twofold significance in terms of decision characteristics. If the problem is viewed as
a deterministic decision problem, whenever the preferred weighting values of objec-
tives and constraint levels are known, the resulting optimal solutions can be identified
from the potentially optimal solutions of the model. Then, the corresponding optimal
transfer prices can be adopted to handle the business situation (recall the above trade-
off analysis). If the problem is viewed as a probabilistic decision problem, it involves
the assessment of the likelihood of (g, l) to occur at the various points of the range.
With proper assumptions, the uncertainty may be represented by random variables
with some known probability distribution. A number of known criteria such as
maximizing expected payoff, minimizing the variance of the payoff, maximin
payoff, maximizing the probability of achieving a targeted payoff, stochastic dom-
inance, probability dominance, and mean-variance dominance can be used to choose
the optimal transfer prices (see Shi 1991). In summary, the multiple factor transfer
pricing model fosters flexibility in designing the optimal transfer prices for the
corporation to cope with all possible changes of business competition. This model
is more likely to be a better aid for executives or managers to understand and deal
with their current or future transfer pricing problems.
822 W. Kwak et al.

29.4.5 Conclusions

A multiple factor transfer pricing model has been developed to solve the transfer
pricing problems in a multidivisional corporation. This model can provide
a systematic and comprehensive scenario about all possible optimal transfer prices
depending on multiple-criteria and multiple-constraint levels. The trade-offs of
optimal transfer prices offer a broad basis for managers of a corporation to flexibly
implement the optimal transfer pricing strategy and cope with various business
situations. Furthermore, this method also aids global optimization, division auton-
omy, and performance evaluation.
There are some research problems remaining to be explored. From a practical
point of view, the framework of this model can be applied to other accounting areas
such as capital budgeting, cost allocation, audit sampling objectives, and personnel
planning for an audit corporation if the decision variables and formulation are
expressed appropriately. From a theoretical point of view, the decomposition
algorithm of linear programming (Dantzig and Wolfe 1960) can be incorporated
into the MC2-simplex method (Seiford and Yu 1979) to sharpen the multiple factor
transfer pricing model’s capability of solving the large-scale transfer pricing prob-
lem. Thus, this incorporation may result in the development of more effective
solution procedures. How to incorporate other trade-off techniques, such as the
satisficing trade-off method (Nakayama 1994), into the MC2 framework for design-
ing optimal transfer pricing strategies is another interesting research problem.

29.5 Capital Budgeting with Multiple Criteria and Multiple

Decision Makers

Capital budgeting is not a trivial task if a firm is to maintain competitive advantages

by adopting new information or manufacturing systems. A firm may implement
innovative accounting systems such as activity-based costing (ABC) or balanced
scorecard to generate more useful information for better economic decision
making in the ever-changing business environment. ABC can provide value-adding
and non-value-adding activity information about new capital investments. Invest-
ment justification in the new manufacturing environment, however, requires a
comprehensive decision-making process that involves competitive analysis, overall
firm strategy, and evaluation of uncertain cash flows (Howell and Schwartz 1994).
The challenge here is to measure cash flows as well as intangible benefits that these
new systems will bring. Furthermore, conflicts of goals, limited resources, and
uncertain risk factors may complicate the capital budgeting problem (see Hillier
(1963), Lee (1993), Karanovic et al. (2010) for details). These problems of conflicts
of goals among decision makers and limited resources in a typical organization
support the use of multiple-criteria and multiple-constraint levels (MC2) linear
programming (Seiford and Yu 1979).
While traditional techniques such as payback or accounting rate of return are
used as a secondary method, discounted cash flow (DCF) methods, including net
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 823

present value (NPV) and internal rate of return (IRR), are the primary quantitative
methods in capital budgeting (Kim and Farragher 1981). The payback method
estimates how long it will take to recover the original investment. However,
this method incorporates neither the cash flows after the payback period nor the
variability of those cash flows (Boardman et al. 1982). The accounting rate of
return method measures a return on the original cost of the investment. Both of
the above methods ignore the time value of money. DCF methods may not
be adequate to evaluate new manufacturing or information systems, because of
a bias in favor of short-term investments with quantifiable benefits (Mensah and
Miranti 1989).
A current trend in capital budgeting methods utilizes mathematical program-
ming and higher discount rates to incorporate higher risk factors (Pike 1983;
Palliam 2005). Hillier (1963) and Huang (2008) suggested useful ways to eval-
uate risky investments by estimating expected values and standard deviations of
net cash flows for each alternative investment. They showed that the standard
deviation of cash flows is easily obtainable. With this information, a complete
description of the risky investment is possible via probability distribution of the
IRR, NPV, or annual cost of the proposed investment under the assumption of the
net cash flows from the investment, which are normally distributed. Similarly,
Turney (1990) suggested a stochastic dynamic adjustment model to incorporate
greater risk premiums when significant additional funds are required in multiple
time periods. Lin (1993) also proposed a multiple-criteria capital budgeting
model under risk. His model used chance constraints of uncertain cash flows
and accounting earnings as risk factors. Pike (1988) empirically tested the
correlation between sophisticated capital budgeting techniques and decision-
making effectiveness and found that management believed that sophisticated
investment techniques improve effectiveness in the evaluation and control of
large capital projects. Weingartner (1963) introduced a mathematical program-
ming approach in the capital budgeting problem. Other researchers have extended
Weingartner’s work with different directions (e.g., Baumol and Quandt 1965;
Bernard 1969; Howe and Patterson 1985). The development of chance-
constrained programming (CCP) by Charnes and Cooper (1961) also enriched
with applications of mathematical programming models in the capital budgeting
problem. They also developed an approximation solution method to the CCP with
zero-one variables using a linear constraint.
A typical mathematical capital budgeting approach maximizes DCFs that mea-
sure a project’s desirability on the basis of its expected net present value as
a primary goal. DCF analysis, however, ignores strategic factors such as future
growth opportunities (Cheng 1993). Furthermore, management can change their
plans if operating conditions change. For example, they can change input and
output mixes or abandon the project in a multi-period situation. The increasing
involvement of stakeholders, other than shareholders, in a business organization
supports a multiple-objective approach (Bhaskar 1979). Other empirical
studies also found that firms used multiple criteria in their capital budgeting
problems (e.g., Bhaskar and McNamee 1983; Thanassoulis 1985). The goal
824 W. Kwak et al.

programming approach has been used to handle multiple-objective problems. It

emphasizes weights of importance of the multiple objectives with respect to the
decision maker’s (DM) preference (Bhaskar 1979; Deckro et al. 1985). Within the
framework of hierarchical goal optimization, several goal programming models
have been suggested to unravel multiple-objective capital budgeting problems (e.g.,
Ignizio 1976; Lee and Lerro 1974). Similarly, Santhanam et al. (1989) used a zero-
one goal programming approach for information system project selection, but their
article lacks a multiple time horizon (learning curve) effect. Choi and Levary
(1989) investigated the use of a chance-constrained goal programming
approach to reflect multiple goals for a multinational capital budgeting problem.
Reeves and Hedin (1993) suggested interactive goal programming (IGP) which
allows more flexibility for DMs in considering trade-offs and adjusting goal target
levels. However, because of the difficulties of measuring the preferences and
priorities of decision makers, Reeves and Franz (1985) developed a simplified
interactive multiple-objective linear programming (SIMOLP) method which uses
an interactive procedure for a DM to identify a preferred solution and Gonzalez
et al. (1987) applied this procedure in a capital budgeting problem (see Corner
et al. (1993) and Reeves et al. (1988) for other examples). Interactive multiple-
objective techniques such as SIMOLP or IGP can reduce the difficulty of
the solution process. Such studies suggest that most capital budgeting problems
require the analysis of multiple criteria that better reflect a real-world situation.
Thanassoulis (1985) addressed multiple objectives such as the maximization
of shareholder wealth, maximization of firm growth, minimization of financial
risk, maximization of the firm liquidity, and minimization of environmental
pollution.
However, these existing multiple-criteria approaches, including goal program-
ming, implicitly assume that there is only one decision maker setting up the
constraint (budget availability) level of a capital budgeting problem. This assump-
tion is not realistic because in most real-world capital budgeting problems, such as
constructing a major highway or shopping mall, multiple decision makers
must involve the decision of the constraint levels (see Sect. 29.5.2 for detailed
discussion). To remove this assumption of a single decision maker from models of
capital budgeting with multiple criteria, this article attempts to incorporate multiple
decision makers’ preferences using (1) the analytic hierarchy process (AHP)
approach and (2) MC2 linear programming in a capital budgeting situation.
Our approach shows its strength in quantifying strategic and nonfinancial factors
that are important in the current competitive business environment. The problem
of incommensurable units in the selection criteria, because of nonfinancial
and qualitative measures, can be resolved by using the AHP approach. The MC2
approach fosters modeling flexibility by incorporating decision makers’ prefer-
ences as multiple-constraint levels.
The rest of the article is organized as follows. Sect. 29.2 introduces the AHP
and MC2 framework. Sect. 29.3 demonstrates the managerial significance and
implications of capital budgeting problems by illustrating an example. Sect. 29.4
concludes the article with several future research avenues.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 825

29.5.1 AHP and MC2 Framework

In general, an organization has limited resources. Furthermore, each manager’s risk

assessments and preferences about a new project may be different. For example,
a financial manager may think that his or her company has only 50 million dollars
available for a new project and he will not approve this project unless he is sure
about substantial financial benefits. In contrast, a production manager may think his
company should have at least 60 million dollars to implement just-in-time (TIT)
manufacturing. The production manager strongly believes that this new system will
produce high-quality products with lower costs to maintain competitive advantages
against competitors. It is clear that there is a conflict of interests between managers.
Even top management’s goal may be different from other managers’. To increase
other managers’ involvement and motivation, top management should induce other
managers’ inputs. There must be trade-offs between goals of different managers in
this group decision-making process. AHP and MC2 linear programming can be used
to resolve this type of group dilemma.

29.5.1.1 Analytical Hierarchy Process (AHP)

AHP is a practical measurement technique that has been widely applied in model-
ing the human judgment process (Saaty 1980). AHP enables decision makers to
structure a complex problem in the form of a hierarchy of its elements according to
an organization’s structure or ranks of management levels. It captures managerial
decision preferences through a series of comparisons of relevant criteria. This
feature of the AHP minimizes the risk of inconsistent decisions due to incommen-
surable units in the selection criteria. Recently, AHP has been applied to several
accounting problems (e.g., capital budgeting (Liberatore et al. 1992), real estate
investment (Kamath and Khaksari 1991), and municipal government capital invest-
ment (Chan 2004)).
The preference of each manager may be analyzed through the use of AHP or
multiple attribute utility technique (MAUT). Both AHP and MAUT have their own
strengths and weaknesses. For a recent debate regarding the two methods, readers
are referred to Dyer and Forman (1991). In this article, the AHP is employed mainly
due to its capability of reducing computational complexity and availability of
software.

29.5.1.2 MC2 Linear Programming

Linear programming has been applied to capital budgeting problems such as
maximizing NPV with a single objective. However, the linear programming
approach has limitations in handling multiple conflicting real-world goals. For
instance, linear programming cannot solve the problem in which a firm ties to
maximize overall profits and total market share simultaneously. This dilemma is
overcome by a technique called multiple-criteria (MC) linear programming
(Goicoechea et al. 1982; Steuer 1986; Yu 1985; Zeleny 1974). To extend the
framework of MC linear programming, Sieford and Yu (1979) and Yu (1985)
formulated a model of MC2 linear programming. This model is based on two
826 W. Kwak et al.

premises. First, from the linear system structure’s perspective, the criteria and
constraint levels may be interchangeable. Thus, like multiple criteria, multiple-
constraint (resource availability) levels can be considered. Second, from the appli-
cation’s perspective, it is more realistic to consider multiple resource availability
levels (discrete right-hand sides) than a single resource availability level in isola-
tion. This recognizes that the availability of resources can fluctuate depending on
the decision situation forces, such as the preferences of the different managers. The
concept of multiple resource levels corresponds to the typical characteristic of
capital budgeting situations where the decision-making process should reflect
each manager’s preference on the new project. A theoretical connection between
MC linear programming and MC2 linear programming can be found in Gyetvan and
Shi (1992) and decision problems related to MC2 linear programming have been
extensively studied (see, e.g., Lee et al. (1990), Shi (1991), and Shi and Yu (1992)).
Key ideas of MC2 linear programming, as a primary theoretical foundation of this
article, are described as follows.
An MC2 linear programming problem can be formulated as

max lt Cx

s:t: Ax  Dg

x0

where C 2 Rqxn, A 2 Rmxn, and D 2 Rmxp are matrices of qxn, mxn, and mxp
dimensions, respectively; x 2 Rn are decision variables; l 2 Rq is called the criteria
parameter; and g 2 Rp is called the constraint level parameter. Both (g, l) are
assumed unknown.
The above MC2 problem has q criteria (objectives) and p constraint levels. If the
constraint level parameter g is known, then the MC2 problem reduces to an MC
linear programming problem (e.g., Yu and Zeleny 1975). In addition, if the criteria
parameter l is known, it reduces to a linear programming problem (e.g., Charnes
and Cooper 1961; Dantzig 1963).
Denote the index set of the basic variables {xj1, . . . , xjm} for the MC2 problem by
J ¼ {j1, . . . , jm}. Note that the basic variables may contain some slack variables.
Without confusion, J is also called a basis for the MC2 problem. Since a basic
solution J depends on parameters (g, l), define that (1) a basic solution J is feasible
for the MC2 problem if and only if there exists a g0 > 0 such that J is a feasible
solution for the MC2 problem with respect to g0 and (2) J is potentially optimal for
the MC2 problem if and only if there exist a g0 > 0 and a l0 > 0 such that J is an
optimal solution for the MC2 problem with respect to (g0, l0). For an MC2 problem,
there may exist a number of potentially optimal solutions {J} as parameters (g, l)
vary depending on decision situations. Seiford and Yu (1979) derived a simplex
method to systematically locate the set of all potentially optimal solutions {J}.
In summary, a model within the framework of AHP and MC2 is proposed
to formulate and solve the multiple-objective capital budgeting problems with
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 827

multiple decision makers as follows. Note that because of complexity of the

problem, we employed AHP to derive weights of l and g. This solution procedure
decreases computational complexity.

29.5.1.3 A Model of MC2 Decision-Making Capital Budgeting

The linear programming approach has limitations for solving real-world problems
with multiple objectives. The goal programming or MOLP approach provides only
one optimal trade-off among several objectives of the firm. None of the past
mathematical models supports multiple decision makers in a capital budgeting
decision-making process. As discussed before, in a real-world situation, the deci-
sion makers (DMs) can have different opinions on the same issue. These different
interpretations can be represented by different constraint levels in mathematical
models. Because both linear programming and goal programming presume a fixed
single constraint level, they fail to mathematically describe such an organizational
differentiation problem.
The MC2 linear programming framework can resolve the above shortcomings in
current capital budgeting models. The framework is flexible enough to include any
objectives depending on problem situation. However, for the sake of clear presen-
tation, we address four groups of objectives that are common in capital budgeting:
(1) maximization of net present value, (2) profitability, (3) growth, and (4) flexibility
of financing. Some group objectives may have multi-time periods. Net present
value measures the expected net monetary gain or loss from a project by
discounting all expected future cash flows to the present point in time, using the
desired rate of return. If we want to incorporate risk factors, we can estimate the
standard deviation of net cash flows as Hillier (1963) suggested. Profitability can be
measured by return on investment (ROI). Growth can be measured by sales growth
or a nonfinancial measure of market share growth, each of which focuses on the
long-term success of a firm. Flexibility of financing (leverage) can be measured by
the debt to equity ratio. As a firm’s debt to equity ratio goes up, the firm’s cost of
borrowing becomes more expensive. Even though our model contains only these
four specific objective groups, the flexibility of the modeling process fits in all
capital budgeting problems with multiple-criteria and multiple-constraint levels.
This model integrates AHP with a mathematical model. The strengths of such an
integration have been shown in several areas (e.g., advertisement media choice
(Dyer and Forman 1991), R&D portfolio selection (Suh et al. 1993), and telecom-
munication hub design (Lee et al. 1995)). The use of both AHP and MC2, as we will
explore in this article, is the first in the decision-making literature.
In general, let i be the time period for a firm to consider capital budgeting and
j be the index number of the projects that the firm can select.
Define xj as the jth project that can be selected by the firm, j ¼ 1, . . . , t. For the
coefficients of the objectives, let vj be the net present value generated by the jth
project; gij be the net sales increase generated by the jth project in the ith time
period to measure sales growth, i ¼ 1, . . . , s; rij be the ROI generated from the jth
project in the ith time period; and lij be the firm’s equity to debt ratio in the ith time
period after jth project is selected to measure liquidity. Here, the optimum debt to
828 W. Kwak et al.

equity ratio for the firm is assumed to be 1 and the inverse of debt to equity ratio is
used to measure leverage. For the coefficients of the constraints, let bij be the cash
outlay required by the jth project in the ith time period. For the coefficients of the
constraint levels, let Cki be the budget availability level believed by the kth manager
(or executive) for the firm in the ith time period, k ¼ 1, . . . , u.
In this article, for illustration, if we treat t ¼ 10, i ¼ 5, and u ¼ 5, then the
model is
X
10
max vj Xj
j¼1
X
10
max gij Xj 8i ¼ 1, . . . , 5
j¼1
X
10
max r ij Xj 8i ¼ 1, . . . , 5
j¼1
X
10
max lij Xj 8i ¼ 1, . . . , 5
j¼1
X
10
subject to bij Xj  ðcli ; . . . ; c5i Þ 8i ¼ 1, . . . , 5
j¼1
and Xj ¼ f0; 1g

where k ¼ 1, . . . , 5, for cki, represents the five possible DMs (i.e., president,
controller, production manager, marketing manager, and engineering manager).
The weights of 16 total objectives can be expressed by (l1, . . . , l16) with
Slq ¼ 1, q ¼ 1, . . . , 16, 0 < lq < 1, and the weights of five DMs can be expressed
by (l1, . . . , l5) with Sgk ¼ 1, k ¼ 1, . . . , 5, 0 < gk < 1.
The above model is an integer MC2 problem with 16 objectives and five
constraint levels. Solving this problem by using a currently available solution
technique (Seiford and Yu 1979) is not a trivial task because of its substantial
computational complexity. To overcome this difficulty, we propose a two-phased
solution procedure as depicted in Fig. 29.6.
The first phase is referred to as the AHP phase in the sense that AHP is applied to
derive weights of l and g, which reduces the model’s complexity. The computation
of weights is effectively handled by Expert Choice (Forman et al. 1985), commer-
cial software for AHP. In this phase, we first induce preferences about objectives
from all five DMs involved. Note that each DM may have different preferences
about the four goals. The president may think maximization of ROI is
the most important goal, while the controller may think maximization of NPV is
the most important goal. AHP generates relative weights for each goal. Then,
multiple-constraint levels are incorporated using the MC2 framework. For instance,
DMs may estimate different budget availability levels for each time period. The
controller may believe that $40 million is available for the first year, while the
production manager may believe that the company can spend $50 million for this
time period. This scenario is more realistic if we consider the characteristics of
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 829

Fig. 29.6 A two-phased Initial Data

capital budgeting model

AHP

Weights of
λ and γ

HC2
Framework IP

Capital
Budgeting

No
Accept ?

Yes

Final Solution

a capital budgeting problem from the perspective of the different utility functions or
risk factors of each DM. An example using AHP to generate relative weights for the
five DMs’ constraint levels (preferences) can be found in Appendix 1.
The second phase is called the integer programming (IP) phase. After applying
the AHP phase to the model, the model is reduced to a linear IP problem, in
which the potentially optimal projects must be selected. To solve this problem,
we employ the ZOOM software (Singhal et al. 1989) that was originally introduced
for solving zero-one integer linear programming problems.

29.5.2 Model Implications

29.5.2.1 Numerical Example

A prototype of our capital budgeting model demonstrates the implications for
multiple criteria and multiple decision makers. We use a Lorie-Savage (1955)
type of problem as follows. The Lorie-Savage Corporation has ten projects under
consideration. All projects have 5-year tune periods.
The executives (president, controller, production manager, marketing manager,
and engineering manager) agree on the fallowing multiple objectives:
1. Maximize net present value of each project. Net present value is computed as the
sum of all the discounted, estimated future cash flows, using the desired rate of
return, minus the initial investment.
830 W. Kwak et al.

Table 29.7 Net present value data for the Lorie-Savage Corporation (in millions)
Cash outlays for each period
Project Net present value Period 1 Period 2 Period 3 Period 4 Period 5
1 $20 $24 $16 $6 $6 $8
2 16 12 10 2 5 4
3 11 8 6 6 6 4
4 4 6 4 7 5 3
5 4 1 6 9 2 3
6 18 18 18 20 15 15
7 7 13 8 10 8 8
8 19 14 8 12 10 8
9 24 16 20 24 16 16
10 4 4 6 8 6 4

2. Maximize the growth of the firm due to each project. Growth can be measured by net
sales increase from each project. This can measure a strategic success factor of a firm.
3. Maximize the profitability of the company. Profitability can be measured by ROI
for each time period generated by each project.
4. Maximize the flexibility of financing. The flexibility of financing (leverage) can
be measured by debt to equity ratio. If a firm’s debt to equity ratio is higher than
the industry average or the optimum level, the firm’s cost of debt financing will
become more expensive. In this example, we use the inverse of the debt to equity
ratio and assume 1 as the optimum debt to equity ratio.
The data related to these objectives is given in Table 29.7. In this example, all
projects are assumed to be independent. In Table 29.7, the firm’s cost of capital is
assumed to be known a priori and to be independent of the investment decisions.
Based on these assumptions, the net present value of each project can be defined as
the sum of the cash flows discounted by the cost of capital. Cash outlay is the
amount of expenditure required for project j, j ¼ 1, 2, 3, . . . , 10, in each time period.
To measure growth, the net sales increase for each time period for each project is
estimated. These data are provided in Table 29.8.
To measure the profitability of each project, ROI is estimated after reflecting
additional income and capital expenditures from each investment for each time
period. These data are provided in Table 29.9.
To measure leverage, the inverse of debt to equity ratio is used after adopting
each project. Here, the optimum debt to equity ratio is assumed to be one for the
Lorie-Savage Corporation. These data are provided in Table 29.10.
The five key DMs in this company (president, controller, production manager,
marketing manager, and engineering manager) have different beliefs regarding
resource availability. For example, for budget availability levels, each DM may
have a different opinion. Of course, the president will make the final decision based
on the opinions of other managers. However, the DMs’ preferences of collection
process should improve the quality of the final decision. Budget availability level
data are provided in Table 29.11.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 831

Table 29.8 Net sales increase data for the Lorie-Savage Corporation (in millions)
Net sales increase for each project
Project Period 1 Period 2 Period 3 Period 4 Period 5
1 $120 $130 $145 $150 $170
2 100 120 140 150 160
3 80 90 95 95 100
4 40 50 50 55 60
5 40 45 50 55 60
6 110 120 140 150 165
7 60 70 65 70 80
8 110 120 100 110 120
9 150 170 180 190 200
10 35 40 40 50 50

Table 29.9 Return on investment data for the Lorie-Savage Corporation (in percentage)
Return on investment for each project
Project Period 1 Period 2 Period 3 Period 4 Period 5
1 10 12 14 15 17
2 10 12 18 16 17
3 12 15 15 15 18
4 8 15 10 8 12
5 15 10 8 20 18
6 12 12 10 15 15
7 8 12 10 12 12
8 14 16 13 15 16
9 12 10 9 12 12
10 10 8 9 8 12

Table 29.10 Debt to equity data for the Lorie-Savage Corporation

Inverse of debt to equity ratio
Project Period 1 Period 2 Period 3 Period 4 Period 5
1 0.85 0.90 0.98 0.95 0.95
2 0.90 0.98 0.95 0.95 0.96
3 0.96 0.97 0.98 0.92 0.92
4 0.98 0.95 0.96 0.92 0.95
5 0.90 0.95 0.95 0.98 0.95
6 0.90 0.95 0.94 0.95 0.95
7 0.96 0.98 0.98 0.98 0.98
8 0.96 0.95 0.90 0.92 0.95
9 0.90 0.88 0.85 0.95 0.95
10 0.98 0.95 0.95 0.98 0.95
832 W. Kwak et al.

Table 29.11 Budget availability level data for the Lorie-Savage Corporation (in millions)
Estimate budget availability
Decision maker Period 1 Period 2 Period 3 Period 4 Period 5
President $50 $30 $30 $35 $30
Controller 40 45 30 30 20
Production manager 55 40 20 30 35
Marketing manager 45 30 40 45 30
Engineering 50 40 45 30 35

The parameters for budget availability levels are derived by using AHP. Here,
one level of an AHP is used. The aggregated objective function also can be obtained
by using AHP. The difference here is that two levels of an AHP must be used to
provide preferences for the objective functions. The first level of this AHP corre-
sponds to the four groups of objectives. The second level corresponds to time
periods within each objective group. Note that the second level is not required for
the objective of maximizing the net present value. Hence, four pairwise comparison
matrices are required for eigenvalue computation. More detailed information about
the formulation of this problem is presented in the appendix. In sum, the objective
function aggregated from the 16 objectives is written as

Maximize 56:34x1 þ 51:74x2 þ 40:54x3 þ 25:43x4 þ 25:44x5

þ 53:63x6 þ 31:98x7 þ 50:59x8 þ 67:62x9 þ 23:17x10 :

The optimum IP solution for this numerical example is selecting projects 2, 3,

4, and 8. This solution reflects the preferences of multiple DMs and satisfies budget
constraints. If the DMs are not satisfied with the solution, we have to use AHP to
assess and compute new values of weights g and l according to Fig. 29.6. This
process will terminate whenever all DMs agree to the compromise solution.

29.5.2.2 Managerial Implications

The model and solution procedure proposed in this study have several managerial
implications. First, the model integrates the multiple-objective capital budgeting
problem with multiple decision makers. The most common problem in a capital
investment decision-making situation is how to estimate cash flows and to incor-
porate the risk factors of decision makers (Hillier 1963; Lee 1993). In our model,
we allow multiple-constraint levels to incorporate each DM’s preference about
budget availability. Neither linear programming nor goal programming models can
handle these characteristics.
Second, our model can be applied to solve capital budgeting problems not only
with the representation of a complex structure but also with motivational implica-
tions. Our method facilitates DMs’ participation in order to minimize
suboptimization of overall company goals (Locke et al. 1981). The model can aid
coordination and synthesis of multiple objectives, some in conflict. This feature can
be quite effective in a capital budgeting situation, in which many objectives are
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 833

contradictory to each other and these objectives can be measured differently by

a number of decision participants. This model can foster more autonomous control
than any other mathematical programming models by allowing DMs to express
their own priority structures and weights.
Third, adoption of the AHP reduces the solution complexity of the resulting MC2
IP program. The MC2 IP is a very complex problem to solve. Real-life size MC2 IP
problems are intractable by using current computational technology. Furthermore,
the computing software for MC2 IP is still under development (Shi and Lee 1992).
Even though the use of AHP may lose some possible trade-offs among objectives
and/or DMs’ preferences, the AHP reduces MC2 IP to a traditional zero-one IP that
is much easier to be solved using currently available IF software.
Lastly, our two-phased framework (Fig. 29.6) for capital budgeting problem is
highly flexible. The DMs can reach an agreement interactively. For example,
alternative budgeting policy can be obtained by providing a different preference
matrix of objectives and/or resource availability levels. Furthermore, the
two-phased framework may be attempted iteratively until all of the DMs are
satisfied with the final budgeting policy.

29.5.3 Conclusions

A decision-making process with a multiple-criteria model has been addressed to

solve capital budgeting problems. This model can foster a high-quality decision in
capital budgeting under multiple criteria and multiple decision makers. This
decision-making strategy reflects each decision maker’s preference and limits
suboptimization of overall company goals. This method can also better handle
real-world problems that may include uncertain factors. By incorporating informa-
tion from each influential manager, our model is more likely to provide better
budgeting solutions than the previous linear programming or goal programming
approaches.
There are other research problems remaining to be explored. From a practical
point of view, the estimation of future cash flows, determining a firm’s cost of
capital, measuring intangible benefits, and measuring residual value of assets are
still important issues. Models like ours can reduce the chance of an ineffective
decision making by incorporating multiple decision makers’ preferences. This
framework can be applied to other accounting problems, such as cost allocation,
audit sampling objectives, and personnel planning for an audit, if the decision
variables and formulation are expressed appropriately.

29.6 Conclusions

We have introduced group decision-making tools that can be applied in accounting

and finance. By the nature of today’s dynamic business environment, there will be
more than one decision maker and business conditions keep changing. We showed
834 W. Kwak et al.

AHP and MC2 applications in performance evaluation, banking performance eval-

uation, international transfer pricing, and capital budgeting. Our last paper shows
the combination of AHP and MC2 in capital budgeting to deal with multiple
objectives and multiple constraints.
Our performance evaluation model shows advantages such as flexibility,
feedbacks, group evaluation, and computing simplicity. A prototype was built via
a personal computer so that the model can be applied to any business situations. The
iterative process of getting input data in the AHP procedure helps each manager as
well as employee to be aware of the importance of strategic factors of each perfor-
mance measure of the bank from our second paper. Our transfer pricing model can
provide a systematic and comprehensive scenario about all possible optimal transfer
prices depending on multiple-criteria and multiple-constraint levels. The trade-offs
of optimal transfer prices offer a broad basis for managers of a corporation to flexibly
implement the optimal transfer pricing strategy and cope with various business
situations. Furthermore, this method also aids global optimization, division auton-
omy, and performance evaluation. Our last paper shows that our capital budgeting
model is more likely to provide better budgeting solutions than the previous
approaches by incorporating information from each influential manager.

Appendix 1

For illustrative purposes, we show how to use AHP to induce the five DMs’
preferences of budget availability and to compute the relative weights. Generally,
AHP collects input judgments of DMs in the form of a matrix by pairwise compar-
isons of criteria (i.e., their budget availability levels). An eigenvalue method is then
used to scale weights of such criteria. That is, the relative importance of each
criteria is computed. The result from all of pairwise comparison is stored in an input
matrix as follows:
President Controller Production Marketing Engineering
Manager Manager Manager
2 3
1 3 4 5 6
6 7
6 1 2 5 57
6 7
6 1 3 47
6 7
6 7
4 1 25
1

Applying an eigenvalue method to the above input matrix results in a vector

Wi ¼ (0.477, 0.251, 0.154, 0.070, 0.048). In addition to the vector, the inconsis-
tency ratio (g) is obtained to estimate the degree of inconsistency in pairwise
comparisons. In this example, the inconsistency ratio is 0.047. A common guideline
is that if the ratio surpasses 0.1, a new input matrix must be generated. Therefore,
this input matrix is acceptable.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 835

A similar computing process can be applied for the 16 objective functions.

However, two hierarchical levels are required for this case. The first level of AHP
corresponds to the four groups of objectives and the second level corresponds to the
time periods within each objective group.
Let xj be the jth project that can be selected by the firm. Using data in Tables 29.7,
29.8, 29.9, and 29.10, the model for capital budgeting with multiple criteria and
multiple DMs is formulated as

maximize 20x1 þ 16x2 þ 11x3 þ 4x4 þ 4x5

þ 18x6 þ 7x7 þ 9x8 þ 24x9 þ 4x10

maximize 120x1 þ 100x2 þ 80x3 þ 40x4 þ 40x5

þ 110x6 þ 60x7 þ 110x8 þ 150x9 þ 35x10

maximize 130x1 þ 120x2 þ 90x3 þ 50x4 þ 45x5

þ 120x6 þ 70x7 þ 120x8 þ 170x9 þ 40x10

maximize 145x1 þ 140x2 þ 95x3 þ 50x4 þ 50x5

þ 140x6 þ 65x7 þ 100x8 þ 180x9 þ 40x10

maximize 150x1 þ 150x2 þ 95x3 þ 55x4 þ 55x5

þ 150x6 þ 70x7 þ 110x8 þ 190x9 þ 50x10

maximize 170x1 þ 160x2 þ 100x3 þ 60x4 þ 60x5

þ 165x6 þ 80x7 þ 120x8 þ 200x9 þ 50x10

maximize 10x1 þ 10x2 þ 12x3 þ 8x4 þ 15x5

þ 12x6 þ 8x7 þ 14x8 þ 12x9 þ 10x10

maximize 12x1 þ 12x2 þ 15x3 þ 15x4 þ 10x5

þ 12x6 þ 12x7 þ 16x8 þ 10x9 þ 8x10

maximize 14x1 þ 18x2 þ 15x3 þ 10x4 þ 8x5

þ 10x6 þ 10x7 þ 13x8 þ 9x9 þ 9x10

maximize 15x1 þ 16x2 þ 15x3 þ 8x4 þ 20x5

þ 15x6 þ 12x7 þ 15x8 þ 12x9 þ 8x10

maximize 17x1 þ 17x2 þ 8x3 þ 12x4 þ 18x5

þ 15x6 þ 12x7 þ 16x8 þ 12x9 þ 12x10

maximize 0:85x1 þ 0:90x2 þ 0:96x3 þ 0:98x4 þ 0:90x5

þ 0:90x6 þ 0:96x7 þ 0:96x8 þ 0:90x9 þ 0:98x10
836 W. Kwak et al.

maximize 0:90x1 þ 0:98x2 þ 0:97x3 þ 0:95x4 þ 0:95x5

þ 0:95x6 þ 0:98x7 þ 0:95x8 þ 0:88x9 þ 0:95x10

maximize 0:98x1 þ 0:95x2 þ 0:98x3 þ 0:96x4 þ 0:95x5

þ 0:94x6 þ 0:98x7 þ 0:90x8 þ 0:85x9 þ 0:95x10

maximize 0:95x1 þ 0:95x2 þ 0:92x3 þ 0:92x4 þ 0:98x5

þ 0:95x6 þ 0:98x7 þ 0:92x8 þ 0:95x9 þ 0:98x10

maximize 0:95x1 þ 0:96x2 þ 0:92x3 þ 0:95x4 þ 0:95x5

þ 0:95x6 þ 0:98x7 þ 0:95x8 þ 0:95x9 þ 0:95x10

subject to 24x1 þ 12x2 þ 8x3 þ 6x4 þ x5

þ 18x6 þ 13x7 þ 14x8 þ 16x9 þ 4x10  49:37

16x1 þ 10x2 þ 6x3 þ 4x4 þ 6x5

þ 18x6 þ 8x7 þ 8x8 þ 20x9 þ 6x10  35:30

6x1 þ 2x2 þ 6x3 þ 7x4 þ 9x5

þ 20x6 þ 10x7 þ 12x8 þ 24x9 þ 8x10  28:91

6x1 þ 5x2 þ 6x3 þ 5x4 þ 2x5

þ 15x6 þ 8x7 þ 10x8 þ 16x9 þ 6x10  33:44

8x1 þ 4x2 þ 4x3 þ 3x4 þ 3x5

þ 15x6 þ 8x7 þ 8x8 þ 16x9 þ 4x10  29:96

and Xj ¼ f0; 1g for j ¼ 1, . . . , 10:

References
Abdel-khalik, A. R., & Lusk, E. J. (1974). Transfer pricing – A synthesis. The Accounting Review, 8–24.
Arrington, E. E., Hillison, W., & Jensen, R. E. (1984). An application of analytical hierarchy
process to model expert judgment on analytic review procedures. Journal of Accounting
Research, 22, 298–312.
Ayvaz, E., & Pehlivanli, D. (2011). The use of time driven activity based costing and analytic
hierarchy process method in the balanced scorecard implementation. International Journal of
Business and Management, 6, 146–158.
Bailey, A. D., & Boe, W. J. (1976). Goal and resource transfers in the multigoal organization. The
Accounting Review, 559–573.
Bapna, R., Goes, P., & Gupta, A. (2005). Pricing and allocation for quality-differentiated online
services. Management Science, 51, 1141–1150.
Baumol, W. J., & Fabian, T. (1964). Decomposition, pricing for decentralization and external
economies. Management Science, 11, 1–32.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 837

Baumol, W. J., & Quandt, R. E. (1965). Investment and discount rate under capital rational –
A programming approach. Economic Journal, 75(298), 173–329.
Bernard, R. H. (1969). Mathematical programming models for capital budgeting – A survey
generalization and critique. Journal of Financial and Quantitative Analysis, 4, 111–158.
Bhaskar, K. A. (1979). Multiple objective approach to capital budgeting. Accounting and Business
Research, 10, 25–46.
Bhaskar, I. C., & McNamee, P. (1983). Multiple objectives accounting and finance. Journal of
Business Finance and Accounting, 10, 595–621.
Boardman, W., Reinhart, I., & Celec, S. E. (1982). The role of payback period in the theory and
application of duration to capital budgeting. Journal of Business Finance and Accounting, 9,
511–522.
Borkowski, S. C. (1990). Environmental and organizational factors affecting transfer pricing:
A survey. Journal of Management Accounting Research, 2, 78–99.
Boucher, T. O., & MacStravic, E. L. (1991). Multi-attribute evaluation within a present value
framework and its relation to the analytic hierarchy process. The Engineering Economist, 37, 1–32.
Calvert, J. (1990). Quality customer service: A strategic marketing weapon for high performance
banking. Industrial Engineering, 54–57.
Campi, J. P. (1992). It’s not as easy as ABC. Journal of Cost Management Summer, 6, 5–11.
Chan, Y. L. (2004). Use of capital budgeting techniques and an analytic approach to capital
investment decisions in Canadian municipal governments. Public Budgeting & Finance, 24,
40–58.
Chan, Y. L., & Lynn, B. E. (1991). Performance evaluation and the analytic hierarchy process.
Journal of Management Accounting Research, 3, 57–87.
Charnes, A., & Cooper, W. W. (1961). Management models and industrial applications of linear
programming (Vols. I & II). New York: Wiley.
Cheng, J. K. (1993). Managerial flexibility in capital investment decision: Insights from the real-
options. Journal of Accounting Literature, 12, 29–66.
Chien, I. S., Shi, Y., & Yu, P. L. (1989). MC2 program: A Pascal program run on PC or Vax
(revised version). Lawrence: School of Business, University of Kansas.
Choi, T. S., & Levary, R. R. (1989). Multi-national capital budgeting using chance-constrained
goal programming. International Journal of Systems Science, 20, 395–414.
Corner, I. L., Decicen, R. F., & Spahr, R. W. (1993). Multiple-objective linear programming in
capital budgeting. Advances in Mathematical Programming and Financial Planning, 3,
241–264.
Curtis, S. L. (2010). Intercompany licensing rates: Implications from principal-agent theory.
International Tax Journal, 36, 25–46.
Dantzig, G. B. (1963). Linear programming and extensions. New Jersey: Princeton University
Press.
Dantzig, G. B., & Wolfe, P. (1960). Decomposition principles for linear programs. Operations
Research, 8, 101–111.
Deckro, R. F., Spahr, R. W., & Herbert, J. E. (1985). Preference trade-offs in capital budgeting
decisions. IIE Transactions, 17, 332–337.
DeWayne, L. S. (2009). Developing a lean performance score. Strategic Finance, 91, 34–39.
Dopuch, N., & Drake, D. F. (1964). Accounting implications of a mathematical programming
approach to the transfer price problem. Journal of Accounting Research, 2, 10–24.
Drucker, P. F. (1993). We need to measure, not count. Wall Street Journal, April 13. Also Expert
Choice Voice, July.
Dyer, R. F., & Forman, E. H. (1991). An analytic approach to marketing decisions. Englewood,
NJ: Prentice Hall.
Eccles, R. G. (1983). Control with fairness in transfer pricing. Harvard Business Review, 61,
149–161.
Eccles, R. G. (1991). The performance measurement manifesto. Harvard Business Review, 69,
131–137.
838 W. Kwak et al.

Eccles, R. G., & Pyburn, P. J. (1992). Creating a comprehensive system to measure performance.
Management Accounting, 74, 41–44.
Eddie, W., Cheng, L., & Li, H. (2001). Analytic hierarchy process: An approach to determine
measures for business performance. Measuring Business Excellence, 5, 30–36.
Fisher, J. (1992). Use of nonfinancial performance measures. Journal of Cost Management, 6,
31–38.
Forman, E. H., Saaty, T. L., Selly, M. A., & Waldron, R. (1985). Expert choice. Pittsburgh:
Decision Support Software.
Goicoechea, A., Hansen, D. R., & Duckstein, L. (1982). Multi-objective decision analysis appli-
cations. NewYork: Wiley.
Gonzalez, G., Reeves, R., & Franz, L. S. (1987). Capital budgeting decision making: An interac-
tive multiple objective linear integer programming search procedure. Advances in Mathemat-
ical Programming and Financial Planning, 1, 21–44.
Gould, J. R. (1964). Internal pricing in corporations when there are costs of using an outside
market. The Journal of Business, 37, 61–67.
Gyetvan, F., & Shi, Y. (1992). Weak duality theorem and complementary slackness theorem for
linear matrix programming problems. Operations Research Letters, 11, 244–252.
Hardy, C., & Reeve, R. (2000). A study of the internal control structure for electronic data interchange
systems using the analytic hierarchy process. Accounting and Finance, 40, 191–210.
Harker, P. T., & Vargas, L. G. (1987). The theory of ratio scale estimation: Saaty’s analytic
hierarchy process. Management Science, 33, 1383–1403.
Hillier, F. S. (1963). The derivation of probabilities information for the evaluation of risky
investments. Management Science, 443–457.
Hirshleifer, J. (1956). On the economies of transfer pricing. The Journal of Business, 172–184.
Howe, K. M., & Patterson, I. H. (1985). Capital investment decisions under economies of scale in
flotation costs. Financial Management, 14, 61–69.
Howell, R. A., & Schwartz, W. A. (1994). Asset deployment and investment justification, In
Handbook of cost management. Boston: Warren Gorham Lamont.
Huang, X. (2008). Mean-variance model for fuzzy capital budgeting. Computers & Industrial
Engineering, 55, 34.
Huang, S., Hung, W., Yen, D. C., Chang, I., & Jiang, D. (2011). Building the evaluation model of the
IT general control for CPAs under enterprise risk management. Decision Support Systems, 50, 692.
Ignizio, J. P. (1976). An approach to the capital budgeting problem with multiple objectives. The
Engineering Economist, 21, 259–272.
Ishizaka, A., Balkenborg, D., & Kaplan, T. (2011). Influence of aggregation and measurement
scale on ranking a compromise alternative in AHP. The Journal of the Operational Research
Society, 62, 700–711.
Kamath, R. R., & Khaksari, S. Z. (1991). Real estate investment: An application of the analytic
hierarchy process. The Journal of Financial and Strategic Decisions, 4, 73–100.
Kaplan, R. S. (1986). Accounting lag: The obsolescence of cost accounting systems. California
Management Review, 28, 175–199.
Kaplan, R. S., & Norton, D. P. (1992). The balanced scorecard-measures that drive performance.
Harvard Business Review, 69, 71–79.
Karanovic, G., Baresa, S., & Bogdan, S. (2010). Techniques for managing projects risk in capital
budgeting process. UTMS Journal of Economics, 1, 55–66.
Kim, S. H., & Farragher, E. I. (1981). Current capital budgeting practices. Management Account-
ing, 6, 26–30.
Lecraw, D. J. (1985). Some evidence on transfer pricing by multinational corporations. In
A. M. Rugman & L. Eden (Eds.), Multinationals and transfer pricing. New York:
St. Martin’s Press.
Lee, C. F. (1993). Statistics for business and financial economics. Lexington: Heath.
Lee, H. (1993). A structured methodology for software development effort prediction using the
analytic hierarchy process. Journal of Systems and Software, 21, 179–186.
29 Group Decision-Making Tools for Managerial Accounting and Finance Applications 839

Lee, S. M., & Lerro, A. (1974). Capital budgeting for multiple objectives. Financial Management,
3, 51–53.
Lee, Y. R., Shi, Y., & Yu, P. L. (1990). Linear optimal designs and optimal contingency plans.
Management Science, 36, 1106–1119.
Lee, H., Nazem, S. M., & Shi, Y. (1995). Designing rural area telecommunication networks via
hub cities. Omega: International Journal of Management Science, 22(3), 305–314.
Li, Y., Huang, M., Chin, K., Luo, X., & Han, Y. (2011). Integrating preference analysis and
balanced scorecard to product planning house of quality. Computers & Industrial Engineering,
60, 256–268.
Liberatore, M. J., Mohnhan, T. F., & Stout, D. E. (1992). A framework for integrating capital
budgeting analysis with strategy. The Engineering Economist, 38, 31–43.
Lin, T. W. (1993). Multiple-criteria capital budgeting under risk. Advances in Mathematical
Programming and Financial Planning, 3, 231–239.
Lin, L., Lefebvre, C., & Kantor, J. (1993). Economic determinants of international transfer pricing
and the related accounting issues, with particular reference to Asian Pacific countries. The
International Journal of Accounting, 49–70.
Locke, E. A., Shaw, K. N., & Lathom, G. P. (1981). Goal setting and task performance:
1969–1980. Psychological Bulletin, 90, 125–152.
Lorie, J. H., & Savage, L. I. (1955). Three problems in rationing capital. Journal of Business, 28,
229–239.
Mensah, Y. M., & Miranti, P. J., Jr. (1989). Capital expenditure analysis and automated
manufacturing systems: A review and synthesis. Journal of Accounting Literature, 8, 181–207.
Merville, L. J., & Petty, W. J. (1978). Transfer pricing for the multinational corporation. The
Accounting Review, 53, 935–951.
Muralidhar, K., Santhanam, R., & Wilson, R. L. (1990). Using the analytic hierarchy process for
information system project selection. Information and Management, 18, 87–95.
Naert, P. A. (1973). Measuring performance in a decentralized corporation with interrelated
divisions: Profit center versus cost center. The Engineering Economist, 99–114.
Nakayama, H. (1994). Some remarks on trade-off analysis in multi-objective programming. The XIth
International conference on multiple criteria decision making, Coimbra, Portugal, Aug. 1–6.
Nanni, A. J., Dixon, J. R., & Vollmann, T. E. (1990). Strategic control and performance measure-
ment. Journal of Cost Management, 4, 33–42.
Natarajan, T., Balasubramanian, S. A., & Manickavasagam, S. (2010). Customer’s choice amongst
self-service technology (SST) channels in retail banking: A study using analytical hierarchy
process (AHP). Journal of Internet Banking and Commerce, 15, 1–15.
Onsi, M. (1970). A transfer pricing system based on opportunity cost. The Accounting Review, 45,
535–543.
Palliam, R. (2005). Application of a multi-criteria model for determining risk premium. The
Journal of Risk Finance, 6, 341–348.
Pike, R. (1983). The capital budgeting behavior and corporate characteristics of capital-
constrained firms. Journal of Business Finance and Accounting, 10, 663–671.
Pike, R. (1988). An empirical study of the adoption of sophisticated capital budgeting practices
and decision-making effectiveness. Accounting and Business Research, 18, 341–351.
Reeves, G. R., & Franz, L. S. (1985). A simplified interactive multiple objective linear program-
ming procedure. Computers and Operations Research, 12, 589–601.
Reeves, G. A., & Hedin, S. R. (1993). A generalized interactive goal programming procedure.
Computers and Operations Research, 20, 747–753.
Reeves, G. F., Lawrence, S. M., & Gonzalez, J. J. (1988). A multiple criteria approach to aggregate
industrial capacity expansion. Computers and Operations Research, 15, 333–339.
Ronen, J., & McKinney, G. (1970). Transfer pricing for divisional autonomy. Journal of Account-
ing Research, 8, 99–112.
Rotch, W. (1990). Activity-based costing in service industries. Journal of Cost Management,
4–14.
840 W. Kwak et al.

Ruefli, T. W. (1971). A generalized goal decomposition model. Management Science, 17,

505–518.
Saaty, T. L. (1980). The analytic hierarchy process. New York: McGraw-Hill.
Santhanam, R., Muralidhar, K., & Schniederjans, M. (1989). Zero-one goal programming
approach for information system project selection. Omega: International Journal of Manage-
ment Science, 17, 583–598.
Schniederjans, M. J., & Garvin, T. (1997). Using the analytic hierarchy process and multi-
objective programming for the selection of cost drivers in activity-based costing. European
Journal of Operational Research, 100, 72–80.
Schniederjans, M. J., & Wilson, R. L. (1991). Using the analytic hierarchy process and goal program-
ming for information system project selection. Information and Management, 20, 333–342.
Seiford, L., & Yu, P. L. (1979). Potential solutions of linear systems: The multi-criteria multiple
constraint level program. Journal of Mathematical Analysis and Applications, 69, 283–303.
Shi, Y. (1991). Optimal linear production systems: Models, algorithms, and computer support
systems. Ph.D. dissertation, School of Business, University of Kansas.
Shi, Y., & Lee, H. (1992). A binary integer linear programming with multi-criteria and multi-
constraint levels (Working paper #92–21). School of Business, University of Nebraska at Omaha.
Shi, Y., & Yu, P. L. (1992). Selecting optimal linear production systems in multiple criteria
environments. Computer and Operations Research, 19, 585–608.
Singhal, J., Marsten, R. E., & Morin, T. L. (1989). Fixed order brand-and-bound methods for
mixed-integer programming: The ZOOM system. ORSA Journal on Computing, 1, 44–51.
Steuer, R. (1986). Multiple criteria optimization: Theory. Computation and application. New York:
Wiley.
Suh, C.K., Suh, E.H., Choi, J. (1993). A two-phased DSS for R&D portfolio selection. Pro-
ceedings of International DSI, 351–354.
Tang, R. Y. W. (1992). Transfer pricing in the 1990s. Management Accounting, 73, 22–26.
Tapinos, E., Dyson, R. G., & Meadows, M. (2011). Does the balanced scorecard make a difference
to the strategy development process? The Journal of the Operational Research Society, 62,
888–899.
Thanassoulis, E. (1985). Selecting a suitable solution method for a multi-objective programming
capital budgeting problem. Journal of Business Finance and Accounting, 12, 453–471.
Thomas, A. L. (1980). A behavioral analysis of joint-cost allocation and transfer pricing.
Champaign: Stipes Publishing.
Turney, S. T. (1990). Deterministic and stochastic dynamic adjustment of capital investment
budgets. Mathematical Computation and Modeling, 13, 1–9.
Watson, D. J. H., & Baumler, J. V. (1975). Transfer pricing: A behavioral context. The Accounting
Review, 50, 466–474.
Weingartner, H. M. (1963). Mathematical programming and the analysis of capital budgeting
problems. Englewood Cliffs: Prentice-Hall.
Wu, C.-R., Lin, C.-T., & Tsai, P.-H. (2011). Financial service sector performance measurement
model: AHP sensitivity analysis and balanced scorecard approach. The Service Industries
Journal, 31, 695.
Yu, P. L. (1985). Multiple criteria decision making: Concepts, techniques and extensions.
New York: Plenum.
Yu, Y. L., & Zeleny, M. (1975). The set of all nondominated solutions in the linear cases and
a multicriteria simplex method. Journal of Mathematical Analysis and Applications, 49,
430–458.
Yunker, P. J. (1983). Survey study of subsidiary, autonomy, performance evaluation and transfer
pricing in multinational corporations. Columbia Journal of World Business, 19, 51–63.
Zeleny, M. (1974). Linear multi-objective programming. Berlin: Springer.
Statistics Methods Applied in Employee
Stock Options 30
Li-jiun Chen and Cheng-der Fuh

Contents
30.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 842
30.2 Preliminary Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 844
30.2.1 Change of Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 844
30.2.2 Hierarchical Clustering with K-Means Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
30.2.3 Standard Errors in Finance Panel Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 846
30.3 Employee Stock Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847
30.3.1 Model-Based Approach to Subjective Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847
30.3.2 Compensation-Based Approach to Subjective Value . . . . . . . . . . . . . . . . . . . . . . . . . 851
30.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 852
30.4.1 Exercise Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 852
30.4.2 Factors’ Effects on ESOs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853
30.4.3 Offsetting Effect of Sentiment and Risk on ESO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
30.5 Empirical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
30.5.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
30.5.2 Preliminary Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
30.5.3 Implications of Regression Results and Variable Sensitivities . . . . . . . . . . . . . . . 861
30.5.4 Subjective Value and Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 864
30.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 866
Appendix 1: Derivation of Risk-Neutral Probability by Change of Measure . . . . . . . . . . . . . . . . . 867
Appendix 2: Valuation of European ESOs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 869
Appendix 3: Hierarchical Clustering with a K-Means Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 870
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 871

L.-j. Chen (*)

Department of Finance, Feng Chia University, Taichung City, Taiwan
e-mail: lijiunchen@fcu.edu.tw
C.-d. Fuh
Graduate Institute of Statistics, National Central University, Zhongli City, Taiwan
e-mail: stcheng@stat.sinica.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 841
DOI 10.1007/978-1-4614-7750-1_30,
# Springer Science+Business Media New York 2015
842 L.-j. Chen and C.-d. Fuh

Abstract
This study presents model-based and compensation-based approaches to
determining the price-subjective value of employee stock options (ESOs).
In the model-based approach, we consider a utility-maximizing model in
which the employees allocate their wealth among company stock, a market
portfolio, and risk-free bonds, and then we derive the ESO formulas, which
take into account illiquidity and sentiment effects. By using the method of
change of measure, the derived formulas are simply like those of the market
value with altered parameters. To calculate the compensation-based subjec-
tive value, we group employees by hierarchical clustering with a K-means
approach and back out the option value in an equilibrium competitive
employment market.
Further, we test illiquidity and sentiment effects on ESO values by running
regressions that consider the problem of standard errors in the finance panel data.
Using executive stock options and compensation data paid between 1992 and
2004 for firms covered by the Compustat Executive Compensation Database, we
find that subjective value is positively related to sentiment and negatively related
to illiquidity in all specifications, consistent with the offsetting roles of sentiment
and risk aversion. Moreover, executives value ESOs at a 48 % premium to the
Black-Scholes value and ESO premiums are explained by a sentiment level of
12 % in risk-adjusted, annualized excess return, suggesting a high level of
executive overconfidence.

Keywords
Employee stock option • Sentiment • Subjective value • Illiquidity • Change of
measure • Hierarchical clustering with K-means approach • Standard errors in
finance panel data • Exercise boundary • Jump diffusion model

30.1 Introduction

Employee stock options (ESOs) are a popular method of compensation.

According to the Compustat Executive Compensation database, the number of
option grants increased from roughly 0.25 billion in 1992 to 1.4 billion in 2001.
Specifically, in fiscal 2001, 53 % of total pay came from granted options,
compared with 33 % in 1992.1 Moreover, executives receive averagely more
than ten thousand options after 1998. ESOs can help firms retain talent and reduce
agency costs (Jensen and Meckling 1976). They also mitigate risk-related incen-
tive problems (Agrawal and Mandelker 1987; Hemmer et al. 2000) and provide an
alternative to cash compensation, which is especially important for firms facing

1
The option values are estimated by the calculation in ExecuCompustat database.
30 Statistics Methods Applied in Employee Stock Options 843

financial constraints (Core and Guay 2001). In addition, ESOs attract highly
motivated and able employees (Core and Guay 2001; Oyer and Schaefer 2005).
All of these factors contribute to the importance of ESOs for corporate gover-
nance and finance research.
The illiquidity problem of ESOs cannot be neglected. ESOs usually have
a vesting period during which they cannot be exercised, and hence they cannot
be redeemed for a fixed period of time. Furthermore, ESOs are generally not
publicly traded and their sale is not permitted. Standard methods for valuing
options are difficult to apply to ESOs. Because of the illiquidity of ESOs, many
employees hold undiversified portfolios that include large stock options of their
own firms. Because of the impossibility of full diversification, the value per-
ceived by employees (subjective value) may be quite different from the traded
option. We consider the subjective value to be what a constrained agent would
pay for the ESOs and the market value to be the value perceived by an
unconstrained agent. Many papers address the differences between subjective
and market value, each concluding that the subjective value of ESOs should be
less than the market value (Lambert et al. 1991; Hall and Murphy 2002;
Ingersoll 2006).
If ESOs are generally worth less than market value, why do employees continue
to accept and, indeed, sometimes prefer ESO compensation? One reasonable
explanation is employee sentiment. Sentiment means positive private information
or behavioral overconfidence regarding the future risk-adjusted returns of the firm.
Simply, the employee believes that he possesses private information and can benefit
from it. Or the employee overestimates the return of the firm and believes ESOs are
valuable. Some empirical evidence supports this conjecture. Oyer and Schaefer
(2005) and Bergman and Jenter (2007) posit that employees attach a sentiment
premium to their stock options; firms exploit this sentiment premium to attract and
retain optimistic employees. Hodge et al. (2009) provide survey evidence and find
that managers subjectively value the stock option greater than its Black-Scholes
value.
Three statistics methods are applied in our ESO study (Chen and Fuh 2011;
Chang et al. 2013), including change of measure, hierarchical clustering with a
K-means approach, and estimation of standard errors in finance panel data.
We derive a solution for ESO value that is a function of both illiquidity
and sentiment in a world where employees balance their wealth between the
company’s stock, the market portfolio, and a risk-free asset. By using the method
of change of measure, we find a probability measure and then the ESO formulas are
derived easily. In addition, from the ESO pricing formulas, we are able to not only
estimate the subjective values but also study the exercise policies. Early exercise is
a pervasive phenomenon and, importantly, the early exercise effect is critical in
valuation of ESOs, especially for employees who are more risk averse and when
there are more restrictions on stock holding.
Applying a comprehensive set of executive options and compensation data,
this study empirically determines subjective value, grouping employees by
hierarchical clustering with a K-means approach and backing out the option
844 L.-j. Chen and C.-d. Fuh

value in an equilibrium competitive employment market. Specifically, we group

executives according to position, the firm’s total market value, nonoption
compensation, and the immediate exercise value of the options for each industry
by hierarchical clustering with a K-means approach. We then calculate
the empirical value of ESO that each executive places on their ESOs in order
for total compensation to be equivalent in the same cluster. Further, we model
both illiquidity and sentiment effects and test them with executive options
data. We regress the empirical ESO values on the proportion of total wealth
held in illiquid firm-specific holdings and sentiment, which is estimated
from our pricing formula or Capital Asset Pricing Model (CAPM) risk-adjusted
alpha under controlling key options pricing variables such as moneyness, time to
maturity, volatility, and dividend payout. As we know, when the residuals are
correlated across observations, the standard errors of estimated coefficients
produced by Ordinary Least Squares (OLS) may be biased and then lead to
incorrect inference. Petersen (2009) compares the different methods used in the
literature and gives researchers guidance for their use. The data we collected are
from multiple firms over several years. Hence, we consider the problem of
standard errors in finance panel data and run the regressions, including standard
errors clustered by firm, year, and both firm and year.
The remainder of this chapter is organized as follows. Section 30.2 introduces
some preliminary knowledge. Section 30.3 develops our model and addresses
the approaches to price-subjective value of ESOs. Section 30.4 presents the
simulation results. Section 30.5 shows the empirical study; Sect. 30.6 concludes
the chapter.

30.2 Preliminary Knowledge

30.2.1 Change of Measure

In our ESO study, we assume stock price follows a jump-diffusion process. Here,
we briefly introduce change of measure for a compound Poisson and Brownian
motion. Suppose that we have a probability space (O, F , ℙ) on which is defined
a Brownian motion Wt. Suppose that on this same probability space there is defined
a compound Poisson process

X
Nt
Qt ¼ Yi
i¼0

with intensity l and jumps having density function f(y). Assume that there is
a single filtration F t, t0, for both the Brownian motion and the compound Poisson
process. In this case, the Brownian motion and compound Poisson process must be
independent (see Shreve 2008).
30 Statistics Methods Applied in Employee Stock Options 845

Let el be a positive number, let ef ðyÞ be another density function with

the property that ef ðyÞ ¼ 0 whenever f(y) ¼ 0, and let Y(t) be an adapted process.
We define  ðt ð 
 1 t 2
Z1 ðtÞ ¼ exp  Yðu dW u  Y ðuÞdu ,
0 2 0
 Y N t ee
e l f ðY i Þ
Z2 ðtÞ ¼ e ll t ,
i¼0
lf ðY i Þ
 
ZðtÞ ¼ Z 1 ðt Z 2 ðt :

It can be verified that the process Z(t) is a martingale.

ð
e
Lemma 30.1 Define P ðAÞ ¼ Z ðT ÞdP for all A 2 F . Under the probability
measure ℙ, the process A

ðt
e t ¼ Wt þ
W YðsÞds
0

is a Brownian motion, Qt is a compound Poisson process with intensity e l and

independent, identically distributed jump sizes having density ef ðyÞ, and the pro-
cesses We t and Qt are independent.
The proof of Lemma 30.1 is presented by Shreve (2008). This lemma is useful
for us to derive the ESO formula in Sect. 30.3.

30.2.2 Hierarchical Clustering with K-Means Approach

The main assumption of the compensation-based subjective value is that all

executives within the same group receive the same total compensation. For each
executive in the group, the implied subjective value is derived by comparing the
difference between nonoption compensation and the average compensation. Group-
ing executives appropriately is essential in the compensation-based approach.
Clustering is an unsupervised technique for analyzing data and dividing patterns
(observations, data items, or feature vectors) into groups (clusters). There are two
kind of clustering algorithms: hierarchical and partitional approaches. Hierarchical
methods produce a nested structure of partitions, whereas partitional methods
produce only one partition (Jain et al. 1999).
Hierarchical clustering algorithms repeat the cycle of either merging smaller
clusters into larger ones or dividing larger clusters to smaller ones. An agglomerative
clustering strategy uses the bottom-up approach of merging clusters into larger
ones, whereas divisive clustering strategy uses the top-down approach of splitting
larger clusters into smaller ones. Typically, the greedy approach is used in
deciding which larger/smaller clusters are used for merging/dividing. Euclidean
distance, Manhattan distance, and cosine similarity are some of the most
846 L.-j. Chen and C.-d. Fuh

commonly used metrics of similarity for numeric data. For non-numeric data,
metrics such as the Hamming distance are used. It is important to note that the
actual observations (instances) are not needed for hierarchical clustering, because
only the matrix of distances is sufficient. The user can obtain different clustering
depending on the level that is cut.
A partitional clustering algorithm obtains a single partition of the data instead
of a clustering structure. A problem accompanying the use of a partitional
algorithm is the choice of the number of desired output clusters (usually
called k). One of the most commonly used partitional clustering algorithms is
the K-means clustering algorithm. It starts with a random initial partition and
keeps reestimating cluster centers and reassigning the patterns to clusters based
on the similarity between the pattern and the cluster centers. These two steps are
repeated until a certain intra-cluster similarity objective function and inter-cluster
dissimilarity objective function are optimized. Therefore, sensible initialization
of centers is an important factor in obtaining quality results from partitional
clustering algorithms.
Hierarchical and partitional approaches each have their advantages and disad-
vantages. Therefore, we apply a hybrid approach combining hierarchical and
partitional approaches in this study.

30.2.3 Standard Errors in Finance Panel Data

In finance panel data, the residuals may be correlated across firms or across time,
and OLS standard errors can be biased. Petersen (2009) compares the different
methods used in the literature and provides guidance to researchers as to which
method should be used.
The standard regression for a panel data is

Y it ¼ Xit b þ Eit i ¼ 1, . . . , N; t ¼ 1, . . . , T

where there are observations on firms i across years t. X and ϵ are assumed to be
independent of each other and to have finite variance. OLS standard errors are
unbiased when the residuals are independent and identically distributed. However, it
may result in incorrect inference when the residuals are correlated across observations.
In finance study, there are two general forms of dependence: time-series depen-
dence (firm effect), in which the residuals of a given firm may be correlated across
years for a given firm, and cross-sectional dependence (time effect), in which the
residuals of a given year may be correlated across different firms. Considering the
firm effect, the residuals and independent variable are specified as

Eit ¼ gi þ nit ; Xit ¼ mi þ vit

Both the independent variable and the residual are correlated across observations
of the same firm but are independent across firms. Petersen (2009) shows that OLS,
30 Statistics Methods Applied in Employee Stock Options 847

Fama-MacBeth, and Newey-West standard errors are biased, and only clustered
standard errors are unbiased as they account for the residual dependence created by
the firm effect.
Considering the time effect, the residuals and independent variable are
specified as

Eit ¼ dt þ it ; Xit ¼ zt þ nit :

OLS standard errors still underestimate the true standard errors. The clustered
standard errors are much more accurate, but unlike the results with the firm effect, they
underestimate the true standard error. However, the bias in the clustered standard error
estimates declines with the number of clusters. Because the Fama-MacBeth procedure
is designed to address a time effect, the Fama-MacBeth standard errors are unbiased.
In both firm and time effect, the residuals and independent variable are
specified as

Eit ¼ gi þ dt þ it ; Xit ¼ mi þ zt þ nit :

Standard errors clustered by only one dimension are biased downward.

Clustering by two dimensions produces less biased standard errors. However,
clustering by firm and time does not always yield unbiased estimates. We only
introduce the method we use in this study. Readers wanting to know more
methods for estimation of standard error in financial panel data should refer to
Petersen (2009).

30.3 Employee Stock Options

This section introduces two methods to price the subjective value of ESOs. We call
them the model-based approach and the compensation-based approach. To derive
the model-based subjective value, we use the technique of change of measure and
find a probability measure P* such that the option value can be generated simply. To
calculate compensation-based subjective value, we group employees by hierarchi-
cal clustering with a K-means approach and back out the option value in an
equilibrium-competitive employment market.

30.3.1 Model-Based Approach to Subjective Value

From Chen and Fuh (2011), we have a three asset economy in which the
employee allocates their wealth among three assets: company stock S, market
portfolio M, and risk-free bond B. Because of the illiquidity of ESO, the
employee is constrained to allocate a fixed fraction a of their wealth to company
stock (via some form of ESO). Define the jump-diffusion processes for the three
assets as follows:
848 L.-j. Chen and C.-d. Fuh

8
>
> dS XNt
>
> ¼ m dt þ s dW þ ndW þ d ðY i  1Þ,
>
> s s m s
< S i¼0
dM (30.1)
>
> ¼ ðmm  dm Þdt þ sm dW m ,
> M
>
>
>
: dB ¼ rdt,
B

where ms ¼ m  d  lk. m, mm, r are instantaneous expected rates of return for the
stock, market portfolio, and risk-free bond, respectively. d and dm are dividends for the
stock and market portfolio, respectively. The Brownian motion process Wm represents
the normal systematic risk of the market portfolio. The Brownian motion process Ws
and jump process Nt are the idiosyncratic risk of the company stock, where Nt captures
the jump risk of company stock and follows a Poisson distribution with average
frequency l. Yi  1 represents the percentage of stock variation when the ith jump
occurs. Denote E(Yi  1) ¼ k and E(Yi  1)2 ¼ k2 for all i. ss and sm are the normal
systematic portions of total volatility for the stock and the market portfolio, respec-
tively, whereas n is the normal unsystematic volatility of the stock. The two Brownian
motions and the jump process are presumed independent. For simplicity, we assume
that CAPM holds so that the efficient portfolio is the market.
Following the idea of Ingersoll (2006), we solve the constrained optimization
problem and determine the employee’s indirect utility function. The marginal
utility is then used as a subjective state price density to value compensation. We
assume that the employee’s utility function U(·) is set as U(C) ¼ Cg/g with
a coefficient of relative risk aversion, 1  g. The process of the employee’s
marginal utility or the pricing kernel can be derived as2:
dJ W
^ dW m  ð1  gÞandW s
¼ ^r dt  s
JW
XNt n o (30.2)
þd ½aðY i  1Þ þ 1g1  1 ,
i¼0
∂J ½W ðtÞ, t
where J W ¼ ∂W ðtÞ
is the marginal utility, J[W(t), t] and W(t) are the
employee’s
 2 2 1 2 total  utility and wealth at time t, ^r ¼ r  ð1  gÞ
mm r
a n þ 2 a g l þ a lk , and s ^ ¼ sm .
The rational equilibrium value of the ESO at time t, F(St, t), satisfies the Euler
equation
Et fJ W ½W ðT Þ, T FðST ; T Þg
FðSt ; tÞ ¼ , (30.3)
J W ½W ðtÞ, t
where F(ST, T) is the payoff at the maturity T. To easily calculate the ESO value, we
find a probability measure P* by using change of measure method and then the
second equality in the following Eq. (30.4) is satisfied.

2
Because the process can be similarly derived from Chang et al. (2008), it is omitted.
30 Statistics Methods Applied in Employee Stock Options 849

Et fJ W ½W ðT Þ, T FðST ; T Þg
F ð St ; t Þ ¼
J W ½W ðtÞ, t (30.4)
¼ erðTtÞ Et ½FðST ; T Þ,

dP ¼ ZðtÞ , Z(t) ¼ e
Z ðT Þ r*t
where dP JW[W(t), t], r* is subjective bond yield, and E*t is the
expectation under P and information at time t. Under P*, the stock process can be
*

expressed as

dS XNt
¼ ðr   d Þdt þ sN dW t þ d ðY i  1Þ,
S i¼0

where
1 
r  ¼ r  ð1  gÞða lk þ g lk2 a2 þ a2 n2
2
 lðx  1Þ,
 
1 
d  ¼ d  ð1  gÞ alk þ g lk2 a2  ð1  a a n2
2
 lðx  1Þ þ lk,
s2N ¼ s2s þ n2 , l ¼ l x, x ¼ E½aðY i  1Þ þ 1g1 ,

W*t is the standard Brownian motion and Nt is a Poisson process with rate l*.
A detailed explanation is given in Appendix 1.

30.3.1.1 European ESO

First, we consider the simple ESO contract, the European ESO. The price formula is
presented in Theorem 30.1.
Theorem 30.1 The value of the European ESO with strike price K and time to
maturity t, written on the jump-diffusion process in Eq. (30.1), is as follows:
X1
ðl  tÞj elt
CE ðSt ; tÞ ¼ CðjÞ (30.5)
j¼0
j!
where
( " #
Y
j
 
d  t 
CðjÞ ¼ St e E Y i F d1
i¼0
   

Ker t E F d2
 Yj   
ln St i¼0 Y i =K þ r   d þ 12 s2N t

d1 ¼ pffiffiffi ,
sN t
pffiffiffi
d2 ¼ d1  sN t:

The proof of Theorem 30.1 is in Appendix 2.

850 L.-j. Chen and C.-d. Fuh

30.3.1.2 American ESO

Suppose that the option can be exercised at n time instants. These time instants are
assumed to be regularly spaced at intervals of Dt, and denoted by ti, 0  i  n,
where t0 ¼ 0, tn ¼ T, and ti+1  ti ¼ Dt for all i. Denote CA as the value of the
American call option, CE as the value of the European call option, K as the strike
price, and Si ¼ Sti . The critical price at these time points is denoted by S*i , 0  i  n,
and is the price at which the agent is indifferent between holding the option and
exercising. Denote E*i as the expectation under P* and information at time ti.
Theorem 30.2 (Chen and Fuh 2011) The value of the American ESO exercisable at
n time instants, when the ESO is not exercised, written on the jump-diffusion
process in Eq. (30.1) is as follows:

CA ðS0 ; T Þ
X
n1 
r‘Dt 
 
¼ CE ðS0 ; T Þ þ e E0 ½S‘ 1  edDt
‘¼1

 
 K 1  erDt I fS‘ S g
‘

X  
n
rjDt 
 
 e E0 ½CA ðSj , ðn  jÞDtÞ  Sj  K I fSj1 S g I fSj <S g : (30.6)
j1 j
j¼2

The critical price S*i at time ti for i ¼ 1,   , n is defined as the solution to the
following equation:

Si  K
 
¼ CE Si , ðn  iÞDt
X 
ni1
r ‘Dt 
   
þ e Ei Siþ‘ 1  ed Dt
‘¼1

 r  Dt

K 1e I fSiþ‘ S g
iþ‘

X 
ni
   
 er jDt 
Ei CA Siþj , ðn  i  jÞDt
j¼2

 
 Siþj  K I fSiþj1 S g I fSiþj <S g ,
iþj1 iþj

where CE(S0, T) and CE(S*i , (n  i)Dt) are calculated in Theorem 30.1.

The value of the American call option, when exercise is allowed at any
time before maturity, is obtained by taking the limit as Dt tends to zero in Eq. (30.6).

30.3.1.3 Sentiment Effect

Often, the manager awarded an incentive option may have different beliefs
about the company’s prospects than the investing public does. The manager
30 Statistics Methods Applied in Employee Stock Options 851

believes that they possess private information and can benefit from it. Or
they have behavioral overconfidence regarding future risk-adjusted return
of their firm and believe ESOs are valuable. We consider the impact of sentiment
on ESO values and the exercise decision. Now, the drift term of stock
process in Eq. 30.1 becomes ms ¼ m + s  d  lk, where sentiment level is
denoted by s. In other words, the employee overestimates or rationally adjusts the
risk-adjusted return of the company owing to inside information by s, then
the same analysis in Theorem 30.1 and 30.2 are valid with a simple adjustment
in parameters. The adjusted interest rate and dividend yield used in pricing are

1 
r  ¼ r þ as  ð1  gÞða lk þ g lk2 a2 þ a2 n2  lðx  1Þ,
2
 
 1
d ¼ d  ð1  aÞs  ð1  g alk þ glk2 a2  ð1  aÞa n2  lðx  1Þ þ lk:
2

30.3.2 Compensation-Based Approach to Subjective Value

The subjective value of ESOs implied by total compensation packages is calculated

using a hierarchical clustering with K-means methodology to split executives into
like groups based on industry, rank, year, the firm market value,3 nonoption
compensation, and the immediate exercise value of the options.4 The number of
groups is decided by a cubic clustering criterion and the average total compensation
is calculated. Then, assuming that all executives within the same cluster receive the
same total compensation, for each executive in this cluster, the implied subjective
value is derived by comparing the difference between nonoption compensation and
the average compensation. We then set all negative implied ESO values equal to
zero and recalculate average compensation in each cluster with these subjective
values, repeating until the relative sum of changes in subjective values in a given
cluster is less than 0.01. This eliminates some negative subjective values such that
the final number of negative or zero values is about 5.7 % of our dataset. Worth
noting is the observation that, even in the first iteration of the process, after
grouping, only about 7.9 % of our data has options with a negative or zero value,
lending credence to the stability of our groupings.
To illustrate the intuition, presume that all executives within the same cluster
receive the same compensation on average, where any differences in salaries, bonuses,
and other income should be accounted for by options. If CEOs’ average total annual
compensation in a given year is three million, a particular CEO who receives one

3
Gabaix and Landier (2008) finds that total market value as a proxy for firm size has the strongest
predictive power on compensation. We, however, redo all tests using number of employees as the
size proxy and find qualitatively identical results.
4
The details of hierarchical clustering with a K-means approach and its performance are presented
in Appendix 2.
852 L.-j. Chen and C.-d. Fuh

million in nonoption compensation must then value their options at two million in
order to agree to continued employment. Importantly, it may be the case that the
market value of these options is only $100,000, but the CEO subjectively values them
at $500,000 because they believe the market to have undervalued the options.
Although this method of calculation is clearly not perfectly precise, we tried numerous
robustness checks by using different grouping criteria, all of which arrived at quali-
tatively identical results. Some intangible sources of value, such as training, learning
opportunities, and advantageous work environments, are not controlled here but may
be relatively unimportant given that this is an executive database of listed firms.

30.4 Simulation Results

Section 30.3 provided a pricing formula for ESOs that includes illiquidity of the
options and sentiment of the employees. Moreover, from this ESO pricing formula,
we can not only estimate the subjective values but also study the exercise policies.
The exercise boundary is endogenously derived by finding the minimum stock price
such that the option value equals its intrinsic value for each time. In other words, the
employee exercises the option when the stock price is above the exercise boundary.
To illustrate our model, this section discusses factors that affect ESO values and
exercise decisions including stock holding constraint, sentiment, level of risk
aversion, moneyness, dividend, time to maturity, total volatility, and normal
unsystematic volatility. We also empirically test these effects in next section.
According to the collected data from Compustat, the model parameters stock
price S, strike price K, total volatility s, dividend yield d, interest free rate r, and
time to maturity t are set to 25, 25, 0.3, 2 %, 5 %, 10, respectively. Normal
unsystematic volatility n is two-thirds of the total volatility following calibrations
applied by Compustat and the majority of papers in the area.5 We employ the
common parameterization for the coefficient of relative risk aversion 1  g ¼ 2 and
three jump size models: double exponential, bivariate jump, and Y ¼ 0 (no residual
value).6 Additionally, default intensity l ¼ 0.01, following Duffee (1999) and
Fruhwirth and Sogner (2006), which use US and German bond data, respectively.

30.4.1 Exercise Behavior

Employees exercising their ESOs earlier is a pervasive phenomenon. Considering the

exercise policies is necessary when studying American ESOs. This is an essential
departure from Chang et al. (2008), which considers European-type ESOs. A number

5
See Bettis et al. (2005), Aggarwal and Samwick (2003), Ingersoll (2006), and Bryan et al. (2000).
6
The parameters of double exponential are estimated by daily return data from 1992 to 2004.
A jump occurs if return goes beyond 10 %, which relates to an approximately three-standard
deviation daily return during this period.
30 Statistics Methods Applied in Employee Stock Options 853

of papers link early exercise behavior to under-diversification of employees (Hemmer

et al. 1996; Core and Guay 2001; Bettis et al. 2005). The problem of valuing ESOs
with early exercise is often approximated in practice by simply using the expected
time until exercise in place of the actual time to maturity (Hull and White 2004; Bettis
et al. 2005). The expected time until exercise is estimated from past experience.
However, Ingersoll (2006) mentions that even using an unbiased estimate of the
expected time until exercise will not give a correct estimate of the option’s value.
And this method cannot be used to determine the subjective value because it will be
smaller due to the extra discounting required to compensate the lack of diversification.
A proper calculation must recognize that the decision to exercise is endogenous.
Liao and Lyuu (2009) incorporate the exercise pattern instead of using the expected
time until exercise technique in valuation of ESOs, to which the exercise patterns
are under Chi-square distribution assumption and not derived endogenously. Inger-
soll (2006) derives the exercise boundaries endogenously, while the exercise
policies are restricted constant in time. We extend the method developed in Gukhal
(2001), with a modification to include that an agent faces a constrained portfolio
problem, and derive the time varying exercise policies endogenously.
Which factors cause employees to exercise their options early? Figure 30.1
compares the exercise boundaries for some factors. Note that exercise boundaries
are decreasing function of time in all cases, which are different from the constant
exercise policies in Ingersoll (2006). The more restrictions on the stock holding or
the more risk averse the employee, the lower the exercise boundary. In other words,
because of the impossibility of full diversification, employees who are more
restricted on the stock holding or more risk averse prefer early exercise of their
options. Employees with high sentiment will postpone the exercise timing due to the
brightening prospect of the company. The employees who receive the money-type
options also tend to exercise early. In addition, larger dividends induce employees to
exercise their options sooner. Options with shorter lifetime are more quickly
exercised. Employees do not have much time value in these options and tend to
exercise their options earlier. Employees exercise volatile options early to balance
their portfolio risk, especially for idiosyncratic risk increases. Indeed, our model
findings are consistent with several empirical studies. For instance, Hemmer
et al. (1996), Huddart and Lang (1996), and Bettis et al. (2005) show that early
exercise is a pervasive phenomenon owing to risk aversion and undiversification of
employees. Huddart and Lang (1996) find that exercise is negatively related to the
time to maturity and positively correlated with the market-to-strike ratio and with the
stock price volatility. Hemmer et al. (1996) and Bettis et al. (2005) also find that
stock price volatility has a significant effect on exercise decisions. In high-volatility
firms, employees exercise options much earlier than in low-volatility firms.

30.4.2 Factors’ Effects on ESOs

Understanding the factors that affect ESO values and the exercise decision is
important for firms designing stock option programs. As we mentioned before,
854 L.-j. Chen and C.-d. Fuh

Fig. 30.1 Exercise boundaries. This figure presents the exercise boundaries according to stock
holding constraint a, sentiment s, level of risk aversion 1  g, moneyness In: K ¼ 20, At: K ¼ 25,
Out: K ¼ 30, where K is exercise price, dividend yield d, time to maturity t, total volatility s, and
idiosyncratic risk n, respectively

ESO values and exercise decisions are closely related. Factors affecting the
employees exercise policies will directly influence the valuation of ESOs. This
section discusses the impact of factors on ESOs. The results are shown in
Table 30.1, which presents the studied factors effect on ESO value, discount
ratio, and early exercise premium, where ESO value is calculated by formula
(30.6), discount ratio is defined as one minus the ratio of subjective to market
value, and early exercise premium is the difference between American and Euro-
pean ESO value.
30 Statistics Methods Applied in Employee Stock Options 855

Table 30.1 Factors effect on employee stock options

Panel A: ESO values and discount ratios and early exercise premiums
CA CD Premium CA CD Premium
a ¼ 0.00 9.6487 0.0000 0.1839 s ¼ 0.005 6.1495 0.3120 0.9385
a ¼ 0.25 7.5859 0.2110 0.5796 s ¼ 0.000 6.4611 0.3279 0.8592
a ¼ 0.50 6.4628 0.3278 0.8592 s ¼ 0.005 6.7632 0.3469 0.7504
a ¼ 0.75 5.7706 0.3998 0.9366 s ¼ 0.010 7.1101 0.3655 0.6661
1g¼1 8.5661 0.1122 0.4010 sok ¼ 25/30 6.4228 0.2321 0.4413
1g¼2 7.5859 0.2110 0.5796 sok ¼ 25/25 7.5859 0.2110 0.5796
1g¼3 6.9702 0.2811 0.9958 sok ¼ 25/20 9.3061 0.1733 1.0268
d ¼ 0.00 8.4788 0.3557 0.3372 t¼5 5.4222 0.2545 0.4084
d ¼ 0.01 7.5724 0.3260 0.8031 t ¼ 10 6.7636 0.3018 1.1608
d ¼ 0.02 6.7704 0.3054 1.1729 t ¼ 15 7.4605 0.3257 1.9932
d ¼ 0.03 6.0673 0.2941 1.4674 t ¼ 20 7.8992 0.3363 2.8481
s ¼ 0.15 5.4571 0.1800 0.0677 n ¼ 0.1 7.1077 0.2607 0.2400
s ¼ 0.30 6.5440 0.3216 0.9413 n ¼ 0.2 6.2785 0.3470 1.0146
s ¼ 0.45 7.0701 0.4378 2.1005 n ¼ 0.3 5.6321 0.4142 2.1113
s ¼ 0.60 7.2776 0.5184 3.4397 n ¼ 0.4 5.1440 0.4650 3.0637
Panel B: Vesting effect
CA CD Premium
VP ¼ 0 VP ¼ 2 VP ¼ 4 VP ¼ 0 VP ¼ 2 VP ¼ 4 VP ¼ 0 VP ¼ 2 VP ¼ 4
a ¼ 0.25 7.5859 7.5660 7.4923 0.2110 0.2125 0.2200 0.5796 0.5651 0.4898
a ¼ 0.50 6.4628 6.4029 6.2639 0.3278 0.3335 0.3479 0.8592 0.8041 0.6635
a ¼ 0.75 5.7706 5.6867 5.5299 0.3998 0.4081 0.4243 0.9366 0.8570 0.6989
This table presents the impact of factors on ESOs, including stock holding constraint a, sentiment
s, level of risk aversion 1  g, moneyness sok, dividend yield d, time to maturity t, total volatility
s, idiosyncratic risk n, and vesting period VP. CA, CD and Premium are the ESO value, discount
ratio (1-subjective/market) and early exercise premium, respectively

Unlike traditional options, ESOs usually have a vesting period during which they
cannot be exercised and employees are not permitted to sell their ESOs. In this
situation, employees receive the ESOs in a very illiquid market. Table 30.1 shows
that subjective values (a 6¼ 0) are uniformly smaller than the market values (a ¼ 0).
These results are consistent with Lambert et al. (1991) and Hall and Murphy (2002),
in which the subjective value is lower than market value due to the constrained
fixed holding in the underlying stock. The more risk averse the employee (more
positive 1  g) and more restrictions on the stock holding (larger a), the more they
lean to depreciate the option values and incur the higher early exercise premium.
Note that early exercise effect on ESO values cannot be ignored in these situations.
Because of the restriction of ESOs, many employees have undiversified portfolios
with large stock options for their own firms. Therefore, a risk-averse employee
discounts the ESO values. Discount ratios increase with stock holding constraint
and the degree of risk aversion. In other words, employees who are more risk averse
and more restricted on the stock holdings need to compensate more risk premium.
856 L.-j. Chen and C.-d. Fuh

Employee sentiment enhances the option value and reduces the early exercise
premium. One would expect for options with high sentiment having higher discount
that the option value declines sharply when employees face an undiversification
problem. The moneyness of each option Sok is the stock price at issuance divided
by strike price. If the option is in (out of) the money, Sok is greater (less) than 1. In the
money options have higher values, lower discount and higher early exercise pre-
miums. Interestingly, even in the money options having less discount than out of the
money, employees still tend more to early exercise in the money options to diversify
their wealth portfolio risk. Larger dividends depreciate the option values and induce
employees to exercise their options sooner, even they have lower discount ratios.
More interestingly, the early exercise premium is not zero when no dividends are paid.
This is a departure from traditional option theory, although it is consistent with the
phenomenon that ESOs are exercised substantially before maturity date, even ESOs
not paying dividends, because of the lack of diversification. Options with longer
lifetime have more value; at the same time, they have higher discount ratios and
early exercise premiums. Although not reported in the table, the lifetime of option
may be negatively related to European ESO value. This is due to the longer one must
wait and the greater risk caused by undiversification affecting the ESO value.
Whereas options may provide incentives for employees to work harder, they can
also induce suboptimal risk-taking behavior. General option pricing results show
that value should increase with risk while employees need to compensate more risk
premium at the same time. It is not necessarily the case that subjective value is
positive related to risk, as is the traditional result.7 We have the usual finding that
total volatility increases the option value; however, on the contrary, with respect to
normal unsystematic volatility, the subjective value decreases with it. In the Black-
Scholes framework, this risk is eliminated under a risk-neutral measure. However,
in our model, the employee has an illiquid holding and full diversification is
impossible. Hence, a risk-averse employee depreciates the ESO values. The dis-
count and early exercise premium increasing with the volatility risk also can be
found in Table 30.1. This is intuitive, because the more volatile the stock price, the
higher the opportunity cost of not being able to exercise. Therefore, employees have
more incentives to early exercise volatile options.8
Further we examine the effect of vesting on subjective value. Panel B compares
the ESOs that vest immediately, after 2, and 4 years, respectively. Vesting obvi-
ously reduces the ESO values since it restricts the exercise timing. Discount ratio
increasing with vesting period implies that market value is affected less than the
subjective value. Because the constrained ESOs are usually exercised much earlier
than unconstrained ESOs and more tend to fall afoul of the vesting rule. While

7
Nohel and Todd (2005), Ryan and Wiggins (2001), and others show that option values increase
with risk, however, they do not study the impact of increased idiosyncratic risk. Carpenter (2000)
presents examples where convex incentive structures do not imply that the manager is more
willing to take risks. The model used in Chang et al. (2008) is able to capture this result.
8
While Panel A shows the results of options that vest immediately, we can also consider the
vesting effect and there is no significant qualitative difference.
30 Statistics Methods Applied in Employee Stock Options 857

vesting has negative effect on the American ESOs, it has no effect on the European
ESOs, therefore, early exercise premium decreases with vesting.

30.4.3 Offsetting Effect of Sentiment and Risk on ESO

The level of sentiment is estimated from two perspectives. First, we consider the
sentiment effect on ESO value (SenV), and then the estimated sentiment level can be
calculated whereby subjective value with sentiment is equal to market value.
Secondly, we estimate sentiment level from the early exercise perspective (SenE),
that is, what value of sentiment such that employees exercise their options at the time
that unconstrained investors do. The sentiment level of European ESOs (SenVE) is
also calculated. Due to the limitation of European options (they are not allowed to
early exercise), the sentiment level can only be estimated from value perspective.
Estimation of sentiment is shown in Table 30.2. Here, we only list the estimated
sentiment level at the money option because there is no obvious relationship
between sentiment level and moneyness. Table 30.2 shows that sentiment estimated
from the American and European ESO formulas have similar patterns. The more
risk averse the employee and more restricted on the stock holding, the higher the
sentiment level is needed. SenVE is slight higher than SenV because of more
restrictions in the European contract.

30.5 Empirical Study

Applying a comprehensive set of executive options and compensation data, this

study empirically prices both subjective value discount created by stock holding
constraint and the risk-adjusted excess returns necessary for employees to offset the
ESO risk premium, that is, the sentiment effect.

30.5.1 Data

Data for this study are collected from the Compustat Executive Compensation
(Execucomp) database. From this database, all executive stock options issued
between 1992 and 2004 are collected with stock price at issuance S, strike price
K, maturity date T, implied volatility Vol, and dividend yield Div. While the median
option is issued at the money, the mean is in the money (Sok ¼ 1.012). Note that
virtually all options are issued at the money (Sok ¼ 1). Indeed, this is true for about
90 % of our dataset. Average values of time to maturity, implied volatility, and
dividend yield are 9.3 years,9 0.43 % and 1.37 %, respectively.

9
For some issues for which there is no time stamp, we assume an issuance date of July 1, inasmuch
asthis would be the middle of the fiscal year for the vast majority of firms.
858 L.-j. Chen and C.-d. Fuh

Table 30.2 Estimation of sentiment

SenV SenVE SenE
R¼1 R¼2 R¼3 R¼1 R¼2 R¼3 R¼1 R¼2 R¼3
a ¼ 0.25 0.0100 0.0200 0.0300 0.0100 0.0200 0.0300 0.0097 0.0180 0.0291
a ¼ 0.50 0.0199 0.0399 0.0599 0.0200 0.0400 0.0599 0.0194 0.0369 0.0585
a ¼ 0.75 0.0298 0.0598 0.0896 0.0299 0.0598 0.0897 0.0289 0.0579 0.0866
This table presents the sentiment levels necessary to offset the ESO risk premium. Sentiment levels
SenV and SenVE are calculated while the subjective value with sentiment is equal to market value
for the American and European options, respectively. SenE is the value of sentiment such that an
employee exercises their options at the time that unconstrained investors do. a and R ¼ 1  g
represent the stock holding constraint and level of risk aversion

In addition to options data, we collect total compensation data from the

Execucomp database, which includes salary, bonus, restricted stock, option, long-
term incentive pay, and other income earned by executives each year. As can be
seen in Table 30.3, the mean total annual compensation for executives in this
dataset is a bit over $2 million, with a median of just over $1 million. The mean
and median ESO compensation numbers are roughly $1.2 and $0.4 million, respec-
tively. Not surprisingly, chief executives who were also board members received
the highest compensation ($4.2 million), but options are a substantial portion of that
compensation ($2.4 million). Indeed, options compensation generally substantially
outweighs all other forms of compensation.
Following Dittmann and Maug (2007), we further define the net cash inflow
(NCash) for each year as follows:
NCash
¼ Fixed salary ðafter taxÞ
þ Dividend income from shares held in own company ðafter taxÞ
þ Value of restricted stock granted
 Personal taxes on restricted stock that vest during the year
þ Net value realized from exercising options ðafter taxÞ
 Cash paid for purchasing additional stock
Fixed salary is the sum of five Compustat data types: Salary, Bonus, Other
Annual, All Other Total, and long-term incentive pay (LTIP).10 The year when the
executive enters the database is denoted by tE. The executive’s cumulative wealth
for year t is then
Xt1 Yt  
W ðtÞ ¼ NCasht þ NCash‘ 1 þ r sf :
‘¼tE s¼‘þ1

10
For cash paid for purchasing additional stock, where direct data is unavailable, we use the change
in stock holdings times the year-end stock price to calculate this value.
30 Statistics Methods Applied in Employee Stock Options 859

Table 30.3 Compensation summary statistics

Aggregate Mean by title
Mean Median Std Dev B&C B&NC NB&C NB&NC
Salary 365 300 234 556 481 335 286
Bonus 336 151 816 650 479 289 222
Other annual 24 0 179 44 35 22 16
All other total 70 11 540 94 131 50 45
LTIP 77 0 442 127 128 72 48
Restricted stock 163 0 803 366 220 184 101
Options 1,178 378 3,407 2,382 1,683 1,264 748
Total 2,214 1,074 4,262 4,219 3,158 2,217 1,465
This table presents summary statistics for compensation data for four categories of executives:
board & CEO (B&C), board & not CEO (B&NC), not board & CEO (NB&C), and not board & not
CEO (NB&NC). Numbers are reported in 1,000s and LTIP represents the long-term incentive pay

In other words, assume that the executive has no wealth before entering the firm; all
NCasht are realized at the end of the fiscal year and invested at the risk-free rate rt+1
f
during the next fiscal year. Then, a is the sum of all illiquid firm-specific holdings,
including unvested restricted stocks and options, divided by total cumulative wealth.
Average a is about 35 %, implying that the illiquid firm-specific holdings account for
more than one-third of executive total wealth.11 We also calculate a by an iterated
approach that synchronizes a and subjective value simultaneously. Qualitative findings
with respect to sentiment are identical. Several proxies of sentiment are used in an
empirical study including previous year risk-adjusted CAPM alpha and Fortune mag-
azine’s list of top 100 firms to work for, and estimated from our ESO pricing formula.

30.5.2 Preliminary Findings

Substituting the subjective value implied by compensation data into our model
along with the options variables given in our dataset, we are able to back out
sentiment levels, Sen. Results are presented in Table 30.4. There are about 105,000
options issued by each firm (AvgIss) over the test period, with a total of nearly
2,700 firms and 82,000 total observations accounted for. Industry breakdowns,
while exhibiting some fluctuations in point estimates, show that results across
industries are qualitatively similar. While the mean Black-Scholes value of options,
BSOPM, is about $13.09, with some variation across industries, the mean subjective
value, Sub, is more than $19.38, reflecting a 48 % premium. That is, although
virtually all of the theoretical literature implies a subjective value discount, empir-
ical data show that executives generally value ESOs more highly than their Black-
Scholes values. Though not reported in the table, t-tests show that subjective values

11
Holland and Elder (2006) find that rank-and-file employees exhibit an a close to 10 % and concur
that subjective value is decreasing in a because of risk aversion and under-diversification.
860 L.-j. Chen and C.-d. Fuh

Table 30.4 Summary statistics for subjective value and sentiment by industry
Sector BSOPM Sub Sen AvgCom AvgIss Obs
10 Energy 12.589 16.239 0.089 1,774.08 78.46 4,307
15 Materials 10.822 17.613 0.076 1,399.90 61.89 6,412
20 Industrials 12.569 20.305 0.115 1,591.53 70.64 12,134
25 Con. Dis. 12.177 19.799 0.106 2,030.63 98.64 15,925
30 Con. Sta. 12.053 18.121 0.076 2,405.29 111.40 4,347
35 Health care 16.290 21.098 0.115 2,338.68 105.61 8,883
40 Financials 13.440 21.770 0.066 2,892.28 99.37 10,441
45 Inf. Tec. 15.853 18.499 0.229 2,748.47 164.53 14,614
50 Tel. Ser. 12.768 22.941 0.162 5,310.29 272.27 1,324
55 Utilities 5.475 14.633 0.063 1,370.11 67.66 3,931
Others 9.487 9.583 0.208 1,360.34 111.24 56
Total 13.088 19.385 0.120 2,213.73 105.13 82,374
This table presents, by industry: Black-Scholes value BSOPM, subjective value Sub, sentiment
level Sen, average total compensation AvgCom, number of options issued AvgIss, and number of
observations by individual Obs. AvgCom and AvgIss are reported in thousands. Sen is calculated
using the European ESO formula (30.5), where the distribution of jump size follows y ¼ 0. Con.
Dis., Con. Sta., Inf. Tec., and Tel. Ser. refer to Consumer Discretionary, Consumer Staples,
Information Technology and Telecommunication Services, respectively

are statistically significantly higher than Black-Scholes values at the 1 % level for
almost all industries and in aggregate. The only exception is the others industry,
where Sen is still significantly positive but Sub is about equal to BSOPM owing to
a particularly high a in this industry.
Given the large proportion of executive income that is attributed to illiquid, firm-
specific options holdings, this finding suggests substantial overconfidence or pos-
itive inside information regarding their firm’s future prospects. Indeed, the data
show that the average executive prices ESOs such that the firm should outperform
the market’s expectations by an average of 12 % per annum (Sen). T-tests show that
these values are significantly different from zero at the 1 % level in all industries
and in aggregate.
Table 30.5 shows the mean and median values of Rt and Sub in each subsample,
where Rt is the CAPM alpha. Top is a dummy variable taking value 1 if the
executive works for a firm listed in Fortune magazine’s top 100 companies to
work for. Results show that firms with higher previous-year return tend to have
significantly higher subjective values. This is true of both the mean and median
value. Interestingly, subsequent return momentum is not consistently present in this
data, at least as regards mean values. Firms listed in the top 100 in fact make
significantly lower risk-adjusted returns in the year in which they are so listed.
However, they enjoy substantially higher subjective value. This indicates that
sentiment may generally be independent of performance but does significantly
affect subjective value.
Figure 30.2 shows that relative subjective values are greater than one but
relatively stable over time. In contrast, the number of issuances generally increases.
30 Statistics Methods Applied in Employee Stock Options 861

Table 30.5 Difference tests for subjective value and sentiment

Rt  1 > 0 Rt  1 < 0 P-value Top ¼ 1 Top ¼ 0 P-value
mean(Rt) 0.00042 0.00041 0.4493 0.00038 0.00059 <.0001
median(Rt) 0.00031 0.00028 0.0176 0.00030 0.00044 <.0001
mean(Sub) 21.4131 15.6336 <.0001 24.7420 17.6169 0.0003
median(Sub) 14.1350 10.9487 <.0001 17.4147 10.9794 <.0001
This table shows the mean and median values of Rt and Sub in each subsample, where Rt is the
CAPM alpha at time t. Top equals 1 if the firm is listed as a top 100 firm by Fortune magazine in
a given year. The p-values measure the significance of difference tests

AvgIss Sub/BSOPM
200 4.00
150 3.00
100 2.00
50 1.00
0 0.00
1992 1994 1996 1998 2000 2002 2004
Year

Fig. 30.2 Summary AvgIss and Sub/BSOPM by year. AvgIss and Sub/BSOPM for each year are
graphed in this figure. AvgIss, BSOPM, and Sub are number of options issued, Black-Scholes
value, and subjective value, respectively. The y-axis of the histogram is on the left and that of the
line chart is on the right

The industry with the second highest subjective values (Financials) has a below-
average number of issuances. These observations highlight the importance of
looking at pricing, rather than issuance alone, as high subjective values do not
imply that ESOs will be a more popular financing tool.

30.5.3 Implications of Regression Results and Variable Sensitivities

We now shift our attention to the testable implications of our model, namely
confirming the relations between key options’ variables and subjective value.
Specifically, we apply the following regression equation:

Sub ¼ Int þ ba a þ bSen Sen þ bSok Sok

(30.7)
þbt t þ bVol Vol þ bDiv Div þ e,

where Int is the intercept term and all variables are defined as before. Note that for
all results presented here, the calculation of significance is via clustered standard
errors by firm, though OLS results are nearly identical.
862 L.-j. Chen and C.-d. Fuh

Table 30.6 Regression results for subjective value

Int a Sen Sok t Vol Div Obs
Sen ¼ Rt  1
Coefficient 1.8988 0.2126 0.0356 0.0361 0.2813 0.3983 0.0783 56,602
(p-value) (<.0001) (<.0001) (<.0001) (0.2476) (0.0489) (<.0001) (<.0001)
Insider
Coefficient 1.8650 0.3001 0.0628 0.0860 0.2883 0.3409 0.0844 23,826
(p-value) (<.0001) (<.0001) (<.0001) (0.0401) (0.1712) (0.0004) (0.0017)
True sentiment
Coefficient 1.9071 0.1252 0.0100 0.0039 0.2631 0.4344 0.0904 21,333
(p-value) (<.0001) (<.0001) (<.0001) (0.8583) (0.2858) (<.0001) (<.0001)
Sen ¼ Top
Coefficient 1.5014 0.2244 0.0114 0.0802 0.0560 0.2318 0.0807 49,090
(p-value) (<.0001) (<.0001) (0.0169) (0.6641) (0.7216) (<.0001) (<.0001)
Y¼0
Coefficient 2.2013 0.4364 0.0076 0.0424 0.3265 0.3872 0.1012 82,374
(p-value) (<.0001) (<.0001) (<.0001) (0.1194) (0.3041) (<.0001) (<.0001)
Double exp
Coefficient 2.1671 0.4270 0.0017 0.0429 0.3056 0.3786 0.1005 82,374
(p-value) (<.0001) (<.0001) (<.0001) (0.1155) (0.3310) (<.0001) (<.0001)
Bivariate con
Coefficient 2.1799 0.4267 0.0021 0.0428 0.3075 0.3883 0.1023 82,374
(p-value) (<.0001) (<.0001) (<.0001) (0.1162) (0.3284) (<.0001) (<.0001)
This table presents the estimated coefficients from the following regressions:
Sub ¼ Int + baa + bSenSen + bSokSok + btt + bVolVol + bDivDiv + e
where Sub, Int, a, Sen, Sok, t, Vol, and Div refer to the subjective value, intercept term, proportion
of total wealth held in illiquid firm-specific holdings, sentiment, ratio of stock price to exercise
price, time to maturity, implied volatility, and dividend payout, respectively. In the first three tests,
Sen ¼ Rt1, the CAPM alpha. We split the data into two groups according to the sign of the product
of Rt and Sen. When Sen correctly forecasts the sign of the CAPM alpha for a given year, this is
denoted as an “insider.” When Sen and Rt do not match in sign, we denote this as “true sentiment.”
In the fourth test, Sen is a dummy variable that takes value 1 if the firm is in Fortune’s top 100 and
0 otherwise. In the next three tests, Sen is calculated from European ESO formula (30.5), with the
distribution of jump size following y ¼ 0, a double exponential, and a bivariate constant jump
model, respectively

First, we apply gross subjective value Sub as the dependent variable. The first
three tests in Table 30.6 use CAPM risk-adjusted alpha from the year prior to option
issuance Rt1 as a proxy for sentiment under the conjecture that those stocks that
performed better in the previous year generate more positive sentiment prior to
options being issued. Note that our model implies that only the risk-adjusted excess
return should be priced because the market portion of the firm’s return is eliminated
via the risk-neutral measure. Bergman and Jenter (2007), in contrast, test the gross
prior year return. Because a year’s worth of data is required to calculate these
alphas, the dataset is reduced to about 57,000 observations. We find that a is
significantly negatively related to subjective value. This matches our intuition
30 Statistics Methods Applied in Employee Stock Options 863

that the larger the proportion of one’s portfolio held in options, the less diversified
the portfolio, and the less valuable the ESO. Sen, on the other hand, is positively
related, and significantly so. In other words, positive sentiment is associated with
higher subjective value. Note that these results control for the usual options’ pricing
factors. Whereas Div is significantly negative related as expected, Sok and t are not
consistently significantly related, and Vol is negatively related. As explored more
fully later, this last negative relation is quite telling and is consistent with our model
as the sensitivity of subjective value to idiosyncratic risk is negative.
Further, the data are split into two groups according to the sign of the product of
Rt and Sen. A positive (negative) sign implies that the positive sentiment measure is
(not) accompanied by strong performance. The positive case (insider), then, can be
explained by nonsentiment-related factors. The executive may have private inside
information and hence be able to forecast future returns. They also have the ability
to affect future returns so that optimism may be self-fulfilling. The negative case
(true sentiment), on the other hand, has not such a concern inasmuch as it would
imply that positive (negative) sentiment is followed by poor (good) performance.
As it turns out, similar results are obtained in both cases: sentiment is positively
related to subjective value while a is negatively related, both significantly
so. As a result, it is not likely that insider information explains whole sentiment
effect on subjective value.
Next, the Top dummy is selected as a proxy for sentiment. Once again, sentiment
is significantly positively related to subjective value while a is significantly nega-
tively related. All other relations are as above.
We also back Sen out of the European ESO formula (30.5) under the aforemen-
tioned three different jump size assumptions.12 Because our model itself determines
the relation between subjective value and Sen, the purpose of these tests is simply to
observe the other variable relations as well as the stability of the model to the
specification of the jump process. Results are quite consistent across the three
processes tested here. All other coefficients remain qualitatively as before with
the coefficient of a, importantly, remaining significantly negative in all cases.
Finally, in order to more clearly test the difference in impact of sentiment for
insider versus true sentiment events, we interact the event identification dummy
with our sentiment proxy as follows:

Sub ¼ Int þ ba a þ bInSen DIn Sen þ bTSen DT Sen þ bSok Sok þ bt t

þ bVol Vol þ bDiv Div þ e: (30.8)

All variables are defined as before and Sen is the previous-period CAPM alpha,
also as before. DIn is a dummy variable that takes value 1 if the event is insider and

12
Here, sentiment is estimated from the European option formula. It can also be calculated from
the American option formula but with more exhaustive computations. As we mentioned before,
sentiment estimated from the European and American ESO formulas have similar patterns. It may
not affect the regression results much.
864 L.-j. Chen and C.-d. Fuh

Table 30.7 Regression results for insider versus true sentiment events
Int a DInSen DTSen Sok t Vol Div
Sen ¼ Rt1
Coefficient 1.9070 0.2214 0.0372 0.0044 0.0343 0.2866 0.3872 0.0876
(p-value) (<.0001) (<.0001) (<.0001) (<.0001) (0.3769) (0.0876) (<.0001) (<.0001)
This table presents regression results for insider versus true sentiment events, where DIn is
a dummy variable taking value 1 if the event is insider and 0 otherwise, DT takes value 1 if it is
true sentiment and 0 otherwise, and Sen is again defined as the CAPM alpha

0 otherwise. By analogy, DT takes value 1 if the event is true sentiment and

0 otherwise. The results appear as in Table 30.7. Note that, while sentiment
increases subjectively value significantly in both cases, the impact of sentiment
when the event is likely to be an insider event is much larger. In other words, when
strong prior performance reveals real information regarding future performance
that may be known to managers, the impact on subjective value is strong. When
prior performance proves not to be informative, the impact on subjective value is
small. However, the impact is positive and significant in both cases.

30.5.4 Subjective Value and Risk

We now turn our attention to the sensitivity of subjective value to risk. While we
note that our model implies a positive relation between total risk and subjective
value, it further dictates that the sensitivity of subjective value to idiosyncratic risk
is negative, a notion supported by our empirical findings. This indicates that
increased levels of risk may negatively affect subjective value owing to the inability
of executives to fully diversify their holdings. In contrast, the Black-Scholes as well
as the majority of options pricing models prescribe no role to idiosyncratic risk, that
is, the sensitivity should be zero, and are generally not be able to capture the
empirical finding that subjective value is negatively related to risk.
In applying the empirical data to the formulae for the sensitivities of subjective
value to various forms of risk, our model does indeed generate a negative relation
between firm-specific risk and subjective value, a finding that is consistent also with
the empirical observations of Meulbroek (2001). This finding is particularly impor-
tant as managers can easily affect the firm’s idiosyncratic risk level through various
moral hazard-related activities.
In Table 30.8, risk sensitivities are calculated, vegas, for all options issues in our
dataset assuming there are no illiquid holdings (UV), that is, a ¼ 0, and using our default
value for a (V), with and without consideration of sentiment. The first two columns find
as expected that the sensitivity with respect to total risk is positive, for both UV and V,
regardless of whether sentiment is considered or not. This is true of all jump specifica-
tions. In every case, the sensitivity is higher when sentiment is not considered. Looking
at the vegas with respect to jump frequency, UV(freq) can be either positive or negative
depending on the jump specification, while V(freq) is always negative. Interestingly,
30 Statistics Methods Applied in Employee Stock Options 865

Table 30.8 Summary statistics for vega

Panel A: Y ¼ 0
UV(total) V(total) UV(freq) V(freq) V(idio) V(size)
Without sentiment
Mean 11.237 13.715 12.787 0.739 24.212 0.162
Median 9.173 11.684 10.159 0.143 19.611 0.077
With sentiment
Mean 5.968 7.646 16.486 11.222 41.638 0.415
Median 3.007 4.906 13.191 2.857 31.063 0.144
Panel B: Double exponential jump model
UV(total) V(total) UV(freq) V(freq) V(idio) V(size)
Without sentiment
Mean 13.456 16.791 1.039 0.564 25.877 0.017
Median 10.875 14.129 0.871 0.449 20.396 0.008
With sentiment
Mean 7.397 10.063 14.217 1.213 44.233 0.042
Median 4.173 7.041 1.823 0.806 32.466 0.014
Panel C: Bivariate constant jump model
UV(total) V(total) UV(freq) V(freq) V(idio) V(size)
Without sentiment
Mean 13.366 16.792 1.194 0.677 25.860 0.017
Median 10.879 14.131 1.015 0.542 20.382 0.008
With sentiment
Mean 7.396 10.060 15.984 1.407 44.231 0.042
Median 4.173 7.037 2.103 0.957 32.468 0.014
This table presents test results for vega. In Panels A, B, and C, the distribution of jump sizes are
zero jump, double exponential jump, and bivariate constant jump, respectively. UV(total) and UV
(freq) are total risk and jump frequency risk vegas under our model when all holdings are liquid.
V(total), V(idio), V(freq), and V(size) refer to total risk vega, idiosyncratic risk vega, jump
frequency risk vega, and jump size risk vega, respectively

UV is positive for the constant jump model but negative for the other two models,
pointing out the importance of jump specification when liquidity is not also considered.
The magnitude of UV is always smaller than that of V.
Perhaps the most interesting factor affecting our subjective value in our model is
idiosyncratic risk, for which the estimate is always negative and is significantly
larger in magnitude than the other vegas. While the jump size vega also plays a role
and is likewise always negative, the magnitude of this effect is much smaller. This
finding highlights the role of idiosyncratic risk in our model and explains why the
empirical sensitivity of subjective value to volatility is found to be negative,
contrary to generally accepted moral hazard models that dictate that option com-
pensation encourages risk taking. If agents are sufficiently under-diversified, the
risk premium from taking on excess idiosyncratic risk offsets gains from convexity
and discourages risk-taking behavior. The corresponding UVs for idiosyncratic and
jump size risk are both zero as these do not play a role in determining market value
866 L.-j. Chen and C.-d. Fuh

when there are no under-diversified holdings. Also, the Vs are substantially more
negative when sentiment is introduced, pointing out the sharply offsetting effects of
positive sentiment and risk aversion in this model. Which piece dominates then
depends on the risk aversion parameter and a of the employee.

30.6 Conclusion

This chapter applies three statistics methods in ESO study, including change of
measure, hierarchical clustering with a K-means approach, and estimation
of standard errors in finance panel data. We use a model for employee stock options
that incorporates illiquidity of the options, a jump diffusion for the stock price
evolution that includes various jump processes, and the potential roles of employee
sentiment and insider information in a world where employees balance their wealth
among the company’s stock, the market portfolio, and a risk-free asset. Our option
contract is American type and the optimal exercise boundary is derived endoge-
nously. From the ESO pricing formula, we are able to not only estimate the
subjective values but also study the exercise policies.
The subjective value placed on ESOs implied by compensation data is calculated
by applying empirical data. Specifically, using data provided by Compustat, execu-
tives are grouped by a hierarchical clustering with a K-means approach based on
a number of firm and individual criteria. By assuming that all executives in the same
cluster receive the same total compensation, a notion that relies on the existence of
competitive labor markets, we then back out the valuation placed by each executive on
their respective ESO. These groups include consideration of nonoption compensation,
rank, industry, year, firm size, and immediate exercise value. Though the extant
literature predicts that employees should discount the value of their options, we find
that executives in fact value their options more highly than implied by Black-Scholes,
applying an average premium of 48 %. As such, the cost of issuance for the firm is
vastly lower than the benefit perceived by employees, suggesting that ESO compen-
sation should be an even larger part of executive compensation. We then relate
subjective value to sentiment levels and generate the novel finding that executives
must expect their firm’s risk-adjusted returns to outpace that predicted by the market
by 12 % in order to justify the subjective value placed on ESOs. This expectation may
be the result of private information regarding the growth prospects of the firm.
Testing subjective value and its relation to pertinent variables, subjective value is
negatively related to the proportion of wealth held in illiquid firm-specific holdings and
positively related sentiment. In other words, the larger the illiquid ESO position is, the
larger the discount risk aversion prescribed and the lower the subjective value implied in
the compensation package. On the other hand, the more positive the employee’s view of
future risk-adjusted returns, the more valuable the ESO. Interestingly, subjective value
may be negatively related to risk as the inability of executives to fully diversify their
holdings may lead to risk premia that outweigh the value placed on risk by the convexity
of options payouts. Note that this relation is particularly negative with regard to
idiosyncratic risk and is empirically also negative for risk associated with both jump
30 Statistics Methods Applied in Employee Stock Options 867

frequency and size. Because these aspects of return are precisely those that may be most
directly controlled by executives, traditional moral hazard arguments relating solely to
the convexity of the options payout may not hold.
Firms increasingly grant nontraditional employee stock options to link stock price
performance and managerial wealth and provide greater incentives to employees.
While this study focuses on the traditional employee stock option, the main intuition
can be involved in nontraditional ESOs. Premium stock option, performance-vested
stock option, repriceable stock option, purchased stock option, reload stock option,
and index stock option are the objects of future study. We can derive the option
formulas and compare the value, incentive effect, and cost per unit of subjective
incentive across the nontraditional ESOs and the traditional ones. This future study
provides a firm a proper compensation vehicle according to its characteristics.

Appendix 1: Derivation of Risk-Neutral Probability by Change of

Measure

Here we introduce the idea for deriving the ESO formulas and define using
probability measure P* by the technique of change of measure.
The process of an employee’s marginal utility or the pricing kernel is:
dJ W  
^ dW m þ ^n dW s þ Y^  1 dN t ,
¼ ^r dt  s
JW
where
 
1
^r ¼ r  ð1  gÞ a2 n2 þ a2 glk2 þ alk ,
2
mm  r
^¼
s ,
sm
^n ¼ ð1  gÞa n,
^ ¼ ½aðY  1Þ þ 1g1 ,
Y
then
  YNt
1 2 1 2 ^ i:
J W ½W ðtÞ, t ¼ J W ½W ð0Þ, 0exp ^r  s ^ W m ðtÞ þ ^n W s ðtÞ
^  ^n t þ; s Y
2 2 i¼0

Let B(t, T) be the price of a zero coupon bond with maturity T, then

J W ½W ðT Þ, T 
Bðt; T Þ ¼ E BðT; T ÞjF t g
J W ½W ðtÞ, t
(   Y )
Nt
1 2 1 2 ^i
¼ E exp ^r  s ^  ^n t þ s ^ W m ðtÞ þ ^n W s ðtÞ Y
2 2 i¼0

¼ expf½^r þ lðx  1Þtg,

868 L.-j. Chen and C.-d. Fuh

where
  n o
^ ¼ E ½aðY  1Þ þ 1g1 :
t ¼ T  t, x ¼ E Y

The bond yield

1
r  lnBðt; T Þ
Tt
1 
¼ r  ð1  gÞðalk þ g lk2 a2 þ a2 n2  lðx  1Þ:
2
Let
  YNt
1 2 1 2 ^i ,
ZðtÞ e J W ¼ exp
rt
 s ^ W m ðtÞ þ ^n W s ðtÞ
^  ^n  lðx  1Þt þ s Y
2 2 i¼0

then Z(t) is a martingale under P, and we have

J W ½W ðT Þ, T  Z ðT Þ
¼ erðTtÞ :
J W ½W ðtÞ, t Z ðtÞ

The rational equilibrium value of the ESO at time t, F(St, t), satisfies the Euler
equation,
 
J W ½W ðT Þ, T 
Fð S t ; t Þ ¼ Et Fð ST ; T Þ
J W ½W ðtÞ, t
¼ erðTtÞ Et ½FðST ; T Þ,
dP ¼ Zð0Þ, and Et is the expectation under P and information at time t. Under
 Z ðt Þ * *
where dP

the probability measure P , the processes W m ¼ W m  s
*
^ t and W s ¼ W s  ^n t are
Brownian motions, Nt is a Poisson process with intensity l* ¼ lx, and the jump
sizes follow density f*Y(y),

1
f Y ðyÞ ¼ ½aðy  1Þ þ 1g1 f Y ðyÞ:
x

Therefore,

dS
¼ ms dt þ ss dW m þ ndW s þ ðY  1ÞdN t
S

¼ r  d  ð1  gÞan2  lk dt
þ ss dW m þ ndW s þ ðY  1ÞdN t

ðr   d Þdt þ sN dW t þ ðY  1 dN t ,
30 Statistics Methods Applied in Employee Stock Options 869

where

 
 1  2
d ¼ d  ð1  gÞ a lk þ g lk2 a  ð1  a an
2
2
 lðx  1Þ þ lk,
s2N ¼ s2s þ n2 ,
sN W t ¼ ss W m þ nW s :

In other words,
  YNt
  1 2 
St ¼ S0 exp r  d  sN t þ sN W t Yi:
2 i¼0

Appendix 2: Valuation of European ESOs

The option price at time t is



CE ðSt ;tÞ ¼ erðTtÞ Et ½ST  K þ

(( )þ )
h    i Y
NT
rt    1 2 
¼e Et St exp r  d  sN t þ sN W T  W t
Yi  K
2
i¼N t
(
h  i
1
¼ S t E exp d   s2N t þ sN W t
2
)
Y
Nt n o

Y i I ðW  a1 Þ Kert E I ðW  a1 Þ
t t
i¼0
( (  ))
h i YNt Y Nt
dt   1 2  
¼ St e E E exp  sN t þ sN W t
Y i I ðW  a1 Þ  Y i
2
i¼0
t  i¼0
( "  #)
Y Nt

 Kert E E I ðW  a1 Þ  Y i
t  i¼0
(  )   
Y Nt
a 1  sN t a1
d  t  rt 
¼ St e E Y i F  pffiffiffi  Ke E F  pffiffiffi
i¼0
t t
( " # )
X1
ðlxtÞj elxt Y j
     
d t  rt 
¼ St e E Y i F d 1  Ke E F d2
j¼0
j! i¼0
870 L.-j. Chen and C.-d. Fuh

Fig. 30.3 Natural log of total compensation. The box plots show the natural log of total
compensation for the two largest industries in our sample. Executives are grouped according to
position, the firm’s total market value, nonoption compensation, and the immediate exercise
value of the options for each industry by using a hierarchical clustering with K-means
approach

where
h Yj  i  
ln St i¼0 Y i =K þ r   d þ 12 s2N t
d 1 ¼ pffiffiffi ,
sN t
pffiffiffi pffiffiffi
d 2 ¼ d 1  sN t, a1 ¼ d2 t:

Appendix 3: Hierarchical Clustering with a K-Means Approach

The compensation-based subjective value is calculated by assuming that all

executives within the same cluster receive the same total compensation. For
each executive in this cluster, the implied subjective value is derived by compar-
ing the difference between nonoption compensation and the average compensa-
tion. It is very important to group executives appropriately. We use cubic
clustering criterion to generate number of groups and split executives according
to position, the firm’s total market value, nonoption compensation, and the
immediate exercise value of the options for each industry by hierarchical cluster-
ing with a K-means approach. For each group, we calculate the average total
compensation for all executives. If all executives receive the same compensation,
on average, any differences in salaries, bonuses, and other income should be
accounted for by options.
Figure 30.3 presents box plots of the natural log of total compensation for the
two largest industries in our sample: Consumer Discretionary and Information
Technology. With the exception of some outliers, which are subsequently
removed in our main tests, the boxed areas generally do not overlap from cluster
30 Statistics Methods Applied in Employee Stock Options 871

to cluster, demonstrating the relative homogeneity of firms within each cluster

and generally distinctly separated from other clusters. As a result, we believe
that compensation characteristics within each cluster should be quite compara-
ble, lending a measure of credence to our method of calculating subjective
value.

References
Aggarwal, R. K., & Samwick, A. A. (2003). Performance incentives within firms: The effect of
managerial responsibility. Journal of Finance, 58, 1613–1649.
Agrawal, A., & Mandelker, G. N. (1987). Managerial incentives and corporate investment and
financing decisions. Journal of Finance, 42, 823–837.
Bergman, N. K., & Jenter, D. (2007). Employee sentiment and stock option compensation. Journal
of Financial Economics, 84, 667–712.
Bettis, C. J., Bizjak, J. M., & Lemmon, M. L. (2005). Exercise behavior, valuation, and the
incentive effects of employee stock options. Journal of Financial Economics, 76, 445–470.
Bryan, S., Hwang, L., & Lilien, S. (2000). CEO stock-based compensation: An empirical analysis
of incentive-intensity, relative mix and economic determinants. Journal of Business, 73,
661–693.
Carpenter, J. N. (2000). Does option compensation increase managerial risk appetite? Journal of
Finance, 55, 2311–2331.
Chang, C., Fuh, C. D., & Hsu, Y. H. (2008). ESO compensation: The role of default risk, employee
sentiment, and insider information. Journal of Corporate Finance, 14, 630–641.
Chang, C., Chen, L. J., & Fuh, C. D. (2013). The pricing of risk and sentiment: A study of
executive stock options. Financial Management, 42, 79–99.
Chen, L. J., & Fuh, C. D. (2011). A study of employee stock options and the exercise decision.
International Research Journal of Finance and Economics, 67, 142–159.
Core, J. E., & Guay, W. (2001). Stock option plans for non-executive employees. Journal of
Financial Economics, 61, 253–287.
Dittmann, I., & Maug, E. (2007). Lower salaries and no options? On the optimal structure of
executive pay. Journal of Finance, 62, 303–343.
Duffee, G. R. (1999). Estimating the price of default risk. The Review of Financial Studies, 12,
197–226.
Fruhwirth, M., & Sogner, L. (2006). The Jarrow/Turnbull default risk model-evidence from the
German market. European Journal of Finance, 12, 107–135.
Gabaix, X., & Landier, A. (2008). Why has CEO pay increased so much? Quarterly Journal of
Economics, 123, 49–100.
Gukhal, C. R. (2001). Analytical valuation of American options on jump-diffusion processes.
Mathematical Finance, 11, 97–115.
Hall, B. J., & Murphy, K. J. (2002). Stock options for undiversified executives. Journal of
Accounting and Economics, 33, 3–42.
Hemmer, T., Matsunaga, S., & Shevlin, T. (1996). The influence of risk diversification on the early
exercise of employee stock options by executive officers. Journal of Accounting and Econom-
ics, 21, 45–68.
Hemmer, T., Kim, O., & Verrechia, R. K. (2000). Introducing convexity into optimal compensa-
tion contracts. Journal of Accounting and Economics, 28, 307–327.
Hodge, F. D., Rajgopal, S., & Shevlin, T. (2009). Do managers value stock options and
restricted stock consistent with economic theory? Contemporary Accounting Research, 26,
899–932.
Holland, L. C., & Elder, E. M. (2006). Employee stock options in compensation agreements:
A financing explanation. Journal of Corporate Finance, 12, 367–379.
872 L.-j. Chen and C.-d. Fuh

Huddart, S., & Lang, M. (1996). Employee stock option exercises: An empirical analysis. Journal
of Accounting and Economics, 21, 5–43.
Hull, J., & White, A. (2004). How to value employee stock options. Financial Analysts Journal,
60, 114–119.
Ingersoll, J. E. (2006). The subjective and objective evaluation of incentive stock options. The
Journal of Business, 79, 453–487.
Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: A review. ACM Computing
Surveys, 31, 264–323.
Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs,
and ownership structure. Journal of Financial Economics, 3, 305–360.
Lambert, R. A., Larcker, D. F., & Verrecchia, R. E. (1991). Portfolio considerations in valuing
executive compensation. Journal of Accounting Research, 29, 129–149.
Liao, F. Y., & Lyuu, Y. D. (2009). An expanded model for the valuation of employee stock
options. Journal of Futures Markets, 29, 713–735.
Meulbroek, L. K. (2001). The efficiency of equity-linked compensation: Understanding the full
cost of awarding executive stock options. Financial Management, 30, 5–44.
Nohel, T., & Todd, S. (2005). Compensation for managers with career concerns: The role of stock
options in optimal contracts. Journal of Corporate Finance, 11, 229–251.
Oyer, P., & Schaefer, S. (2005). Why do some firms give stock options to all employees?
An empirical analysis of alternative theories. Journal of Financial Economics, 76, 99–133.
Petersen, M. (2009). Estimating standard errors in finance panel data sets: Comparing approaches.
Review of Financial Studies, 22, 435–480.
Ryan, H. E., & Wiggins, R. A. (2001). The influence of firm- and manager-specific
characteristics on the structure of executive compensation. Journal of Corporate Finance, 7,
101–123.
Shreve, S. E. (2008). Stochastic calculus for finance II: Continuous-time models. New York:
Springer.
Structural Change and Monitoring Tests
31
Cindy Shin-Huei Wang and Yi Meng Xie

Contents
31.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874
31.2 Structural Breaks in Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875
31.2.1 Stationary Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875
31.2.2 Multiple Breaks Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 876
31.2.3 Prediction Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 878
31.2.4 Structural Changes in Long-Memory Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 879
31.3 Persistent Change in Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 884
31.3.1 Tests for a Change in Persistence for I(0) or I(1) Process . . . . . . . . . . . . . . . . . . . . 884
31.3.2 Unit Root with Persistent Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887
31.4 Tests for Special Structural Breaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 888
31.4.1 Bubble Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 888
31.4.2 Cointegration Breakdown Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893
31.5 Monitoring Structural Breaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 896
31.5.1 Monitoring in Comparably Ideal Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 896
31.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 900
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 900

Abstract
This chapter focuses on various structural change and monitoring tests for a class
of widely used time series models in economics and finance, including I(0), I(1),

C.S.-H. Wang (*)

CORE, Université Catholique de Louvain and FUNDP, Academie Louvain, Louvain-la-Neuve,
Belgium
Department of Quantitative Finance, National TsingHwa University, Hsinchu City, Taiwan
e-mail: cindywang9177@gmail.com
Y.M. Xie
School of Business and Administration, Beijing Normal University, Beijing, China
Department of Economics, University of Southern California, Los Angeles, CA, USA
e-mail: yimengxie69@gmail.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 873
DOI 10.1007/978-1-4614-7750-1_31,
# Springer Science+Business Media New York 2015
874 C.S.-H. Wang and Y.M. Xie

I(d) processes and the cointegration relationship. A structural break appears

a change in endogenous relationships. This break could be caused by a shift in
mean, variance, or a persistent change in the data property. In general, structural
change tests can be categorized into two types: one is the classical approach to
testing for structural change, which employs retrospective tests using a historical
data set of a given length; the other one is the fluctuation-type test in
a monitoring scheme, which means for given a history period for which
a regression relationship is known to be stable, we then test whether incoming
data are consistent with the previously established relationship. Several struc-
tural changes such as CUSUM squared tests, the QLR test, the prediction test,
the multiple break test, bubble tests, cointegration breakdown tests, and the
monitoring fluctuation test are discussed in this chapter, and we further illustrate
all details and usefulness of these tests.

Keywords
Cointegration breakdown test • Structural break • Long-memory process •
Monitoring fluctuation test • Boundary function • CUSUM squared test •
Prediction test • Bubble test • Unit root time series • Persistent change

31.1 Introduction

The structural break refers to the phenomenon illustrating the time series comes
across unanticipated and significantly influential shifts. Generally, most macroeco-
nomic and financial time series are subject to occasional structural breaks.
Ignoring the presence of structural breaks can lead to seriously biased estimates
and forecasts and the unreliability of models (see Clements and Hendry 1998,
1999). Since economic activities often experience rather long period, there exist
plausible possibilities that structural breaks could occur. Hence, it is very important
to monitor and modify the changes of the model characteristics. If those model
characteristics were not corrected in real time, it may lead to wrong decisions on the
policy making and investment plans for policy makers and practitioners. Conse-
quently, those wrong decisions could result in the fluctuations or crises in economy
and finance.
Examples in the real world are a large sum of financial crises ever since before
the Industrial Revolution. For instance, ten enormous bubbles took place ahead of
the twenty-first century documented by Kindleberger (2005), and the latest
subprime crisis occurred in 2008. All crises crushed the wealth of the whole
world again and again. After the first reported financial crisis – Dutch tulip bulb
bubble in 1636 – all meltdowns in the economy systems were obviously character-
ized by the deviation from steadily development.
It implies that structural breaks could happen prior to the bubbles or explosions.
Accordingly, an accurate approach to detecting structural breaks correctly can be
treated as a measurement to notice the possible incoming financial crisis or to
monitor market activities. On the other hand, developing a suitable structural
31 Structural Change and Monitoring Tests 875

break-testing mechanism is crucial to regulators and investors to modify decisions

in order to avoid a potential incoming crash.
The first seminal paper of the issue regarding to the structural break tests was
proposed by Chow (1960). Nowadays, structural change tests have been explored
far more than the original Chow test and developed in many different dimensions.
In this chapter, we survey various structural changes and monitoring tests for a class
of widely used time series models in economics and finance, including I(0), I(1),
I(d) processes, the cointegration breakdown test, and bubble tests. The conventional
structural change tests are categorized into three types: the mean shift, the persistent
change, and the parameter change. Since the issues of break numbers and properties
of the data-generating process (DGP) would be the main factors to affect the
performances of the structural tests, it is essential to design the structural tests
with respect to different scenarios.
The rest of this chapter is organized as follows. Section 31.2 illustrates the
changes in mean and coefficients when DGP are stationary I(0) and stationary
long-memory I(d) processes. Section 31.3 focuses on persistent change systemat-
ically. Tests of bubbles and the cointegration breakdown are discussed in Sect. 31.4.
Section 31.5 provides the monitoring tests in real time.

31.2 Structural Breaks in Parameters

Structural break has been an intriguing issue since Chow (1960) test resolves how to
detect a single structural change with known break date, and more ensuing litera-
tures expand that in different dimensions. Break numbers can be multiple while
break dates are no longer required as necessitates. Moreover, econometricians
apply research closely to reality, adjusting these models from classical stationary
process to nonstationary and long memory, which are found more likely to reflect
economic and financial data. Now, we will offer systematical analysis of these
marked developments.

31.2.1 Stationary Processes

It is the Chow (1960) test that firstly detect a potential structural break for a given
break date. Suppose the model is
0 0
yt ¼ bt Xt1 þ et , (31.1)
where our null hypothesis is bt ¼ b for all t, et is a martingale difference sequence
with respect to the s fields, and Xt is a k  1 vector of regressor, which are here
0
assumed to be constant and/or I(0) with EXtXt ¼ SX and, possibly, a nonzero mean.
Particularly, Xt1 can include lagged dependent variables as long as they are
I(0) under the null. When the break date is unknown, a natural solution proposed
by Quandt (1960) calculates break date, and it is extended by Davies (1977) to
general models with parameters unidentified under the null. And Andrews (1993)
876 C.S.-H. Wang and Y.M. Xie

reports asymptotic critical values of this QLR statistics. in the form of Quandt
likelihood ratio (QLR) statistic.
To the contrast of the null, the alternative hypothesis of a single break in some or
all of the coefficients is
bt ¼ b, tr and bt ¼ b þ g, t > r, (31.2)

where r, k + 1 < r < T, is the “break date” (or “change point”) and g 6¼ 0.
When the potential break date is known, a natural test for change in b is the
Chow (1960) test, which can be implemented in asymptotically equivalent Wald,
Lagrange multiplier (LM), and LR forms. In the Wald form, the test for a break at
a fraction r/T through the sample is


SSR1, T  SSR1, r þ SSRrþ1, T
FT ðr=T Þ ¼
 , (31.3)
SSR1, r þ SSRrþ1, T =ðT  2kÞ

where SSR1, r is the sum of squared residuals from the estimation of Eq. 31.1 on
observation 1,. . ., r, etc. For fixed r/T, FT(r/T) has an asymptotic w2k distribution
under the null. However, when break date is unknown, the situation can be more
complicated. An expectable idea is to estimate the break date and then compute
Eq. 31.3 for that break. But since the change point is selected by virtue of an
apparent break at the point, the null distribution of the resulting test is not the same
as if the break date were chosen without regard to the data. Thus means of
determining r/T must be further specified before the distribution of the resulting
test can be obtained. Quandt (1960) and Davies (1977) come up with another FT
statistics called Quandt likelihood ratio, thus QLR statistic:
QLR ¼ MAX FT ðr Þ (31.4)
r¼r 0 , ..., r1

Intuitively, it has power against a change in b even without a known break date.
Although it confuses econometricians for years to calculate its null asymptotic
distribution, Andrews (1993, Table I) reports computable critical values for it. For
more details, please see Stock (1994).

31.2.2 Multiple Breaks Case

This issue of detecting structural breaks becomes even richer after the possible
multiple breaks are considered no matter the numbers and dates of breaks are
known or not. Bai and Perron (1998) proposed a comprehensive treatment for
estimating linear models with multiple structural breaks. They set the multiple
linear regression with m breaks (m + 1 regimes) as follows:

yt ¼ x0t b þ z0 dj þ ut , (31.5)

where j ¼ 1,   , m + 1, T0 ¼ 0, and Tm+1 ¼ T; yt is the observed independent variable,

xt (p  1), and zt(q  1) are vectors of covariates, and b and dj (j ¼ 1,   , m+1) are the
31 Structural Change and Monitoring Tests 877

corresponding vectors of coefficients; ut is the disturbance. Note this is a partial

structural change model in the sense that b is not subject to shifts and is effectively
estimated using the entire sample. When p ¼ 0, however, it is translated to a pure
structural change. Now rewrite the liner regression system in Eq. 31.5 in matrix form as
Y ¼ Xb þ Z þ U , where Y ¼ (y1, . . ., yT)0 , X ¼ (x1, . . ., xT)0 , U ¼ (u1, . . ., uT)0 ,
0 0
d ¼ (d10 , d2, . . ., dm+1)0 , and Z is the matrix which diagonally
partitions Z at the 0
m-partition (T1, . . ., Tm)s, i.e., Z ¼ diagðZ1 ; . . . ; Z mþ1 Þ with Z i ¼ Z T ði1Þþ1 ; . . . ; ZT i .
Following this, apply associated least-squares estimates  of b, and d are
obtained by
minimizing the sum of squared residuals. And let l^ ¼ l^1 ; . . . ; l^m Þ ¼ T^1 =T T^m =T
with corresponding true value l0 ¼ (l01, . . ., l0m). Thus take the sup F type test of no
structural break (m ¼ 0) versus the alternative hypothesis that there are m ¼ k breaks.
Let (T1,. . ., Tk) be a partition such that Ti ¼ [Tli](i ¼ 1, . . ., k). Operationally, to test
a null of no break versus some fixed number of breaks, define
  1 0 1
0
d^ R R Z Mx Z Rd^
0 0
R
ðT  ðk þ 1Þq  pÞ
FT ðl1 , . . . , lk ; qÞ ¼ (31.6)
kq SSRk
0 0 0 0
where R is the conventional matrix such that (Rd) 0 ¼ (d1  d2, . . ., dk  dk + 1) and
Mx ¼ I  X(X0 X)0 X0 . Here SSRk is the sum of squared residuals under the
alternative hypothesis, which depends on (T1, . . ., Tk). Define the following set for
some arbitrary small positive number ϵ: Lϵ ¼ {(l1, . . ., lk); jl(i + 1)  lij  ϵ,
l1  ϵ, lk  1  ϵ}. The reason for this is to restrict each break date to be
asymptotically distinct. And apparently, it is the generation of the sup F test
considered by Andrews (1993) and others for the case k ¼ 1.
Similarly, estimated break points by testing for l versus l + 1 breaks are also
obtained by a global minimization of the sum of squared residuals. Comparing the
difference between the sum squared residuals obtained with l breaks and that
obtained with l + 1 breaks, it is advised to test each (l + 1) segment for the presence
of an additional break. This requires that the magnitude of shifts is fixed
(nonshrinking). More precisely, the test is defined by


 2
FT l þ 1jlÞ ¼ ST T^1 ; . . . ; T^l  min1ilþ1 inf t2Li,  St T^1 ; . . . ; T^i1 ; t; T^i ; . . . ; T^l =^
s , (31.7)

where



Li,  ¼ t; T^i1 þ T^i  T^i1   t  T^i  T^i  T^i1  (31.8)

and s^ 2 is a consistent estimate of s2 under the null hypothesis. Note that for i ¼ 1,



^
ST T 1
, . .., T^ði  1Þ,t, T^i , . .., T^l is understood as ST t; T^1 ;.. .; T^l and for i ¼ l + 1
as ST T^1 ;. ..; T^l ;t .
As well, Bai and Perron (1996) also provide a valid proof for consistency of the
estimated break fractions along with its limiting distributions. They also apply their
model to the situations where autocorrelation could happen and compare it with
sequential estimation.
878 C.S.-H. Wang and Y.M. Xie

31.2.3 Prediction Tests

Prediction tests which are often applied to forecast stock and exchange rate also
lose power when potential structural change is incorporated without reasonable
adjustment. While numerical literatures explore suitable models to fit the reality,
their efforts still fail for rather long period or large forecasting sample as proven in
Welch and Goyal (2008). This means tests for structural stability are equivalently
vital in forecasting. For time series satisfying characters of stationarily I(0),
prediction tests are operated on the difference between observed variables and
predictions and only when it is before the suspected structural change could such
tests be valid. Lutkepohl (1989) proposes a prediction tests for structural stability of
univariate times series and compares its efficacy to the univariate time series
investigated by Lutkepohl (1988).
Lutkepohl (1989) assumes a multiple time series generated by a stationary vector
stochastic process yt ¼ (y1t,. . ., ykt) in autoregressive (AR) representation

X
1
yt ¼ v þ Ai yti þ ui (31.9)
i¼0

and moving average (MA) representation

X
1
yt ¼ m þ Fi uti , F0 ¼ I k , (31.10)
i¼0

Noted that Ik is the (K  K) identity matrix, v is a K-dimensional vector of

constant terms, m is the mean vector of yt, the Ai and F1 are (K  K) coefficient
matrices, and ut ¼ (u1t, . . ., uKt)0 is K-dimensional white noise. In other words, the et
is i.i.d. multivariate normal, ut  N(0, ∑ u). Although the two present models focus
on infinity, it is permitted that they can both be finite-order models. Thus the mean
squared error (MSE) h-step predictor at origin is

X
1
y t ð hÞ ¼ m ¼ Fi utþhi ,
i¼h

and its MSE matrix is

X
h1
SðhÞ ¼ Fi Su F0 :
i¼0

Since yt is Gaussian, the forecast error vector is normally distributed,

eðhÞ ¼ ytþh  yt ðhÞð0, SðhÞÞ: (31.11)

31 Structural Change and Monitoring Tests 879

Therefore, the rest statistics

0
T ðhÞ ¼ eðhÞ SðhÞ1 eðhÞ  w2 ðK Þ (31.12)

for h ¼ 1, 2, . . ., can be used for testing whether a structural change has occurred
after period t. As usual w2(K) denotes the chi-squared distribution with Kdf.
Besides the univariate description of this system, another test relevant for
multivariate systems is also designed by Lutkepohl (1989). Since the vector of
one-step forecast errors to h-step forecast errors is also normally distributed,

eðhÞ ¼ ½eð1ÞðhÞ  N ð0, SðhÞÞ:

Here

SðhÞ ¼ Fh ðI h Su Þ Fh 60 , (31.13)

where denotes the Kronecker product and

_ 1 I k 0F
Fh ¼ I k 0F _ h1 Fh2 I_  (31.14)
k

Hence,

LðhÞ ¼ eðhÞ0 SðhÞ1 eðhÞ  w2 ðhK Þ (31.15)

for h ¼ 1, 2, . . ., is another sequence of statistics that can be used for stability tests.
This L(h) statistics are particularly easy to compute and according to Lutkepohl
(1989), while univariate time series tests can be relevant for multivariate systems,
a structural change in a multivariate system will not be necessarily be reflected in
the univariate models for the individual variables of the system. And Lutkepohl
(1989) points that since the coefficients Fi, Ai, and Su are usually unknown so that
the statistics cannot be computed, the replacement by estimators is needed. Thus the
estimated processes T^ðhÞ ! w2 ðK Þ and L^ðhÞ ! w2 ðhK Þ can match the null hypoth-
esis of structural change. Generally, Lutkepohl (1989) proposes a prediction test for
checking the structural stability of a system of time series. Despite the fact it lacks
power in modelling the form of the change, it serves as a useful tool without
imposing severe restrictions on the data. Another merit is that very few data are
required after the time period or the time point at which a structural change is
suspected.

31.2.4 Structural Changes in Long-Memory Processes

So far long memory has been recognized as one of most common features shared by
macroeconomic and financial time series. Let Xt be a stationary process with auto-
covariances r(k) ¼ cov(Xt, Xt+k), then Xt is said to have long memory if as jkj ! 1,
880 C.S.-H. Wang and Y.M. Xie

 
1
r ðk Þ  L 1 ðk Þj k j 2H2
, H2 ;1
2

where L1(k) is a slowly varying function as jkj ! 1. That is L1(ta)/L1(t) ! 1

as t ! 1 for any a > 0. This property implies that the correlations are not
summable, and the spectral density has a pole at zero. Under suitable conditions
on L1(·), the spectral density

1 X 1
f ðx Þ ¼ r ðkÞeikx L2 ðxÞjxj1=2H
2p k¼1

is jxj ! 0 for some L2(·) slowly varying the origin. The explanation for this
phenomenon is not elusive since these economic data always persist for a rather
long period and dependence between time thus has a lasting influence. Yet tradi-
tional wisdoms usually appear stuck in analysis of long memory, and thus need for
structural break detection requires econometricians to develop singular new esti-
mations for use. Among these manufactory efforts attempt to deal with long
dependence either invariant to time or matching the time future. Hereafter more
methods will be discussed.

31.2.4.1 Stochastic Volatility Model with Long-Memory Property

Macroeconomic and financial data often display long-run dependence between
different economic individuals. Researchers and practitioners prefer the synthe-
sized form to describing these issues, and thus data aggregation becomes another
focus in econometrics literature. Interestingly such time series often turn out
distinct from the originals, for example, the singular stock price return is normally
stationary I(0), while the index often displays long memory. The Standard Poor’s
500 index would be a supporting evidence for it. Furthermore Hsiao and Robinson
(1978) and Granger (1980) have pointed that it makes sense for contemporaneous
aggregation with stationary heterogenous autoregressive-moving average pro-
cesses. Portfolio arrangement in this sense also regards to connect its volatility
with the long-memory feature. As discussed in Zaffaroni (2000) and Kazakevicius
et al. (2004), converse results also exist in the generalized autoregressive condi-
tional heteroskedasticity (GARCH). Thus whether a long memory can be obtained
by aggregation in volatility models has attracted a lot of attentions.
Zaffaroni (2006) investigates a large class of volatility models for the memory
implications. With this aim, he selects the class of square root stochastic
autoregressive volatility (SR-SARV), which nests both GARCH and stochastic
volatility (SV) models. So the exponential SV model of Taylor (1986) and the
nonlinear moving average model (nonlinear MA) of Robinson and Zaffaroni
(1998) would be the referents. The result shows long memory is ruled out for the
former but is permitted for the latter. To put it more evidently, set a stationary square
integrable process xt with increasing filtration Jt if xt is Jt-adapted, E(xtjJt1) ¼ 0, and
var(xtjJt1) ¼: ft1. And the SR – SARV(1) process then satisfies
31 Structural Change and Monitoring Tests 881

f t ¼ o þ gt1 þ vt , (31.16)

where the sequence vt is assumed as E(vtjJt1) ¼ 0. o and g are constant nonneg-

ative coefficients with g < 1. The implication means E((x)2t ) < 1. It is also assumed
that oi, gi are independent identically distributed (i.i.d.), randomly drawn from
a joint distribution such that gi < 1. At each point n heterogeneous units xi,t(1) are
observed as SR-SARV(1).

31.2.4.2 Testing for a Change in the Long-Memory Parameters

The best known models representing long dependence are fractional Gaussian noise
(Mandelbrot van and Ness 1968) and fractional ARIMA (Granger and Joyeux
1980). These models are stationary with a constant long-memory parameter H.
But this assumption could not hold for some time series since the long dependence
structure might change over time. It makes sense in macroeconomic data for some
economic events persist over long period or has a lasting influence. Note that
changes of H are relevant particularly for that the rate of convergence of confidence
intervals of constants and for parameter estimates in regression with certain classes
of design matrices, e.g., polynomial regression would change if H changes. More
details are referred to Yajima (1988) and Beran (1991). There are huge sums of
literature focused on this issue, among which Beran and Terrin (1996) offer
a simple estimation by exact or approximate maximum likelihood. The method
testing such structural break consists of the alternative that H is not constant. They
make use of functional central limit theorem to compute the quadratic forms and
offer more test statistics.
Define the spectral density by a finite dimensional parameter vector
y ¼ (t, ) ¼ (t, H, 2, . . ., m) so that
ðp
f ðx; yÞ ¼ tf fx; ð1; Þg, log f fx; ð1; Þgdx ¼ 0:
p

By definition, t is the expected mean squared error of the best linear prediction
of Xt given Xs, s  t  1. The long-memory behavior is characterized by H, the
additional parameters 2, . . ., m allow for flexible modelling of short-term features.
Define
ðp
ak ð  Þ ¼ eikx f 1 fx; ð1; Þgdx:
p

Given X1, X2, . . ., XN, let ^ be the value of  that minimizes

X
N


Q ð Þ ¼ aij ðÞ Xi  X Xj  X :
i, j¼1

For Gaussian processes, the asymptotic covariance matrix is the same as for the
exact maximum likelihood estimator. As proved by Beran and Terrin (1996), for
0 < t < 1,
882 C.S.-H. Wang and Y.M. Xie

X
bNtc



Q1 ðt; Þ ¼ aij ðÞ Xi  X Xj  X ,
i, j¼1
X N


Q2 ðt; Þ ¼ aij ðÞ Xi  X Xj  X :
bNtcþ1

Let ^ð1Þ ðtÞ and ^ ð2Þ ðtÞ ð j ¼ 1, 2Þ be defined by


^ð jÞ ðtÞ ¼ argmin Qj t; ^ð jÞ Þ ð j ¼ 1, 2Þ,

and denote by H^ ðtÞ ¼ ^ j1 ðtÞ the corresponding estimates of H. Moreover, define

j

k2 ¼ 2D
11 (), where D is the k  k matrix with elements
1

ðp
∂
Dij ¼ ð2pÞ1 log f fx; ð1; Þgdx:
p ∂i

Then the process

n o
e N ðtÞ :¼ N 12 k1 ftð1  tÞg2 H^ ð1Þ ðtÞ  H^ ð2Þ ðtÞ
1
Z

converges in the Skorokhod topology on D[0, 1] to the Gaussian process Z(t)

defined by
 
1 1 1
ZðtÞ ¼ ftð1  tÞg 2 B1 ðtÞ  B2 ð1  tÞ (31.17)
t 1t

where B1 and B2 are two independent standard Brownian motions. Based on this
distribution, the test statistics is suggested as

T N ¼ supd<t<1d Ze N ðt Þ (31.18)

for some 0 < d < 1. Also it is implied that TN in distribution, where

Y ¼ supd<t<1d jZ ðtÞj (31.19)

1
and Z(t) is defined by Eq. 31.17. Note that, due to the standardization by ftð1  tÞg2,
Z(t) is a standard normal random variable for each fixed t.
Comparatively, tests for a change in parameter values at a given time point are
proposed in linear regression models with long-memory errors. Hidalgo and
Robinson (1996) design a structural change test specially for I(d) data. They derive
the new test in terms of stochastic

0 and non-stochastic regressors. Generally, their
model of form is yt ¼ b nt xt þ ut , where xt is a K  1 vector of observable
regressors, the prime indicates transposition, and b(s) is a K  1 vector such that
31 Structural Change and Monitoring Tests 883

bðsÞ ¼ bA , 0  s  t
¼ bB , t < s < 1,

for a known number t between zero and one, where the unobservable error ut has
auto-covariances featuring long memory. Thus the null hypothesis is H0 : bA ¼ bB
against the alternative H1 : bA 6¼ bB.
For the case of non-stochastic regressors, a type of Wald testing procedure is
applied. For a given t, define h ¼ [tn], where [·] indicated integer part. Correspond-
ingly, let X1 ¼ (x1, . . ., xh)0 , X2 ¼ (xh+ 1, . . ., xn)0 , Y1 ¼ (y1, . . ., yh)0 , Y2 ¼ (yh+ 1, . . ., yn)0 ,
and then estimate bA and B by

0 1 0
b^A ¼ X1 X1 X1 Y 1 ,

0 1 0
b^B ¼ X2 X2 X2 Y 2

Put

0 1 0
0 1 0 0
W ¼ X1 X1 X1 ,  X2 X2 X2 ,
0
u ¼ ð u1 ; . . . ; un Þ ,

Set wt the tth column of W0 . Then H0 is equivalent to

b^A  b^B ¼ W u:
0

Assume ut is Gaussian with zero mean, then

 
b^A  b^B  N 0; W
0
(31.20)

where G is the n  n Toeplitz matrix (g(s  t) ).

When xt is stochastic but independent of ut, however, Eq. 31.20 can be translated as

 0 1=2  
W GW b^A  b^B  N ð0, I k Þ:

Hidalgo and Robinson (1996) also allowed for a degree of generality in the
specification of the error structure, and the test is mainly designed for liner
regression model by applying a semiparametric but a parametric model. Besides,
the change point is assumed to be already known. Yet Kuan and Hsu (1998)
point that such a test would incorporate size distortion that could lead to
a change point while there is none. This is not peculiar since according to
Kuan and Hsu (1998) when the memory parameter is in (0, 0.5), many well-
known structural changes are supposed to suggest a nonexistent change. They
also indicate a spurious change which might arise from stationary data with long
884 C.S.-H. Wang and Y.M. Xie

memory while not responsible to nonstationary. They thus infer distinguishing

between long-memory series with a change would not be reliable sometimes.
More specifically, more and more financial literature illustrate that the realized
volatility, which plays a major role in forming the portfolios, displays long-
memory properties; thus, in order to avoid the loss caused by the market
fluctuations, we could use those structural break tests for long-memory processes
to detect the turning points of the realized volatility and further adjust the
portfolio.

31.3 Persistent Change in Time Series

Permanent property shifts of the time series are characters of many key macroeco-
nomic and financial variables in developed economies. Hence, correctly character-
izing the time series whether stationary or nonstationary would be very helpful to
build the accurate models. This issue has been investigated in terms of the data
property change from I(0) to I(1) or the vice versa. The change might be responsible
for portfolio adjustment especially for financial crises. Furthermore, recent studies
indicate that other than those classical structural break types, there could also exist
persistent changes in processes characterized by long memory, for example, the
long-memory parameter changes from d1 to d2 or the vice versa, where d is the
fractional differencing parameter.

31.3.1 Tests for a Change in Persistence for I(0) or I(1) Process

A normal question in testing for the persistent change is about what the process used
to be and what direction it changes to. To solve that systematically, we could set the
data-generating process (DGP) as following:

yt ¼ d t þ vt , (31.21)

vt ¼ rt vt1 þ et , t ¼ 1, . . . T: (31.22)

Eq. 31.21 consists of the deterministic kernel dt which is a constant either plus
linear time trend or nothing. The error term et is stationary.
Within the model (31.21), there exist four possibilities. The first is that yt is all
the time I(0) as rt ¼ 1 for all t, which can be defined as H1, and the second is
denoted as H01, reflecting Yt changing from I(0) to I(1) at time bt*c, where b·c
represents the integer part. In this situation, rt ¼ r, jrj < 1 t  bt*c and rt ¼ 1 for
t > bt*c in the context of (i). Noted that the change-point proportion t* is assumed
to be unknown while set in L ¼ [LL, LU], which is an interval symmetric around 0.5
and between zero and one. The third denoted as H10 is that yt is form I(1) to I(0) at
time t*.
31 Structural Change and Monitoring Tests 885

Extant tests to detect the above structural changes are CUSUM of squares-based
tests and regression-based tests. While the former has been improved by Leybourne
et al. (2006) as displaying no tendency to spurious over-rejection, modified tests
from regression angle by LKSN still lack the power for special structural change,
taking large sample, for instance, where tests reject H1 with probability one even
though there is no change. Owing to these reasons, we here provide the CUSUM
test according to Leybourne et al. (2006) as the ideal way in detecting this kind of
persistent change.
In order to test the null hypothesis of constant I(0) behavior H1 against a change
in persistence from I(0) to I(1), H01, a standardized CUSUM of squared subsample
OLS residuals is needed:
2
X½tT 
½ tT  v^2
t¼1 t, t
K f ð tÞ ¼ (31.23)
w^ 2f ðtÞ

where
X
½tT 
v^t, t ¼ yt  yðtÞ with yðtÞ ¼ ½tT 1 yt :
t¼1

And similarly H(10) equivalent to change from I(0) to I(1) in the reversed series,
let xt
y(T  t + 1), occurring at time (T  [t*T]) and the reversed series can be
obtained as
XT½tT 
ðT  ½tT Þ2 e
v 2t, t
K r ð tÞ ¼ t¼1
(31.24)
^ 2r ðtÞ
o

where
X
T btT 
e
v t, t ¼ xt  xð1  tÞ with xð1  tÞ ¼ ðT  ½tT Þ1 xt :
t¼1

Notice for both forward and reversed series, the long-run variance o is replaced
by the estimator

X
m
^ 2f ðtÞ ¼ ^g 0 þ 2
o ws, m ^g s , (31.25)
s¼1

X
btT c
^g s ¼ btT c1 D^
v t, t D^
v ts, t , (31.26)
t¼1

Therefore, a test against persistent change could be based on the ratio R of the
forward and reverse CUSUMs of squared statistics, i.e.,
886 C.S.-H. Wang and Y.M. Xie

inf t2L K f ðt Þ N
Rð t  Þ ¼ r  ¼:
inf t2L K ðt Þ D

where L is as described before and t is assumed to be unknown. Furthermore,

denote the demeaned, written as zt ¼ 1, and de-trend, zt ¼ (1,t)0 , respectively.
As T ! 1, l2/T ! 0,
inf t2L Lfx ðtÞ
R) : (31.27)
inf t2L Lrx ðtÞ

As shown in Leybourne et al. (2006), this limiting distribution of R is not

dependable on the long-run variance, thus R could still hold if it is formed by the
unstandardized ratio:
þ
inf t2L K
R¼ r (31.28)
inf t2L K ðtÞ
where

þ
X
½tT  X
T ½tT 
K ðtÞ ¼ ½tT 2 K ðtÞ ¼ ðT  ½tT Þ2
r
v^2t, t and e
v 2t, :
t¼1 t¼1

And the large sample behavior of R, N, and D under both H01 and H10 will be
provided in their Theorem 2 along with consistent estimators of t*. Additionally,
the results of R in their Theorem 2 imply a consistent test of H1 against H01 or H10,
respectively, by the left-tail and the right-tail distribution of R, while the
unstandardized form yields a consistent estimate of break date t*. What’s more,
even the direction of the change is unknown, such a test with two tails can be
appropriate against either H01 and H10. As for the limiting distribution of R under
H0, it remains available for both the demeaned and de-trend cases. Thus, we could
further know that under H0, R !p 1.
Although Leybourne et al. (2006) rule out the size distortion when applied to
process displaying constant persistence, they neglect such a test for long-memory
case could cause over-rejection, which is often very serious. Sibbertsen and Kruse
(2009) provide a solution based on simulation, by which they find despite that
critical values may need modified, estimators for break dates still hold consistently.
In this sense, the yt generated from DGP is assumed to follow an ARFIMA(p,d,q)
process, which means only the memory parameter d determine the degree of
integration of yt. Then the hypothesis should be written as

H 0 : d ¼ d0 for all t,
H1 : d ¼ d1 for t ¼ 1, . . . , ½tT 
d ¼ d2 for t ¼ ½tT  þ 1, . . . , T

Typically under H0 the memory parameter d0 is restricted to [0, 3/2), while d1 2

[0, 1/2) and d2 2 (1/2, 3/2]. It is equivalent to changes from stationary to
31 Structural Change and Monitoring Tests 887

nonstationary long memory at some unknown breakpoint with the vice versa also
valid. Similarly, parts constituting R statistics are thus defined by

½tT 
X X
T ½tT 
K f ðtÞ ¼ ½tT 2d0 v^t, t , r ðtÞ ¼ ðT  ½tT Þ2d0 v^t, t :
t¼1 t¼1

However, Sibbertsen and Kruse (2009) find that asymptotic property has
changed since the limiting distribution depends strongly on the memory parameter,
which still leads to heavy size distortion. To deal with this problem, they simulate
the adjusted critical values which are dependent on d. Although in this way
practitioners benefit as such sample size will not influence critical values as
significantly much as memory parameter, they have to decide the exact d for this
simulation.

31.3.2 Unit Root with Persistent Change

Conventional unit root tests usually seem problematic when they are applied to
nonstationary volatility, although they are proved to perform well in other situa-
tions. Breaks following a stationary Markov switching process and time-varying
conditional variances are proved not to cause significant size distortion and impact
on unit root and cointegration tests. Similarly, stationary time-varying conditional
variances including (ARCH) are also known to have no impact on unit root and
cointegration tests. Conversely, it can be different for permanent changes in
volatility (so that volatility is nonstationary), since it is easy to greatly affect unit
root inference. Hamori and Tokihisa (1997) show that a single abrupt increase in the
innovation variance increases the size of augmented Dickey and Fuller (1979,
1981) (ADF) tests when no deterministic components are present. The similar
situation happens to Kim et al. (2002) who report severe oversizing in the
constant-corrected ADF tests in the presence of an early single abrupt decrease in
the innovation variance. Cavaliere (2004) proposes a more general framework for
investigating the effects of permanent changes in volatility by showing that the
limiting null distributions of the Phillips and Perron (1988) (PP) test statistics
depend on a particular function of the underlying volatility process which can
lead to either oversizing or undersizing in the tests.
Cavaliere and Taylor (2006) consider both smooth volatility changes and mul-
tiple volatility shifts and construct a numerical solution to obtain approximate
quantiles from the asymptotic null distributions of the standard unit root statistics.
They focus on M unit root tests by Perron and Ng (1996) which is still robust when
applied to autocorrelation. Set the following model to generate the time series
process
 0
  0

Xt  gt ¼ a Xt1  g Zt1 þ ut t ¼ 1, 2, . . . , T (31.29)
888 C.S.-H. Wang and Y.M. Xie

where zt is a vector of deterministic components, thus set as zt ¼ (1, t, . . .,t p)0 , while
et ¼ st et, et  iid (0, 1) which is responsible for the heterogeneity. And besides X0 is
of Op(1). Furthermore, the M statistics for a given sample {Xt}T0 is defined as

T 1 X^T  T 1 X^0  s2AR ðkÞ

2 2
M Za :¼ XT :
2T 2 t¼1 X^t1
2

where X^ are the OLS residuals from the regression of Xt on zt, t ¼ 0,. . ., T and s2AR(k)
t

is an autoregressive estimator of the (non-normalized) spectral density at frequency

zero of {ut}. Specifically,
 2 Xk
^ 2 = 1  b^ð1Þ ,
s2AR ðkÞ :¼ s b^ð1Þ :¼ i¼1
b^i ,

where b^i , i ¼ 1,. . ., k, and s

^ 2 are, respectively, the OLS slope and variance
Xk
estimators from the regression equation DX^t ¼ X^t1 þ bi DX^ti þ et:k ,
i¼1
where the lag truncation parameter satisfies the following assumption. Details
about the asymptotic distribution are referred to Cavaliere and Taylor (2006).
Moreover, the Cavaliere and Taylor method performs as a general way to deal
with the persistent change in volatility. However, in real world it usually cannot be
identified easily. In other words, it smooths off structural change and hold for any
process, providing another conveniently operative way for this issue.

31.4 Tests for Special Structural Breaks

There also exist several types of structural breaks that have exclusive features
distinctly from common economic phenomena. For example, a financial bubble
being originated or bursting indicates a significant and catastrophical structural
break, which is capable of wiping out the great wealth of people as much as
possible. Additionally, the cointegration, recognized as a common relationship
between economic individuals especially in international finance and macroeco-
nomics, could be with breaks. Thus, testing a break in cointegrating relationship
correctly could be a solution of financial crises.

31.4.1 Bubble Tests

31.4.1.1 Definition of the Bubble

We first present the definition of the bubble. A bubble is always characterized as an
explosive asset price deviation from its fundamentals regardless of whether such
a deviation is positive or negative. However, investors and regulators often ignore
bubbles rather easily, because there exists no efficient methodology to monitor and
detect those. In general, a bubble could be triggered by the irrational behaviors.
Shiller (2000) conducts this conclusion by observing investors who switch their
31 Structural Change and Monitoring Tests 889

attentions from the specific financial asset to another and points out these behaviors
take no consideration of deviations in fundamentals. More importantly, most
historical events coincided with Shiller’s findings; thus those opinions become
more convincing. A famous example is the remark given by Greenspan, the
ex-chairman of the Federal Reserve Board, which illustrates that it would be fairly
difficult to know when irrational exuberance has unduly escalated asset values and
thus the USA experienced the bubble collapse.
Furthermore, those claims have already been proved. Kirman and Teyssiere
(2005) rebuked that the deviations from fundamental assets are not the outcome
of irrational herding but the significant shifting compositions of expectations, which
result from opinion diffusion processes. This view is backed up by an analysis of
two types of agents in financial market, which are fundamentalists and chartists.
The former denotes agents who believe asset price Pt is related to underlying
fundamental Pt , i.e., a constant P. The latter depends on the Pt which is cumulated
by historical prices. Moreover, there also exists heterogeneity between these agents.
Thus for the former,
X
Mf

Ef ðPtþ1 jI t Þ ¼ Pt þ vj Ptjþ1  Ptj (31.30)
j¼1

where vj, j ¼ 1, . . ., Mf are positive constants and Mf is the memory of the funda-
mentalists. And in return for the latter
X
Mc
Ec ðPtþ1 jI t Þ ¼ hj Ptj (31.31)
j¼0

where hj, j ¼ 1, . . ., M are constants, M is the memory of the chartists. And based
c c

on the two opinions, the market view of prices are synthetic by the weighted
average of these forecast


Em Ptþ1 jI t Þ ¼ wt Ef ðPtþ1 jI t Þ þ ð1  wt ÞEc ðPtþ1 jI t Þ (31.32)

where wt is the proportion of fundamentalists. Such a market view is dynamic since

contacts or meetings happen randomly between agents and often help build their
next investment plans. In this sense, consider kt as the number of fundamentalists at
time t and the remaining N  kt as chartists. Allow some fixed number M of
meetings to take place at each time. Set qt ¼ kt/N for each agents as the observation
to decide the opinion chosen by majority. And thus
 
qi, t ¼ qt þ ei, t , e  N 0, s2q , qi, t 2 ½0; 1:

And it can be easily understood that wt functioned in Eq. 31.32 originates from
the following mechanism,

XN  
1
wt ¼ N 1 ♯ i : qi; t  :
i¼1
2
890 C.S.-H. Wang and Y.M. Xie

Apparently wt varies with time as the qt process, and it is believed to be

dependable on kt reciprocally. Given the two types of agents, the market price
can be written as
Pt ¼ cEm ðPtþ1 jI t Þ þ Z t (31.33)

where c is a constant and Zt is an index of a vector of fundamental variables.

Obviously, the time-varying Wt decides the market view of the prices in the future
and thus is determinant in the real price. As theoretically proved by Kirman and
Teyssiere (2005), Wt switching from zero to one and vice versa could lead to
varying prices around the fundamental level where a change-point process in the
conditional mean definitely exist. It also helps explain that such process never herds
on the extreme and sometimes moves gradually.
Therefore the above two given opinions are agreeable in the aspect that when
prices explode there may be a change point. Accordingly, the implicit change point
is another version of structural break; thus testing for bubbles efficiently could be
viewed as a way to detect the explosive behavior in pricing.

31.4.1.2 Tests for Bubbles

According to the above explanation, many literatures propose suitable testing
mechanism for rational bubbles. As among these tests, a basic solution is based
on a unit root test. Since rational bubbles always manifest explosive characteristics
in prices, these tests such as augmented Dickey-Fuller (ADF) could find this
distinction with stationary process and unit root. Commonly regarded as expecta-
tions of prices and dividends in the future, the current price is determined sly by the
fundamental which refers to dividends if there is no bubble. This means pricing
follow the variation of dividends, and often there is cointegration relationship
between them. Conversely, if a bubble occurs in the prices, explosive behaviors
are expected regardless of what character dividends own. This implication moti-
vated Diba and Grossman (1988) to firstly look for the presence of bubble behavior
by applying unit root tests to t. However, these tests are criticized by Evans (1991)
who questions the validity of the empirical tests since none of them have much
power to detect periodically collapsing bubbles. He supports this criticism by
explaining that a periodically collapsing bubble process can behave much like an
I(1) process or even like a stationary linear autoregressive process, as a result of
which the standard unit root and cointegration tests in this context are not reliable.
Phillips et al. (2009) design a new synthetical unit root test and avoid such
problem. Such a refinery ADF test against the alternative of an explosive root(the
right-tailed) is conducted in an autoregressive specification as
X J

xt ¼ mx þ t1 þ fjtj þ ex, t , ex, t  NID 0; s2x (31.34)
j¼1

where xt stands for log stock or log dividend and the certain order of J is suggested
by Campbell and Perron (1991), while NID denotes independent and normal
distribution. Thus the unit root null hypothesis is H0 : d ¼ 1, and the right-tailed
alternative is H1 : d > 1.
31 Structural Change and Monitoring Tests 891

After this, a series of forward recursive regressions are applied to Eq. 31.34
repeatedly with sequential subsets of the sample data incremented by one observa-
tion at each pass. Set number of observations in the first regression as t0 ¼ [nr0] and
subsequent regressions employ the originating data set supplied by successive
observations giving a sample of size t ¼ [nr] for r0  r  1. And the corresponding
t-statistic by ADFr and ADF1 corresponds to the full sample. Under the null,
ðr
WdW
ADFr ) ð r 0
1=2 ,
W2
0
and ðr
WdW
supr2½r0 ;1 ADF ) supr2½r0 ;1 ð 0 1=2
r
2
W
0

where W is the standard Brownian motion. Comparing supr2½r0 ;1 ADF with the
ð r  ð r 1=2 !
2
right-tailed critical values with supr2½r0 ;1 WdW W will lead to
0 0
a unit root test for against explosiveness. And to pinpoint the structural break, the
recursive test statistics ADFr are needed against the right-tailed critical values of
the asymptotic of the standard Dickey-Fuller t-statistic. In particular, if re denotes
the origination date and rf is the collapse date of explosive behavior in the data,
estimates of these dates are as follows:
n o
r^e ¼ INF s : ADFs > cvadf
bn , ^
r f ¼ inf s^
re inf s : ADF s < cv adf
bn (31.35)
sr0

where cvadfbn ðsÞ denotes the right-sided critical value of ADFs corresponding to
a significant level of bn. Notice the consistent estimation of the bubble period r^e, r^f
requires the significance level bn approaching zero asymptotically and correspond-
ingly cvadf
bn ðsÞ diverging to infinity in order to eliminate type I error as n ! 1. This
can be implemented in a convenient way employing cvadf bn ðsÞ ¼ logðlogðnsÞÞ=100.
Particularly, it replaces the explicit setting for bn which is more complicated. This
test referred by Phillips et al. (2009) is advantageous for the reason that it doesn’t
allow for the possibility of periodically collapsing bubbles, which are often
observed in practical economic and financial applications.
What’s more, there are more other attempts that could avoid potential setbacks
in performance of testing for bubbles. Among them, large sums of progressing
improvement are from unit root tests for explosive characters. For example,
bootstrapping is applied to improve the power of tests, of which Paparoditis and
Politis (2003) make use for the heavy-tail problem. For DGP in Eq. 31.34, it is
assumed that
892 C.S.-H. Wang and Y.M. Xie

^e t ¼ y^e t1 þ et (31.36)

For a test against jyj < 1, the residual block bootstrap method is carried as
pffiffiffi k, k ¼ 1,. . ., B,
follows: for each block bootstrap series satisfying H0, set the sample
and the integer b ¼ b(T) < T is employed such that 1=b þ b= T ! 0 as T ! 1.
Define k ¼ [T( 1)/b] and draw with replacement k integers i0,. . ., ik1 from the set
1,. . ., T  b, and build the bootstrap samples ej(k) as follows:

e1 ðkÞ ¼ ^e 1 , ej ðkÞ ¼ ej1 þ e^im þs , j ¼ 2, . . . , l, l ¼ kb þ 1,

m ¼ ½ð j  2Þ=b, s ¼ j  mb  1, k ¼ 1, . . . , B:

Let y^T be the least-squares (LS) estimator of y in Eq. 31.36 and e y l ðk Þ

be the LS estimator
 of the regression of e j(k) on e j1 (k). Thus H0 would be
rejected if T y^T  1Þ < ql, B ðgÞ, where ql,B(g) is the gth quantile of the distribution
  
of l ey ðÞ  1 when the size for testing is g. Since some financial time series
l
might have heavy tails, the tail index a would make Pr(jetj > x)  xa as x ! 1 for
some a > 0. Yet this modification of standard asymptotic theory for unit root tests is
still valid.
Horvath and Kokoszka (2003) and Jach and Kokoszka (2004) make the mild
hypothesis that the E(et) ¼ 0 and that the et are in the domain of attraction of an
a-stable law with a 2 (1, 2), while Kokoszka and Parfionovas (2004) consider the
more general case a 2 (1, 2] which then includes the Gaussian case a ¼ 2. Chan and
Tran (1989) have shown that under the null hypothesis H0,
ð1
  d La ðtÞdla ðtÞ
T y^T  1 ! x :¼ 0 ð 1 :
La ðtÞdt
2
0

The purpose of the unit root and subsampling tests here is to approximate the
distribution of the unit root statistics x without knowledge of the tail index alpha
which is difficult to estimate.
Another method employing subsampling is advocated by Jach and Kokoszka
(2004) consists in constructing T  b processes which satisfies H0,

e1 ðkÞ ¼ e^k , . . . , eb ðkÞ ¼ e^k þ    þ e^kþb1 , k ¼ 2, . . . , T  b þ 1,

where b is the size of the subsampling blocks. Let e y b ðkÞ be the LS estimator of
the
 regression of e j(k) on e j1 (k). Then the distribution of x would be estimated by
b ey b ðÞ  1 .
Moreover Wu and Xiao (2008) have proposed a procedure similar to
cointegration tests by Xiao and Phillips (2002) to help detect collapsible bubbles
against which Evans (1991) pointed out that unit root tests would lose the power.
31 Structural Change and Monitoring Tests 893

Their procedure is based on the magnitude of variation of the partial sum processes
Xk
Sk ¼ ^e of the residuals of regression (31.36). If there is no bubble, the
t¼1
magnitude of fluctuation of the process Sk is proportional to k1/2, while the presence
of a bubble makes the process Sk diverging to 1. Their statistic denoted by R enjoys
the advantage of no influence from serial correlation and the correlation between
the residuals and the fundamentals under the null hypothesis. As showed below, R
converges to the supremum of a functional of Brownian motions, i.e.,

 
k e
R :¼ max1kT pffiffiffi k1 Sþ
k  T 1 þ d
S T ! dsup 0t1 V ðtÞ
^ e, d T
o
ð 1 ð 1 1 ð
where e ðtÞ ¼ W ðtÞ  tW ð1Þ ,
V W d ¼ W 1 ð tÞ  dW 1 S
0
SS
0
,
d d
0 0 0
S(t) 0 ¼ (1, W2(t)), W1(t) and W2(t) are Brownian motions that are independent
^ t, d is a nonparametric long-run variance estimator.
of each other; o

31.4.2 Cointegration Breakdown Tests

In addition to detecting the structural breaks in a time series, tests for a breakdown
in the cointegration relationship between two nonstationary I(1) process are also of
interest. For example, correlation of global financial market becomes more obvious
as a result of rapid international capital flows, which feature I(1) processes and are
generally the linear combinations of nonstationary time series. As shown by most
empirical evidences, the existing cointegration relationship often comes to the end
prior to the finite crisis. Such a structural break is vital since it is an indicator for
finance crash. Traditional wisdom formulates the cointegration breakdown tests
always from the assumption that the post-breakdown period is relatively long, and
this is often strong and unrealistic especially for practitioners, see Hansen (1992)
and Quintos and Phillips (1993). Andrews and Kim (2003) on the contrary abandon
this assumption and introduce tests for cointegration breakdown with fixed post-
breakdown time length m, under the condition that the sample T + m goes to infinity.
Clearly, their implementation concentrating on end-of-sample could be conve-
niently extended to breakdown tests occurring at the beginning or in the middle
of the sample.
The data-generating process is as follows:
 0
x b0 þ ut for t ¼ 1, . . . , T
yt ¼ 0
x bt þ ut for t ¼ T þ 1, . . . , T þ m

where yt, xt 2 R and xt, b0, bt 2 Rk and the regressors for all time periods are
I(1) processes with potential deterministic and stochastic trend or other stationary
random variables. Thus the null and the alternative hypotheses are
894 C.S.-H. Wang and Y.M. Xie


bt ¼ b0 for all t ¼ T þ 1, . . . , T þ m and
H0 :
8 t : t ¼ 1, . . . , T þ m are stationary and ergodic
u
< bt 6¼ b0 for some t ¼ T þ 1, . . . , T þ m and=or :
H 1 : the distribution of uTþ1 , . . . , uTþm differs from
:
the distribution of u1 , . . . , um

The alternative hypothesis H1 represents a break in cointegrating relationship in

systematical perspectives including (i) a shift in the vector b0 to bt and (ii) a shift in
the distribution of ut from being stationary to being a unit root random variable.
They consider a test statistics in the quadratic form of the “post-breakdown”
residuals u^t : t ¼ T þ 1, . . . , T þ m . The critical value of the test statistics is
determined via a parametric subsampling method, which if the test statistics
exceeds then the test rejects the null hypothesis.

0
0
0
For any 1  r  s  T + m, let Y rs yr , . . ., ys , Xrs xr , . . ., xs , U rs ur , . . ., us .
And the quadratic form will be

0

Pj ðb; oÞ ¼ ðY jðjþm1Þ  Xjðjþm1Þ b o ðY jðjþm1Þ  Xjðjþm1Þ b

and for j ¼ 1,. . ., T + 1. For the P tests in Andrews and Kim (2003), o is some
nonsingular m matrix and Im denotes the m dimensional identity matrix. Naturally,
an estimator of b0 denoting as b^ is based on the least squares with observations
t ¼ r,. . ., s for 1 + m as following:
 0 1 0
b^rs ¼ Xrs Xrs Xrs Y rs

Although other estimators like the fully modified estimator of Phillips and
Hansen (1990) and the ML estimator of Johansen can also be applied, the priority
for explanation is given to LS estimator.
Then the first test statistics, Pa is defined as

  X
T þm  2
Pa ¼ PTþ1 b^1T ¼ yt  xt b^1T :
0

t¼Tþ1

Referred to as a predictive statistic, Pa is the post-breakdown sum of squared

residuals. The motivation for considering this is equivalent to the F statistics
employed to test a single change in the regression just like Chow Test (1969). Set
Pj(b) at a “leave-m-out” estimator, b^ðjÞ as to mirror that b^1T is not dependent on
observations after point T, then

estimator of b using observations indexed by t ¼ 1, . . . , T with
b^ðjÞ ¼
t, . . . , j þ m  1:
31 Structural Change and Monitoring Tests 895

Clearly, b^ðjÞ by the estimators mentioned above is consistent for b0 under suitable
assumptions. Define
 
Pa, j ¼ Pj b^ðjÞ for j ¼ 1, . . . , T  m þ 1:

And the empirical df of Pa,j : j ¼ 1,. . ., T  m + 1 is

X
Tmþ1

F^PðaÞ , T ðxÞ ¼ 1  m þ 1 l Pa, j :
t¼1

Then define the test statistic Pa to be the 1  a sample quantile, q^pa , 1  a and

q^pa , 1  a ¼ INF x 2: F^Pa , T ðxÞ  1  a:

Thus one can reject H0 if Pa > q^pa , 1  a.

However, Pa test’s simulation performance is not satisfying since it often over-
rejects the null hypothesis. Instead, Pb test could have better finite-sample proper-
ties as defined by
   
Pb ¼ PTþ1 b^1ðT þdm=2e and Pb, j ¼ Pj b^ðjÞ for j ¼ 1, . . . , T  m þ 1:

The Pb test supersedes the Pa for it is less variable as the estimator b^1ðTþm=2eÞ
depends on the observation indexed by t ¼ T + 1,. . ., T + 2e.
Moreover, a naturally less variable statistic Pc is dependent on the complete
sample estimator b^1ðTþmÞ :
   
Pc ¼ PTþ1 b^1ðTþmÞ Pc, j ¼ P b^2ðjÞ for j ¼ 1, . . . , T  m þ 1:

where

estimator of b using observations indexed by t ¼ 1, . . . , T with
b^2ðjÞ ¼
t 6¼ j, . . . , j þ 2e1

for j ¼ 1,. . ., T  m + 1. The P tests are suitable for models where errors are
uncorrelated, and including weights in the statistics based on an estimator of the
error covariance matrix will be advantageous if the errors are correlated. Tests with
the weights are the same as Pa  Pc except that o ¼ Im is replaced by
!1
X
T þ1 0
^ 1ðTþmÞ ¼
o 1þ1 U^ j, jþm1 U^ j, jþm1
j¼1
896 C.S.-H. Wang and Y.M. Xie

where
U^ j, jþm1 ¼ Y j, jþm1  Xj, jþm1 b^1ðTþmÞ ,

The estimator o ^ 1  ðT þ mÞ is an estimator of the inverse of the m covariance

0
matrix of the errors o1 0 ¼ (EU1mU1m) .
1

As for the presence of unit root errors from t ¼ T + 1 to t ¼ T + m in a linear

regression model with i.i.d. normal errors, the locally best invariant (LBI) test
statistic is applied. According to Andrews and Kim (2003), set Am(k, l) ¼ min{k, l}
for k, l ¼ 1, . . ., m then the R tests which are aimed to solve the breakdown will be
defined as
Andrews and Kim’s (2003) tests are not consistent because m is fixed as T ! 1
while they are simultaneously asymptotically unbiased. The power of the tests is
determined by the magnitude of the breakdown and m. The former includes the
magnitude of the parameter shift and the magnitude of the unit root error variance,
and the larger is m, the greater is the power. That means failure to reject the null
hypothesis should not be interpreted as strong evidence in favor of stable cointegration.

31.5 Monitoring Structural Breaks

Foregoing tests almost deal with in-sample data, and it may be tempting to apply
them to real-time data. Under this distinct condition, however, traditional method
for detecting structural breaks mainly within a historical data set normally generates
comparatively large chance of mistaken instability, which leads to variant over-
rejection. Typically, such a test will signal a nonexistent break with probability one.
Thus due to the law of the iterated logarithm, no matter how perfectly these
methods perform as one-shot type detection, their sound performance could not
be translated to monitor out-of-sample stability. Conversely, monitoring test could
detect the break on time without these concerns. Focusing on monitoring structural
break, this section will explain how this relatively advantageous test works better
timely. Plus, it is also to provide evidence to support that the good performance still
holds even in long memory, as articulated before to some degree an ideal descrip-
tion of economy issues.

31.5.1 Monitoring in Comparably Ideal Conditions

For a linear regression Yt ¼ X0 bt + et, t ¼ 1, 2,. . ., usually a real-time structural break

test is bounded with the assumption: bt ¼ b0 for t ¼ 1, 2, . . ., m. And the null of the
test is bt ¼ b0 for some t  m + 1. To solve this problem, Chu et al. (1996) come up
with a CUSUM monitoring procedure based on the behavior of recursive residuals
and a fluctuation monitoring procedure based on recursive estimates of parameters.
They use sequential testing to develop tests of structural stability for real-time
economic systems and widen the class of boundary functions of previous monitor-
ing tests, which is outstanding as it matches economic research more exactly.
31 Structural Change and Monitoring Tests 897

Let {Sn} be the partial sum process constructed from ei and assume the process ei
follows the FCLT (functional central limiting theorem) as l ! g1 0 m
1/2
S[ml] )
1
l(l), l 2 [0, 1), where Sn ¼ ∑ i ¼ 1ei and s0 ! n E(Sn) < 1. Chu et al. (1996)
n 2 2

extend Robinson and Siegmund’s (1970) limiting relation as their central

mechanism
 pffiffiffiffi
limm!1 P Sn  mgððn=mÞ, forsome n  1 ¼ PfW ðtÞ  gðtÞ, for some t  0g

where Sn ¼ ∑ tn ¼ 1et, W denotes a standard Brownian motion and g is a stopping

boundary satisfying some regularity conditions. This suggests a stopping function
for the monitoring test. Another limiting result similar to Eq. 31.37 is

 pffiffiffiffi
limm!1 P jSn j  mgððn=mÞ, for some n  m ¼ PfjW ðtÞjðtÞ, t  1g
(31.37)
Robinson and Siegmund (1970) give their proof for an i.i.d. ei in Eq. 31.38 could
hold for a certain class of continuous functions g(t) that satisfy t1/2 g(t) is
ultimately nondecreasing as t ! 1 and other assumptions. Though this is useful,
yet such assumption is not applicable all the time for research. Segen and Sanderson
(1980) provide a partial resolution for they show that under the boundary function
t1/2 g(t) is nondecreasing, stochastic sequence Sn that satisfies the FCLT continues
to hold. Yet such condition turns out more restrictive than that of Robbins and
Siegmund. Chu et al. (1996) propose a new limiting similar relation without
imposing these restrictions. Typically, the boundary function g(t) considered to
be pertinent to economy is given as
n

 1=2 o
P jW ðtÞj  t a2 þ ln tÞ , for some t  1 ¼ 2½1  FðaÞ þ afðaÞ
(31.38)
n
1=2 o

P jW ðtÞj  ðt þ 1Þ1=2 a2 þ lnðt þ 1Þ , some t > 0 ¼ exp a2 (31.39)

where F and f stand for the cdf and pdf, respectively, of a standard normal random
variable.
A monitoring is a stopping time, determined by a detecting statistic (detector) Gn
and a threshold g(m,n), according to tg(Gn)n, Gn > g(m, n). Firstly consider the
situation where the detector is in form of CUSUM. Let
Xn  Xn
0 1

^
b¼ XX X Y be the OLS estimator at time n. Also define
i¼1 i i i¼1 i i
0 0
1
recursive residuals as wk ¼ 0 and wn ¼ ^e =nn1=2 , nn ¼ 1 + Xn(∑n1 i¼1 XiXi) Xn,
^e n ¼ Y n  Xn b^n1 , n ¼ k + 1,. . .m,. . .. Thus the nth-cumulated sum of recursive
0


Xn Xkþ em t
1 1
residuals is Qm t ¼ ^
s o
i¼k i
¼ ^
s i¼k
oi, for ðn  kÞ=em < ðn  k þ 1Þ=em,
where s ^ is a consistent if s, m
e ¼ ðm  kÞ is the integer of ðm e Þt. Thus under H0,
898 C.S.-H. Wang and Y.M. Xie

e 1=2 Qm
t!m t , t 2 ½0; 1Þ ) tðtÞ, t 2 ð0; 1Þ (31.40)

where “)” denotes the weak convergence of the associated probability

measures. According to Chu et al. (1996), as the monitoring starts as m + 1, define
Xkþ½ em ð1þtÞ
em ¼ s
Q ^ i¼mþ1 oi , t 2 [0, 1). In particular, for n/(m  k) < (n + 1)/(m  k),
t
 Xmþn Xm 
e ¼s
Q
m
^ 1 o i  oi , n  1. It follows that t ! m em ) t !
e 1=2 Q
n i¼k i¼k n
½W ðt þ 1Þ  W ðtÞ, t 2 [0, 1). Since this is a Brownian motion, it can be written in
the form of Eq. 31.36
n pffiffiffiffiffiffiffiffiffiffiffiffi  n  o
limm!1 P Qm  m  kg , for some n  1

n
mk 
:
¼ P jW ðtÞj  gðtÞ, for some t  0
m
e
Its mechanism can be explained as that once the path of Q n crosses the
boundary (m  k) g(n/m  k), it will reject the null hypothesis and imply that
1/2

the model suitable for historical period is no longer reliable. Moreover, an extension
of this can be applied to FL monitoring procedure as
n pffiffiffiffiffiffiffiffiffiffiffiffi  n  o
limm!1 P Qm  m  kg , for some n  m

n
mk 
(31.41)
¼ P jW ðtÞj  gðtÞ, for some t  1

It needs notice that CUSUM algorithms, except those recursive residuals, also
hold for such mechanism, e.g., LR statistics for dependent sequences. Yet LR
detector sometimes could be superfluously complex in general and leads to
a complicated computation. According to the above statement, a CUSUM moni-
toring can be implemented as follows:
0
Suppose (i) Yt ¼ Xtb0 + et, t ¼ 1, . . ., m + 1, . . ., where Xt is a k  1 random vector
0
such that m ∑ t¼1Xt and m1 ∑m
1 m
t¼1XtXt converge in probability to b, a
non-stochastic k  1 vector and M, a k matrix of full rank, respectively; (ii) et is
a martingale difference sequence with respect to a sequence of s-algebra Ft such
that E(e2) < 1 and E(e2t jFt1) ¼ s20 for all t, where Ft is generated by. . ., (Yt2,
0 0
Xt1), (Y1, Xt); (iii) the sequence Xtet obeys the functional central limit theorem,
then ( )
m pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
n þ m  k 1=2

e
limm!1 P Q n  n þ m  k a þ ln 2
, for some, n  1
mk
¼ expða2 =2Þ;
(31.42)
  n 1=2 h  n i1=2 
e m pffiffiffiffiffiffiffiffiffiffiffiffi
limm!1 P Q n  m  k a þ ln
2
, some n
mk mk
¼ 2½1  F½a þ afðaÞ
(31.43)
31 Structural Change and Monitoring Tests 899

Additionally, by the right-handed side of Eqs. 31.42 and 31.43, the asymptotic
size control of the CUSUM monitoring is fairly effortless. For CUSUM monitoring
based on Q e m , the 10 % and 5 % asymptotic size correspond to a2 ¼ 4.6 and
n
6, respectively. Equally handily, we obtain the 10 % and 5 % asymptotical size of
the CUSUM monitoring based on Qm by setting a2 ¼ 6.25 and 7.78 in Eq. 31.43,
respectively.
Furthermore, from Eq. 31.36 we also can construct the fluctuation monitoring of
sequential parameter estimates. The key condition is that Xtet obeys the multivariate
FCLT as

1=2
X
½ml
l1=2 V 0 Xt et ) l  W ðlÞ, l 2 ½0; 1Þ
t¼1

0
where V0 ¼ limm!1m1E(SmSm) with Sm ¼ ∑m t¼1Xtet, and W(l) is a k-dimensional
Wiener process. Following this, the fluctuation detector can be defined as
 
Z^n ¼ nD1=2 ^  b^ , n
b (31.44)
m n m

where Dm ¼ M1 1
m V 0 MmP is Op(1) and uniformly positive definite such that

!
X
m
0
Xt Xt =m  Mm ! 0:
t¼1

The essential ingredient of this FL detector is the deviation of the updated

parameter estimate b^n from the historical parameter estimate b^m . Then the
conclusion can be obtained as

ðaÞ : l1=2 ^z ½ml ) l ! W 0 ðlÞ, l 2 ½1; 1Þ,

where W0(l) is a k-dimensional Brownian bridge;

 h n h  n ii1=2 
i 1=2 n  m
ðbÞ : limm!1 P ^z ½ml  m a þ In
2
,
m nm nm
ðfor some  m and some iÞ
¼ 1  ½1  2½1  FðaÞ þ afðaÞk ,

z ½iml is the ith component of Z^½ml . This conclusion summarizes the

i
where ^
monitoring procedure based on the fluctuation of b^n ðn > mÞ relative to b^m .
Similarly, given the number of regressors, k, and arbitrary probability of type
I error, the monitoring boundary can be calculated with a derivative a2. More
details about cases of the one-time parameter shift in both FL and CUSUM
monitoring are referred to Chu et al. (1996).
900 C.S.-H. Wang and Y.M. Xie

31.6 Concluding Remarks

In this chapter, we discuss two classes of structural breaks – the retrospective tests
and the monitoring tests. From the investor’s points of view, it is crucial to use
suitable structural break tests to detect turning points of the financial and macro-
economics data accurately and further adjust the portfolios immediately. With the
increase of the data property caused by highly frequent market fluctuations, the
currently existing structural break tests could not fully detect locations of breaks
without the issue of size distortion. More precisely, more and more sophisticated
structural breaks tests with respect to any possible market performances would be
expected to be created in the near future.

References
Andrews, D. W. K. (1993). Tests for parameter instability and structural change with unknown
change point. Econometrica, 61, 821–856.
Andrews, D. W. K., & Kim, J. (2003). End-of-sample cointegration breakdown tests. Cowles
Foundation Discussion Paper No. 1404.
Bai, J., & Perron, P. (1996). Computation and analysis of multiple structural change models.
Manuscript in Preparation, Universite de Montreal.
Bai, J., & Perron, P. (1998). Estimation and testing linear models with multiple structural changes.
Econometrica, 66, 47–78.
Beran, J. (1991). M-estimators of location for data with slowly decaying serial correlations.
Biometrika, 76, 261–269.
Beran, J., & Terrin, N. (1996). Testing for a change of the long-memory parameter. Biometrika,
83, 627–638.
Campbell, J. Y., & Perron, P. (1991). Pitfalls and opportunities: What macroeconomists should
know about unit roots. National Bureau of Economic Research Macroeconomics Annual, 6,
201–210.
Cavaliere, G. (2004). Unit root tests under time-varying variances. Econometric Reviews, 23,
259–292.
Cavaliere, G., & Taylor, A. M. R. (2006). Testing for unit roots in time series models with
non-stationary volatility. Journal of Econometrics, 140, 919–947.
Chan, N. H., & Tran, L. T. (1989). On the first order autoregressive process with infinite variance.
Econometric Theory, 5, 354–362.
Chow, G. C. (1960). Tests of equality between sets of coefficients in two linear regressions.
Econometrica, 28, 591–605.
Chu, C. J., Stinchcombe, M., & White, H. (1996). Monitoring structural change. Econometrica, 64,
1045–1065.
Clements, M. P., & Hendry, D. F. (1998). Forecasting economic processes. International Journal
of Forecasting, 14, 111–131.
Clements, M. P., & Hendry, D. F. (1999). On winning forecasting competitions in economics.
Spanish Economic Review, 1, 123–160.
Davies, R. B. (1977). Hypothesis testing when a nuisance parameter is present only under the
alternative. Biometrika, 64, 247–254.
Diba, B. T., & Grossman, H. I. (1988). Explosive rational bubbles in stock prices? The American
Economic Review, 78, 520–530.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time
series with a unit root. Journal of the American Statistical Association, 74, 427–431.
31 Structural Change and Monitoring Tests 901

Dickey, D. A., & Fuller, W. A. (1981). Likelihood ratio statistics for autoregressive time series
with a unit root. Econometrica, 49, 1057–1072.
Evans, G. W. (1991). Pitfalls in testing for explosive bubbles in asset prices. American Economic
Review, 81, 922–930.
Granger, C. (1980). Long memory relationships and aggregation of dynamics models. Journal of
Econometrics, 14, 227–238.
Granger, C. W. J., & Joyeux, R. (1980). An introduction to long-memory time series and fractional
differencing. Journal of Time Series, 1, 15–29.
Hamori, S., & Tokihisa, A. (1997). Testing for a unit root in the presence of a variance shift.
Economics Letters, 57, 245–253.
Hansen, B. E. (1992). Tests for parameter instability in regressions with I(1) processes. Journal of
Business and Economic Statistics, 10, 321–336.
Hidalgo, J., & Robinson, P. M. (1996). Testing for structural change in a long-memory environ-
ment. Journal of Econometrics, 70, 159–174.
Horvath, L., & Kokoszka, P. S. (2003). A bootstrap approximation to a unit root test statistic for
heavy tailed observations. Statistics and Probability Letters, 62, 163–173.
Hsiao, C., & Robinson, P. M. (1978). Efficient estimation for a dynamic error-shock model.
International Economic Review, 19, 467–480.
Jach, A., & Kokoszka, P. S. (2004). Subsampling unit root tests for heavy-tailed observations.
Methodology and Computing in Applied Probability, 6, 73–97.
Kazakevicius, V., Leipus, R., & Viano, M. (2004). Stability of random coefficient ARCH models
and aggregation schemes. Journal of Econometrics, 120, 139–158.
Kim, T. H., Leybourne, S., & Newbold, P. (2002). Unit root tests with a break in innovation
variance. Journal of Econometrics, 109, 365–387.
Kindleberger, C. P. (2005). Manias, panics, and crashes: A history of financial crises. New Jersey:
Wiley.
Kirman, A., & Teyssiere, G. (2005). Testing for bubbles and change-points. Journal of Economic
Dynamic Control, 29, 765–799.
Kokoszka, P. S., & Parfionovas, A. (2004). Bootstrap unit root tests for heavy tailed time series. In
S. T. Rachev (Ed.), Handbook of computational and numerical methods in finance
(pp. 175–197). Boston: Birkhäuser.
Kuan, C., & Hsu, C. (1998). Change-point estimation of fractionally integrated processes. Journal
of Time Series Analysis, 19, 693–708.
Leybourne, S., Taylor, R., & Kim, T. (2006). Cusum of squares-based tests for a change in
persistence. Journal of Time Series Analysis, 28, 408–433.
Lutkepohl, H. (1988). Prediction tests for structural stability. Journal of Econometrics, 39,
267–296.
Lutkepohl, H. (1989). Prediction tests for structural stability of multiple time series. Journal of
Business Economic Statistics, 7, 129–135.
Mandelbrot, B. B., & Vanness, J. W. (1968). Fractional Brownian motions, fractional noises and
applications. SIAM Review, 10, 422–437.
Paparoditis, E., & Politis, D. (2003). Residual-based block bootstrap for unit root testing.
Econometrica, 71, 813–855.
Perron, P., & Ng, S. (1996). Useful modifications to some unit root tests with dependent errors and
their local asymptotic properties. Review of Economic Studies, 63, 435–463.
Phillips, P. C. B., & Hansen, B. E. (1990). Statistical inference in instrumental variables regression
with I(1) processes. Review of Economic Studies, 57, 99–125.
Phillips, P. C. B., & Perron, P. (1988). Testing for a unit root in time series regression. Biometrika,
75, 335–346.
Phillips, P. C. B., Wu, Y., & Yu, J. (2009). Explosive behavior in the 1990s Nasdaq: When did
exuberance escalate asset values?. Cowles Foundation Discussion Paper No.1699
Phillips P. C. B., Shi, S. & Yu, J. (2011). Testing for multiple bubbles. Singapore Management
University working paper.
902 C.S.-H. Wang and Y.M. Xie

Quandt, R. E. (1960). Tests of the hypothesis that a linear regression system obeys two separate
regimes. Journal of the American Statistical Association, 55, 324–330.
Quintos, C. E., & Phillips, P. C. B. (1993). Parameter constancy in cointegrating regressions.
Empirical Economics, 18, 675–706.
Robinson, H., & Siegmund, D. (1970). Boundary crossing probabilities for the Wiener process and
sample sums. The Annals of Mathematical Statistics, 41, 1410–1429.
Robinson, P. M., & Zaffaroni, P. (1998). Nonlinear time series with long memory: A model for
stochastic volatility. Journal of Statistical Planning and Inference, 68, 359–371.
Segen, J., & Sanderson, A. (1980). Detection of changes in a time series. IEEE Transaction in
Information Theory, 26, 249–257.
Shiller, R. J. (2000). Irrational exuberance. Princeton: Princeton University Press.
Sibbertsen, P., & Kruse, R. (2009). Testing for a break in persistence under long-range dependen-
cies. Journal of Time Series Analysis, 30, 268–285.
Stock, J. H. (1994). Unit roots, structural breaks and trends, Chapter 46. In Handbook of
econometrics (pp. 2740–2841). Amsterdam: Elsevier
Taylor, S. (1986). Modeling financial time series. New Jersey: Wiley.
Welch, I., & Goyal, A. (2008). A comprehensive look at the empirical performance of equity
premium prediction. Review of Financial Studies, 21, 1455–1508.
Wu, G. & Xiao, Z. (2008). Are there speculative bubbles in stock markets? Evidence from an
alternative approach. Statistics and its Interface, 1, 307–320.
Xiao, Z., & Phillips, P. C. B. (2002). A CUSUM test for cointegration using regression residuals.
Journal of Econometrics, 108, 43–61.
Yajima, Y. (1988). On estimation of a regression model with long-memory stationary errors.
Annals of Statistics, 16, 791–807.
Zaffaroni, P. (2000). Contemporaneous aggregation of GARCH processes. Journal of Time Series
Analysis, 28, 521–544.
Zaffaroni, P. (2006). Aggregation and memory of models of changing volatility. Journal of
Econometrics, 136, 237–249.
Consequences for Option Pricing
of a Long Memory in Volatility 32
Stephen J. Taylor

Contents
32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904
32.2 Long Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 906
32.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 906
32.2.2 Fractionally Integrated White Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 907
32.2.3 Evidence for Long Memory in Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 908
32.2.4 Explanations of Long Memory in Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 909
32.3 Long Memory Volatility Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 910
32.3.1 ARCH Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 910
32.3.2 Estimates for the S & P 100 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 911
32.3.3 Stochastic Volatility Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914
32.4 Option Pricing Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914
32.4.1 A Review of SV and ARCH Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914
32.4.2 The ARCH Pricing Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915
32.4.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 916
32.5 Illustrative Long Memory Option Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 917
32.5.1 Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 917
32.5.2 Parameter Selections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 918
32.5.3 Comparisons of Implied Volatility Term Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 919
32.5.4 Comparisons of Smile Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923
32.5.5 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925
32.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 926
Appendix 1: Series Expansions and Truncation Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 929
Appendix 2: Simulation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 930
Appendix 3: Impact of a Volatility Risk Premium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 931
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 932

S.J. Taylor
Lancaster University Management School, Lancaster, UK
e-mail: s.taylor@lancaster.ac.uk

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 903
DOI 10.1007/978-1-4614-7750-1_32,
# Springer Science+Business Media New York 2015
904 S.J. Taylor

Abstract
Conditionally heteroscedastic time series models are used to describe the
volatility of stock index returns. Volatility has a long memory property in the
most general models and then the autocorrelations of volatility decay at
a hyperbolic rate; contrasts are made with popular, short memory specifications
whose autocorrelations decay more rapidly at a geometric rate.
Options are valued for ARCH volatility models by calculating the discounted
expectations of option payoffs for an appropriate risk-neutral measure. Monte
Carlo methods provide the expectations. The speed and accuracy of the calcu-
lations is enhanced by two variance reduction methods, which use antithetic and
control variables. The economic consequences of a long memory assumption
about volatility are documented, by comparing implied volatilities for option
prices obtained from short and long memory volatility processes.
Results are given for options on the S & P 100-share index, with lives up to
2 years. The long memory assumption is found to have a significant impact upon
the term structure of implied volatilities and a relatively minor impact upon
smile shapes. These conclusions are important because evidence for long mem-
ory in volatility has been found in the prices of many assets.

Keywords
ARCH models • Implied volatility • Index options • Likelihood maximization •
Long memory • Monte Carlo • Option prices • Risk-neutral pricing • Smile
shapes • Term structure • Variance reduction methods

32.1 Introduction

Long memory effects in a stochastic process are effects that decay too slowly to be
explained by stationary processes defined by a finite number of autoregressive and
moving-average terms. Long memory is often represented by fractional integration
of shocks to the process, which produces autocorrelations which decrease at
a hyperbolic rate compared with the faster geometric rate of stationary ARMA
processes.
Long memory in volatility occurs when the effects of volatility shocks decay
slowly. This phenomenon can be identified from the autocorrelations of measures
of realized volatility. Two influential examples are the study of absolute
daily returns from stock indices by Ding et al. (1993) and the investigation of
daily sums of squared 5-min returns from exchange rates by Andersen
et al. (2001b).
Stochastic volatility causes option prices to display both smile and term structure
effects. An implied volatility obtained from the Black-Scholes formula then
depends on both the exercise price and the time until the option expires. Exact
calculation of smile and term effects is only possible for special volatility processes,
32 Consequences for Option Pricing of a Long Memory in Volatility 905

with the results of Heston (1993) being a notable example. Monte Carlo methods
are usually necessary when the volatility process has a long memory and these were
first applied by Bollerslev and Mikkelsen (1996, 1999).
This chapter documents the economic consequences of a long memory
assumption about volatility. This is achieved by comparing implied volatilities
for option prices obtained from short and long memory specifications.
It is necessary to use a long history of asset prices when applying a long memory
model and this chapter uses levels of the S & P 100 index from 1984 to 1998. For
this data it is found that a long memory assumption has a significant economic
impact upon the term structure of implied volatilities and a relatively minor impact
upon smile effects. Some related empirical results are provided by Ohanissian
et al. (2004).
Option traders have to make assumptions about the volatility process.
The effects of some assumptions are revealed by the prices of options with long
lives. Bollerslev and Mikkelsen (1999) find that the market prices of exchange
traded options on the S & P 500 index, with lives between 9 months and 3 years,
are described more accurately by a long memory pricing model than by the
short memory alternatives. Thus these option prices reflect the long memory
phenomenon in volatility, although it is found that significant biases remain
unexplained.
Three explanatory sections precede the illustrative option pricing results in
Sect. 32.5. Section 32.2 defines and characterizes long memory and then reviews
the empirical evidence for these characteristics in volatility. The empirical evidence
for the world’s major markets appears compelling and explanations for the source
of long memory effects in volatility are summarized.
Section 32.3 describes parsimonious volatility models which incorporate long
memory, either within an ARCH or a stochastic volatility framework. The former
framework is easier to use and we focus on applying the fractionally integrated
extension of the exponential GARCH model, which is known by the acronym
FIEGARCH. An important feature of applications is the unavoidable truncation
of an autoregressive component of infinite order. Empirical results are provided for
10 years of S & P 100 returns.
Section 32.4 provides the option pricing methodology. Contingent claim prices
are obtained by simulating terminal payoffs using an appropriate risk-neutral
measure. Numerical methods enhance the accuracy of the simulations and these
are described in Appendix 2.
Section 32.5 compares implied volatilities for European option prices obtained
from short and long memory volatility specifications, for hypothetical S & P
100 options whose lives range from 1 month to 2 years. Options are valued on
10 dates, one per annum from 1989 to 1998. The major impact of the long memory
assumption is seen to be the very slow convergence of implied volatilities to a limit
as the option life increases. This convergence is so slow that the limit cannot be
estimated precisely. Section 32.6 contains conclusions.
906 S.J. Taylor

32.2 Long Memory

32.2.1 Definitions

Definitions which categorize stochastic processes as having either a short memory

or a long memory can be found in Brockwell and Davis (1991), Baillie (1996),
Granger and Ding (1996), and Taylor (2005). The fundamental characteristic of
a long memory process is that dependence between variables separated by t time
units does not decrease rapidly as t increases.
Consider a covariance stationary stochastic process {xt} that has variance s2 and
autocorrelations rt, spectral density f(o), and n-period variance ratios Vn defined by

rt ¼ corðxt ; xtþt Þ, (32.1)

s2 X 1
f ðoÞ ¼ r cos ðtoÞ, o > 0, (32.2)
2p t¼1 t

varðxtþ1 þ . . . þ xtþn Þ X
n1
nt
Vn ¼ ¼ 1 þ 2 r (32.3)
ns2 t¼1
n t

Then a covariance stationary process is here said to have a short memory if

X
n
rt converges as n ! 1; otherwise it is said to have a long memory. A short
t¼1
memory process then has

X
n
rt ! C1 , f ðoÞ ! C2 , V n ! C3 , as n ! 1, o ! 0, (32.4)
t¼1

for constants C1, C2, C3. Examples are provided by stationary ARMA processes.
These processes have geometrically bounded autocorrelations, so that jrtj  Cft
for some C > 0 and 1 > f > 0, and hence Eq. 32.4 is applicable.
In contrast to the above results, all the limits given by Eq. 32.4 do not exist for
a typical covariance stationary long memory process. Instead, it is typical that the
autocorrelations have a hyperbolic decay, the spectral density is unbounded for low
frequencies, and the variance ratio increases without limit. Appropriate limits are
then provided for some positive d < 12 by

rt f ðoÞ Vn
! D1 , ! D2 , 2d ! D3 , as n ! 1, o ! 0, (32.5)
t2d1 o2d n

for positive constants D1, D2, D3. The limits given by Eq. 32.5 characterize
the stationary long memory processes that are commonly used to represent
32 Consequences for Option Pricing of a Long Memory in Volatility 907

long memory in volatility. The fundamental

  parameter d can
 be estimated from data
using a regression, either of ln f^ðoÞ on o or of ln V^n on n as in, for example,
Andersen et al. (2001a)

32.2.2 Fractionally Integrated White Noise

An important example of a long memory process is a stochastic process {yt} which

requires fractional differencing to obtain a set of independent and identically
distributed residuals {et}. Following Granger and Joyeux (1980) and Hosking
(1981), such a process is defined using the filter

d ð d  1Þ 2 d ð d  1Þ ð d  2Þ 3
ð1  LÞd ¼ 1  dL þ L  L þ ... (32.6)
2! 3!

where L is the usual lag operator, so that Lyt ¼ yt1. Then a fractionally integrated
white noise (FIWN) process {yt} is defined by

ð1  LÞd yt ¼ et (32.7)
2
with the et assumed to have zero mean and variance se. Throughout this chapter it is
assumed that the differencing parameter d is constrained by 0  d < 1.
The mathematical properties of FIWN are summarized in Baillie (1996). The
process is covariance stationary if d < 12 and then the following results apply. First,
the autocorrelations are given by

d d ð d þ 1Þ d ð d þ 1Þ ð d þ 2Þ
r1 ¼ , r2 ¼ , r3 ¼ , :::::: (32.8)
1d ð1  d Þð2  d Þ ð1  d Þð2  d Þð3  d Þ

or, in terms of the gamma function,

Gð1  dÞGðt þ dÞ
rt ¼ , (32.9)
GðdÞGðt þ 1  dÞ

with

rt G ð1  d Þ
! as t ! 1: (32.10)
t 2d1 Gðd Þ

Second, the spectral density is

s2e  2d s2e h oi2d

f ðoÞ ¼ 1  eio  ¼ 2 sin , o > 0, (32.11)
2p 2p 2
908 S.J. Taylor

so that

s2e 2d
f ðoÞ ffi o for o near 0: (32.12)
2p

Also,

Vn Gð1  dÞ
! as n ! 1: (32.13)
n2d ð1 þ 2d ÞGð1 þ dÞ

When d  12, the FIWN process has infinite variance and thus the
autocorrelations are not defined, although the process has some stationarity
properties for 12  d < 1.

32.2.3 Evidence for Long Memory in Volatility

When returns rt can be represented as rt ¼ m + stut, with st representing volatility

and independent of an i.i.d. standardized return ut, it is often possible to make
inferences about the autocorrelations of volatility from the autocorrelations of
either jrt  mj or (rt  m)2; see Taylor (2005, 2008). In particular, evidence for
long memory in powers of daily absolute returns is also evidence for long memory
in volatility. Ding et al. (1993) observe hyperbolic decay in the autocorrelations of
powers of daily absolute returns obtained from US stock indices. Dacorogna
et al. (1993) observe a similar hyperbolic decay in 20-min absolute exchange rate
returns. Breidt et al. (1998) find that spectral densities estimated from the loga-
rithms of squared index returns have the shape expected from a long memory
process at low frequencies. Ohanissian et al. (2008) accept the null hypothesis of
long memory for exchange rate volatility, by assessing the long memory implica-
tion that d is invariant under temporal aggregation. Further evidence for long
memory in volatility has been obtained by fitting appropriate fractionally integrated
ARCH models and then testing the null hypothesis d ¼ 0 against the alternative
d > 0. Bollerslev and Mikkelsen (1996) use this test to support long memory
models for US stock index volatility.
Direct evidence for long memory in volatility uses high-frequency data to
construct accurate estimates of the volatility process. The estimated quadratic
variation of the logarithm of the price process during a 24-h period denoted by
t can be estimated from intraday returns rt,j by calculating

X
N
^ 2t ¼
s r 2t, j : (32.14)
j¼1

^ 2t will be very close to the integral of the latent volatility during

The estimate s
the same 24-h period providing N is large but not so large that the bid-ask spread
32 Consequences for Option Pricing of a Long Memory in Volatility 909

and other microstructure effects introduce bias into the estimate. Using 5-min
returns provides conclusive evidence for long memory effects in the estimates s ^ 2t
in four studies: Andersen et al. (2001b) for 10 years of DM/$ and Yen/$ rates;
Andersen et al. (2001b) for 5 years of stock prices for the 30 components of the
Dow-Jones index; Ebens (1999) for 15 years of the same index; and Areal and
Taylor (2002) for 8 years of FTSE-100 stock index futures prices. These papers
provide striking evidence that time series of estimates s ^ 2t display all three
properties of a long memory process: hyperbolic decay in the autocorrelations,
spectral densities at low frequencies that are proportional to o2d, and variance
ratios whose logarithms are very close to linear functions of the aggregation
period n. It is also found that estimates of d are between 0.3 and 0.5, with most
estimates close to 0.4.

32.2.4 Explanations of Long Memory in Volatility

Granger (1980) shows that long memory can be a consequence of aggregating

short memory processes; specifically if AR(1) components are aggregated and if
the AR(1) parameters are drawn from a beta distribution, then the aggregated
process converges to a long memory process as the number of components
increases. Andersen and Bollerslev (1997) develop Granger’s theoretical results
in more detail for the context of aggregating volatility components and also
provide supporting empirical evidence obtained from only 1 year of 5-min returns.
It is plausible to assert that volatility reflects several sources of news, that the
persistence of shocks from these sources depends on the source, and hence that
total volatility may follow a long memory process. Scheduled macroeconomic
news announcements are known to create additional volatility that is very short-
lived (Ederington and Lee 1993), while other sources of news that have a longer
impact on volatility are required to explain volatility clustering effects that last
several weeks.
Gallant et al. (1999) estimate a volatility process for daily IBM returns that is the
sum of only two short memory components, yet the sum is able to mimic long
memory. They also show that the sum of a particular pair of AR(1) processes has
a spectral density function very close to that of fractionally integrated white noise
with d ¼ 0.4 for frequencies o  0.01p. Consequently, evidence for long memory
may be consistent with a short memory process that is the sum of a small number of
components whose spectral density happens to resemble that of a long memory
process except at extremely low frequencies. Attempts to distinguish between true
long memory and short memory models which mimic long memory behavior
include Ohanissian et al. (2008) and Pong et al. (2008).
Barndorff-Nielsen and Shephard (2001) model volatility in continuous time as
the sum of a few short memory components. Their analysis of 10 years of 5-min
DM/$ returns shows that the sum of four volatility processes is able to provide an
excellent match to the autocorrelations of squared 5-min returns, which exhibit the
long memory property of hyperbolic decay.
910 S.J. Taylor

32.3 Long Memory Volatility Models

A general set of long memory stochastic processes can be defined by first applying
the filter (1  L)d and then assuming that the filtered process is a stationary ARMA
(p, q) process. This defines the ARFIMA (p, d, q) models of Granger (1980), Granger
and Joyeux (1980), and Hosking (1981). This approach can be used to obtain long
memory models for volatility, by extending various specifications of short memory
volatility processes. We consider both ARCH and stochastic volatility specifications.

32.3.1 ARCH Specifications

The conditional distributions of returns rt are defined for ARCH models using
information sets It1 which are here assumed to be previous returns {rt  i, i  1},
conditional mean functions mt(It1), conditional variance functions ht(It1), and
a probability distribution D for standardized returns zt. Then the terms

rt  m
zt ¼ pffiffiffiffi t (32.15)
ht

are independently and identically distributed with distribution D and have zero
mean and unit variance.
Baillie (1996) and Bollerslev and Mikkelsen (1996) both show how to define
a long memory process for ht by extending either the GARCH models of Bollerslev
(1986) or the exponential ARCH models of Nelson (1991). The GARCH extension
cannot be recommended because the returns process then has infinite variance for
all positive values of d, which is incompatible with the stylized facts for asset
returns. For the exponential extension, however, ln(ht) is covariance stationary for
d < 12 ; it may then be conjectured that the returns process has finite variance for
particular specifications of ht.
Like Bollerslev and Mikkelsen (1996, 1999), this chapter applies the
FIEGARCH(1, d, 1) specification:

lnðht Þ ¼ a þ ð1  fLÞ1 ð1  LÞd ð1 þ cLÞgðzt1 Þ, (32.16)

gðzt Þ ¼ yzt þ gðjzt j  CÞ, (32.17)

with a, f, d, c respectively denoting the location, autoregressive, differencing, and

moving-average parameters of ln (ht). The i.i.d. residuals g(zt) depend on
a symmetric response parameter g and an asymmetric response parameter y
which allows the conditional variances to depend on the signs of the terms zt;
these residuals have zero mean because C is defined to be the expectation of jzt j. The
EGARCH(1,1) model of Nelson (1991) is given by d ¼ 0. If f ¼ c ¼ 0 and d > 0,
then ln (ht)  a is a fractionally integrated white noise process. In general, ln (ht) is
an ARFIMA(1, d, 1) process.
32 Consequences for Option Pricing of a Long Memory in Volatility 911

Calculations using Eq. 32.16 require truncation of a series expansion in the lag
operator L. The relevant formulae are listed in Appendix 1. We apply the following
ARMA(N, 1) approximation:

X
N   
lnðht Þ ¼ a þ bj ln htj  a þ gðzt1 Þ þ cgðzt2 Þ, (32.18)
j¼1

with the coefficients bj defined by Eqs. 32.31 and 32.32.

32.3.2 Estimates for the S & P 100 Index

Representative parameters are used in Sect. 32.5 to illustrate the option pricing
consequences of long memory in volatility. These parameters are estimated from
daily returns rt for the S & P 100 index, excluding dividends, calculated from index
levels pt as rt ¼ ln (pt/pt1).
The conditional variances are evaluated for t  1 by setting N ¼ 1,000 in
Eq. 32.18, with ln (htj) replaced by a and g(ztj) replaced by zero whenever
t  j  0. The log-likelihood function is calculated for the 2,528 trading days
during the 10-year estimation period from 3 January 1989 to 31 December 1998,
which corresponds to the times 1, 221  t  3, 748 for our dataset; thus the first
1,220 returns are reserved for the calculation of conditional variances before 1989
which are needed to evaluate the subsequent conditional variances.
Results are first discussed when returns have a constant conditional mean which
is estimated by the sample mean. The conditional variances are obtained recur-
sively from Eqs. 32.15, 32.17, 32.18, and 32.32. The conditional distributions are
assumed to be normal when defining the likelihood function. This assumption is
known to be false but it is made to obtain consistent parameter estimates (Bollerslev
and Wooldridge 1992). Preliminary maximizations of the p likelihood
ffiffiffiffiffiffiffiffi showed that
a suitable value for C ¼ E[|zt|] is 0.737, compared with 2=p ffi 0:798 for the
standard normal distribution. They also showed that an appropriate value of the
location parameter a of ln (ht) is 9.56; the log-likelihood is not sensitive to minor
X N
deviations from this level because a is multiplied by a term 1  bj in Eq. 32.18
j¼1
which is small for large N. Consequently, the results summarized in Table 32.1 are
given by maximizing the log-likelihood function over some or all of the parameters
y, g, f, c, d.
The estimates of y and g provide the usual result for a series of US stock index
returns that changes in volatility are far more sensitive to the values of negative
returns than those of positive returns, as first reported by Nelson (1991). When zt is
negative, g(zt) ¼ (g  y)(  zt)  gC, otherwise g(zt) ¼ (g + y)zt  gC. The ratio
(g  y)/(g + y) is at least 4 and hence is substantial for the estimates presented in
Table 32.1.
912 S.J. Taylor

Table 32.1 Parameter estimates for short and long memory ARCH models
Model Constraints y g f c d RLL
AR(1) c¼d¼0 0.06 0.10 0.982 0.0
ARMA(1,1) d¼0 0.09 0.15 0.988 0.44 3.0
FI(d) f¼c¼0 0.12 0.19 0.66 13.9
ARFI(1, d) c¼0 0.11 0.17 0.30 0.59 17.2
ARFIMA(1, d, 1) None 0.10 0.15 0.16 0.98 0.57 21.7
ARFIMA(1, d, 1) f+c+d1 0.12 0.18 0.27 0.68 0.59 19.7
ARFIMA(1, d, 1) d ¼ 0.4, f + c  0.6 0.11 0.18 0.64 0.04 0.4 11.4
Parameters are estimated by maximizing the log-likelihood of daily returns from the S & P
100 index, from 3 January 1989 to 31 December 1998. Returns are modelled as
pffiffiffiffi
r t ¼ m þ ht z t ,
lnðht Þ ¼ a þ ð1  fLÞ1 ð1  LÞd ð1 þ cLÞgðzt1 Þ,
gðzt Þ ¼ yzt þ gðjzt j  CÞ
The zt are assumed to be i.i.d., standard normal variables when defining the likelihood function.
The values C ¼ 0.737 and a ¼  9.56 are used for all likelihood calculations. The relative
log-likelihood (RLL) for a model equals the maximum log-likelihood (MLL) for that model minus
the MLL for the AR(1) model. The MLL for the AR(1) model is 8561.6

The first two rows of Table 32.1 report estimates for short memory
specifications of the conditional variance. The AR(1) specification has
a persistence of 0.982 which is typical for this volatility model. The ARMA
(1,1) specification has an additional parameter and increases the log-likelihood
by 3.0. The third row shows that the fractional differencing filter alone
(d > 0, f ¼ c ¼ 0) provides a better description of the volatility process than
the ARMA(1,1) specification; with d ¼ 0.66 the log-likelihood increases by
10.9. A further increase of 7.8 is then possible by optimizing over all three
volatility parameters, d, f, and c, to give the parameter estimates1 in the fifth
row of Table 32.1.
The estimates for the most general specification identify two issues of concern.
First, d equals 0.57 for our daily data which is more than the typical estimate of 0.4
produced by the studies of higher frequency data mentioned in Sect. 32.2.3. The
same issue arises in Bollerslev and Mikkelsen (1996) with d estimated as 0.63
(standard error 0.06) from 9,559 daily returns of the S & P 500 index, from 1953 to
1990. Second, the sum d + f + c equals 1.39. As this sum equals c1 in Eqs. 32.35
and 32.36, more weight is then given to the volatility shock at time t  2 than to the
shock at time t  1 when calculating ln(ht). This is counterintuitive. To avoid this
outcome, the constraint d + f + c  1 is applied and the results given in the
penultimate row of Table 32.1 are obtained. The log-likelihood is then reduced by
2.0. Finally, if d is constrained to be 0.4, then the log-likelihood is reduced by an
additional 8.3.

1
A conservative robust standard error for our estimate of d is 0.12, using information provided by
Bollerslev and Mikkelsen (1996).
32 Consequences for Option Pricing of a Long Memory in Volatility 913

0.9

0.8

0.7
Coefficient

0.6 d=0, phi=0.988, psi=−0.440

0.5 d=0.4, phi=0.64, psi=−0.04
d=0.59, phi=−0.27, psi=0.68
0.4
d=0.65, phi=0.78, psi=−0.68
0.3

0.2

0.1

0
0 50 100 150 200 250 300 350 400 450 500
Lag

Fig. 32.1 Moving-average coefficients for four ARFIMA(1, d, 1) processes

The estimates obtained here for f and c, namely, 0.27 and 0.68 for the most
general specification, are rather different to the 0.78 and 0.68 given by
Bollerslev and Mikkelsen (1999, Table 1), although the estimates of d are similar,
namely, 0.59 and 0.65. However, the moving-average representations obtained
from these sets of parameter estimates are qualitatively similar. This is shown on
Fig. 32.1 which compares the moving-average coefficients cj defined by (32.36).
The coefficients are positive and monotonic decreasing for the four sets of
parameter values used to produce Fig. 32.1. They show the expected hyperbolic
decay when d > 0 and a geometric decay when d ¼ 0. The values of bj in
Eqs. 32.32 and 32.38 that are used to calculate the conditional variances
decay much faster. For each curve on Fig. 32.1, c10 > 0.33 and c100 > 0.07 while
0 < b10 < 0.02 and 0 < b100 < 0.0003.
The results reported in Table 32.1 are for a constant conditional mean, mt ¼ pm.
ffiffiffiffi
Alternative specifications such as mt ¼ m + brt1, mt ¼ m  12ht, and mt ¼ m þ l ht
give similar values of the log-likelihood when the volatility parameters are set to
the values in the final row of Table 32.1. First, including the lagged return rt1 is
not necessary because the first-lag autocorrelation of the S & P 100 returns
equals 0.022 and is statistically insignificant. Second, including the
adjustment 12ht makes the conditional expectation of (pt  pt1)/pt1 constant
when the conditional distribution is normal. The adjustment reduces the
log-likelihood by an unimportant 0.3. Third, incorporating the ARCH-M
parameter l gives an optimal value of 0.10 and an increase in the
log-likelihood of 1.5. This increase is not significant using a non-robust
likelihood-ratio test at the 5 % level.
914 S.J. Taylor

32.3.3 Stochastic Volatility Specifications

Two shocks per unit time characterize stochastic volatility (SV) models, in contrast
to the single shock zt that appears in ARCH models. A general framework for long
memory stochastic volatility models is given for returns rt by

r t ¼ m þ s t ut (32.19)

with ln(st) following an ARFIMA(p, d, q) process. For example, with p ¼ q ¼ 1,

lnðst Þ ¼ a þ ð1  fLÞ1 ð1  LÞd ð1 þ cLÞvt : (32.20)

This framework has been investigated by Breidt et al. (1998), Harvey (1998),
and Bollerslev and Wright (2000), all of whom provide results for the simplifying
assumption that the two i.i.d. processes {ut} and {vt} are independent. This
assumption can be relaxed and has been for short memory applications (Taylor
1994; Shephard 1996).
Parameter estimation is difficult for SV models, compared with ARCH models,
because SV models have twice as many random innovations as observable vari-
ables. Breidt et al. (1998) describe a spectral-likelihood estimator and provide
results for a CRSP index from 1962 to 1989. For the ARFIMA(1, d, 0) specification,
they estimate d ¼ 0.44 and f ¼ 0.93. Bollerslev and Wright (2000) provide detailed
simulation evidence about semiparametric estimates of d, related to the frequency
of the observations.
It is apparent that the ARCH specification (Eqs. 32.15–32.17) has a similar
structure to the SV specification (Eqs. 32.19–32.20). Short memory special cases of
these specifications, given by d ¼ q ¼ 0, have similar moments (Taylor 1994). This
is a consequence of the special cases having the same bivariate diffusion limit when
appropriate parameter values are defined for increasingly frequent observations
(Nelson 1990; Duan 1997). It seems reasonable to conjecture that the multivariate
distributions for returns defined by (32.15–17) and (32.19–20) are similar, with the
special case of independent shocks {ut} and {vt} corresponding to the symmetric
ARCH model that has y ¼ 0 in Eq. 32.17.

32.4 Option Pricing Methodology

32.4.1 A Review of SV and ARCH Methods

The pricing of options when volatility is stochastic and has a short memory has
been studied using a variety of methods. The most popular methods commence
with separate diffusion specifications for the asset price and its volatility. These
are called stochastic volatility (SV) methods. Option prices then depend on
several parameters including a volatility risk premium and the correlation
between the differentials of the Wiener processes in the separate diffusions.
32 Consequences for Option Pricing of a Long Memory in Volatility 915

The closed-form solution of Heston (1993) assumes that volatility follows a

square-root process and permits a general correlation and a nonzero volatility
risk premium; for applications see, for example, Bakshi et al. (1997) and for
extensions see Duffie et al. (2000).
There has been less research into option pricing for short memory ARCH
models. Duan (1995) provides a valuation framework and explicit results for
the GARCH(1,1) process that can be extended to other ARCH specifications.
Ritchken and Trevor (1999) provide an efficient lattice algorithm for GARCH
(1,1) processes and extensions for which the conditional variance depends on the
previous value and the latest return innovation. Recent innovations are provided by
Christoffersen et al. (2008, 2010).
Methods for pricing options when volatility has a long memory have been
described by Comte and Renault (1998) and Bollerslev and Mikkelsen (1996,
1999). The former authors provide analysis within a bivariate diffusion frame-
work. They replace the usual Wiener process in the volatility equation by
fractional Brownian motion. However, their option pricing formula appears to
require independence between the Wiener process in the price equation and the
volatility process which is not consistent with the empirical evidence for stock
returns.
The most practical way to price options with long memory in volatility is
probably based upon ARCH models, as demonstrated by Bollerslev and Mikkelsen
(1999). We follow the same strategy. From the asymptotic results in Duan (1997),
also discussed in Ritchken and Trevor (1999), it is anticipated that insights about
options priced from a long memory ARCH model will be similar to the insights that
can be obtained from a related long memory SV model.

32.4.2 The ARCH Pricing Framework

When pricing options it will be assumed that returns are calculated from prices
(or index levels) as rt ¼ ln (pt/pt1) and hence exclude dividends. A constant risk-
free interest rate and a constant dividend yield will also be assumed and, to simplify
the notation and calculations, it will be assumed that interest and dividends are paid
once per trading period. Conditional expectations are defined with respect to current
and prior price information represented by It ¼ {pti, i  0}.
To obtain fair option prices in an ARCH framework, it is necessary to make
additional assumptions in order to obtain a risk-neutral measure Q. Duan (1995) and
Bollerslev and Mikkelsen (1999) provide sufficient conditions to apply a risk-
neutral valuation methodology. For example, it is sufficient that a representative
agent has constant relative risk aversion and that returns and aggregate growth rates
in consumption have conditional normal distributions. Kallsen and Taqqu (1998)
derive the same solution as Duan (1995) without making assumptions about utility
functions and consumption. Instead, they assume that intraday prices are determined by
geometric Brownian motion with volatility determined once a day from a discrete-time
ARCH model.
916 S.J. Taylor

32.4.3 Implementation

At time t0 , measured in trading periods, the fair price of a European contingent

claim which has value yt0 +n(pt0 +n) at the terminal time t0 + n is given by
  
yt0 ¼ EQ ern yt0 þn pt0 þn jI t0 (32.21)

with r the risk-free interest rate for one trading period. We now specify an
appropriate way to simulate pt0 +n under a risk-neutral measure Q. Monte Carlo
methods are then used to estimate the conditional expectation in Eq. 32.21.
Following Duan (1995), it is assumed that the distribution of observed returns is
defined by some probability measure P, for which

r t jI t1 P N ðmt ; ht Þ, (32.22)

with

rt  m
zt ¼ pffiffiffiffi t P i:i:d:N ð0; 1Þ (32.23)
ht

It is also assumed that the distributions in a risk-neutral framework are defined

by a measure Q, with
 
1
r t jI t1 Q N r  d  ht , ht , (32.24)
2
and

r t  ðr  d  12ht Þ Q
zt ¼ pffiffiffiffi  i:i:d:N ð0; 1Þ: (32.25)
ht

Here d is the dividend yield, which corresponds to a dividend payment of

dt ¼ (ed  1)pt per share at time t. Then EQ[ptjIt1] ¼ erdpt1 and the expected
value at time t of one share and the dividend payment is EQ[pt + dtjIt1] ¼ erpt1, as
required in a risk-neutral framework. Note that the conditional means are different
for measures P and Q, but the functions ht(pt1, pt2,....) that define the conditional
variances for the two measures are identical.
Option prices depend on the specifications for mt and ht. We again follow Duan
(1995) and assume that

1 pffiffiffiffi
mt ¼ r  d  h t þ l ht (32.26)
2

with l representing a risk premium parameter.

pffiffiffiffi Then the conditional expectations of
rt for measures P and Q differ by l ht and

zt  zt ¼ l: (32.27)

32 Consequences for Option Pricing of a Long Memory in Volatility 917

Option prices are evaluated when the conditional variances are given by an
ARMA(N, 1) approximation to the FIEGARCH(1, d, 1) specification. From
Eqs. 32.18 and 32.27,
!
X
N  
1 j
bj L ðlnðht Þ  aÞ ¼ ð1 þ cLÞgðzt1 Þ ¼ ð1 þ cLÞg zt1  l , (32.28)
j¼1

and

gðzt Þ ¼ yzt þ gðjzt j  CÞ (32.29)

The autoregressive coefficients bj are functions of f and d, which are defined in

Appendix 1. Efficient numerical methods for simulating pt0 +n are described in
Appendix 2.

32.5 Illustrative Long Memory Option Prices

32.5.1 Inputs

Many parameters and additional inputs are required to implement the FIEGARCH
option pricing methodology. To apply that methodology to value European options,
we specify 18 numbers, a price history, and a random number generator, as follows:
• Contractual parameters – time until exercise T measured in years, the exercise
price X, and whether a call or a put option.
• The current asset price S ¼ pt0 and a set of previous prices {pt, 1  t < t 0 }.
• Trading periods per annum M, such that consecutive observed prices are sepa-
rated by 1/M years and likewise for simulated prices {pt, t 0 < t  t 0 + n} with
n ¼ MT.
• Risk-free annual interest rate R, from which the trading period rate r ¼ R/M is
obtained.
• Annual dividend yield D giving a constant trading period payout rate of d ¼ D/M;
both R and D are continuously compounded and applicable for the life of the
option contract.
• The risk premium l for investment in the asset during the life of the option,
pffiffiffiffi such
that one-period conditional expected returns are mt ¼ r  d  12ht þ l ht, for the
real-world measure P.
• Parameters m and l0 that define conditional expected pffiffiffiffi returns during the time
period of the observed prices by mt ¼ m  12ht þ l0 ht , again for measure P.
• Eight parameters that define the one-period conditional variances ht. The inte-
gration level d, the autoregressive parameter f, and the truncation level
N determine the terms bj in the AR(N) filter in Eq. 32.28. The mean a and the
moving-average parameter c complete the ARMA(N, 1) specification for ln(ht)
In Eq. 32.28. The values of the shocks to the ARMA(N, 1) process depend on g
918 S.J. Taylor

and y, which, respectively, appear in the symmetric function g(jzt j  C) and the
asymmetric function yzt whose total determines the shock
pffiffiffiffiffiffiffiterm
ffi g(zt); the con-
stant C is a parameter C0 for observed prices but is 2=p when returns are
simulated.
• K, the number of independent simulations of the terminal asset price ST ¼ pt0 +n.
• A set of Kn pseudorandom numbers distributed uniformly between 0 and 1, from
which pseudorandom standard normal variates can be obtained. These numbers
typically depend on a seed value and a deterministic algorithm.

32.5.2 Parameter Selections

Option values are tabulated for hypothetical European options on the S & P
100 index. Options are valued for 10 dates defined by the last trading days of the
10 years from 1989 to 1998 inclusive. For valuation dates from 1992 onwards, the
size of the price history is set at t0 ¼ 2,000; for previous years the price
history commences on 6 March 1984 and t0 < 2,000. It is assumed that there are
M ¼ 252 trading days in 1 year and hence exactly 21 trading days in one simulated
month. Option values are tabulated when T is 1, 2, 3, 6, 12, 18, and 24 months.
Table 32.2 lists the parameter values used to obtain the main results. The
annualized risk-free rate and dividend yield are set at 5 % and 2 %, respectively.
The risk parameter l is set at 0.028 to give2 an annual equity risk premium of 6 %.
The mean return parameter m is set to the historic mean of the complete set of S & P
100 returns from March 1984 to December 1998 and l0 is set to zero.
There are two sets of values for the conditional variance process because the
primary objective here is to compare option values when volatility is assumed to
have either a short or a long memory. The long memory parameter set takes the
integration level to be d ¼ 0.4, because this is an appropriate level based upon the
high frequency reviewed in Sect. 32.2.3. The remaining variance parameters are
then based on Table 32.1; as the moving-average parameter is small, it is set to zero
and the autoregressive parameter is adjusted to retain the unit total, d + f +
c ¼ 1. The AR filter3 is truncated at lag 1,000, although the results obtained will
nevertheless be referred to as long memory results. The short memory parameters
are similar to those for the AR(1) estimates provided in Table 32.1. The parameters
g and y are both 6 % less in Table 32.2 than in Table 32.1 to ensure that selected
moments are matched for the short and long memory specifications; the uncondi-
tional mean and variance of ln(ht) are then matched for the historic measure P,
although the unconditional means differ by approximately 0.10 for the risk-neutral
measure Q as noted in Appendix 3.

pffiffiffiffiffi
2
The conditional
pffiffiffiffi expectations of rt for measures P and Q differ by l ht and a typical average
value of ht is 0.00858. Assuming 253 trading days in 1 year gives the stated value of l.
3
The filter coefficients sum to b1 + . . . + b1,000 ¼ 0.983. After b1 ¼ 1 and b2 ¼  0.12, all of the
coefficients are near zero, with b100 ¼ 0.00017 and b1,000 ¼ 7  106.
32 Consequences for Option Pricing of a Long Memory in Volatility 919

Table 32.2 Parameter values for option price calculations

Trading periods per annum M 252
Risk-free interest rate r 0.05/M
Conditional mean
Historic intercept m 0.161/M
Historic equity risk premium term l0 0
Dividend yield d 0.02/M
Future equity risk premium term l 0.028
Conditional variance
Short memory Long memory
Integration level d 0 0.4
Truncation limit N 1,000
Mean a of EP[ln(ht)] 9.56 9.56
Autoregressive parameter f 0.982 0.6
Moving-average parameter c 0 0
Asymmetric shock parameter y 0.056 0.11
Symmetric shock parameter g 0.094 0.18
Historical value of EP[|z|] 0.737 0.737
Options are valued on the final trading day of ten consecutive years, from 1989 to 1998. The
returns history is drawn from the set of daily returns from the S & P 100 index from 6 March 1984
to 31 December 1998. A set of t0 ¼ 2,000 historical returns is used from 1992 onwards, and as
many as possible before then. The current level of the index is reset to S ¼ 100 when option values
are determined

Option prices are estimated from K ¼ 10, 000 independent simulations of prices
{pt, t 0 < t  t 0 + n} with n ¼ 504. Applying the antithetic and control variate
methods described in Appendix 2 then produces results for a long memory process
in about 12 min, when the processor speed is 2 GHz. Most of the time is spent
evaluating the high-order AR filter; the computation time is about 1 min for the
short memory process.

32.5.3 Comparisons of Implied Volatility Term Structures

The values of all options are reported using annualized implied volatilities rather
than prices. Each implied volatility (IV) is calculated from the Black-Scholes
formula, adjusted for continuous dividends. The complete set of IV outputs for
one set of inputs forms a matrix with rows labelled by the exercise prices X and
columns labelled by the times to expiry T; examples are given in Tables 32.5 and
32.6 and are discussed later.
Initially we only consider at-the-money options, for which the exercise price equals
the forward price F ¼ Se(RD)T, with IV values obtained by interpolation across two
adjacent values of X. As T varies, the IV values represent the term structure of implied
volatility. Tables 32.3 and 32.4, respectively, summarize these term structures for the
short and long memory specifications. The same information is plotted on Figs. 32.2
920 S.J. Taylor

Table 32.3 At-the-money implied volatilities for a short memory volatility process
Implied volatilities for options that expire after 1, 2, 3, 6, 12, 18, and 24 months
0 1 2 3 6 12 18 24
Year
1989 0.1211 0.1238 0.1267 0.1291 0.1340 0.1384 0.1404 0.1414
1990 0.1378 0.1380 0.1388 0.1395 0.1409 0.1421 0.1427 0.1431
1991 0.1213 0.1240 0.1271 0.1296 0.1339 0.1383 0.1403 0.1413
1992 0.1085 0.1129 0.1175 0.1210 0.1288 0.1356 0.1385 0.1401
1993 0.0945 0.1008 0.1069 0.1119 0.1226 0.1321 0.1363 0.1382
1994 0.1165 0.1201 0.1236 0.1263 0.1321 0.1375 0.1398 0.1411
1995 0.1118 0.1158 0.1200 0.1234 0.1304 0.1367 0.1391 0.1405
1996 0.1462 0.1449 0.1446 0.1445 0.1442 0.1438 0.1440 0.1441
1997 0.1883 0.1791 0.1730 0.1683 0.1599 0.1526 0.1500 0.1486
1998 0.1694 0.1640 0.1607 0.1581 0.1531 0.1488 0.1476 0.1469
Mean 0.1315 0.1323 0.1339 0.1352 0.1380 0.1406 0.1419 0.1425
St. dev. 0.0291 0.0243 0.0205 0.0175 0.0116 0.0063 0.0043 0.0032
The parameters of the EGARCH price process are listed in Table 32.2. The half-life of a volatility
shock is 1.8 months
European option prices are estimated from 10,000 simulations of the asset prices on the dates that
the options expire. The implied volatilities are for at-the-money options whose exercise prices
equal the forward rates for the expiry dates. The standard errors of the implied volatilities are
between 0.0001 and 0.0003
Options are valued on the final trading day of ten consecutive years. The returns history is drawn
from the set of daily returns from the S & P 100 index from 6 March 1984–31 December 1998.
A set of t0 ¼ 2,000 historical returns is used from 1992 onwards and as many as possible before
then. The column for T ¼ 0 provides the annualized volatilities on the valuation dates

and 32.3, respectively. The IV values for T ¼ 0 are obtained from the conditional
variances on the valuation dates. The standard errors of the tabulated implied volatil-
ities increase with T. The maximum standard errors for at-the-money options are,
respectively, 0.0003 and 0.0004 for the short and long memory specifications.
The ten IV term structures for the short memory specification commence
between 9.5 % (1993) and 18.8 % (1997) and converge towards the limiting
value of 14.3 %. The initial IV values are near the median level from 1989 to
1991, are low from 1992 to 1995, and are high from 1996 to 1998. Six of the term
structures slope upwards, two are almost flat, and two slope downwards. The shapes
of these term structures are completely determined by the initial IV values because
the volatility process is Markovian.
There are three clear differences between the term structures for the short and
long memory specifications that can be seen by comparing Figs. 32.2 and 32.3.
First, the long memory term structures can and do intersect because the volatility
process is not Markovian. Second, some of the term structures have sharp kinks
for the first month. This is particularly noteworthy for 1990 and 1996 when the
term structures are not monotonic. For 1990, the initial value of 14.1 % is
followed by 15.6 % at 1 month and a gradual rise to 16.2 % at 6 months and
a subsequent slow decline. For 1996, the term structure commences at 15.6 %,
32 Consequences for Option Pricing of a Long Memory in Volatility 921

Table 32.4 At-the-money implied volatilities for a long memory volatility process
Implied volatilities for options that expire after 1, 2, 3, 6, 12, 18, and 24 months
0 1 2 3 6 12 18 24
Year
1989 0.1194 0.1356 0.1403 0.1429 0.1467 0.1496 0.1507 0.1515
1990 0.1413 0.1556 0.1592 0.1609 0.1624 0.1624 0.1614 0.1607
1991 0.1215 0.1368 0.1416 0.1441 0.1478 0.1502 0.1510 0.1516
1992 0.1239 0.1261 0.1283 0.1301 0.1338 0.1375 0.1395 0.1409
1993 0.1080 0.1101 0.1128 0.1150 0.1195 0.1245 0.1272 0.1292
1994 0.1114 0.1189 0.1213 0.1228 0.1256 0.1284 0.1300 0.1314
1995 0.0965 0.1041 0.1067 0.1085 0.1119 0.1160 0.1187 0.1210
1996 0.1564 0.1357 0.1311 0.1295 0.1283 0.1288 0.1297 0.1308
1997 0.1697 0.1650 0.1626 0.1609 0.1574 0.1540 0.1523 0.1515
1998 0.1734 0.1693 0.1682 0.1672 0.1650 0.1623 0.1610 0.1602
Mean 0.1322 0.1357 0.1372 0.1382 0.1399 0.1414 0.1422 0.1429
St. dev. 0.0267 0.0221 0.0211 0.0204 0.0186 0.0165 0.0151 0.0141
The parameters of the FIEGARCH price process are listed in Table 32.2 and include an integration
level of d ¼ 0.4.
European option prices are estimated from 10,000 simulations of the asset prices on the dates that
the options expire. The implied volatilities are for at-the-money options whose exercise prices
equal the forward rates for the expiry dates. The standard errors of the implied volatilities are all
less than 0.0004
Options are valued on the final trading day of ten consecutive years. The returns history is drawn
from the set of daily returns from the S & P 100 index from 6 March 1984–31 December 1998.
A set of t0 ¼ 2,000 historical returns is used from 1992 onwards, and as many as possible before
then. The column for T ¼ 0 provides the annualized volatilities on the valuation dates

falls to 13.6 % after 1 month, and reaches a minimum of 12.8 % after 6 months
followed by a slow incline. The eight other term structures are monotonic and
only those for 1997 and 1998 slope downwards. Third, the term structures
approach their limiting value very slowly.4 The 2-year IVs range from 12.1 %
to 16.1 %, and it is not possible to deduce the limiting value, although
15.0–16.0 % is a plausible range.5 It is notable that the dispersion between the
ten IV values for each T decreases slowly as T increases, from 2.2 % for 1-month
options to 1.4 % for 2-year options.
There are substantial differences between the two IV values that are
calculated for each valuation date and each option lifetime. Figure 32.4
shows the differences between the at-the-money IVs for the long memory

4
The results support the conjecture that IV(T) ffi a1 + a2T2d1 for large T with a2 determined by the
history of observed returns.
5
An estimate of the constant a1 (defined in the previous footnote) is 16.0 %. An estimate of 15.0 %
follows by supposing the long memory limit is 105 % of the short memory limit, based on the limit
of ln(ht) being higher by 0.1 for the long memory process as noted in Appendix 3. The difference in
the limits is a consequence of the risk premium obtained by owning the asset; its magnitude is
mainly determined by the pronounced asymmetry in the volatility shock function g(zt).
922 S.J. Taylor

0.20
At-the-money implied volatility

0.18

0.16

0.14

0.12

0.10

0.08
0 3 6 9 12 15 18 21 24
Option life in months

Fig. 32.2 Ten volatility term structures for a short memory process

0.20
At-the-money implied volatility

0.18

0.16

0.14

0.12

0.10

0.08
0 3 6 9 12 15 18 21 24
Option life in months

Fig. 32.3 Ten volatility term structures for a long memory process with d = 0.4

specification minus the number for the short memory specification. When
T ¼ 0, these differences range from 1.9 % (1997) to 1.5 % (1992), for
3 month options from 1.5 % (1995, 1996) to 2.1 % (1990), and for 2-year
options from 1.9 % (1995) to 1.7 % (1990). The standard deviation of the ten
differences is between 1.1 % and 1.4 % for all values of T considered so it is
common for the short and long memory option prices to have IVs that differ by
more than 1 %.
32 Consequences for Option Pricing of a Long Memory in Volatility 923

0.025
ATM implied when d=0.4 minus value when d=0

0.020

0.015

0.010

0.005

0.000
0 3 6 9 12 15 18 21 24
−0.005

−0.010

−0.015

−0.020

−0.025
Option life in months

Fig. 32.4 Differences between ten pairs of term structures

32.5.4 Comparisons of Smile Effects

The columns of the IV matrix provide information about the strength of the
so-called smile effect for options prices. These effects seem to be remarkably
robust to the choice of valuation date and they are not very sensitive to the choice
between the short and long memory specifications. This can be seen by considering
the ten values of DIV ¼ IV(T,X1)  IV(T,X2) obtained for the ten valuation dates, for
various values of T, various pairs of exercise prices X1, X2, and a choice of volatility
process. First, for 1-month options with S ¼ 100, X1 ¼ 92, and X2 ¼ 108, the values
of DIV range from 3.0 % to 3.3 % for the short memory specification and from
3.7 % to 4.0 % for the long memory specification. Second, for 2-year options with
X1 ¼ 80 and X2 ¼ 120, the values of DIV range from 1.8 % to 2.0 % and from 1.8 %
to 1.9 %, respectively, for the short and long memory specifications.
Figure 32.5 shows the smiles for 3-month options valued using the short memory
model, separately for the ten valuation dates. As may be expected from the above
remarks, the ten curves are approximately parallel to each other. They are almost all
monotonic decreasing for the range of exercise prices considered, so that a U-shaped
function (from which the idea of a smile is derived) cannot be seen. The near
monotonic decline is a standard theoretical result when volatility shocks are nega-
tively correlated with price shocks (Hull 2000). It is also a stylized empirical fact for
US equity index options; see, for example, Rubinstein (1994) and Dumas et al. (1998).
Figure 32.6 shows the 3-month smiles for the long memory specification. The
shapes on Figs. 32.5 and 32.6 are similar, as all the curves are for the same expiry
time, but they are more dispersed on Fig. 32.6 because the long memory effect
induces more dispersion in at-the-money IVs. The minima of the smiles are
generally near an exercise price of 116. Figure 32.7 shows further long memory
smiles, for 2-year options when the forward price is 106.2. The parallel shapes are
924 S.J. Taylor

0.22

0.20

0.18
Implied volatility

0.16

0.14

0.12

0.10

0.08
84 88 92 96 100 104 108 112 116 120
Exercise price

Fig. 32.5 Ten smile shapes for three-month options and a short memory process

0.22

0.20

0.18
Implied volatility

0.16

0.14

0.12

0.10

0.08
84 88 92 96 100 104 108 112 116 120
Exercise price

Fig. 32.6 Ten smile shapes for three-month options and a long memory process with d = 0.4

clear; the two highest curves are almost identical, and the third, fourth, and fifth
highest curves are almost the same.
Tables 32.5 and 32.6 provide matrices of implied volatilities for options valued
on 31 December 1998. When either the call or the put option is deep out-of-the-
money, it is difficult to estimate the option price accurately because the risk-
neutral probability q(X) of the out-of-the-money option expiring in-the-money is
small. Consequently, the IV information has not been presented when the
corresponding standard errors exceed 0.002; estimates of q(X) are less than 3 %.
32 Consequences for Option Pricing of a Long Memory in Volatility 925

0.18

0.17

0.16
Implied volatility

0.15

0.14

0.13

0.12

0.11

0.10
80 84 88 92 96 100 104 108 112 116 120
Exercise price

Fig. 32.7 Ten smile shapes for two-year options and a long memory process with d = 0.4

The standard errors of the IVs are least for options that are near to at-the-money
and most of them are less than 0.0005 for the IVs listed in Tables 32.5 and 32.6.
All the sections of the smiles summarized by Tables 32.5 and 32.6 are monotonic
decreasing functions of the exercise price. The IV decreases by approximately
4–5 % for each tabulated section.

32.5.5 Sensitivity Analysis

The sensitivity of the IV matrices to three of the inputs has been assessed for
options valued on 31 December 1998. First, consider a change to the risk parameter
l which corresponds to an annual risk premium of 6 % for the tabulated results.
From Sect. 32.4.3, option prices should be lower for large T when l is reduced to
zero. Changing l to zero reduces the at-the-money IV for 2-year options from
16.0 % to 15.4 % for the long memory inputs, with a similar reduction for the short
memory inputs. Second, consider reducing the truncation level N in the AR(N) filter
from 1,000 to 100. Although this has the advantage of a substantial reduction in the
computational time, it changes the IV numbers by appreciable amounts and cannot
be recommended; for example, the 2-year at-the-money IV then changes from
16.0 % to 14.7 %.
The smile shapes on Figs. 32.5, 32.6 and 32.7 are heavily influenced by the
negative asymmetric shock parameter y, which is substantial relative to the sym-
metric shock parameter g. The asymmetry in the smile shapes can be expected to
disappear when y is zero, which is realistic for some assets including exchange
rates. Figures 32.8 and 32.9 compare smile shapes when y is changed from the
926 S.J. Taylor

Table 32.5 A matrix of implied volatilities for a short memory volatility process
Implied volatilities for options that expire after 1, 2, 3, 6, 12, 18, and 24 months
1 2 3 6 12 18 24
X
72 0.1705 0.1651
76 0.1744 0.1670 0.1624
80 0.1835 0.1696 0.1637 0.1600
84 0.1994 0.1911 0.1771 0.1655 0.1607 0.1577
88 0.1959 0.1876 0.1818 0.1711 0.1617 0.1579 0.1556
92 0.1845 0.1780 0.1734 0.1653 0.1579 0.1552 0.1535
96 0.1735 0.1693 0.1660 0.1599 0.1546 0.1526 0.1515
100 0.1644 0.1615 0.1593 0.1549 0.1514 0.1502 0.1496
104 0.1577 0.1550 0.1533 0.1502 0.1481 0.1479 0.1478
108 0.1537 0.1498 0.1480 0.1461 0.1452 0.1457 0.1461
112 0.1510 0.1461 0.1437 0.1422 0.1425 0.1436 0.1445
116 0.1525 0.1435 0.1407 0.1388 0.1400 0.1417 0.1429
120 0.1421 0.1384 0.1359 0.1376 0.1399 0.1415
126 0.1325 0.1345 0.1373 0.1394
132 0.1304 0.1317 0.1348 0.1374
138 0.1293 0.1328 0.1353
144 0.1272 0.1305 0.1336
150 0.1255 0.1286 0.1322
156 0.1269 0.1306
162 0.1246 0.1294
168 0.1220 0.1282
174 0.1277
180 0.1274
The parameters of the EGARCH price process are listed in Table 32.2. The half-life of a volatility
shock is 1.8 months
Options are valued on 31 December 1998. The returns history is the set of 2,000 daily returns from
the S & P 100 index from 4 February 1991 to 31 December 1998. European option prices are
estimated from 10,000 simulations of the asset prices on the dates that the options expire
Implied volatilities shown in Roman font have standard errors (s.e.) that are at most 0.0005 and
those shown in italic font have s.e. between 0.0005 and 0.0020; results are not shown when the
s.e. exceeds 0.0020, when options are either deep in- or out-of-the-money

values used previously to zero, with g scaled to ensure the variance of ln(ht) is
unchanged for measure P. Figure 32.8 shows that the 1-month smile shapes become
U-shaped when y is zero, while Fig. 32.9 shows that the IV are then almost constant
for 1-year options.

32.6 Conclusions

The empirical evidence for long memory in volatility is strong, for both equity and
foreign exchange markets. This evidence may more precisely be interpreted as
32 Consequences for Option Pricing of a Long Memory in Volatility 927

Table 32.6 A matrix of implied volatilities for a long memory volatility process
Implied volatilities for options that expire after 1, 2, 3, 6, 12, 18, and 24 months
1 2 3 6 12 18 24
X
72 0.1811 0.1765
76 0.1829 0.1783 0.1743
80 0.1895 0.1792 0.1754 0.1723
84 0.1970 0.1847 0.1760 0.1728 0.1703
88 0.1954 0.1889 0.1798 0.1729 0.1703 0.1683
92 0.1953 0.1860 0.1815 0.1750 0.1699 0.1680 0.1664
96 0.1814 0.1771 0.1746 0.1707 0.1672 0.1657 0.1645
100 0.1699 0.1691 0.1685 0.1666 0.1645 0.1635 0.1627
104 0.1611 0.1623 0.1631 0.1627 0.1617 0.1615 0.1610
108 0.1557 0.1569 0.1581 0.1593 0.1592 0.1596 0.1595
112 0.1525 0.1526 0.1538 0.1559 0.1570 0.1577 0.1579
116 0.1499 0.1505 0.1529 0.1549 0.1559 0.1564
120 0.1480 0.1480 0.1504 0.1528 0.1543 0.1550
126 0.1439 0.1469 0.1502 0.1520 0.1530
132 0.1440 0.1478 0.1497 0.1512
138 0.1421 0.1456 0.1477 0.1492
144 0.1436 0.1456 0.1473
150 0.1417 0.1438 0.1457
156 0.1402 0.1421 0.1444
162 0.1409 0.1428
168 0.1388 0.1416
174 0.1370 0.1405
180 0.1399
The parameters of the FIEGARCH price process are listed in Table 32.2 and include an integration
level of d ¼ 0.4
Options are valued on 31 December 1998. The returns history is the set of 2,000 daily returns from
the S & P 100 index from 4 February 1991 to 31 December 1998. European option prices are
estimated from 10,000 simulations of the asset prices on the dates that the options expire
Implied volatilities shown in Roman font have standard errors (s.e.) that are at most 0.0005, and
those shown in italic font have s.e. between 0.0005 and 0.0020; results are not shown when the
s.e. exceeds 0.0020, when options are either deep in- or out-of-the-money

evidence for long memory effects, because there are short memory processes that
have similar autocorrelations and spectral densities, except at very low frequencies.
The theory of option pricing when volatility follows a discrete-time ARCH
process relies on weak assumptions about the continuous-time process followed
by prices and the numerical implementation of the theory is straightforward.
Application of the theory when the volatility process is fractionally integrated
does, however, require pragmatic approximations because the fundamental filter
(1  L)d is an infinite order polynomial that must be truncated at some power N.
Option prices are sensitive to the truncation point N, so that large values and long
price histories from an assumed stationary process are required.
928 S.J. Taylor

0.22

0.20

0.18
Implied volatility

0.16

0.14

0.12 d=0.4, theta<0

d=0, theta<0
0.10 d=0.4, theta=0
d=0, theta=0
0.08
84 88 92 96 100 104 108 112 116 120
Exercise price

Fig. 32.8 Impact of asymmetric volatility shocks on one-month options

0.20
d=0.4, theta<0
0.19
d=0, theta<0
d=0.4, theta=0
0.18
d=0, theta=0
Implied volatility

0.17

0.16

0.15

0.14

0.13

0.12
84 88 92 96 100 104 108 112 116 120
Exercise price

Fig. 32.9 Impact of asymmetric volatility shocks on one-year options

The term structure of implied volatility for at-the-money options can

be notably different for short and long memory ARCH specifications
applied to the same price history. Long memory term structures have more
variety in their shapes. They may have kinks for short maturity options and
32 Consequences for Option Pricing of a Long Memory in Volatility 929

they may not have a monotonic shape. Also, term structures calculated on
different valuation dates sometimes intersect each other. None of these
possibilities occurs for a Markovian short memory specification. Long
memory term structures do not converge rapidly to a limit as the
lifetime of options increases. It is difficult to estimate the limit for the typical
value d ¼ 0.4.
Implied volatilities as functions of exercise prices have similar shapes for short
and long memory specifications. The differences in these shapes are minor in
comparison to the differences in the term structure shapes.
It is common for the short and long memory implied volatilities to differ by
more than 1 % for options on the S & P 100 index, regardless of the option
lifetime and the exercise price; if the short memory implied is at a typical level
of 14 %, then the long memory implied is often below 13 % or above 15 %.
Consequently, the economic consequences of a long memory assumption are
important.

Acknowledgments The support of Inquire Europe and the helpful advice of Bernard Dumas, Ton
Vorst, and Granville Tunnicliffe-Wilson are gratefully acknowledged.

Appendix 1: Series Expansions and Truncation Approximations

Series expansions in the lag operator L are required to evaluate the following
conditional variance:

lnðht Þ ¼ a þ ð1  fLÞ1 ð1  LÞd ð1 þ cLÞgðzt1 Þ: (32.30)

We note the results:

X
1
jd1
ð1  LÞd ¼ 1  aj Lj , a1 ¼ d, aj ¼ aj1 , j  2, (32.31)
j¼1
j

X
1
ð1  fLÞð1  LÞd ¼ 1  bj Lj , b1 ¼ d þ f, bj ¼ aj  faj1 , j  2, (32.32)
j¼1

X
1
ð1  fLÞð1  LÞd ð1 þ cLÞ1 ¼ 1  fj Lj ,
j¼1

X
j1
f1 ¼ d þ f þ c, fj ¼ bj  ðcÞj þ ðcÞjk bk , j  2: (32.33)
k¼1
930 S.J. Taylor

The autoregressive weights in Eqs. 32.33 can be denoted as fj (d,f,c). Also,

X
1
ð1  fLÞ1 ð1  LÞd ð1 þ cLÞ ¼ 1 þ cj Lj , (32.34)
j¼1

c1 ¼ d þ f þ c, cj ¼ fj ðd,  c,  fÞ (32.35)

It is necessary to truncate the infinite summations when evaluating empirical

conditional variances. Truncation after N terms of the summations in Eqs. 32.35,
32.33, and 32.32, respectively, gives the MA(N), AR(N), and ARMA(N, 1)
approximations:
X
N  
lnðht Þ ¼ a þ gðzt1 Þ þ cj g ztj1 , (32.36)
j¼1

X
N   
lnðht Þ ¼ a þ fj ln htj  a þ gðzt1 Þ, (32.37)
j¼1

X
N   
lnðht Þ ¼ a þ bj ln htj  a þ gðzt1 Þ þ cgðzt2 Þ: (32.38)
j¼1

As j ! 1, the coefficients bj and fj converge much more rapidly to zero than the
coefficients cj. Consequently it is best to use either the AR or the ARMA
approximation.

Appendix 2: Simulation Methods

Suppose there are returns observed at times 1  t  t0 , whose distributions are given
by measure P, and that we then want to simulate returns for times t > t0 using
measure Q. Then ln (ht) is calculated for 1  t  t0 + 1 using the observed returns,
with ln (ht) ¼ a and g(zt) ¼ 0 for t < 1, followed by simulating z*t  Q N(0,1) and
hence obtaining rt and ln (ht+1) for t > t0 .
Care is required when calculating the constant C in Eq. 32.29 because the
observed conditional distributions of the terms zt are not normal while the simula-
tions assume that they are. Consequently, we define

C ¼ C0 , t  t0 ,
pffiffiffiffiffiffiffiffi (32.39)
¼ 2=p, t > t0 ,

for a constant C0 estimated from observed returns. An alternative method, described

by Bollerslev and Mikkelsen (1999), is to simulate from the sample distribution of
standardized observed returns.
Standard variance reduction techniques substantially enhance the accuracy of
Monte Carlo estimates of contingent claim prices. Our antithetic method uses one
32 Consequences for Option Pricing of a Long Memory in Volatility 931

i.i.d. N(0, 1) sequence {z*t } to define the further i.i.d. N(0, 1) sequences, { zt*},
{z•t }, and { •
 zt }, with
•
 the terms zt chosen so that there
 isnegative
 correlation
 
   
between zt and zt ; this is achieved by defining F z t þ F zt ¼ 1 þ 12sign zt .
The four sequences provide claim prices whose average, y say, is much less variable
than the claim price from a single sequence. An overall average y^ is then obtained
from a set of K values fyk , 1  k  K g.
The control variate method makes use of an unbiased estimate y^CV of a known
parameter yCV, such that y^ is positively correlated with y^CV. A suitable parameter,
when pricing a call option in an ARCH framework, is the price of a call option when
volatility is deterministic. The deterministic volatility process is defined by
replacing all terms ln(ht), t > t 0 + 1, by their expectations under P conditional on
the history It0 . Then yCV is given by the obvious modification of the Black-Scholes
formula, while y^CV is obtained by using the same 4K sequences of i.i.d. variables
that define y^. Finally, a more accurate estimate of the option price is then given by
y~ ¼ y^  b(^ y CV  yCV) with b chosen to minimize the variance of y~.

Appendix 3: Impact of a Volatility Risk Premium

On average the term structure of implied volatilities will slope upwards for the
FIEGARCH option pricing model. This occurs because the expectation of ln(ht)
depends on the measure6 when l 6¼ 0. The unconditional expectation equals a for
measure P. It is different for measure Q because

      pffiffiffiffiffiffiffiffi l2 g
EQ g zt  l ¼ ly þ g EQ zt  l  2=p ffi ly þ pffiffiffiffiffiffi (32.40)
2p

when l is small, and this expectation is in general not zero.7 For a fixed t0 , as t ! 1,
!1
X
N   
E ½lnðht Þ jI t0
! a þ
Q
1 bj ð1 þ cÞEQ g zt  l (32.41)
j¼1

The difference between the P and Q expectations of ln(ht) could be interpreted as

a volatility risk premium. This premium is typically negative, because typically
l > 0, y  0 and g > 0. Furthermore, when y is negative, the dominant term in
Eq. 32.40 is  ly, because l is always small, and then the premium reflects the
degree of asymmetry in the volatility shocks g(zt).

6
The dependence of moments of ht on the measure is shown by Duan (1995, p. 19) for the GARCH
(1,1) model.
pffiffiffiffiffiffiffiffi  
7
When z  N(0,1), E½jz  lj
¼ 2=pexp 12l2 þ lð2FðlÞ  1Þ with F the cumulative distri-
bution function of z.
932 S.J. Taylor

The magnitude of the volatility risk premium can be important and, indeed, the
quantity defined by the limit in Eq. 32.41 becomes infinite8 as N ! 1 when d is
positive. A plausible value of l for the S & P 100 index is 0.028, obtained by assuming
that the equity risk premium is 6 % per annum. For the short memory parameter values
in the first row of Table 32.1, when d ¼ 0 and N ¼ 1,000, the limit of EQ[ln(ht) |It0 ]  a
equals 0.10. This limit increases to 0.20 for the parameter values in the final row of
Table 32.1, when d ¼ 0.4 andpNffiffiffiffi¼ 1,000. The typical effect of adding 0.2 to ln(ht) is to
multiply standard deviations ht by 1.1 so that far-horizon expected volatilities, under
Q, are slightly higher than might be expected from historical standard deviations.

References
Andersen, T. G., & Bollerslev, T. (1997). Heterogeneous information arrivals and return volatility
dynamics: Uncovering the long-run in high frequency returns. Journal of Finance, 52, 975–1005.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Ebens, H. (2001a). The distribution of stock
return volatility. Journal of Financial Economics, 61, 43–76.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2001b). The distribution of realized
exchange rate volatility. Journal of the American Statistical Association, 96, 42–55.
Areal, N. M. P. C., & Taylor, S. J. (2002). The realized volatility of FTSE-100 futures prices.
Journal of Futures Markets, 22, 627–648.
Baillie, R. T. (1996). Long memory processes and fractional integration in econometrics. Journal
of Econometrics, 73, 5–59.
Bakshi, G., Cao, C., & Chen, Z. (1997). Empirical performance of alternative option pricing
models. Journal of Finance, 52, 2003–2049.
Barndorff-Nielsen, O. E., & Shephard, N. (2001). Non-Gaussian Ornstein-Uhlenbeck based
models and some of their uses in financial economics. Journal of the Royal Statistical
Society B, 63, 167–241.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroscedasticity. Journal of
Econometrics, 31, 307–327.
Bollerslev, T., & Mikkelsen, H. O. (1996). Modeling and pricing long memory in stock market
volatility. Journal of Econometrics, 73, 151–184.
Bollerslev, T., & Mikkelsen, H. O. (1999). Long-term equity anticipation securities and stock
market volatility dynamics. Journal of Econometrics, 92, 75–99.
Bollerslev, T., & Wooldridge, J. M. (1992). Quasi-maximum likelihood estimation and inference
in dynamic models with time varying covariances. Econometric Reviews, 11, 143–172.
Bollerslev, T., & Wright, J. H. (2000). Semiparametric estimation of long-memory volatility
dependencies: The role of high-frequency data. Journal of Econometrics, 98, 81–106.
Breidt, F. J., Crato, N., & de Lima, P. (1998). The detection and estimation of long memory in
stochastic volatility. Journal of Econometrics, 83, 325–348.
Brockwell, P. J., & Davis, R. A. (1991). Time series: Theory and methods (2nd ed.). New York:
Springer.
Christoffersen, P., Jacobs, K., Ornthanalai, C., & Wang, Y. (2008). Option valuation with long-run
and short-run volatility components. Journal of Financial Economics, 90, 272–297.
Christoffersen, P., Dorion, C., Jacobs, K., & Wang, Y. (2010). Volatility components: Affine restric-
tions and non-normal innovations. Journal of Business and Economic Statistics, 28, 483–502.
Comte, F., & Renault, E. (1998). Long memory in continuous-time stochastic volatility models.
Mathematical Finance, 8, 291–323.

X
1 X1
As (1  L)d1 ¼ 0 for d > 0, it follows from Eqs. 32.31 and 32.32 that
8
aj ¼ bj ¼ 1:
j¼1 j¼1
32 Consequences for Option Pricing of a Long Memory in Volatility 933

Dacorogna, M. M., Muller, U. A., Nagler, R. J., Olsen, R. B., & Pictet, O. V. (1993).
A geographical model for the daily and weekly seasonal volatility in the foreign exchange
market. Journal of International Money and Finance, 12, 413–438.
Ding, Z., Granger, C. W. J., & Engle, R. F. (1993). A long memory property of stock market
returns and a new model. Journal of Empirical Finance, 1, 83–106.
Duan, J.-C. (1995). The GARCH option pricing model. Mathematical Finance, 5, 13–32.
Duan, J.-C. (1997). Augmented GARCH(p, q) process and its diffusion limit. Journal of Econo-
metrics, 79, 97–127.
Duffie, D., Pan, J., & Singleton, K. J. (2000). Transform analysis and asset pricing for affine jump-
diffusions. Econometrica, 68, 1343–1376.
Dumas, B., Fleming, J., & Whaley, R. E. (1998). Implied volatility functions: Empirical tests.
Journal of Finance, 53, 2059–2106.
Ebens, H. (1999). Realized stock volatility. Working paper, John Hopkins University.
Ederington, L. H., & Lee, J. H. (1993). How markets process information: News releases and
volatility. Journal of Finance, 49, 1161–1191.
Gallant, A. R., Hsu, C. T., & Tauchen, G. E. (1999). Using daily range data to calibrate volatility
diffusions and extract the forward integrated variance. Review of Economics and Statistics, 81,
617–631.
Granger, C. W. J. (1980). Long memory relationships and the aggregation of dynamic models.
Journal of Econometrics, 14, 227–238.
Granger, C. W. J., & Ding, Z. (1996). Varieties of long memory models. Journal of Econometrics,
73, 61–77.
Granger, C. W. J., & Joyeux, R. (1980). An introduction to long-memory time series models and
fractional differencing. Journal of Time Series Analysis, 1, 15–29.
Harvey, A. C. (1998). Long-memory in stochastic volatility. In J. Knight & S. Satchell (Eds.),
Forecasting volatility in financial markets. London: Butterworth-Heinemann.
Heston, S. (1993). A closed-form solution for options with stochastic volatility with applications to
bond and currency options. Review of Financial Studies, 6, 327–343.
Hosking, J. (1981). Fractional differencing. Biometrika, 68, 165–176.
Hull, J. (2000). Options, futures and other derivative securities (4th ed.). New Jersey: Prentice-Hall.
Kallsen, J., & Taqqu, M. (1998). Option pricing in ARCH-type models. Mathematical Finance, 8,
13–26.
Nelson, D. B. (1990). ARCH models as diffusion approximations. Journal of Econometrics, 45, 7–38.
Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.
Econometrica, 59, 347–370.
Ohanissian, A., Russell, J. R., Tsay, R. S. (2004). True or spurious long memory in volatility: Does
it matter for pricing options? Working paper, University of Chicago.
Ohanissian, A., Russell, J. R., & Tsay, R. S. (2008). True or spurious long memory? A new test.
Journal of Business and Economic Statistics, 26, 161–175.
Pong, S., Shackleton, M. B., & Taylor, S. J. (2008). Distinguishing short and long memory
volatility specifications. Econometrics Journal, 11, 617–637.
Ritchken, P., & Trevor, R. (1999). Pricing options under generalized GARCH and stochastic
volatility processes. Journal of Finance, 54, 377–402.
Rubinstein, M. (1994). Implied binomial trees. Journal of Finance, 49, 771–818.
Shephard, N. (1996). Statistical aspects of ARCH and stochastic volatility. In D. R. Cox, D. V.
Hinkley, & O. E. Barndorff-Nielsen (Eds.), Likelihood, time series with econometric and other
applications (pp. 1–67). London: Chapman and Hall.
Taylor, S. J. (1994). Modeling stochastic volatility: A review and comparative study. Mathemat-
ical Finance, 4, 183–204.
Taylor, S. J. (2005). Asset price dynamics, volatility, and prediction. Princeton: Princeton Uni-
versity Press.
Taylor, S. J. (2008). Modelling financial time series (2nd ed.). Singapore: World Scientific
Publishing.
Seasonal Aspects of Australian Electricity
Market 33
Vikash Ramiah, Stuart Thomas, Richard Heaney, and
Heather Mitchell

Contents
33.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 936
33.2 Australian Electricity Spot Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 938
33.3 Australian Demand for Electricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
33.3.1 Demand Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
33.3.2 Demand Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
33.3.3 Modelling Seasonality in Demand for Electricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 952
33.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 952
33.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 955
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 956

Abstract
Australian electricity spot prices differ considerably from equity spot prices in
that they contain an extremely rapid mean-reversion process. The electricity spot
price could increase to a market cap price of AU$12,500 per megawatt hour
(MWh) and revert back to a mean level (AUD$30) within a half-hour interval.
This has implications for derivative pricing and risk management. For example,
while the Black and Scholes option pricing model works reasonably well for
equity market-based securities, it performs poorly for commodities like electric-
ity. Understanding the dynamics of electricity spot prices and demand is also

V. Ramiah (*)
School of Economics, Finance and Marketing, RMIT University, Melbourne, Australia
e-mail: vikash.ramiah@rmit.edu.au
S. Thomas • H. Mitchell
RMIT University, Melbourne, VIC, Australia
e-mail: stuart.thomas@rmit.edu.au; heather.mitchell@rmit.edu.au
R. Heaney
Accounting and Finance, The University of Western Australia, Perth, Australia
e-mail: richard.heaney@uwa.edu.au

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 935
DOI 10.1007/978-1-4614-7750-1_33,
# Springer Science+Business Media New York 2015
936 V. Ramiah et al.

important in order to correctly forecast electricity prices. We develop econo-

metric models for seasonal patterns in both price returns and proportional
changes in demand for electricity. We also model extreme spikes in the data.
Our study identifies both seasonality effects and dramatic price reversals in
the Australian electricity market. The pricing seasonality effects include
time-of-day, day-of-week, monthly, and yearly effects. There is also evidence
of seasonality in demand for electricity.

Keywords
Electricity • Spot price • Seasonality • Outlier • Demand • Econometric modelling

33.1 Introduction

On Monday 11th of January 2010, the temperature in the state of Victoria in

Australia peaked to 42.8
C. This led to an increase in demand for electricity as
households turned on their air conditioning systems. During that half-hour interval,
the wholesale price of electricity surged to over $9,000 and then reverted to the
average price of around $30 in the subsequent half-hour period. Exercising a call
option around that time would have been extremely profitable. What causes the spot
price of electricity to exhibit such erratic behavior? We provide a partial explana-
tion of this behavior in this study.
Unpredictable jumps in electricity spot price are usually referred to as “spikes”
because jumps tend to happen very quickly, with equally quick reversion to mean
levels. These events create a challenge for quantitative analysts when it comes to
pricing electricity derivatives. Before one can price these derivatives, one has to be
able to effectively forecast the spot price distribution. To date, this problem remains
an unresolved matter. For instance, the Black and Scholes (1973) option pricing
model fails to adequately price an option written on the electricity spot price
because of the incidence of these pricing spikes. The Black and Scholes model
works reasonably well in the equity markets whereby the equity time series takes
a longer time to trend upwards or downwards. Given the rapid price reversals
evident in the electricity spot prices, the Black and Scholes model does not perform
particularly well. The key characteristic of non-storability of electricity partially
explains why electricity spot price increases significantly during a short-term
interval. As the possibility of storing large amounts of electricity does not exist,
this product has to be consumed and produced simultaneously. Large price swings
occur in this market mainly because supply is inelastic.
The electricity pricing literature provides clear evidence of the spikes. Thomas
et al. (2011), Huisman and Huurman (2003), and Goto and Karolyi (2004) show
non-normality manifested as positive skewness and extreme leptokurtosis; Johnson
and Barz (1999) observe mean reversion in the long run; Kaminski (1997),
Clewlow and Strickland (2000), and Eichler et al. (2012) find evidence of extreme
behavior with fast-reverting spikes; and Bunn and Karakatsani (2003) demonstrate
excessive volatility. Further, Escribano et al. (2002) show that electricity volatility
33 Seasonal Aspects of Australian Electricity Market 937

can be time varying in a number of countries like in Argentina, New Zealand,

Norway, Sweden, and Spain.1 Other researchers devote their time to modelling the
electricity time series itself. For instance, Knittel and Roberts (2001) apply mean-
reversion, time-varying mean, jump-diffusion, time-dependent jump intensity,
ARMAX, and exponential generalized autoregressive conditional heteroske-
dasticity (EGARCH) models to hourly prices in the Californian electricity market
and conclude that forecasting performance is relatively poor. Kaminski (1997)
addresses the spiky characteristic with a random walk jump-diffusion model, but
this model ignores the persistent mean reversion, which is a feature of electricity
prices, and this avenue is explored further in Clewlow and Strickland (2000). One
of the limitations of the jump-diffusion approaches is the assumption that all shocks
affecting the price series die out at the same rate. Escribano et al. (2002) identify
two additional price components, namely, volatility clustering in the form of
GARCH effects and seasonality (emphasized by Lucia and Schwartz 2002), both
in the deterministic component of prices and the jump intensity.
The seasonal aspects recognized within the electricity market are also important.
The results observed in one geographic region cannot be applied to other areas as
the climatic conditions are different. For example, summer time occurs in different
months around the globe, and for that reason, it is important to investigate the
seasonal aspects within the Australian electricity market.
Seasonality in the electricity market is not limited to price series as seasonal
patterns in demand or system load are well documented in the literature. Harvey
and Koopman (1993) document intra-daily and intra-week effects and incorporate
them into their demand model using splines. Other early studies considered longer-
term load forecasting horizons several months into the future, using daily, weekly,
or monthly demand data (e.g., Engle et al. 1989). Pardo et al. (2002) employ daily
data in a study of Spanish electricity demand and emphasize the importance of daily
and monthly seasonal structures. More recent studies consider modelling and
forecasting demand over shorter periods using intraday data. In the Australian
context, Smith (2000) documents intraday patterns in demand in the New South
Wales electricity market and incorporates diurnal variation into a Bayesian
semiparametric regression framework to model intraday electricity load data and
obtain short-term load forecasts. We extend his analysis by testing whether sea-
sonality exists in other Australian regions.
In the early 1990s, following the release of the Hilmer Report, the Australian
electricity industry embarked on a progressive program of deregulation. The Hilmer
reforms led to the disaggregation of vertically integrated, government-owned elec-
tricity authorities into separate generation transmission and distribution and retail
sales sectors in each state. This gave rise to a wholesale market for electricity
which is currently managed by the Australian Energy Market Operator (AEMO).
AEMO is responsible for five regions, namely, VIC1 (Victoria), NSW1 (New
South Wales), QLD1 (Queensland), SA1 (South Australia), and TAS1 (Tasmania).

1
See Montero et al. (2011) for further evidence in the Spanish market.
938 V. Ramiah et al.

Western Australia and the Northern Territory are geographically remote and to date
have not been integrated into the “national” market. Physical transmission of power
between regions is achieved via interconnectors that physically link the five different
states. The spot electricity market in the AEMO is where all market generators and
market customers settle their electricity sales and purchases based on a spot price.
The market participants are generally large consumers of electricity, generators, and
speculators. Large consumers of electricity like car manufacturers and supermarkets
enter this market to enjoy a lower cost of production and to hedge their positions.
Speculators, on the other hand, trade future contracts, forwards, options, caps, floors,
and swaps for profit motives.

33.2 Australian Electricity Spot Price

The half-hourly pool price data is sourced directly from AEMO (previously known
as National Electricity Market Management Company, NEMMCO) for the period
from 1 January 1999 to 31 January 2006. Prices are expressed in Australian dollars
per megawatt hour ($/MWh). The sample size is 124,224 observations for each
region except for TAS1 that joined the pool at a later stage. Figure 33.1 shows the
data series for NSW1 with evidence of considerable spiked behavior. Descriptive
statistics for the price and return series which are shown in Table 33.1 tend to
confirm these outliers.

10000
NSW

8000

6000

4000

2000

−2000
25000 50000 75000 100000 125000

Fig. 33.1 Electricity price series for NSW1 ($/MWh) for the period 01.01.1999–31.01.2006
 33

Table 33.1 The descriptive statistics of spot prices and returns for the different Australian regions
Range Minimum Maximum Mean Std. deviation Skewness Kurtosis JB stats
Spot prices
NSW1 9,912.13 3.10 9,909.03 34.02 178.11 35.37 1468.54 89701
QLD1 9,098.74 156.14 8,942.60 36.65 158.45 28.56 1058.17 46527
SA1 9,822.43 822.45 8,999.98 41.91 155.18 25.40 800.35 26598
VIC1 7,746.07 329.91 7,416.16 30.19 103.26 35.38 1502.93 93950
TAS1 10,332.00 332.00 10,000.00 87.30 367.26 21.46 505.16 10584
Seasonal Aspects of Australian Electricity Market

Returns
RNSW1 453.00 222.00 231.00 0.0301 1.36 41.46 15,480.38 9981513
RQLD1 380.49 11.25 369.24 0.0611 1.74 111.49 19,122.19 15233050
RSA1 423.74 32.94 390.80 0.0603 1.68 123.00 24,820.63 25665638
RVIC1 351.93 209.50 142.43 0.0262 1.05 15.41 17,959.48 13434834
RTAS1 2416.00 83.50 2,332.50 0.6681 31.42 59.14 3,754.10 586865
939
940 V. Ramiah et al.

We report the range, minimum, maximum, mean, standard deviation, skewness,

kurtosis, and Jarque-Bera statistics (JB stats) for each region’s price and return
series. Mean prices vary between regions from $30.19 for VIC1 to $41.91 for SA1
and to $87.30 for TAS1 as the factor inputs (low-cost coal, high-cost gas and water)
vary from one state to another. TAS1 was at the early stage of joining the pool, and
these statistics depict and are consistent with the immature stage of the market in
this region. The standard deviation of prices is generally high, is widely dispersed
across the regions, and is broadly consistent with the pattern of means, ranging from
$103.26 for VIC1 to $367.26 for TAS1. The lowest maximum price of $7416.16 is
observed in VIC1 and the highest maximum price in TAS1 at $10,000. No prices
above $10,000 were observed as this was the market price cap during the period of
the study. This amount is smaller than other international markets like the state of
California in the United States, which recorded over US$50,000 during a period
when no market cap was in force. These extremely high values are the features that
we are looking for and are referred to as spikes in this study. These instances are
generally viewed as lucrative occurrences by arbitrageurs. The most commonly
traded call options on the Australian electricity market have exercise prices of
either $100 or $300. If we assume that market participants will exercise their rights
when the option is in the money, then VIC1 traders will exercise when the spot price
is above $340 (mean price of around $30 plus three standard deviations away) on
average. To ensure that we capture a spike, we adopt a conservative approach
whereby we define a spike as the mean plus four standard deviations. We observe
a total of 505 spikes in NSW1, QLD1, SA1, and VIC1. QLD1 exhibits the greatest
incidence of extreme price spikes by state with 173 occurrences (34 %), followed
by SA1 with 159 (31 %) and VIC1 with 96 (19 %). By day of the week, Monday
shows the highest incidence with 115 (23 %) tapering gradually to Sunday with
45 occurrences (9 %). June shows the highest incidence by month with 81 (16 %).
The highest incidences by year occur in 2002 with 147 spikes (29 %) and 2000 with
116 spikes (23 %), both markedly higher than any other full year in the study
period. It should be noted that the incidence of extreme price spikes appears to be
declining from 2003 onwards as legislation on collusive behavior was passed in
Queensland. There is evidence of a concentration of spikes occurring between the
hours 06:30 and approximately 10:00 and between 15:30 and 19:00 h, with a marked
increase in frequency concentrated at the 18:00 trading interval.
Another unique characteristic of this dataset is negative price. All five regions
exhibit negative minimum prices. As shown in Table 33.1, the negative prices
for NSW1, QLD1, SA1, VIC1, and TAS1 are $3.10, $156.14, $822.45,
$329.91, and $332.00, respectively. To understand the possibility of such rare
and short-lived occurrence, it is important to understand how the spot price is
derived. It is a derived price per trading interval, calculated by a two-step procedure
based on the offers to supply made by generators in the pool. The trading day is
divided into 48 half-hour “trading intervals” which is then subdivided into 5-min.
“dispatch intervals.” A “dispatch price” is recorded as the marginal price of supply
to meet demand for each 5-min. interval in a given half-hour period. This marginal
price is typically the dispatch offer price of the last generator brought into
33 Seasonal Aspects of Australian Electricity Market 941

production to meet demand at that interval. The spot price is then calculated as an
arithmetic average of the six dispatch prices in a half hour. All generators who are
called into production during a given half-hour trading interval receive this spot
price for the quantity of electricity delivered during the trading interval. A generator
may bid a negative price into the pool for its self-dispatch quantity as a tactical
move to ensure that they are among the first to be called in to generate as closing
down their operations due to not being called into production may be more costly
than producing at negative prices. The implication of a negative price is that the
standard return calculation for prices may contain this bias, and to that end, the
following return formula is used:

ðPt  Pt1 Þ
RET t ¼ : (33.1)
jPt1 j

where RETt represents the half-hourly discrete proportionate change in price

(“return”) at time t, Pt is half-hourly price at time t, and |Pt1| is the absolute
value of the previous half-hourly price, i.e., at time t1. The denominator is
specified as the absolute value to allow for the presence of negative prices. We
prefer a discrete return specification over log returns because the spot market in the
AEMO trades at discrete half-hourly intervals – it is not a continuous market in the
way of most conventional financial markets. Further, a log return specification will
dampen the extreme spike effects we are attempting to capture and is incompatible
with negative prices. Descriptive statistics for the half-hourly return series are
shown in Table 33.1. Mean half-hourly returns vary widely between regions,
from 2.6 % for VIC1 to 6.1 % for QLD1. High maximum returns are observed
and are consistent with the spike behavior discussed above.
One of the clear patterns discussed so far is the spike behavior, and, based on the
earlier discussion, it is fair to say that these spikes are more likely to occur at some
specific periods like 18.00 h, Mondays, and in the year 2002. We develop a method
to test if there is seasonality in the return series with half-hourly return as the
dependent variable. The explanatory variables consist of seasonal dummy
variables, spike behavior, and negative prices. We add lagged returns to the list
of independent variables to control for serial correlation in the return series. The
model employed in this study is presented as Eq. 33.2:

X
5 X6 X12 X
2006
RET R, t ¼ a0 þ b1, i RET R, t1 þ b2, j DAY j þ b3, k MTH k þ b4, l YRl
i¼1 j1 k¼1, 6¼9 l¼1999, 6¼2001

X
48 X
N R, S
X
N R, N
þ b5, m HH m þ b6, o SPIKER, o þ b7, p NEGR, p þ et
m¼1, 6¼24 o¼1 p¼1

(33.2)
where RETR,t represents the discrete return for region R at time t, a0 represents the
constant term, and DAYj represents the dummy variable for each day of the week
942 V. Ramiah et al.

(j ¼ 1 for Monday, 2 for Tuesday, . . . , 6 for Saturday). The dummy variable for
Sunday is dropped on the basis that the returns are relatively lower on that day. We
use similar justification when it comes to discarding other periods. MTHk represents
the dummy variable for each month (k ¼ 1 for January, 2 for February, . . . , 12 for
December). The dummy variable for September is dropped; YRl represents the
dummy variable for each year included in the sample period (l ¼ 1999, . . . , 2006).
The dummy variable for 2001 is dropped; HHm represents the dummy variable for
each half-hourly trading interval (m ¼ 1 for 00:00 h, 2 for 00:30 h, . . . , 48 for
23:30 h). The dummy variable for 11:30 h is dropped; SPIKER,o represents a set of
NR,S dummy variables, one for each extreme spike as previously defined, with NR,S
representing the number of extreme returns observed in region R for the period of the
study; NEGR,N represents the dummy variable for the return associated with an
occurrence of a negative price (p ¼ 1, . . . , NR,N), with NR,N representing the number
of occurrences of a negative price for region R during for the period of the study.
The values for the trading interval at 11:30 h, Sunday, September, and the year
2001 were dropped to avoid exact collinearity and to allow comparison of these
values with the remaining seasonal coefficients for HHm, DAYj, MTHk, and YRl. The
equation was initially estimated for each region with 20 lagged returns (RETRt1, . . . ,
RETRt20). F-tests for redundant variables were performed for all regions, and AIC
and SBC values support the finding that lags 1 through 5 were significant. Lags
6 onwards were not found to be significant and were discarded. Standard tests and
residual diagnostics revealed no misspecification in the above model.
Results of the regression analysis are presented in Table 33.2. Coefficients and
t-statistics (t-stats) are presented for each seasonal dummy variable. The results
generated from this model are consistent with Kaminski (1997), Clewlow and
Strickland (2000) and De Jong and Huismann (2002) with regard to the seasonal
aspects and spike behavior of the price series. Seasonal effects vary between
regions, and time-of-day effects are generally more significant than other
seasonalities. Positive returns are observed at times of peak population activity in
the morning and early evening, and negative returns observed at most other times.
In general, significant negative returns are found for the small hours of the morning
between 12:30 a.m. and approximately 4:00 a.m. in all regions. NSW1 exhibits an
unexplained positive return at 1:30 a.m., reverting to negative returns for the
remainder of the early morning. The hours between 5:00 a.m. and 9:30 a.m.
inclusively exhibit significant positive effect in all regions, reverting to generally
negative returns in the late morning. Negative returns are found in SA1 during
midafternoon. Significant positive effects are observed for all regions in the early
evening, generally between the hours of 5:00 and 7:00 p.m., reverting to significant
negative effects for the remainder of the evening, until positive effects emerge in
the late evening at 10:30 p.m. and at midnight. The periods of positive return in the
morning and early evening are consistent with peaks in activity in the population.
The positive returns observed around 11:00 p.m. are consistent with increased
demand for electricity arising from off-peak hot water systems generally switching
on at 11:00 p.m.
33 Seasonal Aspects of Australian Electricity Market 943

Table 33.2 The seasonality effects in the different regions

SA1 VIC1 NSW1 QLD1
Variable Coeff t-stats Coeff t-stats Coeff t-stats Coeff t-stats
Mon 0.0113 3.36 0.0279 4.00 0.0218 3.13 0.0154 4.05
Tues 0.0105 3.14 0.0249 3.57 0.0190 2.72 0.0137 3.61
Wed 0.0120 3.56 0.0246 3.53 0.0178 2.56 0.0118 3.09
Thurs 0.0109 3.25 0.0232 3.33 0.0213 3.05 0.0137 3.59
Fri 0.0090 2.67 0.0233 3.35 0.0176 2.52 0.0121 3.18
Sat 0.0018 0.53 0.0137 1.97 0.0092 1.33 0.0010 0.26
Jan 0.0003 0.06 0.0022 0.25 0.0015 0.16 0.0172 3.42
Feb 0.0063 1.40 0.0043 0.46 0.0017 0.18 0.0102 1.99
Mar 0.0068 1.54 0.0008 0.09 0.0015 0.16 0.0220 4.39
Apr 0.0019 0.42 0.0021 0.23 0.0012 0.13 0.0029 0.57
May 0.0039 0.87 0.0051 0.56 0.0048 0.52 0.0116 2.31
June 0.0018 0.41 0.0047 0.51 0.0064 0.70 0.0125 2.47
July 0.0089 2.01 0.0023 0.25 0.0043 0.47 0.0134 2.68
Aug 0.0074 1.67 0.0041 0.45 0.0035 0.38 0.0124 2.48
Oct 0.0016 0.37 0.0289 3.16 0.0235 2.56 0.0062 1.24
Nov 0.0048 1.07 0.0059 0.64 0.0025 0.27 0.0020 0.40
Dec 0.0041 0.93 0.0005 0.06 0.0007 0.07 0.0110 2.20
Y1999 0.0406 12.02 0.0128 1.83 0.0072 1.03 0.0227 5.91
Y2000 0.0234 6.93 0.0043 0.61 0.0136 1.95 0.0361 9.43
Y2002 0.0043 1.29 0.0008 0.11 0.0085 1.22 0.0131 3.43
Y2003 0.0001 0.04 0.0036 0.51 0.0002 0.02 0.0087 2.28
Y2004 0.0012 0.37 0.0010 0.14 0.0041 0.58 0.0061 1.58
Y2005 0.0025 0.75 0.0037 0.54 0.0029 0.41 0.0082 2.14
Y2006 0.0210 2.32 0.0031 0.16 0.0018 0.10 0.0080 0.78
H0000 0.0133 1.51 0.1022 5.61 0.1002 5.49 0.1598 16.03
H0030 0.0383 4.36 0.0783 4.29 0.0587 3.22 0.1123 11.27
H0100 0.0402 4.58 0.0819 4.49 0.0637 3.49 0.0930 9.33
H0130 0.0778 8.85 0.1873 10.27 0.0875 4.80 0.0335 3.36
H0200 0.1838 20.93 0.1379 7.57 0.1087 5.96 0.0922 9.25
H0230 0.1172 13.34 0.1147 6.29 0.0814 4.46 0.0716 7.18
H0300 0.1528 17.41 0.0976 5.36 0.0650 3.56 0.0577 5.78
H0330 0.1384 15.76 0.0969 5.32 0.0673 3.69 0.0511 5.13
H0400 0.1016 11.58 0.0619 3.39 0.0391 2.14 0.0417 4.18
H0430 0.0310 3.54 0.0066 0.36 0.0226 1.24 0.0096 0.96
H0500 0.0192 2.19 0.0174 0.95 0.0279 1.53 0.0051 0.52
H0530 0.1118 12.73 0.1614 8.86 0.1441 7.90 0.0533 5.35
H0600 0.0723 8.23 0.0870 4.77 0.0769 4.22 0.0122 1.22
H0630 0.1757 20.00 0.2156 11.82 0.1587 8.70 0.0983 9.86
H0700 0.1988 22.63 0.1196 6.56 0.0752 4.12 0.1173 11.77
H0730 0.0456 5.19 0.0587 3.22 0.0583 3.20 0.0605 6.07
H0800 0.1350 15.37 0.1741 9.55 0.1550 8.50 0.1343 13.46
(continued)
944 V. Ramiah et al.

Table 33.2 (continued)

SA1 VIC1 NSW1 QLD1
Variable Coeff t-stats Coeff t-stats Coeff t-stats Coeff t-stats
H0830 0.1454 16.55 0.1281 7.02 0.1061 5.82 0.1089 10.91
H0900 0.0286 3.26 0.0284 1.56 0.0123 0.67 0.0371 3.72
H0930 0.0160 1.83 0.0733 4.02 0.0659 3.61 0.0588 5.89
H1000 0.0173 1.97 0.0023 0.13 0.0019 0.10 0.0256 2.56
H1030 0.0396 4.51 0.0015 0.08 0.0015 0.08 0.0228 2.29
H1100 0.0120 1.37 0.0150 0.82 0.0142 0.78 0.0213 2.13
H1200 0.0219 2.50 0.0067 0.37 0.0106 0.58 0.0072 0.72
H1230 0.0111 1.27 0.0170 0.93 0.0232 1.27 0.0227 2.27
H1300 0.0020 0.23 0.0030 0.17 0.0033 0.18 0.0254 2.55
H1330 0.0020 0.23 0.0432 2.37 0.0429 2.35 0.0153 1.53
H1400 0.0051 0.58 0.0046 0.25 0.0022 0.12 0.0180 1.80
H1430 0.0402 4.57 0.0063 0.35 0.0052 0.28 0.0005 0.05
H1500 0.0277 3.15 0.0011 0.06 0.0049 0.27 0.0416 4.17
H1530 0.0373 4.25 0.0046 0.25 0.0018 0.10 0.0125 1.26
H1600 0.0105 1.20 0.0160 0.88 0.0171 0.94 0.0014 0.14
H1630 0.0403 4.59 0.0063 0.34 0.0109 0.60 0.0081 0.81
H1700 0.0002 0.03 0.0353 1.94 0.0290 1.59 0.0436 4.37
H1730 0.0355 4.04 0.0685 3.75 0.0853 4.67 0.1127 11.30
H1800 0.1622 18.44 0.2035 11.13 0.2390 13.06 0.2770 27.69
H1830 0.0675 7.68 0.0584 3.20 0.0574 3.14 0.0789 7.91
H1900 0.0408 4.65 0.0235 1.29 0.0237 1.30 0.0380 3.81
H1930 0.1043 11.88 0.0743 4.07 0.0820 4.49 0.1058 10.60
H2000 0.0473 5.39 0.0344 1.89 0.0492 2.70 0.0593 5.94
H2030 0.0698 7.95 0.0336 1.84 0.0382 2.10 0.1199 12.02
H2100 0.0858 9.77 0.0672 3.69 0.0679 3.72 0.0876 8.79
H2130 0.0493 5.61 0.0253 1.39 0.0018 0.10 0.0524 5.26
H2200 0.0994 11.32 0.0998 5.47 0.0901 4.94 0.0910 9.13
H2230 0.0502 5.72 0.1401 7.69 0.1923 10.54 0.1460 14.64
H2300 0.0635 7.23 0.0659 3.62 0.0690 3.79 0.0764 7.66
H2330 0.1348 15.35 0.2989 16.40 0.1890 10.36 0.1219 12.22
C 0.0145 1.91 0.0206 1.31 0.0215 1.37 0.0015 0.17
Return(1) 0.0071 13.23 0.0006 0.34 0.0047 3.44 0.0048 8.25
Return(2) 0.0078 4.38 0.0062 4.54 0.0042 7.11
Return(3) 0.0077 4.34 0.0059 4.31 0.0003 0.47
Return(4) 0.0009 1.48
Return(5) 0.0085 14.37
R-squared 0.9646 0.6119 0.7681 0.9577
Adj R-squared 0.9645 0.6114 0.7678 0.9576
a
F-test was carried to exclude redundant lag returns (from lag 6 to lag 20)
b
Akaike info criterion and Schwarz criterion were used to determine the optimal lags for the
different regions
33 Seasonal Aspects of Australian Electricity Market 945

Day-of-week effects generally appeared stronger for Monday, Tuesday, and

Wednesday than other days of the week. Monthly effects are not found to be
consistent across regions, nor are yearly effects. There appears to be no clear
pattern in monthly effect across regions, although a small but significant negative
effect is noted for October in NSW1 and VIC1. Positive effects are noted for winter
months in SA1, and QLD demonstrates significant positive effects in spring and
summer months. As expected, there is considerable evidence of positive spikes in
the return series, and the evidence shows that negative prices within the Australian
electricity market cannot be ignored. Such results cast doubts on the application of
lognormal prices to the Australian electricity market.

33.3 Australian Demand for Electricity

Given the instantaneous market-clearing nature of prices in the AEMO, a logical

extension of this study is to investigate the prevalence of seasonal effects and spike
behavior in electricity demand, with a view to examining the extent to which these
effects are transmitted from demand to price and how efficiently the spot market
absorbs demand-side shocks.

33.3.1 Demand Data

The demand data used in this study are half-hourly observations of total demand.
The data is obtained from the same source as the price data and covers the same
period and regions. The basic quantity of interest in demand modelling and
forecasting is typically the periodic “total system demand” or “total demand.”
The total demand value reported by AEMO is a derived value, somewhat different
from demand (as may be represented by traded volume), as it may be understood in
conventional financial markets. Suppliers and distributors lodge schedules and
bids for the sale and purchase of electricity with AEMO at 12:30 p.m. on the day
prior to actual dispatch of electricity for each interval. AEMO compiles this data
and mates it with a short-term forecast of system demand and grid capacity
to determine an expected dispatch quantity and dispatch order of generators
(Smith 2000).
This study uses AEMO’s reported “total demand” values for each region,
expressed in megawatts (MW) by half-hour trading interval for the sample period.
Total demand is defined by AEMO as the total forecast regional demand against
which a dispatch solution is performed. For any particular interval and region, this
is determined as described by Eq. 33.3:

X
n X
n X
n
DT ¼ Gi  Li þ NI i þ AIL þ FðDÞ þ ADE (33.3)
i¼1 i¼1 i¼1
946 V. Ramiah et al.

7500

7000

6500

6000
MW

5500

5000

4500
VIC1 Demand - 4/6/00 to 11/6/00
4000

Fig. 33.2 Plot of VIC1 demand for the week commencing Monday 4/6/00, illustrating the regular
intraday and daily seasonal patterns in the demand series

where:
DT is total demand.
Gi is “generator initial MW (SCADA),” the sum of initial MW values for all
scheduled generation units within the region, measured at their generator termi-
nals and reported by SCADA.
Li is “load initial MW (SCADA),” the scheduled base-load generation level for the
interval.
NIi is “net interconnector initial MW into region,” the net of all interconnector flows
into and out of the region.
AIL is “total allocated interconnector losses” represented by ∑(MW losses
X regional loss allocation). “MW losses” represent actual power losses due to
physical leakage from the transmission system. Regional loss allocation is an
NEMMCO predetermined static loss factor for each interconnector.
F(D) is demand forecast, a per-interval demand adjustment that relates the demand
at the beginning of the interval to the target at the end of the interval.
ADE is “aggregate dispatch error,” an adjustment value used by the NEM to
account for disparities between scheduled and actual dispatch for all scheduled
generation units in the region.
Figure 33.2 is provided to illustrate the presence of seasonal patterns in the
intraday behavior of electricity demand in Victoria over a 10-day period in 2000.
Descriptive statistics for the demand series are shown in Table 33.3.
We report the mean, standard deviation, minimum, maximum, range, skewness,
kurtosis, Jarque-Bera statistic, and Augmented Dickey-Fuller2 statistics for each

2
Additional statistics are provided for the demand series but not for the price series. This is because
the original paper on the price series reports these values.
33 Seasonal Aspects of Australian Electricity Market 947

Table 33.3 Descriptive statistics for demand by region, January 1999 to January 2006
Demand NSW1 QLD1 SA1 VIC1
Meana 8123.89 5197.74 1451.56 5411.02
S.D.a 1306.02 864.33 264.45 757.26
Maximuma 12884.15 8231.95 2873.03 8545.39
Minimuma 4624.03 2945.96 778.00 2726.88
Skewness 0.03 0.12 0.74 0.04
Kurtosis 2.64 2.67 4.44 2.83
JB stat 700.17 874.81 22026.19 175.72
ADFb 22.83 17.50 28.30 24.77
N 124224 124224 124224 124224
a
Mean, standard deviation, maximum, and minimum are expressed in megawatts (MW)
b
Augmented Dickey-Fuller (ADF) statistic rejects the hypothesis of a unit root at the 1 % level of
confidence

region’s demand series. NSW1 has the highest mean, median, and maximum demand
observations of the five regions for the period. New South Wales is Australia’s most
populous state so we would expect that demand for electric power to be highest in the
NSW1 region. The other regions follow generally in order of state population, with
VIC1 next highest, followed by QLD1, and SA1. The standard deviation and range of
demand levels are generally high, widely dispersed across the regions, and broadly
consistent with the pattern of means, ranging from 264 MW for SA1 to 1,306 MW for
NSW1. The distributions of demand observations are slightly positively skewed for
all regions. NSW1, QLD1, and VIC1 are slightly platykurtic, while SA1 is
leptokurtic. Jarque-Bera (JB) statistics reject the hypothesis of normal distribution
at the 1 % level of significance for all four regions, and Augmented Dickey-Fuller
statistics reject the hypothesis of a unit root at the 1 % level of significance.

33.3.2 Demand Returns

In this section, we consider the proportional changes in demand over a price

interval, which for convenience we refer to as “demand returns.” The demand
return series are of interest because there are a number of over-the-counter and
exchange-traded derivative products available for hedgers and speculators in the
Australian and overseas electricity markets. Pricing models for derivatives are
informed by the behavior of returns on the spot price. The half-hourly pool price
and its associated returns exhibit strong seasonal and outlier effects as a result of the
occurrence of price spikes. Demand is widely regarded as a major influence on price
(and therefore returns), and we are interested in investigating the extent to which
the seasonalities observed in half-hourly returns on spot price are present in the
equivalent returns on demand. Figure 33.3 shows demand and demand return
over a 10-day period and indicates that demand returns appear to exhibit some
time-of-day effects but also suggests the presence of sudden and fast-reverting
spikes in the demand return series.
948 V. Ramiah et al.

Percent
8000 0

−1
7000

−2
MW

6000

5000

VIC1 Demand vs Demand Return - 4/6/00 to 11/6/00

4000

Demand Demand Return

Fig. 33.3 Plot of VIC1 demand and returns on demand for the week commencing 4/6/00.
Demand is in MW and returns are percentage returns

In light of the fact that AEMO’s total demand is reported at half-hourly intervals
in discrete time, the demand return series used in this study were generated as half-
hourly discrete returns rather than log returns, according to Eq. 33.4 as follows:

ðDt  Dt1 Þ
RDt ¼ (33.4)
Dt1

where RDt is discrete demand return at time t, Dt is half-hourly demand at time t, and
Dt1 is the previous half-hourly total demand, i.e., at time t1. The results of tests for
the presence of a unit root give us confidence that the demand and return series are
stationary, and we prefer this discrete return specification over log returns, as a log
return specification will dampen the spike effects we are attempting to capture.
We define a spike in demand returns as any observed demand return greater than
four standard deviations larger than the mean. Tables 33.4 and 33.5 collates the
occurrences of spikes as defined. Panel (a) shows the occurrence of spikes by region
and in aggregate for weekday, month, and year. Panels (b) and (c) show the
occurrence of spikes by half-hourly trading interval.
Table 33.4 Panel (a) shows that in aggregate there are 208 spikes in demand
returns observed across all regions during the sample period. VIC1 shows the
highest incidence of demand spikes with 92 (44 %) of the 208 observed, followed
by NSW1 with 81 (38 %), and SA1 with 22 (10 %) spikes during the sample period.
By day of the week, Monday shows the highest incidence with 67 (32 %)
tapering gradually to Sunday with 15 occurrences (7 %). August shows the highest
incidence by month with 42 (20 %), and of these 37 occur in NSW1. The next
33 Seasonal Aspects of Australian Electricity Market 949

Table 33.4 Panel (a) Summary of occurrences of extreme demand spikes by region, day of week,
month, and year
Interval NSW1 QLD1 SA1 VIC1 Total
Sun 1 4 4 6 15
Mon 39 3 8 17 67
Tue 12 2 1 21 36
Wed 12 3 4 13 32
Thu 13 0 3 15 31
Fri 4 0 0 15 19
Sat 0 1 2 5 8
Jan 0 0 0 23 23
Feb 0 0 0 17 17
Mar 1 1 5 31 38
Apr 0 0 0 8 8
May 13 0 2 0 15
Jun 20 2 3 0 25
Jul 3 4 4 0 11
Aug 37 1 1 3 42
Sep 4 1 0 1 6
Oct 2 0 1 1 4
Nov 1 4 4 3 12
Dec 0 0 2 5 7
1999 15 4 4 78 101
2000 13 0 2 9 24
2001 17 0 2 4 23
2002 16 3 0 1 20
2003 10 1 0 0 11
2004 7 4 3 0 14
2005 3 1 11 0 15
2006 0 0 0 0 0
Total 81 13 22 92 208

highest incidences by month are March (38) and June (25), with spikes predomi-
nantly in VIC1 for March and in NSW1 in June. The highest incidence by year
occurs in 1999 with 101 spikes (49 %), dropping markedly in 2000 (24) and 2001
(23). The incidence of spikes appears to have settled somewhat from 2003 onwards
at around 15–16 spikes per year.3
Table 33.5 Panel (b) shows the incidence of extreme spikes in demand returns by
half-hourly trading interval. There are concentrations of spikes occurring at the 06:30
(91 spikes, of which 78 occur in NSW1) and 23:30 (80 spikes, all of which occur in
VIC1). A subperiod analysis of demand returns suggests that sharp peaks in demand

3
Sample data for 2006 only includes the month of January and is unlikely to be representative of
the full year.
950 V. Ramiah et al.

Table 33.5 Panel (b) Occurrence of extreme demand spikes by half-hourly trading interval
Interval NSW1 QLD1 SA1 VIC1 Total
H0000 0 0 1 0 1
H0030 0 0 0 0 0
H0100 0 0 0 0 0
H0130 0 0 0 0 0
H0200 0 0 0 0 0
H0230 0 0 0 0 0
H0300 0 0 0 0 0
H0330 0 0 0 0 0
H0400 0 0 0 0 0
H0430 0 0 0 0 0
H0500 0 1 0 0 1
H0530 1 0 0 2 3
H0600 0 0 0 0 0
H0630 78 7 0 6 91
H0700 0 0 5 0 5
H0730 0 1 2 0 3
H0800 0 1 1 1 3
H0830 0 0 1 0 1
H0900 1 1 2 1 5
H0930 0 0 0 0 0
H1000 0 0 0 0 0
H1030 0 0 0 0 0
H1100 0 0 0 0 0
H1130 0 0 0 0 0
H1200 0 0 0 1 1
H1230 0 0 0 0 0
H1300 0 0 1 0 1
H1330 0 0 2 0 2
H1400 0 0 0 0 0
H1430 0 0 1 0 1
H1500 0 0 0 0 0
H1530 0 0 0 0 0
H1600 0 0 0 0 0
H1630 0 0 0 0 0
H1700 0 0 0 0 0
H1730 0 1 0 0 1
H1800 1 1 1 0 3
H1830 0 0 5 1 6
H1900 0 0 0 0 0
H1930 0 0 0 0 0
H2000 0 0 0 0 0
H2030 0 0 0 0 0
(continued)
33 Seasonal Aspects of Australian Electricity Market 951

Table 33.5 (continued)

Interval NSW1 QLD1 SA1 VIC1 Total
H2100 0 0 0 0 0
H2130 0 0 0 0 0
H2200 0 0 0 0 0
H2230 0 0 0 0 0
H2300 0 0 0 0 0
H2330 0 0 0 80 80

Victoria Intra-Day Effect (1999–2005)

0.1

0.08

0.06
Changes in demand

0.04

0.02

0
0

0
:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3

:3
00

01

02

03

04

05

06

07

08

09

10

11

12

13

14

15

16

17

18

19

20

21

22

23
−0.02

−0.04

−0.06
Time

Fig. 33.4 Half-hourly demand returns for VIC1 for the period 1999–2005

returns are persistent throughout the sample period, as illustrated by Fig. 33.4.
Figure 33.4 shows the pattern of demand returns for VIC1 for the period
1999–2005 and is illustrative of the pattern in NSW1. The 06:30 peak in demand
appears consistent with the commencement of the morning peak in activity in the
population. We believe that the 2,330 peak coincides with the activation of off-peak
hot water systems set to take advantage of overnight off-peak retail electricity tariffs.
Descriptive statistics for the half-hourly demand return series are shown in
Table 33.6. We report the mean, standard deviation, minimum, maximum, range,
skewness, kurtosis, and Augmented Dickey-Fuller statistics for each region’s
demand return series.
Mean, standard deviation, maximum, and minimum are expressed in terms of
half-hourly percentage return and are broadly consistent across NSW1, QLD1, SA1,
and VIC1. The distributions of demand returns for all four regions demonstrate
positive skewness and high positive kurtosis. Jarque-Bera (JB) statistics reject the
null hypothesis of normal distribution at the 1 % level of significance for all four
regions. This fat-tailed character is consistent with studies on price behavior (see
Huisman and Huurman 2003; Higgs and Worthington 2005; Wolack 2000) and
appears driven by the presence of spikes in demand returns. Augmented Dickey-
Fuller (ADF) statistics robustly reject the hypothesis of a unit root at the 1 % level of
significance for all five regions, again consistent with the findings of the earlier studies.
952 V. Ramiah et al.

Table 33.6 Descriptive statistics for half-hourly demand returns, by region, January 1999 to
January 2006
Returns NSW1 QLD1 SA1 VIC1
Meana 0.05 0.04 0.05 0.04
S.D.a 3.15 2.87 3.31 2.98
Maximuma 44.50 52.64 38.41 49.52
Minimuma 26.55 31.94 38.97 33.81
Skewness 1.13 0.94 0.39 1.11
Kurtosis 5.00 5.43 4.45 5.01
JB stat 47,301.83 49,025.43 14,003.32 46,573.67
ADFb 42.57 43.29 41.51 41.51
N 1,24,223 1,24,223 1,24,223 1,24,223
a
Mean, standard deviation, maximum, and minimum are expressed as percentage values
b
Augmented Dickey-Fuller (ADF) statistic rejects the hypothesis of a unit root at the 1 % level of
confidence

33.3.3 Modelling Seasonality in Demand for Electricity

In this section, we adapt and adjust Eq. 33.2 to model seasonality in demand return.
The equation is as follows:

X
6 X
11 X
2006
RDit ¼ f0 þ b1 LRt þ b2 DAY i þ b3 MTH i þ b4 YRi
n¼1, i6¼Sun n¼1, i6¼Sep n¼1999, i6¼2001
(33.5)
X
47 XNS
þ b5 HH i þ b6 DDSPIKEi þ et
n¼1, i6¼1, 130 hrs n¼1

where RDit represents the discrete demand return for region i at time t, f0 represents
the constant term, LRit represents the lagged demand return for region i at time t,
DDSPIKEi represents the dummy variable set for each occurrence of extreme return
as previously defined, and the remaining variables are defined as in Eq. 33.2. The
trading interval at 11:30 h, Sunday, September, and the year 2001 were incorpo-
rated into the constant term a in the model as the base case for each dummy series.
These base cases were selected as the trading interval, day, month, and year, in
which demand return activity was consistently lowest in all four regions.

33.4 Empirical Results

Results of the regression analysis are presented in Tables 33.7 and 33.8. Coeffi-
cients and t-statistics are presented for each seasonal dummy variable and for
lagged returns. In view of the very large number of individual spikes in demand
returns (208 spikes identified for the sample period across all regions), coefficients
for individual spikes are not explicitly reported though results are discussed.
33 Seasonal Aspects of Australian Electricity Market 953

Table 33.7 Panel (a) – Results Of regression analysis for half-hourly demand return against
seasonal dummy variables, by region for day, month, and year
NSW1 QLD1 SA1 VIC1
Coeff t-stats Coeff t-stats Coeff t-stats Coeff t-stats
C .00603 8.05 .00170 3.66 .00356 4.02 .00228 7.43
Mon 0.00099 7.88 0.00084 3.24 0.00091 4.57 0.00058 4.24
Tue 0.00014 1.41 0.00020 0.80 0.00011 0.70 0.00004 0.30
Wed 0.00010 0.96 0.00014 0.57 0.00005 0.34 0.00006 0.42
Thu 0.00004 0.30 0.00020 0.79 0.00047 2.36 0.00002 0.17
Fri 0.00099 7.27 0.00105 4.05 0.00092 4.28 0.00062 4.58
Sat 0.00025 1.87 0.00033 1.28 0.00018 0.85 0.00021 1.58
Jan 0.00018 0.38 0.00001 0.03 0.00009 0.12 0.00008 0.46
Feb 0.00005 0.10 0.00005 0.13 0.00016 0.23 0.00004 0.20
Mar 0.00003 0.05 0.00008 0.22 0.00006 0.09 0.00018 0.98
Apr 0.00009 0.19 0.00025 0.68 0.00014 0.20 0.00011 0.61
May 0.00009 0.20 0.00009 0.24 0.00022 0.31 0.00004 0.24
Jun 0.00003 0.07 0.00010 0.27 0.00038 0.54 0.00003 0.18
Jul 0.00011 0.23 0.00012 0.34 0.00016 0.23 0.00002 0.12
Aug 0.00003 0.06 0.00005 0.13 0.00040 0.61 0.00000 0.00
Oct 0.00000 0.01 0.00001 0.02 0.00009 0.14 0.00001 0.03
Nov 0.00008 0.18 0.00001 0.02 0.00013 0.19 0.00004 0.22
Dec 0.00012 0.26 0.00010 0.28 0.00041 0.59 0.00000 0.03
1999 0.00009 0.20 0.00006 0.24 0.00019 0.31 0.00031 2.26
2000 0.00003 0.06 0.00007 0.25 0.00002 0.03 0.00000 0.01
2002 0.00002 0.05 0.00001 0.03 0.00004 0.06 0.00001 0.10
2003 0.00001 0.02 0.00001 0.03 0.00003 0.05 0.00002 0.14
2004 0.00005 0.12 0.00001 0.04 0.00004 0.07 0.00002 0.15
2005 0.00002 0.05 0.00001 0.03 0.00009 0.15 0.00003 0.18
2006 0.00013 0.13 0.00008 0.11 0.00129 0.86 0.00005 0.15
R2 0.87 0.83 0.82 0.82 0.87
Adj R2 0.87 0.83 0.82 0.82 0.87

We find that coefficients are relatively small for NSW1, QLD1, SA1, and VIC1.
Day-of-week effects are positive and significant for Monday in NSW1, QLD1, SA1,
and VIC1. Thursday shows significant negative effect in demand returns in SA1 only.
There is significant negative Friday effect in all regions. There appears to be no clear
pattern evident for monthly effect across regions. There is no clear pattern evident in
yearly effect, although significant negative effect is observed for 1999 in VIC1. Half-
hourly time-of-day effects offer more interesting results and are broadly more
consistent across regions than the seasonal effects previously discussed. In general,
significant negative demand returns are found for the small hours of the morning
from 12:30 a.m. until approximately 4:00–5:00 a.m. in all regions. VIC1 exhibits an
unexplained highly significant positive return at 1:30 a.m., reverting to negative
returns until for the remainder of the early morning. There is wide variation in the
pattern of demand returns during the waking day. NSW1 and VIC1 show broadly
954 V. Ramiah et al.

Table 33.8 Panel (b): Results of regression analysis for demand return against seasonal dummy
variables, by halfhourly trading interval by region, 0000–2,330 h
NSW1 QLD1 SA1 VIC1
Coeff t-stats Coeff t-stats Coeff t-stats Coeff t-stats
H0000 0.013 14.30 0.040 91.44 0.044 49.76 0.055 139.42
H0030 0.008 8.91 0.057 114.74 0.013 14.63 0.017 48.79
H0100 0.014 15.48 0.047 89.84 0.004 4.89 0.010 26.90
H0130 0.014 15.96 0.043 87.43 0.062 70.67 0.086 238.13
H0200 0.020 22.39 0.036 74.58 0.064 68.82 0.066 169.17
H0230 0.019 21.03 0.029 61.46 0.053 55.84 0.023 62.27
H0300 0.009 10.63 0.019 41.36 0.040 42.74 0.012 33.13
H0330 0.005 5.91 0.010 23.47 0.031 33.90 0.006 17.25
H0400 0.005 5.71 0.004 9.95 0.023 25.48 0.002 5.04
H0430 0.015 17.58 0.005 12.12 0.013 15.02 0.007 20.52
H0500 0.020 22.54 0.013 29.54 0.004 4.49 0.009 26.51
H0530 0.048 54.16 0.034 76.97 0.005 5.53 0.027 77.09
H0600 0.039 43.14 0.045 93.50 0.022 24.88 0.021 59.57
H0630 0.049 53.66 0.079 157.08 0.028 31.42 0.045 123.52
H0700 0.045 48.87 0.084 133.88 0.052 58.70 0.017 44.80
H0730 0.011 12.11 0.082 133.22 0.041 44.66 0.022 57.89
H0800 0.036 40.06 0.058 94.63 0.023 25.87 0.028 78.38
H0830 0.020 22.05 0.035 67.65 0.031 35.51 0.009 24.48
H0900 0.003 3.91 0.018 38.06 0.019 21.19 0.002 6.33
H0930 0.015 17.25 0.014 30.12 0.007 8.13 0.007 20.14
H1000 0.001 1.24 0.004 8.62 0.005 5.81 0.001 3.97
H1030 0.002 2.45 0.001 1.36 0.002 3.36 0.002 5.85
H1100 0.002 3.03 0.002 6.49 0.001 0.98 0.003 7.98
H1200 0.004 5.80 0.001 5.05 0.001 2.53 0.002 6.15
H1230 0.003 3.82 0.004 10.98 0.002 2.08 0.001 2.39
H1300 0.002 2.03 0.005 10.64 0.000 0.23 0.001 3.86
H1330 0.004 5.03 0.001 3.23 0.002 2.92 0.010 28.51
H1400 0.002 2.14 0.001 2.88 0.006 6.97 0.005 12.89
H1430 0.005 5.82 0.000 0.89 0.010 11.42 0.000 0.61
H1500 0.005 5.44 0.003 6.23 0.006 6.46 0.002 4.52
H1530 0.007 7.57 0.001 1.98 0.010 11.66 0.003 7.70
H1600 0.011 12.34 0.004 9.22 0.009 9.88 0.005 14.29
H1630 0.008 8.69 0.008 17.99 0.004 4.93 0.002 6.33
H1700 0.019 22.22 0.016 36.28 0.003 3.59 0.009 23.96
H1730 0.014 15.77 0.021 46.66 0.005 5.65 0.005 13.52
H1800 0.027 30.31 0.027 58.50 0.010 11.86 0.010 28.01
H1830 0.001 1.18 0.013 28.47 0.015 16.97 0.006 16.80
H1900 0.001 1.01 0.006 14.25 0.002 2.06 0.004 10.01
H1930 0.004 5.07 0.010 21.73 0.008 8.85 0.005 13.97
H2000 0.002 2.52 0.014 32.78 0.012 14.03 0.002 5.10
(continued)
33 Seasonal Aspects of Australian Electricity Market 955

Table 33.8 (continued)

NSW1 QLD1 SA1 VIC1
Coeff t-stats Coeff t-stats Coeff t-stats Coeff t-stats
H2030 0.010 11.60 0.023 51.22 0.010 11.32 0.007 20.68
H2100 0.008 8.61 0.022 48.41 0.017 19.81 0.008 23.12
H2130 0.004 4.81 0.021 47.15 0.021 23.79 0.014 39.32
H2200 0.012 13.21 0.026 57.72 0.021 24.30 0.011 29.98
H2230 0.028 31.77 0.016 35.83 0.024 27.00 0.004 10.35
H2300 0.014 15.49 0.019 42.48 0.000 0.10 0.000 0.52
H2330 0.000 0.14 0.005 11.30 0.009 10.29 0.085 234.83

similar intraday patterns, with significant positive demand returns dominating

between 04:30 and 18:00 and with minor variation reverting to negative demand
returns for the remainder of the evening. QLD1 and SA1 are broadly similar,
demonstrating significant positive effect between 04:30 and 10:30 and predominantly
negative effect from 10:30 to 15:30, reverting to positive effect from 16:00 to 19:30
when we see reversion to negative effect for the remainder of the evening. The
periods of positive return in the early morning and early evening are consistent with
peaks in activity in the population. We would expect to observe positive demand
returns arising from off-peak hot water systems generally switching on at 11:00 p.m.,
but curiously positive returns around that hour are evident only in NSW1.

33.5 Conclusions

The current literature establishes that electricity time series differ from traditional
financial data having greater incidence of spikes than is generally observed in
financial data and this results in extreme volatility. The work done overseas is of
limited practical application to the Australian market as our market structure is
different. The lesson learned from our study is that even within the same country,
variation exists as different states reflect variation in geographic differences. The
implication is that specific models must be developed for each geographic region. Our
work is innovative in that we develop an econometric time series seasonal model that
can be applied in areas where seasonality is suspected. For instance, the model was
initially applied to price and return series and then later on to demand return series.
This model has the capability to be applied in other financial time series models.
Developing models to explain and to predict electricity prices is a significant
task. This is important both for the market participants who operate in the physical
market and for those trading in electricity derivatives. Modelling electricity prices
and trading in this market is challenging, so much so that investment banks avoid
trading electricity derivatives because of the incidence of spikes. This provides
a strong incentive for this research and future research into the electricity market.
A better understanding of the time series nature of the electricity prices may help in
the development of more efficient forecasting models which will help to lower risk
956 V. Ramiah et al.

management costs and thus reduce the cost of managing electricity price exposures.
Electricity generators, electricity retailers, transmission and distribution network
suppliers, electricity retailer companies, and electricity consumers will all benefit
from development of more accurate models in this area.

Acknowledgment The authors wish to acknowledge the financial support of the Australian
Centre for Financial Studies, via grant 18/2005. The research on the seasonality within the spot
price has been published in 2011, “Seasonal factors and outlier effects in rate of return in
Australia’s National Electricity Market,” Applied Economics, Vol. 43, Issue 3; pp. 355–369.
Binesh Seetanah provided invaluable research assistance. Any remaining errors are the responsi-
bility of the authors.

References
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81(3), 637.
Bunn, D., & Karakatsani, N. (2003). Forecasting electricity prices. Working paper, London
Business School.
Clewlow, L., & Strickland, C. (2000). Valuing energy options in a one-factor model fitted to
forward prices. Working paper, University of Sydney.
Eichler, M., Grothe, O., Manner, H., & Tuerk, D. (2012). Modeling spike occurrences in electricity
spot prices for forecasting. Working paper, RM/12/029, Maastricht University.
Engle, R., Granger, C., & Hallman, J. (1989). Merging short and long-run forecasts: An applica-
tion of seasonal cointegration to monthly electricity sales forecasting. Journal of Economet-
rics, 50, 297–327.
Escribano, A. J., Peaea J., & Villaplana, P. (2002). Modelling electricity prices: International
evidence. Working paper, Universidad Carlos III de Madrid.
Goto, M., & Karolyi, G. A. (2004). Understanding electricity price volatility within and across
markets. Dice Centre working paper 2004–12, pp. 1–41.
Harvey, A., & Koopman, S. J. (2002). Forecasting hourly electricity demand using time-varying
splines. Journal of the American Statistical Association, 88(424), 1228–1236.
Huisman, R., & Huurman, C. (2003). Fat tails in power prices. ERIM report series reference
no. ERS-2003-059-F&A. SSRN: http://ssrn.com/abstract¼450992
Johnson, B., & Barz, G. (1999). Selecting stochastic processes for modelling electricity prices. In
Energy modelling and the management of uncertainty. Risk Publications.
Kaminski, V. (1997). The challenge of pricing and risk managing electricity derivatives. In The
U.S. Power Market (pp. 149–71).
Knittel, C. R., & Roberts, M. (2001). An empirical examination of deregulated electricity prices.
Working paper, PWP-087, University of California Energy Institute.
Montero, J. M., Garcia, M. C., & Fernandez-Aviles, G. (2011). Modelling the volatility of the
Spanish wholesale electricity spot market. Asymmetric GARCH models vs. Threshold ARSV
model. Revista de Estudios de Economı́a Aplicada, 29(2), 1–20.
Pardo, A., Meneu, V., & Valor, E. (2002). Temperature and seasonality influences on Spanish
electricity load. Energy Economics, 24, 55–70.
Smith, M. (2000). Modelling and short-term forecasting of New South Wales electricity system
load. Journal of Business & Economic Statistics, 18(4), 465–478.
Thomas, S., Ramiah, V., Heaney, R., & Mitchell, H. (2011). Seasonal factors and outlier effects in
rate of return in Australia’s national electricity market. Applied Economics, 43(3), 355–369.
Pricing Commercial Timberland Returns
in the United States 34
Bin Mei and Michael L. Clutter

Contents
34.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 958
34.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 960
34.2.1 CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 960
34.2.2 Fama-French Three-Factor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 961
34.2.3 CAPM and Fama-French Three-Factor Model Under the State Space
Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 961
34.2.4 Stochastic Discount Factor Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 962
34.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963
34.3.1 Timberland Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963
34.3.2 Basis Assets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964
34.3.3 Other Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964
34.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 965
34.4.1 Estimation of the CAPM and the Fama-French Three-Factor Model . . . . . . . . 965
34.4.2 State Space Estimation of the CAPM and the Fama-French
Three-Factor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 965
34.4.3 Abnormal Performance Measured by the SDF Approach . . . . . . . . . . . . . . . . . . . . 966
34.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 967
Appendix 1: State Space Model with the Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 970
Appendix 2: Heuristic Proof of Equation 34.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 971
Appendix 3: NCREIF Timberland Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 972
Appendix 4: EViews Code for Estimating the CAPM and the Fama-French
Three-Factor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 972
Appendix 5: Steps for the SDF Approach Using Excel Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975

B. Mei (*) • M.L. Clutter

Warnell School of Forestry and Natural Resources, University of Georgia, Athens, GA, USA
e-mail: bmei@uga.edu; mclutter@warnell.uga.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 957
DOI 10.1007/978-1-4614-7750-1_34,
# Springer Science+Business Media New York 2015
958 B. Mei and M.L. Clutter

Abstract
Commercial timberland assets have attracted more attention in recent decades.
One unique feature of this asset class roots in the biological growth, which is
independent of traditional financial markets. Using both parametric and non-
parametric approaches, we evaluate private- and public-equity timberland
investments in the United States. Private-equity timberland returns are proxied
by the NCREIF Timberland Index, whereas public-equity timberland returns are
proxied by the value-weighted returns on a dynamic portfolio of the US publicly
traded forestry firms that had or have been managing timberlands. The results
from parametric analysis reveal that private-equity timberland investments
outperform the market and have low systematic risk, whereas public-equity
timberland investments fare similarly as the market. The nonparametric
stochastic discount factor analyses reveal that both private- and public-equity
timberland assets have higher excess returns.
Static estimations of the capital asset pricing model and Fama-French
three-factor model are obtained by ordinary least squares, whereas dynamic
estimations are by state space specifications with the Kalman filter. In estimating
the stochastic discount factors, linear programming is used.

Keywords
Alternative asset class • Asset pricing • Evaluation • Fama-French three-factor
model • Nonparametric analysis • State space model • Stochastic discount factor •
Timberland investments • Time series • Time-varying parameter

34.1 Introduction

Timberland investments have been unprecedentedly active in the United States

the past 30 years. A number of factors have motivated public attention toward
timberland assets. On the supply side, because of the internal subsidies from
timber divisions to processing mills, timberland properties managed by traditional
vertically integrated forest products firms have been undervalued by the Wall
Street. To fix this mispricing, these firms began divesting their timberlands as
a strategic move. For example, International Paper, a global leading forest products
firm, has disposed of most of its timberlands and focused on its core business of
paper and packaging products manufacturing in recent decades. Currently, almost
no forest products firms in the United States still own timberlands (Harris
et al. 2010). On the demand side, institutional investors, e.g., organizations with
fiduciary obligations such as pension funds, university endowments, foundations,
and trusts, have diversified into nonfinancial asset classes such as commercial
timberlands on the passage of Employee Retirement Income Security Act
(ERISA) in 1974.
34 Pricing Commercial Timberland Returns in the United States 959

There are several ways to invest in commercial timberlands. High-net-wealth

families and individuals can participate in commingled (pooled) funds, or they can
own and manage timberlands directly. Others can buy stocks and bonds of
publicly traded forestry firms that focus on timberland business. Most institutional
investors hold timberland properties via timberland investment management
organizations (TIMOs). TIMOs manage their institutional assets in either sepa-
rately managed accounts (individually managed accounts) or pooled funds.
A separately managed account holds timberland properties of one investor in
a single portfolio, whereas a pooled fund collects capital from a number of
investors and allocates it to a portfolio of timberland properties. Investors tend to
have more discretion with separate accounts than pooled funds (Zinkhan
and Cubbage 2003). In 2008, there were about 30 TIMOs in the United
States, and the total value of their timberland assets exceeded $35 billion
(Zinkhan 2008).
Since the public recognition of commercial timberland as an alternative asset
class, a number of studies have been conducted to assess the financial performance
of timberland investments. The major findings of previous research can be summa-
rized as follows. (1) Timberland has countercyclical returns or low (even negative
in some cases) correlation with the financial assets (Binkley et al. 1996; Cascio and
Clutter 2008; Redmond and Cubbage 1988; Washburn and Binkley 1990; Zinkhan
1988). (2) Timberland can be an effective hedge against unexpected inflation
(Fortson 1986; Washburn and Binkley 1993). (3) If timberland investors can exploit
the biological growth of timber thus time the market, they can get higher and better
returns (Caulfield 1998; Conroy and Miles 1989; Haight and Holmes 1991).
(4) Relative inefficiency tends to exist in timberland markets (Caulfield 1998),
although this situation has been alleviated through time (Washburn 2008; Zinkhan
2008). (5) Among a variety of forestry-related investment vehicles, institutional
timberland investments and timberland limited partnerships have low-risk levels
but excess returns (Sun and Zhang 2001). (6) In the long run, timber and/or
timberland returns are cointegrated with other nontimber financial instruments
(Heikkinen 2002; Liao et al. 2009).
Almost all of the above studies are based on the single-period capital asset
pricing model (CAPM). Sun and Zhang (2001) extended the literature in timberland
investments by employing the arbitrage pricing theory (APT), and Heikkinen
(2002) and Liao et al. (2009) expanded the literature by using cointegration
analysis. Nevertheless, all those methods are parametric in nature. This study has
several contributions. First, timberland assets are considered separately in private
and public markets, and their returns are compared. Second, supplementary to
the ordinary least squares (OLS) estimation of the CAPM and the Fama-French
three-factor model, a state space model with the Kalman filter is employed to
examine the time-varying risk-adjusted excess return (a) and systematic risk (b).
Finally, the nonparametric stochastic discount factor (SDF) approach is introduced
for pricing timberland returns.
960 B. Mei and M.L. Clutter

The major results are that private-equity timberland investments have

significant excess returns but low systematic risk, whereas public-equity timberland
investments fare similarly as the market, and that intertemporal consumption
decisions affect the intertemporal marginal rate of substitution of timberland
investors and thus impact the rational pricing of timberland assets. These results
can further our understanding of the financial aspects of commercial timberland
assets in the United States. The next two sections describe the methodologies and
the data. Section 34.4 explains the empirical results, and the last section makes
some concluding remarks.

34.2 Methods

For the parametric method, an explicit model is needed. Two candidate models
prevalent in the finance literature are the CAPM and the Fama-French three-factor
model. The parametric method is often criticized for the “joint hypothesis tests”
problem, i.e., testing the asset pricing model and the abnormal performance (market
efficiency) simultaneously. The nonparametric method does not require such an
explicit model specification and is therefore not subject to these critiques. The SDF
approach is a general, nonparametric asset pricing approach and is complement to
the parametric approaches.

34.2.1 CAPM

Built on Markowitz’s (1952) groundwork of mean-variance efficient portfolio,

Sharpe (1964) and Lintner (1965) developed its economy-wide implications – the
CAPM. The CAPM states that the expected return on an asset or a portfolio E[Ri]
equals a risk-free rate Rf plus a premium that depends on the asset’s bi and the
expected risk premium on the market portfolio E[Rm]
Rf, i.e.,


E ½ R i  ¼ R f þ bi E ½ R m 
R f : (34.1)

In empirical regression analysis, the CAPM is estimated in the excess return

form


Ri
Rf ¼ ai þ bi Rm
Rf þ mi , (34.2)

where ex post realized returns Ri and Rm rather than ex ante expected returns E[Ri]
and E[Rm] are used. The intercept ai is called Jensen’s (1968) alpha. A positive a
suggests that the individual asset outperforms the market and earns a higher than
risk-adjusted return, whereas a negative a suggests that the individual asset
underperforms the market and earns a lower than risk-adjusted return. Therefore,
Jensen’s alpha has become a commonly used measure of abnormal performance,
and testing whether it is zero has been widely used in the empirical asset pricing
literature.
34 Pricing Commercial Timberland Returns in the United States 961

34.2.2 Fama-French Three-Factor Model

Given the empirical evidence that small size stocks outperform large size stocks,
and value (high book-to-market) stocks outperform growth (low book-to-market)
stocks on average, Fama and French (1993) develop a model that includes these
extra two factors to adjust for risk:

E½Ri 
Rf ¼ bRMRF, i E½RRMRF  þ bSMB, i E½RSMB  þ bHML, i E½RHML , (34.3)

where RRMRF ¼ Rm
Rf is the same market factor as in the CAPM, representing the
market risk premium; RSMB ¼ Rsmall
Rbig is the size factor, representing the return
difference between a portfolio of small stocks and a portfolio of large stocks (SMB
stands for “small minus big”); RHML ¼ RhighBM
RlowBM is the book-to-market
factor, representing the return difference between a portfolio of high book-to-
market stocks and a portfolio of low book-to-market stocks (HML stands for
“high minus low”); and b’s are called factor loadings, representing each asset’s
sensitivity to these factors. When estimating Fama-French three-factor model, ex
post realized returns are used, as in the case of the CAPM, and an intercept is added
to capture the abnormal performance,

Ri
Rf ¼ ai þ bRMRF, i RRMRF þ bSMB, i RSMB þ bHML, i RHML þ ei : (34.4)

34.2.3 CAPM and Fama-French Three-Factor Model Under the State

Space Framework

The CAPM (Eq. 34.2) and the Fama-French three-factor model (Eq. 34.4) are
usually estimated by OLS, possibly with some correction for the autocorrelations
in the errors. One restrictive nature of the OLS method is that the coefficients in the
regression are imposed to be constant. This condition may be unrealistic in real
asset pricing modeling. For instance, one would suspect that both a’s and b’s should
be time varying. To solve this problem, we can estimate the CAPM and the Fama-
French three-factor model in the state space framework with the Kalman filter
(Appendix 1). Using the CAPM as an example, in the state space framework, the
system of equations is specified as


Ri, t
Rf , t ¼ ai, t þ bi, t Rm, t
Rf , t þ mi, t
ai, tþ1 ¼ ai, t þ xt (34.5)
bi, tþ1 ¼ bi, t þ tt

where mi,t, xt, and tt are normally and independently distributed mean-zero
error terms. In the state space model, the first equation in (34.5) is called
the observation or measurement equation, and the second and third equations are
called the state equations. In this particular case, each state variable follows
a random walk.
962 B. Mei and M.L. Clutter

One advantage of the state space approach with time-varying parameters is that
it can incorporate external shocks, such as policy and regime shifts, economic
reforms, and political uncertainties, into the system, especially when the shocks
are diffuse in nature (Sun 2007). This approach has been applied to a variety of
issues, including demand systems (e.g., Doran and Rambaldi 1997), aggregate
consumptions (e.g., Song et al. 1996), policy analysis (e.g., Sun 2007), and price
modeling and forecasting (e.g., Malaty et al. 2007).

34.2.4 Stochastic Discount Factor Approach

The single-period asset pricing models ignore the consumption decisions. In effect,
investors make their consumption and portfolio choices simultaneously in an
intertemporal setting. In the framework of an exchange economy in which an
investor maximizes the expectation of a time-separable utility function (Lucas
1978), it can be proved that (Appendix 2)

 
Et 1 þ Ri, tþ1 Mtþ1 ¼ 1, (34.6)

where Ri,t+1 is the return on asset i in the economy and Mt+1 is known as the
stochastic discount factor, or intertemporal marginal rate of substitution, or pricing
kernel (e.g., Campbell et al. 1997).
Hansen and Jagannathan (1991) demonstrated how to identify the SDF from
a set of basis assets, i.e., the derivation of the volatility bounds. These bounds
are recognized as regions of admissible mean-standard deviation pairs of the
SDF. Their major assumptions are the law of one price and the absence of arbitrage
opportunities. Accordingly, there are two particular solutions for the SDF: the law
of one price SDF and the no-arbitrage SDF. The process of retrieving the reverse-
engineered law of one price SDF is equivalent to the following constrained
optimization problem:
" #1=2
1 X T
Min sMt ¼ ðM t
v Þ
Mt T
1 t¼1
1X T
s:t: Mt ¼ v (34.7)
T t¼1
1X T

Mt 1 þ Ri, t ¼ 1
T t¼1

for a range of selected v (mean of Mt) and for all assets i ¼ 1,2,  ,N. Under the
stronger condition of no arbitrage, another positivity constraint on Mt is needed.
Therefore, the only difference between the law of one price SDF and the
no-arbitrage SDF is whether Mt is allowed to be negative. In this study,
no-arbitrage SDF is used. Following Hansen and Jagannathan (1991),
34 Pricing Commercial Timberland Returns in the United States 963

nonnegativity instead of positivity restriction Mt  0 is added to retrieve the

no-arbitrage SDF. Last, sample size T should be sufficiently large such that the
time-series version of law of large numbers applies; that is, the sample moments on
a finite record converge to their population counterparts as the sample size becomes
large (Hansen and Jagannathan 1991).
Provided the existence of a risk-free asset, it can be shown that

 
Et Ri, tþ1
Rf Mtþ1 ¼ 0: (34.8)

This equation presents the basis for testing the risk-adjusted performance of
a portfolio (Chen and Knez 1996). Namely, one can test whether
  
 
ai ¼ Et ai, t ¼ Et Ri, tþ1
Rf Mtþ1 ¼ 0: (34.9)

Ahn et al. (2003) pointed out that this measure generalizes Jensen’s alpha
and does not count on a specific asset pricing model. Based on this method,
they reassess the profitability of momentum strategies and found that their non-
parametric risk adjustment explains almost half of the anomalies.

34.3 Data

34.3.1 Timberland Returns

Returns for both private- and public-equity timberland investments are analyzed.
Although TIMOs have become the major timberland investment management
entities for institutional investors as well as high-net-wealth families and individ-
uals, their financial data are rarely publicly available. To provide a performance
benchmark, several TIMOs, together with National Council of Real Estate
Investment Fiduciaries (NCREIF) and the Frank Russell Company, initiated the
NCREIF Timberland Index in early 1992 (Binkley et al. 2003) (Appendix 3).
NCREIF members can be divided into data contribution members, professional
members, and academic members. Data contribution members include investment
managers and plan sponsors who own or manage real estate in a fiduciary setting.
Professional members include providers of accounting, appraisal, legal, consulting,
or other services to the data contribution members. Academic members include
full-time professors of real estate. Data contribution members submit their data on
a quarterly basis for computation of the NCREIF Property Index. Regarding
the NCREIF Timberland Index, it is some TIMOs that are the major data
contribution members. The quarterly NCREIF Timberland Index is reported at
both regional (the South, the Northeast, and the Pacific Northwest) and national
levels, and extends back to 1987. In this study, the national-level NCREIF
Timberland Index (1987Q1–2010Q4) is used as a return proxy for the US
private-equity timberland investments.
964 B. Mei and M.L. Clutter

Returns on public-equity timberland investments are proxied by the value-

weighted returns on a dynamic portfolio of the US publicly traded forestry firms
that had or have been managing timberlands. These firms include Deltic Timber,
the Timber Co, IP Timberlands Ltd., Plum Creek, Pope Resources, Potlatch,
Rayonier, and Weyerhaeuser. Deltic Timber and Pope Resources are natural
resources companies focused on the ownership and management of timberland;
the Timber Co and IP Timberlands Ltd. are subsidiaries of Georgia-Pacific and
International Paper that track the value and performance of their timberland prop-
erties; Plum Creek, Potlatch, and Rayonier are publicly traded real estate invest-
ment trusts (REITs) that are engaged in timberland management; and
Weyerhaeuser is a forest products firm that has a significant portion of its business
in timberlands. The market value of each firm is calculated as the product of stock
price and total shares outstanding at the end of each quarter. Financial data for these
forestry firms are obtained from the Center for Research in Security Prices (CRSP).
To be consistent with the NCREIF Timberland Index, the sample spans from
1987Q1 to 2010Q4.

34.3.2 Basis Assets

To mimic the complete investment opportunity set that is available to investors,

a parsimonious set of basis assets needs to be specified. King (1966) proved that
industry groupings maximize intragroup correlation and minimize intergroup cor-
relation and concluded that market and industry factors capture most of the com-
mon variation in stock returns. Following Hansen and Jagannathan (1991), we
construct the reference set by forming industry portfolios according to SIC code.
In this study, two sets of basis assets are chosen – one is the five-industry portfolios
plus long-term treasury bonds, and the other is the ten-industry portfolios plus long-
term treasury bonds. The industry groups are derived from stocks listed on NYSE,
AMEX, and NASDAQ based on their four-digit SIC codes. The five industries are
classified as consumer goods, manufacturing, Hi-Tech, healthcare, and others,
whereas the ten industries are classified as consumer nondurables, consumer dura-
bles, manufacturing, energy, Hi-Tech, telephone and television transmission,
shops, healthcare, utilities, and others. Value-weighted returns on the industry
portfolios are obtained from Kenneth R. French’s website, and returns on the
portfolio of long-term treasury bonds are obtained from CRSP. Presuming that
the basis assets are rationally priced, the SDF can be retrieved.

34.3.3 Other Indices

Market returns are approximated by the value-weighted returns on all NYSE,

AMEX, and NASDAQ stocks from CRSP. Risk-free rate, as approximated by the
1-month treasury bill rate from Ibbotson Associates, Inc., and Fama-French factors
are available on Kenneth R. French’s website.
34 Pricing Commercial Timberland Returns in the United States 965

34.4 Empirical Results

34.4.1 Estimation of the CAPM and the Fama-French

Three-Factor Model

Panel A of Table 34.1 presents the OLS estimation of the CAPM and the Fama-
French three-factor model using the quarterly NCREIF Timberland Index after
adjustment for the seasonality. A significant positive a from the CAPM suggests
that private-equity timberland investments have a risk-adjusted excess return of
about 8.20 % (2.05 %  4) per year. This excess return is slightly larger
after accounting for Fama-French factors. Market b’s from both models are insig-
nificantly different from zero, but significantly less than one. This means that
private-equity timberland investments are not only weakly correlated with the
market but also less risky than the market. The small magnitudes with high
p-values of the coefficients for SMB and HML signify that these two extra factors
add limited explanatory power to the CAPM in pricing private-equity timberland
returns.
In contrast, the CAPM and the Fama-French three-factor model fit the returns
on the dynamic portfolio of forestry firms much better, as implied by the higher R2
values (Panel A of Table 34.2). This is within our expectation since these forestry
firms are publicly traded and are more exposed to the market. However, a’s
are insignificant albeit positive, indicating no abnormal performance. Market b’s
are significantly different from zero, but not from one. In addition, b’s for SMB
and HML in Fama-French three-factor model are significant at the 5 % level or
better, meaning these factors capture some variations in the portfolio returns that
are not explained by the market premium. As a result, the abnormal performance (a)
has dropped by 53 %. The magnitudes of b’s indicate that the dynamic portfolio is
dominated by mid-large firms with middle book-to-market ratios.

34.4.2 State Space Estimation of the CAPM and the Fama-French

Three-Factor Model

Panel B of Table 34.1 presents the state space estimation of the CAPM and the
Fama-French three-factor model using the NCREIF Timberland Index. Those OLS
coefficient estimates are used as the starting values. Only a is specified as a state
variable (stochastic level) in that little time variation is observed in b, and both AIC
and SBC favor the deterministic-b model. Back to the model specification in system
(Eq. 34.5), this is equivalent to restrict tt ¼ 0. The AIC and SBC are marginal-
ly larger than those for the OLS estimation because of the relatively small sample
size. Figure 34.1 depicts the evolution of the risk-adjusted excess returns of the
NCREIF Timberland Index estimated from the CAPM. For most time in the last
22 years, the NCREIF Timberland Index has achieved positive abnormal returns
with an average of 10.20 % per year (calculated from the estimated a series).
Nevertheless, in certain years (2001–2003), the a is low and even negative,
966 B. Mei and M.L. Clutter

Table 34.1 Estimation of the CAPM and the Fama-French three-factor model using the NCREIF
Timberland Index (1987Q1–2010Q4)
CAPM FF3
Coefficient Estimate p-value Coefficient Estimate p-value
Panel A: OLS estimation
a 2.05 0.001 a 2.11 0.001
b 0.01 0.773 bRMRF 0.02 0.675
bSMB
0.04 0. 634
bHML
0.05 0.318
H0: b ¼ 1 0.000 H0: bRMRF ¼ 1 0.000
R2 0.14 R2 0.15
Log likelihood
251.24 Log likelihood
250.59
S.E. of regression 3.78 S.E. of regression 3.82
Durbin-Watson stat. 1.94 Durbin-Watson stat. 1.96
Akaike info. criterion 5.53 Akaike info criterion 5.56
Schwarz criterion 5.61 Schwarz criterion 5.69
F-stat. 7.09 F-stat. 3.82
Panel B: State space estimation
a 0.54 0.692 a 0.79 0.549
b
0.02 0.919 bRMRF 0.01 0.856
bSMB
0.11 0.263
bHML
0.05 0.581
H0: b ¼ 1 0.000 H0: bRMRF ¼ 1 0.000
Log likelihood
275.00 Log likelihood
273.59
Akaike info. criterion 5.86 Akaike info criterion 5.80
Schwarz criterion 5.94 Schwarz criterion 5.94
(1) OLS estimates after correction for the fourth-order autocorrelation in the residuals. (2) Only a
is specified stochastic under the state space framework, while b is specified deterministic due to its
lack of variation and AIC criterion

indicating no abnormal performance. Although not reported here, the time-varying

a’s estimated from Fama-French three-factor model exhibit similar patterns.
For the dynamic portfolio, however, only b is specified to be stochastic since
little time variation is observed in a, and both AIC and SBC favor the deterministic-
a model. The time-varying b of the dynamic portfolio of forestry firms is plotted in
Fig. 34.2. Overall, there is a decreasing trend in the market b. The average b over
the sample period is 1.05, which is not significantly different from the market risk.

34.4.3 Abnormal Performance Measured by the SDF Approach

The mean of the no-arbitrage SDF Mt is specified in the selected range of [0.9750, 1]
with an increment step of 0.0025. When the five-industry portfolios plus
the long-term treasury bonds are used as the basis assets, the global minimum
variance of Mt is identified at v ¼ 0.9800; when the ten-industry portfolios plus
34 Pricing Commercial Timberland Returns in the United States 967

Table 34.2 Estimation of the CAPM and the Fama-French three-factor model using returns on
a dynamic portfolio of publicly traded forestry firms in the United States (1987Q1–2010Q4)
CAPM FF3
Coefficient Estimate p-value Coefficient Estimate p-value
Panel A: OLS estimation
a 0.59 0.521 a 0.28 0.750
b 0.95 0.000 bRMRF 0.87 0.000
bSMB 0.39 0.031
bHML 0.31 0.001
H0: b ¼ 1 0.330 H0: bRMRF ¼ 1 0.249
R2 0.48 R2 0.54
Log likelihood
344.00 Log likelihood
337.27
S.E. of regression 8.80 S.E. of regression 8.29
Durbin-Watson stat. 2.19 Durbin-Watson stat. 2.24
Akaike info. criterion 7.21 Akaike info criterion 7.11
Schwarz criterion 7.26 Schwarz criterion 7.22
F-stat. 85.16 F-stat. 36.58
Panel B: State space estimation
a 0.59 0.500 a 0.24 0.786
b 0.95 0.000 bRMRF 0.84 0.000
bSMB 0.39 0.021
bHML 0.32 0.001
H0: b ¼ 1 0.390 H0: bRMRF ¼ 1 0.446
Log likelihood
355.38 Log likelihood
348.61
Akaike info. criterion 7.47 Akaike info criterion 7.37
Schwarz criterion 7.55 Schwarz criterion 7.50
Only b is specified stochastic under the stochastic framework, while a is specified deterministic
due to its lack of variation and AIC criterion

the long-term treasury bonds are used instead, the global minimum variance of Mt is
identified at v ¼ 0.9750.
The SDF performance measures for both the NCREIF Timberland Index, and the
returns on the dynamic portfolio of publicly traded timber firms are reported in
Table 34.3. The a values for both return indices have increased, and the latter has
become marginally significant. This indeed implies that intertemporal consumption
decisions play a key role in pricing timberland assets. In a word, there is clear
evidence of statistically as well as economically significant excess returns for the
NCREIF Timberland Index, but only some evidence of economically significant
excess returns for the portfolio of publicly traded timber firms.

34.5 Conclusion

Using both parametric and nonparametric techniques, in this study, we reexamined

the financial performance of timberland investments. Private-equity timberland
968 B. Mei and M.L. Clutter

10 % Alpha
+/− 2 RMSE
8

−2

−4

−6
1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010

Fig. 34.1 Evolution of a over time from the state space estimation of the CAPM using the
NCREIF Timberland Index (1987Q1–2010Q4). Note: The time-varying a estimated from Fama-
French three-factor model exhibits similar patterns, thus is not shown separately. The graph is
available from the authors upon request

3.0
Beta
+/− 2 RMSE

2.5

2.0

1.5

1.0

0.5
1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010

Fig. 34.2 Evolution of b over time from the state space estimation of the CAPM using returns on
a dynamic portfolio of publicly traded forestry firms in the United States (1987Q1–2010Q4). Note:
The time-varying b estimated from Fama-French three-factor model exhibits similar patterns, thus
is not shown separately. The graph is available from the authors upon request
34 Pricing Commercial Timberland Returns in the United States 969

Table 34.3 Performance measures of timberland returns by the nonparametric SDF approach
(1987Q1–2010Q4)
Performance measure (a) p-value (one tail)
Mean of Mt (v) S.D. of Mt ðsMt Þ (1) (2) (1) (2)
Panel A: Five-industry portfolios plus long-term T-bonds
0.9775 0.199 2.63 1.59 0.000 0.119
0.9800 0.176 2.60 1.36 0.000 0.156
0.9825 0.217 2.57 1.13 0.000 0.202
Panel B: Ten-industry portfolios plus long-term T-bonds
0.9725 0.244 2.76 2.03 0.000 0.056
0.9750 0.237 2.75 1.82 0.000 0.082
0.9775 0.255 2.77 1.60 0.000 0.116
Column (1) is for the NCREIF Timberland Index, and Column (2) is for the returns on a dynamic
portfolio of the US publicly traded forestry firms that had or have been managing timberlands

returns are approximated by the NCREIF Timberland Index, whereas public-equity

timberland returns are approximated by the value-weighted returns on a dynamic
portfolio of the US publicly traded timber firms. The parametric analyses reveal that
private-equity timberland assets outperform the market but have low systematic
risk, whereas public-equity timberland assets perform similarly as the market.
Therefore, inclusion of private-equity timberland properties can improve the effi-
cient frontier, albeit such potential is limited for public-equity timberland proper-
ties. Unlike the parametric methods, the nonparametric SDF approach does not rely
on any specific asset pricing models and hence are not subject to the “joint
hypothesis test” criticisms. The results from the SDF approach suggest higher
excess returns for both private- and public-equity timberland investments, which
in turn signify the important role of intertemporal consumption decisions in rational
pricing of timberland assets.
The positive a of private-equity timberland returns may be associated with the
patience of institutional investors toward embedded strategic options for timber-
lands (Zinkhan 2008). If a timberland property has potential for higher and better
use such as residential or commercial development opportunities, or if it is suitable
for conservation easements, or if it has mineral or gas opportunities, it may have
extra income sources, and the land value can be dramatically higher. The positive a
may also be related to the liquidity risk that institutional investors bear since
a typical TIMO has an investment time horizon of 10–15 years or even longer. In
contrast, stocks of publicly traded timber firms can be easily traded on the stock
exchanges. Moreover, initiation of a TIMO-type separately managed account
usually requires a capital commitment of $25–$50 million, while participation in
a TIMO-type pooled fund generally requires a minimum capital commitment of
$1–$3 million (Zinkhan and Cubbage 2003). The large capital amount may enable
the investors to achieve some degree of diversification.
The lower excess returns of the NCREIF Timberland Index around 2001–2003
may be associated with its relative weak performance during that time. In 2001Q4,
970 B. Mei and M.L. Clutter

the NCREIF Timberland Index fell by 6.5 %, the largest drop it ever had, which was
primarily caused by the capital loss from the shrinking timberland values. In the
same period, the S&P 500 index went up by 7.8 %. The overall decreasing trend in
b for the dynamic portfolio of forestry firms may be related to the massive
restructurings of these firms. For instance, Plum Creek, Potlatch, and Rayonier
have converted themselves into timber REITs in recent years. With improved tax
efficiency and increased concentration on timberland management, these timber
REITs are expected to be less risky.
Another interesting fact noted in this study is that, despite the current economic
downturn triggered by the subprime residential mortgage blowup, private-equity
timberland returns remain relatively strong. While the CRSP market index went
down 39 % in 2008, the NCREIF Timberland Index achieved a 9 % return, or on the
risk-adjusted basis, an excess return of 10 % (calculated using the estimated a series
in 2008). In contrast, the portfolio value of publicly traded timber firms fell 39 %
just like the market. However, it should be noted that most of those forestry firms do
have non-timberland business, such as paper and lumber mills, which may be more
sensitive to the overall economic conditions. A close examination of the three
publicly traded timber REITs reveals that they were less affected by the gloomy
market. Looking ahead, global economic crisis will last for some time, multiple
factors will affect timberland returns, and the net effect on timberland properties
has yet to be observed (Washburn 2008).
It should be noted that there have been some concerns about the data and method
consistency of the NCREIF Timberland Index. As pointed out by Binkley
et al. (1996), there are no standardized appraisal and valuation practice in forestry,
so heterogeneity may exist in the data. In addition, due to lack of quarterly
appraisals for many properties in the NCREIF Timberland Index, quarterly return
series may be less useful than the annual ones. Finally, the NCREIF Timberland
Index is a composite performance measure of a very large pool of commercial
forestland properties acquired in the private market for investment purposes.
Hence, individual investors should use cautions when interpreting the NCREIF
Timberland Index.

Appendix 1: State Space Model with the Kalman Filter

The multivariate time-series model can be represented by the following state space
form:

yt ¼ Z t at þ et , et  NIDð0; H t Þ (34.10)

atþ1 ¼ T t at þ Rt t , t  NIDð0; Qt Þ (34.11)

for t ¼ 1,   , N, where yt is p  1 vector of observed values at time t, Zt is

a p  m matrix of variables, at is m  1 state vector, Tt is called the transition matrix
of order m  m, and Rt is an m  r selection matrix with m  r. The first equation is
34 Pricing Commercial Timberland Returns in the United States 971

called the observation or measurement equation, and the second is called state
equation. The parameters at, Ht, and Qt in the system of equations can be estimated
jointly by the maximum likelihood method with the recursive algorithm Kalman
filter. The intention of filtering is to update the information of the system each time
a new observation yt is available, and the filtering equations are

v t ¼ y t
Z t at ,
0
Ft ¼ Z t Pt Z t þ H t ,
0
K t ¼ T t Pt Z t F
1
t , (34.12)
Lt ¼ T t
K t Z t ,
atþ1 ¼ T t at þ K t vt ,
0 0
Ptþ1 ¼ T t Pt Lt þ Rt Qt Rt ,

For t ¼ 1,   , N. The mean vector a1 and the variance matrix P1 are known for
the initial state vector a1 (Durbin and Koopman 2001; Harvey 1989).

Appendix 2: Heuristic Proof of Equation 34.6

In a pure exchange economy with identical consumers, a typical consumer wishes

to maximize the expected sum of time-separable utilities
" #
X
1
Max Et b UðCtþi Þ
i
Ct
i¼0
  (34.13)
X
N X
N
s:t: xjt pjt þ Ct ¼ W t þ xjt
1 pjt þ djt
j¼1 j¼1

where x jt is the amount of security j purchased at time t, pjt is the price of security j at
time t, Wt is the individual’s endowed wealth at time t, Ct is the individual’s
consumption at time t, djt is the dividend paid by security j at time t, and b is time
discount. Express Ct in terms of xjt , and differentiate the objective function with
respect to xjt , then we can get the following first-order condition:
h i h  i
Et U 0 ðCt Þpjt ¼ Et bU0 ðCtþ1 Þ pjtþ1 þ d jtþ1 (34.14)

for all j. After rearranging the terms, we can reach Eq. 34.6, where

bU 0 ðCtþ1 Þ
Mt ¼
U 0 ðC t Þ
(34.15)
pj þ d j
Rtþ1 ¼ tþ1 j tþ1
1:
pt
972 B. Mei and M.L. Clutter

Appendix 3: NCREIF Timberland Index

The NCREIF Timberland Index has two components, the income return and the
capital return. The income return is also known as EBITDDA return, which
represents earnings before interest expenses, income taxes, depreciation, depletion,
and amortization. The capital return is derived from land appreciation. The formu-
las to calculate these returns are

EBITDDAt
IRt ¼ (34.16)
MV t
1 þ 0:5ðCI t
PSt þ PPt
EBITDDAt Þ

MV t
MV t
1
CI t þ PSt
PPt

CRt ¼ (34.17)
MV t
1 þ 0:5ðCI t
PSt þ PPt
EBITDDAt Þ

where IRt and CRt are the income return and capital return, respectively; EBITDDAt
equals the net operating revenue obtained from the tree farm (primarily from timber
sales); CIt equals the capitalized expenditures on the tree farm (e.g., forest regen-
eration and road construction); PSt equals the net proceeds from sales of land from
the tree farm; PPt equals the gross costs of adding land to the tree farm; and MVt
equals the market value of the tree farm (Binkley et al. 2003).

Appendix 4: EViews Code for Estimating the CAPM and the

Fama-French Three-Factor Model

Texts after the single quotation marks are notations.

’ Specify the file location
CD "C:\Mei Bin\Publication\Handbook of FES"

’ Create a workfile in EViews and read in quarterly data

1987Q1-2010Q4
Workfile Timberland_Jul2011 q 1987 2010
’ 7 is the total number of series to be read in
Read(t ¼ dat, s) quarterly.csv 7
’ Group the 7 series
group quarterly NCREIF MktRf SMB HML RF port
’ Estimate the CAPM for NCREIF, adjusted for autocorrelation
’ LS means OLS. Equation given by dependent variable
followed by a list of independent variables
’ In excess returns
equation CAPM1.LS (NCREIF-RF) C MktRf AR(4)
’ Estimate the Fama-French three-factor model for the
NCREIF Timberland Index
equation FF31.LS (NCREIF-RF) C MktRf SMB HML AR(4)
34 Pricing Commercial Timberland Returns in the United States 973

’ Estimate the CAPM for the portfolio, AR(4) term dropped

due to its insignificance.
equation CAPM2.LS (port-RF) C MktRf
’ Estimate the Fama-French three-factor model for the
portfolio
equation FF32.LS (port-RF) C MktRf SMB HML

’ State space estimation of time-varying parameters for

the CAPM
’ Define a state space model for the NCREIF Timberland Index
sspace KFcapm1
’ Signal equation, the CAPM
’ Define alpha to be time-varying (state variable)
KFcapm1.append @signal (NCREIF-RF) ¼ sv1 + c(1)*MktRf +
[var ¼ exp(c(2))]
’ State equation as a random walk
KFcapm1.append @state sv1 ¼ sv1(-1) + [var ¼ exp(c(3))]
’ Starting values for the state space model. Values come
the OLS estimation
KFcapm1.append @param c(1) 0.01 c(2) 0 c(3) 0
’ Maximum likelihood estimation
KFcapm1.ml(showopts, m ¼ 500, c ¼ 0.0001, m)
Show KFcapm1.output
’ Save the time-varying alphas and its RMSEs
KFcapm1.makestates(t ¼ filt) CAPM1filt*
KFcapm1.makestates(t ¼ filtse) CAPM1filtse*

’ Generate the graph of time-varying alphas with the 95 %

confidence intervals
series CAPM1_a_bandplus ¼ CAPM1filtsv1 + 2*CAPM1filtsesv1
series CAPM1_a_bandminus ¼ CAPM1filtsv1 - 2*CAPM1filtsesv1
’ Group the series to be shown in a graph
Group CAPM1_a_curves CAPM1filtsv1 CAPM1_a_bandplus
CAPM1_a_bandminus

’ State space model for the portfolio of public forest

products firms
sspace KFcapm2
’ Define beta to be time-varying (state variable)
KFcapm2.append @signal (port-RF) ¼ c(1) + sv1*MktRf +
[var ¼ exp(c(2))]
KFcapm2.append @state sv1 ¼ sv1(-1) + [var ¼ exp(c(3))]
KFcapm2.append @param c(1) 0.59 c(2) 4.3 c(3) -30
KFcapm2.ml(showopts, m ¼ 500, c ¼ 0.0001, m)
974 B. Mei and M.L. Clutter

Show KFcapm2.output

KFcapm2.makestates(t ¼ filt) CAPM2filt*

KFcapm2.makestates(t ¼ filtse) CAPM2filtse*
series CAPM2_b_bandplus ¼ CAPM2filtsv1 + 2*CAPM2filtsesv1
series CAPM2_b_bandminus ¼ CAPM2filtsv1 - 2*CAPM2filtsesv1

Group CAPM2_b_curves CAPM2filtsv1 CAPM2_b_bandplus

CAPM2_b_bandminus

sspace KFff1
’ Time-varying alpha in the Fama-French three-factor
model
KFff1.append @signal (NCREIF-RF) ¼ sv1 + c(1)*MktRf + c(2)
*SMB + c(3)*HML + [var ¼ exp(c(4))]
KFff1.append @state sv1 ¼ sv1(-1) + [var ¼ exp(c(5))]
KFff1.append @param c(1) 0.02 c(2) -0.04 c(3) -0.05 c(4) 0
c(5) 0
KFff1.ml(showopts, m ¼ 500, c ¼ 0.0001, m)
Show KFff1.output

sspace KFff2
’ Time-varying beta in the Fama-French three-factor model
KFff2.append @signal (port-RF) ¼ c(1) + sv1*MktRf + c(2)
*SMB + c(3)*HML + [var ¼ exp(c(4))]
KFff2.append @state sv1 ¼ sv1(-1) + [var ¼ exp(c(5))]
KFff2.append @param c(1) 0 c(2) 0 c(3) 3 c(4) 3 c(5) 0
KFff2.ml(showopts, m ¼ 500, c ¼ 0.0001, m)
Show KFff2.output
’ Save the workfile
wfsave Timberland_Jul2011

Appendix 5: Steps for the SDF Approach Using Excel Solver

First, choose minimizing the standard deviation of the SDFs as the objective
function.
Second, set the mean of the SDFs equal to a predetermined value. This is
constraint No.1.
Third, for each basis asset (industry group) in the industry portfolio, add
1X T

one constraint according to Mt 1 þ Ri, t ¼ 1 . That is, add five more
T t¼1
constraints when using the five-industry portfolio, whereas add ten more constraints
34 Pricing Commercial Timberland Returns in the United States 975

when using the ten-industry portfolio. Fourth, specify the solutions to be nonneg-
ative and solve for the SDFs. Fifth, use the SDFs to price timberland returns
according to Eq. 34.9. Repeat steps 1–5 with a different value as the given mean of
the SDFs.

References
Ahn, D.-H., Conrad, J., & Dittmar, R. F. (2003). Risk adjustment and trading strategies. Review of
Financial Studies, 16, 459–485.
Binkley, C. S., Raper, C. F., & Washburn, C. L. (1996). Institutional ownership of U.S. Timberland:
History, rationale, and implications for forest management. Journal of Forestry, 94, 21–28.
Binkley, C. S., Washbrun, C. L., & Aronow, M. E. (2003). The NCREIF timberland property
index: Research notes 2003. Boston: Hancock Timber Resource Group.
Campbell, J. Y., Lo, A. W., & MacKinlay, A. C. (1997). The econometrics of financial markets.
Princeton: Princeton University Press.
Cascio, A. J., & Clutter, M. L. (2008). Risk and required return assessments of equity Timberland
investments in the United States. Forest Products Journal, 58, 61–70.
Caulfield, J. P. (1998). Timberland return drivers and investing styles for an asset that has come of
age. Real Estate Finance, 14, 65–78.
Chen, Z., & Knez, P. J. (1996). Portfolio performance measurement: Theory and applications.
The Review of Financial Studies, 9, 511–555.
Conroy, R., & Miles, M. (1989). Commercial forestland in the pension portfolio: The biological
beta. Financial Analysts Journal, 45, 46–54.
Doran, H. E., & Rambaldi, A. N. (1997). Applying linear time-varying constraints to
econometric models: With an application to demand systems. Journal of Econometrics, 79,
83–95.
Durbin, J., & Koopman, S. J. (2001). Time series analysis by state space methods. New York:
Oxford University Press.
Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics, 33, 3–56.
Fortson, J. C. (1986). Factors affecting the discount rate for forestry investments. Forest Products
Journal, 36, 67–72.
Haight, R. G., & Holmes, T. P. (1991). Stochastic price models and optimal tree cutting: Results
for Loblolly pine. Natural Resource Modeling, 5, 424–443.
Hansen, L. P., & Jagannathan, R. (1991). Implications of security market data for models of
dynamic economies. Journal of Political Economy, 99, 225–262.
Harris, T., Baldwin, S., & Siry, J. (2010). United States Timberland markets 2000 to mid-2010:
Transactions, values & market research. Athens: Timber Mart-South.
Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter.
Cambridge, UK: Cambridge University Press.
Heikkinen, V. P. (2002). Co-integration of timber and financial markets-implications for portfolio
selection. Forest Science, 48, 118–128.
Jensen, M. C. (1968). The performance of mutual funds in the period 1945–1964. The Journal of
Finance, 23, 389–416.
Kenneth, R. French’s website. http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.
html.
King, B. F. (1966). Market and industry factors in stock price behavior. The Journal of Business,
39, 139–190.
Liao, X., Zhang, Y., & Sun, C. (2009). Investments in timberland and softwood timber as parts of
portfolio selection in the United States: A cointegration analysis and capital asset pricing
model. Forest Science, 55, 471–479.
976 B. Mei and M.L. Clutter

Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets. The Review of Economics and Statistics, 47, 13–37.
Lucas, R. E., Jr. (1978). Asset prices in an exchange economy. Econometrica, 46, 1429–1445.
Malaty, R., Toppinen, A., & Viitanen, J. (2007). Modeling and forecasting Finnish pine sawlog
stumpage prices using alternative time-series methods. Canadian Journal of Forest Research,
37, 178–187.
Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7, 77–91.
Redmond, C. H., & Cubbage, F. W. (1988). Portfolio risk and returns from timber asset invest-
ments. Land Economics, 64, 325–337.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance, 19, 425–442.
Song, H., Liu, X., & Romilly, P. (1996). A time varying parameter approach to the Chinese
aggregate consumption function. Economics of Planning, 29, 185–203.
Sun, C. (2007). Variation of federal cost-share programs in the United States and the inducement
effects on tree planting. Journal of Forest Economics, 12, 279–296.
Sun, C., & Zhang, D. (2001). Assessing the financial performance of forestry-related investment
vehicles: Capital Asset Pricing Model vs. Arbitrage Pricing Theory. American Journal of
Agricultural Economics, 83, 617–628.
Washburn, C. (2008). Forestry investments in a slowing global economy. London/Athens: Inter-
national Forest Investment/Center for Forest Business.
Washburn, C. L., & Binkley, C. S. (1990). On the use of period-average stumpage prices to
estimate forest asset pricing models. Land Economics, 66, 379–393.
Washburn, C. L., & Binkley, C. S. (1993). Do forest assets hedge inflation? Land Economics, 69,
215–224.
Zinkhan, F. C. (1988). Forestry projects, modern portfolio theory, and discount rate selection.
Southern Journal of Applied Forestry, 12, 132–135.
Zinkhan, F. C. (2008). History of modern timberland investments. In International forest invest-
ment: An executive education conference (Munich, Germany), Center for Forest Business,
Athens.
Zinkhan, F. C., & Cubbage, F. W. (2003). Financial analysis of timber investments. In E. O. Sills
& K. L. Abt (Eds.), Forests in a market economy. Dordrecht: Kluwer.
Optimal Orthogonal Portfolios with
Conditioning Information 35
Wayne E. Ferson and Andrew F. Siegel

Contents
35.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 978
35.2 Optimal Orthogonal Portfolios: The Classical Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 979
35.3 The Conditional Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 982
35.4 The Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 983
35.5 Empirical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 987
35.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 992
Appendix 1: Theorems and Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993
Efficient Portfolio Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993
Proofs of Propositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 997
Proof of Propositions 2 and 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 997
Proof of Proposition 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
Appendix 2: Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
The Parametric Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1001

Abstract
Optimal orthogonal portfolios are a central feature of tests of asset pricing models
and are important in active portfolio management problems. The portfolios com-
bine with a benchmark portfolio to form ex ante mean variance efficient portfolios.
This paper derives and characterizes optimal orthogonal portfolios in the presence
of conditioning information in the form of a set of lagged instruments. In this
setting, studied by Hansen and Richard (1987), the conditioning information is
used to optimize with respect to the unconditional moments. We present an

W.E. Ferson (*)

University of Southern California, Los Angeles, CA, USA
e-mail: ferson@marshall.usc.edu; wayne.ferson@marshall.usc.edu
A.F. Siegel
University of Washington, Seattle, WA, USA
e-mail: asiegel@uw.edu; asiegel@u.washington.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 977
DOI 10.1007/978-1-4614-7750-1_35,
# Springer Science+Business Media New York 2015
978 W.E. Ferson and A.F. Siegel

empirical illustration of the properties of the optimal orthogonal portfolios. From

an asset pricing perspective, a standard stock market index is far from efficient
when portfolios trade based on lagged interest rates and dividend yields. From an
active portfolio management perspective, the example shows that a strong tilt
toward bonds improves the efficiency of equity portfolios.
The methodology in this paper includes regression and maximum likelihood
parameter estimation, as well as method of moments estimation. We form
maximum likelihood estimates of nonlinear functions as the functions evaluated
at the maximum likelihood parameter estimates. Our analytical results also
provide economic interpretation for test statistics like the Wald test or multivar-
iate F test used in asset pricing research.

Keywords
Asset pricing tests • Conditioning information • Minimum-variance efficiency •
Optimal portfolios • Predicting returns • Portfolio management • Stochastic
discount factors • Generalized method of moments • Maximum likelihood •
Parametric bootstrap • Sharpe ratios

35.1 Introduction

The optimal orthogonal portfolio, also known as the most mispriced portfolio or the
active portfolio, is a central concept in asset pricing tests and in modern portfolio
management. In asset pricing problems, it represents the difference between the
performance of a benchmark portfolio and the maximum potential performance in
a sample of assets (Jobson and Korkie 1982). In modern portfolio management, it
shows how to actively tilt away from a given benchmark portfolio to achieve
portfolio efficiency (Gibbons et al. 1989; Grinold and Kahn 1992).
Optimal orthogonal portfolios are studied by Roll (1980), MacKinlay (1995),
Campbell et al. (1987), and others. However, these studies restrict the analysis to
a setting where the portfolio weights are fixed over time. In contrast, studies in asset
pricing use predetermined variables to model conditional expected returns, corre-
lations, and volatility. Portfolio weights may be functions of the predetermined
variables, and they will generally vary over time. Quantitative portfolio managers
routinely use conditioning information in optimized portfolio strategies. Therefore,
it is important to understand optimal orthogonal portfolios in a conditional setting.
This paper derives, characterizes, and illustrates optimal orthogonal portfolios in
a conditional setting. The setting is one where the conditional means and variances
of returns are time varying and optimal time-varying portfolio weights achieve
unconditional mean variance efficiency with respect to the information, as
described by Hansen and Richard (1987) and Ferson and Siegel (2001).1 Ferson

1
An alternative is to study conditional efficiency, where the weights minimize the conditional
variance. This may be handled by simply reinterpreting the classical analysis.
35 Optimal Orthogonal Portfolios with Conditioning Information 979

and Siegel (2001) argue that this setting is interesting from the perspective of active
portfolio management, where the client cannot observe the information that
a portfolio manager may have. Ferson and Siegel (2003, 2009) show that this
setting is also interesting from the perspective of testing asset pricing models, but
they do not develop the optimal orthogonal portfolio.
We show that the optimal orthogonal portfolio has time-varying weights, and
we derive the weights in a closed form. The portfolio weights for unconditionally
efficient portfolios in the presence of conditioning information are derived by
Ferson and Siegel (2001). They consider the case with no risk-free asset and the
case with a fixed risk-free asset whose return is constant over time. We generalize
these solutions to the case with a “conditionally” risk-free asset whose return is
known at the beginning of the period and is thus included in the lagged condi-
tioning information and may vary over time. We derive solutions for the optimal
orthogonal portfolios with conditioning information, including cases where there
is no risk-free asset, a constant risk-free rate, or a time-varying conditionally risk-
free rate. We show that a “law of conservation of squared Sharpe ratios” holds,
implying that the optimal orthogonal portfolio’s squared unconditional Sharpe
ratio is the difference between that of the benchmark portfolio and the maximum
unconditional squared Sharpe ratio that is possible using the assets and condi-
tioning information. Empirical examples illustrate the performance of the optimal
orthogonal portfolios with conditioning information and the behavior of the
portfolio weights.
Section 35.2 briefly reviews optimal orthogonal portfolios in the classical case
with no conditioning information. Section 35.3 describes the setting for our analysis
with conditioning information. Section 35.4 presents the main theoretical results,
and Sect. 35.5 presents our empirical examples. Section 35.6 concludes. Appendix 1
includes the proofs of the main results, and Appendix 2 describes the methodology
of our empirical examples in detail, including a general description of the
parametric bootstrap.

35.2 Optimal Orthogonal Portfolios: The Classical Case

In the classical case, portfolio weights are fixed constants over time, and there is no
conditioning information. Optimal orthogonal portfolios are tied to mean variance
efficiency. Mean variance efficient portfolios maximize the expected return, given
the variance of the return. Since Markowitz (1952) and Sharpe (1964), such
portfolios have been at the core of financial economics.
The mean variance efficiency of a given portfolio can be described using a system
of time-series regressions. If rt ¼ Rt  g0 is the vector of N excess returns at time t,
measured in excess of a given risk-free or zero-beta return, g0, and rpt ¼ Rpt  g0 is
the excess return on a benchmark portfolio, the regression system is

r t ¼ a þ br pt þ ut ; t ¼ 1, . . . , T, (35.1)
980 W.E. Ferson and A.F. Siegel

where T is the number of time-series observations, b is the N-vector of regression

slopes or betas, and a is the N-vector of intercepts or alphas. The tested portfolio
rpt is represented among the returns in rt, so the covariance matrix of the residuals
in Eq. 35.1 is singular (rpt might be included explicitly or might be a fixed-weight
portfolio of the assets in rt). The portfolio rp is minimum-variance efficient
and has the given zero-beta return only if a ¼ 0.2 The mean variance
efficiency of a given portfolio is of normative investment interest, as an efficient
portfolio maximizes a concave utility function defined solely over the
mean and variance of the portfolio return, as would follow from normally
distributed returns in a single-period model, for example. Equation 35.1
may be interpreted as referring to multiple factor portfolios, where rp is
a K-vector and b is an N  K matrix. Then, the benchmark portfolio is
a linear combination of the K returns in the vector rp (e.g., Shanken 1987;
Gibbons et al. 1989).
Definition The most mispriced (or optimal orthogonal) portfolio with respect to rp,
when there is no conditioning information, has excess return rc ¼ x0 cr, where the
weights xc satisfy.

0 2

xa
xc ¼ Arg Max : (35.2)
x Varðx0 r Þ

It is clear from the definition in Eq. 35.2 why the portfolio is referred to as the
most mispriced. The vector a captures the “mispricing” of the tested asset returns in
Eq. 35.1 when evaluated using the benchmark rp, and the portfolio xc has the largest
squared alpha relative to its variance. This interpretation also reveals why the
portfolio is of central interest in active portfolio management. Given a benchmark
portfolio rp, an active portfolio manager places bets by deviating from the portfolio
weights that define the benchmark. The manager is rewarded for bets that deliver
higher returns and penalized for increasing the volatility. The portfolio in
(2) describes the active bets that achieve the largest amount of extra return for the
variance. Thus, the solution is also referred to as the active portfolio by Gibbons
et al. (1989). (See Grinold and Kahn (1992) for an in-depth treatment of modern
portfolio management.)
In the
classical case of fixed portfolio weights, the solution to Eq. 35.2 is given
0 1
by xc ¼ 1 V 1 a V 1 a , where 1 is an N-vector of ones and V ¼ Cov(r), the
covariance matrix of the returns. Using this solution, several well-known properties

2
Note the distinction between minimum variance efficient portfolios, which minimize the variance
for the given mean return, and mean variance efficient, which maximize the mean return given its
variance. The latter set of portfolios is a subset of the former, typically depicted as the positively
sloped portion of the minimum variance efficient boundary when graphed with mean return on the
y-axis and standard deviation or variance of return on the x-axis. The portfolio rp is mean variance
efficient when a ¼ 0 and E(rp) exceeds the expected excess return of the global minimum variance
portfolio.
35 Optimal Orthogonal Portfolios with Conditioning Information 981

of the optimal orthogonal portfolio follow.3 For example, a combination of the

portfolio rc and the benchmark portfolio rp is optimal, that is, minimum-variance
efficient (Jobson and Korkie 1982). The portfolio is orthogonal to the benchmark
0
portfolio in the sense that Cov(xcr, rp) ¼ 0.
The optimal orthogonal portfolio is central for the interpretation of tests of
portfolio efficiency. Classical test statistics for the hypothesis that a ¼ 0 in
Eq. 35.1 can be written in terms of squared Sharpe ratios (e.g., Jobson and Korkie
1982). Consider the Wald statistic:

!
0 1
^S 2 ðRÞ  ^S 2 Rp
W ¼ T^a ½Covð^a Þ ^a ¼ T 2

_ 2 ðN Þ
w (35.3)
^
1 þ S Rp

where ^a is the OLS or maximum likelihood (ML) estimator of a (after removing rp

or another asset from the vector r to avoid singularity of the covariance matrix) and
Cov(^a) is its asymptotic covariance matrix. Upper case R’s refer to gross returns, and
lower case r’s refer to returns in excess of the zero-beta rate. The term S^2(Rp) is the
sample value of the squared Sharpe ratio of Rp when the zero-beta rate is g0 so that
S2(Rp) ¼ [E(rp)/s(rp)]2. The term S^2(R) is the sample value of the maximum squared
Sharpe ratio that can be obtained by portfolios of the assets in R (including Rp):

(
0 2 )
E xr
S ðRÞ ¼ max
2
: (35.4)
x Varðx0 r Þ

The Wald statistic has an asymptotic chi-squared distribution with N degrees of

freedom, under the null hypothesis that Rp is efficient with the given zero-beta
return. Scaled with a degrees of freedom adjustment, the statistic has an
F distribution under normally distributed returns (Gibbons et al. 1989).
It can be shown that the squared Sharpe ratios can be decomposed using the optimal
orthogonal portfolio as S2(R) ¼ S2(Rp) + S2(Rc). A similar decomposition holds at the
sample values. This decomposition, a “law of conservation of squared Sharpe ratios,” is
used by Jobson and Korkie (1982) to derive the second equality in (35.3). Since the
Sharpe ratio is the slope of a line in the mean-standard deviation space, Eq. 35.3
suggests a graphical representation for the statistic in the sample mean-standard
deviation space. It measures the distance between the sample mean-standard deviation
frontier and the location of the tested portfolio, inside the frontier. This distance is
proportional to the squared Sharpe ratio of the optimal orthogonal portfolio. Kandel
(1984), Roll (1985), Gibbons et al. (1989), and Kandel and Stambaugh (1987, 1989)
further develop this interpretation.

3
See Roll (1980), Gibbons et al. (1989), MacKinlay (1995), and Campbell et al. (1987) for
analyses of optimal orthogonal portfolios in the classical case with no conditioning information.
982 W.E. Ferson and A.F. Siegel

35.3 The Conditional Setting

We use conditioning information in a setting similar to that of Hansen and Richard

(1987) and Ferson and Siegel (2001), where minimum-variance efficiency is defined
in terms of the unconditional means and variances of the portfolios that result from
the use of conditioning information. Ferson and Siegel (2009) refer to this as
efficiency with respect to the information, Z. This setting has proven useful in asset
pricing tests (Ferson and Siegel 2003, 2009; Bekaert and Liu 2004), in forming
hedging portfolios (Ferson et al. 2006), and in portfolio management problems
(Ahkbar et al. 2007; Chiang 2012). We study the optimal orthogonal portfolio in
this setting. The distinction between mean variance efficiency and minimum-variance
efficiency, as in the classical setting, applies in this setting as well.
Consider a portfolio of N assets with gross returns, Rt+1, where the weights that
determine the portfolio at time t are functions of the information, Zt. The gross
return on such a portfolio with weight x(Zt) is x0 (Zt)Rt+1. The restrictions on the
portfolio weight function are that the weights must sum to 1 (almost surely in Zt),
and that the unconditional expected value and second moments of the portfolio
return are well defined. Consider now all portfolio returns that may be formed, for
a given set of asset returns Rt+1 and given conditioning information, Zt, with well-
defined first and second moments. This set determines a mean-standard deviation
frontier, as shown by Hansen and Richard (1987). This frontier depicts the uncon-
ditional means versus the unconditional standard deviations of the portfolio returns.
A portfolio is defined to be efficient with respect to the information Zt if and only if
it is on this mean-standard deviation frontier.
Ferson and Siegel (2001) derive solutions for efficient-with-respect-to-Z portfolios
in closed form. They consider the case with no risk-free asset and the case with a fixed
risk-free asset whose return is constant over time. In Theorem 1 of Appendix 1, we
derive the solution for the case with a risk-free asset whose return is known at the
beginning of the period and is thus included in the information Z and may vary over
time. In this case the variation in the risk-free rate over time affects the unconditional
variance of the portfolio return.
Ferson and Siegel (2001) argue that efficiency with respect to the information
is especially relevant in a portfolio management context. It is reasonable to
assume that the portfolio manager has more information about asset returns than
the client. Assume that the client desires an unconditionally mean-variance
efficient portfolio. The manager observes conditioning information that is rele-
vant about future returns, and by conditioning on this information, he or she can
expand the investor’s opportunity set. The manager maximizes the investor’s
mean variance opportunity set by using his information to maximize the uncon-
ditional mean for a given unconditional variance. The efficient-with-respect-to-
Z strategy is therefore the strategy that the investor would wish the portfolio
manager to use.
Ferson and Siegel (2009) show how asset pricing theories make statements about
portfolios that are efficient with respect to information Z and develop tests of the
hypothesis that a portfolio is efficient with respect to Z. The optimal orthogonal
35 Optimal Orthogonal Portfolios with Conditioning Information 983

portfolio with respect to information Z is a useful concept in these portfolio efficiency

tests. We begin the analysis with a result from Hansen and Richard (1987).
Proposition 1 (Hansen and Richard 1987, Corollary 3.1.) Given N asset gross
returns, Rt+1, a given portfolio with gross return Rp,t+1 is minimum-variance efficient
with respect to the information Zt if and only if Eq. 35.5 is satisfied (equivalently,
0
there exists constants g0 and g1 such that Eq. 35.6 is satisfied) for all x(Zt) such that x ðZ t Þ
1 ¼ 1 almost surely, where the relevant unconditional expectations exist and are finite:

 h 0 i
 h 0 i
Var Rp, tþ1  Var x ðZ t ÞRtþ1 if E Rp, tþ1 ¼ E x ðZ t ÞRtþ1 (35.5)

h 0 i h 0 i
E x ðZt ÞRtþ1 ¼ g0 þ g1 Cov x ðZ t ÞRtþ1 , Rp, tþ1 : (35.6)

Equation 35.5 is the definition of efficiency with respect to Z. It states that Rp, t+1
is on the minimum-variance boundary formed by all possible portfolios that use the
assets in R and the conditioning information. Equation 35.6 states that the familiar
expected return-covariance relation from Fama (1973) and Roll (1977) must hold
with respect to the efficient portfolio. In Eq. 35.6, the coefficients g0 and g1 are fixed
scalars that do not depend on the functions x() or the realizations of Zt.

35.4 The Main Results

The optimal orthogonal portfolio with conditioning information plays roles

analogous to the classical setting with no conditioning information. Thus, for exam-
ple, restricting the maximization in Eq. 35.4 to fixed-weight portfolios where x is
a constant vector, we obtain efficiency in the classical case. In contrast, an efficient
portfolio with respect to the information Z maximizes the squared Sharpe ratio over
all portfolio weight functions, x(Z). Maximizing over a larger set of weights expands
the investment opportunity set and produces a larger maximum Sharpe ratio.
With conditioning information, the optimal orthogonal portfolio’s weight
function is time varying, and we derive this portfolio weight for three cases. First,
with no risk-free asset in which case a fixed unconditional “zero-beta” rate g0 is
arbitrarily chosen. By varying the zero-beta rate, the solutions can describe any
point on the efficient-with-respect-to-Z frontier. Second, we consider
a conditionally time-varying risk-free asset whose return Rf ¼ Rf (Z ) is measureable
and thus known as part of the information set Z so that Var[Rf (Z )|Z ] ¼ 0, but which
is unconditionally risky in the sense that Var[Rf (Z )] > 0. Here, we again choose an
arbitrary zero-beta rate g0 to describe the frontier. In the third case, there exists an
unconditional risk-free asset with fixed return Rf ¼g0. In this case, the efficient-
with-respect-to-Z frontier becomes a line passing through g0 (at risk zero) and the
point representing the mean and standard deviation of a particular portfolio
strategy’s return.
984 W.E. Ferson and A.F. Siegel

We consider portfolios formed from the risky assets using weights xq ¼ xq(Z)
where the weights must sum to 1 (for all Z) when there is no risk-free asset. This
constraint is relaxed when there is a conditional or unconditional risk-free asset
(where the implicit weight in the risk-free asset is then set at 1 minus the sum of the
weights in the risky assets). When there is no risk-free asset, the portfolio return
0
is Rq,t+1 ¼ xq(Zt)Rt+1, and when there is a risk-free asset, the portfolio return is
0  
Rq, tþ1 ¼ Rf ðZ t Þ þ xq ðZt Þ Rtþ1  Rf ðZt Þ1 whether or not Rf (Z) is constant. In either
case, we denote the (unconditional) portfolio mean mq ¼ E(Rq) and variance
s2q ¼ Var(Rq). When there exists a conditional time-varying risk-free asset, the
portfolio takes advantage of the ability to adapt both the percentage invested in
risky assets and their portfolio weights in response to the information Z. This may
be interpreted as “market timing” and security selection, respectively. We define
the optimal orthogonal portfolio with respect to a given benchmark portfolio
P formed from the risky assets using (possibly) time-varying weights xp ¼ xp(Z).
Definition The most mispriced (or optimal orthogonal) portfolio, Rc, with respect
to the benchmark portfolio Rp and conditioning information Z, with portfolio
weight denoted xc(Z), uses the conditioning information to maximize a2c /s2c where
the unconditional variance of Rc is s2c , the unconditional mean is mc ¼ E(Rc), the
unconditional alpha of Rc with respect to Rp is ac ¼ mc  [g0 + (mp  g0)scp/s2p], the
zero-beta rate is g0, and the unconditional covariance is scp ¼ Cov(Rc, Rp).
Proposition 2 The unique most mispriced (or optimal orthogonal) portfolio Rc with
respect to a given benchmark portfolio Rp (with weights xp and expected return mp),
conditioning information Z and given zero-beta rate, g0, has the following portfolio
weight in each of the following cases. If there is no risk-free rate, then the weights
conditionally sum to 1 as defined by
( 0
! )
L1 L1 1 L
xc ðZ Þ ¼ A 0 þ ½ðc þ 1Þms þ b L  0 mðZÞ þ Bxp , (35.7)
1 L1 1 L1
Rf (Z) (which may be either constant or time varying, but if constant,4 then we must
choose g0 ¼ Rf along with any ms 6¼ Rf), then the solution is
   
xc ðZÞ ¼ A ðc þ 1Þms þ b  Rf Q mðZ Þ  Rf 1 þ Bxp , (35.8)
where
ðms  g0 Þ=s2s
A¼
 , (35.9)
ðms  g0 Þ=s2s  mp  g0 =s2p

4
Equation 35.11 cannot be used to determine ms when Rf (Z) is almost surely constant due to
division by zero, and, in this case, every choice ms 6¼ Rf uses the same (rescaled) portfolio of risky
assets Q mðZ Þ  Rf 1 in the formation of an efficient portfolio xs(Z).
35 Optimal Orthogonal Portfolios with Conditioning Information 985


mp  g0 =s2p
B¼
 , (35.10)
ðms  g0 Þ=s2s  mp  g0 =s2p

ms ¼ ða þ bg0 Þ=ðb þ cg0 Þ, (35.11)

s2s ¼ a þ 2bms þ cm2s , (35.12)

  
1


0 
Q ¼ Q ðZ Þ  E R  Rf 1 R  Rf 1 Z

1 (35.13)
  0
¼ mðZÞ  Rf 1 mðZÞ  Rf 1 þ S
= e ðZ Þ ,

n h 0  io1 h i1
L ¼ LðZ Þ  E RR Z
0
¼ mðZÞm ðZÞ þ S
= e ðZ Þ : (35.14)

The constants a, b, and c are defined in Appendix 1 in Theorem 1 (when there

exists a risk-free rate) and in Theorem 2 (when there is no risk-free rate).
Proof: See Appendix 1.

The term ms represents the unconditional expected return of the efficient-with-

respect-to-Z portfolio, Rs, that maximizes the squared unconditional Sharpe ratio in
Eq. 35.4 over all portfolio weight functions with respect to the given value of g0.
When there exists a risk-free rate Rf (Z) (that may be time varying because its value is
included in the information set Z at the beginning of the period), the conditional mean
variance boundary, the tangency intercept Rf (Z), and thus the location of the condi-
tionally mean variance efficient portfolio may vary over time as Rf (Z) varies. When
there is no risk-free rate, or when the risk-free rate is time varying, we use the
parameter g0 to determine a fixed location on the (curved) unconditionally efficient-
with-respect-to-Z boundary; however, when there is a fixed risk-free rate, this
boundary is a degenerate hyperbola, and every portfolio on the upper line is efficient.
We next show that the optimal orthogonal portfolio Rc can be formed by combining
the benchmark portfolio Rp with the efficient-with-respect-to-Z portfolio Rs, and from
this result, it then follows that Rp and Rc can be combined to produce the efficient-
with-respect-to-Z portfolio Rs.
Proposition 3 The most mispriced or optimal orthogonal portfolio Rc may be found
as a fixed linear combination of the benchmark portfolio Rp and the efficient-with-
respect-to-Z portfolio, Rs (that maximizes the squared Sharpe ratio for the given
zero-beta rate, g0), as follows:
986 W.E. Ferson and A.F. Siegel

  h
 i
ðms  g0 Þ=s2s Rs  mp  g0 =s2p Rp
Rc ¼   h
 i (35.15)
ðms  g0 Þ=s2s  mp  g0 =s2p

Or

s2p Rs  sps Rp
Rc ¼ ¼ ARs þ BRp with A þ B ¼ 1, (35.16)
s2p  sps

where we assume that Rs and Rp are not perfectly correlated and that s2p 6¼ ssp.
Proof: See Appendix 1.
Propositions 2 and 3 extend the concept of the active or optimal orthogonal
portfolio to the setting of efficiency with respect to conditioning information. Given
a benchmark portfolio Rp, its optimal orthogonal portfolio with respect to Z shows
how to tilt away from the benchmark weights to obtain efficiency with respect to Z.
The portfolio Rc has weights that depend on Z. Thus, the optimal tilt away from
a benchmark uses the manager’s information Z in a dynamic way.
Equation 35.16 shows how the optimal orthogonal portfolio Rc can be formed by
combining an efficient-with-respect-to-Z portfolio Rs with Rp. The portfolio Rc is
the regression error of Rs projected on Rp, normalized so that the weights sum to
1, as can be seen by solving Eq. 35.16 for Rs. Thus, the portfolio Rc is uncorrelated
with Rp.
We defined the most mispriced portfolio with conditioning information as
maximizing the squared alpha relative to the unconditional variance of Rc. Since
Rc is orthogonal to Rp, its residual variance in regression (Eq. 35.1) is the same as its
total variance. Thus, we can think of the optimal orthogonal portfolio as maximiz-
ing alpha given its residual variance among all orthogonal portfolios.
Given a benchmark portfolio with return Rp, the optimal orthogonal portfolio
with conditioning information is useful for active portfolio management. It might
seem natural for a manager with information Z to simply reinterpret the classical
analysis, where all the moments are the conditional moments given Z. This, in fact,
is the interpretation that much of the literature on active portfolio management has
used (e.g., Jorion 2003; Roll 1982). This approach produces conditionally mean
variance efficient portfolios given Z. However, as shown by Dybvig and Ross
(1985), a conditionally efficient portfolio is likely to be seen as inefficient from
the (unconditional) perspective of a client without access to the information Z.
The optimal orthogonal portfolio describes the active portfolio bets that deliver
optimal performance from the client’s perspective.
Let S2(R) be the maximum squared Sharpe ratio obtained by the efficient-with-
respect-to-Z portfolio and let Rc be the optimal orthogonal portfolio with respect to
Rp and information Z.
Proposition 4 Law of conservation of squared Sharpe ratios. For a given zero-beta
or risk-free rate, if S2s ¼ S2(R) is the maximum squared Sharpe ratio obtained by
35 Optimal Orthogonal Portfolios with Conditioning Information 987

portfolios x(Z) and Rc is the optimal orthogonal portfolio with respect to Rp and
information Z at that zero-beta rate, then S2s ¼ S2p + S2c , where S2i denotes the squared
Sharpe ratio of portfolio i.
Proof: See Appendix 1.
Proposition 4 shows that if we test the hypothesis that a portfolio is efficient with
respect to Z using versions of the test statistic in Eq. 35.3, as developed in Ferson
and Siegel (2009), then the role of the optimal orthogonal portfolio with informa-
tion Z is analogous to the role of the optimal orthogonal portfolio in the case of the
classical test statistics. The Sharpe ratio of the optimal orthogonal portfolio with
conditioning information indicates how far the tested benchmark portfolio is from
the efficient-with-respect-to-Z frontier. The squared Sharpe ratio of the optimal
orthogonal portfolio is the numerator of the test statistic. This numerator and thus
the test statistic is zero only if the tested portfolio is efficient in the sample, and it
grows larger as the tested portfolio is further from efficiency.

35.5 Empirical Examples

We present empirical examples using data on portfolios of common stocks, where

the firms are grouped according to conventional criteria. We use the returns of
common stocks sorted according to market capitalization and book-to-market
ratios, focusing on value-weighted decile portfolios of small capitalization stocks,
value stocks (high book/market), and growth stocks (low book/market), as provided
on Ken French’s website. We also include a long-term government bond return,
splicing the Ibbotson Associates 20-year US government bond return series for
1931–1971, with the CRSP greater than 120 month US government bond return
after 1971. The market portfolio, measured as the CRSP value-weighted stock
return index, is the benchmark or tested portfolio, Rp. The risk-free return is the
return from rolling over 1-month Treasury bills from CRSP. We use its sample
average, 3.8 % per year, as the fixed zero-beta rate in all of the examples. As
conditioning information in Z, its return is lagged by 1 year. All of the returns are
discretely compounded annual returns, and the sample period is 1931–2007.
The lagged instruments, Z, are the lagged Treasury return and the log of the
market price/dividend ratio at the end of the previous year. In calculating the
price/dividend ratio, the stock price is the real price of the S & P 500 Index, and
the dividend is the real dividends accruing to the index over the past year. These
data are from Robert Shiller’s website.
We treat the risk-free asset in three distinctly different ways to highlight the three
different versions of our solutions for the optimal orthogonal portfolios. In the first
case, the risk-free rate is assumed to be a fixed constant. Here, we do not include the
Treasury return as a lagged instrument, and we set g0 ¼ 3.8 % to be the
average Treasury return during the sample. The target mean ms of the efficient-
with-respect-to-Z portfolio is set equal to the sample mean return of the market
index in this case, which determines the amount of leverage the portfolio uses at
the fixed risk-free rate.
988 W.E. Ferson and A.F. Siegel

In the second case, no risk-free asset exists. Here, we again set g0 ¼ 3.8 % to pick
a point on the mean variance boundary, and we do not allow the portfolio to take
a position in a risk-free asset. We do allow the lagged Treasury return as condi-
tioning information in Z, which highlights the information in lagged Treasury
returns about the future risky asset returns. In the third case, there is
a conditionally risk-free return that is contained in Z. Here, we use the lagged
Treasury return in the conditioning information, and we allow the portfolio to trade
the subsequent Treasury return in addition to the other risky assets. The subsequent
Treasury return is not really known ex ante as the formula assumes, but its
correlation with the lagged return is 0.92 during our sample, and this allows us to
implement the time-varying risk-free rate example in a somewhat realistic way.
(In reality, there is no ex ante risk-free asset given the importance of inflation risk.)
In practical terms, this example highlights the effects of “market timing,” or
varying the amount of “cash” in the portfolio, in addition to varying the allocation
among the risky assets.
Table 35.1 summarizes the results. The rows show results for the benchmark
index (Market), three equity portfolios, and the government bond return. The
CAPM a refers to the intercept in the regression (1), of the portfolio returns in
excess of the Treasury bill returns, on the excess return of the market index. The
small stock portfolio has the largest alpha, at 3.95 % per year, while the growth
stock portfolio has a negative alpha. The symbol su refers to the standard deviation
of the regression residuals. The small stock portfolio has the largest su or
nonmarket risk, at more than 25 % per year.
The bottom four rows of Panel A summarize the optimal orthogonal portfo-
lios when the market index is the benchmark. The fixed-weight portfolio Rc uses
no conditioning information. Its alpha is larger than any of the separate assets, at
4.66 % per year, and its residual standard deviation is also relatively large, at
16.5 % per year. Since the portfolio is orthogonal to the market index, its
residual standard deviation is the same as its total standard deviation, or vola-
tility of return. The ratio of the alpha to the residual volatility is known as the
appraisal ratio (Treynor and Black 1973) or the information ratio (Grinold and
Kahn 1992). Optimal orthogonal portfolios try to maximize the square of this
ratio. The fixed-weight portfolio Rc delivers an information ratio of 0.282,
substantially larger than those of the small stock or value portfolios, at 0.155
and 0.177, respectively, and also larger than the bond portfolio, which has an
information ratio of 0.183 by virtue of its relatively small volatility.
Table 35.1 summarizes performance statistics for the optimal orthogonal port-
folios with conditioning information. There are three versions with (1) a fixed risk-
free rate, (2) no risk-free rate, and (3) a time-varying conditional risk-free rate. The
information ratios in all three cases are larger than those of the individual assets or
the fixed-weight active portfolio, which illustrates the potential value of using
conditioning information explicitly in a portfolio management context (see also,
Chiang 2009). The improvement over fixed weights is modest for the case with no
risk-free asset. However, the information ratio is about three times as large, at
0.955, in the example with a time-varying risk-free rate. This illustrates the
 35

Table 35.1 Optimal orthogonal portfolios

Panel A: summary statistics
Average active portfolio weights Avg [Xc(Z)]
Asset CAPM a su Information Ratio Fixed Rc Fixed Rf No Rf Varying Rf
Market index 0 0 0.000 1.260 0.179 0.210 0.308
Small stocks 3.95 25.5 0.155 0.353 0.262 0.133 0.097
Value stocks 3.14 17.8 0.177 0.345 0.137 0.079 0.142
Growth stocks 1.00 8.3 0.121 0.155 0.028 0.098 0.020
Bonds 1.66 9.1 0.183 1.410 0.918 0.900 0.432
Fixed Rc 4.66 16.5 0.282 1.000
Xc(Z) portfolios:
Fixed Rf 8.66 25.5 0.340
No Rf 3.12 10.5 0.297
Varying Rf 8.77 9.2 0.955
Panel B: squared Sharpe ratios
S2(Rp) S2(Rc) Sum
Fixed Rc 0.189 0.079 0.267
Optimal Orthogonal Portfolios with Conditioning Information

Xc(Z) portfolios:
Fixed Rf 0.189 0.115 0.304
No Rf 0.189 0.088 0.277
Varying Rf 0.182 0.909 1.191
Annual returns on portfolios of common stocks and long-term US government bonds cover the 1931–2007 period. The CAPM a refers to the intercept in the
regression Eq. 35.1, of the portfolio return in excess of a zero-beta parameter, on that of the broad value-weighted stock market index (market index). The
symbol su refers to the standard deviation of the regression residuals. The fixed portfolio, Rc, ignores the conditioning information. The alphas and residual
standard deviations are annual percentage units. The information ratio is the ratio a/ su
989
990 W.E. Ferson and A.F. Siegel

potential usefulness of interest rate information and the ability to hold “cash” as
a function of market conditions, in a portfolio management context.
The averages over time of the optimal orthogonal portfolios’ weights on the
risky assets are shown in the right-hand columns of Panel A. The weights are
normalized to sum to 1.0. The weights xc of the optimal orthogonal portfolios
combine with the benchmark (whose weights, xp, are 100 % in the market index)
to determine an efficient portfolio. The overall efficient portfolio weights xs
therefore vary over time and depend on how xc and xp are combined, which is
determined by the coefficients A and B in Eqs. 35.9 and 35.10. The estimated
weights are as follows. In the fixed risk-free rate case, xs ¼ 0.60 xc + 0.40 xp. In the
no risk-free rate case, xs ¼ 0.23 xc + 0.77 xp. In the time-varying risk-free rate
case, xs ¼ 0.44 xc + 0.56 xp. Thus, the efficient portfolio is formed as a convex
combination of the benchmark and the active portfolio, with reasonable weights
in each case.
The four right-hand columns of Panel A show that two of the active portfolios
take short positions in the market benchmark, indicating an optimal tilt away from
the market index. The fixed-weight portfolio Rc takes an extreme short
position of 126 %, while on average the portfolio using the conditioning
information but no risk-free asset takes a position of 21 % in the market index.
These short positions finance large long positions in the US government bond and
also long positions (in most cases) in small stocks, value stocks, and growth stocks.
It is interesting that all the portfolios tilt positively, although by small amounts, into
growth stocks even though growth stocks have negative CAPM alphas. This occurs
because of the correlations among the asset classes.
All the versions of the optimal orthogonal portfolio suggest strong tilts into
government bonds. The bond tilt is the most extreme for the fixed-weight solution,
at 141 %, and is relatively modest, at 43.2 %, for the portfolio assuming a time-
varying conditional risk-free rate. This makes sense, as that portfolio can hold
short-term Treasuries in addition to government bonds. The large weights in bonds
reflect various features of the data and the value of the zero-beta rate. With larger
values for the zero-beta rate, provided that the rate remains below the mean of the
global minimum-variance portfolio of risky assets, the optimal orthogonal portfo-
lios become more aggressive as the target expected return of the efficient-with-
respect-to-Z portfolio increases.
The squared Sharpe ratios in Panel B of Table 35.1 indicate how far the stock
market index is from efficiency. The squared Sharpe ratio for the market is 0.189,
measured relative to the fixed zero-beta rate of 3.8 %. The market portfolio’s
squared Sharpe ratio is slightly smaller, at 0.182, for returns measured in excess
of a time-varying risk-free rate. This is because the negative covariance between the
risk-free rate and stock returns increases the variance of the excess return. For the
fixed-weight orthogonal portfolio Rc, the squared Sharpe ratio is 0.079, and for
the portfolios using conditioning information, it varies between 0.088 and 0.909.
According to the law of conservation of squared Sharpe ratios in Proposition 4,
the sum of the index and optimal orthogonal portfolios’ squared Sharpe ratios is
the squared slope of the tangency from the given zero-beta rate to the relevant
35 Optimal Orthogonal Portfolios with Conditioning Information 991

Fig. 35.1 Optimal 1.5

orthogonal portfolio weights:
the no risk-free rate case Bonds
1.0

Portfolio Weight
Growth Small
0.5
Value

0.0

−0.5
SP500
−1.0
1930 1950 1970 1990 2010
Year

mean variance frontier. The sum is 0.267 when no conditioning information is

used and is 0.304–1.091 when the information is used. With a fixed risk-free rate,
we condition only on the lagged dividend yield, which is a relatively weak
predictor for stock returns (see Ferson et al. 2003). However, in the case with
no risk-free rate, the portfolio is not allowed to hold short-term Treasuries, which
substantially weakens the performance. In the time-varying risk-free rate case, the
portfolio strategy is allowed to “market time” by holding short-term Treasuries. In
the other two cases, we use the information in the lagged Treasury rate, which is
a relatively strong predictor, and the efficient-with-respect-to-Z boundary is far
above the mean variance boundary that ignores the conditioning information.
(Ferson and Siegel 2009, present an analysis of the statistical significance of
differences like these.)
Table 35.1 suggests that the portfolio weights of the fixed-weight Rc portfolio are
extreme and would not be realistic in practical portfolio management settings. This
reflects well-known issues with mean classical variance optimal solutions in practice
(e.g., Michaud 1989; Siegel and Woodgate 2007). To obtain practical portfolio
weights in the classical mean variance problem, it is generally necessary to constrain
the weights (e.g., Frost and Savarino 1988) or shrink them toward a benchmark e.g.,
Jorion (2003) or Jagannathan and Ma (2003). In this context, note that the optimal
orthogonal portfolios using Z take less extreme positions on average than the fixed-
weight solution, yet still are able to generate larger information ratios.
Figures 35.1 and 35.2 present time-series plots of the weights for the optimal
orthogonal portfolios using Z. Like the fixed-weight Rc case, the weights in the case
with a fixed risk-free rate assumption appear too volatile and noisy to be of practical
interest, likely reflecting the poor predictive ability based solely on the dividend
yield. We do not plot them here to save space. The weights in the other two examples
are shown in Figs. 35.1 and 35.2. They generally vary relatively smoothly over time,
suggesting that they would not involve prohibitive trading costs in practice.
Figure 35.1 depicts the weights for the optimal orthogonal portfolio in the case
with no risk-free asset. The market index weights are negative through much of the
992 W.E. Ferson and A.F. Siegel

1.0
Bonds
0.8

0.6
Portfolio Weight
0.4 Small

0.2

0.0

−0.2
Growth Value SP500
−0.4

−0.6
1930 1950 1970 1990 2010
Year

Fig. 35.2 Optimal orthogonal portfolio weights: the time-varying risk-free rate case

sample and near 25 %, indicating that the overall efficient-with-respect-to-Z

portfolio keeps about 71 % of its money in the market index for much of the
sample, computed as (0.23)(0.25) + 0.77 using the estimated weights for the no
risk-free rate case, xs ¼ 0.23 xc + 0.77 xp, as reported earlier. Starting in the
mid-1990s, the market weights decrease to around 50 % for the optimal orthog-
onal portfolio weight and 66 % for the overall efficient-with-respect-to-Z portfolio.
The weight of this optimal orthogonal portfolio in small stocks is positive for much
of the sample, but turns slightly negative in the early 1970s, then positive again in
the early 1990s. The strategy shorts value stocks in the 1930s and 1940s, but holds
positive positions through most of the rest of the sample. The government bond gets
the largest weight, starting near 80 % and growing sharply in the 1990s to near
100 % in the latter parts of the sample.
Figure 35.2 depicts the weights for the optimal orthogonal portfolio in the case
with a time-varying, conditional risk-free rate. This strategy keeps positive weights
in the market index until 1999, then it holds small short positions for most of the rest
of the sample. The weight in small stocks is positive for much of the sample, turning
slightly negative in the early 1970s and then positive again in the early 1990s. The
strategy holds value stocks long until the early 2000s and shorts growth stocks
during much of the 1980s. The government bond again gets the largest weight,
starting at 32 % and growing to almost 70 % at the end of the sample.

35.6 Conclusions

This chapter derives, characterizes, and illustrates optimal orthogonal portfolios in

the presence of conditioning information in the form of a set of lagged instruments.
Optimal orthogonal portfolios combine with a benchmark portfolio to form mean
variance efficient portfolios. We generalize previously published solutions for
35 Optimal Orthogonal Portfolios with Conditioning Information 993

optimal orthogonal portfolios with information to include the case of a time-

varying, but conditionally known, risk-free asset return.
In a portfolio management context, it is reasonable to assume that the portfolio
manager has more information about asset returns than the client. The manager
observes information about future returns and, by conditioning on this information,
can expand the investor’s opportunity set. Starting from a given benchmark port-
folio, the optimal orthogonal portfolio with conditioning information describes the
active portfolio bets that, when combined with the benchmark, deliver mean
variance optimal performance from the uninformed client’s perspective.
We present empirical examples using a broad stock market index as the bench-
mark and portfolios featuring small capitalization, value and growth stocks, and
a long-term US government bond return. We examine three versions of the optimal
orthogonal portfolio with (1) a fixed risk-free rate, (2) no risk-free rate, and (3) a
time-varying conditional risk-free rate. The optimal orthogonal portfolios with
conditioning information have larger information ratios than the orthogonal port-
folio that does not use the conditioning information. At the same time, they take less
extreme positions than the fixed-weight solution.
From an asset pricing perspective, a standard stock market index is far from
efficient when portfolios trade based on lagged interest rates and dividend yields.
From an active portfolio management perspective, the example shows that a strong
tilt toward bonds improves the efficiency of equity portfolios. Our results should be
useful in future asset pricing and portfolio management applications.

Appendix 1: Theorems and Proofs

Efficient Portfolio Solutions

Portfolio weights for efficient portfolios in the presence of conditioning information

are derived by Ferson and Siegel (2001). They consider the case with no risk-free
asset and the case with a fixed risk-free asset whose return is constant over time.
In Theorem 1, we generalize to consider the case with a risk-free asset whose return
is known at the beginning of the period, and thus is included in the information Z,
and may vary over time. We then, in Theorem 2, reproduce the case with no risk-
free asset from Ferson and Siegel (2001) for future reference.
Consider N risky assets with returns R. In N  1 column vector notation, we have

R ¼ mðZ Þ þ e:

The noise term e is assumed to have conditional mean zero given Z and
nonsingular conditional covariance matrix S = e ðZ Þ. The conditional expected return
vector is m(Z) ¼ E(R|Z). Let the 1  N row vector x0 (Z) ¼ (x1(Z), . . . , xN(Z)) denote
the portfolio share invested in each of the N risky assets, investing (or borrowing) at
0
the risk-free rate the amount 1  x ðZÞ1, where 1  (1 ,. . .,1)0 denotes the column
vector of ones. We allow for a conditional risk-free asset returning Rf ¼ Rf (Z).
994 W.E. Ferson and A.F. Siegel

0

The return on the portfolio is Rs ¼ Rf þ x ðZÞ R  Rf 1 , with unconditional
expectation and variance as follows:

 n 0  o
ms ¼ E Rf þ E x ðZ Þ mðZ Þ  Rf 1 ,

 

s2s ¼ E R2s  m2s ¼ E E R2s jZÞ  m2s ,
  h 0 i n 0  o
s2s ¼ E R2f þ E x ðZÞQ1 xðZÞ þ 2E Rf x ðZÞ mðZÞ  Rf 1  m2s , (35.17)

where we have defined the N  N matrix

n h

0  io1
Q ¼ QðZÞ  E R  Rf 1 R  Rf 1 Z
n  0 o1
¼ mðZÞ  Rf 1 mðZ Þ  Rf 1 þ S
= e ðZ Þ :

Also, define the constants:

n 0  o
zE mðZ Þ  Rf 1 Q mðZ Þ  Rf 1 ,
n  0  o
’  E Rf mðZÞ  Rf 1 Q mðZ Þ  Rf 1 ,

and
n  0  o
c  E R2f mðZÞ  Rf 1 Q mðZÞ  Rf 1 :

Theorem 1 Given a target unconditional expected return ms, N risky assets, instru-
ments Z, and a conditional risk-free asset with rate Rf ¼ Rf (Z) that may vary over
time, the unique portfolio having minimum unconditional variance is determined
by the weights

 
m s  E Rf þ ’  
xs ðZ Þ ¼  Rf Q mðZ Þ  Rf 1
z (35.18)
   
¼ ðc þ 1Þms þ b  Rf Q mðZ Þ  Rf 1 ,

and the optimal portfolio variance is

s2s ¼ a þ 2bms þ cm2s

  E R ’ 2
½ ð fÞ  ’EðRf Þ
where a ¼ E R2f þ z  c, b ¼ z , and c ¼ 1z  1. When the risk-free
asset return is constant, then these formulas simplify to Theorem 2 of Ferson and
Siegel (2001) with
35 Optimal Orthogonal Portfolios with Conditioning Information 995

ms  Rf  
xs ðZ Þ ¼ Q mðZ Þ  Rf 1
z

2
and with optimal portfolio variance s2s ¼ 1z
z ms  Rf .
Proof Our objective is to minimize, over the choice of

xs(Z), the
 portfolio variance
0
Var(Rs) subject to E(Rs) ¼ ms, where Rs ¼ Rf þ x ðZ Þ R  Rf 1 and the variance is
given by Eq. 35.17. We form the Lagrangian:
 0   0  
L½xðZÞ ¼ E x ðZ ÞQ1 xðZ Þ þ 2E Rf x ðZÞ mðZÞ  Rf 1
 0  
þ2lE ms  Rf  x ðZ Þ mðZ Þ  Rf 1
and proceed using a perturbation argument. Let q(Z) ¼ x(Z) + dy(Z), where x(Z) is
the conjectured optimal solution, y(Z) is any regular function of Z, and d is a scalar.
Optimality of x(Z) follows when the partial derivative of L[q(Z)] with respect to d is
identically zero when evaluated at d ¼ 0. Thus,
 0 
  
0 ¼ E y ðZÞ Q1 xðZÞ þ Rf  l mðZÞ  Rf 1

 
for all functions y(Z), which implies that Q1 xðZ Þ þ Rf  l mðZ Þ  Rf 1 ¼ 0
almost surely in Z. Solve this expression for x(Z) to obtain Eq. 35.18, where the
Lagrange multiplier l is evaluated by solving for the target mean, ms. The expression
for the optimal portfolio variance follows by substituting the optimal weight function
into Eq. 35.17. Formulas for fixed Rf then follow directly. QED.
When the risk-free asset’s return is time varying and contained in the informa-
tion set Z at the beginning of the portfolio formation period, the conditional mean
variance efficient boundary varies over time with the value of Rf (Z) along with the
conditional asset means and covariances . In this case, a zero-beta parameter, g0,
may be chosen to fix a point on the unconditionally efficient-with-respect-to-
Z boundary. The choice of the zero-beta parameter corresponds to the choice of
a target unconditional expected return ms. For a given value of g0, the target mean
maximizes the squared Sharpe ratio (ms  g0)2/s2s along the mean variance bound-
ary, which implies ms ¼  (a + bg0)/(b + cg0).
When there is a risk-free asset that is constant over time, the unconditionally
efficient-with-respect-to-Z boundary is linear (a degenerate hyperbola) and reaches
the risk-free asset at zero risk. In this case, we use g0 ¼ Rf and can obtain any ms
larger or smaller than Rf, levering the efficient portfolio up or down with positions
in the risk-free asset.
When there is no risk-free asset, we define portfolio s by letting x0 ¼ x0 (Z) ¼ [x1(Z),
. . ., xN(Z)] denote the shares invested in each of the N risky assets, with the
constraint that the weights sum to 1.0 almost surely in Z. The return on this
portfolio, Rs ¼ x0 (Z)R, has expectation and variance as follows:
h 0 i
ms ¼ E x ðZ ÞmðZ Þ ,
n 0 o
s2s ¼ E x ðZÞL1 xðZÞ  m2s ,
996 W.E. Ferson and A.F. Siegel

where we have defined the N  N matrix

n h 0  io1 h 1
L ¼ LðZÞ  E RR Z
0
¼ mðZ Þm ðZÞ þ S
= e ðZÞ :

Also, define the constants:


1
d1 ¼ E 0 ,
1 L1

0
!
1 LmðZÞ
d2 ¼ E ,
10 L1

and
" 0
! #
0 L1 1 L
d3 ¼ E m ð Z Þ L  0 mðZÞ :
1 L1

Theorem 2 (Ferson and Siegel 2001, Theorem 3) Given N risky assets and no risk-
free asset, the unique portfolio having minimum unconditional variance and uncon-
ditional expected return ms is determined by the weights:

0 0
!
0 1L m  d2 0 L1 1 L
xs ðZ Þ ¼ 0 þ s m ðZ Þ L  0
1 L1 d 3 1 L1
0 0
!
1L 0 L1 1 L
¼ 0 þ ½ðc þ 1Þms þ bm ðZ Þ L  0 , (35.19)
1 L1 1 L1

and the optimal portfolio variance is

s2s ¼ a þ 2bms þ cm2s ,

where a ¼ d1 + d22/d3, b ¼  d2/d3, and c ¼ (1  d3)/d3.

The efficient-with-respect-to-Z boundary is formed by varying the value
of the target mean return ms in Eq. 35.19. Note that the second term on
the right-hand side of Eq. 35.19 is proportional to the vector of weights of
an excess return, or zero net investment portfolio (post multiplying that term
by a vector of ones implies that the weights sum to zero). The first
term in Eq. 35.19 is the weight of the global minimum conditional second
moment portfolio. Thus, Eq. 35.19 illustrates two-fund separation: any
35 Optimal Orthogonal Portfolios with Conditioning Information 997

efficient-with-respect-to-Z portfolio can be found as a combination of the

global minimum conditional second moment portfolio and some weight on
the unconditionally efficient excess return described by the second term.

Proofs of Propositions

Proof of Propositions 2 and 3

We begin with Proposition 3. To see that Eqs. 35.15 and 35.16 are equivalent, use
s
the fact that efficiency of Rs implies that mp ¼ g0 þ sps2 ðms  g0 Þ and thus mssg
2
0
s s
m g
¼ psps 0 . Next, given any portfolio Rq 6¼ Rc, we will show that a2q/s2q < a2c /s2c .
Beginning with Eq. 35.16, we compute:

s2p s2s  s2ps

scs ¼
s2p  sps

and
 
s2p s2p s2s  s2ps
s2c ¼  2 :
s2p  sps
s
The efficiency of Rs implies that mc ¼ g0 þ sscs2 ðms  g0 Þ, mp ¼ g0 þ sps2 ðms  g0 Þ,
s s s
and mq ¼ g0 þ sq2s ðms  g0 Þ. Substituting these expressions and using the fact that
s
scp ¼ 0, which follows from (35.16), we compute:


 2

 2
a2c a2q mc  go þ mp  go scp =sp 2 mq  go þ mp  go sqp =sp 2
 ¼ 
s2c s2q s2c s2q
0  2 1
s 2 sqs s2p  sps sqp
B C
¼ ðms  g0 Þ2 @ 4cs2  A
ss sc s4s s4p s2q

0  1
s2p s2s  s2ps s2qs s4p þ s2ps s2qp  2sqs s2p sps sqp
¼ ðms  g0 Þ 2 @  A
s4s s2p s4s s4p s2q

ðms  g0 Þ2  4 2 2 
¼ sp ss sq  s2p s2q s2ps  s2qs s4p  s2ps s2qp þ 2s2p sqs sps sqp :
ss sp sq
4 4 2

We now use the fact that s2sp < s2s s2p (because, by assumption, Rp and Rs are not
perfectly correlated) to see that
998 W.E. Ferson and A.F. Siegel

a2c a2q ðms  g0 Þ2 h 4 2 2   i

 s s s  s s s
2 2 2
 s s
4 2
 s 2
s s
2 2
þ 2s 2
s s s
p qs qp ps
s2c s2q s4p s2q s4s p q s p q sp p qs qp s p

ðm  g Þ2  
¼ s2 2 04 s2p s2q s2s  s2q s2sp  s2p s2qs  s2qp s2s þ 2sqs sqp sps 0
sp sq ss

where the final inequality follows from recognizing that the variance-covariance
terms in parentheses are equal to the determinant of the (necessarily nonnegative
definite) covariance matrix of (Rp,Rq,Rs). This establishes the maximal property of Rc.
To show uniqueness, note further that the inequality will be strict (and we will have
a2c/s2c  a2q/s2q > 0) unless we have both of the following conditions corresponding to
the two inequalities in the final calculation: (1) sqp ¼ 0 so that Rq and Rp are
orthogonal, and (2) the covariance matrix of (Rp,Rq,Rs) is singular so that Rq is
a linear combination of Rp and Rs. However, there is only one portfolio orthogonal
to Rp that can be formed as a linear combination lRp + (1  l)Rs, and this solution
is Rc. This establishes Proposition 3, which holds in the case of both Theorem 1 and
Theorem 2, that is, whether or not there is a conditionally risk-free asset.
The expressions in Proposition 2 for the optimal weights, xc(Z), follow from
substituting portfolio weights from Theorem 1 (if there exists a conditional risk-free
asset that may be time varying) or Theorem 2 (otherwise) into Eq. 35.15 and noting
that the constants A and B represent the combining portfolio weights implied by
(35.15) as Rc ¼ ARs + BRp. Substituting, we see that Eq. 35.15 implies that the
portfolio weight function Axs(Z) + Bxp(Z) generates returns Rc, completing the
proof of Proposition 2. QED.

Proof of Proposition 4

Let Rs denote the efficient-with-respect-to-Z portfolio corresponding to zero-beta

s
rate g0. The efficiency of Rs implies that we may substitute mp  g0 ¼ ðms  g0 Þ sps2
scs s
and mc  g0 ¼ ðms  g0 Þ s2 to find
s

2 !
mp  g0 ðmc  g0 Þ2 ðms  g0 Þ2 s2ps s2cs
S2p þ S2c ¼ þ ¼ þ 2 :
s2p s2c s4s s2p sc

Next, substituting for scs and s2c from the above expressions, we verify that this
expression reduces to S2s . QED.

Appendix 2: Methodology

We estimate the conditional mean functions, m(Z), by ordinary least squares

regressions of the returns on the lagged values of the conditioning variables. On
the assumption that the conditional mean returns are linear functions of Z, these
35 Optimal Orthogonal Portfolios with Conditioning Information 999

are the optimal generalized method of moments (GMM, see Hansen 1982)
estimators. The covariance matrix of the residuals is used as the estimate of
S
= e ðZÞ , which is assumed to be constant. These are the maximum likelihood
estimates (MLE) under joint normality of (R,Z). In general, the conditional
covariance matrix of the returns given Z will be time varying as a function
of Z, as in conditional heteroskedasticity. Ferson and Siegel (2003) model
conditional heteroskedasticity in alternative ways and find using parametric
bootstrap simulations that this increases the tendency of the efficient-with-
respect-to-Z portfolio weights to behave conservatively in the face of extreme
realizations of Z.
The optimal orthogonal portfolio weights in Table 35.1 and Fig. 35.1 are
estimated from Eqs. 35.7 to 35.8 in the text where, in the time-varying risk-free
rate case, the Treasury bill return is assumed to be conditionally risk-free in
Eq. 35.8. The benchmark portfolio xp is a vector with a 1.0 in the place of the
market index and zeros elsewhere. The matrix Q is estimated by using the MLE
estimates of m(Z) and S= e ðZ Þ in the function given by Eq. 35.13. The parameter mp is
estimated as the sample excess return on the market index, and g0 is the sample
mean of the Treasury return, 3.8 %.

The Parametric Bootstrap

The parametric bootstrap is a special case of the simple, or nonparametric, boot-

strap, itself an example of a resampling scheme. Introduced by Efron (1979), the
bootstrap is useful when we wish to conduct statistical inferences, but when we
either don’t have an analytical formula for the sampling variation of a statistic,
don’t wish to assume normality or some other convenient distribution that allows
for an analytical formula or have a sample too small to trust asymptotic distribution
theory. The basic idea is to build a sampling distribution by resampling from the
data at hand. In the simplest example, we have some statistic that we have estimated
from a sample, and we want to know its sampling distribution. We resample from
the original data, randomly with replacement, to generate an artificial sample of the
same size, and we compute the statistic on the artificial sample. Repeating this
many times, the histogram of the statistics computed on the artificial samples is an
estimate of the sampling distribution for the original statistic. This distribution can
be used to estimate standard errors, confidence intervals, etc. We can think of the
bootstrap samples as being related to the original sample as the original sample is to
the population. There are many variations on the bootstrap, and a good overview is
provided by Efron and Tibshirani (1993).
In the simple, or nonparametric, bootstrap, no assumptions are made about the
form of the distribution. It is assumed, however, that the sample accurately reflects
the underlying population distribution, and this is critical for reliable inferences.
For example, suppose that the true distribution was a uniform on [0, M]. In a sample
drawn from this distribution, the maximum value is likely to be smaller than M, so
that the bootstrap will likely understate the true variability of the data. This problem
1000 W.E. Ferson and A.F. Siegel

is obviously worse if the original sample has fewer observations. If the data are
contaminated with measurement errors, in contrast, the extent of the true variability
can be overstated. Even with large sample sizes, the bootstrap can be unreliable. For
example, if the true distribution has infinite variance, the bootstrap distribution for
the sample mean is inconsistent (Athreya 1987).
With a parametric bootstrap, we can sometimes do better than with
a nonparametric bootstrap, where “do better” means, for example, obtain more
accurate confidence intervals (e.g., Andrews et al. 2006). The idea of the parametric
bootstrap is to use some of the parametric structure of the data. This might be as
simple as assuming the form of the probability distribution. For example, assuming
that the data are independent and normally distributed, we can generate artificial
samples from a normal distribution using the sample mean and variance as the
parameters. This is not exactly the right thing to do, because we should be sampling
from a population with the true parameter values, not their estimated values. But, if
the estimates of the mean and variance are good enough, we should be able to
obtain reliable inferences.
To illustrate and further suggest the flexibility of the parametric bootstrap,
consider an example, similar to the setting in our paper, where we have
a regression of stock returns on a vector of lagged instruments, Z, which are highly
persistent over time. Obviously, sampling from the Z’s randomly with replacement
would destroy their strong time-series dependence. Time-series dependence can be
accommodated in a nonparametric way by using a block bootstrap. Here, we sample
randomly a block of consecutive observations, where the block length is set to
capture the extent of memory in the data.
In order to capture the time-series dependence of the lagged Z in a parametric
bootstrap, we can model the lagged instruments as vector AR(1), for example,
retaining the estimator of the AR(1) coefficient and the model residual, which we
call the shocks, Uz. Regressing the future stock returns on the lagged instruments,
we retain the regression coefficient and the residuals, which we call the shocks, Ur.
We generate a sample of artificial data, with the same length as the original sample,
as follows. We concatenate the shocks as v ¼ (Uz,Ur). Resampling rows from v,
randomly with replacement, retains the covariances between the stock return
shocks and the instrument Z shocks. This can be important for capturing features
like the lagged stochastic regressor bias described by Stambaugh (1999). Drawing
an initial row from v, we take the Uz shock and add it to the initial value of the
Z (perhaps, drawn from the unconditional sample distribution) to produce the first
lagged instrument vector, Zt–1. We draw another row from v and construct the first
observation of the remaining data as follows. The stocks’ returns are formed by
adding the Ur shock to bzt–1, where b is the “true” regression coefficient that defines
the conditional expected return. The contemporaneous values of the Zs are formed
by multiplying the VAR coefficient by Zt–1 and adding the shock, Uz. The next
observation is generated by taking the previous contemporaneous value of the Z as
Zt–1 and repeating the process. In this way, the Z values are built up recursively,
which captures their strong serial correlation.
35 Optimal Orthogonal Portfolios with Conditioning Information 1001

References
Abhyankar, A., Basu, D., & Streme, A. (2007). Portfolio efficiency and discount factor bounds
with conditioning information: An empirical study. Journal of Banking and Finance, 31,
419–437.
Andrews, D. W. K., Lieberman, O., & Marmer, V. (2006). Higher order improvements of the
parametric bootstrap for long-memory gaussian processes. Journal of Econometrics, 133,
673–702.
Athreya, K. B. (1987). Bootstrap of the mean in the infinite variance case. Annals of Statistics, 15,
724–731.
Bekaert, G., & Liu, J. (2004). Conditioning information and variance bounds on pricing kernels.
Review of Financial Studies, 17, 339–378.
Campbell, J., Lo, A., & MacKinlay, A. C. (1987). The econometrics of financial markets.
Princeton: Princeton University Press.
Chiang, E. (2012). Modern portfolio management with conditioning information. Working paper.
University of North Carolina, Charlotte.
Dybvig, P., & Ross, S. (1985). Differential information and performance measurement using
a security market line. Journal of Finance, 40, 383–399.
Efron, B. (1979). Bootstrap methods: Another looks at the jackknife. Annals of Statistics, 7, 1–26.
Efron, B., & Tibshirani, R. J. (1993). Introduction to the bootstrap. New York: Chapman and Hall.
Fama, E. F. (1973). A note on the market model and the two-parameter model. Journal of Finance,
28, 1181–1185.
Ferson, W. E., & Siegel, A. F. (2001). The efficient use of conditioning information in portfolios.
Journal of Finance, 56, 967–982.
Ferson, W. E., & Siegel, A. F. (2003). Stochastic discount factor bounds with conditioning
information. Review of Financial Studies, 16, 567–595.
Ferson, W. E., & Siegel, A. F. (2009). Testing portfolio efficiency with conditioning information.
Review of Financial Studies, 22, 2735–2758.
Ferson, W. E., Sarkissian, S., & Simin, T. (2003). Spurious regressions in financial economics?
Journal of Finance, 58, 1393–1414.
Ferson, W. E., Siegel, A. F., & Xu, P. (2006). Mimicking portfolios with conditioning information.
Journal of Financial and Quantitative Analysis, 41, 607–635.
Frost, P. A., & Savarino, J. E. (1988). For better performance: Constrain portfolio weights. Journal
of Portfolio Management, 15, 29–34.
Gibbons, M. R., Ross, S. A., & Shanken, J. (1989). A test of the efficiency of a given portfolio.
Econometrica, 57, 1121–1152.
Grinold, R. C., & Kahn, R. N. (1992). Active portfolio management (2nd ed.). New York: McGraw
Hill.
Hansen, Lars P. (1982). Large Sample Properties of Generalized Method of Moments Estimators.
Econometrica, 50, 1029–1054.
Hansen, L. P., & Richard, S. F. (1987). The role of conditioning information in deducing testable
restrictions implied by dynamic asset pricing models. Econometrica, 55(3), 587–613.
Jagannathan, R., & Ma, T. (2003). Risk reduction in large portfolios: Why imposing the wrong
constraints helps. Journal of Finance, 58, 1651–1684.
Jobson, J. D., & Korkie, B. (1982). Potential performance and tests of portfolio efficiency. Journal
of Financial Economics, 10, 433–466.
Jorion, P. (2003). Portfolio optimization with tracking error constraints. Financial Analysts
Journal, 59, 70–82.
Kandel, S. A. (1984). The likelihood ratio test statistic of mean variance efficiency without
a riskless asset. Journal of Financial Economics, 13, 575–592.
Kandel, S. A., & Stambaugh, R. F. (1987). On correlations and inferences about mean-variance
efficiency. Journal of Financial Economics, 18, 61–90.
1002 W.E. Ferson and A.F. Siegel

Kandel, S. A., & Stambaugh, R. F. (1989). A mean-variance framework for tests of asset pricing
models. Review of Financial Studies, 2, 125–156.
MacKinlay, A. C. (1995). Multifactor models do not explain deviations from the CAPM. Journal
of Financial Economics, 38, 3–28.
Markowitz, H. M. (1952). Portfolio selection. Journal of Finance, 7, 77–91.
Michaud, R. O. (1989). The Markowitz optimization enigma: Is ‘optimized’ optimal? Financial
Analysts Journal, 45, 31–42.
Roll, R. R. (1977). A critique of the asset pricing theory’s tests – part 1: On past and potential
testability of the theory. Journal of Financial Economics, 4, 129–176.
Roll, R. R. (1980). Orthogonal portfolios. Journal of Financial and Quantitative Analysis, 15,
1005–1023.
Roll, R. R. (1982). A mean variance analysis of tracking error. Journal of Portfolio Management,
18, 13–22.
Roll, R. R. (1985). A note on the geometry of Shanken’s CSR T2 test for mean/variance efficiency.
Journal of Financial Economics, 14, 349–357.
Shanken, J. (1987). Multivariate proxies and asset pricing relations: Living with the Roll critique.
Journal of Financial Economics, 18, 91–110.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance, 19, 425–442.
Siegel, A. F., & Woodgate, A. (2007). Performance of portfolios optimized with estimation error.
Management Science, 53, 1005–1015.
Stambaugh, R. F. (1999). Predictive regressions. Journal of Financial Economics, 54, 375–421.
Treynor, J. L., & Black, F. (1973). How to use security analysis to improve portfolio selection.
Journal of Business, 46, 66–86.
Multifactor, Multi-indicator Approach
to Asset Pricing: Method and 36
Empirical Evidence

Cheng-Few Lee, K. C. John Wei, and Hong-Yi Chen

Contents
36.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004
36.2 The MIMIC Model and the Tests of CAPM and APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005
36.2.1 The MIMC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005
36.2.2 The Testing Model of CAPM by the MIMIC Approach . . . . . . . . . . . . . . . . . . . . 1007
36.2.3 The Testing Model of APT by the MIMIC Approach . . . . . . . . . . . . . . . . . . . . . . . 1008
36.3 Test of CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1008
36.4 Tests of APT by MIMIC Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
36.4.1 Macroeconomic Variables as the Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
36.4.2 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
36.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1020
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1022

Abstract
This paper uses a multifactor, multi-indicator approach to test the capital asset
pricing model (CAPM) and the arbitrage pricing theory (APT). This approach is
able to solve the measuring problem in the market portfolio in testing CAPM;
and it is also able to directly test APT by linking the common factors to the
macroeconomic indicators. Our results from testing CAPM support

C.-F. Lee (*)

Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu; lee@business.rutgers.edu
K.C.J. Wei
Hong Kong University of Science and Technology, Kowloon, Hong Kong
e-mail: johnwei@ust.hk
H.-Y. Chen
Department of Finance, National Central University, Taoyuan, Taiwan
e-mail: fnhchen@ncu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1003
DOI 10.1007/978-1-4614-7750-1_36,
# Springer Science+Business Media New York 2015
1004 C.-F. Lee et al.

Stambough’s (Journal of Financial Economics, 10, 237–268, 1982) argument

that the inference about the tests of CAPM is insensitive to alternative market
indexes.
We propose a MIMIC approach to test CAPM and APT. The beta estimated
from the MIMIC model by allowing measurement error on the market portfolio
does not significantly improve the OLS beta, while the MLE estimator does
a better job than the OLS and GLS estimators in the cross-sectional regressions
because the MLE estimator takes care of the measurement error in beta. There-
fore, the measurement error problem on beta is more serious than that on the
market portfolio.

Keywords
Capital asset pricing model, CAPM • Arbitrage pricing theory • Multifactor
multi-indicator approach • MIMIC • Measurement error • LISREL approach •
Ordinary least square, OLS • General least square, GLS • Maximum-likelihood
estimation, MLE

36.1 Introduction

Roll (1977) has shown that the capital asset pricing model (CAPM) can never be
tested unless the market portfolio is capable of being measured and identified.
However, the market portfolio is actually unobservable. Stated differently, since
the market portfolio is subject to measurement error, Sharpe (1964), Lintner
(1965), and Mossin’s (1966) type of CAPM can never be tested directly. In contrast,
the test of Ross’s (1976, 1977) arbitrage pricing theory (APT) does not rely upon the
identifications of the market portfolio or the true factors. Nevertheless, Shanken
(1982) argues that Ross’s contention that APT is inherently more easily tested is
questionable. If we can directly link these unobservable factors to some observable
indicators, the Shanken criticism of the test of APT can be avoided or reduced.
Fortunately, a multiple indicators and multiple causes (MIMIC) model, pro-
posed by Zellner (1970), Goldberger (1972a, b), Joreskog and Goldberger (1975),
and others, is an attractive methodology in dealing with the problem of
unobservable variables. Goldberger (1972b) conceptually described that structural
equation model is a combination of factor analysis and econometrics model.
Goldberger (1972b) and Joreskog and Goldberger (1975) develop a structural
equation model with multiple indicators and multiple causes of a single latent
variable, MIMIC model, and obtain maximum likelihood estimates of parameters.
The MIMIC model displays a mixture of econometric and factor analysis themes.
This concept has successfully been used to test economic models, such as the
structure of price expectation and inflationary expectation (see Turnovsky 1970;
Lahiri 1976). In addition, structural equation model has also been used in finance-
related studies. Titman and Wessels (1988), Chang et al. (2009), and Yang
et al. (2009) apply structural equation models (e.g., LISREL approach and
MIMIC model) in determining capital structure decision. Maddala and
36 Multifactor, Multi-indicator Approach to Asset Pricing 1005

Nimalendran (1996) use structural equation model to examine the effect of earnings
surprises on stock prices, trading volumes, and bid-ask spread. However, the
structural equation model, especially MIMIC model, has not been used in capital
asset pricing determination.
The purpose of this paper is twofold: (i) to use the MIMIC model to reexamine
CAPM and (ii) to use the MIMIC model to investigate the relationship between the
factors in APT and the macroeconomic indicators directly. APT is attractive to both
academicians and practitioners, because the model allows more than one factor.
However, to date the practical applications of APT are still limited since previous
studies in testing the model do not directly link the factors to the indicators. If the
linkage between the factors and the indicators can be derived, practical applications
will be much improved.
The outline of this paper is as follows. In Sect. 36.2, the MIMIC model is
reviewed and CAPM and APT in terms of the MIMIC model are demonstrated.
Section 36.3 shows how MIMIC can be used to test CAPM, and Sect. 36.4
investigates the MIMIC applied to test APT. Finally, a brief summary is contained
in Sect. 36.5.

36.2 The MIMIC Model and the Tests of CAPM and APT

36.2.1 The MIMC Model

Suppose that a system has k unobservable latent variables z ¼ (z1, . . .,zk)0 , p observable
exogenous indicators x ¼ (x1,. . ., xp)0 , and m observable endogenous variables
y ¼ (y1,. . ., ym)0 .1 The specification of this extended MIMIC model of Jöreskog and
Goldberger (1975) is as follows. The latent factors z are linearly determined, subject to
disturbances e ¼ (e1, . . .,ek)0 , by observable exogenous indicators x:2

z ¼ ax þ e, (36.1)

where
2 3
a11 ,   , a1p
6    7
6 7
a¼6
6    77 is a k  p matrix:
4    5
ak1 ,   , akp

1
Fogler et al. (1981) and Chen et al. (1983) indirectly link the factors extracted from the APT to
economic indicators. Jöreskog and Goldberger (1975) have shown that this kind of indirect
estimation method is not as efficient as the direct estimation method to be explored in this section.
2
Here, we use different terminologies in defining the factors and indicators compared with those
used in traditional MIMIC model.
1006 C.-F. Lee et al.

Fig. 36.1 Multiple causes

and indicators of
unobservable variables. This
figure shows the path diagram
illustrating MIMIC model of
the structural equation
system. In this figure,
observable variables
X1, . . . ,XP are causes of the
latent variables Z1, . . . ,Zk,
while Y1, . . . ,Ym are
indicators of Z1, . . . ,Zk

In addition, the latent factors z linearly determine the components of endogenous

variables y subject to disturbances u ¼ (u1, . . . ,um)0 :
y ¼ bz þ u, (36.2)
where
2 3
b11 ,   , b1k
6    7
6 7
b¼6
6    77 is an m  k matrix:
4    5
bm1 ,   , bmk
The disturbances are assumed to be mutually independent and normally
disturbed with mean zero, namely, e  N(0,S), u  N(0,y2), where S ¼ diag(s21 ,
. . .,s2k ) and y2 ¼ diag(y21 , . . .,y2m ). For convenience, all variables are taken to have
mean zero. The system of Eqs. 36.1 and 36.2 are shown in Fig. 36.1.
Solving the equation systems of Eqs. 36.1 and 36.2, we have the following
reduced form connecting the observable variables:

y ¼ bax þ be þ u ¼ hx þ v, (36.3)

where the reduced-form coefficient matrix is

h ¼ ba: (36.4)

The reduced-form disturbance matrix is

v ¼ be þ u, (36.5)

which has a covariance matrix of

 
O ¼ Eðvv0 Þ ¼ E ðbe þ uÞðbe þ uÞ0 ¼ bSb0 þ y2 , (36.6)
where E presents expectation operator.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1007

There are two types of restrictions on the reduced form: (i) the m  p regression
coefficient matrix h has rank k, the m  p components of h being expressed in terms
of kp + mk elements of a and b, and (ii) the m  m residual covariance matrix Ω
satisfies a factor analysis model with k common factors, the m(m + 1)/2 distinct
elements of Ω being expressed in terms of the k + km + m elements of s2, b, and y2.
The first restriction, which is the same as the simultaneous equation model, is
familiar to econometricians. The second restriction, which is the same as the factor
analysis model, is familiar to psychometricians. In Eq. 36.5, e, b, and u are regarded
as the common factors, the factor loadings, and the unique disturbances in the factor
analysis model, respectively.
We observe that the reduced-form parameters remain unchanged, when any
column, say j, of b is multiplied by a scalar and the j th row of a and sj are both
divided by the same scalar. To remove this indeterminacy of the model, we can
normalize the model through (i) s2, or (ii) b, or (iii) a. After normalization, the
maximum likelihood estimation (MLE) procedure can be used to obtain consistent
estimators of the elements in parameters a, b, and y2 (see Attfield 1983; Chen 1981;
Joreskog and Goldberger 1975; and others). In the following, we demonstrate how
to apply the MIMIC model to test CAPM and APT.

36.2.2 The Testing Model of CAPM by the MIMIC Approach

The CAPM can be rewritten, in terms of MIMIC model, as follows:

er i ¼ bier m þ e
ui
i ¼ 1, . . . , N, (36.7)
er m ¼ er m þ e em,

where
er i ¼ the realized excess return (total return less risk-free rate) on security i in
a deviation form
er m ¼ the realized excess return of the NYSE Composite Index
er m ¼ the unobservable excess return on the market portfolio
In this special one-factor case, we remove the indeterminacy by setting
the coefficient on er m equal to one. Equation 36.7 is a simultaneous equation
model, in which there are N equations linking the individual security
(or portfolio) return to the unobservable true market return and one equation linking
the unobservable true market portfolio return to the realized return of the NYSE
composite index. After obtaining the estimated betas from simultaneous equation
system of (36.7), a cross-sectional regression of the security return against its risk
(b) will be used to test CAPM or to estimate the riskless rate and the market risk
premium as follows:

r it ¼ ^a 0t þ ^a 1t bi , (36.8)
1008 C.-F. Lee et al.

where
rit ¼ the excess return on security i at time t
^
a 0t ¼ the estimate of the intercept which is supposed to be zero
^
a 1t ¼ the estimate of the market risk premium
Four different estimation procedures will be used to estimate Eq. 36.8. They are:
(i) stationary OLS, (ii) nonstationary OLS, (iii) GLS, and (iv) MLE.3

36.2.3 The Testing Model of APT by the MIMIC Approach

The testing model of APT, in terms of the MIMIC model, can be rewritten as
follows:

er i ¼ bi1ef 1 þ . . . þ bikef k þ eui, ði ¼ 1, . . . , N Þ

ef j ¼ aj1eI 1 þ . . . þ ajpeI p þ ee j , , (36.9)
ðj ¼ 1, . . . , kÞ

where
ef j ¼ the jth unobservable factor
eI h ¼ the hth macroeconomic indicator, h ¼ 1,    ,p
For convenience and easy explanation, each factor ef is assumed to have
different set of explained indicators eI ’s. Note that there are N return equations
plus k factor equations in the simultaneous equation system (36.9). The LISREL
computer program of Jöreskog and Sörbon (1981) is used to estimate the parame-
ters, a and b, in Eq. 36.9. A cross-sectional regression is also used to test APT and to
estimate the riskless rate and the factor risk premia by regressing the security return
against its risks, b’s. The a’s coefficients in Eq. 36.9 can be used to explain the
relationship between factors and indicators.

36.3 Test of CAPM

This section tests CAPM using the market model and the MIMIC model described in
Sect. 36.2. The objective is to investigate whether the MIMIC method yields
a different inference from the market model. Nineteen industry common stock
portfolios are formed with the same manner used by Schipper and Thompson
(1981), and Stambough (1982).4 The return on a portfolio is the arithmetic average
of returns for firms on the CRSP monthly tape with the appropriate two-digit SEC
code for the given month. The tests use returns from the period 1963–1982, and this
total period is divided into two equal subperiods: (i) subperiod 1, 1963–1972 and

3
The terminologies stationary OLS and nonstationary OLS have been used by Friend and West-
field (1980). The GLS and MLE methods have been discussed by Litzenberger and
Ramaswamy (1979).
4
See Chen (1981).
36 Multifactor, Multi-indicator Approach to Asset Pricing 1009

(ii) subperiod 2, 1973–1982. Portfolios are formed primarily because they provide
a convenient way to limit the computational dimensions of the MIMIC method. As
mentioned by Stambough (1982), industry portfolios also allow rejection of CAPM
due to the presence of additional industry-related variables in the risk-return relation.
Table 36.1 indicates the number of securities in each portfolio, the SEC codes,
and betas calculated from the market model and also from the MIMIC model. When
the MIMIC model was used to estimate betas by the portfolio excess returns in
subperiod 1, convergent problems were encountered during minimization so that
the result is inappropriate. Consequently, raw returns in deviation form on the
portfolios are used to estimate betas for subperiod 1. The betas estimated from
the MIMIC model are very close to those from the market model in both periods.
This evidence supports Stambough’s discovery that inferences about CAPM are
very insensitive to alternative market indexes.
Table 36.2 presents return-risk trade-off from the cross-sectional relationship in
which the average monthly excess portfolio returns (monthly portfolio returns less
monthly return on 3-month Treasury bills) is regressed on a beta estimated either
from the MIMIC model or the market model from two different 120-month periods.
Four different estimation procedures are used to estimate the intercepts and the
market risk premia. The OLS method presents two sets of t-statistics shown in
the parentheses under the same relevant regression coefficients. The first set
(denoted S) assumes that the regression coefficients are constant or stationary
over each 120-month period. The second set (denoted NS) of the OLS t-statistics
allows the nonstationarities of the regression coefficients by computing the
cross-sectional regression coefficients in each month and deriving the appropriate
standard errors from the time series 120 estimates of the OLS regression coeffi-
cients. The GLS and MLE methods also permit the nonstationary coefficients.
Thus, their t-statistics are derived from the same procedure in the OLS
(NS) method. Although the OLS and the GLS estimators are biased and inconsistent
due to measurement error in beta (see Litzenberger and Ramaswamy 1979), the
maximum likelihood estimators are consistent simply because MLE takes care of
measurement error in beta.
In Table 36.2, the coefficients in the NS regression of OLS, GLS, and MLE are
obviously characterized by much larger standard errors so that they lose any
significance shown in the stationary OLS regression.5
However, different estimated betas cause little changes in return-risk relation-
ships. From the results of the OLS stationary method, there exists a significant
return-risk relationship in subperiod 1, but not in subperiod 2 in both MIMIC and
market models. The poor return-risk relationship in subperiod 2 may be due to the
poor performance of CAPM in determining a pricing relation. In the next section,
APT will be used to examine an alternative pricing relationship. Even though the
null hypothesis of CAPM that a0 ¼ 0 cannot be rejected at the 5 % level in all four
cases, all coefficients are positive. In addition, the intercept in subperiod 2, about

5
The similar results were also found in the Friend and Westfield’s (1980) study of co-skewness.
1010 C.-F. Lee et al.

Table 36.1 Industry portfolio SEC codes, number of firms, and estimated betas
Portfolio description SEC code # of firmsEstimated betas
Period 1 Period 2
Market MIMC Market MIMC
1972 1982 model model model model
1. Mining 10–14 56 71 1.056 1.009 0.922 0.916
2. Food and beverages 20 75 51 0.894 0.841 0.803 0.803
3. Textile and apparel 22,23 58 45 1.264 1.220 1.081 1.082
4. Paper products 26 30 30 1.030 1.029 0.910 0.909
5. Chemical 28 87 83 0.949 0.947 0.847 0.846
6. Petroleum 29 28 22 0.706 0.792 0.745 0.740
7. Stone, clay, glass 32 43 31 1.050 1.106 1.045 1.045
8. Primary metals 33 56 49 1.136 1.193 0.932 0.930
9. Fabricated metals 34 45 46 1.145 1.155 1.102 1.102
10. Machinery 35 93 104 1.234 1.231 1.104 1.104
11. Appliance and 36 87 82 1.384 1.401 1.179 1.180
elec. equip.
12. Transport. equip. 37 64 50 1.209 1.275 1.150 1.151
13. Misc. manufactrng. 38,39 64 59 1.375 1.314 1.197 1.198
14. Railroads 40 18 11 1.294 1.229 0.899 0.895
15. Other transport. 41, 42, 44 34 35 1.335 1.447 1.203 1.203
45, 47
16. Utilities 49 138 152 0.443 0.467 0.564 0.562
17. Department stores 53 35 28 1.149 1.104 1.125 1.125
18. Other retail trades 50–52, 103 97 1.144 1.088 1.123 1.124
54–59
19. Banking, finance, 60–67 184 240 0.968 1.021 1.069 1.068
real estate
This table presents the number of firms and betas in each industry portfolio. The numbers of firms
are obtained in the end of 1972 and the end of 1982. The betas shown in the first column are
estimated from the market model, while those in the second column are estimated from the MIMIC
model. Period 1 represents the sample period from 1963 to 1972 and period 2 represents the sample
period from 1973 to 1982

4.3 % annual rate, is too high. This is consistent with prior tests of the traditional
version of CAPM.
All nonstationary estimates of a0 and a1 in OLS, GLS, and MLE in Table 36.2
are insignificant. Because of the low test power for all nonstationary procedures
(the standard errors are too high), in the following, only the magnitudes of the
estimated coefficients are discussed. The MLE and GLS estimates of a0 are much
lower and much closer to zero than the corresponding OLS estimates in all four cases.
In addition, the MLE estimates of a1 is greater than the corresponding GLS estimates
and is closer to the realized market risk premia in all four cases. The realized market
risk premia are 0.751 % and 0.636 % monthly for periods 1 and 2, respectively.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1011

Table 36.2 Return-risk cross-sectional relationships of CAPM: 1963–1982

Panel A: MIMC model Panel B: Market model
2 2
Procedure a0 a1 R a0 a1 R
Period 1: 1963–1972
OLS-S 0.159 0.539 0.258 0.187 0.516 0.243
(0.71) (2.70)* (0.85) (2.61)*
OLS-NS 0.159 0.539 0.258 0.187 0.516 0.243
(0.41) (1.03) (0.48) (1.02)
GLS 0.029 0.660 0.095 0.599
(0.06) (1.29) (0.19) (1.22)
MLE 0.106 0.793 0.082 0.770
(0.24) (1.29) (0.13) (1.22)
Period 2: 1973–1982
OLS-S 0.370 0.266 0.004 0.363 0.273 0.001
(1.33) (0.97) (1.29) (0.99)
OLS-NS 0.370 0.266 0.004 0.363 0.273 0.001
(0.60) (0.31) (0.59) (0.32)
GLS 0.106 0.532 0.107 0.531
(0.24) (0.71) (0.24) (0.71)
MLE 0.081 0.556 0.081 0.556
(0.17) (0.71) (0.17) (0.71)
This table presents return-risk trade-off from the cross-sectional relationship in which the average
monthly excess portfolio returns are regressed on a beta estimated either from the MIMC model or
the market model from two different 120-month periods. For different estimation procedures,
stationary OLS, nonstationary OLS, GLS, and MLE, are used to estimate the intercepts
and the market risk premia. The stationary OLS version corresponds to the regression Ri  Rf
¼ a0 þ a1 bi þ ee i , (i¼1,  ,N). The other versions correspond to the regression Ri  Rf ¼ a0 þ a1
bi þ ee i, (i ¼ 1,  ,N; and t ¼ 1,  T. The monthly returns are multiplied by 100 before regressing.
The reported coefficients are arithmetic average of the time series, while t-statistics are in
parentheses under each relevant coefficient. *represents significant at the 5 % level

This evidence proves that the MLE estimator in the return-risk cross-sectional
regressions is more appropriate than OLS or GLS estimator in testing CAPM.
In sum, we have proposed an alternative estimator of betas by the MIMIC model
in which measurement error in a market portfolio is allowed. Nevertheless, this
reasonable alternative method does not gain much from the traditional OLS
estimator. However, some interesting results have surfaced. This evidence supports
Stambough’s conclusion that the tests of CAPM are insensitive to different market
indexes. In return-risk cross-sectional regressions, our evidence shows that the
MLE estimator is more appropriate than the OLS or GLS estimator due to
measurement error in beta. From these two interesting results, we conclude that
measurement error on beta is more serious than measurement error on the market
portfolio in testing CAPM.
1012 C.-F. Lee et al.

36.4 Tests of APT by MIMIC Approach

This section tests APT using the MIMIC model demonstrated in Sect. 36.2. The
objective is to investigate that (i) the proper number of factors is used to explain the
data and (ii) the relationships between factors and indicators which are measured by
macroeconomic variables. The same 19 industry portfolios described in previous
section are used here. The macroeconomic variables are selected from those most
likely related to common stock returns. In the following, the indicators selected in
this study will be discussed.

36.4.1 Macroeconomic Variables as the Indicators

In early 1970s, several studies attempt to employ economic methods to investigate

the relationship between money supply and aggregated common stock prices. Models
developed by Keran (1971), Homa and Jaffee (1971), and Hamburger and Kochin
(1972) appear to have met with considerable success in explaining the behavior of
Standard and Poor’s Composite Index. However, Pesando (1974) reexamines above
models using different periods. He finds that the extraordinary success of these
methods in tracking the behavior of stock prices during the sample period may be
illusory. We believe that the above spurious regression phenomenon results from
ignoring the autocorrelated errors in time series regression equations as pointed out
by Granger and Newbold (1974). Gargett (1978) used a qualitative method to study
the relationship between these two variables. He discovered that the Dow Jones
Industrial Index follows changes in money supply with a lag of 3 months.
The relationship between stock returns and inflation has been extensively
studied. In particular, Bodie (1976), Nelson (1976), and Fama and Schwert
(1977) all present evidence that monthly returns to NYSE Composite Index are
negatively related to the inflation rate as indicated by the consumer price index
(CPI) since 1953. Cohn and Lessard (1981), Gultekin (1983), and Solnik (1983)
also find that stock prices are negatively related to nominal interest rate and
inflation in a number of countries. Fama (1981) suggests a reason why the stock
market reaction to unexpected inflation is weak. He argues that unexpected inflation
is contemporaneously correlated with unexpected movements in important “real”
variables, such as capital expenditures or real GNP, so that the correlation between
stock returns and unexpected inflation is spurious. After extensively reexamining
the relationship between the stock returns and inflation, Geske and Roll (1983)
conclude that only Nelson’s (1979) and Fama’s (1981) money demand explanation
is logically consistent, but it seems unable to fully explain all of the empirical
phenomena. Geske and Roll (1983) therefore propose the fiscal and monetary
explanation. They argue that the basic underlying relation is between stock returns
and changes in inflationary expectations.
In exploring the common stocks as hedges against the investment opportunity
sets, Schipper and Thompson (1981) select two candidates for state variables.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1013

They are price level as measured by consumer price index (CPI) and the real gross
national product less corporate profit (GNP). They find that hedge portfolios offer
meaningful hedging potential in portfolio-formation period. In addition, CAPM or
the market model indicates that the return on a security or a portfolio most likely
co-move with the return on the market portfolio.
In summary, the variables most likely correlated with a stock return would be
classified as five categories: (1) money supply, (2) real production, (3) inflation,
(4) interest rate, and (5) market return. Further, Brigham (1982) decomposes a risk
premium into maturity risk premium and default risk premium. Thus, these two
indicators are also included in our study. According to the above discussion, the
following 11 variables are selected as the indicators.
1. Return on the market portfolio (RM): the return on NYSE common stock
composite index.
2. Transaction volume (VL): the change rate in the transaction volume (shares)
for all of the NYSE common stocks.
3. Real riskless rate (RF): the real interest rate on 3-month Treasury bills.
4. Maturity risk premium (MP): the difference between the real interest rates on
long-term Treasury bonds (10 or more years) and on 3-month Treasury bills.
5. Default risk premium (DP): the difference between the real interest rates on
new AA corporate bonds and on 3-month Treasury bills.
6. Consumer price index inflation rate (CPI): the change rate in urban consumer
price index for all items.
7. Money supply (M2): the real change rate in money stock as measured by M2
(M1 + time deposits).
8. Velocity of money supply (PI/M2): the ratio of personal income to money
supply M2. This is an alternative measure of money supply.
9. Real industrial production (IP): the change rate in real total industrial
productions.
10. Real auto production (IPA): the change rate in real automotive products.
11. Real home production (IPH): the change rate in real home goods.
Since the automobile and housing industries generally lead the rest of the
economy, the last two indicators, IPA and IPH, are used to catch up the first two
biggest industries. The reason to select industrial production instead of GNP in this
study is that all other indicators are published monthly while GNP is published
quarterly. Since the industrial production is a very good proxy for GNP, we sacrifice
GNP measure to gain the number of time periods.

36.4.2 Empirical Results

After carefully selecting the indicator candidates, the time lag or leading problem
needs to be solved. The correlation coefficients between the market portfolio
return and the other indicators with lags and leadings of zero to 5 months for both
periods were examined. All indicators are contemporaneously correlated with the
1014 C.-F. Lee et al.

Table 36.3 Eigenvalue as a percentage of the first eigenvalue for 19 industry portfolios:
1963–1982
Factor PRC ALP SCF ULS
Panel A: Period: 1963–1972
1 100 % 100 % 100 % 100 %
2 3.8 4.4 3.9 2.9
3 3.1 1.5 2.6 1.7
4 2.1 1.0 2.3 0.8
5 1.9 0.7 2.0 0.7
6 1.6 0.6 1.9 0.6
Panel B: Period: 1973–1982
1 100 % 100 % 100 % 100 %
2 7.9 6.9 7.1 7.2
3 2.1 2.8 2.9 1.8
4 1.3 1.0 1.9 0.9
5 1.0 0.6 1.6 0.5
6 1.0 0.4 1.4 0.4
This table shows the eigenvalues as a percentage of the first eigenvalue. Four methods are used to
determine the number of factors in APT model. They are principal component analysis (PRC),
alpha factor analysis (ALP), simple common factor analysis (SCF), and unweighted least squares
method (ULS)

market portfolio except three real production indicators. The real production
indicators follow the market portfolio return with a lag of 2 or 3 months.
Therefore, in this study, all indicators are contemporaneous except three real
production indicators which is a 2-month lag.
Before APT is directly tested by the MIMIC model, factor analysis is prelimi-
narily used to determine the number of factors in both periods. Table 36.3 shows the
eigenvalues as a percentage of the first eigenvalue. Clearly, it is only one factor in
period 1, while it is perhaps two factors in period 2 by “scree” test described in
Chapter 4 of Wei (1984).6 Consequently, at most a two-factor model is enough to
explain the historical data. Three alternative MIMIC models are proposed to test
APT: (i) one-factor 11-indicator model, (ii) one-factor six-indicator model, and
(iii) two-factor six-indicator model. When the two one-factor models are used to
test APT in period 1, there is little difference between 11-indicator and six-indicator
model. Thus, only six indicators are used in the two-factor model to save the
computer time.7
The structural coefficients of APT in the MIMIC model for period 1 are reported
in Table 36.4a. In both one-factor models, only the stock market-related variables,

6
In his dissertation, Wei (1984) has shown that the “scree” test is a powerful test in identifying the
number of relevant factors in the APT. By using simulation study, Wei has shown that Roll and
Ross’s (1980) ML method in estimating factors are inferior to methods listed in Table 36.3.
7
It is very expensive to run LISREL program, especially for more than two factor models.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1015

the market return (RM), and the transaction volume (VL) are significant at the 5 %
level. From this evidence, if the pricing relation in period 1 is a one-factor APT, this
common factor would be most likely related to only the stock market-related
indicators, namely, the market portfolio return and the transaction volume. Other
indicators may be correlated with this single common factor, but they are obviously
not as important as the stock market-related indicators. Now, let us closely examine
other indicators with an absolute t-value of greater than one for 11-indicator model.
The real riskless interest rate (RF), CPI inflation rate (CPI), and the real auto
production (IPA) are negatively correlated with this common factor, while the
velocity of money supply (PI/M2) is positively related to this common factor.
If we regard this common factor as a proxy of market portfolio because most of
the weight is on the market portfolio, then, except for real auto production, this
evidence supports previous studies done on the relationship between common stock
returns and other indicators (Keran 1971; Homa and Jaffee 1971; Hamburger and
Kochin 1972; Pesando 1974; Bodie 1976; Nelson 1976; Fama and Schwert 1977;
and others). However, there is no previous study which examines the relationship
between stock returns and the real auto production.
Some might argue that weak relationship between non-stock market indicators
and the common factor is due to multicollinearities among the indicators. There-
fore, six of the 11 indicators are selected to represent each category indicator. They
are the market portfolio return (RM), the transaction volume (VL), real riskless
interest rate (RF), CPI inflation rate (CPI), money supply (M2), and the real total
industrial production (IP). This is the one-factor six-indicator model. The result of
this model is shown in Table 36.4a column 2. The result of this one-factor
six-indicator model is very close to that of the one-factor 11-indicator model. As
mentioned before, only the stock market-related indicators are significantly corre-
lated with the common factor. Real riskless interest rate, inflation, money supply,
and real production are all negatively but insignificantly related to the common
factor. For factor equation, the 11-indicator model has only a little high R-square
than the six-indicator model. They are 0.5611 and 0.5412, respectively. Comparing
the betas of these two one-factor models in Table 36.4a, they are very highly
correlated with a correlation coefficient of about 1.000. In addition, the average
R-square of each return equation in both one-factor models is the same with a value
of 0.811. Up to this point, there is not much difference between the one-factor
11-indicator and the one-factor six-indicator models. Later on, the cross-sectional
regression will be used to double check this result.
Even though we have already discussed that the appropriate model is a one-factor
model for period 1 by the scree test of factor analysis, we want to use a two-factor
model to double check whether the second factor is significant or not.
A predetermined two-factor six-indicator model will be used to test APT. Factor
analysis is employed to classify these six indicators into two groups. The first group
includes the market portfolio and the transaction volume, while the second group
includes other four indicators: real riskless interest rate (RF), inflation rate (CPI),
money supply (M2), and real total industrial production (IP). The two-factor result
shown in Table 36.4a columns 3 and 4 displays that only the first factor is
1016 C.-F. Lee et al.

significantly related to RM and VL, whereas the second factor is very insignifi-
cantly correlated with the second group indicators. This is also evident by exam-
ining from the second betas in the table. All of the second betas are insignificant.
Furthermore, the first beta coefficient is very highly correlated with the betas in both
one-factor models with both correlation coefficients of 0.996. This is further

Table 36.4 Structural coefficients of APT in the MIMIC model

Period 1: 1963–1972
a’s Coefficients
One-factor 11-indicator One-factor 6-indicator Two-factor 6-indicator
Indicator f1 f1 f1 f2
RM 0.908(8.08)* 0.925(7.55)* 0.923(8.85)*
VL 0.040(2.55)* 0.043(2.60)* 0.040(2.28)*
RF 17.24(1.2) 3.451(0.76) 0.300(–)
MP 15.84(0.90)
DP 1.693(0.12)
CPI 16.51(1.2) 4.596(1.1) 0.119(0.12)
M2 0.039(0.03) 0.344(0.31) 0.458(0.58)
PI/M2 22.23(1.29)
IP 0.141(0.20) 0.469(1.1) 0.054(0.06)
IPA 0.078(1.3)
IPH 0.189(0.92)
R-square 0.5611 0.5412 0.5912 0.0774
b’s Coefficients
One-factor 11-indicator One-factor 6-indicator Two-factor 6-indicator
Industry f 1 f1 f1 f2
1 1.000(–) 0.974(12.1)* 1.000(–) 0.410( 0.06)
2 0.834(16.5)* 0.812(13.6)* 0.857(16.3)* 0.555(0.06)
3 1.210(16.2)* 1.178(13.4)* 1.211(16.1)* 0.390(0.06)
4 1.020(14.7)* 0.993(12.5)* 1.059(14.4)* 1.075(0.06)
5 0.939(17.4)* 0.914(14.1)* 0.942(17.4)* 0.270( 0.06)
6 0.785(10.5)* 0.764(9.62)* 0.831(10.3)* 1.575(0.06)
7 1.097(15.3)* 1.068(12.9)* 1.122(15.2)* 0.555(0.06)
8 1.183(15.8)* 1.151(13.2)* 1.190(15.7)* 0.120( 0.05)
9 1.146(18.1)* 1.115(14.4)* 1.126(17.6)* 1.205(0.06)
10 1.221(18.2)* 1.188(14.5)* 1.192(17.4)* 1.670(0.06)
11 1.389(16.6)* 1.352(13.6)* 1.335(15.0)* 2.875(0.06)
12 1.264(17.7)* 1.231(14.2)* 1.236(17.0)* 1.675(0.06)
13 1.302(17.2)* 1.268(13.9)* 1.268(16.2)* 1.995(0.06)
14 1.218(13.2)* 1.186(11.5)* 1.215(13.0)* 0.485(0.06)
15 1.435(14.0)* 1.396(12.1)* 1.405(13.5)* 1.760(0.06)
16 0.463(7.66)* 0.450(7.30)* 0.527(7.33)* 2.340(0.06)
17 1.094(14.1)* 1.065(12.2)* 1.112(14.1)* 0.265(0.06)
18 1.078(16.5)* 1.050(13.6)* 1.094(16.4)* 0.180(0.06)
19 1.012(14.7)* 0.985(12.5)* 1.056(14.3)* 1.260(0.06)
(continued)
36 Multifactor, Multi-indicator Approach to Asset Pricing 1017

Table 36.4 (continued)

Period 2: 1973–1982
a’s Coefficients
One-factor 11-indicator Two-factor 6- indicator
Indicator f1 f1 f2
RM 0.859(6.38)* 1.000(6.67)*
VL 0.073(3.34)* 0.075(2.83)*
RF 5.586(1.1) 5.600(–)
MP 2.254(0.40)
DP 7.417(0.70)
CPI 9.786(1.6)* 1.566(0.32)
M2 1.820(1.2) 1.810(0.72)
PI/M2 28.60(1.16)
IP 0.290(0.42)
IPA 0.142(1.6)* 0.368(1.7)*
IPH 0.065(0.20)
R-square 0.5811 0.5525 0.1836
b’s Coefficients
One-factor 11-indicator Two-factor 6-indicator
Industry f1 f1 f2
1 0.845(8.85)* 0.993(7.86)* 0.502(1.95)*
2 0.786(13.4)* 0.724(8.39)* 0.184(1.7)*
3 1.073(13.3)* 0.948(7.98)* 0.379(1.9)*
4 0.884(13.2)* 0.850(14.4)* 0.081(1.0)
5 0.829(13.5)* 0.804(8.81)* 0.054(0.77)
6 0.676(7.96)* 0.797(7.41)* 0.408(1.93)*
7 1.026(14.2)* 0.948(8.62)* 0.218(1.7)*
8 0.898(12.4)* 0.896(8.67)* 0.029(0.38)
9 1.083(14.8)* 1.026(8.97)* 0.146(1.4)
10 1.082(14.5)* 1.054(9.11)* 0.053(0.63)
11 1.171(14.7)* 1.098(8.87)* 0.192(1.6)
12 1.139(14.4)* 1.058(8.72)* 0.222(1.7)*
13 1.188(14.7)* 1.115(8.86)* 0.208(1.6)
14 0.850(11.0)* 0.869(8.33)* 0.089(0.99)
15 1.186(13.0)* 1.088(8.21)* 0.288(1.7)*
16 0.534(9.76)* 0.506(7.32)* 0.076(1.2)
17 1.122(12.8)* 0.961(7.55)* 0.506(2.0)*
18 1.113(14.3)* 1.001(8.40)* 0.341(1.9)*
19 1.042(13.1)* 0.959(8.29)* 0.242(1.7)*
This table presents the structural coefficients of the APT in the MIMIC model for period 1 and
period 2. The structural model is written as
er i ¼ bi1ef 1 þ bi2ef 2 þ e
ui, i ¼ 1, . . . , 19
ef j ¼ aj1 ðRMÞ þ aj2 ðVLÞ þ aj3 ðRFÞ þ aj4 ðMPÞ þ aj5 ðDPÞ þ aj6 ðCPI Þ þ aj7 ðM2Þ
þ aj8 ðPI=M2Þ þ aj9 ðIPÞ þ aj10 ðIPAÞ þ aj11 ðIPH Þ þ ee j , j ¼ 1 or 1, 2:
Variables used as indicators in the structural model are the market return (RM), transaction volume
(VL), real risk interest rate (RF), maturity risk premium (MP), default risk premium (DP), CPI inflation
rate (CPI), money supply (M2), velocity of money supply (PI/M2), real total industrial production (IP),
real auto production (IPA), and real home production (IPH). *represents significant at the 5 % level
1018 C.-F. Lee et al.

supported by the R-square criteria. The average R-square of each return equation in
the one-factor model is 0.811, while 0.844 in the two-factor model. If the adjusted
R-square is used as the criteria, the increase in R-square will be insignificant. Thus,
a one-factor APT is good enough to explain the first period data. We will reconfirm
this argument by the cross-sectional regression.
Table 36.5 Panel A reports the return-risk cross-sectional relationship in period 1
for APT in the MIMIC model. Both one-factor models have a very similar result.
Their adjusted R-squares are the same of 0.258. The intercepts are both insignifi-
cantly different from zero but positive (recall that the LHS variable is excess return).
The factor risk premia for both one-factor models are positive and significant. This is
exactly the result that should have been concluded from APT. On the other hand, the
intercept in the two-factor model is much higher than those in the one-factor model,
and the two factor risk premia are both insignificant.8 The adjusted R-square is a little
lower than those in the one-factor models. From the results of the MIMIC model and
the cross-sectional regressions, it is probable that one-factor APT with the market
portfolio and the transaction volume as the determinants of the single factor is the
appropriate pricing model for period 1. Comparing the MIMIC CAPM and CAPM in
Table 36.2 with this one-factor model, the MIMIC CAPM is closer to the one-factor
model. In addition, the factor risk premium is closer to the realized market risk
premium than that in the MIMIC CAPM model.
Now, let us examine APT in period 2. The structural coefficients of APT in the
MIMIC model for period 2 are represented in Table 36.4b. Because six indicators
do not make much difference from 11 indicators in the one-factor model in period 1,
only the 11-indicator model is used to test APT for the one-factor model in period 2
in order to save the computer time. From the one-factor model, the result is similar
to that in period 1. The stock market return and the transaction volume are the most
significant indicators. However, inflation and the real auto production are also
significant even just at the marginal level. This reinforces our suspicion that there
may be more than one factor in period 2.
Let us closely examine the indicators with an absolute t-value greater than one.
Real interest riskless rate, inflation, money supply, and real auto production are all
negatively correlated with the common factor, while the velocity of money supply is
positively correlated with this common factor. Overall, this result again supports
previous studies. Now, let us turn to the two-factor six-indicator model. Follow the
same procedures done for period 1. The result is displayed in Table 36.4b columns
2 and 3. The market portfolio (RM) and the transaction volume (LV) are significantly
related to the first factor. However, the real auto production is significantly correlated
with the second factor; this time even only at the marginal level. Comparing other
indicators with those in period 1, real riskless interest rate is again negatively related
to the second factor, but inflation rate and money supply are positively related to the
second factor in the period. From the result of the relationship between the factors
and the indicators, the second factor is still important even though less important than

8
The loss of the significance of the first factor risk premium is due to the multicollinearity problem.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1019

Table 36.5 Return-risk cross-sectional relationships of APT in the MIMIC model: 1963–1982
2
^
a0 ^
a1 ^
a2 R
Model Panel A: Period 1, 1963–1972
One-factor 0.159 0.543* 0.258
11-indicator (0.710) (2.693)
One-factor 0.159 0.558* 0.258
6-indicator (0.710) (2.692)
Two-factor 0.507 0.206 0.071 0.257
6-indicator (1.219) (0.529) (1.221)
Panel B: Period 2, 1973–1982
One-factor 0.370 0.285 0.002
11-indicator (1.328) (0.967)
Two-factor 0.040 0.684* 0.362* 0.215
6-indicator (0.150) (2.377) (2.066)
This table presents the return-risk cross-sectional relationship for the APT in the MIMIC model.
The cross-sectional relationship between return and risk can be written as Ri ¼ ^ a0 þ ^a 1 b1 þ ^
a 2 b2
þ ee i, i¼1,. . .,19. Panel A reports the relationship in period 1 and Panel B reports the relationship in
period 2. The average monthly returns are multiplied by 100 before regressing. *represents
significant at the 5 % level

the first factor for period 2. This can also be checked by the significance of betas in
Table 36.4b. Ten out of the 19 s factor betas are significant although the relationship
is not as strong as those in the first factor betas. Because the first factor is correlated
with stock market-related indicator, this factor can be regarded as a proxy of the
market portfolio. The correlation coefficient between the beta in the first factor and
the beta in the MIMIC CAPM is very high with a coefficient of 0.923. Further, the
average R-square of each return equation in the one-factor model is 0.818, while
0.885 in the two-factor model. When the adjusted R-square is used as a criteria, the
increase in R-square should not be trivial and will be significant. In addition, the
R-square for the second factor equation is as high as 0.1836. All of these results
indicate that the second factor should be important. We further check whether the
second factor is important or not by the cross-sectional regression.
The risk-return cross-sectional relationship of APT in the MIMIC model for
period 2 is shown in Table 36.5 Panel B. For the one-factor model, both the
intercept and the factor risk premium are insignificant. The intercept is too high
with an annual rate of 4.3 %, while the factor risk premium is far below the realized
market risk premium with a monthly rate of 0.636 %. The adjusted R-square is
negative. On the other hand, the intercept for the two-factor model is insignificant
and very near to zero, while the two factor premia are significant with the first factor
premium very close to the realized market risk premium. The adjusted R-square is
as high as 0.215. The results of APT in the MIMIC model confirm the scree test of
the factor analysis and the poor performance of CAPM in previous section for
period 2. Furthermore, comparing the result in Table 36.5 Panel A with that in
Table 36.2, we can conclude that the two-factor APT outperforms the one-factor
APT, the MIMIC CAPM, and CAPM in period 2. But the one-factor APT does
1020 C.-F. Lee et al.

a better job than the two-factor APT in period 1, and there is not much difference
among the one-factor APT, the MIMIC CAPM, and CAPM. This evidence supports
Ross’s argument that APT is more general than CAPM because APT allows more
than one factor in the pricing relation.
It will be interesting to see what will happen if market variables (the market
portfolio and transaction volume) are excluded from the model. Table 36.6 shows
the structural coefficients of the one-factor APT without market variables.9 The
b coefficients (factor loadings) shown in Table 36.6 for both periods are very close
to those estimated from the one-factor APT with market variables shown in
Tables 36.4a and 36.4b, respectively.10 However, the results from the factor
equation have dramatically changed.
For period 1, the R-square is only 0.0611 which is much lower than 0.5611 from
the one-factor APT with market variables. This result indicates that the market
variables play the major role in the pricing behavior during period 1. When the
market variables excluded from the model, among nine indicators, only the two
money supply variables, M2 and PI/M2, are positively related to the unique
common factor at the marginal level. For period 2, the R-square of the factor
equation for the model without market variables is 0.3915 which is higher than
the one in period 1. This evidence denotes that, in addition to the market variables,
some other macroeconomic indicators play a relatively important role in the pricing
behavior during period 2. Among the nine nonmarket variables, real risk-free rate
(RF), default risk premium (DF), and inflation (CPI) are significantly negatively
related to the common factor, while the velocity of money supply (PI/M2) and the
real auto production (IPA) are significantly positively related to the common factor.
All of the relationships are as we expect. The results from the model without market
variables also confirm our previous evidence that the 1963–1972 data can be
described by a one-factor APT with the market variables as its indicators, while
the 1973–1982 data should be explained by more than one-factor APT.

36.5 Concluding Remarks

This paper bases on parts of Wei’s (1984) dissertation using a MIMIC approach to test
CAPM and APT. The results support the conclusion that APT outperforms CAPM,
especially for the period from 1973 to 1982. The beta estimated from the MIMIC
model by allowing measurement error on the market portfolio does not significantly
improve the OLS beta. However, the MLE estimator does a better job than the OLS
and GLS estimators in the cross-sectional regressions because the MLE estimator
takes care of the measurement error in beta. Therefore, the measurement error problem

9
Only the one-factor APT is used to investigate the difference between the models shown in
Table 36.4 and in Table 36.6.
10
If we normalize the one-factor 11-indicator model for period 2 shown in Table 36.4b by setting
b1 ¼ 1.00, it is easily seen that the b coefficients of one-factor model shown in Tables 36.4b and
36.6 column 2 are very similar.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1021

Table 36.6 Structural coefficients of the one-factor APT in the MIMIC model without market
variables: 1963–1982
a’s Coefficients
Indicator 1963–1972 1973–1982
RF 25.541(1.26) 18.666(3.39)*
MP 26.061(1.18) 1.067(0.17)
DP 4.350(0.23) 30.745(2.68)*
CPI 25.989(1.32) 23.l46(3.52)*
M2 3.652(1.85)* 1.630(0.98)
PI/M2 45.126(1.91)* 110.380(4.22)*
IP 0.932(0.97) 0.124(0.16)
I PA 0.046(0.57) 0.214(2.18)*
IPH 0.298(1.06) 0.120(0.33)
R-square 0.0611 0.3915
b’s Coefficients
Industry 1963–1972 1973–1982
1 1.000(–) 1.000(–)
2 0.833(16.45)* 0.930(10.10)*
3 1.209(16.20)* 1.270(10.03)*
4 1.019(14.70)* 1.045(9.99)*
5 0.939(17.43)* 0.981(10.11)*
6 0.785(10.53)* 0.800(9.64)*
7 1.097(15.33)* 1.214(10.41)*
8 1.183(15.79)* 1.063(9.64)*
9 1.146(18.11)* 1.283(10.63)*
10 1.221(18.25)* 1.282(10.53)*
11 1.389(16.58)* 1.386(10.60)*
12 1.265(17.74)* 1.349(10.50)*
13 1.302(17.16)* 1.407(10.63)*
14 1.218(13.19)* 1.006(8.96)*
15 1.436(14.05)* 1.404(9.92)*
16 0.463(7.66)* 0.632(8.23)*
17 1.094(14.13)* 1.329(9.84)*
18 1.078(16.46)* 1.318(10.47)*
19 1.012(14.70)* 1.233(9.95)*
This table shows estimated coefficients of the one-factor APT in the MIMIC model without market
variables. The structural model is shown as follows. *represents significant at the 5 % level
er i ¼ bi1ef 1 þ bi2ef 2 þ e
ui, i ¼ 1, . . . , 19
ef j ¼ a1 ðRFÞ þ a2 ðMPÞ þ a3 ðDPÞ
þ a4 ðCPI Þ þ a5 ðM2Þ þ a6 ðPI=M2Þ þ a7 ðIPÞ
þ a8 ðIPAÞ þ a9 ðIPH Þ þ ee j ,

on beta is more serious than that on the market portfolio. This evidence supports
Stambough’s (1982) argument that the inference about the tests of CAPM is insensi-
tive to alternative market indexes. When the one-factor APT with market variables is
compared with the model without market variables, we found that the market variables
1022 C.-F. Lee et al.

play a major role in pricing behavior. Therefore, we conclude that it is inappropriate

for the study of the relationship between the common factors extracted from APT and
the macroeconomic variables without including the market variables.

References
Attfield, C. (1983). Consistent estimation of certain parameters in the unobservable variable model
when there is a specification error. Review of Economics and Statistics, 65, 164–167.
Bodie, Z. (1976). Common stocks as a hedge against inflation. Journal of Finance, 31, 459–740.
Brigham, S. (1983). Fundamentals of financial management (3rd ed.). Chicago: Dryden Press.
Brigham, E. F., & Houston, J. F. (2012). Fundamentals of financial management (13th ed.).
Singapore: Cengage Learning.
Chang, C., Lee, A. C., & Lee, C. F. (2009). Determinants of capital structure choice: A structural
equation modeling approach. Quarterly Review of Economics and Finance, 49, 197–213.
Chen, C. (1981). The EM approach to the multiple indicators and multiple causes model via the
estimation of the latent variable. Journal of the American Statistical Association, 76, 702–708.
Chen, N. F., Roll, R., & Ross, S. A. (1983). Economic forces and the stock market: Testing the APT
and alternative asset pricing theories. Working paper, University of Chicago.
Cohn, R., & Lessard, D. (1981). Are markets efficient? Tests of alternative hypothesis. On the
effect of inflation on stock prices: International evidence. Journal of Finance, 36, 277–289.
Fama, E. (1981). Stock returns, real activity, inflation, and money. The American Economic
Review, 71, 545–565.
Fama, E., & Schwert, G. (1977). Asset returns and inflation. Journal of Financial Economics, 5,
115–146.
Fogler, H. R., John, K., & Tipton, J. (1981). Three factor interest rate differentials and stock
groups. Journal of Finance, 36, 323–336.
Friend, I., & Westerfield, R. (1980). Co-skewness and capital asset pricing. Journal of Finance, 35,
897–913.
Gargett, D. (1978). The link between stock prices and liquidity. Financial Analysis Journal, 34, 50–54.
Geske, R., & Roll, R. (1983). The fiscal and monetary linkage between stock returns and inflation.
Journal of Finance, 38, 1–33.
Goldberger, A. (1972a). Maximum-likelihood estimation of regressions containing unobservable
independent variables. International Economic Review, 13, 1–15.
Goldberger, A. (1972b). Structural equation methods in the social sciences. Econometrica, 40,
979–1001.
Granger, C., & Newbold, P. (1974). Spurious regression in econometrics. Journal of
Econometrics, 2, 111–120.
Gultekin, N. (1983). Stock market returns and inflation: Evidence from other countries. Journal of
Finance, 38, 49–65.
Hamburger, M., & Kochin, L. (1972). Money and stock prices: The channels of influence. Journal
of Finance, 27, 231–249.
Homa, K., & Jaffee, D. (1971). The supply of money and common stock prices. Journal of
Finance, 26, 1045–1066.
Joreskog, K., & Goldberger, A. (1975). Estimation of a model with multiple indicators and multiple
causes of a single latent variable. Journal of the American Statistical Association, 70, 631–639.
Joreskog, K., & Sorbom, D. (1981). LISREL V: Analysis of linear structural relationships by the
method of maximizing likelihood, user’s guide. Chicago: National Educational Resources.
Keran, M. (1971). Expectations, money, and stock market. Review, Federal Reserve Bank of
St. Louis, 53, 16-31.
Lahiri, K. (1976). Inflationary expectations: Their formation and interest rate effects.
The American Economic Review, 66, 124–131.
36 Multifactor, Multi-indicator Approach to Asset Pricing 1023

Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets. Review of Economics and Statistics, 47, 13–37.
Litzenberger, R., & Ramaswamy, K. (1979). The effects of personal taxes and dividends on capital
asset prices: Theory and empirical evidence. Journal of Financial Economics, 7, 163–195.
Maddala, G. S., & Nimalendran, M. (1996). Error-in-variables problems in financial models.
In G. S. Maddala & C. R. Rao (Eds.), Handbook of statistics. New York: Elsevier Science.
Mossin, J. (1966). Equilibrium in a capital asset market. Econometrica, 34, 768–783.
Nelson, C. (1976). Inflation and rate of returns on common stocks. Journal of Finance, 31, 471–483.
Nelson, C. (1979). Recursive structure in U.S. income, prices and output. Journal of Political
Economy, 87, 1307–1327.
Pesando, J. (1974). The supply of money and common stock prices: Further observations on
econometric evidence. Journal of Finance, 29, 909–921.
Roll, R. (1977). A critique of the asset pricing theory’s tests. Journal of Financial Economics, 4,
129–176.
Roll, R., & Ross, S. (1980). An empirical investigation of the arbitrage pricing theory. Journal of
Finance, 35, 1073–1103.
Ross, S. (1976). The arbitrage theory of capital asset pricing. Journal of Economic Theory, 13, 341–360.
Ross, S. (1977). Return, risk, and arbitrage. In I. Friend & J. L. Bicksler (Eds.), Risk and return in
finance (Vol. I). Ballinger Pub. Co.: Cambridge, MA.
Schipper, K., & Thompson, R. (1981). Common stocks as hedges against shifts in the consumption
on investment opportunity set. Journal of Business, 54, 305–328.
Shanken, J. (1982). The arbitrage pricing theory: Is it testable? Journal of Finance, 37, 1129–1140.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance, 19, 425–442.
Solnik, B. (1983). The relation between stock prices and inflationary expectations: The interna-
tional evidence. Journal of Finance, 38, 35–48.
Stambough, R. (1982). On the exclusion of assets from tests of the two-parameter model:
A sensitivity analysis. Journal of Financial Economics, 10, 237–268.
Titman, S., & Wessels, R. (1988). The determinants of capital structure choice. Journal of
Finance, 43, 1–19.
Turnovsky, S. (1970). Empirical evidence on the formation of price expectations. Journal of the
American Statistical Association, 65, 1441–1454.
Wei, K. (1984). The arbitrage pricing theory versus the generalized intertemporal capital asset
pricing model: Theory and empirical evidence. Unpublished Ph.D. dissertation, University of
Illinois.
Yang, C. C., Lee, C. F., Gu, Y. X., & Lee, Y. W. (2009). Co-determinantion of capital structure
and stock returns – A LISREL approach: An empirical test of Taiwan stock markets. Quarterly
Review of Economics and Finance, 50, 222–233.
Zellner, A. (1970). Estimation of regression relationships counting un-observable variables.
International Economic Review, 11, 441–454.
Binomial OPM, Black–Scholes OPM,
and Their Relationship: Decision Tree 37
and Microsoft Excel Approach

John C. Lee

Contents
37.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026
37.2 Call and Put Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1027
37.3 One-Period Option Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1028
37.4 Two-Period Option Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
37.5 Using Microsoft Excel to Create the Binomial Option Trees . . . . . . . . . . . . . . . . . . . . . . . . . 1036
37.6 Black–Scholes Option Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1040
37.7 Relationship Between the Binomial OPM and the Black–Scholes OPM . . . . . . . . . . . . . 1044
37.8 Decision Tree Black–Scholes Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
37.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
Appendix 1: Excel VBA Code: Binomial Option Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1059

Abstract
This chapter will first demonstrate how Microsoft Excel can be used to
create the decision trees for the binomial option pricing model. At the same
time, this chapter will discuss the binomial option pricing model in a less
mathematical fashion. All the mathematical calculations will be taken care by
the Microsoft Excel program that is presented in this chapter. Finally, this
chapter also uses the decision tree approach to demonstrate the relationship
between the binomial option pricing model and the Black–Scholes option
pricing model.

J.C. Lee
Center for PBBEF Research, North Brunswick, NJ, USA
e-mail: johnleejohnlee@yahoo.com; leeleeassociates@gmail.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1025
DOI 10.1007/978-1-4614-7750-1_37,
# Springer Science+Business Media New York 2015
1026 J.C. Lee

Keywords
Binomial option pricing model • Decision trees • Black–Scholes option pricing
model • Call option • Put option • Microsoft Excel • Visual Basic for Applica-
tions • VBA • Put-call parity • Sigma • Volatility • Recursive programming

37.1 Introduction

The binomial option pricing model derived by Rendleman and Barter (1979)
and Cox et al. (1979) is one the most famous models used to price options. Only
the Black–Scholes model (1973) is more famous. One problem with learning the
binomial option pricing model is that it is computationally intensive. This results in
a very complicated formula to price an option.
The complexity of the binomial option pricing model makes it a challenge to
learn the model. Most books teach the binomial option model by describing
the formula. This is not very effective because it usually requires the learner to
mentally keep track of many details, many times to the point of information
overload. There is a well-known principle in psychology that the average number
of things that a person can remember at one time is seven.
This chapter will first demonstrate the power of Microsoft Excel. It will do this
by demonstrating that it is possible to create large decision trees for the
binomial pricing model using Microsoft Excel. A ten-period decision tree would
require 2,047 call calculations and 2,047 put calculations. This chapter will also
show the decision tree for the price of a stock and the price of a bond, each requiring
2,047 calculations. Therefore, there would be 2,047 * 4 ¼ 8,188 calculations for
a complete set of ten-period decision trees.
Secondly, this chapter will present the binomial option model in a less mathe-
matical matter. It will try to make it so that the reader will not have to keep track of
many things at one time. It will do this by using decision trees to price call and put
options.
Finally, this chapter will show the relationship between the binomial option
pricing model and the Black–Scholes option pricing model.
This chapter uses a Microsoft Excel workbook called binomialBS_OPM.xls
that contains the VBA code to create the decision trees for the binomial option
pricing model. The VBA code is published in Appendix 1. The password for the
workbook is bigsky for those who want to study the VBA code. E-mail me at
JohnLeeExcelVBA@gmail.com and indicate that the password is “bigsky” to get
a copy of this Microsoft Excel workbook.
Section 37.2 discusses the basic concepts of call and put options. Section 37.3
demonstrates the one-period call and put option pricing models. Section 37.4
presents the two-period option pricing model. Section 37.5 demonstrates how to
use the Microsoft Excel workbook binomialBS_OPM.xls to create the decision trees
for an n-period binomial option pricing model. Section 37.6 demonstrates the use
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1027

Fig. 37.1 Value of AMZN Value of AMZN Call Option Strike Price = $200
call option 900
800
700
600

Value
500
400
300
200
100
0
0 200 300 400 500 600 700 800 900 1000
Strike Price

of the Black–Scholes model. Section 37.7 shows the relationship between the
binomial option pricing model and the Black–Scholes option pricing model.
Section 37.8 demonstrates how to use the Microsoft Excel workbook
binomialBS_OPM.xls to demonstrate the relationship between the binomial option
pricing model and the Black–Scholes option pricing model.

37.2 Call and Put Options

A call option gives the owner the right but not the obligation to buy the
underlying security at a specified price. The price in which the owner can buy the
underlying price is called the exercise price. A call option becomes valuable when
the exercise price is less than the current price of the underlying stock price.
For example, a call option on an AMZN stock with an exercise price of $200
when the stock price of an Amazon stock is $250 is worth $50. The reason
it is worth $50 is because a holder of the call option can buy the AMZN stock at
$200 and then sell the AMZN stock at the prevailing price of $250 for a profit of
$50. Also, a call option on an AMZN stock with an exercise price of $300 when the
stock price of an AMZN stock is $150 is worth $0.
A put option gives the owner the right but not the obligation to sell the
underlying security at a specified price. A put option becomes valuable when the
exercise price is more than the current price of the underlying stock price.
For example, a put option on an AMZN stock with an exercise price of $200
when the stock price of an AMZN stock is $150 is worth $50. The reason it is
worth $50 is because a holder of the put option can buy the AMZN stock at the
prevailing price of $150 and then sell the AMZN stock at the put price of $200
for a profit of $50. Also, a put option on an AMZN stock with an exercise price of
$200 when the stock price of the AMZN stock is $250 is worth $0.
Figures 37.1 and 37.2 are charts showing the value of call and put options of the
above GE stock at varying prices.
1028 J.C. Lee

Fig. 37.2 Value of AMZN Value of AMZN Put Option Strike Price = $200
put option 250

200

150

Value
100

50

0
0 50 100 150 200 250 300 350 400 450
Strike Price

37.3 One-Period Option Pricing Model

What should be the value of these options? Let’s look at a case where we are only
concerned with the value of options for one period. In the next period, a stock price
can either go up or go down. Let’s look at a case where we know for certain that an
AMZN stock with a price of $200 will either go up 5 % or go down 5 % in the next
period and the exercise after one period is $200. Figures 37.3, 37.4, and 37.5 show
the decision tree for the AMZN stock price, the AMZN call option price, and the
AMZN put option price, respectively.
Let’s first consider the issue of pricing an AMZN call option. Using a one-period
decision tree, we can illustrate the price of an AMZN stock if it goes up 5 % and the
price of a stock AMZN if it goes down 5 %. Since we know the possible ending
values of the AMZN stock, we can derive the possible ending values of a call
option. If the stock price increases to $210, the price of the AMZN call option will
then be $10 ($210  $200). If the AMZ stock price decreases to $190, the value of
the call option will worth $0 because it would be below the exercise price of $200.
We have just discussed the possible ending value of an AMZN call option in
period 1. But what we are really interested in is what the value is now of the
AMZN call option knowing the two resulting values of the AMZN call option.
To help determine the value of a one-period AMZN call option, it’s useful to
know that it is possible to replicate the resulting two states of the value of the
AMZN call option by buying a combination of stocks and bonds. Below is the
formula to replicate the situation where the price increases to $210. We will assume
that the interest rate for the bond is 3 %:
210S þ 1:03B ¼ 10
190S þ 1:03B ¼ 0

We can use simple algebra to solve for both S and B. The first thing that we need
to do is to rearrange the second equation as follows:

1:03B ¼ 190S
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1029

Fig. 37.3 AMZN stock price Period 0 Period 1

210
200
190

Period 0 Period 1
10
??
Fig. 37.4 AMZN call option
price 0

Period 0 Period 1
0
??
Fig. 37.5 AMZN put option
price 10

With the above equation, we can rewrite the first equation as

210S þ ð190SÞ ¼ 10
20S ¼ 10
S ¼ 0:5
We can solve for B by substituting the value 0.05 for S in the first equation:

210ð0:5Þ þ 1:03B ¼ 10
105 þ 1:03B ¼ 10
1:03B ¼ 95
B ¼ 92:23
Therefore, from the above simple algebraic exercise, we should at period 0 buy
0.05 shares of AMZN stock and borrow 9.223 at 3 % to replicate the payoff
of the AMZN call option. This means the value of an AMZN call option should
be 0.5 * 200  92.23 ¼ 7.77.
If this were not the case, there would then be arbitrage profits. For example, if
the call option were sold for $30, there would be a profit of 22.23. This would result
in the increase in the selling of the AMZN call option. The increase in the supply of
AMZN call options would push the price down for the call options. If the call option
were sold for $5, there would be a saving of $2.77. This saving would result in the
increase demand for the AMZN call option. This increase demand would result in
the price of the call option to increase. The equilibrium point would be $7.77.
1030 J.C. Lee

Using the above mentioned concept and procedure, Benninga (2000) has derived
a one-period call option model as

C ¼ qu Max½Sð1 þ uÞX, 0 þ qd Max½Sð1 þ d Þ  X, 0 (37.1)

where
id
qu ¼
ð1 þ i Þðu  d Þ
ui
qd ¼
ð1 þ i Þðu  d Þ
u ¼ increase factor
d ¼ down factor
i ¼ interest rate
If we let i ¼ r, p ¼ (r  d)/(u  d), 1  p ¼ (u  r)/(u  d), R ¼ 1/(1 + r),
Cu ¼ Max[S(1 + u)  X, 0], and Cd ¼ Max[S(1 + d)  X, 0], then we have

C ¼ ½pCu þ ð1  pÞCd =R, (37.2)

where
Cu ¼ call option price after increase
Cd ¼ call option price after decrease
Equation 37.2 is identical to Eq. 6B.6 in Lee et al. (2000, p. 234).1
Below calculates the value of the above one-period call option where the strike
price, X, is $200 and the risk-free interest rate is 3 %. We will assume that the price
of a stock for any given period will either increase or decrease by 5 %:

X ¼ $200
S ¼ $200
u ¼ 1:05
d ¼ 0:95
R ¼ 1 þ r ¼ 1 þ 0:03
p ¼ ð1:03  0:95Þ=ð1:05  0:95Þ
C ¼ ½0:8ð10Þ þ 0:2ð0Þ=1:03 ¼ $7:77

Therefore, from the above calculations, the value of the call option is $7.77.
Figure 37.6 shows the resulting decision tree for the above call option.

1
Please note that in Lee et al. (2000, p. 234) u ¼ 1 + percentage of price increase and
d ¼ 1 – percentage of price increase.
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1031

Fig. 37.6 Call option price Period 0 Period 1

10.000
7.770
0

Like the call option, it is possible to replicate the resulting two states of the value
of the put option by buying a combination of stocks and bonds. Below is the
formula to replicate the situation where the price decreases to $190:
210S þ 1:03B ¼ 0
190S þ 1:03B ¼ 10

We will use simple algebra to solve for both S and B. The first thing we will do is
to rewrite the second equation as follows:

1:03B ¼ 10  190S

The next thing to do is to substitute the above equation to the first put option
equation. Doing this would result in the following:

210S þ 10  190S ¼ 0

The following solves for S:

20S ¼ 10
S ¼ 0:5

Now let us solve for B by putting the value of S into the first equation. This is
shown below:
210ð0:5Þ þ 1:03B ¼ 0
1:03B ¼ 105
B ¼ 101:94

From the above simple algebra exercise, we have S ¼ 0.5 and B ¼ 101.94.
This tells us that we should in period 0 lend $101.94 at 3 % and sell 0.5 shares of
stock to replicate the put option payoff for period 1. And the value of the AMZN put
option should be 200(0.5) + 101.94 ¼ 1.94.
Using the same arbitrage argument that we used in the discussion of the call
option, 0.194 has to be the equilibrium price of the put option.
As with the call option, Benninga (2000) has derived a one-period put option
model as

P ¼ qu Max½X  Sð1 þ uÞ, 0 þ qd Max½X  Sð1 þ dÞ, 0 (37.3)

1032 J.C. Lee

where

id
qu ¼
ð1 þ i Þðu  d Þ
ui
qd ¼
ð1 þ i Þðu  d Þ
u ¼ increase factor
d ¼ down factor
i ¼ interest rate

If we let i ¼ r, p ¼ (r  d)/(u  d), 1  p ¼ (u  r)/(u  d), R ¼ 1/(1 + r),

Pu ¼ Max[X  S(1 + u), 0], and Pd ¼ Max[X  S(1 + d), 0], then we have

P ¼ ½pPu þ ð1  pÞPd =R, (37.4)

where
Pu ¼ put option price after increase
Pd ¼ put option price after decrease
Below calculates the value of the above one-period put option where the strike
price, X, is $20 and the risk-free interest rate is 3 %:

P ¼ ½0:8ð0Þ þ 0:2ð10Þ=1:03 ¼ $1:94

From the above calculation, the put option pricing decision tree would look like
the following.
Figure 37.7 shows the resulting decision tree for the above put option.
There is a relationship between the price of a put option and the price of all call
option. This relationship is called the put-call parity. Equation 37.5 shows the
relationship between the price of a put option and the price of a call option:

P ¼ C þ X=R  S (37.5)

where
C ¼ call price
X ¼ strike price
R ¼ 1 + interest rate
S ¼ stock price

Period 0 Period 1
0
1.940
Fig. 37.7 AMZN put option
10
price
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1033

The following uses the put-call parity to calculate the price of the AMZN put
option:
P ¼ $7:77 þ $200=ð1:03Þ  $200
¼ 7:77 þ 194:17  200
¼ 1:94

37.4 Two-Period Option Pricing Model

We now will look at pricing options for two periods. Figure 37.8 shows the
stock price decision tree based on the parameters indicated in the last section.
This decision tree was created based on the assumption that a stock price will either
increase by 5 % or decrease by 5 %.
How do we price the value of a call and put option for two periods?
The highest possible value for our stock based on our assumption is $220.5.
We get this value first by multiplying the stock price at period 0 by 105 % to get the
resulting value of $210 of period 1. We then again multiply the stock price in
period 1 by 105 % to get the resulting value of $220.5. In period two, the value
of a call option when a stock price is $220.5 is the stock price minus the exercise
price, $220.5  $200, or $20.5. In period two, the value of a put option when a stock
price is $220.5 is the exercise price minus the stock price, $200  $220.5,
or $20.5. A negative value has no value to an investor so the value of the put option
would be $0.
The lowest possible value for our stock based on our assumptions is $180.5.
We get this value first by multiplying the stock price at period 0 by 95 %
(decreasing the value of the stock by 5 %) to get the resulting value of $190 of
period 1. We then again multiply the stock price in period 1 by 95 % to get the
resulting value of $180.5. In period two, the value of a call option when a stock
price is $180.5 is the stock price minus the exercise price, $180.5  $200, or
$19.5. A negative value has no value to an investor so the value of a call option
would be $0. In period two, the value of a put option when a stock price is $18.05 is
the exercise price minus the stock price, $200  $180.5, or $19.5. We can derive the
call and put option values for the other possible values of the stock in period 2 in
the same fashion.
Figures 37.9 and 37.10 show the possible call and put option values for period 2.
We cannot calculate the value of the call and put options in period 1 the
same way we did in period 2 because it’s not the ending value of the stock.
In period 1, there are two possible call values. One value is when the stock price
increased, and one value is when the stock price decreased. The call option decision
tree shown in Fig. 37.9 shows two possible values for a call option in period 1. If we
just focus on the value of a call option when the stock price increases from period
one, we will notice that it is like the decision tree for a call option for one period.
This is shown in Fig. 37.11.
1034 J.C. Lee

Fig. 37.8 AMZN stock price Period 0 Period 1 Period 2

220.5
210.0
199.5
200.0
199.5
190.0
180.5

Period 0 Period 1 Period 2

20.5

0.0

0.0

Fig. 37.9 AMZN call option 0.0

Period 0 Period 1 Period 2

0.0

0.5

0.5

Fig. 37.10 AMZN put

option 19.5

Period 0 Period 1 Period 2

20.5000

0
Fig. 37.11 AMZN call
option
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1035

Using the same method for pricing a call option for one period, the price of a call
option when stock price increases from period 0 will be $15.922. The resulting
decision tree is shown in Fig. 37.12.
In the same fashion, we can price the value of a call option when a stock
price decreases. The price of a call option when a stock price decreases from
period 0 is $0. The resulting decision tree is shown in Fig. 37.13.
In the same fashion, we can price the value of a call option in period 0.
The resulting decision tree is shown in Fig. 37.14.

Fig. 37.12 AMZN call Period 0 Period 1 Period 2

option
20.5000
15.9220
0

Period 0 Period 1 Period 2

20.5000
15.9220
0

0
0
0
Fig. 37.13 AMZN call
option

Period 0 Period 1 Period 2

20.5000
15.9220
0.0000
12.3670
0.0000
0.0000
0.0000
Fig. 37.14 AMZN call
option
1036 J.C. Lee

Fig. 37.15 AMZN put Period 0 Period 1 Period 2

option
0.0000
0.0970
0.5000
0.8860
0.5000
4.1750
19.5000

Table 37.1 Periods Calculations

1 3
2 7
3 17
4 31
5 63
6 127
7 255
8 511
9 1,023
10 2,047
11 4,065
12 8,191

We can calculate the value of a put option in the same manner as we did
in calculating the value of a call option. The decision tree for a put option is
shown in Fig. 37.15.

37.5 Using Microsoft Excel to Create the Binomial Option Trees

In the previous section, we priced the value of a call and put option by pricing
backwards, from the last period to the first period. This method of pricing call and
put options will work for any n-period. To price the value of a call option for
two periods required seven sets of calculations. The number of calculations
increases dramatically as n increases. Table 37.1 lists the number of calculations
for specific number of periods.
After two periods, it becomes very cumbersome to calculate and create the
decision trees for a call and put option. In the previous section, we saw that
calculations were very repetitive and mechanical. To solve this problem,
this chapter will use Microsoft Excel to do the calculations and create the
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1037

decision trees for the call and put options. We will also use Microsoft Excel to
calculate and draw the related decision trees for the underlying stock and bond.
To solve this repetitive and mechanical calculation of the binomial option
pricing model, we will look at a Microsoft Excel file called binomialBS_OPM.
xls. We will use this Microsoft Excel workbook to produce four decision trees
for the GE stock that was discussed in the previous sections. The four decision
trees are:
1. Stock price
2. Call option price
3. Put option price
4. Bond price
This section will demonstrate how to use the binomialBS_OPM.xls Excel file to
create the four decision trees. Figure 37.16 shows the Excel file binomialBS_OPM.
xls after the file is opened. Pushing the button shown in Fig. 37.16 will get the
dialog box shown in Fig. 37.17.
The dialog box shown in Fig. 37.17 shows the parameters for the binomial
option pricing model. These parameters are changeable. The dialog box in
Fig. 37.17 shows the default values.

Fig. 37.16 Excel file BinomialBS_OPM.xls

1038 J.C. Lee

Fig. 37.17 Dialog box

showing parameters for the
binomial option pricing
model

Pushing the calculate button shown in Fig. 37.17 will produce the four decision
trees shown in Figs. 37.18, 37.19, 37.20, and 37.21.
The table at the beginning of this section indicated 31 calculations were required
to create a decision tree that has four periods. This section showed four decision
trees. Therefore, the Excel file did 31 * 4 ¼ 121 calculations to create the four
decision trees.
Benninga (2000, p. 260) has defined the price of a call option in a binomial
option pricing model with n-periods as
Xn   h i
n i ni
C¼ qu qd max Sð1 þ uÞi ð1 þ dÞni , 0 (37.6)
i¼0
i

and the price of a put option in a binomial option pricing model with n-periods as
X n   h i
n i ni
P¼ qu qd max X  Sð1 þ uÞi ð1 þ dÞni , 0 (37.7)
i¼0
i

Lee et al. (2000, p. 237) have defined the pricing of a call option in a binomial
option pricing model with n-period as
1 X n
n! h i
C¼ n pk ð1  pÞnk max 0, ð1 þ uÞk ð1 þ dÞnk , S  X (37.8)
R k¼0 k!ðn  k!Þ
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1039

Fig. 37.18 Stock price Stock Price

decision tree Decision Tree
Price = 200, Exercise = 200, U = 1.0500, D = 0.9500, N = 4, R = 0.03
Number of calculations: 31
243.1012
231.5250
219.9487
220.5000
219.9487
209.4750
199.0012
210.0000
219.9487
209.4750
199.0012
199.5000
199.0012
189.5250
180.0487
200.0000
219.9487
209.4750
199.0012
199.5000
199.0012
189.5250
180.0487
190.0000
199.0012
189.5250
180.0487
180.5000
180.0487
171.4750
162.9012

The definition of the pricing of a put option in a binomial option pricing model
with n-period would then be defined as
1 X n
n! h i
P¼ n pk ð1  pÞnk max 0, X  ð1 þ uÞk ð1 þ dÞnk , S (37.9)
R k¼0 k!ðn  kÞ!
1040 J.C. Lee

Fig. 37.19 Call option Call Option Pricing

pricing decision tree Decision Tree
Price = 200, Exercise = 200, U = 1.0500, D = 0.9500, N = 4, R = 0.03
Number of calculations: 31
Binomial Call Price = 22.9454
43.1012
37.3502
19.9487
32.0184
19.9487
15.4941
0.0000
27.2054
19.9487
15.4941
0.0000
12.0343
0.0000
0.0000
0.0000
22.9454
19.9487
15.4941
0.0000
12.0343
0.0000
0.0000
0.0000
9.3470
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000

37.6 Black–Scholes Option Pricing Model

The most famous option pricing model is the Black–Scholes option pricing model.
In this section, we will demonstrate the usage of the Black–Scholes option pricing
model. In latter sections, we will demonstrate the relationship between the binomial
option pricing model and the Black–Scholes pricing model. The Black–Scholes
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1041

Fig. 37.20 Put option Put Option Pricing

pricing decision tree Decision Tree
Price = 200, Exercise = 200, U = 1.0500, D = 0.9500, N = 4, R = 0.03
Number of calculations: 31
Binomial Put Price: 0.6428
0.0000
0.0000
0.0000
0.0377
0.0000
0.1939
0.9988
0.2338
0.0000
0.1939
0.9988
1.0535
0.9988
4.6498
19.9513
0.6428
0.0000
0.1939
0.9988
1.0535
0.9988
4.6498
19.9513
2.3754
0.9988
4.6498
19.9513
8.0192
19.9513
22.6998
37.0988
1042 J.C. Lee

Fig. 37.21 Bond pricing Bond Pricing

decision tree Decision Tree
Price = 200, Exercise = 200, U = 1.0500, D = 0.9500, N = 4, R = 0.03
Number of calculations: 31
1.1255
1.0927
1.1255
1.0609
1.1255
1.0927
1.1255
1.0300
1.1255
1.0927
1.1255
1.0609
1.1255
1.0927
1.1255
1.0000
1.1255
1.0927
1.1255
1.0609
1.1255
1.0927
1.1255
1.0300
1.1255
1.0927
1.1255
1.0609
1.1255
1.0927
1.1255
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1043

model prices European call and put options. The Black–Scholes model for
a European call option is

C ¼ SNðd1Þ  XerT Nðd2Þ (37.10)

where
C ¼ call price
S ¼ stock price
r ¼ risk-free interest rate
T ¼ time to maturity of option in years
N ¼ standard normal distribution
s ¼ stock volatility
 
s2
lnðS=XÞ þ r þ T
2
d1 ¼ pffiffiffi
s T
pffiffiffi
d2 ¼ d1  s T

Let’s manually calculate the price of a European call option in terms of Eq. 37.10
with the following parameter values, S ¼ 200, X ¼ 200, r ¼ 3 %, T ¼ 4, and
s ¼ 20 %.
Solution
   
s2 :22
lnðS=XÞ þ r þ T lnð200=200Þ þ :03 þ ð4Þ
2 2 ð:03 þ :02Þ  4 :2
d1 ¼ pffiffiffi ¼ pffiffiffi ¼ ¼ ¼ :5,
s T :2 4 :4 :4
pffiffiffi
d2 ¼ :5  :2 4 ¼ :1

Nðd1Þ ¼ 0:69146, Nðd2Þ ¼ 0:5398, erT ¼ 0:8869

C ¼ ð200Þ  ð0:69146Þ  ð200Þ  ð0:8869Þ  0:5398
¼ 138:292  95:74972 ¼ 42:5422

The Black–Scholes put-call parity equation is

P ¼ C  S þ XerT

The put option value for the stock would be

P ¼ 42:54  200 þ 200ð0:8869Þ

¼ 42:54  200 þ 177:38 ¼ 19:92
1044 J.C. Lee

37.7 Relationship Between the Binomial OPM

and the Black–Scholes OPM

We can use either the binomial model or Black–Scholes to price an option. They
both should result in similar numbers. If we look at the parameters in both models,
we will notice that the binomial model has an Increase Factor (U), a Decrease
Factor (D), and n-period parameters that the Black–Scholes model does not have.
We also notice that the Black–Scholes model has the s and T parameters that the
binomial model does not have. Benninga (2008) suggests the following translation
between the binomial and Black–Scholes parameters:
pffiffiffiffi pffiffiffiffi
Dt ¼ T=n R ¼ erDt U ¼ es Dt D ¼ es Dt

In the Excel program, shown in Appendix 1, we use Benninga’s (2008) Increase

Factor and Decrease Factor definitions. They are defined as follows:

RD UR
qU ¼ , qD ¼
RðU  DÞ RðU  DÞ

where
U ¼ 1 + percentage of price increase
D ¼ 1  percentage of price increase
R ¼ 1 + interest rate

Fig. 37.22 Dialog box

showing parameters for the
binomial option pricing
model
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1045

Call Option Pricing

Decision Tree
Price = 200, Exercise = 200, U = 1.2214, D = 0.8187, N = 4, R = 0.03
Number of calculations: 31
Binomial Call Price = 40.6705
Black-Scholes Call Price = 42.5356, d1 = 0.5000, d2 = 0.1000, N(d1) = 0.6915, N(d2) = 0.5398
245.1082
170.3347
98.3650
110.0120
98.3650
50.1915
0.0000
67.9201
98.3650
50.1915
0.0000
25.6106
0.0000
0.0000
0.0000
40.6705
98.3650
50.1915
0.0000
25.6106
0.0000
0.0000
0.0000
13.0680
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000

Fig. 37.23 Decision tree approximation of Black–Scholes call pricing

1046 J.C. Lee

Fig. 37.24 Decision Put Option Pricing

tree approximation of Decision Tree
Black–Scholes put pricing Price = 200, Exercise = 200, U = 1.2214, D = 0.8187, N = 4, R = 0.03
Number of calculations: 31
Binomial Put Price: 18.0546
Black-Scholes Put Price: 19.9197
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
6.4258
0.0000
0.0000
0.0000
13.9635
0.0000
30.3430
65.9360
18.0546
0.0000
0.0000
0.0000
13.9635
0.0000
30.3430
65.9360
32.1081
0.0000
30.3430
65.9360
54.2889
65.9360
84.3268
110.1342
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1047

37.8 Decision Tree Black–Scholes Calculation

We will now use the BinomialBS_OPM.xls Excel file to calculate the binomial and
Black–Scholes call and put values illustrated in Sect. 37.5. Notice that in Fig. 37.22
the Binomial Black–Scholes Approximation check box is checked. Checking this
box will cause T and Sigma parameters to appear and will adjust the Increase
Factor – u and Decrease Factor – d parameters. The adjustment was done as
indicated in Sect. 37.7.
Notice in Figs. 37.23 and 37.24 that the binomial option pricing model value
does not agree with the Black–Scholes option pricing model. The binomial OPM
value will get very close to the Black–Scholes OPM value once the binomial
parameter n gets very large. Benninga (2008) demonstrated that the
binomial value will be close to the Black–Scholes when the binomial n parameter
gets larger than 500.

37.9 Summary

This chapter demonstrated, with the aid of Microsoft Excel and decision trees, the
binomial option model in a less mathematical fashion. This chapter allowed the
reader to focus more on the concepts by studying the associated decision trees,
which were created by Microsoft Excel. This chapter also demonstrates that using
Microsoft Excel releases the reader from the computation burden of the binomial
option model.
This chapter also published the Microsoft Excel Visual Basic for Application
(VBA) code that created the binomial option decision trees. This allows for those
who are interested in studying the many advance Microsoft Excel VBA program-
ming concepts that were used to create the decision trees. One major computer
science programming concept used by the Excel VBA program in this chapter is
recursive programming. Recursive programming is the ideal of a procedure
calling itself many times. Inside the procedure, there are statements to decide
when not to call itself.
This chapter also used decision trees to demonstrate the relationship between the
binomial option pricing model and the Black–Scholes option pricing model.

Appendix 1: Excel VBA Code: Binomial Option Pricing Model

It is important to note that the thing that makes Microsoft Excel powerful is
that it offers a powerful professional programming language called Visual
Basic for Applications (VBA). This section shows the VBA code that
1048 J.C. Lee

generated the decision trees for the binomial option pricing model. This code is
in the form frmBinomiaOption. The procedure cmdCalculate_Click is the first
procedure to run.

’/**************************************************
’/ Relationship Between the Binomial OPM
’/ and Black-Scholes OPM:
’/ Decision Tree and Microsoft Excel Approach
’/
’/ by John Lee
’/ JohnLeeExcelVBA@gmail.com
’/ All Rights Reserved
’/**************************************************
Option Explicit
Dim mwbTreeWorkbook As Workbook
Dim mwsTreeWorksheet As Worksheet
Dim mwsCallTree As Worksheet
Dim mwsPutTree As Worksheet
Dim mwsBondTree As Worksheet
Dim mdblPFactor As Double
Dim mBinomialCalc As Long
Dim mCallPrice As Double ’jcl 12/8/2008
Dim mPutPrice As Double ’jcl 12/8/2008
’/**************************************************
’/Purpose: Keep track the numbers of binomial calc
’/**************************************************
Property Let BinomialCalc(l As Long)
mBinomialCalc ¼ l
End Property
Property Get BinomialCalc() As Long
BinomialCalc ¼ mBinomialCalc
End Property
Property Set TreeWorkbook(wb As Workbook)
Set mwbTreeWorkbook ¼ wb
End Property
Property Get TreeWorkbook() As Workbook
Set TreeWorkbook ¼ mwbTreeWorkbook
End Property
Property Set TreeWorksheet(ws As Worksheet)
Set mwsTreeWorksheet ¼ ws
End Property
Property Get TreeWorksheet() As Worksheet
Set TreeWorksheet ¼ mwsTreeWorksheet
End Property
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1049

Property Set CallTree(ws As Worksheet)

Set mwsCallTree ¼ ws
End Property
Property Get CallTree() As Worksheet
Set CallTree ¼ mwsCallTree
End Property
Property Set PutTree(ws As Worksheet)
Set mwsPutTree ¼ ws
End Property
Property Get PutTree() As Worksheet
Set PutTree ¼ mwsPutTree
End Property
Property Set BondTree(ws As Worksheet)
Set mwsBondTree ¼ ws
End Property
Property Get BondTree() As Worksheet
Set BondTree ¼ mwsBondTree
End Property
Property Let CallPrice(dCallPrice As Double)
’12/8/2008
mCallPrice ¼ dCallPrice
End Property
Property Get CallPrice() As Double
Let CallPrice ¼ mCallPrice
End Property
Property Let PutPrice(dPutPrice As Double)
’12/10/2008
mPutPrice ¼ dPutPrice
End Property
Property Get PutPrice() As Double
’12/10/2008
Let PutPrice ¼ mPutPrice
End Property
Property Let PFactor(r As Double)
Dim dRate As Double
dRate ¼ ((1 + r) - Me.txtBinomialD) / (Me.txtBinomialU -
Me.txtBinomialD)
Let mdblPFactor ¼ dRate
End Property
Property Get PFactor() As Double
Let PFactor ¼ mdblPFactor
End Property
Property Get qU() As Double
Dim dblDeltaT As Double
1050 J.C. Lee

Dim dblDown As Double

Dim dblUp As Double
Dim dblR As Double
dblDeltaT ¼ Me.txtTimeT / Me.txtBinomialN
dblR ¼ Exp(Me.txtBinomialr * dblDeltaT)
dblUp ¼ Exp(Me.txtSigma * VBA.Sqr(dblDeltaT))
dblDown ¼ Exp(-Me.txtSigma * VBA.Sqr(dblDeltaT))
qU ¼ (dblR - dblDown) / (dblR * (dblUp - dblDown))
End Property
Property Get qD() As Double
Dim dblDeltaT As Double
Dim dblDown As Double
Dim dblUp As Double
Dim dblR As Double
dblDeltaT ¼ Me.txtTimeT / Me.txtBinomialN
dblR ¼ Exp(Me.txtBinomialr * dblDeltaT)
dblUp ¼ Exp(Me.txtSigma * VBA.Sqr(dblDeltaT))
dblDown ¼ Exp(-Me.txtSigma * VBA.Sqr(dblDeltaT))
qD ¼ (dblUp - dblR) / (dblR * (dblUp - dblDown))
End Property
Private Sub chkBinomialBSApproximation_Click()
On Error Resume Next
’Time and Sigma only BlackScholes parameter
Me.txtTimeT.Visible ¼ Me.chkBinomialBSApproximation
Me.lblTimeT.Visible ¼ Me.chkBinomialBSApproximation
Me.txtSigma.Visible ¼ Me.chkBinomialBSApproximation
Me.lblSigma.Visible ¼ Me.chkBinomialBSApproximation
txtTimeT_Change
End Sub
Private Sub cmdCalculate_Click()
Me.Hide
BinomialOption
Unload Me
End Sub
Private Sub cmdCancel_Click()
Unload Me
End Sub
Private Sub txtBinomialN_Change()
’jcl 12/8/2008
On Error Resume Next
If Me.chkBinomialBSApproximation Then
Me.txtBinomialU ¼ Exp(Me.txtSigma * Sqr(Me.txtTimeT /
Me.txtBinomialN))
Me.txtBinomialD ¼ Exp(-Me.txtSigma * Sqr(Me.txtTimeT /
Me.txtBinomialN))
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1051

End If
End Sub
Private Sub txtTimeT_Change()
’jcl 12/8/2008
On Error Resume Next
If Me.chkBinomialBSApproximation Then
Me.txtBinomialU ¼ Exp(Me.txtSigma * Sqr(Me.txtTimeT /
Me.txtBinomialN))
Me.txtBinomialD ¼ Exp(-Me.txtSigma * Sqr(Me.txtTimeT /
Me.txtBinomialN))
End If
End Sub
Private Sub UserForm_Initialize()
With Me
.txtBinomialS ¼ 20
.txtBinomialX ¼ 20
.txtBinomialD ¼ 0.95
.txtBinomialU ¼ 1.05
.txtBinomialN ¼ 4
.txtBinomialr ¼ 0.03
.txtSigma ¼ 0.2
.txtTimeT ¼ 4
Me.chkBinomialBSApproximation ¼ False
End With
chkBinomialBSApproximation_Click
Me.Hide
End Sub
Sub BinomialOption()
Dim wbTree As Workbook
Dim wsTree As Worksheet
Dim rColumn As Range
Dim ws As Worksheet
Set Me.TreeWorkbook ¼ Workbooks.Add
Set Me.BondTree ¼ Me.TreeWorkbook.Worksheets.Add
Set Me.PutTree ¼ Me.TreeWorkbook.Worksheets.Add
Set Me.CallTree ¼ Me.TreeWorkbook.Worksheets.Add
Set Me.TreeWorksheet ¼ Me.TreeWorkbook.Worksheets.Add
Set rColumn ¼ Me.TreeWorksheet.Range("a1")
With Me
.BinomialCalc ¼ 0
.PFactor ¼ Me.txtBinomialr
.CallTree.Name ¼ "Call Option Price"
.PutTree.Name ¼ "Put Option Price"
.TreeWorksheet.Name ¼ "Stock Price"
.BondTree.Name ¼ "Bond"
1052 J.C. Lee

End With
DecisionTree rCell:¼rColumn, nPeriod:¼Me.txtBinomialN
+ 1, _
dblPrice:¼Me.txtBinomialS, sngU:¼Me.txtBinomialU, _
sngD:¼Me.txtBinomialD
DecitionTreeFormat
TreeTitle wsTree:¼Me.TreeWorksheet, sTitle:¼"Stock
Price "
TreeTitle wsTree:¼Me.CallTree, sTitle:¼"Call Option
Pricing"
TreeTitle wsTree:¼Me.PutTree, sTitle:¼"Put Option
Pricing"
TreeTitle wsTree:¼Me.BondTree, sTitle:¼"Bond Pricing"
Application.DisplayAlerts ¼ False
For Each ws In Me.TreeWorkbook.Worksheets
If Left(ws.Name, 5) ¼ "Sheet" Then
ws.Delete
Else
ws.Activate
ActiveWindow.DisplayGridlines ¼ False
ws.UsedRange.NumberFormat ¼ "#,##0.0000_);(#,##0.0000)"
End If
Next
Application.DisplayAlerts ¼ True
Me.TreeWorksheet.Activate
End Sub
Sub TreeTitle(wsTree As Worksheet, sTitle As String)
wsTree.Range("A1:A5").EntireRow.Insert (xlShiftDown)
With wsTree
With.Cells(1)
.Value ¼ sTitle
.Font.Size ¼ 20
.Font.Italic ¼ True
End With
With.Cells(2, 1)
.Value ¼ "Decision Tree"
.Font.Size ¼ 16
.Font.Italic ¼ True
End With
With.Cells(3, 1)
.Value ¼ "Price ¼ " & Me.txtBinomialS & _
",Exercise ¼ " & Me.txtBinomialX & _
",U ¼ " & Format(Me.txtBinomialU, "#,##0.0000") & _
",D ¼ " & Format(Me.txtBinomialD, "#,##0.0000") & _
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1053

",N ¼ " & Me.txtBinomialN & _

",R ¼ " & Me.txtBinomialr
.Font.Size ¼ 14
End With
With.Cells(4, 1)
.Value ¼ "Number of calculations: " & Me.BinomialCalc
.Font.Size ¼ 14
End With
If wsTree Is Me.CallTree Then
With.Cells(5, 1)
.Value ¼ "Binomial Call Price¼ " & Format(Me.CallPrice,
"#,##0.0000")
.Font.Size ¼ 14
End With
If Me.chkBinomialBSApproximation Then
wsTree.Range("A6:A7").EntireRow.Insert (xlShiftDown)
With.Cells(6, 1)
.Value ¼ "Black-Scholes Call Price¼ " & Format(Me.
BS_Call, "#,##0.0000") _
& ",d1¼" & Format(Me.BS_D1, "#,##0.0000") _
& ",d2¼" & Format(Me.BS_D2, "#,##0.0000") _
& ",N(d1)¼" & Format(WorksheetFunction.NormSDist
(BS_D1), "#,##0.0000") _
& ",N(d2)¼" & Format(WorksheetFunction.NormSDist
(BS_D2), "#,##0.0000")
.Font.Size ¼ 14
End With
End If
ElseIf wsTree Is Me.PutTree Then
With.Cells(5, 1)
.Value ¼ "Binomial Put Price: " & Format(Me.PutPrice,
"#,##0.0000")
.Font.Size ¼ 14
End With
If Me.chkBinomialBSApproximation Then
wsTree.Range("A6:A7").EntireRow.Insert (xlShiftDown)
With.Cells(6, 1)
.Value ¼ "Black-Scholes Put Price: " & Format(Me.BS_PUT,
"#,##0.0000")
.Font.Size ¼ 14
End With
End If
End If
End With
1054 J.C. Lee

End Sub
Sub BondDecisionTree(rPrice As Range, arCell As Variant,
iCount As Long)
Dim rBond As Range
Dim rPup As Range
Dim rPDown As Range
Set rBond ¼ Me.BondTree.Cells(rPrice.Row, rPrice.
Column)
Set rPup ¼ Me.BondTree.Cells(arCell(iCount - 1).Row,
arCell(iCount - 1).Column)
Set rPDown ¼ Me.BondTree.Cells(arCell(iCount).Row,
arCell(iCount).Column)
If rPup.Column ¼ Me.TreeWorksheet.UsedRange.Columns.
Count Then
rPup.Value ¼ (1 + Me.txtBinomialr) ^ (rPup.Column - 1)
rPDown.Value ¼ rPup.Value
End If
With rBond
.Value ¼ (1 + Me.txtBinomialr) ^ (rBond.Column - 1)
.Borders(xlBottom).LineStyle ¼ xlContinuous
End With
rPDown.Borders(xlBottom).LineStyle ¼ xlContinuous
With rPup
.Borders(xlBottom).LineStyle ¼ xlContinuous
.Offset(1, 0).Resize((rPDown.Row - rPup.Row), 1). _
Borders(xlEdgeLeft).LineStyle ¼ xlContinuous
End With
End Sub
Sub PutDecisionTree(rPrice As Range, arCell As Variant,
iCount As Long)
Dim rCall As Range
Dim rPup As Range
Dim rPDown As Range
Set rCall ¼ Me.PutTree.Cells(rPrice.Row, rPrice.Column)
Set rPup ¼ Me.PutTree.Cells(arCell(iCount - 1).Row,
arCell(iCount - 1).Column)
Set rPDown ¼ Me.PutTree.Cells(arCell(iCount).Row,
arCell(iCount).Column)
If rPup.Column ¼ Me.TreeWorksheet.UsedRange.Columns.
Count Then
rPup.Value ¼ WorksheetFunction.Max(Me.txtBinomialX -
arCell(iCount - 1), 0)
rPDown.Value ¼ WorksheetFunction.Max(Me.txtBinomialX -
arCell(iCount), 0)
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1055

End If
With rCall
’12/10/2008
If Not Me.chkBinomialBSApproximation Then
.Value ¼ (Me.PFactor * rPup + (1 - Me.PFactor) * rPDown) /
(1 + Me.txtBinomialr)
Else
.Value ¼ (Me.qU * rPup) + (Me.qD * rPDown)
End If
Me.PutPrice ¼.Value ’12/8/2008
.Borders(xlBottom).LineStyle ¼ xlContinuous
End With
rPDown.Borders(xlBottom).LineStyle ¼ xlContinuous
With rPup
.Borders(xlBottom).LineStyle ¼ xlContinuous
.Offset(1, 0).Resize((rPDown.Row - rPup.Row), 1). _
Borders(xlEdgeLeft).LineStyle ¼ xlContinuous
End With
End Sub
Sub CallDecisionTree(rPrice As Range, arCell As Variant,
iCount As Long)
Dim rCall As Range
Dim rCup As Range
Dim rCDown As Range
Set rCall ¼ Me.CallTree.Cells(rPrice.Row, rPrice.
Column)
Set rCup ¼ Me.CallTree.Cells(arCell(iCount - 1).Row,
arCell(iCount - 1).Column)
Set rCDown ¼ Me.CallTree.Cells(arCell(iCount).Row,
arCell(iCount).Column)
If rCup.Column ¼ Me.TreeWorksheet.UsedRange.Columns.
Count Then
With rCup
.Value ¼ WorksheetFunction.Max(arCell(iCount - 1) - Me.
txtBinomialX, 0)
.Borders(xlBottom).LineStyle ¼ xlContinuous
End With
With rCDown
.Value ¼ WorksheetFunction.Max(arCell(iCount) - Me.
txtBinomialX, 0)
.Borders(xlBottom).LineStyle ¼ xlContinuous
End With
End If
With rCall
1056 J.C. Lee

If Not Me.chkBinomialBSApproximation Then

.Value ¼ (Me.PFactor * rCup + (1 - Me.PFactor) * rCDown) /
(1 + Me.txtBinomialr)
Else
.Value ¼ (Me.qU * rCup) + (Me.qD * rCDown)
End If
Me.CallPrice ¼.Value ’12/8/2008
.Borders(xlBottom).LineStyle ¼ xlContinuous
End With
rCup.Offset(1, 0).Resize((rCDown.Row - rCup.Row), 1). _
Borders(xlEdgeLeft).LineStyle ¼ xlContinuous
End Sub
Sub DecitionTreeFormat()
Dim rTree As Range
Dim nColumns As Integer
Dim rLast As Range
Dim rCell As Range
Dim lCount As Long
Dim lCellSize As Long
Dim vntColumn As Variant
Dim iCount As Long
Dim lTimes As Long
Dim arCell() As Range
Dim sFormatColumn As String
Dim rPrice As Range
Application.StatusBar ¼ "Formatting Tree.. "
Set rTree ¼ Me.TreeWorksheet.UsedRange
nColumns ¼ rTree.Columns.Count
Set rLast ¼ rTree.Columns(nColumns).EntireColumn.
SpecialCells(xlCellTypeConstants, 23)
lCellSize ¼ rLast.Cells.Count
For lCount ¼ nColumns To 2 Step -1
sFormatColumn ¼ rLast.Parent.Columns(lCount).
EntireColumn.Address
Application.StatusBar ¼ "Formatting column " &
sFormatColumn
ReDim vntColumn(1 To (rLast.Cells.Count / 2), 1)
Application.StatusBar ¼ "Assigning values to array for
column " & _
rLast.Parent.Columns(lCount).EntireColumn.Address
vntColumn ¼ rLast.Offset(0, -1).EntireColumn.Cells(1).
Resize(rLast.Cells.Count / 2, 1)
rLast.Offset(0, -1).EntireColumn.ClearContents
ReDim arCell(1 To rLast.Cells.Count)
lTimes ¼ 1
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1057

Application.StatusBar ¼ "Assigning cells to arrays.

Total number of cells: " & lCellSize
For Each rCell In rLast.Cells
Application.StatusBar ¼ "Array to column " &
sFormatColumn & " Cells " & rCell.Row
Set arCell(lTimes) ¼ rCell
lTimes ¼ lTimes + 1
Next
lTimes ¼ 1
Application.StatusBar ¼ "Formatting leaves for column "
& sFormatColumn
For iCount ¼ 2 To lCellSize Step 2
Application.StatusBar ¼ "Formatting leaves for cell "
& arCell(iCount).Address
If rLast.Cells.Count <> 2 Then
Set rPrice ¼ arCell(iCount).Offset(-1 * ((arCell
(iCount).Row - arCell(iCount -1).Row) / 2), -1)
rPrice.Value ¼ vntColumn(lTimes, 1)
Else
Set rPrice ¼ arCell(iCount).Offset(-1 * ((arCell
(iCount).Row - arCell(iCount -1).Row) / 2), -1)
rPrice.Value ¼ vntColumn
End If
arCell(iCount).Borders(xlBottom).
LineStyle ¼ xlContinuous
With arCell(iCount - 1)
.Borders(xlBottom).LineStyle ¼ xlContinuous
.Offset(1, 0).Resize((arCell(iCount).Row - arCell
(iCount - 1).Row), 1). _
Borders(xlEdgeLeft).LineStyle ¼ xlContinuous
End With
lTimes ¼ 1 + lTimes
CallDecisionTree rPrice:¼rPrice, arCell:¼arCell,
iCount:¼iCount
PutDecisionTree rPrice:¼rPrice, arCell:¼arCell,
iCount:¼iCount
BondDecisionTree rPrice:¼rPrice, arCell:¼arCell,
iCount:¼iCount
Next
Set rLast ¼ rTree.Columns(lCount - 1).EntireColumn.
SpecialCells(xlCellTypeConstants, 23)
lCellSize ¼ rLast.Cells.Count
Next ’ / outer next
rLast.Borders(xlBottom).LineStyle ¼ xlContinuous
Application.StatusBar ¼ False
1058 J.C. Lee

End Sub
’/**************************************************
’/Purpse: To calculate the price value of every state
of the binomial
’/ decision tree
’/**************************************************
Sub DecisionTree(rCell As Range, nPeriod As Integer, _
dblPrice As Double, sngU As Single, sngD As Single)
Dim lIteminColumn As Long
If Not nPeriod ¼ 1 Then
’Do Up
DecisionTree rCell:¼rCell.Offset(0, 1), nPeriod:¼
nPeriod - 1, _
dblPrice:¼dblPrice * sngU, sngU:¼sngU, _
sngD:¼sngD
’Do Down
DecisionTree rCell:¼rCell.Offset(0, 1), nPeriod:¼
nPeriod - 1, _
dblPrice:¼dblPrice * sngD, sngU:¼sngU, _
sngD:¼sngD
End If
lIteminColumn ¼ WorksheetFunction.CountA(rCell.
EntireColumn)
If lIteminColumn ¼ 0 Then
rCell ¼ dblPrice
Else
If nPeriod <> 1 Then
rCell.EntireColumn.Cells(lIteminColumn + 1) ¼ dblPrice
Else
rCell.EntireColumn.Cells(((lIteminColumn + 1) * 2) -
1) ¼ dblPrice
Application.StatusBar ¼ "The number of binomial calcs are
: " & Me.BinomialCalc _ & " at cell " & rCell.EntireColumn.
Cells(((lIteminColumn + 1) * 2) - 1).Address
End If
End If
Me.BinomialCalc ¼ Me.BinomialCalc + 1
End Sub
Function BS_D1() As Double
Dim dblNumerator As Double
Dim dblDenominator As Double
On Error Resume Next
dblNumerator ¼ VBA.Log(Me.txtBinomialS / Me.
txtBinomialX) + _
37 Binomial OPM, Black–Scholes OPM, and Their Relationship 1059

((Me.txtBinomialr + Me.txtSigma ^ 2 / 2) * Me.txtTimeT)

dblDenominator ¼ Me.txtSigma * Sqr(Me.txtTimeT)
BS_D1 ¼ dblNumerator / dblDenominator
End Function
Function BS_D2() As Double
On Error Resume Next
BS_D2 ¼ BS_D1 - (Me.txtSigma * VBA.Sqr(Me.txtTimeT))
End Function
Function BS_Call() As Double
BS_Call ¼ (Me.txtBinomialS * WorksheetFunction.
NormSDist(BS_D1)) _
- Me.txtBinomialX * Exp(-Me.txtBinomialr * Me.txtTimeT) * _
WorksheetFunction.NormSDist(BS_D2)
End Function
’Used put-call parity theorem to price put option
Function BS_PUT() As Double
BS_PUT ¼ BS_Call - Me.txtBinomialS + _
(Me.txtBinomialX * Exp(-Me.txtBinomialr * Me.txtTimeT))
End Function

References
Benninga, S. (2000). Financial modeling. Cambridge: MIT Press.
Benninga, S. (2008). Financial modeling. Cambridge: MIT Press.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 31, 637–659.
Cox, J., Ross, S. A., & Rubinstein, M. (1979). Option pricing: A simplified approach. Journal of
Financial Economics, 7, 229–263.
Daigler, R. T. (1994). Financial futures and options markets concepts and strategies. New York:
Harper Collins.
Jarrow, R., & TurnBull, S. (1996). Derivative securities. Cincinnati: South-Western College
Publishing.
Lee, J. C. (2001). Using Microsoft excel and decision trees to demonstrate the binomial option
pricing model. Advances in Investment Analysis and Portfolio Management, 8, 303–329.
Lee, C. F. (2009). Handbook of quantitative finance. New York: Springer.
Lee, C. F., & Lee, A. C. (2006). Encyclopedia of finance. New York: Springer.
Lee, C. F., Lee, J. C., & Lee, A. C. (2000). Statistics for business and financial economics. New
Jersey: World Scientific.
Lee, J. C., Lee, C. F., Wang, R. S., & Lin, T. I. (2004). On the limit properties of binomial
and multinomial option pricing models: Review and integration. In C. F. Lee (Ed.), Advances
in quantitative analysis of finance and accounting new series (Vol. 1). Singapore: World
Scientific.
Rendleman, R. J., Jr., & Barter, B. J. (1979). Two-state option pricing. Journal of Finance, 34(5),
1093–1110.
Walkenbach, J. (2010). Excel 2010 power programming with VBA. Indianapolis: Wiley.
Wells, E., & Harshbarger, S. (1997). Microsoft excel 97 developer’s handbook. Redmond:
Microsoft Press.
Dividend Payments and Share
Repurchases of US Firms: 38
An Econometric Approach

Alok Bhargava

Contents
38.1 Introduction and Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1063
38.2 Specific Hypotheses, Financial Databases, and the Formulation of
Econometric Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065
38.3 Analytical Framework for Modeling the Dividends and
Share Repurchases Interrelationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1067
38.3.1 Background Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1067
38.3.2 Some Conceptual Aspects of Dividends and Share Repurchase
Interrelationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1067
38.4 Processing the Compustat Database from the United States . . . . . . . . . . . . . . . . . . . . . . . . . 1068
38.5 Comprehensive Empirical Models for Dividends and Share Repurchase
Interrelationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1069
38.6 The Econometric Framework for Addressing Simultaneity and
Between-Firm Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1071
38.7 Descriptive Statistics from the Annual Compustat Database . . . . . . . . . . . . . . . . . . . . . . . . 1072
38.8 Results from Simple Dynamic Models for Dividends and
Share Repurchase Controlling for Between-Firm Heterogeneity . . . . . . . . . . . . . . . . . . . . 1075
38.9 Results from Comprehensive Dynamic Models for Dividends and
Share Repurchases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1078
38.9.1 Results from Dynamic and Static Models for Firms’
Dividends per Share . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1078
38.9.2 Results from Dynamic and Static Models for Firms’
Share Repurchases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1080
38.9.3 Results from Dynamic and Static Models for Dividends
per Share with Share Repurchases and Intangible Assets Included as
Explanatory Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1082
38.9.4 Results from Dynamic and Static Models for Share Repurchases with
Intangible Assets Included as Explanatory Variables . . . . . . . . . . . . . . . . . . . . . . 1085

A. Bhargava
School of Public Policy, University of Maryland, College Park, MD, USA
e-mail: bhargava@umd.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1061
DOI 10.1007/978-1-4614-7750-1_38,
# Springer Science+Business Media New York 2015
1062 A. Bhargava

38.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1086

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1087
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1090

Abstract
The analyses of dividends paid by firms and decisions to repurchase their own
shares require an econometric approach because of the complex dynamic
interrelationships. This chapter begins by, first, highlighting the importance
of developing comprehensive econometric models for these interrelation-
ships. It is common in finance research to spell out “specific hypotheses”
and conduct empirical research to investigate validity of the hypotheses.
However, such an approach can be misleading in situations where variables
are simultaneously determined as is often the case in financial applications.
Second, financial and accounting databases such as Compustat are complex
and contain useful longitudinal information on variables that display
considerable heterogeneity across the firms. Empirical analyses of financial
databases demand the use of econometric and computational methods in
order to draw robust inferences. For example, using longitudinal data on the
same US firms, it was found that dividends were neither “disappearing”
nor “reappearing” but were relatively stable in the period 1992–2007
Bhargava (Journal of the Royal Statistical Society A, 173, 631–656, 2010).
Third, the econometric methodology tackled the dynamics of relationships
and investigated endogeneity of certain explanatory variables. Identification
of the model parameters is achieved in such models by exploiting the
cross-equations restrictions on the coefficients in different time periods.
Moreover, the estimation entails using nonlinear optimization methods
to compute the maximum likelihood estimates of the dynamic random
effects models and for testing statistical hypotheses using likelihood ratio
tests. For example, share repurchases were treated as endogenous
explanatory variable in the models for dividend payments, and dividends
were treated as endogenous variables in the models for share repurchases.
The empirical results showed that dividends are decided quarterly at the
first stage, and higher dividends payments lowered share repurchases
by firms that are made at longer intervals. These findings cast
some doubt on evidence for the simple “substitution” hypothesis
between dividends and share repurchases. The appendix outlines some of
the econometric estimation techniques and tests that are useful for research
in finance.

Keywords
Compustat database • Corporate policies • Dividends • Dynamic random effects
models • Econometric methodology • Endogeneity • Maximum likelihood •
Intangible assets • Model formulation • Nonlinear optimization • Panel data •
Share repurchases
38 Dividend Payments and Share Repurchases of US Firms 1063

38.1 Introduction and Background

The proximate determinants of dividends payments by firms are of interest to

researchers in economics and finance. From a historical perspective, Tinbergen
(1939) proposed a model for predicting stock prices based on firms’ earnings and
dividends per share. Later, Lintner (1956) proposed a partial adjustment scheme for
dividends and represented it by a first-order autoregressive (“dynamic”) model where
dividends per share in the current time period were a function of the previous levels.
Such models were also employed in the economics literature to reflect “habit
persistence” in the consumption function (Duesenberry 1949), i.e., current consump-
tion depends on the past consumption. The work by Lintner (1956) was extended to
data on US firms by Fama and Babiak (1968). However, the methods used in the
early research for estimation of model parameters were rather elementary.
While data on macroeconomic variables such as consumption and income
typically span for long periods and facilitate the estimation of econometric models
(Koopmans 1950), the estimation of models using firm level data presents addi-
tional complications. For example, the large amount of heterogeneity across firms
needs to be taken into account. Such issues are important for finance research
because the neglect of unobserved between-firm differences can vitiate the consis-
tency properties of the estimated model parameters. Moreover, financial databases
such as (Compustat 2008) contain elaborate information at the firm level and
several variables affecting dividend payments by firms need to be accounted for
in the econometric models.
Further, shares repurchase, where firms purchase their own shares from stock
holders and place them in “Treasury stock,” have become popular since the 1980s
in the United States and in the European Union (von Eije and Megginson 2008).
Share repurchases reduce the numbers of shares held by investors (“common shares
outstanding”), thereby increasing share values and hence are a form of payment to
stock holders. It has also become popular for executives and employees of firms to
receive part of their remuneration as options to purchase company stock at the
current price in the future. The longitudinal information in financial data sets such
as Compustat can facilitate the investigation of substantive issues such as the
proximate determinants of dividend payments and the interrelationships between
dividends and share repurchases by firms (e.g., Jagannathan et al. 2000; Fama and
French 2001; Grullon and Michaely 2002; De Angelo et al. 2006; Skinner 2008;
Bhargava 2010; Jones and Wu 2010).
Further, while analyses of the Compustat data can provide useful insights,
methodological aspects are critical. For example, most previous analyses have
not exploited the longitudinal (“panel”) nature of the Compustat data. Instead,
cross-sectional regressions have been estimated in different time periods for
modeling firms’ chances of paying dividends, i.e., dependent variables take the
values 0 or 1 (Jagannathan et al. 2000; Fama and French 2001; De Angelo
et al. 2006). In applications where longitudinal analyses were conducted for
dividends payments (Skinner 2008), the model parameters were likely to be incon-
sistently estimated because the appropriate estimators for dynamic models were not
1064 A. Bhargava

employed. This chapter discusses several difficulties in the previous literature and
presents findings on the interrelationships between dividends and share repurchases
by over 2,000 US industrial firms (see Bhargava 2010).
First, it is essential to model the time structure of the interrelationships between
dividends and share repurchases using longitudinal data on firms. While Lintner
(1956) argued that firms may be reluctant to alter dividend rates, firms often adjust
dividends per share to the market circumstances. Thus, classification of firms
between dividend “payers” and “non-payers” is not a concrete one since a majority
of firms may pay dividends during some periods. For example, using eight 2-yearly
averages of dividends paid by a panel of 3,154 US industrial firms in the period
1992–2007 (i.e., for 1992–1993, 1994–1995, 1996–1997, 1998–1999, 2000–2001,
2002–2003, 2004–2005, 2006–2007), Bhargava (2010) found that 42 % of the firms
never paid dividends, 33 % paid dividends in all 8 periods, and 25 % paid dividends
in 1–7 periods. Moreover, dividends per share ranged from $0.00 to $67.61 so that it
is appropriate to model the payments via continuous random variables (Fama and
Babiak 1968). Similarly, firms’ share repurchase are continuous variables and can be
defined using the variables “Treasury stock,” “purchase of common stock,” and “sale
of common stock.” Using longitudinal data on 2,907 firms in 8 periods, 20 % of firms
made no repurchases, 5 % repurchased shares in all 8 periods, and 75 % made 1–7
repurchases (Bhargava 2010). Thus, models with continuous dependent variables for
dividends and share repurchases are suitable for analyses of payout decisions.
Moreover, it is essential to use longitudinal data on the same firms for analyzing
the interrelationships between dividends and share repurchases.
Second, from a methodological standpoint, the estimation of dynamic models
from longitudinal data covering large number of firms observed for a few time
periods requires the treatment of lagged dependent variables as “endogenous”
variables that are correlated with firm-specific random effects (Anderson and
Hsiao 1981; Bhargava and Sargan 1983); this issue will be discussed below and
in the Appendix. Despite the popularity of Lintner (1956) type models in finance
literature, such issues have seldom been addressed by researchers and critically
affect consistency properties of the estimated model parameters. Moreover, it is
informative to model the actual magnitudes of dividends and share repurchases,
since these are likely to be related to firm characteristics such as size, earnings, and
debt. It would also be useful to consider their joint determination; models for
continuous dependent variables using longitudinal data can handle situations
where some explanatory variables are correlated with errors affecting the models
(Bhargava and Sargan 1983; Bhargava 1991).
Third, there has been a discussion in the finance literature regarding possible
“substitution” between dividends payments and share repurchases by firms. While
some researchers have argued that these payout methods may not be substitutes
(e.g., John and Williams 1985), substitution of dividends by share repurchases
(or vice versa) are consistent with certain economic formulations (Miller and
Modigliani (1961). For example, Grullon and Michaely (2002) interpreted negative
correlations between forecasting errors in changes in dividends and share
repurchases as evidence of substitution. However, these issues demand an
38 Dividend Payments and Share Repurchases of US Firms 1065

econometric approach for addressing the possible asymmetries and endogeneity in

the relationships. For example, using the Compustat data on 2,800 industrial firms
in the eight time periods noted above, bivariate correlations between dividends per
share and repurchases were 0.11, 0.11, 0.08, 0.07, 0.11, 0.03, 0.05, and 0.06,
respectively. These positive correlations were statistically significant; similar
results were obtained for correlations between total dividend payments and share
repurchases. Thus, investigation of possible substitution between dividends and
share repurchases demands a comprehensive econometric framework (see below).
This chapter outlines a comprehensive analysis of the interrelationships between
dividends and share repurchases using longitudinal Compustat data at 2-year inter-
vals on over 2,000 US firms for the period 1992–2007 (Bhargava 2010). A brief
discussion of issues arising in econometric methodology that is important for formu-
lation of models and testing hypotheses in finance research is presented in Sect. 38.2.
The analytical framework for dividends and share repurchase interrelationships is
developed in Sect. 38.3, and econometric models embodying various hypotheses are
outlined. In Sect. 38.4, Compustat variables are described and construction of the
longitudinal data set is explained. For example, 2-yearly averages for variables are
well suited for the analyses because share repurchases occur roughly at 2-year
intervals (e.g., Stephens and Weisbach 1998); averaging also reduces the impact of
missing values. The comprehensive econometric models for dividends and share
repurchase relationships are outlined in Sect. 38.5. The econometric framework is
outlined in Sect. 38.6 and it yields consistent and efficient estimates of model
parameters; diagnostic tests for model adequacy and certain steps in the applications
of the methods are outlined in the Appendix. The descriptive results from the
Compustat data are in Sect. 38.7; results from estimating simple dynamic models
for dividends and share repurchases are in Sect. 38.8. In Sect. 38.9, the results from
estimating comprehensive dynamic econometric models for dividends and share
repurchases are discussed. It is emphasized that one can draw robust conclusions
from the estimated parameters of models that capture salient aspects of the various
relationships. The conclusions are summarized in Sect. 38.10.

38.2 Specific Hypotheses, Financial Databases, and the

Formulation of Econometric Models

It would be useful to discuss the role of conceptual aspects in the formulation of

econometric models for finance relationships especially in the context of interrela-
tionships between dividends and share repurchases. First, note that in biomedical
sciences, randomized controlled trials are designed to investigate “specific hypoth-
eses.” The design of experiments is influenced by previous findings that are treated
as assumptions in the trial (Cox 1958) and enable the investigation of specific
hypotheses under consideration (Bhargava 2008). The validity of specific hypoth-
eses is investigated using the emerging data from control and treatment groups of
the trial. In contrast, finance researchers often spell out hypotheses that are
influenced by economic postulates and investigate the support for the hypotheses
1066 A. Bhargava

using existing databases such as Compustat for the United States. However, the
hypotheses driving the economic and finance investigation are often based on
untested assumptions and/or “stylized facts” (Granger 1992). By contrast, assump-
tions invoked in biomedical sciences reflect the knowledge accumulated from
previous experimental studies, and hypotheses are tested using the new data.
Further, in social sciences research, most phenomena being explained result
from the interactions between economic and social factors. Because different
dimensions of the relationships merit different emphasis depending on the context,
social science researchers are forced to address several hypotheses simultaneously.
Typically, this is done via formulation of comprehensive models where, for exam-
ple, the systematic part of the relationships incorporates the relevant variables. The
stochastic properties of the dependent variable and error terms are also tackled in
the formulation of comprehensive econometric models.
The analyses of databases such as Compustat entail the modeling of accounting,
finance, and economic variables so that financial research often demands econo-
metric modeling of variables that are jointly determined. Moreover, economic
theory is helpful in identifying variables and often implies restrictions on certain
model parameters. Thus, investigation of hypotheses in finance applications entails
the development of comprehensive econometric models within which several
hypotheses can be embedded and tested using the estimated model parameters.
This is in contrast with the approach in biomedical sciences where trials are
designed to investigate specific hypotheses, and researchers are not concerned
with interdependence between the variables. Thus, the currently popular approach
in finance of spelling out a few specific hypotheses needs to be augmented by
development of comprehensive econometric models reflecting the hypotheses. It is
only after the estimation of comprehensive econometric models from databases
such as Compustat that investigators can satisfactorily test the validity of specific
hypotheses using statistical procedures.
As an illustration of the importance of methodological aspects for drawing infer-
ences, it would be helpful to reappraise the work by Grullon and Michaely (2002)
claiming empirical support for the “substitution” hypothesis between
dividend payments and share repurchases. First, as noted in the Introduction, the
correlations between dividend payments and share repurchases estimated from
Compustat data were positive so the interrelationships are likely to be complex.
Second, if dividends and share repurchases are substitutes, then they are likely to be
simultaneously determined; modeling the interrelationships would require the use of
econometric methods for handling endogenous explanatory variables in longitudinal
analyses. Also, longitudinal data need to cover the same firms over time in order to
shed light on the interrelationships. While such issues were not considered by Grullon
and Michaely (2002), the authors explained forecasting errors in changes in dividends
based on share repurchases and interpreted negative correlations as evidence for
substitution hypothesis. However, dividends are announced quarterly, whereas share
repurchases are made at 2–3-yearly intervals. Thus, the stochastic properties of vari-
ables cast doubt on adequacy of the models that explain a higher frequency variable
such as dividends by the slowly changing share repurchases. In fact, it is likely that
38 Dividend Payments and Share Repurchases of US Firms 1067

dividend payments that are made on a quarterly basis affect share repurchases but not
vice versa. As shown below, firms’ share repurchases are constrained by dividend
payments and repurchases are relatively flexible (Bhargava 2010). Overall, from
a methodological standpoint, it is important in financial analyses to develop compre-
hensive models prior to the testing of specific hypotheses.

38.3 Analytical Framework for Modeling the Dividends and

Share Repurchases Interrelationships

38.3.1 Background Issues

The earliest dynamic model for dividends is given by Lintner (1956):

Di t ¼ ai t þ b Pi t þ d Di t1 þ ui t ði ¼ 1, 2, . . . , H; t ¼ 2, 3, . . . , TÞ (38.1)

For firm i in time period t, Di t is the dividend payment, Pi t is profit or earnings,

and ui t is a random error term; it is assumed that there are H firms in the sample that
are observed in T time periods. The subsequent work by Fama and Babiak (1968)
expressed dividends and earnings in per share terms. An important feature of these
models is that the short-run effect of a unit increase in profit is given by b, while the
long-run equilibrium impact is [b/(1  d)] with |d| < 1.
Explanatory variables such as firms’ debt, assets, and investments can be
included in Eq. 38.1. However, the estimation methods employed in previous
research (Lintner 1956; Fama and Babiak 1968; Lee et al. 1987) were appropriate
for time series data on a single firm rather than for a panel of firms. Because firms in
Compustat and other databases are heterogeneous, it is important to include firm-
specific “random” or “fixed” effects in the model. For example, treating the
unobserved between-firm differences as randomly distributed variables, error
terms ui t can be decomposed as:

ui t ¼ di þ vi t (38.2)

where, di are firm-specific random effects that follow some distribution (e.g., normal)
with zero mean and constant variance, and vi t are distributed with zero mean and
constant variance (Anderson and Hsiao 1981; Bhargava and Sargan 1983). More-
over, one can invoke the general assumption that ui t’s are drawings from
a multivariate distribution with a symmetric and positive definite dispersion matrix.

38.3.2 Some Conceptual Aspects of Dividends and Share

Repurchase Interrelationships

From a conceptual standpoint, one would expect comprehensive models for

dividends or dividends per share to resemble the dynamic model in Eq. 38.1,
with additional explanatory variables such as firms’ assets, debt, and
1068 A. Bhargava

investments (Bhargava 2010). Similar models can be developed for share

repurchases that have become popular following the 1986 Tax Reform Act in
the United States. However, there are likely to be asymmetries in the interrela-
tionships between dividend payments and share repurchases. For example, firms
regularly paying dividends may be reluctant to lower dividends in order to
increase share repurchases since that might send ambiguous signals to investors.
Thus, higher dividend payments are likely to lower firms’ ability to make
repurchases, i.e., in a model for repurchases, dividend payments are likely to
be estimated with a negative coefficient. By contrast, decisions to make
repurchases may be influenced by firms’ unexpectedly large cash holdings
(Guay and Harford 2000; Brav et al. 2005; Chen and Wang 2012) so that in
a model for dividends, coefficient of share repurchases need not be statistically
significant. These conceptual aspects can be embedded in a simultaneous equa-
tion model taking into account endogeneity of dividend payments and share
repurchases (Sect. 38.5, below).

38.4 Processing the Compustat Database from the

United States

Databases such as Compustat compile detailed information on accounting and

finance variables for large numbers of firms listed on stock exchanges. However,
many firms merge or exit after some years and are removed from the database. Cross-
sectional studies analyzing data on firms for different years are therefore likely to
include many firms that will be dropped in subsequent periods. From the standpoint
of modeling the proximate determinants of dividends and share repurchases, this
raises several issues. First, firms staying on the stock market for only a few time
periods may typically pay small or no dividends. Also, the interrelationships between
dividend payments and share repurchases require a longer time frame since dividends
per share are announced on a quarterly basis, while share repurchases are made
roughly at 2-year intervals. Thus, while cross-sectional analyses can provide insights,
the interrelationships between dividend payments and share repurchases require that
the same firms are observed for a certain number of years.
Second, there are trade-offs between analyzing longitudinal data that span long
versus short time periods. If, for example, the data cover long periods of (say) over
30 years, then the number of firms in the sample is likely to be small and mainly the
well-established firms will be included. By contrast, if one analyzes longitudinal
data for only 5 years, then many firms included in the sample will be dropped from
the stock exchange in later periods. In view of the fact that share repurchase
decisions are made roughly at 2-year intervals, a reasonable approach would be
to analyze data covering around 15 years. Moreover, one can create 2-yearly
averages so that if observations were missing for a firm in one of the years, the
firm could still be included in the sample. Since the data on executive compensation
are available from 1992 (ExecuComp 2008), the sample period 1992–2007 is
appealing (Bhargava 2011).
38 Dividend Payments and Share Repurchases of US Firms 1069

Third, even when the same firms are retained in the sample, there is large intra-
and interfirm variation in financial variables. In the context of dividends, some
problems can be tackled by expressing variables in per share terms (Fama and
Babiak 1968). However, as recognized in the early statistical literature by Pearson
(1897) and Neyman (1952), transforming variables into ratios can induce spurious
correlations. Such practices were criticized by Kronmal (1993), Bhargava (1994)
developed a likelihood ratio statistic for testing the restrictions on coefficients that
enable expressing variables as ratios in empirical models. While research in finance
often expresses variables in ratio forms, such transformation can affect the magni-
tudes and signs of estimated regression coefficients. It is important to conduct
robustness check to ensure that transformation of variables do alter the conclusions.
Finally, there is often high inter- and intra-firm variation in variables such as
share repurchases, assets, and debt that are measured in millions of dollars in the
Compustat database. Transformations to natural logarithms can reduce internal varia-
tion in the data. For example, share repurchase can be expressed in dollars and then
transformed into natural logarithms with zero dollar values set equal to one. This
procedure facilitates the estimation of model parameters using numerical optimization
techniques and obviates the need for arbitrarily truncating large values of variables
(Bhargava 2010).

38.5 Comprehensive Empirical Models for Dividends and Share

Repurchase Interrelationships

A dynamic random effects model for dividends per share adjusted for “stock splits”
using cumulative adjustment factor (#27), with Compustat item numbers next to the
variables, can be written as (Bhargava 2010):

ðDividends per share; #26, #27Þi t ¼ a0 þ a1 ðEarnings; #18, #17Þi t þ a2 ln ðTotal assets; #6Þi t
þ a3 ðMarket-to-book value; #199, #25, #60Þi t þ a4 ln ðLong  term debt; #9Þi t
þ a5 ln ðShort-term investments; #193Þi t þ a6 ðTime dummy period 3Þi t
þ a7 ðTime dummy period 4Þi t þ a8 ðTime dummy period 5Þi t
þ a9 ðTime dummy period 6Þi t þ a10 ðTime dummy period 7Þi t
þ a11 ðTime dummy period 8Þi t þ a12 ðDividends per share; #26, #27Þi t1
þ ui t ði ¼ 1, 2, . . . , H; t ¼ 2, 3, . . . , 8Þ
(38.3)

In the empirical model in Eq. 38.3, it is recognized that firms sometimes split
their shares especially if share prices are high. Earnings are defined as “income
before extraordinary items” (#18) minus 0.6 times “special items” (#17), measured
in million of dollars (Skinner 2008). Income before extraordinary items is the firms’
incomes taking into account all expenses, while special items reflect adjustments
1070 A. Bhargava

for debts and losses. “Total assets” (#6), “long-term debt” (#9), and “short-term
investments” (#193) were expressed in dollars and then converted into natural
logarithms. The “market-to-book value” variable was constructed by multiplying
“share price” (#199) by number of “common shares outstanding” (#25) and
dividing this product by “common equity total” (#60). “Dividends per share” in
period (t1) will be referred to as “lagged dependent variable.” Models of the type in
Eq. 38.3 were also estimated for dividends expressed as ratios to firms’ book values,
while dropping the explanatory variable market-to-book value from Eq. 38.3.
It is assumed in Eq. 38.3 that 8-time observations at 2-yearly intervals
{1992–1993, 1994–1995, 1996–1997, 1998–1999, 2000–2001, 2002–2003,
2004–2005, 2006–2007} were available on H firms. Moreover, as explained in
Sect. 38.6 and the Appendix, initial values of the dependent variable in time period 1
were modeled using a “reduced form” equation that includes a separate coefficient,
c0, for the constant term. Thus, the model in Eq. 38.3 includes separate coefficients
of the constant term for each of the 8-time periods. This formulation allows the
variables in the model to have different means in the 8-time periods and is useful if
there are trends in the dependent variables in the observation period.
Further, ui t’s are error terms that can be decomposed in the simple random effect
fashion as in Eq. 38.2. The variance of vi t is the “within” (or intra-) firm variance, and
[var (di)/var (vit)] is the “between-to-within” variance ratio that can be estimated when
errors are decomposed as in Eq. 38.2. In another version of the model in Eq. 38.3,
firms’ intangible assets (Compustat item #33) were included in the model, though
greater numbers of observations were missing for this variable. Also, firms’ share
repurchases were introduced as potentially endogenous explanatory variables to test if
they influenced dividend payments. Last, static versions of the model in Eq. 38.3 that
excluded the lagged dividends were estimated to assess robustness of the results.
The model for share repurchases was similar to that in Eq. 38.3 except that
dividends per share were included as a potentially endogenous explanatory variable:

ln ðShare repurchases; #226, #115, #108Þi t ¼ b0 þ b1 ðEarnings; #18, #17Þi t

þ b2 ln ðTotal assets; #6Þi t þ b3 ðMarket-to-book value; #199, #25, #60Þi t
þ b4 ln ðLong-term debt; #9Þi t þ b5 ln ðShort-term investments; #193Þi t
þ b6 ðTime dummy period 3Þi t þ b7 ðTime dummy period 4Þi t þ b8 ðTime dummy period 5Þi t
þ b9 ðTime dummy period 6Þi t þ b10 ðTime dummy period 7Þi t þ b11 ðTime dummy period 8Þi t
þ b12 ðDividends per share; #26, #27Þi t þ b13 ln ðShare repurchases; #226,#115, #108Þi t1
þ u2 i t ði ¼ 1, 2, .. . , H;t ¼ 2, 3, . . ., 8Þ
(38.4)

Note that share repurchases were defined as the change in “Treasury stock”
(#226), and if this was less than or equal to zero for the firm, the nonnegative
difference between “purchase of common stock” (#115) and “sale of common
stock” (#108) was used (Fama and French 2001). This definition covered the
38 Dividend Payments and Share Repurchases of US Firms 1071

situation where firms “retire” shares from Treasury stock so that when the change in
Treasury stock was less than or equal to zero, nonnegative difference between
purchase of common stock and sale of common stock was used as the measure of
repurchases. Banyi et al. (2008) have advocated the use of purchase of common stock
as a measure of repurchases, and this variable was also modeled to assess robustness
of the results. Finally, alternative models for share repurchases were estimated with
intangible assets included as an explanatory variable, and also where dividends per
share was replaced by dividends expressed as ratios to firms’ book values.

38.6 The Econometric Framework for Addressing Simultaneity

and Between-Firm Heterogeneity

The methodology used for estimation of dynamic and static random effect models,
where some explanatory variables are endogenous, was developed in Bhargava and
Sargan (1983) and Bhargava (1991). Let the dynamic model be given by:

X
m X
n1
yi t ¼ z i j gj þ x1 i j t bj
j¼1 j¼1

X
n
þ x2 i j t bj þ a yi t1 þ ui t ði ¼ 1, 2, . . . , H; t ¼ 2, 3, . . . , TÞ
j¼n1 þ1

(38.5)

where z’s are time-invariant variables, x1 and x2 are, respectively, n1 exogenous and
n2 endogenous time-varying variables (n1 + n2 ¼ n). In the model for share
repurchases, for example, dividends per share is potentially an endogenous time-
varying variable; unobserved factors affecting share repurchases, reflected in firm-
specific random effects (di) in Eq. 38.2, can influence dividend payments. For
exposition purposes, first assuming that all time-varying variables are exogenous
so that n2 ¼ 0 and subscripts on the x variables can be dropped, the dynamic model
can be written in a simultaneous equations framework by defining a “reduced form”
equation for initial observations that do not include any endogenous explanatory
variables, and a “triangular” system of (T-1) “structural” equations for the
remaining periods (Bhargava and Sargan 1983):

X
m n X
X T
yi 1 ¼ z i j zj þ uj k xi j k þ ui 1 ði ¼ 1, . . . , HÞ (38.6)
j¼1 j¼1 k¼1

And
0 0 0
B Y þ Cz 0 þ Cx X ¼ U
Z
ðT  1ÞxT TxH ðT  1Þx m m x H ðT 1Þx nT nTxH ðT  1ÞxH (38.7)
1072 A. Bhargava

Here, Y, Z, and X are, respectively, matrices containing observations on the

dependent, time-invariant, and time-varying explanatory variables; dimensions of
the matrices are written below the respective symbols. B is a (T1) x T lower
triangular matrix of coefficients:

Bi i ¼ a, Bi, iþ1 ¼ 1, Bi j ¼ 0 otherwise ði ¼ 1, . . . , T  1; j ¼ 1, . . . , TÞ

(38.8)

where a is the coefficient of lagged dependent variable. Matrices Cz and Cx contain

coefficients of time-invariant and time-varying regressors, respectively; U contains
the error terms.
Note that in the reduced form Eq. 38.6, all T realizations of the n exogenous
time-varying variables appear as explanatory variables for modeling the systematic
or predicted part of yi 1. Because maximum likelihood or instrumental variable
methods treat {yi 1, yi 2, . . ., yi T} as jointly determined variables, it is essential to
explain the reduced form Eq. 38.6 using all exogenous variables in the model. The
estimation of model parameters and certain econometric tests are described in the
Appendix.

38.7 Descriptive Statistics from the Annual Compustat

Database

The Compustat annual data on firms for the period 1992–2007 were processed for
the analyses (Bhargava 2010). As is customary in the dividends literature, financial
firms with Standard Industry Classification (SIC) codes between 6,000 and 6,999
and utility firms with SIC codes between 4,900 and 4,949 were dropped since there
are regulations on dividend payments by such firms (Fama and French 2001).
Moreover, observations were retained on firms that were in the Compustat database
for at least 14 years, and 2-yearly averages were created at 8-time points, i.e., for
{1992–1993, 1994–1995, 1996–1997, 1998–1999, 2000–2001, 2002–2003,
2004–2005, 2006–2007}. An alternative data set created 3-yearly averages at
5-time points, i.e.,{1992–1994, 1995–1997, 1998–2000, 2001–2003, 2004–2006},
though the analyses of the 2-yearly averages led to more robust parameter estima-
tion. Averaging over 2 years led to a sample of 3,290 industrial firms observed in
8-time periods. Also, estimation of dynamic models requires “balanced” panels,
i.e., where firms are included at all eight time points. While it is possible to analyze
“unbalanced” panels using static models in software packages such as Stata (2008),
it is difficult to address endogeneity issues in the analyses.
The sample means and standard deviations of 2-year averages of variables are
reported in Table 38.1 for some of the years with Compustat annual item numbers
noted next to the variables. The mean dividends per share were approximately
$0.32 for the period 1992–1999, and declined slightly to $0.27 for 2000–2003, and
increased to $0.39 in 2006–2007. While a simple paired t-test on the set of firms’
observations showed significant (P < 0.05) increase in dividends per share between
 38

Table 38.1 Sample means and standard deviations of 2-yearly averages of selected variables from the Compustat database for up to 3,290 US industrial firms
in 1992–2007
1992–1993 1998–1999 2002–2003 2006–2007
Variable Mean SD Mean SD Mean SD Mean SD
Income before extraordinary items (#18), $ million 53.99 317.55 138.05 702.98 124.09 1,015.57 309.43 1,818.94
Special items (#17), $ million 15.99 215.30 1.56 372.44 55.10 552.09 28.20 731.89
Treasury common stock (#226), $ million 37.08 398.01 108.79 852.50 194.16 1,461.74 406.76 2,994.53
Purchase of common stock (#115), $ million 10.39 66.72 54.54 282.34 51.85 349.32 221.96 1,538.69
Sale of common stock (#108), $ million 21.66 112.12 31.07 160.90 25.69 104.93 72.96 1,022.77
Share repurchases (#226, #115, #108), $ million 9.95 111.13 44.44 316.76 49.49 467.39 174.77 1,564.85
Dividends (#21), $ million 32.40 166.22 63.22 291.33 63.09 376.21 113.18 629.87
Dividends per share (#26), $ 0.304 0.902 0.298 0.882 0.281 1.272 0.390 1.095
Total assets (#6), $ million 2,292.84 13,195.74 3,845.47 23,638.40 5,491.57 33,186.31 8,083.80 61,202.42
Cumulative adjustment factor (#27) 2.48 4.48 1.41 1.24 1.18 0.72 1.03 0.89
Share price (#199), $ 17.55 19.61 19.40 27.03 17.21 25.29 23.55 42.62
Long-term debt (#9), $ million 429.26 2,417.02 695.61 3,664.86 1,070.41 6,085.70 1,519.05 11,825.75
Dividend Payments and Share Repurchases of US Firms

Short-term investments (#193), $ million 88.31 1,181.56 173.91 2,767.14 238.97 3,735.60 368.76 8,385.86
Intangible assets (#33), $ million 99.91 746.35 279.68 1,279.77 651.12 4,629.00 980.21 5,507.11
Common shares outstanding (#25), million 39.53 133.66 84.22 279.68 159.55 1,682.32 177.41 1,684.17
Common equity total (#60), $ million 541.70 2,214.41 958.11 3,917.14 1,351.86 6,265.71 1,985.57 8,140.76
Percentages of firms paying dividendsa 36.50 – 40.79 – 39.89 – 44.00 –
Percentages of firms making repurchasesa 26.29 – 49.82 – 37.93 – 43.46 –
Means and standard deviations of variables using 2-yearly averages at four time points for up to 3,290 US industrial firms observed for at least 14 years in
1992–2007 (Source: Bhargava 2010); specific number of firms included depends on the extent of unavailable information for each variable in each period;
Compustat items and units of measurements are listed next to variables
a
Percentages are based on firms’ with nonzero payments
1073
1074 A. Bhargava

450

400

350

300

250
$million

200

150

100

50

0
1992–93 1994–95 1996–97 1998–99 2000–01 2002–03 2004–05 2006–07
Year
Treasury common stock, (#226) Purchase of common stock, (#115)
Sale of common stock, (#108) Share repurchases, (#226, #115, #108)
Dividends, (#21)

Fig. 38.1 Sample means of dividends and components of share repurchases by US industrial
firms for 1992–2007. Share repurchases were calculated as the change in Treasury stock (#226),
and if this was less than or equal to zero, the nonzero difference between purchase of common
stock (#115) and sale of common stock (#108) was used (Source: Bhargava 2010)

1992–1993 and 2006–2007, the pattern over time was more complex partly because
of tax changes in 2003. Mean share repurchases increased from $10 million in
1992–1993 to $175 million in 2006–2007.
Figure 38.1 plots the sample means (in $ millions) of Treasury stock, purchase
of common stock, sale of common stock, dividends, and share repurchases.
There were upward trends in Treasury stock, share repurchases, and purchases
of common stock. There was an increase in dividends in the period from 2002–2003
to 2006–2007, where dividends increased from $63 million to $113 million.
This could be due to reductions in 2003 in tax rates on dividends (Julio and
Ikenberry 2004).
Further, focusing on 2,880 firms with non-missing observations on dividends
and share repurchases, percentages of firms paying dividends and making share
repurchases in the 8-time periods, were {36.5, 39.9, 40.6, 40.8, 39.6, 39.9, 43.7,
44.0} and {26.3, 30.7, 38.2, 49.8, 48.1, 37.9, 38.1, 43.5}, respectively. While
percentages of firms making share repurchases increased from 26.3 in 1992–1993
to 43.5 in 2006–2007, the peak of 49.8 % was reached in the 1998–1999 period.
The mean share repurchases in 1998–1999 were 44.44 million, while they
38 Dividend Payments and Share Repurchases of US Firms 1075

amounted to $174.77 million in 2006–2007 so that share repurchases were

significantly higher in 2006–2007 after adjusting for the price level. The other
salient feature of the Compustat data was that the mean intangible assets increased
from $99.9 million in 1992–1993 to $980.2 million in 2006–2007 constituting
a tenfold increase.

38.8 Results from Simple Dynamic Models for Dividends and

Share Repurchase Controlling for Between-Firm
Heterogeneity

The results from estimating simple dynamic random effects models for dividends
per share are in Table 38.2 for the sample of 3,113 industrial firms (Bhargava 2010).
The results are presented for the cases where firms paid nonzero dividends at least
in one of the eight time periods and where firms paid dividends in all 8-time periods.
The results for ratios of dividends to book value of firms are presented in the last
column of Table 38.2. A set of six dummy variables for time periods 3–8 was
included to account for differences in dividend payments in the 8-time periods. The
models were estimated by maximum likelihood and provide consistent and efficient
estimates of the parameters.
For the pooled sample, coefficients of the dummy variables in the model for
dividends per share were positive and significant (P < 0.05) in time periods 7 and
8 corresponding to 2004–2005 and 2006–2007, respectively. The positive coeffi-
cients were consistent with reported increases in dividends following reduction in
tax rates in 2003. Coefficients of the dummy variables were qualitatively similar for
the reduced sample of 1,827 firms that paid dividends at least in one of the eight
periods and in the case where 1,035 firms paid dividends in all eight periods.
Moreover, coefficients of the dummy variables for time periods 7 and
8 corresponding to 2004–2005 and 2006–2007, respectively, were approximately
twice as large for firms paying dividends than for firms that never paid dividends.
Thus, reduction in tax rates on dividends in 2003 appeared to have increased
dividends per share. These results also show the robustness of dynamic models in
situations where firms decide against paying dividends in some time periods, i.e., to
zero values of the dependent variables. This is not surprising since one is modeling
the deviations of dividends per share from an overall mean.
The coefficients of the lagged dependent variables for the first three cases
presented in Table 38.2 were close to 0.28. Moreover, the ratios of between-to-
within variances for the three cases, i.e., pooled sample, firms paying nonzero
dividends at least once, and firms paying dividends in all 8 periods were, 0.38,
0.33, and 0.45, respectively. These estimates were close despite reduction in sample
size from 3,113 firms in the pooled sample to 1,035 in the case of firms paying
nonzero dividends in all 8 periods. The within firm variances for the three cases
were 0.52, 0.88, and 1.07, respectively. Finally, in the last column, the results for
the ratio of dividends to firms’ book values were similar to those for the pooled
sample for dividends per share. While the estimated coefficient of lagged dependent
 1076

Table 38.2 Maximum likelihood estimates of simple dynamic models for 2-yearly averages of dividend paid by US industrial firms in 1992–2007
Dependent variable:dividends/
Dependent variable: dividends per share (Compustat items #26/#27) book value (#21,#60)
Firms paying nonzero Firms paying nonzero
All firms dividends at least once dividends in all 8 periods All firms
Explanatory variables Coefficient SE Coefficient SE Coefficient SE Coefficient SE
Constant 0.172 0.015 0.294 0.018 0.421 0.031 0.018 0.004
Time period 3 dummy variable 0.017 0.020 0.029 0.026 0.029 0.046 0.001 0.006
Time period 4 dummy variable 0.033 0.020 0.058* 0.026 0.081 0.046 0.003 0.006
Time period 5 dummy variable 0.002 0.019 0.003 0.026 0.010 0.046 0.0002 0.006
Time period 6 dummy variable 0.017 0.019 0.030 0.026 0.082 0.046 0.00002 0.006
Time period 7 dummy variable 0.087* 0.020 0.148* 0.027 0.130* 0.046 0.005 0.006
* *
Time period 8 dummy variable 0.133 0.019 0.228 0.027 0.235* 0.046 0.025* 0.006
Lagged dependent variable 0.285* 0.009 0.283* 0.010 0.258* 0.013 0.290* 0.018
(between-to-within) variance ratio 0.375* 0.017 0.328* 0.019 0.451* 0.034 0.016* 0.007
Within variance 0.518 – 0.879 – 1.073 – 0.040 –
2 x (maximized log-likelihood function) 11,552.55 – 741.58 – 2,301.34 58,929.64
Number of firms 3,113 1,827 1,035 2,315
Dependent variable is dividends per share (#26) divided by cumulative adjustment factor (#27) (Source: Bhargava 2010); 8-time observations at 2-yearly
intervals on firms were used; slope coefficients and standard errors are reported
*
P < 0.05
A. Bhargava
38 Dividend Payments and Share Repurchases of US Firms 1077

Table 38.3 Maximum likelihood estimates of simple dynamic models for 2-yearly averages of
share repurchases of US industrial firms in 1992–2007
Dependent variable: ln (share repurchases $) (using Compustat
items #226, #115, #108)
Firms making
Firms making nonzero
nonzero repurchases repurchases in all
All firms at least once 8 periods
Explanatory variables Coefficient SE Coefficient SE Coefficient SE
Constant 3.216 0.061 4.057 0.153 4.376 0.587
Time period 3 dummy variable 1.008* 0.122 1.262* 0.205 0.054 0.192
Time period 4 dummy variable 2.428* 0.117 3.048* 0.200 0.067 0.193
Time period 5 dummy variable *
1.426 0.126 1.818* 0.208 0.128 0.197
Time period 6 dummy variable 0.095 0.117 0.083 0.203 0.540* 0.183
Time period 7 dummy variable *
0.738 0.122 *
0.938 0.208 0.017 0.192
Time period 8 dummy variable 1.730* 0.123 2.180* 0.205 0.171 0.194
Lagged dependent variable 0.362* 0.008 0.354* 0.009 0.754* 0.038
(between-to-within) variance 0.153* 0.010 0.066* 0.008 0.184* 0.092
ratio
Within variance 39.176 – 48.341 – 1.841 –
2 x (maximized log-likelihood 88,260.15 – 73,671.73 – 886.97
function)
Number of firms 2,907 2,329 127
Dependent variable share repurchases were calculated from Compustat items #226, #115, and
#108 and expressed in dollars and transformed into natural logarithms with zero value assigned to
zero purchases (source: Bhargava 2010); 8-time observations at 2-yearly intervals on the firms in
1992–2007 were used; slope coefficients and standard errors are reported
*
P < 0.05

variable was very close (0.29), the estimated between-to-within variance ratio was
0.016 that was lower than the corresponding estimate (0.38) in the model for
dividends per share.
The results from simple dynamic models for logarithms of US firms’ share
repurchases are in Table 38.3 for the three cases, i.e., pooled sample, firms making
at least one repurchase, and firms making repurchase in all 8-time periods. For the
pooled sample of 2,907 firms, coefficients of dummy variables for five of the time
periods were positive and statistically significant; coefficient of the dummy variable
for time period six corresponding to 2002–2003 was not significant. The results
were very similar for the case where 2,329 firms made repurchases at least once.
The dummy variables were generally insignificant in the third case for firms making
repurchases in all 8 periods; there were 127 firms in this group which was a small
sample size for estimating dynamic models.
The coefficients of lagged dependent variables in the three cases were 0.36, 0.35,
and 0.75, respectively; these estimates were significantly less than unity thereby
indicating that share repurchase series were stationary. The between-to-within
variance ratios for the three cases were 0.15, 0.07, and 0.18, respectively; within
1078 A. Bhargava

firm variances were quite large in these models. The higher coefficient of lagged
dependent variable in the model where 127 firms made repurchases in all 8 periods
may be due to the small sample size. Moreover, such firms may have been less
heterogeneous in some respects and had a smooth pattern of share repurchases.
Such issues can be systematically investigated by controlling for firm characteris-
tics and are addressed below.

38.9 Results from Comprehensive Dynamic Models for

Dividends and Share Repurchases

The results from empirical models for dividends per share and share repurchases,
outlined in Eqs. 38.3 and 38.4, are presented in this section. Dynamic and static
random effects models were estimated and the results are initially presented for the
pooled samples and for the case where firms paid dividends or made repurchases at
least once during 8-time periods.

38.9.1 Results from Dynamic and Static Models for Firms’ Dividends
per Share

Table 38.4 presents the results from dynamic and static models for dividends per
share paid by US industrial firms; a set of six dummy variables for time periods was
included in the models though the coefficients are not reported (Bhargava 2010).
The firms’ total assets, long-term debt, and short-term investments were expressed
in dollars and converted to natural logarithms. Because firms’ earnings were
defined as the difference between income before extraordinary items and special
items, this variable sometimes assumed negative values and was not transformed
into logarithms.
For the dynamic model estimated using the pooled sample, coefficient of
earnings was estimated with a positive coefficient that was statistically significant.
The coefficients of earnings and total assets were also significant in the static model
for the pooled sample and in dynamic and static models for the subsample where
firms paid dividends at least once during the 8-time periods. These results provide
evidence that earnings and total assets of US firms were positively and significantly
associated with dividends per share. By contrast, Skinner (2008, Table 8) found
these variables to be generally insignificant predictors of the ratio of dividends to
total payout (dividends plus repurchases) using cross-sectional regressions and data
on 345 US firms. When some of the coefficients were statistically significant, they
were positive for some years and negative for others. The contradictory findings
were likely to be due to combining dividends and share repurchases into a single
variable and using estimation methods that are appropriate for cross-sectional
analyses. The results in Table 38.4, however, were consistent with evidence from
the European Union (von Eije and Megginson 2008), where static random effects
models were estimated using unbalanced panel data on approximately 3,000 firms.
 38

Table 38.4 Maximum likelihood estimates from comprehensive dynamic and static random effects models for 2-yearly averages of dividends per share paid
by US industrial firms in 1992–2007
Dependent variable: dividends per share (adjusted for stock splits; Compustat items # 26/#27)
Dynamic model: all Dynamic model: firms paying Static model: firms paying at
firms Static model: all firms at least once least once
Explanatory variables: Coefficient SE Coefficient SE Coefficient SE Coefficient SE
Constant 9.383 2.664 9.169 0.464 9.357 1.306 11.091 0.817
Earnings (#18, #17), $million 0.0001* 0.00001 0.0001* 0.00001 0.0002* 0.00001 0.0001* 0.00001
* * *
ln (total assets) (#6), $ 0.058 0.019 0.060 0.003 0.054 0.010 0.072* 0.006
Market-to-book value (#199, #25, #60) 0.0004 0.004 0.0004 0.0004 0.0005 0.0008 0.001 0.001
ln (long-term debt) (#9), $ 0.0007 0.009 0.0008 0.0006 0.001 0.003 0.002 0.001
ln (short-term investments) (#193), $ 0.002 0.007 0.0006 0.0005 0.005* 0.002 0.002* 0.001
* *
Lagged dependent variable 0.201 0.021 – 0.193 0.014 –
(between-to-within) variance ratio 0.126* 0.014 – 0.114* 0.014 –
Dividend Payments and Share Repurchases of US Firms

Within variance 0.611 – – 0.980 – –

2x (maximized log-likelihood function) 5,878.43 – 630.16 – –
Number of firms 1,854 1,854 1,144 1,144
Dependent variable is dividends per share (#26) divided by cumulative adjustment factor (#27) (Source: Bhargava 2010); 8-time observations at 2-yearly
intervals on firms in 1992–2007 were used; slope coefficients and standard errors are reported; time dummies for 6 periods were included though their
coefficients are not reported in the tables
*
P < 0.05
1079
1080 A. Bhargava

The variable market-to-book value was not statistically significant in the

dynamic and static models estimated from the pooled or disaggregated samples.
Because market-to-book value is a composite variable based on Compustat items
price, common shares outstanding, and common equity total, it may be difficult to
unscramble the effects on dividend payments. Coefficients of long-term debt were
not statistically significant in the four cases reported in Table 38.4. Because firms
may be reluctant to frequently alter dividend policy, it is likely that long-term debt
might influence share repurchases rather than dividends. Coefficients of short-term
investments were positively and significantly associated with dividends per share in
the dynamic and static models estimated using the subsample of firms that paid
dividends at least once.
The coefficients of lagged dependent variables from the pooled sample and the
subsample of firms paying dividends at least once were 0.20 and 0.19, respectively.
These coefficients implied that the long-run effects of explanatory variables were
1.25 times the respective short-run effects reported in Table 38.4. Moreover, the
ratios of between-to-within variance in the two dynamic models were 0.13 and
0.11, respectively. While these ratios were smaller than those reported in Table 38.2
for simple dynamic models for dividends per share (0.38 and 0.33, respectively),
the estimates suggest that some heterogeneity across firms could not be accounted
for by explanatory variables in the models. The within firm variances in the two
dynamic models were 0.61 and 0.98. Overall, the results in Table 38.4 supported the
view that larger firms paid higher dividends per share.

38.9.2 Results from Dynamic and Static Models for Firms’ Share
Repurchases

The results for logarithms of share repurchases by US industrial firms are in

Table 38.5 for the pooled sample and for firms making repurchases at least once in
the sample period. As in the results from the model for dividends per share, firms’
earnings and total assets were estimated with significant positive coefficients in
dynamic and static models for the pooled sample and for the subsample of firms
making repurchases at least once. Coefficients of the logarithm of total assets in
dynamic models were the short-run “elasticities” of share repurchases with respect to
total assets, i.e., 0.61 and 0.78, respectively, for the pooled sample and subsample of
firms making repurchases. Because coefficients of lagged dependent variables were
approximately 0.40 in this model, the long-run effects were about 1.67 times the
respective short-run effects. For example, the long-run elasticity of share repurchases
with respect to total assets was 1.02 and this was larger than the point estimate (0.70)
from the static model for the pooled sample. Similarly, short- and long-run elasticities
of share repurchases with respect to total assets for firms making repurchases at least
once were 0.78 and 1.33, respectively. Thus, the size of the firm was an important
determinant of magnitudes of share repurchases. However, there were no nonlinear-
ities apparent with respect to firms’ earnings and total assets, i.e., squared terms of
these variables were not significant predictors of share repurchases.
 38

Table 38.5 Maximum likelihood estimates from comprehensive dynamic and static random effects models for 2-yearly averages of share repurchases by US
industrial firms in 1992–2007
Dependent variable: ln (share repurchases $) (using Compustat items #226, #115, #108)
Dynamic model: all Static model: all Dynamic model: firms Static model: firms
firms firms repurchasing least once repurchasing at least once
Explanatory variables Coefficient SE Coefficient SE Coefficient SE Coefficient SE
Constant 99.008 6.428 114.142 7.960 116.688 7.620 147.576 9.851
Earnings (#18, #17), $million 0.0004* 0.0001 0.0008* 0.0001 0.0004* 0.0001 0.0007* 0.0001
* * *
ln (total assets) (#6), $ 0.613 0.041 0.703 0.051 0.776 0.052 0.961* 0.066
Market-to-book value (#199, #25, #60) 0.010* 0.004 0.010* 0.003 0.011* 0.005 0.012* 0.005
ln (long-term debt) (#9), $ 0.021* 0.008 0.020* 0.010 0.054* 0.012 0.053* 0.017
ln (short-term investments) (#193), $ 0.001 0.006 0.003 0.008 0.020 0.008 0.019 0.012
Dividends per share (#26/#27), $ 0.201* 0.057 0.139* 0.059 0.179* 0.078 0.153* 0.077
* *
Lagged dependent variable 0.404 0.011 – 0.393 0.011 –
(between-to-within) variance ratio 0.168* 0.012 – 0.014* 0.005 –
Dividend Payments and Share Repurchases of US Firms

Within variance 29.619 – – 46.627 – –

2 x (maximized log-likelihood function) 52,088.12 – 35,401.14 –
Chi-square test exogeneity of dividends (8df) 10.15 11.53 12.14 11.26
Number of firms 1,845 1,845 1,141 1,141
Dependent variable share repurchases were based on items #226, #115, and #108 (Source: Bhargava 2010); 8-time observations at 2-yearly intervals on the
firms in 1992–2007 were used; slope coefficients and standard errors are reported; time dummies for 6 periods were included though their coefficients are not
reported in the tables
*
P < 0.05
1081
1082 A. Bhargava

In all the models in Table 38.5, firms’ market-to-book value ratio was estimated
with negative and significant coefficients in the dynamic and static models using the
pooled and disaggregated samples. Thus, firms with higher market-to-book value
ratio made lower share repurchases. Also, firms’ long-term debt was estimated with
negative and significant coefficients in dynamic models from the pooled sample and
from the subsample of firms making repurchases at least once. The short-run
elasticity from the pooled sample was 0.02, while the long-run elasticity was
0.03. Doubling of firms’ long-term debt predicted a 3 % decline in long-run share
repurchases. The coefficients of short-term investments were not significant in any
of the models in Table 38.5.
An important aspect of the results in Table 38.5 was that the estimated coeffi-
cients of dividends per share were negative and statistically significant in the
dynamic and static models for the pooled sample and for the subsample of firms
that made at least one repurchase. Thus, controlling for factors such as earnings and
size of the firm, firms paying higher dividends per share made significantly lower
share repurchases. The short-run effects of dividends per share on share repurchases
were 0.20 and 0.18, respectively, for the pooled sample and subsample of firms
making repurchases; the respective long-run effects were 0.33 and 0.30. The
relatively large magnitudes of these effects indicated that firms with higher divi-
dend payments were likely to make smaller share repurchases.
Last, exogeneity hypotheses for dividends per share were accepted at the 5 %
level using test statistics that were distributed as Chi-square variables with 8 of
freedom (5 % critical limit of Chi-square (8) ¼ 15.5). The values of Chi-square
statistics around 12 in Table 38.5 indicated that there might be some dependence in
the unobserved components of firms’ decisions to pay dividends and make share
repurchases though it was not significant at the 5 % level. Such hypotheses will be
tested below for the effects of share repurchases on firms’ dividend payments.

38.9.3 Results from Dynamic and Static Models for Dividends

per Share with Share Repurchases and Intangible Assets
Included as Explanatory Variables

The results from models for dividends per share and ratios of dividends to firms’
book value, with share repurchases included as a potentially endogenous explana-
tory variable, are in Table 38.6; firms’ intangible assets were included in these
models. Because short-term investments were not significant predictors of share
repurchases in Table 38.5, this variable was dropped from the models for dividends
and share repurchases in Tables 38.6 and 38.7, respectively. There were greater
number of missing observations on intangible assets and sample sizes in Table 38.6
were lower than sample sizes in Table 38.4.
The noteworthy feature of the results for dividends per share in Table 38.6 was
that coefficients of share repurchases was estimated with a small negative coeffi-
cient (0.004) in the dynamic model that reached statistical significance at the 5 %
level. However, this coefficient from the static model was 0.0002 and was not
 38

Table 38.6 Maximum likelihood estimates from comprehensive dynamic and static random effects models for 2-yearly averages of dividends per share paid
by US industrial firms in 1992–2007 with intangible assets and share repurchases included as explanatory variables
Dependent variable: dividends per share (#26/#27) Dependent variable: (dividends/book value) (#21/#60)
Dynamic model Static model Dynamic model Static model
Explanatory variables Coefficient SE Coefficient SE Coefficient SE Coefficient SE
Constant 11.348 0.844 10.855 0.612 0.077 0.027 0.068 0.010
Earnings (#18, #17), $million 0.0001* 0.00001 0.0001* 0.00001 0.00001* 0.000003 0.00001* 0.000001
* *
ln (total assets) (#6), $ 0.075 0.006 0.067* 0.004 0.005 0.002 0.005* 0.001
Market-to-book value (#199, #25, #60) 0.001 0.001 0.0004 0.0004 – –
ln (long-term debt) (#9), $ 0.0004 0.002 0.0003 0.001 0.00003 0.0005 0.0004* 0.0001
ln (intangible assets) (#33), $ 0.004* 0.001 0.003* 0.001 0.0002 0.0004 0.0001 0.0001
ln (share repurchases) (#226, #115, #108),$ 0.004* 0.002 0.0002 0.0005 0.0003 0.0004 0.0002 0.0001
Lagged dependent variable 0.191* 0.012 – 0.255* 0.023 –
(between-to-within) variance ratio 0.122* 0.013 – 0.012* 0.006 –
Dividend Payments and Share Repurchases of US Firms

Within variance 0.761 – 0.063 – –

2x (maximized log-likelihood function) 2,063.30 – 31,108.69 – –
Chi-square test exogeneity repurchases (8df) 6.235 6.74 13.26 11.23
Number of firms 1,429 1,429 1,418 1,418
Dependent variables are dividends per share (#26) divided by cumulative adjustment factor (#27) and the ratio of dividends to book value (#21/#60) (Source:
Bhargava 2010); 8-time observations at 2-yearly intervals on firms in 1992–2007 were used; slope coefficients and standard errors are reported; time dummies
for 6 periods were included though their coefficients are not reported
*
P < 0.05
1083
 1084

Table 38.7 Maximum likelihood estimates from dynamic and static random effects models for 2-yearly averages of share repurchases by US industrial firms
in 1992–2007 with intangible assets included as an explanatory variable
Dependent variable: ln (Share repurchases $) (Using Compustat items #226, #115, #108)
Specification 1: Specification 2:
Dynamic model Static model Dynamic model Static model
Explanatory variables: Coefficient SE Coefficient SE Coefficient SE Coefficient SE
Constant 89.993 5.547 103.897 8.953 81.508 4.659 94.684 8.243
Earnings (#18, #17), $million 0.0004* 0.0001 0.001* 0.0001 0.0004* 0.0001 0.001* 0.0001
* * *
ln (total assets) (#6), $ 0.524 0.039 0.602 0.058 0.517 0.036 0.599* 0.058
* *
Market-to-book value (#199, #25, #60) 0.012 0.0004 0.011 0.003 – –
ln (long-term debt) (#9), $ 0.027* 0.009 0.019 0.013 0.028* 0.009 0.021* 0.013
* * *
ln (intangible assets) (#33), $ 0.044 0.003 0.044 0.011 0.046 0.004 0.046* 0.011
* *
Dividend per share (#26/#27), $ 0.217 0.054 0.183 0.058 – –
Dividends/Book value (#21/#60) – – 0.583* 0.248 0.537* 0.207
Lagged dependent variable 0.412* 0.012 – 0.412* 0.011 –
*
(between-to-within) variance ratio 0.171 0.014 – 0.166* 0.014 –
Within variance 28.123 – – 28.310 – –
2x (maximized log-likelihood function) 39,778.71 – 39,523.88 –
Chi-square test exogeneity of dividends (8df) 9.62 11.30 10.34 15.40*
Number of firms 1,429 1,429 1,418 1,418
Dependent variable share repurchases were based on items #226, #115, and #108 (Source: Bhargava 2010); Specification 1 included dividends per share as
explanatory variable, while Specification 2 included dividends expressed in terms of book values; 8-time observations at 2-yearly intervals on firms in
1992–2007 were used; slope coefficients and standard errors are reported; time dummies for 6 periods were included though their coefficients are not reported
*
P < 0.05
A. Bhargava
38 Dividend Payments and Share Repurchases of US Firms 1085

significant. Also, coefficients of share repurchases in the model for the ratio of
dividends to book value were statistically not different from zero. Exogeneity
hypotheses for share repurchases could not be rejected using the Chi-square statis-
tics in dynamic and static models for dividends per share and for the ratio of
dividends to firms’ book value. Because share repurchases in the comprehensive
models generally did not affect dividend payments, it seems likely that firms
decided their dividends payments at an earlier stage than share repurchases. By
contrast, dividends per share were significant predictors of share repurchases in
Table 38.5.
Another important finding in Table 38.6 was that intangible assets were esti-
mated with negative and significant coefficients in dynamic and static models for
dividends per share. The results for other explanatory variables were similar to
those presented in Table 38.4, thereby showing robustness of the estimates to
changes in model specification and to reductions in sample sizes due to greater
number of missing observations on intangible assets.

38.9.4 Results from Dynamic and Static Models for Share

Repurchases with Intangible Assets Included as
Explanatory Variables

Table 38.7 presents the results for logarithms of share repurchases with intangible
assets included as explanatory variables in the models; in Specification 2, dividends
per share were replaced by the ratio of dividends to firms’ book values. The results
in Table 38.7 showed that intangible assets were positively and significantly
associated with share repurchases. This was in contrast with the results in Table 38.6
where intangible assets were negatively associated with dividends per share. Thus,
firms possessing higher intangible assets paid lower dividends per share and made
greater share repurchases; these results indicate the inappropriateness of modeling
the ratio of dividend payments to total payouts (Skinner 2008). Furthermore, while
intangible assets have received attention in the literature (Lev 2001), their effects on
dividends and share repurchases have not been rigorously investigated. For example,
it has been suggested that information technology firms may pay low (or zero)
dividends and make frequent repurchases in part because they have higher invest-
ments in research and development. A dummy variable was created using SIC
codes for information technology firms in manufacturing, transportation, and com-
munications sectors. However, the estimated coefficient was not significantly
different from zero. Thus, the reported coefficients in Table 38.7 appear not to
suffer from biases due to omission of such variables.
Further, dividends per share and the ratio of dividends to book value were
negatively and significantly associated with share repurchases, with large magni-
tudes of the estimated coefficients. Thus, the models with intangible assets as an
explanatory variable again showed that firms paying higher dividends made smaller
share repurchases. Also, exogeneity hypothesis for dividends-to-book value ratio
was close to rejection using the Chi-square statistic at 5 % level in the static model.
1086 A. Bhargava

The rejection of the null indicated that there was some dependence in the
unobserved factors affecting the firms’ decisions to pay dividends and make share
repurchases. For example, if firms made “unexpectedly” large share repurchases,
then such decisions in turn can affect dividend payments.
Finally, in view of the literature on choice of measures for share repurchases
(Banyi et al. 2008), the variable “purchase of common stock” was used as a proxy
for share repurchases, and dynamic and static models were estimated. The results
from modeling purchase of common stock were similar to those reported in
Tables 38.5 and 38.7 using the more complex definition of share repurchases.
However, greater numbers of observations were missing on purchase of common
stock variable and these were reduced in the more complex definition because
changes in Treasury stock were often positive so that the purchase of common stock
variable was not utilized. Overall, the results indicated that these two measures for
share repurchases yield similar results, perhaps because the models captured many
salient features of the interrelationships.

38.10 Conclusion

This chapter presented a detailed analysis of the interrelationships between divi-

dend payments and share repurchases by US industrial firms and modeled the
proximate determinants of the payout methods (Bhargava 2010). The usefulness
of analyzing longitudinal Compustat data on large number of firms and of appro-
priate econometric estimators was underscored. The empirical results provided
several insights that can be summarized as follows. First, the magnitudes of
dividends per share and dividend payments were broadly stable over the period of
1992–2007 using the data on the same firms. Moreover, simple dynamic models
showed that dividends were higher in the period 2003–2007, following reduction in
tax rates in 2003. These results were in some contrast with previous cross-sectional
findings where entry and exit by firms in the Compustat database affected the
numbers of firms paying dividends and magnitudes of dividend payments. For
example, dividends were not seen to be “disappearing” (Fama and French 2001)
and, apart from the effects of the tax cut in 2003, dividends were not “reappearing”
(Julio and Ikenberry 2004). The simple dynamic models for share repurchases
indicated steady increases in the period 1992–2007.
Second, the analytical and econometric frameworks led to comprehensive
dynamic models for dividends per share and repurchases. The empirical results
showed that variables such as firms’ earnings, total assets, and investments were
positively and significantly associated with dividends per share. Similar results
were found for share repurchases, though firms’ long-term debt was negatively and
significantly associated with repurchases but was significantly associated with
dividends only in a few models. Moreover, while dividends per share were nega-
tively and significantly associated with share repurchases, coefficients of share
38 Dividend Payments and Share Repurchases of US Firms 1087

repurchases were generally insignificant in the models for dividends. Exogeneity

hypotheses for dividends per share were accepted in most models for repurchases,
and exogeneity of share repurchases was also accepted in the models for dividends.
These results suggest that firms’ decisions to pay dividends at particular rates are
taken prior to the decisions to repurchase shares that are more strongly influenced
by the firms’ current financial situations. For example, while firms can adjust
repurchases given their debt levels, it may be more difficult to alter dividend
payments. Thus, interpreting negative correlations between dividend forecasting
errors and share repurchases as evidence of “substitution” between payment forms
(Grullon and Michaely 2002) seems inconsistent with the elaborate two stage
decision process documented by econometric analyses of the longitudinal
Compustat database.
Third, the effects of firms’ intangible assets on dividends per share were negative
and significant, while these coefficients were positive and significant in the models
for share repurchases. Such findings indicate that it is appropriate to model the two
payout methods via separate models rather than combining them into a single
variable. These results support the insights of Pearson (1897) and Neyman (1952)
cautioning against employing ratios of variables in applied work. Such problems are
exacerbated in longitudinal data analyses because stochastic properties of variables
combined can evolve differently over time. It might have been useful to disaggregate
intangible assets into components such as goodwill since they may differentially
affect dividends and repurchase decisions. Because of missing observations on such
variables in the Compustat database, further analyses were not pursued.
Finally, while econometric methods were useful for modeling the dynamic
interactions between dividend payments and share repurchases, it is important in
future research to investigate the role played by executive remuneration or com-
pensation in making such allocations. Options to purchase company stock at some
point in the future are regularly granted to the top executives and other employees
of US firms and these may have potentially increased firms’ share repurchases. If
such decisions in turn reduce funds available for investment and research and
development, then tax policies should discourage excessive repurchase activity.
The ExecuComp (2008) database contains information on salary, bonus, and stock
options granted to top executives of approximately 1,500 US firms from 1992.
A recent analysis of the merged ExecuComp and Compustat databases in fact
showed that beyond certain thresholds, share repurchases lower firms’ expenditures
on research and development and investments and should be discouraged
(Bhargava 2012).

Appendix

The concentrated or “profile” log-likelihood functions of the model in Eq. 38.3 was
computed by a FORTRAN program and was optimized using the numerical scheme
1088 A. Bhargava

E04 JBF from Numerical Algorithm Group (1991). Note that the data were read
firm-by-firm and several second moment matrices were created. Thus, there was no
limit on the number of firms in the sample. The profile likelihood function, which
depends only on the structural form parameters of interest, was computed using the
second moment matrices. FORTRAN programs for the computations of the likeli-
hood functions are available on request from the author.
The likelihood functions were separately computed for the case where the
variance-covariance matrix of the errors was unrestricted and where the errors
were decomposed in a random effect fashion, as in Eq. 38.2. Likelihood ratio
tests were applied to discriminate between these two alternative formulations.
Assuming that the number of firms (H) is large but the number of time observations
is fixed, asymptotic standard errors of the parameters were obtained by approxi-
mating second derivatives of the functions at the maximum.
Further, assuming that n1 and n2 time-varying variables are exogenous and
endogenous, respectively, as in Eq. 38.5, one can tackle the correlation between
the errors u’s and x2’s. It is reasonable to assume in short panels that a variable such
as dividends per share in Eq. 38.5 may be correlated only with firm-specific random
effects di. Thus, the correlation pattern can be decomposed as:

x2 i j t ¼ lj di þ x 2i j t (38.9)

where x*2 i j t are uncorrelated with di, and di are randomly distributed variables
with zero mean and finite variance as in Eq. 38.2. This correlation pattern was
invoked by Bhargava and Sargan (1983) and has the advantage that deviations of
x2 i j t’s from their time means:

xþ 2 i j t ¼ x2 i j t  x
2ij ðt ¼ 2, . . . , T; j ¼ n1 þ 1, . . . , n; i ¼ 1, . . . , HÞ
(38.10)

Where

X
T
x
2ij ¼ x2 i j t =T ðj ¼ n1 þ 1, . . . n; i ¼ 1, . . . , HÞ (38.11)
t¼1

can be used as additional [(T  1) n2] instrumental variables to facilitate identifi-

cation and estimation of parameters. Because the T deviations from means (x+2 i j t)
in Eq. 38.10 will sum to zero for every firm, one time observation per variable needs
to be omitted so that the (T1) instrumental variables are linearly independent. The
reduced form Eq. 38.6 for y1 is modified in this case as:

X
m Xn1 X
T n2 X
X T
yi 1 ¼ z i j zj þ uj k x1i j k þ mj k xþ 2i j k þ ui 1 ði ¼ 1, . . . , HÞ
j¼1 j¼1 k¼1 j¼1 k¼2

(38.12)
38 Dividend Payments and Share Repurchases of US Firms 1089

Further, likelihood ratio tests can be applied to test for exogeneity of time means
of the n2 endogenous variables (x2) in Eq. 38.11 by testing if correlations between
the T errors (ui t) affecting the dependent variables and n2 time means in Eq. 38.11
are zero. For example, given 8-time observations (T ¼ 8), likelihood ratio statistic
for testing zero correlation between errors affecting share repurchases in Eq. 38.4
(n2 ¼ 1) and time means of dividends per share is distributed for large H as
a Chi-square variable with 8 degrees of freedom.
The identification of parameters is achieved via the n1 time-varying exogenous
variables in the model (e.g., Sargan 1958). For example, in the system representing
(T1) time periods in Eq. 38.7, only the exogenous explanatory variables in time
period t explain the dependent variable in period t. The remaining (T1) n2 vari-
ables are excluded from the t th equation and are used in the set of instrumental
variables for identifying the coefficients of endogenous variables. Sufficient con-
ditions for identification, exploiting the time structure of longitudinal models, were
developed in Bhargava and Sargan (1983). For example, each exogenous time-
varying variable in the model for share repurchases in Eq. 38.4 effectively provides
eight exogenous variables of which seven are excluded from the equations. The null
hypothesis that dividends per share do not affect share repurchases can be tested
using the estimated coefficient b12 of dividends per share in Eq. 38.4. Also, one can
test if share repurchases affect dividends per share by including repurchases as
a potentially endogenous variable in Eq. 38.3.
For static version of the models not containing lagged dependent variables, one
can use stepwise estimation procedures with equality restrictions on coefficients
across time periods (Bhargava 1991). Efficient instrumental variables estimators
were used to estimate model parameters of Eq. 38.4, assuming the correlation
patterns for x2 i j t as in Eq. 38.9 and without restricting variance-covariance
matrices of the errors, i.e., the T x T dispersion matrix was assumed to be symmetric
and positive definite but not of the simple random effects form as in Eq. 38.2. This
formulation has the advantage that the errors vi t in Eq. 38.2 may be serially
correlated which was likely to be the case in static models since the lagged
dependent variables were omitted. Exogeneity hypotheses can be tested in the
stepwise estimation procedures via Chi-square tests that are asymptotically equiv-
alent to likelihood ratio tests.
Finally, while one can use “fixed” effects estimators (with dummy variables for
each firm) to circumvent certain endogeneity problems, increase in the number of
parameters with sample size leads to the problem of “incidental parameters”
(Neyman and Scott 1948). For example, coefficient of the lagged dependent variable
cannot be consistently estimated due to the incidental parameters in fixed effects
models unless the number of time observation is large. Moreover, from a modeling
standpoint, the use of random effects models obviates the need for estimating the
coefficients of large numbers of dummy variables for firms thereby enhancing the
efficiency of estimates. Exogeneity hypotheses for variables such as dividends per
share can be tested in the model for share repurchases. If the null were rejected, then
the models were estimated under the appropriate assumption that dividends per share
were correlated with errors affecting the models for share repurchases.
1090 A. Bhargava

References
Anderson, T. W., & Hsiao, C. (1981). Estimation of dynamic models with error components.
Journal of the American Statistical Association, 76, 598–606.
Banyi, M., Dyl, E., & Kahle, K. (2008). Errors in estimating share repurchases. Journal of
Corporate Finance, 14, 460–474.
Bhargava, A. (1991). Identification and panel data models with endogenous regressors. Reviews of
Economic Studies, 58, 129–140.
Bhargava, A. (1994). Modeling the health of Filipino children. Journal of the Royal Statistical
Society A, 157, 417–432.
Bhargava, A. (2008). Randomized controlled experiments in health and social sciences. Econom-
ics and Human Biology, 6, 293–298.
Bhargava, A. (2010). An econometric analysis of the dividends and share repurchases by
U.S. firms. Journal of the Royal Statistical Society A, 173, 631–656.
Bhargava, A. (2012, in press). Executive compensation, share repurchases and investment expen-
ditures: Econometric evidence from U.S. firms. Review of Quantitative Finance and
Accounting.
Bhargava, A., & Sargan, J. D. (1983). Estimating dynamic random effects models from panel data
covering short time periods. Econometrica, 51, 1635–1660.
Brav, A., Graham, J. R., Harvey, C. R., & Michaely, R. (2005). Payout policy in the 21st century.
Journal of Financial Economics, 77, 483–527.
Chen, S. -S., & Wang, Y. (2012, in press). Financial constraints and share repurchases. Journal of
Financial Economics.
Compustat. (2008). Standard & Poor’s Compustat. http://www.compustat.com/
Cox, D. (1958). The planning of experiments. New York: Wiley.
De Angelo, H., De Angelo, L., & Stultz, R. (2006). Dividend policy and the earned/contributed
capital mix: A test of the life cycle theory. Journal of Financial Economics, 81, 227–254.
Duesenberry, J. (1949). Income, saving, and the theory of consumer behavior. Cambridge:
Harvard University Press.
ExecuComp. (2008). Standard & Poor’s ExecuComp. http://www.compustat.com/products.aspx?
id¼2147492873
Fama, E., & Babiak, H. (1968). Dividend policy: An empirical analysis. Journal of the American
Statistical Association, 63, 1132–1161.
Fama, E., & French, K. (2001). Disappearing dividends: Changing firm characteristics or lower
propensity to pay? Journal of Financial Economics, 60, 3–43.
Fenn, G. W., & Liang, N. (2001). Corporate payout policy and managerial stock incentives.
Journal of Financial Economics, 60, 45–72.
Granger, C. (1992). Evaluating economic theory. Journal of Econometrics, 51, 3–5.
Grullon, G., & Michaely, R. (2002). Dividends, share repurchases and the substitution hypothesis.
Journal of Finance, 57, 1649–1684.
Gu, F., & Wang, W. (2005). Intangible assets, information complexity, and analysts’ earnings
forecasts. Journal of Business Finance and Accounting, 32, 1673–1702.
Guay, W., & Harford, J. (2000). The cash-flow permanence and information content of dividend
increases versus repurchases. Journal of Financial Economics, 57, 385–415.
Jagannathan, M., Stephens, C., & Weisbach, M. (2000). Financial flexibility and the choice
between dividends and repurchases. Journal of Financial Economics, 57, 355–384.
John, K., & Williams, J. (1985). Dividends, dilution, and taxes: A signaling equilibrium. Journal
of Finance, 40, 1053–1070.
Jones, R., & Wu, Y. (2010). Executive compensation, earnings management and share holder
litigation. Review of Quantitative Finance and Accounting, 35, 1–20.
Julio, B., & Ikenberry, D. (2004). Reappearing dividends. Journal of Applied Corporate Finance,
16, 89–100.
38 Dividend Payments and Share Repurchases of US Firms 1091

Kahle, K. (2002). When a buyback isn’t a buyback: Open market repurchases and employee
options. Journal of Financial Economics, 63, 235–261.
Koopmans, T. C. (Ed.). (1950). Statistical inference in dynamic economic models. New York:
Wiley.
Kronmal, R. (1993). Spurious correlation and the fallacy of the ratio standard revisited. Journal of
the Royal Statistical Society A, 156, 379–392.
Lee, C., Wu, C., & Djarraya, M. (1987). A further empirical investigation of the dividend
adjustment process. Journal of Econometrics, 35, 267–285.
Lev, B. (2001). Intangibles. Washington, DC: Brookings Institution Press.
Lintner, J. (1956). Distributions of incomes of corporations among dividends, retained earnings,
and taxes. American Economic Review, 46, 97–113.
Miller, M., & Modigliani, F. (1961). Dividend policy, growth, and the valuation of shares. Journal
of Business, 34, 411–433.
Neyman, J. (1952). Lectures and conferences on mathematical statistics and probability (2nd ed.,
pp. 143–154). Washington, DC: U.S. Department of Agriculture.
Neyman, J., & Scott, E. (1948). Consistent estimates based on partially consistent observations.
Econometrica, 16, 1–32.
Numerical Algorithm Group. (1990). Numerical Algorithm Group library. Mark 13. Oxford:
Oxford University.
Pearson, K. (1897). Mathematical contributions to the theory of evolution- on a form of spurious
correlation which may arise when indices are used in the measurement of organs. Proceedings
of the Royal Society of London, 60, 489–497.
Sargan, J. D. (1958). The estimation of economic relationships using instrumental variables.
Econometrica, 26, 393–415.
Skinner, D. J. (2008). The evolving relation between earnings, dividends, and stock repurchases.
Journal of Financial Economics, 87, 582–609.
Stata. (2008). Stata version 10. College Station: StataCorp.
Stephens, C., & Weisbach, M. (1998). Actual share reacquisition in open-market repurchase
programs. Journal of Finance, 53, 313–333.
Tinbergen, J. (1939). The dynamics of share-price formation. Review of Economics and Statistics,
21, 153–160.
Von Eije, H., & Megginson, W. (2008). Dividends and share repurchases in the European Union.
Journal of Financial Economics, 89, 347–374.
Term Structure Modeling and Forecasting
Using the Nelson-Siegel Model 39
Jian Hua

Contents
39.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1094
39.2 Modeling the Term Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1094
39.3 Fitting the Yield Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
39.3.1 Data Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
39.3.2 Estimation Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
39.4 Forecasting the Yield Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
39.5 Other Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1099
39.6 Generalized Duration Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1100
39.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
Appendix 1: One-Step Estimation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
Appendix 2: Two-Step Estimation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103

Abstract
In this chapter, we illustrate some recent developments in the yield curve
modeling by introducing a latent factor model called the dynamic Nelson-Siegel
model. This model not only provides good in-sample fit, but also produces
superior out-of-sample performance. Beyond Treasury yield curve, the model
can also be useful for other assets such as corporate bond and volatility.
Moreover, the model also suggests generalized duration components
corresponding to the level, slope, and curvature risk factors.
The dynamic Nelson-Siegel model can be estimated via a one-step procedure,
like the Kalman filter, which can also easily accommodate other variables of
interests. Alternatively, we could estimate the model through a two-step process
by fixing one parameter and estimating with ordinary least squares. The model is
flexible and capable of replicating a variety of yield curve shapes: upward

J. Hua
Baruch College (CUNY), New York, NY, USA
e-mail: jian.hua@baruch.cuny.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1093
DOI 10.1007/978-1-4614-7750-1_39,
# Springer Science+Business Media New York 2015
1094 J. Hua

sloping, downward sloping, humped, and inverted humped. Forecasting the yield
curve is achieved through forecasting the factors and we can impose either
a univariate autoregressive structure or a vector autoregressive structure on the
factors.

Keywords
Term structure • Yield curve • Factor model • Nelson-Siegel curve • State-space
model

39.1 Introduction

There have been major advances in theoretical term structure models, as well as
their econometric estimation. Two popular approaches to term structure modeling
are no-arbitrage models and equilibrium models. The no-arbitrage tradition focuses
on eliminating arbitrage opportunities by perfectly fitting the term structure at
a point in time, which is important for pricing derivatives; see Hull and White
(1990) and Health et al. (1992), among others. The equilibrium tradition models the
dynamics of the instantaneous rate typically through affine models so that yields at
other maturities can be derived under various assumptions about the risk premium
(e.g., Vasicek 1977; Cox et al. 1985; Duffie and Kan 1996).
For many finance questions, such as bond portfolio management, derivatives pricing,
and risk management, it is crucial to both produce accurate estimates of the current term
structure as well as forecast future interest rate dynamics. One class of models that has
one potential satisfactory answer to these questions is that of the Nelson-Siegel class of
models (see, Nelson and Siegel 1987). Here, we survey some recent developments of
this model. This model not only provides good in-sample fit, but also produces superior
out-of-sample performance. Moreover, the model also suggests generalized duration
components corresponding to the level, slope, and curvature risk factors.

39.2 Modeling the Term Structure

Let Pt(t) be the date t price of a zero-coupon riskless bond that pays $1 in t periods.
Then,

Pt ðtÞ ¼ expðtyt ðtÞÞ: (39.1)

In practice, yield curves, discount curves, and forward curves are not observed.
Instead, they must be estimated from observed bond prices. A popular approach is
to estimate forward rates at the observed maturities, and then, construct
unsmoothed yields by averaging appropriate estimated forward rates. These yields
exactly price the included bonds (see Fama and Bliss 1987).
The original Nelson and Siegel (1987) framework is a convenient and parsimo-
nious three-component exponential approximation. They work with the forward
39 Term Structure Modeling and Forecasting Using the Nelson-Siegel Model 1095

Fig. 39.1 Factor loadings.

This figure plots the factor 1
loadings when l is prefixed β1 Loadings
at 0.0609
0.8

Loadings
0.6 β2 Loadings

0.4

0.2 β3 Loadings

0
0 20 40 60 80 100 120
τ (Maturity, in Months)

rate, which can be viewed as a constant plus a Laguerre function. Diebold and
Li (2006) made it dynamic, and thus, the corresponding yield curve is



1  elt 1  elt lt
yt ðtÞ ¼ b1t þ b2t þ b3t e þ et , (39.2)
lt lt

and call it DNS hereafter. b1 changes all yields uniformly, and it can be called the
level factor (Lt). b2 loads the short rate more heavily, its loading decays to zero as
maturity lengthens, and it can be called the slope factor (St). b3 loads the medium
term more heavily, its loading starts at zero and decays back to zero as maturity
increases, and it can be called the curvature factor (Ct). In Fig. 39.1, we plot the
factor loadings for a specific value of l. l determines the maturity at which the
medium-term (or the curvature factor) loading achieves its maximum, which effec-
tively controls the location of the hump of the curve. The DNS model is different
from the factor analysis, in which one estimates both the unobserved factors and
factor loadings. Here, the framework imposes structure on the factor loadings.

39.3 Fitting the Yield Curve

39.3.1 Data Construction

The standard way of measuring the term structure is by means of the spot rate curve
(or equivalently the yield-to-maturity) on zero-coupon bonds.1 The problem with

1
Zero coupon bonds are chosen to limit the “coupon effect,” which implies that two bonds that are
identical in every respect except for bearing different coupon rates can have a different yield-to-
maturity.
1096 J. Hua

zero-coupon yields is that these are not usually directly observed. For example, the
US Treasury Bills do not bear coupon, but they are only available for maturities of
1-year or less, and the longer maturity zero-coupon yields need to derived from
coupon bearing Treasury Notes and Bonds. A popular approach to resolve such an
issue is the bootstrapping procedure by Fama and Bliss (1987), which sequentially
extracts forward rates from bond prices with successively longer maturities and
then takes advantages of the interchangeable nature of the spot rate curve, discount
curve, and forward rate curve.

39.3.2 Estimation Procedure

Because of the structure the model imposes, estimation can be achieved with high
precision. There are two popular approaches: the one-step and two-step procedures.
The one-step procedure can be achieved in two ways. One method is simply to
estimate the model by nonlinear least squares for each month t. The other way is to
transform the system into a state-space representation and estimate l and the factors
via a Kalman filter. Alternatively, for a two-step procedure, Diebold and Li (2006)
advocate prefixing the l, and estimating the factors via ordinary least squares (OLS).
The appendix provides more discussion of these two approaches.
As we can see in Fig. 39.2, the model is flexible and is capable of replicating
a variety of yield curve shapes: upward sloping, downward sloping, humped, and
inverted humped. Figure 39.3 displays the autocorrelations of the estimated factors
and residuals. The level factor displays high persistence and is, of course, positive.
In contrast, the slope and curvature factors are less persistent and assume both
positive and negative values.

39.4 Forecasting the Yield Curve

Forecasting the yield curve can be achieved through forecasting the factors. We can
impose either a univariate autoregressive structure or a vector autoregressive
structure on the factors.
The yield forecasts based on underlying univariate AR(1) factor specifications
are



1  elt t 1  elt t
^y tþh=t ðtÞ ¼ b1, tþh=t þ b2, tþh=t þ b3, tþh=t  elt t , (39.3)
lt t lt t

where b forecasts are made by the following:

^ ^
c 1 þ ^g 1 b
b 1, tþh=t ¼ ^ 1t
^
b ¼ ^
c þ ^g ^
b (39.4)
2, tþh=t 2 2 2t
^
b ¼ ^ þ ^g ^
3, tþh=t c 3 3 b 3t
39 Term Structure Modeling and Forecasting Using the Nelson-Siegel Model 1097

Yield Curve on 3/31/1989 Yield Curve on 7/31/1989

9.8 8
9.7 7.9
9.6

Yield (Percent)
Yield (Percent)

7.8
9.5
9.4 7.7
9.3 7.6
9.2
7.5
9.1
9 7.4
8.9 7.3
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Maturity (Months) Maturity (Months)
Yield Curve on 5/30/1997 Yield Curve on 8/31/1998
6.8 5.05
6.6
6.4 5
Yield (Percent)

Yield (Percent)

6.2
4.95
6
5.8
4.9
5.6
5.4 4.85
5.2
5 4.8
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Maturity (Months) Maturity (Months)

Fig. 39.2 Selected fitted yield curves. This figure presents fitted yield curve for selected dates,
together with actual yields (Source: Diebold and Li 2006)

The yield forecasts based on an underlying multivariate autoregressive specifi-

cation (Vector AR(1)) are



1  elt t 1  elt t
^y tþh=t ðtÞ ¼ b1, tþh=t þ b2, tþh=t þ b3, tþh=t  elt t , (39.5)
lt t lt t

where

^ ^,
c þ ^g b
b tþh=t ¼ ^ t (39.6)

and b ¼ {b1,b2,b3,}0 .
The random walk model,

^y tþh=t ðtÞ ¼ yt ðtÞ, (39.7)

is a no-change forecast model, which is used as the benchmark.2

2
Diebold and Li (2006) conduct a more extensive comparisons with competing models.
1098 J. Hua

Autocorrelation of βˆ 1t Autocorrelation of εˆ 1t
1.0 0.2

0.8
0.1

Autocorrelation
Autocorrelation

0.6

0.4
0.0
0.2

0.0 −0.1
−0.2

−0.4 −0.2
5 10 15 20 25 30 35 40 45 50 55 60 5 10 15 20 25 30 35 40 45 50 55 60
Displacement Displacement
Autocorrelation of βˆ 2t Autocorrelation of εˆ 2t
1.0 0.2

0.8
0.1
Autocorrelation

Autocorrelation

0.6

0.4
0.0
0.2

0.0 −0.1
−0.2

−0.4 −0.2
5 10 15 20 25 30 35 40 45 50 55 60 5 10 15 20 25 30 35 40 45 50 55 60
Displacement Displacement
Autocorrelation of βˆ 3t Autocorrelation of εˆ 3t
1.0 0.2

0.8
0.1
Autocorrelation
Autocorrelation

0.6

0.4
0.0
0.2

0.0 −0.1
−0.2

−0.4 −0.2
5 10 15 20 25 30 35 40 45 50 55 60 5 10 15 20 25 30 35 40 45 50 55 60
Displacement Displacement

Fig. 39.3 Autocorrelation and residual autocorrelation. This figure presents sample autocorrela-
tions of the level, slope, and curvature factors, as well as the sample autocorrelations of AR
(1) models t to the three estimated factors, along with Barletts approximate 95 condence bands
(Source: Diebold and Li 2006)

We define forecast errors at t + h as ytþh ðtÞ  ^y tþhjt ðtÞ. Table 39.1 reports the
root mean-squared errors of the out-of-sample performance of the DNS model
versus the random walk model. An error of 0.235 indicates a mean-squared error
of 23.5 basis points in the yield prediction. For 1 month ahead, the random walk is
39 Term Structure Modeling and Forecasting Using the Nelson-Siegel Model 1099

Table 39.1 Out-of-sample forecasting results

3-month 1-year 2-year 3-year 5-year 10-year 30-year
1 month a head
Random walk 0.235 0.259 0.290 0.300 0.288 0.257 0.211
Unrestricted VAR 0.213 0.261 0.307 0.302 0.310 0.314 0.419
Univariate AR 0.248 0.282 0.315 0.301 0.311 0.314 0.418
6 months ahead
Random walk 0.910 0.904 0.938 0.857 0.813 0.655 0.570
Unrestricted VAR 0.909 1.050 1.078 0.999 0.949 0.796 0.717
Univariate AR 0.694 0.890 0.908 0.881 0.827 0.651 0.486
1 years ahead
Random walk 1.552 1.505 1.343 1.190 1.065 0.858 0.622
DNS Unrestricted VAR 1.643 1.731 1.641 1.483 1.338 1.123 0.979
DNS Univariate AR 1.383 1.445 1.338 1.182 1.041 0.819 0.615
This table presents the results of out-of-sample forecasts using VAR and univariate AR specifi-
cations of the DNS factors. I estimate all models recursively from 1987:1 to the time the forecast
is made, beginning in 1997:1 and extending through 2002:12. I define forecast errors at t + h as
ytþh ðtÞ  ^y tþh=t ðtÞ and report the root mean-squared errors versus random walk.

hard to beat, but matters improve dramatically. The DNS model with an AR
(1) specification produces smaller errors, so it clearly has advantages (see Diebold
and Li 2006 for more details). Remember here, we only illustrate the point forecast,
and it can be extended easily to interval forecasting since they are useful for risk
management.

39.5 Other Applications

Beyond the US Treasury yield curve, the Nelson-Siegel model has also shown
success in fitting and forecasting other assets and global yield curves (see Krishnan
et al. 2010; Diebold et al. 2008; Hua 2010). Since the model produces three factors
that capture the information in the entire term structure, we can analyze the
dynamic relationship among them. For example, Hua (2010) analyzes the dynamic
interaction between the credit spread term structure and equity option implied
volatility term structure.
Moreover, the DNS framework can be easily augmented with other variables
of interest. We could also analyze how macroeconomic variables are linked
with the interest rate term structure; see Auroba et al. (2006). This can be done
easily in a state-space framework. Frequently, yield variation reacts to
macrovariables, and macrovariables can also be impacted by the yields. Poten-
tially, we can improve our forecasts of one by incorporating information from
the other.
1100 J. Hua

39.6 Generalized Duration Measure

Traditional interest rate risk management focuses on duration and duration man-
agement, which only considers parallel shifts of the yield curve. However, in
practice, nonparallel shifts do exist and these are a significant source of risk. The
DNS model presents a generalized duration component that corresponds to the
level, slope, and curvature risk factors.
We can define a bond duration measure as follows. Let the cash flows from bond
be C1,C2, . . .,CI, and define the associated maturities to be t1,t2,. . .,tI. We assume
that the yield curve is linear in some arbitrary factors f1, f2, and f3,

yt ðtÞ ¼ B1 ðtÞf 1t þ B2 ðtÞf 2t þ B3 ðtÞf 3t : (39.8)

This is consistent with the DNS setup. Since the price of the bond can be
expressed as,

X
I
P¼ Ci eti yt ðti Þ , (39.9)
i¼1

for an arbitrary change of the yield curve, the price change is

X XI h i
I
∂P
dP ¼ dyi ðti Þ ¼ Ci eti yt ðti Þ ðti Þ dyt ðti Þ, (39.10)
i¼1
∂yt ðti Þ i¼1

Where yt(ti) are treated as independent variables. Rearranging terms, we can

express the percentage change in bond price as a function of changes in the factors

( ) ( )
I  
dP X 3 X 1 ti yt ðti Þ
X3 XI
 ¼ Ci e ti Bj ðti Þ df it ¼ wi ti Bj ðti Þ df it ,
P j¼1 i¼1
P j¼1 i¼1

(39.11)

where wi is the weight associated with Ci.

Since we have decomposed the bond price changes into risk factor changes, the
duration component associated with each risk factor is, for j ¼ 1,2,3,

X
I
Dj ¼ wi ti Bj ðti Þ: (39.12)
i¼1
39 Term Structure Modeling and Forecasting Using the Nelson-Siegel Model 1101

Moreover, based on the dynamic Nelson-Siegel model, the vector duration is

X
I
D1 ¼ wi ti
i¼1
XI
1  elti
D2 ¼ wi (39.13)
i¼1
l
X

I
1  elti
D3 ¼ wi  ti elti :
i¼1
l

The vector duration measure has the following properties:

• D1, D2, and D3 increase with maturity t.
• D1, D2, and D3 decrease with coupon rate.
• D1, D2, and D3 decrease with yield-to-maturity.
Note that D1 is exactly the tradition Macaulay duration.
Diebold et al. (2006) apply this vector duration measure as immunization tools in
a practical bond portfolio management context. They report that hedging based on
the vector duration outperforms hedging based on Macaulay duration in almost all
samples and all holding periods. Moreover, it also outperforms polynomial vector
duration during unusual market situations such as the monetary regime of the early
80s. Therefore, the vector duration measure seems to be an appealing risk
management tool.

39.7 Summary

We have presented the DNS model as a capable and flexible model that can fit the
term structure of interest rates in-sample and predict them out-of-sample. It can
easily be extended to a different asset class or accommodate other variables into the
specifications. Given the advantages of the dynamic Nelson-Siegel model in many
dimensions, it is gaining in popularity. Recently, the Board of Governors of the
Federal Reserve System has started publishing the daily estimated factors of the
Treasury yield curve on their website.3
One potential drawback of the DNS model is that it does not rule out arbitrage
opportunities. Recently, an arbitrage free version of the model has been developed;
see Christensen et al. (2009, 2011).

Appendix 1: One-Step Estimation Method

We illustrate the one-step method to estimate the DNS model. If the dynamics of
betas (factors) follow a vector autoregressive process of first order, the model

3
The data can be found at http://www.federalreserve.gov/econresdata/researchdata.htm
1102 J. Hua

immediately forms a state-space system since ARMA state vector dynamics of any
order may be transformed into a state-space form. Thus, the transition equation is
0 1 0 10 1 0 1
b1, t  mb1 a11 a12 a13 b1, t1  mb1 1, t
@ b2, t  mb2 A ¼ @ a21 a22 a23 A@ b2, t1  mb2 A þ @ 2, t A, (39.14)
b3, t  mb3 a31 a32 a33 b3, t1  mb3 3, t

t ¼ 1, . . ., T. The measurement equation, which relates N yields to the three

unobserved factors, is
0 1
1  et1 l 1  et 1 l t 1 l
B 1  e C
0 1 B t1 l t1 l C 0 1
B 1  et l 1  et 2 l C
y t ð t1 Þ B1 t 2 l C0 1 e1, t
B y t ð t2 Þ C B
2
e C b1 , t
B C¼B t2 l t2 l C@ b A B e2, t C
@ ⋮ A B B C 2, t þB C
@ ⋮ A, (39.15)
C
B⋮ ⋮ ⋮ C b3, t
y t ð tN Þ B C eN , t
B C
@ 1  et l 1  et N l A
1 1
 et N l
tN l tN l

t ¼ 1, . . ., T. In matrix notation, we rewrite the state-space system as measure-

ment equation,

yt ¼ Gft þ et ; (39.16)

state equation,

ðft  mÞ ¼ Aðft1  mÞ þ t : (39.17)

For optimality of the Kalman filter, we assume that the white noise transaction
and measurement errors be orthogonal to each other.


   
t 0 Q 0
 WN ; , (39.18)
et 0 0 H
 0
E f0 t ¼ 0, (39.19)
 0
E f0 et ¼ 0: (39.20)

The state-space set up with the application of the Kalman filter delivers maximum-
likelihood estimates and smoothed underlying factors, where all parameters are
estimated simultaneously. Such representation also allows for heteroskedasticity,
missing data, or heavy-tailed measurement errors. Moreover, other useful variables,
such as macroeconomic variables, can be augmented into the state equation to
understand the dynamic interactions between the yield curve and the macroeconomy.
39 Term Structure Modeling and Forecasting Using the Nelson-Siegel Model 1103

Appendix 2: Two-Step Estimation Method

The estimation for the two-step procedure requires us to choose a l value first. Once
the l is fixed, the values of the two regressors (factor loadings) can be computed, so
ordinary least squares can be applied to estimate the betas (factors) at each period t.
Doing so is not only simple and convenient, but also eliminates the potential for
numerical optimization challenges. The question is: What is the appropriate value
of l. Recall that l determines the maturity at which the loading achieves its
maximum. For example, for Treasury data, 2 or 3 years are commonly considered
medium term, so a simple average of the two is 30-month. The l value that
maximizes the loadings on the medium-term factor at exactly 30-month is
0.0609. For a different market, the medium maturity could be different, so the
corresponding l value could be different as well.

References
Aruoba, B. S., Diebold, F. X., & Rudebusch, G. D. (2006). The macroeconomy and yield curve:
A dynamic latent factor approach. Journal of Econometrics, 127, 309–338.
Christensen, J. H. E., Diebold, F. X., & Rudebusch, G. D. (2009). An arbitrage-free generalized
Nelson-Siegel term structure model. Econometrics Journal, 12, 33–64.
Christensen, J. H. E., Diebold, F. X., & Rudebusch, G. D. (2011). The affine arbitrage-free class of
Nelson-Siegel term structure models. Journal of Econometrics, 164, 4–20.
Cox, J. C., Ingersoll, J. E., & Ross, S. A. (1985). A theory of the term structure of interest rates.
Econometrica, 53, 385–407.
Diebold, F. X., Ji, L., & Li, C. (2006). A three-factor yield curve model: Non-affine structure,
systematic risk sources, and generalized duration. In L. R. Klein (Ed.), Macroeconomics,
finance and econometrics: Essays in memory of Albert Ando (pp. 240–274). Cheltenham:
Edward Elgar.
Diebold, F. X., & Li, C. (2006). Forecasting the term structure of government bond yields. Journal
of Econometrics, 127, 337–364.
Diebold, F. X., Li, C., & Yue, V. (2008). Global yield curve dynamics and interactions:
A generalized Nelson-Siegel approach. Journal of Econometrics, 146, 351–363.
Duffie, D., & Kan, R. (1996). A yield-factor model of interest rates. Mathematical Finance,
6, 379–406.
Fama, E., & Bliss, R. (1987). The information in long-maturity forward rates. American Economic
Review, 77, 680–692.
Heath, D., Jarrow, R., & Morton, A. (1992). Bond pricing and the term structure of interest rates:
A new methodology for contingent claims valuation. Econometrica, 60, 77–105.
Hua, J. (2010). Option implied volatilities and corporate bond yields: A dynamic factor approach
(Working Paper). Baruch College, City University of New York.
Hull, J., & White, A. (1990). Pricing interest-rate-derivative securities. Review of Financial
Studies, 3, 573–592.
Krishnan, C. N. V., Ritchken, P., & Thomson, J. (2010). Predicting credit spreads. Journal of
Financial Intermediation, 19(4), 529–563.
Nelson, C. R., & Siegel, A. F. (1987). Parsimonious modeling of yield curve. Journal of Business,
60, 473–489.
Vasicek, O. (1977). An equilibrium characterization of the term structure. Journal of Financial
Economics, 5, 177–188.
The Intertemporal Relation Between
Expected Return and Risk on Currency 40
Turan G. Bali and Kamil Yilmaz

Contents
40.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106
40.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1109
40.3 Estimation Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114
40.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117
40.5 Time-Varying Risk Aversion in the Foreign Exchange Market . . . . . . . . . . . . . . . . . . . . . . 1125
40.6 Testing Merton’s (1973) ICAPM in Currency Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1130
40.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
Appendix 1: Maximum Likelihood Estimation of GARCH-in-Mean Models . . . . . . . . . . . . . . . 1135
Appendix 2: Estimation of a System of Regression Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1137
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1139

Abstract
The literature has so far focused on the risk-return trade-off in equity markets
and ignored alternative risky assets. This paper examines the presence and
significance of an intertemporal relation between expected return and risk in
the foreign exchange market. The paper provides new evidence on the
intertemporal capital asset pricing model by using high-frequency intraday
data on currency and by presenting significant time variation in the risk aversion
parameter. Five-minute returns on the spot exchange rates of the US dollar
vis-à-vis six major currencies (the euro, Japanese yen, British pound sterling,
Swiss franc, Australian dollar, and Canadian dollar) are used to test the existence
and significance of a daily risk-return trade-off in the FX market based on the

T.G. Bali (*)

McDonough School of Business, Georgetown University, Washington, DC, USA
e-mail: tgb27@georgetown.edu
K. Yilmaz
College of Administrative Sciences and Economics, Koc University, Istanbul, Turkey
e-mail: kyilmaz@ku.edu.tr

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1105
DOI 10.1007/978-1-4614-7750-1_40,
# Springer Science+Business Media New York 2015
1106 T.G. Bali and K. Yilmaz

GARCH, realized, and range volatility estimators. The results indicate a positive
but statistically weak relation between risk and return on currency.
Our empirical analysis relies on the maximum likelihood estimation of the
GARCH-in-mean models as described in Appendix 1. We also use the seem-
ingly unrelated (SUR) regressions and panel data estimation to investigate the
significance of a time-series relation between expected return and risk on
currency as described in Appendix 2.

Keywords
GARCH • GARCH-in-mean • Seemingly unrelated regressions (SUR) • Panel
data estimation • Foreign exchange market • ICAPM • High-frequency data •
Time-varying risk aversion • High-frequency data • Daily realized volatility

40.1 Introduction

Merton’s (1973) intertemporal capital asset pricing model (ICAPM) indicates that
the conditional expected excess return on a risky market portfolio is a linear
function of its conditional variance plus a hedging component that captures the
investor’s motive to hedge for future investment opportunities. Merton (1980)
shows that the hedging demand component becomes negligible under certain
conditions and the equilibrium relation between risk and return is defined as


Et ðRtþ1 Þ ¼ b  Et s2tþ1 (40.1)

where Et(Rt+1) and Et(s2t+1) are, respectively, the conditional mean and variance of
excess returns on a risky market portfolio and b > 0 is the risk aversion parameter
of market investors. Equation 40.1 establishes the dynamic relation that investors
require a larger risk premium at times when the market is riskier.
Many studies investigate the significance of an intertemporal relation between
expected return and risk in the aggregate stock market. However, the existing
literature has not yet reached an agreement on the existence of a positive risk-
return trade-off for stock market indices.1 Due to the fact that the conditional mean
and volatility of the market portfolio are not observable, different approaches,
different data sets, and different sample periods used by previous studies in
estimating the conditional mean and variance are largely responsible for the
contradictory empirical evidence.
The prediction of Merton (1973, 1980) that expected returns should be related to
conditional risk applies not only to the stock market portfolio but also to any risky
portfolio. However, earlier studies have so far focused on the risk-return trade-off in

1
See French et al. (1987), Campbell (1987), Nelson (1991), Campbell and Hentschel (1992), Chan
et al. (1992), Glosten et al. (1993), Scruggs (1998), Harvey (2001), Goyal and Santa-Clara (2003),
Brandt and Kang (2004), Ghysels et al. (2005), Bali and Peng (2006), Christoffersen and Diebold
(2006), Guo and Whitelaw (2006), Lundblad (2007), and Bali (2008).
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1107

equity markets and ignored other risky financial assets. Although there are a few
studies testing the significance of a time-series relation between risk and return
in international equity markets, the focus is generally on the US stock market. It is
also important to note that earlier studies assume a constant risk-return trade-off and
ignore time variation in the risk aversion parameter b.2 This paper examines the
intertemporal relation between expected return and risk in currency markets. The
paper not only investigates ICAPM in the foreign exchange market but examines
the significance of time-varying risk aversion as well.
The foreign exchange market includes the trading of one currency against
another between large banks, central banks, currency speculators, multinational
corporations, governments, and other financial markets and institutions. The
FX market is an interbank or inter-dealer network first established in 1971 when
many of the world’s major currencies moved towards floating exchange rates. It is
considered an over-the-counter (OTC) market, meaning that transactions are
conducted between two counterparties that agree to trade via telephone or elec-
tronic network. Because foreign exchange is an OTC market where brokers/dealers
negotiate directly with one another, there is no central exchange or clearing house.3
The FX market has grown rapidly since the early 1990s. According to the
triennial central bank surveys conducted by the Bank for International Settlements
(BIS), the April 2007 data show an unprecedented rise in activity in traditional
foreign exchange markets compared to 2004. As shown in Table 40.1, average daily
turnover rose to US $3.1 trillion in April 2007, an increase of 69 % (compared to
April 2004) at current exchange rates and 63 % at constant exchange rates.4 Since
April 2001, average daily turnover in foreign exchange markets worldwide
(adjusted for cross-border and local double-counting and evaluated at April 2007
exchange rates) increased by 58 % and 69 % between two consecutive triennial
surveys. Comparing the average daily turnovers of US $500 billion in 1988 and US
$3.1 trillion in 2007 indicates that trading volume in FX markets increased by more
than five times over the past two decades.
The FX market has become the world’s largest financial market, and it is not
uncommon to see over US $3 trillion traded each day. By contrast, the New York
Stock Exchange (NYSE) – the world’s largest equity market with daily trading
volumes in the US $60–80 billion dollar range – is positively dwarfed when
compared to the FX market. Daily turnover in FX markets is now more than ten
times the size of the combined daily turnover on all the world’s equity markets.

2
A few exceptions are Chou et al. (1992), Harvey (2001), and Lettau and Ludvigson (2010).
3
As FX trading has evolved, several locations have emerged as market leaders. Currently, London
contributes the greatest share of transactions with over 32 % of the total trades. Other trading
centers – listed in order of volume – are New York, Tokyo, Zurich, Frankfurt, Hong Kong, Paris,
and Sydney. Because these trading centers cover most of the major time zones, FX trading is a true
24-h market that operates 5 days a week.
4
In addition to “traditional” turnover of US $3.1 trillion in global foreign exchange market, US
$2.1 trillion was traded in currency derivatives.
1108 T.G. Bali and K. Yilmaz

Table 40.1 Reported foreign exchange market turnover by currency paira (Daily averages in
April, in billions of US dollars and percent)
2001 2004 2007
Amount % share Amount % share Amount % share
US dollar/euro 354 30 503 28 840 27
US dollar/yen 231 20 298 17 397 13
US dollar/British pound 125 11 248 14 361 12
US dollar/Australian dollar 47 4 98 5 175 6
US dollar/Swiss franc 57 5 78 4 143 5
US dollar/Canadian dollar 50 4 71 4 115 4
US dollar/other 195 17 295 16 628 21
Euro/yen 30 3 51 3 70 2
Euro/sterling 24 2 43 2 64 2
Euro/Swiss franc 12 1 26 1 54 2
Euro/other 21 2 39 2 112 4
Other currency pairs 26 2 42 2 122 4
All currency pairs 1,173 100 1,794 100 3,081 100
a
Adjusted for local and cross-border double-counting

Even when combining the US bond and equity markets, total daily volumes still do
not come close to the values traded on the currency market.
The FX market is unique because of its trading volumes; the extreme liquidity
of the market; the large number of, and variety of, traders in the market; its
geographical dispersion; its long trading hours (24 h a day except on weekends);
the variety of factors that affect exchange rates; the low margins of profit compared
with other markets of fixed income (but profits can be high due to very large trading
volumes); and the use of leverage.
Earlier studies have so far focused on the US stock market when investigating
the ICAPM. However, with an average daily trading volume of US $3 trillion per
day, Forex is far and away the most enormous financial market in the world,
dwarfing the trading volumes of other markets. We contribute to the existing
literature by examining for the first time the significance of an intertemporal
relation between expected return and risk on currency. We also test whether
aggregate risk aversion in the FX market changes through time.
We utilize 5-min returns on the spot exchange rates of the US dollar vis-à-vis six
major currencies (the euro, Japanese yen, British pound sterling, Swiss franc,
Australian dollar, and Canadian dollar) to construct the daily returns, realized
volatility, and range volatility estimators. Then, using the intraday data-based
daily returns as well as the GARCH, realized, and range-based volatility measures,
we test for the presence and significance of a risk-return trade-off in the FX market.
By sampling the return process more frequently, we improve the accuracy of
the conditional volatility estimate and measure the risk-return relationship at the
daily level. When we assume a constant risk-return trade-off in currency
markets, we find a positive but statistically weak relation between expected return
and risk on currency.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1109

We estimate the dependence of expected returns on the lagged realized variance

over time using rolling regressions. This also allows us to check whether our results
are driven by a particular sample period. Two different rolling regression
approaches provide strong evidence on the time variation of risk aversion
parameters for all currencies considered in the paper. However, the direction
of a relationship between expected return and risk is not clear for the entire FX
market.
The paper is organized as follows. Section 40.2 provides the descriptive statistics
for the daily and 5-min returns on exchange rates as well as the daily realized and
range-based volatility measures. Section 40.3 explains the estimation methodology.
Section 40.4 presents the empirical results on a constant risk-return trade-off in
the FX market. Section 40.5 examines the significance of time-varying risk
aversion. Section 40.6 investigates whether the covariances of individual
exchange rates with the FX market are priced in currency market. Section 40.7
concludes the paper.

40.2 Data

To test the significance of a risk-return trade-off in currency markets, we use daily

returns on the spot exchange rates of the US dollar vis-à-vis six major currencies:
the euro (EUR), Japanese yen (JPY), British pound sterling (GBP), Swiss franc
(CHF), Australian dollar (AUD), and Canadian dollar (CAD). According to the BIS
(2007) study, on the spot market the most heavily traded currency pairs were
EUR/USD (27 %), JPY/USD (13 %), GBP/USD (12 %), AUD/USD (6 %),
CHF/USD (5 %), and CAD/USD (4 %). As reported in Table 40.2, the US dollar
has been the dominant currency in both the spot and the forward and the swap
transactions. Specifically, the US currency was involved in 88.7 % of transactions,
followed by the euro (37.2 %), the Japanese yen (20.3 %), the pound sterling
(16.9 %), the Swiss franc (6.1 %), Australian dollar (5.5 %), and Canadian
dollar (4.2 %). The sum of the six major currencies (EUR, JPY, GBP, CHF,
AUD, CAD) accounts for a market share approximately equal to that of the US
dollar (90.2 %).5
The raw 5-min data on six exchange rates (EUR/USD, JPY/USD, GBP/USD,
CHF/USD, AUD/USD, and CAD/USD) are obtained from Olsen and Associates.
The full sample covers 2,282 days, from January 1, 2002 to March 31, 2008.
Following Bollerslev and Domowitz (1993) and Andersen et al. (2001b), we define
the day as starting at 21:05 pm on one night and ending at 21:00 pm the next night.
The total number of 5-min observations for each exchange rate is therefore equal to
2,282  288 ¼ 657,216. However, we are not able to use all of these observations

5
Note that volume percentages should add up to 200 %; 100 % for all the sellers and 100 % for all
the buyers. As shown in Table 40.2, the market shares of seven major currencies add up to 180 %.
The remaining 20 % of the total (200 %) market turnover has been accounted by other currencies
from Europe and from other parts of the world.
1110 T.G. Bali and K. Yilmaz

Table 40.2 Most traded Currency Symbol % daily share

currencies: currency
distribution of reported FX US dollar USD ($) 88.7 %
market turnover Euro EUR (€) 37.2 %
Japanese yen JPY (¥) 20.3 %
British pound sterling GBP (£) 16.9 %
Swiss franc CHF (Fr) 6.1 %
Australian dollar AUD ($) 5.5 %
Canadian dollar CAD ($) 4.2 %
Swedish krona SEK (kr) 2.3 %
Hong Kong dollar HKD ($) 1.9 %
Norwegian krone NOK (kr) 1.4 %
Other 15.5 %
Total 200 %

because the trading activity in FX markets slows down substantially during the
weekends and the major US official holidays. Following Andersen et al. (2001b),
along with the weekends, we removed the following holidays from our sample:
Christmas (December 24–26), New Year’s (December 31–January 2), July 4th,
Good Friday, Easter Monday, Memorial Day, Labor Day, Thanksgiving Day, and
the day after. In addition to official holidays and weekends, we removed 3 days
(March 4, 2002, April 14, 2003, and January 30, 2004) from our sample as these
days contained the longest zero or constant 5-min return sequences that might
contaminate the daily return and variance estimates. As a result, we end up with
a total of 1,556 daily observations.
Panel A of Table 40.3 presents the mean, median, maximum, minimum,
standard deviation, skewness, kurtosis, and autoregressive of order one, AR(1),
statistics for daily returns on the six exchange rates. The standard errors of the
skewness and kurtosis estimates provide evidence that the empirical distributions of
returns on exchange rates are generally symmetric and fat tailed. More specifically,
the skewness measures are statistically insignificant for all currencies, except for
the Japanese yen. The kurtosis measures are statistically significant without any
exception. The Jarque-Bera, JB ¼ n[(S2/6) + (K–3)2/24], is a formal statistic
with the Chi-square distribution for testing whether the returns are normally
distributed, where n denotes the number of observations, S is skewness, and K is
kurtosis. The JB statistics indicate significant departures from normality for
the empirical return distributions of six exchange rates. As expected, daily
returns on exchange rates are not highly persistent, as shown by the negative AR
(1) coefficients which are less than 0.10 in absolute value. Although the
economic significance of the AR(1) coefficients is low, they are statistically
significant at the 5 % or 1 % level for all currencies, except for the British pound
and Australian dollar.
The daily intertemporal relation between expected return and risk on
currency is tested using the daily realized variance of returns on exchange rates.
In very early work, the daily realized variance of asset returns is measured
 40

Table 40.3 Descriptive statistics

Panel A. Descriptive statistics for the daily returns on exchange rates
EUR JPY GBP CHF AUD CAD
Mean 0.00033 0.000076 0.00020 0.00029 0.00034 0.00025
Median 0.00024 0.00004 0.00029 0.00012 0.00064 0.00032
Maximum 0.01963 0.02515 0.01756 0.02126 0.03496 0.02563
Minimum 0.01837 0.02686 0.02048 0.02223 0.02259 0.01734
Std. dev. 0.00563 0.00580 0.00504 0.00635 0.00681 0.00520
Skewness 0.0783 0.1522** 0.0776 0.0306 0.5377** 0.1091
Kurtosis 3.6142** 4.2687** 3.4508** 3.5184** 4.5571** 3.8636**
** **
Jarque-Bera 26.05 110.36 14.74** 17.67** 232.18** 51.44**
** *
AR(1) 0.0643 0.0504 0.0117 0.0788** 0.0164 0.0642**
Panel B. Descriptive statistics for the daily realized variance measures
EUR JPY GBP CHF AUD CAD
Mean 3.47  105 4.07  105 2.77  105 4.31  105 6.00  105 3.60  105
5 5
Median 3.13  10 3.39  10 2.50  105 3.77  105 4.86  105 3.09  105
Maximum 0.000253 0.000505 0.000156 0.000367 0.000904 0.00022
Minimum 6.51  106 4.05  106 5.42  106 7.54  106 1.48  105 5.12  106
5 5
Std. dev. 1.93  10 3.12  10 1.37  105 2.48  105 4.39  105 2.30  105
** **
Skewness 2.5435 6.2585 2.2476** 3.1791** 6.8653** 2.5195**
** **
Kurtosis 19.957 71.480 13.655** 28.010** 104.839** 13.591**
** **
Jarque-Bera 20319.0 314192.0 8670.8** 43174.2** 684614.4** 8918.8**
** **
AR(1) 0.50 0.55 0.51** 0.49** 0.62** 0.64**
Panel C. Descriptive statistics for the daily range variance measures
The Intertemporal Relation Between Expected Return and Risk on Currency

EUR JPY GBP CHF AUD CAD

Mean 2.76  105 3.15  105 2.30  105 3.56  105 4.20  105 2.63  105
5 5
Median 1.91  10 2.14  10 1.61  105 2.52  105 2.68  105 1.83  105
1111

(continued)
 1112

Table 40.3 (continued)

Panel C. Descriptive statistics for the daily range variance measures
EUR JPY GBP CHF AUD CAD
Maximum 0.000204 0.000558 0.000189 0.000364 0.000777 0.000297
Minimum 1.37  106 3.58  107 1.46  106 1.55  106 2.58  106 1.66  106
Std. dev. 2.62  105 3.69  105 2.09  105 3.41  105 5.20  105 2.70  105
** **
Skewness 2.2886 5.2166 2.3097** 2.5960** 5.4669** 3.3357**
** **
Kurtosis 9.953 50.502 11.119** 13.832** 51.593** 22.051**
** **
Jarque-Bera 4492.6 153348.5 5657.0** 9354.9** 160839.7** 26417.1**
** **
AR(1) 0.09 0.25 0.19** 0.12* 0.34** 0.28**
This table reports statistics for daily returns on US dollar exchange rates (Panel A), daily realized variance (Panel B), and daily range variance (Panel C) for six
major currencies: the euro (EUR), Japanese yen (JPY), British pound sterling (GBP), Swiss franc (CHF), Australian dollar (AUD), and Canadian dollar
(CAD). Mean, median, maximum, minimum, standard deviation, skewness, kurtosis, and AR(1) statistics are reported each
pffiffiffiffiffiffiffi for p currency.
ffiffiffiffiffiffiffiffi
ffi Standard errors of
skewness and kurtosis estimates, computed under the null hypothesis that the returns are normally distributed, are 6 =n and 24 =n, respectively. Jarque-Bera,
JB ¼ n[(S2/6) + (K–3)2/24], is a formal statistic for testing whether the returns are normally distributed, where n denotes the number of observations, S is
skewness, and K is kurtosis. The JB statistic distributed as the Chi-square with 2 of freedom measures the difference of the skewness and kurtosis of the series
with those from the normal distribution. The sample period is from January 3, 2002 to March 31, 2008, yielding a total of 1,556 daily observations
*, **
denote statistical significance at the 5 % and 1 % level, respectively
T.G. Bali and K. Yilmaz
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1113

by the squared daily returns, where the asset return is defined as the natural
logarithm of the ratio of consecutive daily closing prices. A series of papers by
Andersen et al. (2001a, b, 2003, 2004) indicate that these traditional measures are
poor estimators of day-by-day movements in volatility, as the idiosyncratic
component of daily returns is large. They demonstrate that the realized volatility
measures based on intraday data provide a dramatic reduction in noise and a radical
improvement in temporal stability relative to realized volatility measures based on
daily returns. Andersen et al. (2003) show formally that the concept of realized
variance is, according to the theory of quadratic variation and under suitable
conditions, an asymptotically unbiased estimator of the integrated variance, and
thus it is a canonical and natural measure of daily return volatility.
Following the recent literature on integrated volatility, we use the high-
frequency intraday data to construct the daily realized variance of exchange
rates. To set forth notation, let Pt denote the time t (t  0) exchange rate with the
unit interval t corresponding to 1 day. The discretely observed time-series process
of logarithmic exchange rate returns with q observations per day, or a return horizon
of 1/q, is then defined by

RðqÞ, t ¼ ln Pt  ln Pt1=q (40.2)

where t ¼ 1/q, 2/q, . . . . We calculate the daily realized variance of exchange rates
using the intraday high-frequency (5-min) return data as
qX
i 1

VARrealized
t ¼ R2ðqÞ, ti=q (40.3)
i¼0

where qi,t is the number of 5-min intervals on day t and Ri,t is the logarithmic
exchange rate return in 5-min interval i on date t.
On a regular trading day, there are 288 5-min intervals. The exchange rate of the
most recent record in a given 5-min interval is taken to be the exchange rate of that
interval. A 5-min return is then constructed using the logarithmic exchange rate
difference for a 5-min interval. With 1,556 days in our full sample, we end up with
using a total of 1,556  288 ¼ 448,128 5-min return observations to calculate daily
return and variance estimates.
Panel B of Table 40.3 presents the summary statistics of the daily realized-
variances of exchange rate returns. The average daily realized variance is 6  105
for AUD/USD, 4.31  105 for CHF/USD, 4.07  105 for JPY/USD, 3.60  105
for CAD/USD, 3.47  105 for EUR/USD, and 2.77  105 for GBP/USD. These
measures correspond to an annualized volatility of 12.30 % for AUD/USD, 10.42 %
for CHF/USD, 10.13 % for JPY/USD, 9.52 % for CAD/USD, 9.35 % for
EUR/USD, and 8.35 % for GBP/USD. A notable point in Panel B is that the daily
realized variances are highly persistent, as shown by the AR(1) coefficients which
are in the range of 0.49–0.64. Consistent with Andersen et al. (2001a, b), the
distributions of realized variances are skewed to the right and have much thicker
tails than the corresponding normal distribution.
1114 T.G. Bali and K. Yilmaz

Market microstructure noises in transaction data such as the bid-ask bounce may
influence our risk measures based on the realized volatility and GARCH volatility
forecasts, even though the data we use contain very liquid financial time series
and thus are least subject to biases created by market microstructure effects.
An alternative volatility measure that utilizes information contained in the high-
frequency intraday data is Parkinson’s (1980) range-based estimator of the daily
integrated variance:


2
VARrange
t ¼ 0:361 ln Pmax
t  ln Pmin
t (40.4)

where Pmax
t and Pmin
t are the maximum and minimum values of the exchange rate on
day t. Alizadeh et al. (2002) and Brandt and Diebold (2006) show that the
range-based volatility estimator is highly efficient, approximately Gaussian, and
robust to certain types of microstructure noise such as bid-ask bounce. In addition,
range data are available for many assets over a long sample period.
Panel C of Table 40.3 presents the summary statistics of the daily range variances
of exchange rate returns. The average daily range variance is 4.20  105 for
AUD/USD, 3.56  105 for CHF/USD, 3.15  105 for JPY/USD, 2.63  105 for
CAD/USD, 2.76  105 for EUR/USD, and 2.0  105 for GBP/USD. These
measures correspond to an annualized volatility of 10.29 % for AUD/USD, 9.47 %
for CHF/USD, 8.91 % for JPY/USD, 8.14 % for CAD/USD, 8.34 % for EUR/USD,
and 7.61 % for GBP/USD. These results indicate that the daily realized volatility
estimates are somewhat higher than the daily range volatilities. Another notable
point in Panel C is that the daily range variances are less persistent than the daily
realized variances. Specifically, the AR(1) coefficients are in the range of 0.09–0.34
for the daily range variances. Similar to our findings for the daily realized variances,
the distributions of range variances are skewed to the right and have much thicker
tails than the corresponding normal distribution.

40.3 Estimation Methodology

The following GARCH-in-mean process is used with the conditional normal density
to model the intertemporal relation between expected return and risk on currency:

Rtþ1  a þ b  s2tþ1jt þ etþ1 (40.5)

etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1Þ, Eðetþ1 Þ ¼ 0 (40.6)


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t , (40.7)
"
#
  1 1 Rtþ1  mtþ1jt 2
f Rtþ1 ; mtþ1jt ; stþ1jt ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp  (40.8)
2ps2tþ1jt 2 stþ1jt
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1115

where Rt+1 is the daily return on exchange rates for period t+1, mt+1|t  a + b  st+1|t
2
is
the conditional mean for period t + 1 based on the information set up to time t,
et+1 ¼ zt+1  st+1|t is the error term with E(et+1) ¼ 0, st+1|t, is the conditional standard
deviation of daily returns on currency, and zt+1  N(0,1) is a random variable drawn
from the standard normal density and can be viewed as information shocks in the
FX market. st+1|t 2
is the conditional variance of daily returns based on the
information set up to time t denoted by Ot. The conditional variance, st+1|t 2
, follows
a GARCH(1,1) process as defined by Bollerslev (1986) to be a function of the last
period’s unexpected news (or information shocks), zt, and the last period’s variance,
st2. f(Rt+1;mt+1|t,st+1|t) is the conditional normal density function of Rt+1 with the
conditional mean of mt+1|t and conditional variance of st+1|t 2
. Our focus is
to examine the magnitude and statistical significance of the risk aversion parameter
b in Eq. 40.5.
Campbell (1987) and Scruggs (1998) point out that the approximate relationship
in Eq. 40.1 may be misspecified if the hedging term in ICAPM is important. To
make sure that our results from estimating Eq. 40.5 are not due to model misspeci-
fication, we added to the specifications a set of control variables that have been used
in the literature to capture the state variables that determine changes in the
investment opportunity set. Several studies find that macroeconomic variables
associated with business cycle fluctuations can predict the stock market.6 The
commonly chosen variables include Treasury bill rates, federal funds rate, default
spread, term spread, and dividend-price ratios. We study how variations in the fed
funds rate, default spread, and term spread affect the intertemporal risk-return
relation.7 Earlier studies also control for the lagged return in the conditional mean
specification.
We obtain daily data on the federal funds rate, 3-month Treasury bill, 10-year
Treasury bond yields, and BAA-rated and AAA-rated corporate bond yields from
the H.15 database of the Federal Reserve Board. The federal funds (FED) rate is the
interest rate at which a depository institution lends immediately available funds
(balances at the Federal Reserve) to another depository institution overnight. It is
a closely watched barometer of the tightness of credit market conditions in the
banking system and the stance of monetary policy. In addition to the fed funds rate,
we use the term and default spreads as control variables. The term spread (TERM) is
calculated as the difference between the yields on the 10-year Treasury bond and
the 3-month Treasury bill. The default spread (DEF) is computed as the difference
between the yields on the BAA-rated and AAA-rated corporate bonds. We test the
significance of the risk aversion parameter, b, after controlling for macroeconomic
variables and lagged return:

6
See Keim and Stambaugh (1986), Chen et al. (1986), Campbell and Shiller (1988), Fama and
French (1988, 1989), Campbell (1987, 1991), Ghysels et al (2005), and Guo and Whitelaw (2006).
7
We could not include the aggregate dividend yield (or the dividend-price ratio) because the data
on dividends are available only at the monthly frequency while our empirical analyses are based on
the daily data.
1116 T.G. Bali and K. Yilmaz

Rtþ1  a þ b  s2tþ1jt þ l1  FEDt þ l2  DEFt þ l3  TERMt þ l4  Rt þ etþ1

Rtþ1  a þ b  s2tþ1jt þ l1  FEDt þ l2  DEF þ l3  TERMt þ l4  Rt þ etþ1 (40.9)

etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1Þ, Eðetþ1 Þ ¼ 0 (40.10)


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t (40.11)

Earlier studies that investigate the daily risk-return trade-off generally rely on
the GARCH-in-mean methodology. In risk-return regressions, it is not common to
use the realized variance measures obtained from the intraday data. In this paper,
we first generate the daily realized variance based on the 5-min returns on exchange
rates and then estimate the following risk-return regression:
 
Rtþ1 ¼ a þ b  Et VARrealized
tþ1 þ etþ1 (40.12)

where Rt+1 is the 1-day ahead return on exchange rate and Et[VARrealized
t+1 ] is proxied
by the lagged realized variance measure, i.e., Et[VARrealized
t+1 ]¼VARrealized
t defined in
Eq. 40.3. As reported in Panel B of Table 40.3, VARrealized t has significant
persistence measured by the first-order serial correlation that makes VARrealizedt
a reasonable proxy for the 1-day ahead expected realized variance. The slope
coefficient b in Eq. 40.12, according to Merton’s (1973) ICAPM, is the relative
risk aversion coefficient which is expected to be positive and statistically
significant.
To control for macroeconomic variables and lagged returns that may
potentially affect the fluctuations in the FX market, we estimate the risk aversion
coefficient, b, after controlling for the federal funds rate, term spread, default
spread, and lagged return:

Rtþ1  a þ b  VARrealized
t þ l1  FEDt þ l2  DEFt þ l3  TERMt þ l4  Rt þ etþ1
(40.13)

and test the statistical significance of b.

In addition to the GARCH-in-mean and realized volatility models, we use the
range-based volatility estimator with and without control variables to test the
significance of risk aversion b:

Rtþ1 ¼ a þ b  VARrange
t þ etþ1 (40.14)

Rtþ1  a þ b  VARrange
t þ l1  FEDt þ l2  DEFt þ l3  TERMt þ l4  Rt þ etþ1
(40.15)

where VARrange
t is the Parkinson’s (1980) range-based estimator of the daily inte-
grated variance defined in Eq. 40.4.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1117

The uncovered interest rate parity indicates that the appreciation

(or depreciation) rate of a currency is related to the interest rate differential of
two countries.8 Therefore, the hedging demand component of the ICAPM is
proxied by the short-term interest rates of the two countries. Specifically, the
intertemporal relation is tested based on the GARCH-in-mean, realized, and
range volatility estimators along with the London Interbank Offer Rate (LIBOR)
for the US and the corresponding foreign country:

foreign
Rtþ1  a þ b  s2tþ1jt þ l1  LIBORUS
t þ l2  LIBORt þ l3  Rt þ etþ1 (40.16)

foreign
Rtþ1  a þ b  VARrealized
t þ l1  LIBORUS
t þ l2  LIBORt þ l3  Rt þ etþ1 (40.17)

Rtþ1  a þ b  VARrange
t þ l1  LIBORUS foreign
t þ l2  LIBORt þ l3  Rt þ etþ1 (40.18)

where LIBORUS t and LIBORforeign

t are the LIBOR rates for the US and the
corresponding foreign country. To control for a potential first-order serial correla-
tion in daily returns on exchange rates, we include the lagged return (Rt) to the
conditional mean specifications.

40.4 Empirical Results

Table 40.4 presents the maximum likelihood parameter estimates and the t-statistics
in parentheses for the GARCH-in-mean model. The risk aversion parameter (b) is
estimated to be positive for all currencies considered in the paper, but the parameter
estimates are not statistically significant, except for the British pound and
the Canadian dollar. Specifically, b is estimated to be 5.18 for the euro, 4.42 for
the Japanese yen, 29.07 for the British pound, 0.87 for the Swiss franc, 11.04 for
the Australian dollar, and 22.40 for the Canadian dollar. Based on the Bollerslev-
Wooldridge (1992) heteroscedasticity consistent covariance t-statistics reported in
Table 40.4, the risk aversion coefficient has a t-statistic of 1.83 for the Canadian
dollar and t-statistic of 1.77 for the British pound. Although we do not have
a strong statistical significance, we can interpret this finding as a positive risk-
return trade-off in the US/Canadian dollar and US dollar/lb exchange rate markets.
Overall, these results indicate a positive but statistically weak relation between
expected return and risk on currency.

8
Assuming that the interest rate is 5 % per annum in the US and 2 % per annum in Japan, the
uncovered interest rate parity predicts that the US dollar would depreciate against the Japanese yen
by 3 %.
1118 T.G. Bali and K. Yilmaz

Table 40.4 Daily risk-return trade-off in foreign exchange markets based on the GARCH-in-
mean model
Parameters EUR JPY GBP CHF AUD CAD
a 0.0005 0.0002 0.0009 0.0003 0.0008 0.0008
(1.46) (0.30) (2.25) (0.72) (2.26) (2.70)
b 5.1772 4.4206 29.065 0.8693 11.042 22.399
(0.47) (0.28) (1.77) (0.08) (1.28) (1.83)
g0 1.09  107 1.69  106 3.37  107 2.60  107 6.97  107 2.45  107
(0.65) (1.76) (2.32) (1.00) (0.92) (2.21)
g1 0.0301 0.0587 0.0430 0.0331 0.0541 0.0420
(4.25) (4.57) (4.51) (4.24) (3.81) (4.82)
g2 0.9672 0.8922 0.9443 0.9617 0.9318 0.9502
(116.10) (35.85) (73.13) (107.35) (50.04) (93.08)
The following GARCH-in-mean process is used with conditional normal density to model the
intertemporal relation between expected return and risk on currency

Rtþ1  a þ b  s2tþ1jt þ etþ1

etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1Þ, Eðetþ1 Þ ¼ 0


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t

where Rt+1 is the daily return on exchange rates for period t+1, mt+1|t  a+b  st+1|t
2
is the conditional
mean for period t+1 based on the information set up to time t denoted by Ot, et+1 ¼ zt+1  st+1|t is the
error term with E(et+1) ¼ 0, st+1|t is the conditional standard deviation of daily returns on currency,
and zt+1  N(0,1) is a random variable drawn from the standard normal density and can be viewed
as information shocks in FX markets. s2t+1|t is the conditional variance of daily returns based on the
information set up to time t denoted by Ot. The conditional variance, st+1|t
2
, follows a GARCH(1,1)
process defined as a function of the last period’s unexpected news (or information shocks), zt, and
the last period’s variance, st2. The table presents the maximum likelihood parameter estimates and
the t-statistics in parentheses

Another notable point in Table 40.4 is the significance of volatility clustering. For
all currencies, the conditional volatility parameters (g1, g2) are positive, between zero
and one, and highly significant. The results indicate the presence of rather extreme
conditionally heteroskedastic volatility effects in the exchange rate process because
the GARCH parameters, g1 and g2, are found to be not only highly significant, but
also the sum (g1 + g2) is close to one for all exchange rates considered in the paper.
This implies the existence of substantial volatility persistence in the FX market.
Table 40.5 reports the daily risk aversion parameter estimates and their statistical
significance for each currency after controlling for macroeconomic variables and
lagged return. The risk-return coefficient estimates are similar to our earlier findings
in Table 40.4. The relationship between expected return and conditional risk is
positive but statistically weak for all exchange rates, except for the British pound
and the Canadian dollar where we have a risk aversion parameter of 36.51 with
t-stat. ¼1.89 for the British pound and 27.86 with t-stat. ¼ 2.15 for the Canadian
dollar. These results indicate that controlling for the hedging demand component of
the ICAPM does not alter our findings.
 40

Table 40.5 Daily risk-return trade-off in foreign exchange markets based on the GARCH-in-mean model with control variables
Parameters EUR JPY GBP CHF AUD CAD
a 0.0003 0.0024 0.0012 0.0006 0.0007 0.0013
(0.13) (1.54) (0.73) (0.29) (0.39) (0.95)
b 6.9731 15.986 36.514 17.913 16.478 29.329
(0.68) (0.96) (1.89) (1.17) (1.56) (2.15)
l1 2.19  104 3.76  104 1.00  104 2.43  104 3.11  104 4.64  105
(0.66) (1.13) (0.35) (0.69) (0.90) (0.18)
l2 5.58  105 7.24  104 8.06  104 1.00  104 8.66  105 7.33  104
(0.07) (0.96) (1.13) (0.09) (0.10) (1.03)
l3 5.33  104 6.79  104 2.71  104 5.82  104 5.64  104 1.62  104
(1.45) (1.65) (0.82) (1.43) (1.27) (0.50)
l4 0.055 0.042 0.008 0.062 0.010 0.038
(2.04) (1.48) (0.30) (2.48) (0.35) (1.45)
g0 1.14  107 1.64  106 3.13  107 2.37  107 7.32  107 2.42  107
(0.99) (2.94) (2.30) (1.83) (2.63) (2.15)
g1 0.0311 0.0581 0.0408 0.0331 0.0566 0.0417
(4.56) (4.41) (4.52) (3.81) (4.03) (4.80)
g2 0.9660 0.8938 0.9475 0.9622 0.9285 0.9506
(97.28) (33.74) (77.76) (88.65) (53.11) (92.60)
The following GARCH-in-mean process is used with control variables to estimate the intertemporal relation between expected return and risk on currency

Rtþ1  a þ b  s2tþ1jt þ l1  FEDt þ l2  DEFt þ l3  TERMt þ l4  Rt þ etþ1

etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1Þ, Eðetþ1 Þ ¼ 0


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t
The Intertemporal Relation Between Expected Return and Risk on Currency

where FEDt, DEFt, and TERMt are macroeconomic variables that proxy for the hedging demand component of ICAPM, and Rt is the lagged daily return. The
federal funds rate (FED) is the interest rate at which a depository institution lends immediately available funds (balances at the Federal Reserve) to another
depository institution overnight. The term spread (TERM) is calculated as the difference between the yields on the 10-year Treasury bond and the 3-month
Treasury bill. The default spread (DEF) is computed as the difference between the yields on the BAA-rated and AAA-rated corporate bonds. The table
1119

presents the maximum likelihood parameter estimates and the t-statistics in parentheses
1120 T.G. Bali and K. Yilmaz

Table 40.5 shows that the slope coefficient (l4) on the lagged return is negative
for all currencies, but it is statistically significant only for the euro (with
t-stat. ¼ –2.04) and the Swiss franc (with t-stat. ¼ –2.48).9 We find a negative
but insignificant first-order serial correlation for the Japanese yen, British pound,
Australian dollar, and Canadian dollar.
Table 40.6 presents the parameter estimates and their Newey and West (1987)
adjusted t-statistics from the risk-return regressions with realized daily variance.
Panel A reports results without the control variables and tests whether the realized
variance obtained from the sum of squared 5-min returns can predict 1-day ahead
returns on exchange rates. The risk aversion parameter (b) is estimated to be
positive for five out of six currencies considered in the paper, but only two of
these parameter estimates are statistically significant at the 10 % level. Specifically,
b is estimated to be 11.39 for the euro, 7.91 for the Japanese yen, 7.91 for the British
pound, 13.78 for the Swiss franc, –1.09 for the Australian dollar, and 11.89 for the
Canadian dollar. Based on the Newey-West (1987) t-statistics reported in
Table 40.6, the Swiss franc has a risk aversion parameter of 13.78 (t-stat. ¼ 2.21)
and the euro has a risk aversion coefficient of 11.39 (t-stat. ¼ 1.77). These results
indicate that the daily realized variance measures obtained from intraday data
positively predict future returns on exchange rates, but the link between risk and
return is generally statistically insignificant.
Panel B of Table 40.6 presents the risk aversion coefficient estimates after
controlling for the federal funds rate, term spread, default spread, and lagged return.
Similar to our findings in Panel A, the risk aversion parameter is estimated to be
18.76 with t-stat. ¼ 2.42 for the euro, 8.77 with t-stat. ¼ 1.80 for the Japanese yen,
and 18.89 with t-stat. ¼ 2.70 for the Swiss franc, indicating a positive and
significant link between the realized variance and the 1-day ahead returns on the
US dollar/euro, US dollar/yen, and US dollar/Swiss franc exchange rates. There
is also a positive but statistically weak relation for the British pound and the
Canadian dollar.
Table 40.7 reports the parameter estimates and their Newey-West t-statistics
from the risk-return regressions with the daily range variance of Parkinson (1980).
As shown in both panels, with and without control variables, the risk aversion
parameter (b) is estimated to be positive but statistically insignificant, except for
the marginal significance of b for the Canadian dollar in Panel B. These results
provide evidence that the daily range volatility obtained from the intraday data
positively predict future returns on exchange rates, but there is no significant
relation between risk and return on currency.
The estimates in Tables 40.6 and 40.7 present a negative and significant
autocorrelation for the euro, Japanese yen, Swiss franc, and Canadian dollar.
The first-order autocorrelation coefficient is negative but statistically insignificant
for the British pound and the Australian dollar.

9
Jegadeesh (1990), Lehmann (1990), and Lo and MacKinlay (1990) provide evidence for the
significance of short-term reversal (or negative autocorrelation) in short-term stock returns.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1121

Table 40.6 Daily risk-return trade-off in foreign exchange markets based on the realized
variance
Panel A. Daily risk-return trade-off without control variables
Parameters EUR JPY GBP CHF AUD CAD
a 0.00073 0.00040 0.00042 0.00088 0.00027 0.00068
(2.84) (1.59) (1.77) (2.90) (0.77) (2.25)
b 11.393 7.9064 7.9145 13.777 1.0895 11.885
(1.77) (1.61) (1.07) (2.21) (0.21) (1.38)
Panel B. Daily risk-return trade-off with control variables
Parameters EUR JPY GBP CHF AUD CAD
a 0.00124 0.00257 0.00080 0.00169 0.00030 0.00074
(0.82) (1.58) (0.57) (0.96) (0.17) (0.49)
b 18.759 8.7656 10.429 18.886 0.0858 13.608
(2.42) (1.80) (1.30) (2.70) (0.01) (1.55)
l1 0.00028 0.00035 0.00025 0.00033 4.4  106 0.00007
(0.99) (1.06) (0.90) (1.03) (0.01) (0.25)
l2 0.00051 0.00112 0.00013 0.00076 0.00059 0.00041
(0.72) (1.33) (0.18) (0.91) (0.63) (0.60)
l3 0.00054 0.00054 0.00041 0.00065 0.00004 0.00012
(1.54) (1.36) (1.19) (1.67) (0.07) (0.37)
l4 0.068 0.048 0.013 0.077 0.017 0.071
(2.81) (1.81) (0.49) (3.00) (0.56) (2.88)
The following regression is estimated with and without control variables to test the significance of
the intertemporal relation between expected return and risk on currency

Rt+ 1  a + b  VARtrealized + l1  FEDt + l2  DEFt + l3  TERMt + l4  Rt + et+1

where VARtrealized is the daily realized variance computed as the sum of squared 5-min returns on
exchange rates. The table presents the parameter estimates and their Newey and West (1987)
t-statistics in parentheses

An interesting observation in Tables 40.5, 40.6, and 40.7 is that the slope
coefficients (l1, l2, l3) on the lagged macroeconomic variables are found to be
statistically insignificant, except for some marginal significance for the term spread
in the regressions with the Swiss franc. Although one would think that unexpected
news in macroeconomic variables could be viewed as risks that would be rewarded in
the FX market, we find that the changes in federal funds rate and term and default
spreads do not affect time-series variation in daily exchange rate returns. Our inter-
pretation is that it would be very difficult for macroeconomic variables to explain daily
variations in exchange rates. If we examined the risk-return trade-off at lower fre-
quency (such as monthly or quarterly frequency), we might observe significant impact
of macroeconomics variables on monthly or quarterly variations in exchange rates.
Panel A of Table 40.8 presents the maximum likelihood parameter estimates and
the t-statistics in parentheses for the GARCH-in-mean model with LIBOR rates for
the US and the corresponding foreign country. The risk aversion parameter (b) is
estimated to be positive for all currencies, but the parameter estimates are
1122 T.G. Bali and K. Yilmaz

Table 40.7 Daily risk-return trade-off in foreign exchange markets based on the range volatility
Panel A. Daily risk-return trade-off without control variables
Parameters EUR JPY GBP CHF AUD CAD
a 0.00051 0.00022 0.00038 0.00053 0.00053 0.00048
(2.54) (1.11) (2.15) (2.28) (2.15) (2.67)
b 6.6224 4.6907 7.8004 6.8730 4.4630 8.4727
(1.32) (1.15) (1.31) (1.56) (1.05) (1.61)
Panel B. Daily risk-return trade-off with control variables
Parameters EUR JPY GBP CHF AUD CAD
a 0.00179 0.00250 0.00088 0.00238 0.00042 0.00053
(1.22) (1.57) (0.65) (1.40) (0.23) (0.37)
b 8.6053 5.0792 8.3969 7.3411 5.3794 9.2256
(1.58) (1.24) (1.36) (1.52) (1.14) (1.69)
l1 0.00036 0.00030 0.00026 0.00045 0.00010 0.00005
(1.26) (0.95) (0.93) (1.41) (0.25) (0.16)
l2 0.00041 0.00106 0.00015 0.00049 0.00038 0.00033
(0.57) (1.29) (0.22) (0.59) (0.42) (0.48)
l3 0.00056 0.00049 0.00040 0.00070 0.00019 0.00009
(1.59) (1.25) (1.18) (1.81) (0.40) (0.28)
l4 0.067 0.051 0.012 0.078 0.026 0.067
(2.77) (1.89) (0.45) (3.13) (0.87) (2.71)
The following regression is estimated with and without control variables to test the significance of
the intertemporal relation between expected return and risk on currency

Rt+1  a + b  VARrange
t + l1  FEDt + l2  DEFt + l3  TERMt + l4  Rt + et+1

where VARranget ¼ 0.361[ln(Pmax

t )  ln(Pmin 2
t )] is Parkinson’s (1980) range-based estimator of the
daily integrated variance. The table presents the parameter estimates and their Newey-West (1987)
t-statistics in parentheses

statistically significant only for the British pound, Australian dollar, and
Canadian dollar. Specifically, b is estimated to be 30.87 for the British pound,
17.75 for the Australian dollar, and 32.90 for the Canadian dollar. Based on
the Bollerslev-Wooldridge heteroscedasticity consistent covariance t-statistics
reported in Table 40.8, the risk aversion coefficient has a t-statistic of 1.82 for the
British pound, 1.84 for the Australian dollar, and 2.36 for the Canadian dollar.
Although we do not have a strong statistical significance, we can interpret this finding
as a positive risk-return trade-off in the US dollar/British pound, US/Australian dollar,
and US/Canadian dollar markets. Overall, the results indicate a positive but statisti-
cally weak relation between expected return and risk on currency.
Another point worth mentioning in Panel A is that the slope coefficients on
the US LIBOR rate are estimated to be positive and statistically significant at the
5 % level for the euro, Japanese yen, and Swiss franc and significant at the 10 %
level for the Canadian dollar. As expected, the slope coefficients on the
LIBOR rates of the corresponding foreign country turn out to be negative but
statistically insignificant.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1123

Table 40.8 Daily risk-return trade-off in foreign exchange markets with LIBOR interest rates
Panel A. GARCH-in-mean
Parameters EUR JPY GBP CHF AUD CAD
a 0.0004 0.0011 0.0019 0.0004 0.0004 0.0006
(0.62) (1.82) (1.97) (0.31) (0.27) (0.58)
b 6.7962 13.422 30.871 18.036 17.753 32.898
(0.66) (0.84) (1.82) (1.33) (1.84) (2.36)
l1 2.16  104 2.86  104 1.41  3.30  104 2.57  104 1.95  104
105
(1.97) (2.53) (0.12) (2.19) (1.47) (1.66)
l2 2.61  104 8.94  1.83  3.44  3.35  4.81 
104 104 104 104 104
(1.10) (1.32) (0.78) (1.13) (0.78) (1.73)
l3 0.055 0.042 0.008 0.062 0.010 0.039
(2.22) (1.62) (0.31) (2.46) (0.36) (1.48)
g0 1.15  107 1.64  106 3.11  2.64  107 7.35  107 2.44  107
107
(1.22) (1.30) (2.25) (0.73) (0.99) (2.13)
g1 0.0300 0.0575 0.0408 0.0338 0.0563 0.0424
(4.44) (4.49) (4.51) (3.72) (3.93) (4.87)
g2 0.9672 0.8946 0.9476 0.9608 0.9287 0.9498
(96.57) (32.83) (77.50) (79.92) (50.57) (92.31)
Panel B. Realized variance
Parameters EUR JPY GBP CHF AUD CAD
a 0.00074 0.00108 0.00150 0.00171 0.00118 0.00024
(1.08) (2.52) (1.61) (2.86) (0.77) (0.37)
b 17.558 10.272 10.499 18.730 0.708 15.199
(2.24) (1.95) (1.31) (2.63) (0.11) (1.69)
l1 2.5  104 3.1  104 3.14  4.5  104 4.6  106 1.4  104
106
(2.39) (2.47) (0.03) (2.96) (0.026) (1.00)
l2 3.3  1.6  2.2  104 6.5  1.6  104 3.0 
104 103 104 104
(1.81) (2.01) (0.90) (2.19) (0.42) (1.03)
l3 0.069 0.048 0.013 0.078 0.016 0.073
(2.85) (1.79) (0.51) (3.10) (0.53) (2.91)
Panel C. Range variance
Parameters EUR JPY GBP CHF AUD CAD
a 0.00015 0.00079 0.00136 0.00090 0.00079 0.00018
(0.26) (2.13) (1.49) (1.86) (0.52) (0.28)
b 8.181 5.558 8.416 7.400 5.249 9.761
(1.49) (1.29) (1.37) (1.51) (1.03) (1.77)
l1 1.9  104 2.8  104 1.1  3.5  104 7.26  105 9.5  105
105
(1.92) (2.32) (0.09) (2.51) (0.46) (0.68)
(continued)
1124 T.G. Bali and K. Yilmaz

Table 40.8 (continued)

Panel C. Range variance
Parameters EUR JPY GBP CHF AUD CAD
l2 3.4  1.3  2.1  104 6.3  1.08  106 1.9 
104 103 104 104
(1.86) (1.84) (0.88) (2.20) (0.003) (0.63)
l3 0.068 0.051 0.012 0.080 0.025 0.067
(2.82) (1.87) (0.47) (3.22) (0.85) (2.71)
The following regressions with GARCH-in-mean, realized variance, and range variance are
estimated to test the significance of the intertemporal relation between expected return and risk
on currency

foreign
Rtþ1  a þ b  s2tþ1jt þ l1  LIBORUS
t þ l2  LIBORt þ l3  Rt þ etþ1 ,
foreign
Rtþ1  a þ b  VARrealized
t þ l1  LIBORUS
t þ l2  LIBORt þ l3  Rt þ etþ1 ,
Rtþ1  a þ b  VARrange
t þ l1  LIBORUS foreign
t þ l2  LIBORt þ l3  Rt þ etþ1 ,

where LIBORUS
t and LIBORforeign
t are the LIBOR rates for the US and the corresponding foreign
country

Panel B of Table 40.8 reports the parameter estimates and their Newey-West
adjusted t-statistics from the risk-return regressions with daily realized variance
after controlling for the LIBOR rates and the lagged return. The results indicate
a positive and significant link between the realized variance and the 1-day ahead
returns on the euro, Japanese yen, Swiss franc, and Canadian dollar. There is also
a positive but statistically weak relation for the British pound.
Panel C of Table 40.8 shows the parameter estimates and their Newey-West
t-statistics from the risk-return regressions with the daily range variance of
Parkinson (1980). With LIBOR rates and the lagged return, the risk aversion
parameter (b) is estimated to be positive for all currencies but statistically signif-
icant only for the Canadian dollar. Overall, the results provide evidence that after
controlling for the interest rate differential of two countries, there is a positive but
statistically weak relation between risk and return on currency.
Similar to our earlier findings from the GARCH-in-mean model, Panels B and
C of Table 40.8 show that the slope coefficients on the US LIBOR rate are generally
positive, whereas the slopes on the corresponding foreign LIBOR rates are negative
with a few exceptions.
Many studies fail to identify a statistically significant intertemporal relation
between risk and return of the stock market portfolios. French et al. (1987) find
that the coefficient estimate is not significantly different from zero when they use
past daily returns to estimate the monthly conditional variance.10 Chan et al. (1992)
employ a bivariate GARCH-in-mean model to estimate the conditional variance,

10
When testing monthly risk-return trade-off, French et al. (1987) use the monthly realized
variance obtained from the sum of squared daily returns within a month.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1125

and they also fail to obtain a significant coefficient estimate for the United States.
Campbell and Hentchel (1992) use the quadratic GARCH (QGARCH) model of
Sentana (1995) to determine the existence of a risk-return trade-off within an
asymmetric GARCH-in-mean framework. Their estimate is positive for one
sample period and negative for another sample period, but neither is statistically
significant. Glosten et al. (1993) use monthly data and find a negative but statisti-
cally insignificant relation from two asymmetric GARCH-in-mean models. Based
on semi-nonparametric density estimation and Monte Carlo integration, Harrison
and Zhang (1999) find a significantly positive risk and return relation at a 1-year
horizon, but they do not find a significant relation at shorter holding periods such as
1 month. Using a sample of monthly returns and implied and realized volatilities for
the S and P 500 index, Bollerslev and Zhou (2006) find an insignificant
intertemporal relation between expected return and realized volatility, whereas
the relation between return and implied volatility turns out to be significantly
positive.
Several studies find that the intertemporal relation between risk and return is
negative (e.g., Campbell 1987; Breen et al. 1989; Turner et al. 1989; Nelson 1991;
Glosten et al. 1993; Harvey 2001; Brandt and Kang 2004). Some studies do provide
evidence supporting a positive and significant relation between expected return and
risk on stock market portfolios (e.g., Bollerslev et al. 1988; Scruggs 1998;
Ghysels et al. 2005; Bali and Peng 2006; Guo and Whitelaw 2006; Lundblad
2007; Bali 2008).
Merton’s (1973) ICAPM provides a theoretical model that gives a positive
equilibrium relation between the conditional first and second moments of excess
returns on the aggregate market portfolio. However, Abel (1988), Backus and
Gregory (1993), and Gennotte and Marsh (1993) develop models in which
a negative relation between expected return and volatility is consistent with equi-
librium. As summarized above, there has been a lively debate on the existence and
direction of a risk-return trade-off, and empirical studies are still not in agreement
for the stock market portfolios. The empirical results presented in Tables 40.4, 40.5,
40.6, 40.7, and 40.8 indicate that the intertemporal relation between expected return
and risk on currency is positive but in most cases statistically insignificant. Hence,
our findings from the FX market are in line with some of the earlier studies that
investigated the significance of a risk-return trade-off for the stock market.

40.5 Time-Varying Risk Aversion in the Foreign Exchange

Market

Chou et al. (1992), Harvey (2001), and Lettau and Ludvigson (2010) suggest that
the risk-return relation for the stock market may be time varying. In the existing
literature, there is no study investigating the presence and significance of time-
varying risk aversion in the FX market. We have so far assumed a constant
risk-return trade-off in currency markets and found a positive but statistically
insignificant relation between expected return and risk on exchange rates.
1126 T.G. Bali and K. Yilmaz

We now estimate the dependence of expected returns on the lagged realized

variance over time using rolling regressions. This also allows us to check whether
our results are driven by a particular sample period. We estimate the risk-return
relation specified in Eqs. 40.12 and 40.13 for the six exchange rates with rolling
samples. We used two different rolling regression approaches. The first one uses
a fixed rolling window of 250 business days (i.e., approximately 1 year), whereas
the second one starts with the in-sample period of 250 business days and then
extends the sample by adding each daily observation to the estimation while
keeping the start date constant.
Figure 40.1 plots the estimated relative risk aversion parameters (b) and their
statistical significance over time from the fixed rolling window of 250 days.11
Specifically, the first 250 daily return observations of exchange rates and their
realized variances (from 1/3/2002 to 1/8/2003) are used for estimation of the
relative risk aversion parameter for 1/8/2003. The sample is then rolled forward
by removing the first observation of the sample and adding one to the end, and
another 1-day ahead risk-return relationship is measured. This recursive estimation
procedure is repeated until the sample is exhausted on March 31, 2008. The
estimated relative risk aversion parameter over the fixed rolling sample period
represents the average degree of risk aversion over that sample period. Computation
of the relative risk aversion parameters using a rolling window of data allows us to
observe the time variation in an investors’ average risk aversion.
A common observation in Fig. 40.1 is that there is a strong time-series variation
in the risk aversion estimates for all currencies considered in the paper. The first
panel in Fig. 40.1 indicates that in the US dollar/euro FX market, the aggregate risk
aversion is generally positive with some exceptions in the second half of 2006 and
from May to August 2007. For the out-of-sample period of January 2003 to March
2008, only 208 out of 1,306 daily risk aversion estimates are negative. Based on the
Newey-West adjusted t-statistics, all of these negative risk aversion estimates are
statistically insignificant. 143 (291) out of 1,098 positive risk aversion estimates
turn out to be statistically significant at least at the 5 % level (10 % level). These
results indicate a positive but statistically insignificant time-varying risk aversion in
the US dollar/euro market.
The second panel in Fig. 40.1 displays that in the Japanese yen market, the
aggregate risk aversion is generally positive, but there are quite a lot of days in
which we observe a negative relation between expected return and risk in the
US dollar/yen market. 431 out of 1,306 daily risk aversion estimates are negative,
but about one third is statistically significant at the 10 % level. 185 (314) out of
875 positive risk aversion estimates turn out to be statistically significant at least at
the 5 % level (10 % level). These results indicate that there is a positive but not
strong time-varying risk aversion in the US dollar/yen exchange rate market.

11
Since the time-varying risk aversion coefficients from estimating Eqs. 40.12 and 40.13 with and
without control variables turn out to be very similar, we only report results from the full
specification of Eq. 40.13. Time-varying risk aversion estimates obtained from the parsimonious
specification of Eq. 40.12 are available from the authors upon request.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1127

Third panel in Fig. 40.1 shows that in the US dollar/lb sterling market, the risk
aversion is generally positive, but there is a long period of time in which we
observe a negative relation between expected return and risk in the US dollar/lb
market. Specifically, 872 out of 1,306 daily risk aversion estimates are positive, but
only 5 out of 872 are marginally significant. Similarly, only 46 out of 434 negative
risk aversion estimates turn out to be statistically significant at the 10 % level.
These results provide evidence that although there is a significant time variation in
the aggregate risk aversion, it is not clear whether the currency trade
generates a larger or smaller risk premium at times when the US dollar/lb FX
market is riskier.
The fourth panel in Fig. 40.1 indicates that in the Swiss franc market, the risk
aversion is estimated to be positive throughout the sample period (2002–2008),
except for a few months in 2006. Only 71 out of 1,306 daily risk aversion
estimates are negative, but none of them is statistically significant. 353 (467) out
of 1,235 positive risk aversion estimates turn out to be statistically significant at
least at the 5 % level (10 % level). These results indicate a positive and relatively
strong time-varying risk aversion, implying that the currency trade generates
a larger risk premium at times when the US dollar/Swiss franc trade becomes
riskier.
The fifth panel in Fig. 40.1 indicates that in the Australian dollar market
736 out of 1,306 daily risk aversion estimates are positive, but none of them is
statistically significant. Only 65 out of 570 negative risk aversion estimates
turn out to be marginally significant at the 10 % level. The figure indicates
a strong time-varying risk aversion, but there is no significantly positive or
negative relation between risk and return in the US/Australian dollar exchange
rate market.
The last panel in Fig. 40.1 demonstrates that in the US/Canadian dollar market,
for slightly more than half of the sample, the risk aversion is estimated to be
positive and slightly less than half of the sample it turns out to be negative.
However, based on the t-statistics of these risk aversion estimates, there is no
evidence for a significantly positive or negative link between expected return and
risk on currency. Only 35 out of 757 positive risk aversion coefficients and
only 46 out of 549 negative risk aversion parameters are found to be significant
at the 10 % level. Although there is a significant time-series variation in the
aggregate risk aversion, trading in the US/Canadian dollar FX market does not
provide clear evidence for a larger or smaller risk premium at times when the
market is riskier.
Figure 40.2 plots the estimated relative risk aversion parameters (b) and their
statistical significance over time from the rolling regressions with a fixed starting
date. Specifically, the first 250 daily return observations of exchange rates and
their realized variances (from 1/3/2002 to 1/7/2003) are used for estimation of the
relative risk aversion parameter for 1/8/2003. The sample is then extended by
adding one observation to the end (from 1/3/2002 to 1/8/2003), and the 1-day
ahead risk-return relation is measured for 1/9/2003. This recursive estimation
procedure is repeated until March 31, 2008.
1128 T.G. Bali and K. Yilmaz

U.S.Dollar – Euro U.S.Dollar – Japanese Yen

120 60
risk aversion risk aversion
t-statistic t-statistic 40
80

time -varying risk aversion

time -varying risk aversion

4 4 20
40
3
t-statistic

t-statistic
0
2
2 0
−20
1 0
−40 −40
0
−2
−1

−2 −4
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007
U.S.Dollar – British Pound U.S.Dollar – Swiss Franc
80 80
risk aversion risk aversion
t-statistic time -varying risk aversion t-statistic 60

time -varying risk aversion

40
2 4 40

0 20
t-statistic

1
t-statistic

3
0
0 −40 2
−20
−1 1
−80 −40

−2 0

−3 −1
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

U.S.Dollar – Australian Dollar U.S.Dollar – Canadian Dollar

60 60
risk aversion risk aversion
t-statistic 40 t-statistic 40

time- varying risk aversion

time- varying risk aversion

20 20
3
2 0 0
t-statistic
t-statistic

2
1
−20 −20
0 1
−40
−40
−1 0 −60
−2
−1
−3
−4 −2
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

Fig. 40.1 Rolling regression estimates from the fixed length window of 250 days

Similar to our findings from the fixed rolling window regressions, Fig. 40.2
provides evidence for a significant time variation in the risk aversion estimates for
all currencies considered in the paper. The first panel in Fig. 40.2 shows that in the
US dollar/euro market, the aggregate risk aversion is positive with a few exceptions
in January 2003. Only 16 out of 1,306 risk aversion estimates are negative, but none
of these estimates is statistically significant based on the Newey-West t-statistics.
795 (870) out of 1,290 positive risk aversion estimates turn out to be statistically
significant at least at the 5 % level (10 % level). These results indicate a positive and
strong time-varying risk aversion in the US dollar/euro market.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1129

U.S.Dollar – Euro U.S.Dollar – Japanese Yen

30 20
risk aversion risk aversion
t-statistic t-statistic 15
20

time-varying risk aversion

time-varying risk aversion

2.5 10

2.0 10 5
2.0
t-statistic

t-statistic
1.5 1.6 0
0
1.0 1.2 −5
0.8
0.5 −10
0.4
0.0
0.0
−0.5 −0.4
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

U.S.Dollar – British Pound U.S.Dollar – Swiss Franc

20 30
risk aversion risk aversion
2.0 t-statistic t-statistic
15
20
time-varying risk aversion

time-varying risk aversion

1.5 3
10
10
t-statistic

t-statistic

1.0 5 2
0
0
0.5 1
−5 −10
0.0 0
−10
−0.5 −1
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

U.S.Dollar – Australian Dollar U.S.Dollar – Canadian Dollar

20 25
risk aversion risk aversion
t-statistic t-statistic 20
15
2.5

time-varying risk aversion

time-varying risk aversion

15
3 10
2.0 10
t-statistic

t-statistic

5
2 1.5 5
0 1.0 0
1
−5 −5
0.5
0 −10
0.0

-1 −0.5
2003 2004 2005 2006 2007 2003 2004 2005 2006 2007

Fig. 40.2 Rolling regression estimates from the windows with fixed starting point

The second panel in Fig. 40.2 shows that in the US dollar/yen FX market, the
aggregate risk aversion is positive with a few exceptions from March to June 2004.
Only 68 out of 1,306 risk aversion estimates are negative and all of them are
statistically insignificant. 129 out of 1,238 positive risk aversion estimates turn
out to be marginally significant at the 10 % level. These results imply a positive but
statistically weak time-varying risk aversion in the US dollar/yen market.
The third panel in Fig. 40.2 depicts that in the pound sterling market, the risk
aversion is positive throughout the sample, except for a short period of time in 2003.
Only 90 out of 1,306 risk aversion estimates are negative, but they are not
statistically significant. Although there is a significant time variation in the risk
1130 T.G. Bali and K. Yilmaz

aversion and most of the risk-return coefficients are positive, only 41 out of 1,216
positive risk aversion estimates turn out to be significant at the 10 % level.
Therefore, it is not clear whether the currency trade generates a larger or smaller
risk premium at times when the US dollar/lb market is riskier.
The fourth panel in Fig. 40.2 provides evidence that in the Swiss franc market,
the risk aversion is estimated to be positive throughout the sample period
(2002–2008), except for a few days in January 2003. Only 20 out of 1,306 risk
aversion estimates are negative but statistically insignificant. 816 (934) out of 1,286
positive risk aversion estimates turn out to be statistically significant at least at the
5 % level (10 % level). These results suggest a positive and strong time-varying risk
aversion, implying that the currency trade generates a larger risk premium at times
when the US dollar/Swiss franc exchange rate market is riskier.
The fifth panel in Fig. 40.2 shows that in the Australian dollar market only
138 out of 1,306 risk aversion estimates are negative with no statistical significance
even at the 10 % level. Only 106 out of 1,168 positive risk aversion coefficients are
found to be marginally significant at the 10 % level. Although there is significant
time variation in the aggregate risk aversion, the results do not suggest a strong
positive or negative link between expected return and risk in the US/Australian
dollar market.
The last panel in Fig. 40.2 demonstrates that in the US/Canadian dollar market,
the risk aversion is estimated to be positive, except for a few days in May, October,
and November 2003. Similar to our earlier findings, only 41 out of 1,308 risk
aversion estimates are negative with very low t-statistics. However, based on the
statistical significance of positive risk aversion estimates, there is no evidence for
a strong positive link between expected return and risk on currency either.
Only 276 out of 1,265 positive risk aversion coefficients are found to be significant
at the 10 % level. Although there is a significant time-series variation in the
aggregate risk aversion, trading in the US/Canadian dollar FX market does not
provide clear evidence for a larger or smaller risk premium at times when the
market is riskier.

40.6 Testing Merton’s (1973) ICAPM in Currency Market

Merton’s (1973) ICAPM implies the following equilibrium relation between risk
and return for any risky asset i:

mi  r ¼ A  sim þ B  six , (40.19)

where r is the risk-free interest rate, mi  r is the expected excess return on the risky
asset i, sim denotes the covariance between the returns on the risky asset i and the
market portfolio m, and six denotes a (1  k) row of covariances between the
returns on risky asset i and the k state variables x. A denotes the average relative risk
aversion of market investors, and B measures the market’s aggregate reaction to
shifts in a k-dimensional state vector that governs the stochastic investment
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1131

opportunity. Equation 40.19 states that in equilibrium, investors are compensated in

terms of expected return for bearing market risk and for bearing the risk of
unfavorable shifts in the investment opportunity set.
Merton (1980) shows that the intertemporal hedging demand component (B  six)
is economically and statistically smaller than the market risk component (A  sim) of
ICAPM. While testing the significance of A and B at daily frequency, Bali and
Engle (2010) provide supporting evidence for Merton (1980) that the conditional
covariances of individual stocks with the market portfolio have positive and
statistically significant loading, whereas the innovations in state variables are not
priced in the stock market. That is, the conditional covariances of stock returns with
the unexpected news in state variables have insignificant loadings.
We examine Merton’s (1973) ICAPM based on the following system of
equations:

Ri, tþ1 ¼ Ci þ A  sim, t þ ei, tþ1

(40.20)
Rm, tþ1 ¼ Cm þ A  s2m, t þ em, tþ1

where the expected conditional covariance of individual exchange rates with the
currency market, Et(sim,t+1), is represented by the 1-day lagged realized covariance,
i.e., Et(sim,t+1) ¼ sim,t. Similarly, the expected conditional variance of the
currency market, Et(sm,t+12), is represented by the 1-day lagged realized variance,
i.e., Et(sm,t+12) ¼ sm,t2.12
The currency market portfolio is measured by the “value-weighted” average
returns on EUR, JPY, GBP, CHF, AUD, and CAD. The weights are obtained from
the “US Dollar Index.” Just as the Dow Jones Industrial Average reflects the general
state of the US stock market, the US Dollar Index (USDX) reflects the general
assessment of the US dollar. USDX does it through exchange rates averaging of US
dollar and six most tradable global currencies. The weights are 57.6 % for EUR,
13.6 % for JPY, 11.9 % for GBP, 9.1 % CAD, 4.2 % for AUD, and 3.6 % for CHF. In
our empirical analysis, daily returns on the currency market, Rm,t+1, are calculated by
multiplying daily returns on the six exchange rates by the aforementioned weights.
We estimate the system of Eq. 40.20 using an ordinary least square (OLS) as
well as a weighted least square method that allows us to place constraints on
coefficients across equations. We constrain the slope coefficient (A) on the lagged
realized variance-covariance matrix (sim,t, sm,t2) to the same value across all the
currencies for cross-sectional consistency. We allow the intercepts (Ci, Cm) to differ
across all the currencies. Under the null hypothesis of the ICAPM, the intercepts
should be jointly zero, and the common slope coefficient (A) should be positive and
statistically significant. We use insignificant estimates of A and the deviations from
zero of the intercept estimates as a test against the validity and sufficiency of

12
Daily realized covariances between the exchange rates and the currency market and daily
realized variance of the currency market are computed using 5-min returns in a day.
1132 T.G. Bali and K. Yilmaz

the ICAPM. In addition to the OLS panel estimates, we compute the t-statistics of
the parameter estimates accounting for heteroskedasticity and autocorrelation as
well as contemporaneous cross-correlations in the error terms. This estimation
methodology for the system of Eq. 40.20 can be regarded as an extension of the
seemingly unrelated regression (SUR) method.
Table 40.9 presents the OLS and SUR panel regression estimates of the
currency-specific intercepts, common slope coefficients on the lagged realized
variance-covariance matrix, and their t-statistics. The parameters and their
t-statistics are estimated using the daily returns on the currency market and the
six exchange rates. The last row reports the Wald statistics with p-values from
testing the joint hypothesis that all intercepts equal zero: H0:
C1 ¼ C2 ¼ . . . ¼ C6 ¼ Cm ¼ 0. A notable point in Table 40.9 is that the common
slope coefficient (A) is positive and statistically significant. Specifically, the risk
aversion coefficient on the realized variance-covariance matrix is estimated to be
23.33 with a t-statistic of 5.40. After correcting for heteroscedasticity, autocorrela-
tion, and contemporaneous cross-correlations, the SUR estimate of the risk aversion
coefficient turns out to be 15.80 with t-stat. ¼ 2.57. These results indicate
a positive and significant relation between risk and return on the currency market.
Another notable point in Table 40.9 is that for both the OLS and SUR estimates,
the Wald statistics reject the hypothesis that all intercepts equal zero. This implies
that the market risk alone cannot explain the entire time-series variation in
exchange rates.
According to the original ICAPM of Merton (1973), the relative risk aversion
coefficient (A) is restricted to the same value across all risky assets, and it is
positive and statistically significant. The common slope estimates in Table 40.9
provide empirical support for the positive risk-return trade-off.
We now test whether the slopes on (sim, sm2) are different across currencies.
We examine the sign and statistical significance of different slope coefficients
(Ai, Am) on (sim, sm2) in the following system of equations:
Ri, tþ1 ¼ Ci þ Ai  sim, t þ ei, tþ1 ,
(40.21)
Rm, tþ1 ¼ Cm þ Am  s2m, t þ em, tþ1
To determine whether there is a common slope coefficient (A) that corresponds
to the average relative risk aversion, we first estimate the currency-specific
slope coefficients (Ai, Am) and then test the joint hypothesis that H0:
A1 ¼ A 2 ¼ . . . ¼ A 6 ¼ Am .
Table 40.10 presents the OLS and SUR parameter estimates using daily returns
on the six exchange rates and the value-weighted currency market index. As
compared to Eq. 40.20, we have additional six-slope coefficients to estimate in
Eq. 40.21. As shown in Table 40.10, all of the slope coefficients (Ai, Am) are
estimated to be positive and highly significant without any exception. These
results indicate a positive and significant intertemporal relation between risk and
return on the currency market. We examine the cross-sectional consistency of the
intertemporal relation by testing the equality of slope coefficients based on the
Wald statistics. As reported in Table 40.10, the Wald statistics, from testing
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1133

Table 40.9 Testing Merton’s (1973) ICAPM with a common slope coefficient
OLS panel regression SUR panel regression
Intercept t-stat. Intercept t-stat.
AUD 0.00076 4.69 0.00063 3.15
EUR 0.00097 5.24 0.00077 3.53
GBP 0.00062 3.82 0.00049 2.95
CAD 0.00049 3.24 0.00041 2.77
CHF 0.00085 4.72 0.00066 2.98
JPY 0.00045 2.79 0.00032 1.79
Market 0.00078 4.50 0.00061 3.35
Risk aversion Slope t-stat. Slope t-stat.
23.33 5.40 15.80 2.57
H0: Intercepts ¼ 0 Wald p-value Wald p-value
52.20 0.00 17.98 0.0121
Entries report the OLS and SUR panel regression estimates based on the following system of
equations
Ri, tþ1 ¼ Ci þ A  sim, t þ ei, tþ1 ,
Rm, tþ1 ¼ Cm þ A  s2m, t þ em, tþ1 ,
where sim,t is the 1-day lagged realized covariance between the exchange rate and the currency
market. sm,t2 is the 1-day lagged realized variance of the currency market. A is a common slope
coefficient on the lagged realized variance-covariance matrix. (Ci, Cm) denotes currency-specific
intercepts for AUD, EUR, GBP, CAD, CHF, JPF, and the currency market. The last row reports the
Wald statistics with p-values from testing the joint hypothesis that all intercepts equal zero: H0:
C1 ¼ C2 ¼ . . . ¼ C6 ¼ Cm ¼ 0

the joint hypothesis that H0: A1 ¼ A2 ¼. . . ¼ A6 ¼ Am, is 1.41 for OLS and 5.24
for SUR, which fail to reject the null hypothesis. These results indicate the
equality of positive slope coefficients across all currencies, which empirically
validates the ICAPM.

40.7 Conclusion

There is an ongoing debate in the literature about the intertemporal relation between
stock market risk and return and the extent to which expected stock returns are
related to expected market volatility. Recently, some studies have provided evi-
dence for a significantly positive link between risk and return in the aggregate stock
market, but the risk-return trade-off is generally found to be insignificant and
sometimes even negative. This paper is the first to investigate the presence and
significance of an intertemporal relation between expected return and risk in the
foreign exchange market. The paper provides new evidence on the ICAPM by
using high-frequency intraday data on currency and by presenting significant time
variation in the risk aversion parameter. We utilize daily and 5-min returns on the
spot exchange rates of the US dollar vis-à-vis six major currencies (the euro, Japanese
yen, British pound sterling, Swiss franc, Australian dollar, and Canadian dollar)
 1134

Table 40.10 Testing Merton’s (1973) ICAPM with different slope coefficients
OLS panel regression SUR panel regression
Intercept t-stat. Slope t-stat. Intercept t-stat. Slope t-stat.
AUD 0.00084 2.95 28.51 2.08 0.00072 2.85 21.01 1.90
EUR 0.00080 2.80 17.12 1.83 0.00071 3.13 13.62 2.06
GBP 0.00060 2.35 22.19 1.73 0.00030 1.38 4.60 0.46
CAD 0.00063 3.04 35.92 2.47 0.00041 2.23 15.67 1.36
CHF 0.00082 2.66 21.92 2.10 0.00070 2.80 17.32 2.26
JPY 0.00048 1.89 25.25 2.06 0.00050 2.30 26.29 2.77
Market 0.00075 3.08 21.92 2.31 0.00059 3.14 14.53 2.29
H0: Intercepts ¼ 0 Wald p-value Wald p-value
51.40 0.00 15.15 0.0341
H0: Equal slopes Wald p-value Wald p-value
1.41 0.9655 5.24 0.5131
Entries report the OLS and SUR panel regression estimates based on the following system of equations

Ri, tþ1 ¼ Ci þ Ai  sim, t þ ei, tþ1 ,

Rm, tþ1 ¼ Cm þ Am  s2m, t þ em, tþ1 ,

where sim,t is the 1-day lagged realized covariance between the exchange rate and the currency market. s2m,t is the 1-day lagged realized variance of the
currency market. (Ai, Am) denotes currency-specific slope coefficients on the lagged realized variance-covariance matrix. (Ci, Cm) denotes currency-specific
intercepts for AUD, EUR, GBP, CAD, CHF, JPF, and the currency market. The Wald statistics with p-values are reported from testing the joint hypothesis that
all intercepts equal zero: H0:C1 ¼ C2 ¼ . . . ¼ C6 ¼ Cm ¼ 0. The last row presents the Wald statistics from testing the equality of currency-specific slope
coefficients H0: A1 ¼ A2 ¼ . . . ¼ A6 ¼ Am
T.G. Bali and K. Yilmaz
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1135

and test the existence and significance of a risk-return trade-off in the FX market
using the GARCH, realized, and range-based volatility measures. The maximum
likelihood parameter estimates of the GARCH-in-mean model and the risk-return
regressions with daily realized and range volatility estimators indicate that the
intertemporal relation between risk and return is generally positive but statistically
weak in the FX market.
We provide strong evidence on the time variation of risk aversion parameters for
all currencies considered in the paper. However, the direction of a relationship
between expected return and risk is not clear. The results indicate a positive but
not strong time-varying risk aversion in the US dollar/euro exchange rate market.
The risk-return regressions with realized variance provide evidence for a positive
but statistically weak risk aversion estimates in the US dollar/yen market.
Although there is a significant time variation in risk aversion estimates for the
British pound, it is not clear whether the currency trade generates a larger or smaller
risk premium at times when the US dollar/lb market is riskier. The risk aversion
parameter is estimated to be positive but marginally significant throughout the
sample period for the Swiss franc, implying that the currency trade generally yields
a larger risk premium at times when the US dollar/Swiss franc market is riskier. For
most of the sample, the risk-return coefficients are estimated to be positive but
statistically insignificant for the Canadian dollar, suggesting that the intertemporal
relation between risk and return is flat for the US/Canadian dollar market.

Appendix 1: Maximum Likelihood Estimation of GARCH-in-Mean

Models

Modeling and estimating the volatility of financial time series has been high on
the agenda of financial economists since the early 1980s. Engle (1982) put forward
the Autoregressive Conditional Heteroskedastic (ARCH) class of models for
conditional variances which proved to be extremely useful for analyzing financial
return series. Since then an extensive literature has been developed for modeling the
conditional distribution of stock prices, interest rates, exchange rates, and futures
prices. Following the introduction of ARCH models by Engle (1982) and their
generalization by Bollerslev (1986), there have been numerous refinements of this
approach to estimating conditional volatility. Most of the refinements have been
driven by empirical regularities in financial data.
Engle (1982) introduces ARCH(p) model:

Rtþ1  a þ etþ1 (40.22)

etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1Þ,

(40.23)
Eðetþ1 Þ ¼ 0


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 e2t1 þ . . . þ gp e2tp (40.24)
1136 T.G. Bali and K. Yilmaz

"
#

 1 1 Rtþ1  m 2
f Rtþ1 ; m; stþ1jt ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp  (40.25)
2ps2tþ1jt 2 stþ1jt

where Rt+1 is the daily return for period t+1, m ¼ a is the constant conditional mean,
et+1 ¼ zt+1  st+1|t is the error term with E(et +1) ¼ 0, st+1|t is the conditional standard
deviation of daily returns, and zt+1  N(0,1) is a random variable drawn from the
standard normal density and can be viewed as information shocks or unexpected
news in the market. st+1|t2 is the conditional variance of daily returns based on the
information set up to time t denoted by Ot. The conditional variance, st+1|t2, follows
an ARCH(p) process which is a function of the last period’s unexpected news
(or information shocks). f(Rt+1;m,st+1|t) is the conditional normal density function of
Rt+1 with the conditional mean of m and conditional variance of st+1|t2.
Given the initial values of et and, the parameters in Eqs 40.22 and 40.24 can be
estimated by maximizing the log-likelihood function over the sample period.
The conditional normal density in Eq. 40.25 yields the following log-likelihood
function:
n

n n 1X Rtþ1  m 2
LogLARCH ¼  lnð2pÞ  lnstþ1jt  (40.26)
2 2 2 t¼1 stþ1jt
Bollerslev (1986) extends the original work of Engle (1982) and defines the
current conditional variance as a function of the last period’s unexpected news as
well as the last period’s conditional volatility:
Rtþ1  a þ etþ1 (40.27)
etþ1 ¼ ztþ1  stþ1jt , ztþ1  N ð0; 1ÞEðetþ1 Þ ¼ 0 (40.28)


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t (40.29)

where the conditional variance, st+1|t2, in Eq. 40.29 follows a GARCH(1,1) process
as defined by Bollerslev (1986) to be a function of the last period’s unexpected
news (or information shocks), zt, and the last period’s variance, s2t . The parameters
in Eqs. 40.27, 40.28, and 40.29 are estimated by maximizing the conditional
log-likelihood function in Eq. 40.26.
Engle et al. (1987) introduce the ARCH-in-mean model in which the conditional
mean of financial time series is defined as a function of the conditional variance. In
our empirical investigation of the ICAPM for exchange rates, we use the following
GARCH-in-mean process to model the intertemporal relation between expected
return and risk on currency
Rtþ1  a þ b  s2tþ1jt þ etþ1 (40.30)
etþ1 ¼ ztþ1  stþ1jt ; ztþ1  N ð0; 1Þ;
(40.31)
Eðetþ1 Þ ¼ 0


E e2tþ1 jOt ¼ s2tþ1jt ¼ g0 þ g1 e2t þ g2 s2t (40.32)
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1137

"
#
  1 1 Rtþ1  mtþ1jt 2
f Rtþ1 ; mtþ1jt ; stþ1jt ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp  (40.33)
2ps2tþ1jt 2 stþ1jt

where Rt+1 is the daily return on exchange rates for period t+1, mt+1|t  a + b  st+1|t
2
is
the conditional mean for period t+1 based on the information set up to time t,
et+1 ¼ zt+1  st+1|t is the error term with E(et+1) ¼ 0, st+1|t is the conditional standard
deviation of daily returns on currency, and zt+1  N(0,1) is a random variable drawn
from the standard normal density and can be viewed as information shocks in the FX
market. st+1|t
2
is the conditional variance of daily returns based on the information set
up to time t denoted by Ot. The conditional variance, st+1|t 2
, follows a GARCH(1,1)
process as defined by Bollerslev (1986) to be a function of the last period’s
unexpected news (or information shocks), zt, and the last period’s variance,
st2. f(Rt+1;mt+1|t,st+1|t) is the conditional normal density function of Rt+1 with the
conditional mean of mt+1|t and conditional variance of st+1|t
2
.
Given the initial values of et and, the parameters in Eqs. 40.30 and 40.32 can be
estimated by maximizing the log-likelihood function over the sample period. The
conditional normal density in Eq. 40.33 yields the following log-likelihood function

n n
LogLARCH ¼  lnð2pÞ  lnstþ1jt
2 2

2 (40.34)
1 X Rtþ1  mtþ1jt
n

2 t¼1 stþ1jt

where the conditional mean mt+1|t  a + b  s2t+1|t has two parameters and the
conditional variance st+1|t2 ¼ g0 + g1et2 + g2s2t has three parameters. Maximizing
the log-likelihood in Eq. 40.34 yields the parameter estimates (a, b, g0, g1, g2).
The interested reader may wish to consult Enders (2009), Chap. 3, and Tsay
(2010), Chap. 3, for comprehensive analysis of ARCH/GARCH models and their
maximum likelihood estimation. Chapter 3 in Enders (2009) provides a detailed
coverage of the basic ARCH and GARCH models, as well as the GARCH-in-mean
processes and multivariate GARCH in some detail. Chapter 3 in Tsay (2010) pro-
vides a detailed coverage of the ARCH, GARCH, GARCH-M, the exponential
GARCH, and Threshold GARCH models.

Appendix 2: Estimation of a System of Regression Equations

Consider a system of n equations, of which the typical ith equation is

y i ¼ X i b i þ ui (40.35)

where yi is a N  1 vector of time-series observations on the i th dependent variable,

Xi is a N  ki matrix of observations of ki independent variables, bi is a ki  1 vector
1138 T.G. Bali and K. Yilmaz

of unknown coefficients to be estimated, and ui is a N  1 vector of random

disturbance terms with mean zero. Parks (1967) proposes an estimation procedure
that allows the error term to be both serially and cross-sectionally correlated.
In particular, he assumes that the elements of the disturbance vector u follow an
AR(1) process

uit ¼ ri uit1 þ eit ; ri < 1 (40.36)

where eit is serially independently but contemporaneously correlated:



Cov eit ejt ¼ sij , 8i, j, and Cov eit ejs ¼ 0, for s 6¼ t (40.37)

Equation 40.35 can then be written as

yi ¼ Xi bi þ Pi ei , (40.38)

with
2
1=2 3
1  r2i 0 0 ... 0
6 7
6 r 1  r2 1=2


... 07
6 i
i

1 0 7
6 2 7
6 ri 1  r2i 1=2 ri 0 ... 07
6
Pi ¼ 6 7 (40.39)
: 7
6 7
6 : 7
6 7
4 5

: 2 1=2
rN1
i 1  ri rN2
i rN3
i ... 1

Under this setup, Parks presents a consistent and asymptotically efficient three-
step estimation technique for the regression coefficients. The first step uses single
equation regressions to estimate the parameters of autoregressive model. The
second step uses single equation regressions on transformed equations to estimate
the contemporaneous covariances. Finally, the Aitken estimator is formed using the
estimated covariance,


^ ¼ XT O1 X 1 XT O1 y
b (40.40)

Where O  E[uuT] denotes the general covariance matrix of the innovation. In

my application, I use the aforementioned methodology with the slope coefficients
restricted to be the same for all portfolios. In particular, we use the same three-step
procedure and the same covariance assumptions as in Eqs. 40.36, 40.37, 40.38,
40.39, and 40.40 to estimate the covariances and to generate the t-statistics for the
parameter estimates.
The interested reader may wish to consult Wooldridge (2010), Chaps. 10.4, 10.5,
and 10.6 for recent developments on panel data estimation. Chapter 10 in
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1139

Wooldridge (2010) presents Basic Linear Unobserved Effects Panel Data

Models, Chap. 10.4 provides Random Effects Methods, Chap. 10.5 contains
Fixed Effects Methods, and Chap. 10.6 First Differencing Methods. Bali (2008)
and Bali and Engle (2010) follow SUR estimation to investigate the
empirical validity of the conditional intertemporal capital asset pricing models
(ICAPM).

References
Abel, A. B. (1988). Stock prices under time-varying dividend risk: An exact solution in an infinite-
horizon general equilibrium model. Journal of Monetary Economics, 22, 375–393.
Alizadeh, S., Brandt, M. W., & Diebold, F. X. (2002). Range-based estimation of stochastic
volatility models. Journal of Finance, 57, 1047–1092.
Andersen, T. G., Bollerslev, T., & Diebold, F. X. (2004). Parametric and nonparametric volatility
measurement. In Y. Aı̈t-Sahalia & L. P. Hansen (Eds.), Handbook of financial econometrics.
Amsterdam: North Holland.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Ebens, H. (2001a). The distribution of realized
stock return volatility. Journal of Financial Economics, 61, 43–76.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2001b). The distribution of realized
exchange rate volatility. Journal of the American Statistical Association, 96, 42–55.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2003). Modeling and forecasting
realized volatility. Econometrica, 71, 579–626.
Backus, D. K., & Gregory, A. W. (1993). Theoretical relations between risk premiums and
conditional variance. Journal of Business and Economic Statistics, 11, 177–185.
Bali, T. G. (2008). The intertemporal relation between expected returns and risk. Journal of
Financial Economics, 87, 101–131.
Bali, T. G., & Engle, R. F. (2010). The intertemporal capital asset pricing model with dynamic
conditional correlations. Journal of Monetary Economics, 57, 377–390.
Bali, T. G., & Peng, L. (2006). Is there a risk-return tradeoff? Evidence from high frequency data.
Journal of Applied Econometrics, 21, 1169–1198.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroscedasticity. Journal of
Econometrics, 31, 307–327.
Bollerslev, T., & Domowitz, I. (1993). Trading patterns and prices in the interbank foreign
exchange market. Journal of Finance, 48, 1421–1443.
Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1988). A capital asset pricing model with time-
varying covariances. Journal of Political Economy, 96, 116–131.
Bollerslev, T., & Wooldridge, J. M. (1992). Quasi-maximum likelihood estimation and inference
in dynamic models with time varying covariances. Econometric Reviews, 11, 143–172.
Bollerslev, T., & Zhou, H. (2006). Volatility puzzles: A simple framework for gauging return-
volatility regressions. Journal of Econometrics, 131, 123–150.
Brandt, M. W., & Diebold, F. X. (2006). A no-arbitrage approach to range-based estimation of
return covariances and correlations. Journal of Business, 79, 61–73.
Brandt, M. W., & Kang, Q. (2004). On the relationship between the conditional mean and
volatility of stock returns: A latent VAR approach. Journal of Financial Economics, 72,
217–257.
Breen, W., Glosten, L. R., & Jagannathan, R. (1989). Economic significance of predictable
variations in stock index returns. Journal of Finance, 44, 1177–1189.
Campbell, J. Y. (1987). Stock returns and the term structure. Journal of Financial Economics, 18,
373–399.
Campbell, J. Y. (1991). A variance decomposition for stock returns. Economic Journal, 101,
157–179.
1140 T.G. Bali and K. Yilmaz

Campbell, J. Y., & Hentchel, L. (1992). No news is good news: An asymmetric model of changing
volatility in stock returns. Journal of Financial Economics, 31, 281–318.
Campbell, J. Y., & Shiller, R. (1988). The dividend-price ratio and expectations of future
dividends and discount factors. Review of Financial Studies, 1, 195–228.
Chan, K. C., Karolyi, G. A., & Stulz, R. M. (1992). Global financial markets and the risk premium
on U.S. equity. Journal of Financial Economics, 32, 137–167.
Chen, N.-F., Roll, R., & Ross, S. A. (1986). Economic forces and the stock market. Journal of
Business, 59, 383–403.
Chou, R., Engle, R. F., & Kane, A. (1992). Measuring risk aversion from excess returns on a stock
index. Journal of Econometrics, 52, 201–224.
Christoffersen, P. F., & Diebold, F. X. (2006). Financial asset returns, direction-of-change,
forecasting, and volatility dynamics. Management Science, 52, 1273–1287.
Enders, W. (2009). Applied econometric time series (3rd ed.). New York: Wiley.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance
of United Kingdom inflation. Econometrica, 50, 987–1007.
Engle, R. F., Lilien, D. M., & Robins, R. P. (1987). Estimating time varying risk premia in the term
structure: The ARCH M model. Econometrica, 55, 391–407.
Fama, E. F., & French, K. R. (1988). Dividend yields and expected stock returns. Journal of
Financial Economics, 22, 3–25.
Fama, E. F., & French, K. R. (1989). Business conditions and expected returns on stocks and
bonds. Journal of Financial Economics, 25, 23–49.
French, K. R., Schwert, G. W., & Stambaugh, R. F. (1987). Expected stock returns and volatility.
Journal of Financial Economics, 19, 3–29.
Gennotte, G., & Marsh, T. A. (1993). Variations in economic uncertainty and risk premiums on
capital assets. European Economic Review, 37, 1021–1041.
Ghysels, E., Santa-Clara, P., & Valkanov, R. (2005). There is a risk-return tradeoff after all.
Journal of Financial Economics, 76, 509–548.
Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the relation between the expected
value and the volatility of the nominal excess returns on stocks. Journal of Finance, 48,
1779–1801.
Goyal, A., & Santa-Clara, P. (2003). Idiosyncratic risk matters! Journal of Finance, 58, 975–1008.
Guo, H., & Whitelaw, R. F. (2006). Uncovering the risk-return relation in the stock market.
Journal of Finance, 61, 1433–1463.
Harrison, P., & Zhang, H. (1999). An investigation of the risk and return relation at long horizons.
Review of Economics and Statistics, 81, 399–408.
Harvey, C. R. (2001). The specification of conditional expectations. Journal of Empirical Finance,
8, 573–638.
Jegadeesh, N. (1990). Evidence of predictable behavior of security returns. Journal of Finance, 45,
881–898.
Keim, D. B., & Stambaugh, R. F. (1986). Predicting returns in the stock and bond markets. Journal
of Financial Economics, 17, 357–390.
Lehmann, B. (1990). Fads, martingales, and market efficiency. Quarterly Journal of Economics,
105, 1–28.
Lettau, M., & Ludvigson, S. C. (2010). Measuring and modeling variation in the risk-return
tradeoff. In Handbook of Financial Econometrics by Y. Ait-Sahalia & L. P. Hansen (Eds.),
Vol 1, pp. 617–690. Amsterdam: North Holland.
Lo, A. W., & MacKinlay, A. C. (1990). When are contrarian profits due to stock market
overreaction? Review of Financial Studies, 3, 175–205.
Lundblad, C. (2007). The risk return tradeoff in the long run: 1836–2003. Journal of Financial
Economics, 85, 123–150.
Merton, R. C. (1973). An intertemporal capital asset pricing model. Econometrica, 41, 867–887.
Merton, R. C. (1980). On estimating the expected return on the market: An exploratory investi-
gation. Journal of Financial Economics, 8, 323–361.
40 The Intertemporal Relation Between Expected Return and Risk on Currency 1141

Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.

Econometrica, 59, 347–370.
Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica, 55, 703–708.
Parkinson, M. (1980). The extreme value method for estimating the variance of the rate of return.
Journal of Business, 53, 61–65.
Parks, R. W. (1967). Efficient estimation of a system of regression equations when disturbances
are both serially and contemporaneously correlated. Journal of the American Statistical
Association, 62, 500–509.
Scruggs, J. T. (1998). Resolving the puzzling intertemporal relation between the market risk
premium and conditional market variance: A two-factor approach. Journal of Finance, 53,
575–603.
Sentana, E. (1995). Quadratic ARCH models. Review of Economic Studies, 62, 639–661.
Tsay, R. S. (2010). Analysis of financial time series. New York: Wiley.
Turner, C. M., Startz, R., & Nelson, C. R. (1989). A markov model of heteroskedasticity, risk, and
learning in the stock market. Journal of Financial Economics, 25, 3–22.
Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.). Boston:
MIT Press.
Quantile Regression and Value at Risk
41
Zhijie Xiao, Hongtao Guo, and Miranda S. Lam

Contents
41.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1144
41.2 Traditional Estimation Methods of VaR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1146
41.3 Quantile Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1147
41.4 Autoregressive Quantile Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1149
41.4.1 The QAR Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1149
41.4.2 Nonlinear QAR and Copula-Based Quantile Models . . . . . . . . . . . . . . . . . . . . . . . . 1151
41.4.3 The CaViaR Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1153
41.5 Quantile Regression of Conditional Heteroskedastic Models . . . . . . . . . . . . . . . . . . . . . . . . . 1153
41.5.1 ARCH Quantile Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1153
41.5.2 GARCH Quantile Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1155
41.6 An Empirical Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1159
41.6.1 Data and the Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1159
41.6.2 Model Specification Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1160
41.6.3 Estimated VaRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1163
41.6.4 Performance of the ARCH Quantile Regression Model . . . . . . . . . . . . . . . . . . . . . 1165
41.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1166
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1167

Abstract
This paper studies quantile regression (QR) estimation of Value at Risk (VaR).
VaRs estimated by the QR method display some nice properties. In this paper,
different QR models in estimating VaRs are introduced. In particular, VaR

We thank Prof. C.-F. Lee for helpful comments. This project is partially supported by Boston
College Research fund.
Z. Xiao (*)
Department of Economics, Boston College, Chestnut Hill, MA, USA
e-mail: xiaoz@bc.edu
H. Guo • M.S. Lam
Bertolon School of Business, Salem State University, Salem, MA, USA
e-mail: hguo@salemstate.edu; miranda.lam@salemstate.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1143
DOI 10.1007/978-1-4614-7750-1_41,
# Springer Science+Business Media New York 2015
1144 Z. Xiao et al.

estimations based on quantile regression of the QAR models, copula models,

ARCH models, GARCH models, and the CaViaR models are systematically
introduced. Comparing the proposed QR method with traditional methods based
on distributional assumptions, the QR method has the important property in that
it is robust to non-Gaussian distributions. Quantile estimation is only influenced
by the local behavior of the conditional distribution of the response near the
specified quantile. As a result, the estimates are not sensitive to outlier observa-
tions. Such a property is especially attractive in financial applications since
many financial data like, say, portfolio returns (or log returns) are usually not
normally distributed. To highlight the importance of the QR method in estimat-
ing VaR, we apply the QR techniques to estimate VaRs in International Equity
Markets. Numerical evidence indicates that QR is a robust estimation method
for VaR.

Keywords
ARCH • Copula • GARCH • Non-normality • QAR • Quantile regression • Risk
management • Robust estimation • Time series • Value at risk

41.1 Introduction

The Value at Risk (VaR) is the loss in market value over a given time horizon that is
exceeded with probability t, where t is often set at 0.01 or 0.05. In recent years,
VaR has become a popular tool in the measurement and management of financial
risk. This popularity is spurred both by the need of various institutions for managing
risk and by government regulations (see Blankley et al., 2000; Dowd 1998, 2000;
Saunders 1999). VaR is an easily interpretable measure of risk that summarizes
information regarding the distribution of potential losses. In requiring publicly
traded firms to report risk exposure, the Securities and Exchange Commission
(SEC) lists VaR as a disclosure method “expressing the potential loss in future
earnings, fair values, or cash flows from market movements over a selected period
of time and with a selected likelihood of occurrence.”
Estimation of VaR has attracted much attention from researchers (Duffie and
Pan (1997); Wu and Xiao (2002); Guo et al. (2007)). Many existing methods of VaR
estimation in economics and finance are based on the assumption that financial
returns have normal (or conditional normal) distributions (usually with ARCH or
GARCH effects). Under the assumption of a conditionally normal return distribu-
tion, the estimation of conditional quantiles is equivalent to estimating conditional
volatility of returns. The massive literature on volatility modeling offers a rich
source of parametric methods of this type. However, there is accumulating evidence
that financial time series and return distributions are not well approximated by
Gaussian models. In particular, it is frequently found that market returns display
negative skewness and excess kurtosis. Extreme realizations of returns can
adversely affect the performance of estimation and inference designed for Gaussian
conditions; this is particularly true of ARCH and GARCH models whose
41 Quantile Regression and Value at Risk 1145

estimation of variances is very sensitive to large innovations. For this reason,

research attention has recently shifted toward the development of more robust
estimators of conditional quantiles.
There is also growing interest in nonparametric estimation of conditional quantiles.
However, nearest neighbor and kernel methods are somewhat limited in their ability
to cope with more than one or two covariates. Other approaches to estimating VaR
include the hybrid method and methods based on extreme value theory.
Quantile regression as introduced by Koenker and Bassett (1978) is well suited
to estimating VaR. Value at Risk, as mandated in many current regulatory contexts,
is a conditional quantile by definition. This concept is intimately linked to quantile
regression estimation.
Quantile regression has now become a popular robust approach for statistical
analysis. Just as classical linear regression methods based on minimizing sums of
squared residuals enable one to estimate models for conditional mean, quantile
regression methods offer a mechanism for estimating models for the conditional
quantiles. These methods exhibit robustness to extreme shocks and facilitate
distribution-free inference. In recent years, quantile regression estimation for time-
series models has gradually attracted more attention. Koenker and Zhao (1996)
extended quantile regression to linear ARCH models and estimate conditional
quantiles by a linear quantile regression. Engle and Manganelli (1999) have suggested
a nonlinear dynamic quantile model where conditional quantiles themselves follow an
autoregression, and they call this a Conditional Autoregressive Value at Risk
(CaViaR) specification. Computation of the CaViaR model is challenging and grid
searching is conventionally used in practice. Koenker and Xiao (2006) investigate
quantile autoregressive processes that can capture systematic influences of condition-
ing variables on the location, scale, and shape of the conditional distribution of the
response and therefore constitute a significant extension of classical time-series
models in which the effect of conditioning is confined to a location shift. Xiao and
Koenker (2009) recently studied quantile regression estimation of GARCH models.
GARCH models have proven to be highly successful in modeling financial data and
are arguably the most widely used class of models in financial applications. However,
quantile regression GARCH models are highly nonlinear and thus complicated to
estimate. The quantile estimation problem in GARCH models corresponds to
a restricted nonlinear quantile regression, and conventional nonlinear quantile regres-
sion techniques are not directly applicable, adding an additional challenge to the
already complicated estimation problem. Koenker and Xiao (2009) propose a two-step
approach for quantile regression on GARCH models. The proposed method is
relatively easy to implement compared to other nonlinear estimation techniques in
quantile regression and has good sampling performance in our simulation experiments.
VaRs estimated by the quantile regression approach display some nice properties.
For example, they track VaRs estimated from GARCH volatility models well during
normal market conditions. However, during market turmoils when large market price
drops are followed by either further drops or rebounds, GARCH volatility models
tend to predict implausibly high VaRs. This is due to the fact that volatility and VaRs
are not synonymous. While large positive and negative return shocks indicate
1146 Z. Xiao et al.

higher volatility, only large negative return shocks indicate higher Value at Risk.
GARCH models treat both large positive and negative return shocks as indications of
higher volatility. VaRs estimated by the ARCH/GARCH quantile regression model,
while predicting higher volatility in the ARCH/GARCH component, assign a much
bigger weight to the large negative return shock than the large positive return shock.
The resulting estimated VaRs seem to be closer to reality.
In this chapter, we study quantile regression estimation of VaR. Different quantile
models in estimating VaR are introduced in this paper. In particular, Value at Risk
analysis based on quantile regression of the QAR models, copula models, ARCH
models, GARCH models, and the CaViaR models is systematically introduced. To
highlight the importance of quantile regression method in estimating VaR, we apply
the quantile regression techniques to estimate VaR in International Equity Markets.
Numerical evidence indicates that quantile regression is a robust estimation method
for VaR.
This chapter is organized as follows. We introduce the traditional VaR estima-
tion methods and quantile regression in Sect. 41.2. The quantile autoregression
(QAR) models are given in Sect. 41.3; nonlinear QAR models based on copula and
the CaViaR models are also introduced. Section 41.4 introduces quantile regression
estimation on ARCH and GARCH models. Section 41.5 contains an empirical
application of quantile regression estimation of VaRs. Section 41.6 concludes.

41.2 Traditional Estimation Methods of VaR

For a time series of returns on an asset, {rt}nt¼1, the t (or 100t%) VaR at time t,
denoted by VaRt, is defined by

Prðr t < VaRt jF t1 Þ ¼ t (41.1)

where F t1 denotes the information set at time t1, including past values of returns
and possibly the value of some covariates Xt.
If we assume that the time series of returns are modeled by

r t ¼ mt þ st et ,

where rt is the return of an asset at time t and mt, st 2 F t1. The random variables et are
martingale difference sequences. The Conditional Value at Risk of rt given F t1 is
VaRt ðtÞ ¼ mt þ st Qe ðtÞ,

where Qe(t) denotes the Unconditional Value at Risk of the error term et. Assuming
conditional normality, the 5 % VaR at time t can be computed as

VaRt ð0:05Þ ¼ mt þ 1:65st

where mt and st are the conditional mean and conditional volatility for rt.
41 Quantile Regression and Value at Risk 1147

RiskMetrics takes a simple and pragmatic approach to modeling the conditional

volatility. The forecast for time t variance in RiskMetrics method is a weighted
average of the previous forecast, using weight l, and of the latest squared innova-
tion, using weight (1l):

s2t ¼ ls2t1 þ ð1  lÞr 2t1 (41.2)

where the parameter l is called the decay factor (1 > l > 0). Conceptually l should
be estimated using a maximum likelihood approach. RiskMetrics simply set it
optimally at 0.94 for daily data and 0.97 for monthly data. Our analysis is on
weekly data and we set l at 0.95.
There are extensive empirical evidences supporting the use of ARCH and
GARCH models in conditional volatility estimation. Bollerslev et al. (1992) pro-
vide a nice overview of the issue. Sarma et al. (2000) showed that at the 5 % level,
an AR(1)-GARCH(1,1) model is a preferred model under the conditional normality
assumption. The AR(1)-GARCH(1,1) model is specified:



r tþ1 ¼ a0 þ a1 r t þ 2tþ1 2tþ1 I t  N 0; s2t

s2t ¼ o0 þ o1 s2t1 þ o2 22t : (41.3)

The conditional mean equation is modeled as an AR(1) process to account for

the weakly autoregressive behavior of returns.

41.3 Quantile Regression

Quantile regression was introduced by Koenker and Bassett (1978) and has
received a lot of attention in econometrics and statistics research in the past two
decades. The quantile function of a scalar random variable Y is the inverse of its
distribution function. Similarly, the conditional quantile function of Y given X is the
inverse of the corresponding conditional distribution function, i.e.,

QY tjXÞ ¼ F1
Y ðtjXÞ ¼ inf fy : FY ðyjXÞ  tg,

where FY(yjX) ¼ P(Y  yjX). By definition, the t VaR at time t is the t-th
conditional quantile of rt giving information at time t1.
Consider a random variable Y characterized by its distribution function F(y), the
t-th quantile of Y is defined by

QY ðtÞ ¼ inf fyjFðyÞ  tg:

1148 Z. Xiao et al.

If we have a random sample {y1, . . ., yn} from the distribution F, the t-th sample
quantile can be defined as
  
^ Y ðtÞ ¼ inf yF
Q ^ ðyÞ  t ,

^ is the empirical distribution function of the random sample. Note that the
where F
above sample quantile may be found by solving the following minimization
problem:
2 3
X X
min 4 t j y t  bj þ ð1  tÞjyt  bj5: (41.4)
b2<
t2ft:yt bg t2ft:yt <bg

Koenker and Bassett (1978) studied the analogue of the empirical quantile
function for the linear models and generalized the concept of quantiles to the
regression context.
If we consider the linear regression model
0
Y t ¼ b X t þ ut , (41.5)

where ut are iid mean zero random variables with quantile function Qu(t) and Xt are
k-by-1 vector of regressors including an intercept term and lagged residuals, then,
conditional on the regressor Xt, the t-th quantile of Y is a linear function of Xt:

QY t ðtjXt Þ ¼ b0 Xt þ Qu ðtÞ ¼ bðtÞ0 Xt

where b(t)0 ¼ (b1 + Qu(t), b2,   , bk). Koenker and Bassett (1978) show that the
t-th conditional quantile of Y can be estimated by an analogue of Eq. 41.4:

^
^ Y ðtjXt Þ ¼ X0 b
Q t t ðtÞ

where
2 3
X  0 
X  0 
^ ðtÞ ¼ arg min 4
b t y t  x t b  þ ð1  tÞyt  xt b5 (41.6)
b2<k t2ft:yt xt bg t2ft:yt <xt bg

is called as the regression quantiles. Let rt(u) ¼ u(tI(u < 0)), then
X  
^ ðtÞ ¼ arg min
b
0
rt yt  xt b :
b2<k t

Quantile regression method has the important property that it is robust to

distributional assumptions. This property is inherited from the robustness property
of the ordinary sample quantiles. Quantile estimation is only influenced by the local
41 Quantile Regression and Value at Risk 1149

behavior of the conditional distribution of the response near the specified quantile.
As a result, the estimated coefficient vector ^y ðtÞ is not sensitive to outlier obser-
vations. Such a property is especially attractive in financial applications since many
financial data like, say, portfolio returns (or log returns) are usually heavy tailed and
thus not normally distributed.
The quantile regression model has a mathematical programming representation
which facilitates the estimation. Notice that the optimization problem (Eq. 41.6)
may be reformulated as a linear program by introducing “slack” variables to
represent the positive and negative parts of the vector of residuals (see Koenker
and Bassett (1978) for a more detailed discussion). Computation of the regression
quantiles by standard linear programming techniques is very efficient. It is also
straightforward to impose the nonnegativity constraints on all elements of g.
Barrodale and Roberts (1974) proposed the first efficient algorithm for L1-
estimation problems based on modified simplex method. Koenker and d’Orey
(1987) modified this algorithm to solve quantile regression problems. For very
large quantile regression problems, there are some important new ideas which
speed up the performance of computation relative to the simplex approach under-
lying the original code. Portnoy and Koenker (1997) describe an approach that
combines some statistical preprocessing with interior point methods and achieves
faster speed over the simplex method for very large problems.

41.4 Autoregressive Quantile Regression Models

41.4.1 The QAR Models

In many finance applications, the time-series dynamics can be more complicated

than the classical autoregression where past information (Ytj) influences only the
location of the conditional distribution of Yt. For example, it is well known that
the correlations tend to be larger in bear than in bull markets. Recognizing that the
correlation is asymmetric is important for risk management and other
financial applications. Any attempt to diagnose or forecast series of this type
requires that a mechanism be introduced to capture the empirical features of
the series.
An important extension of the classical constant coefficient time-series model is
the quantile autoregression (QAR) model (Koenker and Xiao 2006). Given a time
series {Yt}, let F t be the s-field generated by {Ys, s  t}; {Yt} is a p-th order QAR
process if


QY t tjF t1 Þ ¼ y0 ðtÞ þ y1 ðtÞY t1 þ    þ yp ðtÞY tp ; (41.7)

this implies, of course, that the right-hand side of Eq. 41.7 is monotonically
increasing in t. In the above QAR model, the autoregressive coefficients may be
t-dependent and thus can vary over different quantiles of the conditional
1150 Z. Xiao et al.

distribution. Consequently, the conditioning variables not only shift the location of
the distribution of Yt but also may alter the scale and shape of the conditional
distribution. The QAR models play a useful role in expanding the modeling
territory of the classical autoregressive time-series models, and the classical
AR(p) model can be viewed as a special case of QAR by setting yj(t) (j ¼ 1, . . ., p)
to constants.
Koenker and Xiao (2006) studied the QAR model. The QAR model can be
estimated by
X

^y ðtÞ ¼ min rt Y t  y> Xt , (41.8)
y
t

where Xt ¼ (1, Yt1, . . ., Ytp)> and y(t) ¼ (y0(t), y1(t), . . ., yp(t))>; they show that
under regularity assumptions, the limiting distribution of the QAR estimator is
given by

pffiffiffi^ 

n y ðtÞ  yðtÞ ) N 0, tð1  tÞO1 1
1 O0 O1 ,

where O0 ¼ E(XtX> t ) and O1 ¼ lim n

1
∑nt¼1 ft1[F1 >
t1(t)]XtXt .
The QAR models expand the modeling options for time series that display
asymmetric dynamics and allow for local persistency. The models can capture
systematic influences of conditioning variables on the location, scale, and shape
of the conditional distribution of the response and therefore constitute a significant
extension of classical constant coefficient linear time-series models.
Quantile varying coefficients indicate the existence of conditional heteroske-
dasticity. Given the QAR process (Eq. 41.7), let y0 ¼ E[y0(Ut)], y1 ¼ E[y1(Ut)], . . . ,
yp ¼ E[yp(Ut)], and


V t ¼ y0 ðU t Þ  Ey0 ðU t Þ þ ½y1 ðU t Þ  Ey1 ðU t ÞY t1 þ    þ yp ðU t Þ  Eyp ðU t Þ Y tp ;

the QAR process can be rewritten as

Y t ¼ y0 þ y1 Y t1 þ    þ yp Y tp þ V t (41.9)

where Vt is martingale difference sequence. The QAR process is a weak sense AR

process with conditional heteroskedasticity.
What’s the difference between a QAR process and an AR process with ARCH
(or GARCH) errors? In short, the ARCH type model only focuses on the first two
moments, while the QAR model goes beyond the second moment and allows for
more flexible structure in higher moments. Both models allow for conditional
heteroskedasticity and they are similar in the first two moments, but they can be
quite different beyond conditional variance.
41 Quantile Regression and Value at Risk 1151

41.4.2 Nonlinear QAR and Copula-Based Quantile Models

More complicated functional forms with nonlinearity can be considered for the
conditional quantile function if we are interested in the global behavior of the time
series. If the t-th conditional quantile function of Yt is given by

QY t ðtjF t1 Þ ¼ H ðXt ; yðtÞÞ,

where Xt is the vector containing lagged Ys, we may estimate the vector of
parameters y(t) (and thus the conditional quantile of Yt) by the following nonlinear
quantile regression:
X
min rt ðY t  H ðXt , yÞÞ: (41.10)
y
t

Let ett ¼ yt  H ðxt , yðtÞÞ, H_ y ðxt ; yÞ ¼ @Hðxt ; yÞ=@y; we assume that

1X

f t QY t ðtjXt Þ H_ y ðXt , yðtÞÞH_ y ðXt , yðtÞÞ> ! V ðtÞ,
P
V n ð tÞ ¼
n t

1X _
Hy ðXt , yðtÞÞH_ y ðXt , yðtÞÞ> ! OðtÞ,
P
On ðtÞ ¼
n t

and

1 X _
pffiffiffi Hy ðxt , yðtÞÞY t ðett Þ ) N ð0, tð1  tÞOðtÞÞ,
n t

where V(t) and O(t) are non-singular; then under appropriate assumptions, the
nonlinear QAR estimator ^y ðtÞ defined as solution of Eq. 41.10 is root-n consistent
and
pffiffiffi^   
n y ðtÞ  yðtÞ ) N 0, tð1  tÞV ðtÞ1 OðtÞV ðtÞ1 : (41.11)

In practice, one may employ parametric copula models to generate nonlinear-in-

parameters QAR models (see, e.g., Bouyé and Salmon 2008; Chen et al. 2009).
Copula-based Markov models provide a rich source of potential nonlinear
dynamics describing temporal dependence and tail dependence. If we consider,
for example, a first-order strictly stationary Markov process, {Yt}nt¼1, whose
probabilistic properties are determined by the joint distribution of Yt1 and Yt,
say, G*(yt1, yt), and suppose that G*(yt1, yt) has continuous marginal distribution
function F*(·), then by Sklar’s Theorem, there exists a unique copula function
C*(·, ·) such that

G ðyt1 , yt Þ C ðF ðyt1 Þ, F ðyt ÞÞ,

1152 Z. Xiao et al.

where the copula function C*(·, ·) is a bivariate probability distribution function

with uniform marginals. Differentiating C*(u, v) with respect to u and evaluating at
u ¼ F*(x), v ¼ F*(y), we obtain the conditional distribution of Yt given Yt1 ¼ x:

@C ðu; vÞ
Pr½Y t < yjY t1 ¼ x ¼ C1 ðF ðxÞ, F ðyÞÞ:
@u u¼F ðxÞ, v¼F ðyÞ

For any t 2 (0, 1), solving t ¼ Pr[Yt < y|Yt1 ¼ x] C1(F(x), F(y)) for
y (in terms of t), we obtain the t-th conditional quantile function of Yt given
Yt1 ¼ x:


QY t tjxÞ ¼ F1 C1 
1 ðt; F ðxÞÞ ,

where F*1(·) signifies the inverse of F*(·) and C1

1 (;u) is the partial inverse of
C1(u,v) with respect to v ¼ F*(yt).
In practice, neither the true copula function C*(·, ·) nor the true marginal
distribution function F*(·) of {Yt} is known. If we model both parametrically by
C(·, ·; a) and F(y; b), then the t-th conditional quantile function of Yt, QY t ðtjxÞ,
becomes a function of the unknown parameters a and b, i.e.,



QY t tjxÞ ¼ F1 C1
1 ðt; Fðx; bÞ, aÞ, b :

Denoting y ¼ (a0 , b0 )0 and h(x, a, b) C1

1 (t; F(x, b), a), we will write

QY t tjxÞ ¼ F1 ðhðx; a; bÞ, bÞ H ðx; yÞ: (41.12)

For example, if we consider the Clayton copula:

Cðu; v; aÞ ¼ ½ua þ va  11=a , where a > 0,

one can easily verify that the t-th conditional quantile function of Ut given ut1 is
 h  i1=a
QUt tjut1 Þ ¼ ta=ð1þaÞ  1 ua
t1 þ 1 :

See Bouyé and Salmon (2008) for additional examples of copula-based condi-
tional quantile functions.
Although the quantile function specification in the above representation assumes
the parameters to be identical across quantiles, we may permit the estimated
parameters to vary with t, thus extending the original copula-based QAR models
to capture a wide range of systematic influences of conditioning variables on the
conditional distribution of the response. By varying the choice of the copula
specification, we can induce a wide variety of nonlinear QAR(1) dependence, and
41 Quantile Regression and Value at Risk 1153

the choice of the marginal enables us to consider a wide range of possible tail
behavior as well. In many financial time-series applications, the nature of the
temporal dependence varies over the quantiles of the conditional distribution.
Chen et al. (2009) studied asymptotic properties of the copula-based nonlinear
quantile autoregression.

41.4.3 The CaViaR Models

Quantile-based method provides a local approach to directly model the dynamics of

a time series at a specified quantile. Engle and Manganelli (2004) propose the
Conditional Autoregressive Value at Risk (CaViaR) specification for the t-th
conditional quantile of ut:

X
p X
q


Qut ðtjF t1 Þ ¼ b0 þ bi Quti ðtjF ti1 Þ þ aj ‘ Xtj , (41.13)
i¼1 j¼1

where Xtj 2 F tj, F tj is the information set at time tj. A natural choice of Xtj is
the lagged u. When we choose Xtj ¼ jutjj, we obtain GARCH-type CaViaR
models:

X
p X
q  
Qut ðtjF t1 Þ ¼ b0 þ bi Quti ðtjF ti1 Þ þ aj utj :
i¼1 j¼1

If Xtj ¼ 0, we obtain an autoregressive model for the VaRs:

X
p
Qut ðtjF t1 Þ ¼ b0 þ bi Quti ðtjF ti1 Þ:
i¼1

Engle and Manganelli (2004) discussed many choices of ‘(Xtj), leading to

different specifications of the CaViaR model.

41.5 Quantile Regression of Conditional Heteroskedastic

Models

41.5.1 ARCH Quantile Regression Models

ARCH and GARCH models have proven to be highly successful in modeling

financial data. Estimators of volatilities and quantiles based on ARCH and
GARCH models are now widely used in finance applications. Consider the follow-
ing linear ARCH(p) process:
 
ut ¼ st  et , st ¼ g0 þ g1 jut1 j þ    þ gp utp , (41.14)
1154 Z. Xiao et al.

where 0 < g0 < 1, g1, . . ., gp  0 and et are iid(0,1) random variables with pdf f (·)
and CDF F (·). Let Zt ¼ (1,jut1j, . . .,jutqj)> and g(t) ¼ (g0F1(t), g1F1(t), . . .,
gqF1(t))>; the conditional quantiles of ut is given by
 
Qut ðtjF t1 Þ ¼ g0 ðtÞ þ g1 ðtÞjut1 j þ    þ gp ðtÞutp  ¼ gðtÞ> Z t

and can be estimated by the following linear quantile regression of ut on Zt:

X

min rt ut  g> Zt , (41.15)
g
t

where g ¼ (g0, g1,   , gq)>. The asymptotic behavior of the above quantile
regression estimator is given by Koenker and Zhao (1996). In particular,
suppose that ut is given by model (Eq. 41.14), f is bounded and continuous, and
f(F1(t)) > 0 for any 0 < t < 1. In addition, if Ejutj2+d < 1, then the regression
quantile ^g ðtÞ of Eq. 41.15 has the following Bahadur representation:

pffiffiffi S1 1 X n
nð^g ðtÞ  gðtÞÞ ¼
11  pffiffiffi Z> y ðett Þ þ op ð1Þ
f F ð tÞ n t¼1 t t
0
where S1 ¼ EZtZt/st and ett ¼ etF1(t). Consequently,

!
pffiffiffi tð1  tÞ 1 1 0
nð^g ðtÞ  gðtÞÞ ¼ N 0,
1 2 S1 S0 S1 , with S0 ¼ EZ t Z t :
f F ð tÞ

In many applications, conditional heteroskedasticity is modeled on the residuals

of a regression. For example, we may consider the following AR-ARCH model:
0
Y t ¼ a X t þ ut (41.16)

where Xt ¼ (1, Yt1, . . .,Ytp)>, a ¼ (a0, a1, . . ., ap)>, and ut is a linear ARCH(p)
process given by model (Eq. 41.14). The conditional quantiles of Yt is then given by

QY t ðtjF t1 Þ ¼ a Xt þ gðtÞ> Zt :

0
(41.17)

One way to estimate the above model is to construct a joint estimation of

a and g(t) based on nonlinear quantile regression. Alternatively, we may consider
a two-step procedure that estimates a in the first step and then estimates g(t)
based on the estimated residuals. The two-step procedure is usually less efficient
because the preliminary estimation of a may affect the second-step estimation
of g(t), but it is computationally much simpler and is widely used in empirical
applications.
41 Quantile Regression and Value at Risk 1155

41.5.2 GARCH Quantile Regression Models

ARCH models are easier to estimate, but cannot parsimoniously capture the
persistent influence of long past shocks comparing to the GARCH models. How-
ever, quantile regression GARCH models are highly nonlinear and thus compli-
cated to estimate. In particular, the quantile estimation problem in GARCH models
corresponds to a restricted nonlinear quantile regression, and conventional
nonlinear quantile regression techniques are not directly applicable.
Xiao and Koenker (2009) studied quantile regression estimation of the following
linear GARCH(p, q) model:

ut ¼ st  et , (41.18)
 
st ¼ b0 þ b1 st1 þ    þ bp stp þ g1 jut1 j þ    þ gq utq : (41.19)

Let F t1 represents information up to time t1; the t-th conditional quantile of
ut is given by

Qut ðtjF t1 Þ ¼ yðtÞ> Z t , (41.20)

where Zt ¼ (1, st1, . . .,stp,jut1j, . . .,jutqj)> and y(t)> ¼ (b0, b1, . . ., bp,
g1, . . ., gq)F1(t).
Since Zt contains stk(k ¼ 1,  , q) which in turn depends on unknown
parameters y ¼ (b0, b1, . . ., bp, g1, . . ., gq), we may write Zt as Zt(y) to emphasize
the nonlinearity and its dependence on y. If we use the following nonlinear quantile
regression
X

min rt ut  y> Z t ðyÞ , (41.21)
y
t

for a fixed t in isolation, consistent estimate of y cannot be obtained since it ignores

the global dependence of the stk’s on the entire function y(·). If the dependence
structure of ut is characterized by (1) and (1), we can consider the following
restricted quantile regression instead of Eq. 41.21:
( XX

  arg minp, y i t rti ut  p>
p^ , ^y ¼ i Z t ðyÞ
s:t: pi ¼ yðti Þ ¼ yF1 ðti Þ:

Estimation of this global restricted nonlinear quantile regression is complicated.

Xiao and Koenker (2009) propose a simpler two-stage estimator that both
incorporates the global restrictions and also focuses on the local approximation
around the specified quantile. The proposed estimation consists of the following
two steps: (i) The first step considers a global estimation to incorporate the global
dependence of the latent stk’s on y. (ii) Then, using results from the first step, we
1156 Z. Xiao et al.

focus on the specified quantile to find the best local estimate for the conditional
quantile. Let

AðLÞ ¼ 1  b1 L      bp Lp , BðLÞ ¼ g1 þ    þ gq Lq1 ;

under regularity assumptions ensuring that A(L) is invertible, we obtain an

ARCH(1) representation for st:

X
1  
st ¼ a0 þ aj utj : (41.22)
j¼1

For identification, we normalize a0 ¼ 1. Substituting the above ARCH(1)

representation into (1) and (1), we have
!
X
1  
ut ¼ a0 þ aj utj  et , (41.23)
j¼1

and

X
1  
Qut ðtjF t1 Þ ¼ a0 ðtÞ þ aj ðtÞutj ,
j¼1

where aj ðtÞ ¼ aj Qet ðtÞ, j ¼ 0, 1, 2, . . ..

Let m ¼ m(n) be a truncation parameter; we may consider the following
truncated quantile autoregression:

Qut ðtjF t1 Þ

a0 ðtÞ þ a1 ðtÞjut1 j þ    þ am ðtÞjutm j:

By choosing m suitably small relative to the sample size n, but large

enough to avoid serious bias, we obtain a sieve approximation for the GARCH
model.
One could estimate the conditional quantiles simply using a sieve approximation:
∨
Qut (t | Ft−1) = aˆ 0 (t ) + aˆ1 (t ) | ut−1 | + ⋅⋅⋅ + aˆ m (t ) | ut−m |,
where ^
a j ðtÞ are the quantile autoregression estimates. Under regularity assumptions

∨
Qut (t | Ft−1 ) = Qut (t | Ft−1 ) + Op (m / n ).
However, Monte Carlo evidence indicates that the simple sieve approximation
does not directly provide a good estimator for the GARCH model, but it serves as an
adequate preliminary estimator. Since the first step estimation focuses on the global
model, it is desirable to use information over multiple quantiles in estimation.
41 Quantile Regression and Value at Risk 1157

Combining information over multiple quantiles helps us to obtain globally coherent

estimate of the scale parameters.
Suppose that we estimate the m-th order quantile autoregression
!
X
n X
m  
e
a ðtÞ ¼ arg min rt ut  a0  aj utj  (41.24)
a
t¼mþ1 j¼1

at quantiles (t1, . . . , tK) and obtain estimates e

a ðtk Þ, k ¼ 1, . . . , K. Let e
a 0 ¼ 1 in
accordance with the identification assumption. Denote

h i>
a ¼ ½a1 ; . . . ; am ; q1 ; . . . ; qK > , a ðt1 Þ> , . . . , e
p¼ e a ðtK Þ> ,

where qk ¼ Qet ðtk Þ, and

fðaÞ ¼ g  a ¼ ½q1 , a1 q1 , . . . , am q1 , . . . , qK , a1 qK , . . . , am qK > ,

where g ¼ [q1, . . ., qK]> and a ¼ [1, a1, a2, . . ., am]>; we consider the following
estimator for the vector a that combines information over the K quantile estimates
based on the restrictions aj ðtÞ ¼ aj Qet ðtÞ:

a ¼ arg min ðp  fðaÞÞ> An ðp  fðaÞÞ,

e (41.25)
a

where An is a (K(m + 1))  (K(m + 1)) positive definite matrix. Denoting e

a¼
ðe
a0; . . . ; e
a m Þ, st can be estimated by

X
m  
st ¼ e
e a0 þ a j utj :
e
j¼1

In the second step, we perform a quantile regression of ut on

 >
e
Z t ¼ 1, e s tp , jut1 j, . . . , utq  by
s t1 , . . . e
X  
min et ;
rt ut  y> Z (41.26)
y
t

the two-step estimator of y(t)> ¼ (b0(t), b1(t), . . ., bp(t), g1(t), . . ., gq(t)) is then
_
given by the solution of Eq. 41.26, y ðtÞ, and the t-th conditional quantile of ut can
be estimated by

^ u ðtjF t1 Þ ¼ ^y ðtÞ> Z

Q e t:
t

Iteration can be applied to the above procedure for further improvement.

1158 Z. Xiao et al.

Let e
a ðtÞ be the solution of Eq. 41.24; then under appropriate assumptions, we have

a ðtÞ  aðtÞ 2 ¼ Op ðm=nÞ,
ke (41.27)

and for any l 2 Rm+1,

pffiffiffi >
nl ðe a ðtÞ  aðtÞÞ
) N ð0; 1Þ,
sl

where s2l ¼ fe(F1 2 > 1 1

e (t)) l Dn ∑n(t)Dn l, and

" #
1 X n
xt x> 1 X n
Dn ¼ t
, S n ð tÞ ¼ xt xT Y 2 ðutt Þ,
n t¼mþ1 st n t¼mþ1 t t

where xt ¼ (1,jut1j, . . ., jutmj)>.

Define
2 3
 Qet ðt1 Þ
@fðaÞ
G¼ ¼ f_ ða0 Þ ¼ ½g  J m ⋮I K  a0 , g0 ¼ 4    5,
@a> a¼a0 Q ðt Þ et K

where g0 and a0 are the true values of vectors g ¼ [q1, . . ., qK]> and a ¼ [1, a1, a2,
. . ., am]>, and
2 3
0  0
6 1  0 7
Jm ¼ 6
4⋮
7
⋱ ⋮5
0  1

is an (m + 1)  m matrix and IK is a K-dimensional identity matrix; under regularity

assumptions, the minimum distance estimator e a solving (Eq. 41.25) has the
following asymptotic representation:
pffiffiffi
1 pffiffiffi
nð^a  a0 Þ ¼ G> An G G> An nðp  pÞ þ op ð1Þ

where
2 !3
y ðutt Þ
6 D1
n xt
t1 1  7
6 f e F1e ðt1 Þ 7
pffiffiffi 1 X n 6
6
7
7
nðp  pÞ ¼  pffiffiffi  ! 7 þ op ð1Þ,
n t¼mþ1 6
6 7
4 ytm ðuttm Þ 5
D1 x t
1

n
f e F e ð tm Þ
41 Quantile Regression and Value at Risk 1159

and the two-step estimator ^y ðtÞ based on Eq. 41.26 has asymptotic representation:
( )
pffiffiffi^  1 1 1 X pffiffiffi
n y ðtÞ  yðtÞ ¼ 
1  O pffiffiffi Z t yt ðutt Þ þ O1 G nðe a  aÞ þ op ð1Þ,
f e Fe ðtÞ n t

where a ¼ [a1, a2, . . ., am]>, O ¼ E[ZtZt>/st], and

X
p  
Zt
G¼ yk Ck , Ck ¼ E ðjutk1 j; . . . ; jutkm jÞ :
k¼1
st

41.6 An Empirical Application

41.6.1 Data and the Empirical Model

In this section, we apply the quantile regression method to five major world equity
market indexes. The data used in our application are the weekly return series, from
September 1976 to June 2008, of five major world equity market indexes: the US
S&P 500 Composite Index, the Japanese Nikkei 225 Index, the UK FTSE
100 Index, the Hong Kong Hang Seng Index, and the Singapore Strait Times
Index. The FTSE 100 Index data are from January 1984 to June 2008. Table 41.1
reports some summary statistics of the data.
The mean weekly returns of the five indexes are all over 0.1 % per week, with the
Hang Seng Index producing an average return of 0.23 % per week, an astonishing

Table 41.1 Summary statistics of the data

S&P 500 Nikkei 225 FTSE 100 Hang Seng Singapore ST
Mean 0.0015 0.0010 0.0017 0.0023 0.0012
Std. Dev. 0.0199 0.0253 0.0237 0.0376 0.0291
Max 0.1002 0.1205 0.1307 0.1592 0.1987
Min 0.1566 0.1289 0.2489 0.5401 0.4551
Skewness 0.4687 0.2982 1.7105 3.0124 1.5077
Excess kurtosis 3.3494 2.9958 12.867 9.8971 19.3154
AC(1) 0.0703 0.0306 0.0197 0.0891 0.0592
AC(2) 0.0508 0.0665 0.0916 0.0803 0.0081
AC(3) 0.0188 0.0328 0.0490 0.0171 0.0336
AC(4) 0.0039 0.0418 0.0202 0.0122 0.0099
AC(5) 0.0189 0.0053 0.0069 0.0386 0.0519
AC(10) 0.0446 0.0712 0.0138 0.0345 0.0227
This table shows the summary statistics for the weekly returns of five major equity indexes of the
world. AC(k) denotes autocorrelation of order k. The source of the data is the online data service
Datastream
1160 Z. Xiao et al.

increase in the index level over the sample period. In comparison, the average
return of Nikkei 225 index is only 0.1 %. The Hang Seng’s phenomenal rise does
not come without risk. The weekly sample standard deviation of the index is
3.76 %, the highest of the five indexes. In addition, over the sample period the
Hang Seng suffered four larger than 15 % drop in weekly index level, with
maximum loss reaching 35 %, and there were 23 weekly returns below 10 %!
As has been documented extensively in the literature, all five indexes display
negative skewness and excess kurtosis. The excess kurtosis of Singapore Strait
Times Index reached 19.31, to a large extent driven by the huge 1 week loss of
47.47 % during the 1987 market crash. The autocorrelation coefficients for all five
indexes are fairly small. The Hang Seng Index seems to display the strongest
autocorrelation with the AR(1) coefficient equal to 0.0891.
We consider an AR-linear ARCH model in the empirical analysis. Thus, the
return process is modeled as

r t ¼ a0 þ a1 r t1 þ    þ as r ts þ ut , (41.28)

where
 
ut ¼ st et , st ¼ g0 þ g1 jut1 j þ    þ gq utq ,

and the t Conditional VaR of ut is given by

0
Qut ðtjF t1 Þ ¼ gðtÞ Zt


 0
gðtÞ ¼ g0 ðtÞ, g1 ðtÞ, . . . , gq ðtÞ , and Zt ¼ 1; jut1 j; . . . ; utq  :
0

For each time series, we first conduct model specification analysis and choose the
appropriate lags for the mean equation and the quantile ARCH component. Based on
the selected model, we use Eq. 41.28 to obtain a time series of residuals. The residuals
are then used in the ARCH VaR estimation using a quantile regression.

41.6.2 Model Specification Analysis

We conduct sequential tests for the significance of the coefficients on lags. The
inference procedures we use here are asymptotic inferences. For estimation of the
covariance matrix, we use the robust HAC (Heteroskedastic and Autocorrelation
Consistent) covariance matrix estimator of Andrews (1991) with the data-
dependent automatic bandwidth parameter estimator recommended in that paper.
First of all, we choose the lag length in the autoregression,

r t ¼ a0 þ a1 r t1 þ    þ as r ts þ ut ,
41 Quantile Regression and Value at Risk 1161

using a sequential test of significance on lag coefficients. The maximum lag length
that we start with is s ¼ 9, and the procedure is repeated until a rejection occurs.
Table 41.2 reports the sequential testing results for the S&P 500 index. The
t-statistics of all the coefficients are listed for nine rounds of the test. We see that
the t-statistic of the coefficient with the maximum number of lags does not become
significant until s ¼ 1, the ninth round. The preferred model is an AR(1) model. The
selected mean equations for all five indexes are reported in Table 41.4.
Our next task is to select the lag length in the ARCH effect

 
ut ¼ g0 þ g1 jut1 j þ    þ gq utq  et :

Again, a sequential test is conducted. To calculate the t-statistic, we need to

estimate o2 ¼ t(1  t)/f(F1(t))2. There are many studies on estimating f(F1(t)),
including Siddiqui (1960), Bofinger (1975), Sheather and Maritz (1983), and Welsh
(1987). Notice that

dF1 ðtÞ 1
¼
1  ; (41.29)
dt f F ðtÞ

following Siddiqui (1960), we may estimate (Eq. 41.29) by a simple difference

quotient of the empirical quantile function. As a result,
 
f Fd
1 2hn
ðtÞ ¼ (41.30)
^ 1
F ð t þ hn Þ  F^ 1 ðt  hn Þ
1
^ ðtÞ is an estimate of F1(t) and hn is a bandwidth which goes to zero as
where F
n ! 1. A bandwidth choice has been suggested by Hall and Sheather (1988) based
on Edgeworth expansion for studentized quantiles. This bandwidth is of order n1/3
and has the following representation:

h 00
i1=3
hH S ¼ z2=3
a 1:5sðtÞ=s ðtÞ n1=3 ,

where za satisfies F(za) ¼ 1  a/2 for the construction of 1a confidence intervals.
In the absence of additional information, s(t) is just the normal density. Starting
with qmax ¼ 10, a sequential test was conducted and results for the 5 % VaR model
of the S&P 500 Index are reported in Table 41.3. We see that in the fourth round,
the t-statistic on lag 7 becomes significant. The sequential test stops here, and it
suggests that ARCH(7) is appropriate.
Based on the model selection tests, we decide to use the AR(1)-ARCH(7)
regression quantile model to estimate 5 % VaR for the S&P 500 index. We also
conduct similar tests on the 5 % VaR models for other four indexes. To conserve
space we do not report the entire testing process in the paper. Table 41.4 provides
a summary of the selected models based on the tests. The mean equations
 1162

Table 41.2 VaR model mean specification test for the S&P 500 Index
Round 1st 2nd 3rd 4th 5th 6th 7th 8th 9th
a0 3.3460 3.3003 3.2846 3.3248 3.2219 3.7304 3.1723 3.0650 3.8125
a1 1.6941 1.7693 1.8249 1.9987 1.9996 2.0868 2.1536 2.097 2.2094
a2 1.2950 1.3464 1.1555 1.0776 0.0872 1.3106 1.2089 1.0016
a3 0.9235 0.9565 0.9774 1.5521 0.8123 0.8162 0.9553
a4 1.0414 1.0080 0.9947 1.0102 0.9899 0.1612
a5 0.7776 0.7642 0.7865 0.8288 0.7662
a6 0.2094 0.5362 0.7166 0.8931
a7 1.5594 1.5426 1.5233
a8 0.8926 0.8664
a9 0.3816
This table reports the test results for the VaR model mean equation specification for the S&P 500 Index. The number of lags in the AR component of the ARCH
model is selected according to the sequential test. The table reports the t-statistic for the coefficient with the maximum lag in the mean equation
Z. Xiao et al.
41 Quantile Regression and Value at Risk 1163

Table 41.3 5 % VaR model ARCH specification test for the S&P 500 Index
Round 1st 2nd 3rd 4th
g0 16.856 15.263 17.118 15.362
g1 2.9163 3.1891 3.2011 3.1106
g2 1.9601 2.658 2.533 2.321
g3 1.0982 1.0002 0.9951 1.0089
g4 0.6807 0.8954 1.1124 1.5811
g5 0.7456 0.8913 0.9016 0.9156
g6 0.3362 0.3456 0.4520 0.3795
g7 1.9868 2.0197 1.8145 2.1105
g8 0.4866 0.4688 1.5631
g9 1.2045 1.0108
g10 1.1326
This table reports the test results for the 5 % VaR model specification for the S&P 500 Index. The
number of lags in the volatility component of the ARCH model is selected according to the test.
The table reports the t-statistic for the coefficient with the maximum lag in the ARCH equation

Table 41.4 ARCH VaR models selected by the sequential test

Index Mean Lag 5 % ARCH Lag
S&P 500 1 6
Nikkei 225 1 7
FTSE 100 1 6
Hang Seng 3 6
Singapore ST 2 7
This table summarizes the preferred ARCH VaR models for the five global market indexes. The
number of lags in the mean equation and the volatility component of the ARCH model is selected
according to the test

generally have one or two lags, except the Hang Seng Index, which has a lag of
3 and displays more persistent autoregressive effect.
For the ARCH equations, at least six lags are needed for the indexes.

41.6.3 Estimated VaRs

The estimated parameters for the mean equations for all five indexes are reported in
Table 41.5. The constant term for the five indexes is between 0.11 % for the Nikkei
and 0.24 % for the Hang Seng. As suggested by Table 41.1, the Hang Seng seems to
display the strongest autocorrelation, and this is reflected in the four lags chosen by
the sequential test. Table 41.6 reports the estimated quantile regression ARCH
parameters for the 5 % VaR model:
USA – S&P 500 Index. The estimated 5 % VaRs generally range between 2.5 %
and 5 %, but during very volatile periods they could jump over 10 %, as what
happened in October 1987. During high-volatility periods, there is high variation
in estimated VaRs.
1164 Z. Xiao et al.

Table 41.5 Estimated mean equation parameters

Round S&P 500 Nikkei 225 FTSE 100 Hang Seng Singapore ST
a0 0.0019 0.0011 0.0022 0.0024 0.0014
(0.0006) (0.0006) (0.0008) (0.001) (0.0009)
a1 0.0579 0.0827 0.0617 0.1110 0.0555
(0.0233) (0.0305) (0.0283) (0.0275) (0.0225)
a2 0.0796 0.0751
(0.0288) (0.0288)
a3 0.0985
(0.0238)
This table reports the estimated parameters of the mean equation for the five global equity indexes.
The standard errors are in parentheses under the estimated parameters

Table 41.6 Estimated ARCH equation parameters for the 5 % VaR model
Parameter S&P 500 Nikkei 225 FTSE 100 Hang Seng Singapore ST
g0 0.0351 0.0421 0.0346 0.0646 0.0428
(0.0016) (0.0023) (0.0013) (0.0031) (0.0027)
g1 0.2096 0.0651 0.0518 0.1712 0.1119
(0.0711) (0.0416) (0.0645) (0.0803) (0.0502)
g2 0.1007 0.1896 0.0588 0.0922 0.1389
(0.0531) (0.0415) (0.0665) (0.0314) (0.0593)
g3 0.0101 0.1109 0.0311 0.2054 0.0218
(0.0142) (0.0651) (0.0242) (0.0409) (0.0379)
g4 0.1466 0.0528 0.0589 0.0671 0.1102
(0.0908) (0.0375) (0.0776) (0.0321) (0.0903)
g5 0.0105 0.0987 0.0119 0.0229 0.1519
(0.0136) (0.0448) (0.0123) (0.0338) (0.0511)
g6 0.0318 0.0155 0.0876 0.0359 0.0311
(0.0117) (0.0297) (0.0412) (0.0136) (0.0215)
g7 0.2323 0.1123
(0.0451) (0.0517)
This table reports the estimated parameters of the ARCH equation for the 5 % VaR model for the
five global indexes. The standard errors are in parentheses under the estimated parameters

Japan – Nikkei 225 Index. The estimated VaR series is quite stable and remains at
the 4 % and the 7 % level from 1976 till 1982. Then the Nikkei 225 Index took
off and appreciated about 450 % over the next 8 years, reaching its highest level
at the end of 1989. This quick rise in stock value is accompanied by high risk,
manifested here by the more volatile VaR series. In particular, the VaRs fluctu-
ated dramatically, ranging from a low of 3 % to a high of 15 %. This volatility in
VaR may reflect both optimistic market outlook at times and worry about high
valuation and the possibility of a market crash. That crash did come in 1990, and
41 Quantile Regression and Value at Risk 1165

10 years later, the Nikkei 225 Index still hovers around at a level which is about
half off the record high in 1989. The 1990s is far from a rewarding decade for
investors in the Japanese equity market. Average weekly 5 % VaR is about 5 %,
and the variation is also very high.
UK – FTSE 100 Index. The 5 % VaR is very stable and averages about 3 %. They
stay very much within the 2–4 % band, except on a few occasions, such as the
1987 global market crash.
Hong Kong – Hang Seng Index. The Hang Seng Index produces an average return
of 0.23 % per week. The Hang Seng’s phenomenal rise does not come without
risk. We mentioned above that the weekly sample standard deviation of the
index is 3.76 %, the highest of the five indexes. In addition, the Hong Kong stock
market has had more than its fair share of the market crashes.
Singapore – Strait Times Index. Interestingly, the estimated VaRs display a pattern
very similar to that of the UK FTSE 100 Index, although the former is generally
larger than the latter. The higher risk in the Singapore market did not result in
higher return over the sample period. Among the five indexes, the Singapore
market suffered the largest loss during the 1987 crash, a 47.5 % drop in a week.
The market has since recovered much of the loss. Among the five indexes, the
Singapore market only outperformed the Nikkei 225 Index over this period.

41.6.4 Performance of the ARCH Quantile Regression Model

In this section we conduct an empirical analysis to compare VaRs estimated by

RiskMetrics and regression quantiles and those by volatility models with the
conditional normality assumption. There are extensive empirical evidences
supporting the use of the GARCH models in conditional volatility estimation.
Bollerslev et al. (1992) provide a nice overview of the issue. Therefore, we compare
VaR estimated based on RiskMetrics and GARCH(1,1) model and quantile regres-
sion based on ARCH.
To measure the relative performance more accurately, we compute the percent-
age of realized returns that are below the negative estimated VaRs. The results are
reported in Table 41.7. The top panel of the table presents the percentages for the
VaRs estimated by the ARCH quantile regression model, the middle panel for the
VaRs estimated by the GARCH model with the conditional normal return distribu-
tion assumption, and the bottom panel for the VaRs estimated by the RiskMetrics
method. We estimate VaRs using these methods at 1 %, 2 %, 5 %, 10 %. Now we
have a total of four percentage levels. The regression quantile method produces the
closest percentage in general. Both the RiskMetrics method and the GARCH
method seem to underestimate VaRs for the smaller percentages and overestimate
VaRs for the larger percentages.
The five indexes we analyzed are quite different in their risk characteristics as
discussed above. The quantile regression approach seems to be relatively robust and
can consistently produce reasonably good estimates of the VaRs at different
1166 Z. Xiao et al.

Table 41.7 VaR model performance comparison

% VaR 1% 2% 5% 10 %
Quantile regression
S&P 500 1.319 1.925 5.3108 9.656
Nikkei 225 1.350 2.011 5.7210 10.56
FTSE 100 0.714 1.867 5.6019 9.016
Hang Seng 0.799 2.113 4.9011 9.289
GARCH
S&P 500 1.3996 1.7641 4.0114 7.6151
Nikkei 225 1.4974 1.7927 4.3676 8.4098
FTSE 100 1.1980 1.6133 3.3891 6.7717
Hang Seng 1.8962 2.8658 3.6653 7.6439
RiskMetrics
S&P 500 0.3790 0.5199 1.1180 3.2563
Nikkei 225 0.5877 0.9814 1.358 4.1367
FTSE 100 0.2979 0.5796 0.9984 3.5625
Hang Seng 0.7798 0.9822 1.4212 4.1936
This table reports the coverage ratios, i.e., the percentage of realized returns that is below the
estimated VaRs. The top panel reports the performance of the VaRs estimated by the quantile
regression model. The middle panel reports the results for VaRs estimated by the GARCH model
based on the conditionally normal return distribution assumption. The bottom panel reports the
results for VaRs estimated by the RiskMetrics method

percentage (probability) levels. The GARCH model with the normality assumption,
being a good volatility model, is not able to produce good VaR estimates. The
quantile regression model does not assume normality and is well suited to hand
negative skewness and heavy tails.

41.7 Conclusion

Quantile regression provides a convenient and powerful method of estimating

VaR. The quantile regression approach not only provides a method of estimating
the conditional quantiles (VaRs) of existing time-series models; it also substantially
expands the modeling options for time-series analysis. Estimating Value at Risk
using the quantile regression does not assume a particular conditional distribution
for the returns. Numerical evidence indicates that the quantile-based methods have
better performance than the traditional J. P. Morgan’s RiskMetrics method and
other methods based on normality. The quantile regression based method provides
an important tool in risk management.
There are several existing programs for quantile regression applications. For
example, both parametric and nonparametric quantile regression estimations can be
implemented by the function rq() and rqss() in the package quantreg in the
computing language R, and SAS now has a suite of procedures modeled closely
on the functionality of the R package quantreg.
41 Quantile Regression and Value at Risk 1167

References
Andrews, D. W. K. (1991). Heteroskedasticity and autocorrelation consistent covariance matrix
estimation. Econometrica, 59, 817–858.
Barrodale, I., & Roberts, F. D. K. (1974). Solution of an overdetermined system of equations in the
l1 norm. Communications of the ACM, 17, 319–320.
Blanley, A., Lamb, R., & Schroeder, R. (2000). Compliance with SEC disclosure requirements
about market risk. Journal of Derivatives, 7, 39–50.
Bofinger, E. (1975). Estimation of a density function using order statistics. Australian Journal of
Statistics, 17, 1–7.
Bollerslev, T., Chou, R. Y., & Kroner, K. F. (1992). ARCH modeling in finance. Journal of
Econometrics, 52, 5–59.
Bouyé, E., & Salmon, M. (2008). Dynamic copula quantile regressions and tail area dynamic
dependence in forex markets [Manuscript]. Coventry: Financial Econometrics Research Cen-
tre, Warwick Business School.
Chen, X., Koenker, R., & Xiao, Z. (2009). Copula-based nonlinear quantile autoregression. The
Econometrics Journal, 12, 50–67.
Dowd, K. (1998). Beyond value at risk: The new science of risk management. England: Wiley.
Dowd, K. (2000). Assessing VaR accuracy. Derivatives Quarterly, 6, 61–63.
Duffie, D., & Pan, J. (1997). An overview of value at risk. Journal of Derivatives, 4, 7–49.
Engle, R. F., & Manganelli, S. (1999). CAViaR: Conditional autoregressive value at risk by
regression quantiles (Working paper). San Diego: University of California.
Engle, R. F., & Manganelli, S. (2004). CAViaR: Conditional autoregressive value at risk by
regression quantiles (Working paper). San Diego: University of California.
Guo, H., Wu, G., & Xiao, Z. (2007). Estimating value at risk for defaultable bond portfolios by
regression quantile. Journal of Risk Finance, 8(2), 166–185.
Hall, P., & Sheather, S. J. (1988). On the distribution of a studentized quantile. Journal of Royal
Statistical Society B, 50, 381–391.
Koenker, R., & Bassett, G. (1978). Regression quantiles. Econometrica, 84, 33–50.
Koenker, R., & d’Orey, V. (1987). Computing regression quantiles. Applied Statistics, 36,
383–393.
Koenker, R., & Xiao, Z. (2006). Quantile autoregression. Journal of the American Statistical
Association, 101(475), 980–1006.
Koenker, R., & Zhao, Q. (1996). Conditional quantile estimation and inference for ARCH models.
Econometric Theory, 12, 793–813.
Portnoy, S. L., & Koenker, R. (1997). The gaussian hare and the Laplacian tortoise: Computability
of squared-error vs. absolute-error estimator. Statistical Science, 12, 279–300.
Sarma, M., Thomas, S., & Shah, A. (2000). Performance evaluation of alternative VaR models
(Working paper). Indira Gandhi Institute of Development Research.
Saunders, A. (1999). Financial institutions management: A modern perspective, Irwin series in
finance. McGraw-Hill, New York.
Sheather, S. J., & Maritz, J. S. (1983). An estimate of the asymptotic standard error of the sample
median. Australian Journal of Statistics, 25, 109–122.
Siddiqui, M. (1960). Distribution of quantiles in samples from a bivariate population. Journa of
Research Natural Bureau Standards, 64B, 145–150.
Welsh, A. (1987). Asymptotically efficient estimation of the sparsity function at a point. Statistics
and Probability Letters, 6, 427–432.
Wu, G., & Zhijie Xiao (2002). An analysis of risk measures. Journal of Risk, 4(4), 53–75.
Xiao, Z., & Koenker, R. (2009). Conditional quantile estimation and inference for GARCH
models. Journal of the American Statistical Association, 104(488), 1696–1712.
Earnings Quality and Board Structure:
Evidence from South East Asia 42
Kin-Wai Lee

Contents
42.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1170
42.2 Prior Research and Hypotheses Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1174
42.2.1 Equity Ownership of Outside Directors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1174
42.2.2 Board Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1175
42.2.3 Equity-Based Compensation of Outside Directors and
Control Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1176
42.2.4 Board Independence and Control Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1176
42.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1177
42.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1177
42.4.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1177
42.4.2 Discretionary Accruals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1178
42.4.3 Accrual Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1181
42.4.4 Earnings Informativeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1182
42.4.5 Robustness Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1183
42.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1184
Appendix 1: Discretionary Accruals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1184
Appendix 2: Accruals Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1186
Appendix 3: Earnings Informativeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1188
Appendix 4: Adjusting for Standard Errors in Panel Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1189
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1192

Abstract
Using a sample of listed firms in Southeast Asian countries, this paper examines
the association among board structure and corporate ownership structure in
affecting earnings quality. I find that the negative association between separation

K.-W. Lee
Division of Accounting, Nanyang Business School, Nanyang Technological University,
Singapore, Singapore
e-mail: akwlee@ntu.edu.sg

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1169
DOI 10.1007/978-1-4614-7750-1_42,
# Springer Science+Business Media New York 2015
1170 K.-W. Lee

of control rights from cash flow rights and earnings quality varies systematically
with board structure. I find that the negative association between separation of
control rights from cash flow rights and earnings quality is less pronounced in
firms with high equity ownership by outside directors. I also document that in
firms with high separation of control rights from cash flow rights, those firms
with higher proportion of outside directors on the board have higher earnings
quality. Overall, my results suggest that outside directors’ equity ownership and
board independence are associated with better financial reporting outcome,
especially in firms with high expected agency costs arising from misalignment
of control rights and cash flow rights.
The econometric method employed is regressions of panel data. In a panel
data setting, I address both cross-sectional and time-series dependence. Gow
et al. (2010, The Accounting Review 85(2), 483–512) find that in the presence of
both cross-sectional and time-series dependence, the two-way clustering method
which allows for both cross-sectional and time-series dependence produces
well-specified test statistics. Following Gow et al. (2010, The Accounting
Review 85(2), 483–512), I employ the two-way clustering method where the
standard errors are clustered by both firm and year in my regressions of panel
data. Johnston and DiNardo (1997, Econometrics method. New York: Mc-Graw
Hill) and Greene (2000, Econometrics analysis. Upper Saddle River: Prentice-
Hall) are two econometric textbooks that contain a detailed discussion of the
econometrics issues relating to panel data.

Keywords
Earnings quality • Board structure • Corporate ownership structure • Panel data
regressions • Cross-sectional and time-series dependence • Two-way clustering
method of standard errors

42.1 Introduction

In Asia, corporate ownership concentration is high, and many listed firms are
mainly controlled by a single large shareholder (La Porta et al. 1999; Claessens
et al. 2000). Asian firms also show a high divergence between control rights and
cash flow rights, which allows the largest shareholder to control a firm’s opera-
tions with a relatively small direct stake in its cash flow rights. Control is often
increased beyond ownership stakes through pyramid structures, cross-holdings
among firms, and dual class shares (Claessens et al. 2000). It is argued that
concentrated ownership facilitates transactions in weak property rights environ-
ment by providing the controlling shareholders the power and incentive to nego-
tiate and enforce contracts with various stakeholders (Shleifer and Vishny 1997).
As a result of concentrated ownership, the main agency problem in listed firms
in Asia is the conflict of interest between the controlling shareholder and
minority shareholder. Specifically, controlling shareholder has incentives to
42 Earnings Quality and Board Structure: Evidence from South East Asia 1171

expropriate the wealth of minority shareholders by engaging in rent-seeking

activities and to mask their private benefits of control by supplying low-quality
financial accounting information. Empirical evidence also shows that the quality
and credibility of financial accounting information are lower in firms with
high separation of control rights and cash flow rights (Fan and Wong 2002;
Haw et al. 2004).
An important question is how effective are corporate governance mechanisms
in mitigating the agency problems in Asia firms, especially in improving corporate
transparency in firms with concentrated ownership. Controlling shareholders in
Asia typically face limited disciplinary pressures from the market for corporate
control because hostile takeovers are infrequent (La Porta et al. 1999; Fan and
Wong 2002). Furthermore, controlling shareholders face little monitoring pres-
sure from analysts because analysts are less likely to follow firms with potential
incentives to withhold or manipulate information, such as when the family/
management group is the largest control rights blockholder (Lang et al. 2004).
In these environments, external corporate governance mechanisms, in particular
the market for corporate control and analysts’ scrutiny, exert limited disciplinary
pressure on controlling shareholders. Consequently, internal corporate gover-
nance mechanisms such as the board of directors may be important to mitigate
the agency costs associated with the ownership structure of Asian firms. Thus, the
primary research questions in this paper are: (1) Do board of directors play
a corporate governance role over the financial reporting process in listed firms
in Asia? (2) How does the board of director affect financial reporting quality in
firms with high expected agency costs arising from the separation of control rights
and cash flow rights?
Specifically, this paper examines the relation among outside directors’ equity
ownership, board independence, and separation of control rights from cash flow
rights of controlling shareholder in affecting earnings quality. My empirical
strategy is as follows: First, I examine the main effect between earnings quality
and (i) outside directors’ equity ownership, (ii) the proportion of outside directors
on the board, and (iii) the separation of control rights from cash flow rights of
the largest ultimate shareholder. This sheds light on my first research question
on whether the board of directors plays a corporate governance role over the
financial reporting process in listed firms in Asia. Second, I examine the
(i) interaction between outside directors’ equity ownership and the separation of
control rights from cash flow rights and (ii) interaction between the proportion of
outside directors on the board and the separation of control rights from cash flow
rights, in shaping earnings quality. This addresses the second research question on
the effect of board structure (in particular, board independence and equity own-
ership of outside directors) on financial reporting quality in firms with high
expected agency costs arising from the separation of control rights and cash
flow rights.
In this paper, I focus on two important attributes of board monitoring – outside
directors’ equity ownership and board independence – and their association
with financial reporting quality. These attributes are important for two reasons.
1172 K.-W. Lee

First, prior research generally finds that in developed economies such as the United
States and the United Kingdom, there is a positive association between board
independence and earnings quality (Dechow et al. 1996; Klein 2002; Peasnell
et al. 2005). However, there is limited evidence on the effect of board independence
on the financial accounting process in Asia. My paper attempts to fill this gap.
Second, recent research on the monitoring incentives of the board suggests that
equity ownership of outside directors plays an important role in mitigating mana-
gerial entrenchment (Perry 2000; Ryan and Wiggins 2004). An implication of this
stream of research is that even in firms with high board independence, entrenched
managers can weaken the monitoring incentives of outside directors by reducing
their equity-based compensation. In other words, board independence that is not
properly augmented with incentive compensation may hamper the monitoring
effectiveness of independent directors over management.
My sample consists of 2,875 firm-year observations during the period
2004–2008 in five Asian countries comprising Indonesia, Malaysia, the Philippines,
Singapore, and Thailand. These countries provide a good setting to test the
governance potential of the board of directors because shareholders in these
countries typically suffer from misaligned managerial incentives, ineffective legal
protection, and underdeveloped markets for corporate control (La Porta et al. 1999;
Claessens et al. 2000; Fan and Wong 2002).
I measure earnings quality with three financial reporting metrics:
(i) discretionary accruals, (ii) mapping of accruals to cash flow, and (iii) informa-
tiveness of reported earnings. My results are robust across alternative earnings
quality metrics. I find that earnings quality is higher when outside directors have
higher equity ownership. This result suggests that internal monitoring of the quality
and credibility of accounting information is improved through aligning share-
holders’ and directors’ incentives. Consistent with the monitoring role of outside
directors (Fama and Jensen 1983), I also find that earnings quality is positively
associated with the proportion of outside directors on the board. This result supports
the notion that outside directors have incentives to be effective monitors in order to
maintain the value of their reputational capital. Consistent with prior studies
(Fan and Wong 2002; Haw et al. 2004), I also document that earnings quality is
negatively associated with the separation of control rights from cash flow rights of
the largest ultimate shareholder.
More importantly, I document that the negative association between separation
of control rights from cash flow rights and earnings quality is less pronounced in
firms with high equity ownership by outside directors. This result suggests equity
ownership improves the incentives and monitoring intensity of outside directors in
firms with high expected agency costs arising from the divergence of control rights
from cash flow rights. Furthermore, my result indicates the negative association
between separation of control rights from cash flow rights and earnings quality is
mitigated by the higher proportion of outside directors on the board. This result
provides evidence supporting the corporate governance role of outside directors in
42 Earnings Quality and Board Structure: Evidence from South East Asia 1173

constraining managerial discretion over financial accounting process in firms with

high levels of misalignment between control rights and cash flow rights. Collec-
tively, my results suggest that strong internal governance structures can alleviate
agency problems between the controlling shareholder and minority shareholders.
More generally, my results highlight the interplay between board structure and
corporate ownership structure in shaping earnings quality.
I perform several robustness tests. My results are robust across different
economies. In addition, year-by-year regressions yield qualitatively similar results,
suggesting my inferences are not time-period specific. I also include additional
country-level institutional variables such as legal origin, country investor protec-
tion, and enforcement of shareholder rights. My results are qualitatively similar.
Specifically, after controlling for country-level legal institutions, firm-specific
internal governance mechanisms, namely – outside directors’ equity ownership
and board independence – continue to be important in mitigating the negative
effects of the divergence between control rights and cash flow rights on earnings
quality.
My study has several contributions. First, prior studies find that the divergence of
control rights from cash flow rights reduces the informativeness of reported earn-
ings (Fan and Wong 2002) and induces earnings management (Haw et al. 2004).
I extend these studies by demonstrating two specific channels at the firm
level – equity ownership by outside directors and proportion of outside directors
on the board – that mitigate the negative association between earnings quality and
divergence of control rights from cash flow rights. This result suggests that the
board of directors play an important corporate governance role to alleviate agency
problems in firms with entrenched insiders. My findings also complement Fan and
Wong’s (2005) result that given concentrated ownership, a controlling owner may
introduce some monitoring or bonding mechanisms that limit his ability to expro-
priate minority shareholders and hence mitigate agency conflicts. In the Fan and
Wong’s study, high-quality external auditors alleviate agency problems in firms
with concentrated ownership, whereas in my study, strong board of directors
augmented with proper monitoring incentives mitigate agency problems in firms
with concentrated ownership.
Second, my results suggest that there is an incremental role for firm-specific
internal governance mechanisms, beyond country-level institutions, in improving
the quality of financial information. Haw et al. (2004) find that earnings manage-
ment that is induced by the divergence between control rights and cash flow rights is
less pronounced in countries where (i) legal institutions protect minority share-
holder rights (such as legal tradition, minority shareholder rights, efficiency of
judicial system, or disclosure system) and (ii) in countries with effective extralegal
institutions (such as the effectiveness of competition law, diffusion of the press, and
tax compliance). My study shows that after controlling for both country-level legal
and extralegal institutions, firm-specific internal governance mechanisms, namely,
outside directors’ incentive compensation and board independence, continue to be
1174 K.-W. Lee

important in constraining management opportunism over the financial reporting

process in firms with high expected agency costs arising from the divergence
between control rights and cash flow rights. To the extent that changes in
country-level legal institutions are relatively more costly and more difficult than
changes in firm-level governance mechanisms, my result suggests that improve-
ment in firm-specific governance mechanisms can be effective to reduce private
benefits of control. My results complement finding in prior studies (Johnson
et al. 2000; La Porta et al. 1998; Lang et al. 2004) that firms in countries with
weak legal protection substitute with strong firm-level internal governance mech-
anisms to attract investors. My results also extend the finding in Leuz et al. (2003)
that firms located in countries with weaker investor protection have higher
earnings management. An important question is what factors may constrain
managerial opportunism when country-level investor protection is weak?
Because my sample consists of countries with generally weak investor protection,
I shed light on this question by documenting that firm-level governance
structures matter in improving earnings quality in countries with weak investor
protection.
The rest of the paper proceeds as follows. Section 42.2 develops the
hypotheses and places my paper in the context of related research. Section 42.3
describes the sample and method. Section 42.4 presents my results. I conclude the
paper in Sect. 42.5.

42.2 Prior Research and Hypotheses Development

42.2.1 Equity Ownership of Outside Directors

Recent research examines the compensation structure of outside directors, who play
an important monitoring role over management’s actions. The central theme in this
body of research is that incentive compensation leading to share ownership
improves the outside directors’ incentives to monitor. Mehran (1995) finds firm
performance is positively associated with the proportion of directors’ equity-based
compensation. Perry (2000) finds that the likelihood of CEO turnover following
poor performance increases when directors receive higher equity-based compensa-
tion. Shivdasani (1993) finds that probability of a hostile takeover is negatively
associated with the percentage of shares owned by outside directors in target firms.
He interprets this finding as suggesting that board monitoring may substitute for
monitoring from the market of corporate control. Hermalin and Weisbach (1988)
and Gillette et al. (2003) develop models where incentive compensation for
directors increases their monitoring efforts and effectiveness. Ryan and Wiggins
(2004) find that directors in firms with entrenched CEOs receive a significantly
smaller proportion of compensation in the form of equity-based awards. Their
result suggests that entrenched CEOs use their position to influence directors’
42 Earnings Quality and Board Structure: Evidence from South East Asia 1175

compensation, which results in contracts that provide directors with weaker

incentives to monitor management.
Internal monitoring is improved through aligning shareholders’ and directors’
incentives. If higher equity-based compensation contracts provide outside directors
with stronger incentives to act in the interests of shareholders, I predict that
managerial opportunism over the financial reporting process is reduced when
outside directors have higher equity-based compensation. My first hypothesis is:
H1 Earnings quality is positively associated with the outside directors’ equity
ownership.

42.2.2 Board Independence

There is considerable literature on the role of outside directors in reducing agency

problems between managers and shareholders. Fama and Jensen (1983) argue that
outside directors have strong incentives to be effective monitors in order to main-
tain their reputational capital. Prior studies support the notion that board effective-
ness in protecting shareholders’ wealth is positively associated with the proportion
of outside directors on the board (Weisbach 1988; Rosenstein and Wyatt 1990).
In the United States, Klein (2002) finds that firms with high proportion of outside
directors on the board have lower discretionary accruals. Using a sample of
listed firms in the United Kingdom, Peasnell et al. (2005) document that the
greater the board independence, the lower the propensity of managers making
income-increasing discretionary accruals to avoid reporting losses and earnings
reductions. Using US firms subjected to SEC enforcement action for alleged
earnings manipulation, Dechow et al. (1996) and Beasley (1996) find that the
probability of financial reporting fraud is negatively associated with the proportion
of outside directors on the board.
In contrast, in emerging markets, conventional wisdom suggests that the agency
conflicts between controlling owners and the minority shareholders may be difficult to
mitigate through conventional corporate control mechanisms such as boards of direc-
tors (La Porta et al. 1998; Claessens et al. 2000; Fan and Wong 2005; Lee 2007; Lee
et al. 2009). However, since the Asian economic crisis in 1997, many countries in Asia
took steps to improve their corporate governance environment such as implementing
country-specific code of corporate governance. For example, the Stock Exchange of
Thailand Code of Best Practice for Directors of Listed Companies was implemented in
1998, the Code of Proper Practices for Directors for the Philippines was implemented
in 2000, the Malaysian Code of Corporate Governance was implemented in 2000, and
the Singapore Code of Corporate Governance was implemented in 2001. Among the
key provisions of the code of corporate governance in these countries is the recom-
mendation to have sufficient independent directors on the board to improve monitoring
of management. For example, the 2001 Code of Corporate Governance for Singapore
stated that there should be a strong and independent element on the board, which is able
to exercise objective judgment on corporate affairs independently from management.
1176 K.-W. Lee

Although compliance with the code of corporate governance is not legally mandatory,
listed companies are required to explain deviations from the recommendations of the
code of corporate governance.1
I posit that in the waves of corporate governance reform in emerging markets in
the early 2000s and the guidelines of country-specific code of corporate governance
in emphasizing the importance of board independence, there is heightened aware-
ness among outside directors on their increased monitoring responsibilities. To the
extent that outside directors in listed firms in Asia perform a corporate governance
role, I predict that:
H2 Earnings quality is positively associated with the proportion of outside direc-
tors on the board.

42.2.3 Equity-Based Compensation of Outside Directors and

Control Divergence

The preceding discussion suggests that higher equity-based incentive compensation

for outside directors improves their monitoring efforts. Greater monitoring from
outside directors reduces managerial discretion over the financial reporting process
(Lee et al. 2008). The benefits of more effective monitoring arising from
higher equity ownership are likely to be concentrated in firms with high agency
problems arising from the separation of control rights from cash flow rights. Thus,
I predict that:
H3 The negative association between separation of control rights from cash flow
rights and earnings quality is less pronounced in firms with high equity ownership
by outside directors.

42.2.4 Board Independence and Control Divergence

Outside directors play an important corporate governance role in resolving

agency problems between managers and shareholders. Following prior studies
(Beasley 1996; Dechow et al. 1996; Klein 2002; Lee et al. 2007), higher proportion
of outside directors on the board is associated with higher constraints on manage-
ment discretion over the financial reporting process. I extend this notion to posit that
the greater monitoring efforts from a high proportion of outside directors on the

1
To illustrate, the Singapore Exchange Listing Rules require “listed companies to describe in the
annual reports their corporate governance practices with specific reference to the principles of the
Code, as well as disclose and explain any deviation from any guideline of the Code. Companies are
also encouraged to make a positive confirmation at the start of the corporate governance section of
the annual report that they have adhered to the principles and guidelines of the Code, or specify
each area of non-compliance. Many of these guidelines are recommendations for companies to
disclose their corporate governance arrangements.”
42 Earnings Quality and Board Structure: Evidence from South East Asia 1177

board are likely to mitigate the negative effects of the separation of control rights
from cash flow rights on earnings quality. Thus, I predict that:
H4 The negative association between separation of control rights from cash flow
rights and earnings quality is mitigated by the proportion of outside directors.

42.3 Data

I begin with the Worldscope database to identify listed firms in five Asian
countries comprising Indonesia, Malaysia, the Philippines, Singapore, and Thai-
land during the period 2004–2008. I exclude financial institutions because of their
unique financial structure and regulatory requirements. I eliminate observations
with extreme values of control variables such as return-on-assets and leverage.
I obtain stock price data from the Datastream database. I obtain annual reports for
the period 2004–2008 from the Global Report database and company websites.
The sample consists of 617 firms for 2,875 firm-year observations during the
period 2004–2008 in five Asian countries comprising Indonesia, Malaysia, the
Philippines, Singapore, and Thailand.
I collect data on the board characteristics such as board size, the number of
independent directors, and equity ownership of directors from the annual report.
I also examine the annual report to trace the ultimate owners of the firms.
The procedure of identifying ultimate owners is similar to the one used in La
Porta et al. (1999).2 In this study, I measure earnings quality with three financial
reporting metrics: (i) discretionary accruals, (ii) mapping of accruals to cash flow,
and (iii) informativeness of reported earnings.
Appendix 1 contains detailed description on the econometric method.

42.4 Results

42.4.1 Descriptive Statistics

Table 42.1 presents the descriptive statistics. Mean absolute discretionary accrual as
a proportion of lagged assets is 0.062. Mean equity ownership of outside directors
(computed as common stock and stock options held by outside directors divided by

2
In summary, an ultimate owner is defined as the shareholder who has the determining voting
rights of the company and who is not controlled by anyone else. If a company does not have an
ultimate owner, it is classified as widely held. To economize on the data collection task, the
ultimate owner’s voting right level is set at 50 % and not traced any further once that level exceeds
50 %. Although a company can have more than one ultimate owner, we focus on the largest
ultimate owner. We also identify the cash flow rights of the ultimate owners. To facilitate the
measurement of the separation of cash flow and voting rights, the maximum cash flow rights level
associated with any ultimate owner is also set at 50 %. However, there is no minimum cutoff level
for cash flow rights.
1178 K.-W. Lee

Table 42.1 Descriptive statistics. The sample consists of 617 firms for 2,875 firm-year
observations during the period 2004–2008 in five Asian countries comprising Indonesia, Malaysia,
the Philippines, Singapore, and Thailand
Mean 25th percentile Median 75th percentile Standard deviation
DISCAC 0.062 0.009 0.038 0.085 0.053
AQ 0.068 0.029 0.035 0.063 0.037
EBC (%) 2.041 0.837 1.752 2.663 1.035
OUTDIR 0.489 0.206 0.385 0.520 0.217
VOTE 1.198 1.000 1.175 1.326 0.638
BOARDSIZE 7 5 8 10 3
CEODUAL 0.405 0 0 1 –
LNASSET 11.722 9.867 11.993 13.078 2.115
MB 1.851 0.582 1.272 2.195 0.833
LEV 0.261 0.093 0.211 0.335 0.106
ROA 0.086 0.027 0.0705 0.109 0.071
DISCAC ¼ absolute value of discretionary accruals estimated based on the modified Jones model
AQ ¼ accrual quality measured by Dechow and Dichev’s (2002) measure of mapping of accruals
to past, present, and future cash from operations
DIROWN ¼ common stock and stock options held by outside directors divided by number of
ordinary shares outstanding in the firm
OUTDIR ¼ proportion of outside directors on the board
VOTECASH ¼ voting rights divided by cash flow rights of the largest controlling shareholder
CEODUAL ¼ a dummy variable that equals 1 if the CEO is chairman of board and 0 otherwise
BOARDSIZE ¼ number of directors on the board
LNASSET ¼ natural logarithm of total assets
MB ¼ market value of equity divided by book value of equity
LEV ¼ long-term debt divided by total assets
ROA ¼ net profit after tax divided by total assets

number of ordinary shares outstanding in the firm) is 2.04 %. The mean board size
and proportion of outside directors on the board are 7 and 0.489, respectively. CEO
chairs the board in 40 % of the firms. Consistent with Fan and Wong’s (2002) study
of East Asian economies, the firms in my sample also have high divergence of
control rights from cash flow rights (mean VOTE ¼ 1.198).

42.4.2 Discretionary Accruals

Table 42.2 presents the estimates of regressions of unsigned discretionary accruals

on equity-based compensation, proportion of outside directors on the board, and the
separation of control right from cash flow right. Following Gow et al. (2010),
I employ the two-way clustering method where the standard errors are clustered
by both firm and year in my regressions. In column (1), I document a negative
association between the absolute value of discretionary accruals and the equity
ownership of outside directors. Results also indicate that firms with higher propor-
tion of outside directors on the board have lower discretionary accruals. I find that
42 Earnings Quality and Board Structure: Evidence from South East Asia 1179

Table 42.2 Regressions of unsigned discretionary accruals. The sample consists of 617 firms
for 2,875 firm-year observations during the period 2004–2008 in five Asian countries comprising
Indonesia, Malaysia, the Philippines, Singapore, and Thailand. The dependent variable is absolute
discretionary accruals computed based on the modified Jones model. All variables are defined in
Table 42.1. The t-statistics (in parentheses) are adjusted based on standard errors clustered by firm
and year (Petersen 2009). The symbols *, **, and *** denote statistical significance at the 10 %,
5 %, and 1 % levels (two-tailed), respectively
Predicted sign 1 2
EBC  0.2513 (3.07)*** 0.2142 (2.86)***
OUTDIR  0.1862 (2.41)** 0.1053 (2.19)**
VOTE + 0.8173 (2.85)*** 0.9254 (2.94)***
VOTE *EBC  0.4160 (2.73)***
VOTE * OUTDIR  0.1732 (2.29)**
BOARDSIZE +/ 0.0359 (1.57) 0.0817 (1.42)
CEODUAL + 0.4192 (1.61) 0.2069 (1.45)
LNASSET  0.5311 (8.93)*** 0.5028 (8.01)***
MB + 0.1052 (4.25)*** 0.2103 (3.02)***
LEV + 1.2186 (5.38)*** 1.1185 (5.19)***
ROA +/ 3.877 (4.83)*** 4.108 (4.72)***
Adjusted R2 12.5 % 14.1 %

earnings management (as proxied by absolute discretionary accruals) increases as

the separation between control rights and cash flow rights of controlling share-
holders increases. This result is consistent with the finding in Haw et al. (2004).
In column (2), I test whether the positive association between discretionary accruals
and the separation between control rights and cash flow rights of controlling
shareholder is mitigated by the equity ownership of outside directors. The coeffi-
cient on the interaction term between the separation of control rights from cash flow
rights and the equity ownership of outside directors (VOTE* DIROWN) is negative
and significant at the 1 % level, supporting the hypothesis that in firms with high
separation of control rights from cash flow rights, earnings management is reduced
when outside directors have higher equity ownership. This finding suggests that
greater equity-based compensation increases the monitoring effectiveness of out-
side directors over the financial reporting process in firms with agency conflicts
arising from their control rights from cash flow rights. In addition, the coefficient on
the interaction term between the separation of control rights from cash flow rights
and the proportion of outside directors (VOTE*OUTDIR) is negative and signifi-
cant at the 5 % level, supporting the hypothesis that in firms with high separation of
control rights from cash flow rights, earnings management is reduced in firms with
high proportion of outside directors on the board. This result is consistent with the
monitoring role of independent directors to improve the credibility of accounting
information in firms with agency problems arising from their concentrated corpo-
rate ownership structure.
1180 K.-W. Lee

Table 42.3 Regressions of signed discretionary accruals. The sample consists of 617 firms for
2,875 firm-year observations during the period 2004–2008 in five Asian countries comprising
Indonesia, Malaysia, the Philippines, Singapore, and Thailand. In column (1), the sample consists
of firms with income-increasing discretionary accruals, and the dependent variable is positive
discretionary accruals. In column (2), the sample consists of firms with income-decreasing
discretionary accruals, and the dependent variable is negative discretionary accruals. All variables
are defined in Table 42.1. All regressions contain dummy control variables for country, year, and
industry. The t-statistics (in parentheses) are adjusted based on standard errors clustered by firm
and year (Petersen 2009). The symbols *, **, and *** denote statistical significance at the 10 %,
5 %, and 1 % levels (two-tailed), respectively

(1) (2)
Positive DISCAC Negative DISCAC
EBC 0.1865 (2.21)** 0.2017 (2.09)**
OUTDIR 0.1172 (2.08)** 0.1302 (2.11)**
VOTE 0.8103 (3.25)*** 0.6735 (2.23)**
VOTE *EBC 0.3952 (2.49)*** 0.3064 (2.05)**
VOTE * OUTDIR 0.1732 (2.13)** 0.1105 (2.01)**
BOARDSIZE 0.0533 (1.27) 0.0681 (1.09)
CEODUAL 0.1860 (1.32) 0.1562 (1.26)
LNASSET 0.7590 (5.22)*** 0.4463 (6.12)***
MB 0.2019 (2.08)** 0.1085 (1.93)**
LEV 0.9781 (3.20)*** 0.7701 (2.10)**
ROA 3.087 (4.13)*** 4.253 (3.62)***
Adjusted R2 13.8 % 12.4 %

I partition the sample into two groups based on the sign of the firms’ discretion-
ary accruals. Table 42.3 column (1) presents the results using the subsample of
firms with income-increasing discretionary accruals. Results indicate firms with
higher equity ownership by outside directors, higher board independence, and
lower divergence of control rights from cash flow rights, have lower income-
increasing discretionary accruals. More importantly, I find that outside directors’
equity ownership and proportion of outside directors mitigate the propensity of
firms with high separation of control rights from cash flow rights to make higher
income-increasing discretionary accruals. Table 42.3 column (2) presents the
results using the subsample of firms with income-decreasing discretionary accruals.
I find firms with higher equity ownership, higher board independence, and lower
divergence of control rights from cash flow rights, have lower income-decreasing
discretionary accruals. Furthermore, I find that equity ownership by outside direc-
tors and proportion of outside directors mitigate the propensity of firms with high
separation of control rights from cash flow rights to make higher income-decreasing
discretionary accruals.
In summary, when outside directors have equity ownership and when board
independence is high, firms have both lower income-increasing and income-
decreasing discretionary accruals, apparently mitigating earnings management
42 Earnings Quality and Board Structure: Evidence from South East Asia 1181

Table 42.4 Regressions of accrual quality. The sample consists of 617 firms for 2,875 firm-
year observations during the period 2004–2008 in five Asian countries comprising Indonesia,
Malaysia, the Philippines, Singapore, and Thailand. The dependent variable is AQ measured by
Dechow and Dichev’s (2002) measure of mapping of accruals to past, present, and future cash
from operations with higher values of AQ denoting better accrual quality. All variables are defined
in Table 42.1. All regressions contain dummy control variables for country, year, and industry. The
t-statistics (in parentheses) are adjusted based on standard errors clustered by firm and year
(Petersen 2009). The symbols *, **, and *** denote statistical significance at the 10 %, 5 %,
and 1 % levels (two-tailed), respectively
Predicted sign 1 2
EBC + 0.3725 (3.11)*** 0.2133 (3.26)***
OUTDIR + 0.2049 (2.15)** 0.1557 (2.86)***
VOTE  0.5108 (3.74)*** 0.4352 (3.05)***
VOTE *EBC + 0.1751 (2.80)***
VOTE * OUTDIR + 0.1163 (2.09)**
BOARDSIZE +/ 0.1003 (1.29) 0.0642 (1.17)
CEODUAL +/ 0.1890 (0.83) 0.2173 (1.56)
LNASSET + 3.2513 (5.11)*** 2.8764 (4.82)***
OPERCYCLE  3.2941 (4.75)*** 3.0185 (5.01)***
NETPPE + 1.1802 (3.35)*** 0.9926 (3.72)***
STDCFO  1.2981 (2.83)*** 1.8344 (3.32)***
STDSALE  1.0203 (2.77)*** 0.7845 (2.09)**
NEGEARN  0.8306 (2.02)** 1.0345 (1.77)*
Adjusted R2 9.2 % 10.5 %

both on the upside and downside. For firms with greater separation of control
rights from cash flow rights of controlling shareholders, those with high
equity ownership by outside directors and those with high proportion of outside
directors have lower income-increasing and lower income-decreasing discretionary
accruals.

42.4.3 Accrual Quality

Table 42.4 presents regressions of accrual quality on corporate ownership structure

and board characteristics. Following Gow et al. (2010), I employ the two-way
clustering method where the standard errors are clustered by both firm and year
in my regressions. In column (1), the coefficient EBC is positive and significant,
suggesting that that firms whose directors receive higher equity ownership have
higher accrual quality. Firms with high proportion of outside directors have higher
accrual quality. The coefficient on VOTE is negative and significant, indicating the
firms with high misalignment between control rights and cash flow rights have
lower accrual quality. In column (2), the interaction term VOTE*DIROWN is
positive and significant at the 1 % level. This finding suggests that firms with
1182 K.-W. Lee

Table 42.5 Regressions of returns on earnings. The sample consists of 617 firms for 2,875
firm-year observations during the period 2004–2008 in five Asian countries comprising Indonesia,
Malaysia, the Philippines, Singapore, and Thailand. The dependent variable (RET) is 12-month
cumulative raw return ending 3 months after the fiscal year-end. All regressions contain dummy
control variables for country, year, and industry. The t-statistics (in parentheses) are adjusted based
on standard errors clustered by firm and year (Petersen 2009). The symbols *, **, and *** denote
statistical significance at the 10 %, 5 %, and 1 % levels (two-tailed), respectively
Predicted sign 1 2
EARN + 1.1735 (3.85)*** 1.2811 (3.62)***
EARN *EBC + 0.3122 (2.87)*** 0.2983 (2.80)***
EARN * OUTDIR + 0.1094 (2.13)** 0.1105 (2.08)**
EARN * VOTE  0.6817 (3.72)*** 0.7019 (3.50)***
EARN * VOTE *EBC + 0.2602 (2.83)***
EARN * VOTE * OUTDIR + 0.1925 (2.15)**
EARN * BOARDSIZE +/ 0.0836 (1.50) 0.0801 (1.22)
EARN * CEODUAL +/ 0.0405 (1.42) 0.0215 (1.30)
EARN * LNASSET + 0.2011 (2.89)*** 0.3122 (3.07)***
EARN * MB + 0.1573 (1.80)* 0.1806 (1.81)*
EARN * LEV  0.7814 (2.03)** 0.6175 (2.84)***
N 3,172 3,172
Adjusted R2 11.8 % 13.3 %

higher equity ownership by outside directors have a less pronounced negative

association between accrual quality and the separation of control rights from cash
flow rights of controlling shareholders. I then test whether board independence
attenuates the negative association between accrual quality and the separation of
control rights from cash flow rights of controlling shareholders. The interaction
term VOTE*OUTDIR is positive and significant at the 5 % level. Hence, for firms
with high separation of control rights from cash flow rights of controlling share-
holder, those with higher proportion of outside directors have higher accrual
quality. Collectively, my results suggest that stronger directors’ equity ownership
and higher board independence are associated with better financial reporting out-
come, especially in firms with high expected agency costs arising from
misalignment of control rights and cash flow rights.

42.4.4 Earnings Informativeness

Table 42.5 presents the regression results on earnings informativeness. The coeffi-
cient EARN* DIROWN is positive and significant, indicating the greater the equity
ownership by outside directors, the higher informativeness of reported earnings.
The coefficient EARN*OUTDIR is positive and significant, implying that firms
42 Earnings Quality and Board Structure: Evidence from South East Asia 1183

with higher proportion of outside directors have higher informativeness of reported

earnings. Consistent with prior studies (Fan and Wong 2002), the coefficient
EARN*VOTE is negative and significant, indicating that the separation of control
rights from cash flow rights of controlling shareholder reduces the informativeness
of reported earnings.
In column (2), I examine the interaction between effectiveness of board
monitoring and the divergence of control rights from cash flow rights in affecting
earnings informativeness. The interaction term EARN*VOTE* DIROWN is
positive and significant at the 1 % level. In firms with high misalignment between
control rights from cash flow rights, the informativeness of earnings is higher when
outside directors have higher equity ownership. The interaction term
EARN*VOTE*OUTDIR is positive and significant at the 5 % level. The negative
association between earnings informativeness and the separation between
control rights from cash flow rights controlling shareholder is less pronounced
in firms with higher proportion of outside directors. In other words, in firms
with high misalignment between control rights from cash flow rights, the
informativeness of earnings is higher in firms with higher proportion of outside
directors.

42.4.5 Robustness Tests

As a sensitivity analysis, I repeat all my tests at the economy level. The economy-
by-economy results indicate that earnings quality is positively associated with
equity ownership by outside directors and board independence and negatively
associated with the separation of cash flow rights from control rights. More
importantly, the mitigating effects of equity ownership and board independence
on the association between separation of cash flow rights from control rights and
earnings quality are not concentrated in any given economy. Year-by-year regres-
sions yield qualitatively similar results, suggesting my inferences are not time-
period specific.
As a robustness test, I follow Haw et al. (2004) to include legal institutions
that protect minority shareholder rights (proxied by legal tradition, minority
shareholder rights, efficiency of judicial system, or disclosure system) and extrale-
gal institutions (proxied by the effectiveness of competition law, diffusion of the
press, and tax compliance) in my tests. I continue to document firm-specific internal
governance mechanisms, namely, outside directors’ equity ownership and board
independence, still matter in constraining management opportunism over the finan-
cial reporting process, especially in firms with high expected agency costs arising
from the divergence between control rights and cash flow rights. Thus, my results
suggest that there is an incremental role for firm-specific internal governance
mechanisms, beyond country-level institutions, in improving the quality of finan-
cial information by mitigating insiders’ entrenchment.
1184 K.-W. Lee

42.5 Conclusion

Publicly reported accounting information, which measures a firm’s financial

position and performance, can be used as important input information in
various corporate governance mechanisms such as managerial incentive plans.
Whether and how reported accounting information is used in the governance
of a firm depends on the quality and credibility of such information. I provide
evidence that board of directors plays an important corporate governance role
in improving the quality and credibility of accounting information in firms
with high agency conflicts arising from their concentrated ownership structure.
I examine the relation among outside directors’ equity ownership, board
independence, separation of control rights from cash flow rights of controlling
shareholder, and earnings quality. I measure earnings quality with three financial
reporting metrics: (i) discretionary accruals, (ii) mapping of accruals to cash flow,
and (iii) informativeness of reported earnings. I find that earnings quality is
positively associated with outside directors’ equity ownership and the proportion
of outside directors on the board. I document that firms with higher agency
problems arising from the separation of control rights from cash flow rights of
controlling shareholders have lower earnings quality. The negative association
between separation of control rights from cash flow rights and earnings quality is
less pronounced in firms with higher equity ownership by outside directors. This
finding suggests that equity ownership that aligns outside directors’ and share-
holders’ interest is associated with more effective monitoring of managerial
discretion on reported earnings. In addition, the low earnings quality induced by
the separation of control rights from cash flow rights is mitigated by the
proportion of outside directors on the board. Overall, my results suggest that
directors’ equity ownership and board independence are associated with better
financial reporting outcomes, especially in firms with high expected agency costs
arising from misalignment of control rights and cash flow rights.

Appendix 1: Discretionary Accruals

My first proxy for earnings quality is discretionary accruals. A substantial stream

of prior studies uses absolute discretionary accruals as a proxy for earnings
management (Ashbaugh et al. 2003; Warfield et al. 1995; Klein 2002; Kothari
et al. 2005). Absolute discretionary accruals reflect corporate insiders’ propensity to
inflate reported income to conceal private benefits of control and to understate
income in good performance years to create reserves for poor performance in the
future. Accruals are estimated by taking the difference between net income and
cash flow from operations. I employ the modified cross-sectional Jones
(1991) model to decompose total accruals into non-discretionary accruals and
42 Earnings Quality and Board Structure: Evidence from South East Asia 1185

discretionary accruals. Specifically, I estimate the following model for each country
in each year at the one-digit SIC industry:

ACC ¼ g1 ð1=LAG1ASSETÞ þ g2 ðCHGSALE  CHGRECÞ þ g3 ðPPEÞ (42.1)

where:
ACC ¼ total accruals, which are calculated as net income minus operating cash
flows scaled by beginning-of-year total assets.
LAG1ASSET ¼ total assets at beginning of the fiscal year.
CHGSALE ¼ sales change, which is net sales in year t less net sales in year
t  1, scaled by beginning-of-year-t total assets.
CHGREC ¼ change in accounts receivables scaled by beginning-of-year-t total
assets.
PPE ¼ gross property, plant, and equipment in year t scaled by beginning-of-year-t
total assets.
I use the residuals from the annual cross-sectional country-industry regression
model in (A1) as the modified Jones model discretionary accruals.
I use the following regression model to test the association between discretion-
ary accruals and board structure:

DISCAC ¼ b0 þ b1 EBC þ b2 OUTDIR þ b3 VOTE þ b4 VOTE  EBC

þ b5 VOTE  OUTDIR þ b6 BOARDSIZE þ b7 CEODUAL
þ b8 LNASSET þ b9 MB þ b10 LEV þ b11 ROA
þ Year controls þ Country Controls (42.2)

where:
DISCAC ¼ absolute value of discretionary accruals estimated based on the
modified Jones model (see Eq. 42.1).
DIROWN ¼ common stock and stock options held by outside directors divided by
number of ordinary shares outstanding in the firm.
OUTDIR ¼ proportion of outside directors on the board.
VOTE ¼ control rights divided by cash flow rights of the largest controlling
shareholder.
BOARDSIZE ¼ number of directors on the board.
CEODUAL ¼ a dummy variable that equals 1 if the CEO is chairman of board and
0 otherwise.
LNASSET ¼ natural logarithm of total assets.
MB ¼ market value of equity divided by book value of equity.
LEV ¼ long-term debt divided by total assets.
ROA ¼ net profit after tax divided by total assets.
Country Controls ¼ a set of country dummy variables.
Year Controls ¼ a set of year dummy variables.
1186 K.-W. Lee

If high equity ownership for outside directors improves the board monitoring of
managerial discretion over the financial accounting process, I predict the coefficient
b1 to be negative. Similarly, a negative coefficient for b2 suggests that board
independence curtails managerial opportunism on financial reporting. A positive
coefficient b3 indicates greater separation of control rights from cash flow rights of
the largest controlling shareholder induces greater earnings management. I predict
the positive association between absolute discretionary accruals and the separation
of control rights from cash flow rights to be less pronounced in firms with high
equity ownership by outside directors. Thus, I expect coefficient b4 to be negative.
Furthermore, I predict the positive association between absolute discretionary
accruals and the separation of control rights from cash flow rights to be less
pronounced in firms high proportion of outside directors on the board. Thus,
I expect coefficient b5 to be negative.
Other board characteristics include the total number of directors (BOARDSIZE)
and CEO-chairman duality (CEODUAL). The evidence is mixed on whether board
size and CEO duality impairs board effectiveness. Thus, ex ante, there is no
prediction on the sign on both variables. The model controls for the effects of
firm size, growth opportunities, and leverage on discretionary accruals. Large firms
have greater external monitoring, have more stable operations and stronger control
structures, and hence report smaller abnormal accruals (Dechow and Dichev 2002).
Firm size (LNASSET) is measured based on book value of total assets. Because
discretionary accruals are higher for firms with higher growth opportunities,
I employ the market-to-book equity (MB) ratio to control for the effect of growth
opportunities on discretionary accruals (Kothari et al. 2005). I also include financial
leverage (LEV), defined as long-term debt divided by total assets, to control for the
managerial discretion over the financial accounting process to mitigate constraints
of accounting-based debt covenants (Smith and Watts 1992). To control for the
effect of firm performance on discretionary accruals, I include firm profitability
(ROA), defined as net income divided by total assets. Finally, I include
country dummy variables to capture country-specific factors that may affect the
development of capital markets and financial accounting quality. I include dummy
variables for years and industries to control for time effect and industry effects,
respectively.

Appendix 2: Accruals Quality

My second proxy for earnings quality is accruals quality. Dechow and Dichev
(2002) propose a measure of earnings quality that captures the mapping of current
accruals into last-period, current-period, and next-period cash flows. Francis
et al. (2005) find that this measure (which they term accrual quality) is associated
with measures of cost of equity capital. My measure of accrual quality is based on
Dechow and Dichev’s (2002) model relating current accruals to last-period,
current-period, and next-period cash flows:
42 Earnings Quality and Board Structure: Evidence from South East Asia 1187

TCAj, t CFOj, t1 CFOj, t CFOj, tþ1

¼ g0, j þ g1j þ g2j þ g2j þ ej, t (42.3)
Assets Assets Assets Assets

where:
TCAj,t ¼ firm j’s total current accruals in year t ¼ DCAj,t – DCLj,t – DCASHj,t
+ DSTDj,t
Assets ¼ firm j’s average total assets in year t  1 and year t
CFOj,t ¼ cash flow from operations in year t is calculated as net income less total
accruals (TA) where:
TAj,t ¼ DCAj,t – DCLj,t – DCASHj,t + DSTDj,t – DEPNj,t where
DCAj,t ¼ firm j’s change in current assets between year t  1 and year t
DCLj,t ¼ firm j’s change in current liabilities between year t  1 and year t
DCASHj,t ¼ firm j’s change in cash between year t  1 and year t
DSTDj,t ¼ firm j’s change in debt in current liabilities between year t  1 and
year t
DEPNj,t ¼ firm j’s change in depreciation and amortization expense in year t
I estimate Eq. 42.3 for each one-digit SIC industry for each country-year
combination. These estimations yield firm- and year-specific residuals, ejt, which
form the basis for the accrual quality metric. AQ is the standard deviation of firm j’s
estimated residuals multiplied by 1. Hence, large values of AQ correspond to high
accrual quality.
I employ the following model to test the association between accrual quality and
board characteristics:

AQ ¼ b0 þ b1 EBC þ b2 OUTDIR þ b3 VOTE þ b4 VOTE  DIROWN

þ b5 VOTE  OUTDIR þ b6 CEODUAL þ b7 BOARDSIZE
þ b8 LNASSET þ b9 OPERCYCLE þ b10 NETPPE þ b11 STDSALE
þ b12 STDCFO þ b13 NEGEARN þ Country controls
þ Industry Controls þ Year Controls:
(42.4)

where:
AQ ¼ the standard deviation of firm j’s residuals from a regression of
current accruals on lagged, current, and future cash flows from operations.
I multiply the variable by 1 so that higher AQ measure denotes higher accrual
quality.
OPERCYCLE ¼ log of the sum of the firm’s days accounts receivable and days
inventory.
NETPPE ¼ ratio of the net book value of PP&E to total assets.
STDCFO ¼ standard deviation of the firm’s rolling 5-year cash flows from
operations.
STDSALE ¼ standard deviation of the firm’s rolling 5-year sales revenue.
NEGEARN ¼ the firm’s proportion of losses over the prior 5 years.
All other variables are previously defined.
1188 K.-W. Lee

If high equity ownership for outside directors improves the board monitoring of
managerial discretion over the financial accounting process, I predict coefficient b1
to be positive. Similarly, a positive coefficient for b2 suggests that higher board
independence is associated with higher accrual quality. If greater agency costs arise
from the higher separation of control rights from cash flow rights of the largest
controlling shareholder, coefficient b3 should be negative. I predict that the nega-
tive association between accrual quality and the separation of control rights from
cash flow rights is mitigated in firms with high equity ownership by outside
directors. Thus, I expect coefficient b4 to be positive. Furthermore, the negative
effect of the separation of control rights from cash flow on rights accrual quality
should be attenuated in firms with high proportion of outside directors on the board.
Thus, I expect coefficient b5 to be positive.
In Eq. 42.4, the control variables include innate determinants of accrual quality.
Briefly, Dechow and Dichev (2002) find that accrual quality is positively associated
with firm size and negatively associated with cash flow variability, sales variability,
operating cycle, and incidence of losses. Firm size is measured by the natural
logarithm of total assets (LNASSET). Operating cycle (OPERCYCLE) is the log
of the sum of the firm’s days accounts receivable and days inventory. Capital
intensity, NETPPE, is proxied by the ratio of the net book value of PP&E to total
assets. Cash flow variability (STDCFO) is the standard deviation of the firm’s
rolling 5-year cash flows from operations. Sales variability (STDSALE) is the
standard deviation of the firm’s rolling 5-year sales revenue. Incidence of negative
earnings realizations, NEGEARN, is measured as the firm’s proportion of losses
over the prior 5 years.

Appendix 3: Earnings Informativeness

My third proxy of earnings quality is earnings informativeness, measured by the

earnings response coefficients (Warfield et al. 1995; Fan and Wong 2002; Francis
et al. 2005). The following model is adopted to investigate the relation between
earnings informativeness and equity-based compensation, board independence, and
separation of control rights from cash flow rights:
RET ¼ b0 þ b1 EARN þ b2 EARN  DIROWN þ b3 EARN  OUTDIR
þ b4 EARN  VOTE þ b5 EARN  VOTE  EBC
þ b6 EARN  VOTE  OUTDIR þ b7 EARN  BOARDSIZE
(42.5)
þ b8  EARN  CEODUAL þ b9 EARN  LNASSET
þ b10 EARN  MB þ b11 EARN  LEV þ b12 EARN  ROA
þ Year controls þ Country Controls þ Industry Controls þ e

where:
RET ¼ 12-month cumulative raw return ending 3 months after the fiscal year-end.
EARN ¼ net income for year t, scaled by the market value of equity at the end of t  1.
42 Earnings Quality and Board Structure: Evidence from South East Asia 1189

All other variables are as previously defined.

The estimated coefficient on b1 reflects the earnings response coefficient.
A positive estimate on b2 will be consistent with the notion that equity ownership
for outside directors is associated with more informative earnings. A positive
estimate on b3 indicates that the greater proportion of outside directors on the
board, the greater the informativeness of earnings. From Fan and Wong (2002),
I expect coefficient b4 to be negative, indicating that the reported earnings are less
informative when the ultimate shareholder’s control rights exceed his cash flow
rights. If high equity ownership for outside directors improves their monitoring of
management, the negative effects of the divergence of control rights from cash flow
rights should be mitigated in firms with high equity-based compensation. I expect
coefficient b5 to be positive. If monitoring intensity is positively associated with the
proportion of outside directors on the board, the reduced informativeness of
reported earnings in firm with high divergence of control rights from cash flow
rights should be mitigated in firms with higher board independence. I expect
coefficient b6 to be positive.

Appendix 4: Adjusting for Standard Errors in Panel Data

Gow et al. (2010) examine several approaches to address issues of cross-sectional

and time-series dependence in accounting research. They identified a number of
common approaches: Fama-MacBeth, Newey-West, the Z2 statistic, and standard
errors clustered by firm, industry, or time. Gow et al. (2010) review each of these
approaches and discuss the circumstances in which they produce valid inferences.
This section is drawn heavily from Gow et al. (2010). Correcting for cross-
sectional and time-series dependence in accounting research. The Accounting
Review 85(2), 483–512. Reader should refer to the paper for details.
(i) OLS and White Standard Errors
OLS standard errors assume that errors are both homoskedastic and
uncorrelated across observations. While White (1980) standard errors are
consistent in the presence of heteroskedasticity, both OLS and White produce
misspecified test statistics when either forms of dependence is present.
(ii) Newey-West
Newey and West (1987) generalize the White (1980) approach to yield
a covariance matrix estimator that is robust to both heteroskedasticity and
serial correlation. Gow et al. (2010) find that the Newey-West procedure
produces slightly biased estimates of standard errors when time-series depen-
dence alone is present. However, Gow et al. (2010) find that, in the presence of
both cross-sectional and time-series dependence, Newey and West method
produces misspecified test statistics with even moderate levels of cross-
sectional dependence.
(iii) Fama-MacBeth
The Fama-MacBeth approach (Fama and MacBeth 1973) is designed to
address concerns about cross-sectional correlation. The Fama-MacBeth
1190 K.-W. Lee

approach (FM-t) involves estimating T cross-sectional regressions (one for

each period) and basing inferences on a t-statistic calculated as
b 1X T
^
t¼ , where b ¼ b (42.6)
seðbÞ T t¼1 t

and se(b) is the standard error of the coefficients based on their empirical
distribution. When there is no cross-regression (time-series) dependence, this
approach yields consistent estimates of the standard error of the coefficients as
T goes to infinity.
Two common variants of the Fama-MacBeth approach appear in the
accounting literature. The first variant, FM-i, involves estimating firm- or
portfolio-specific time-series regressions with inferences based on the
cross-sectional distribution of coefficients. This modification of the Fama-
MacBeth approach is appropriate if there is time-series dependence but not
cross-sectional dependence. However, FM-i is frequently used when cross-
sectional dependence is likely, such as when returns are the dependent
variable.
The second common variant of the FM-t approach, FM-NW, is intended to
correct for serial correlation in addition to cross-sectional correlation. FM-NW
modifies FM-t by applying a Newey-West adjustment in an attempt to correct
for serial correlation.
Gow et al. (2010) suggest two reasons to believe that FM-NW may not
correct for serial correlation. First, FM-NW involves applying Newey-West to
a limited number of observations, a setting in which Newey-West is known to
perform poorly. Second, FM-NW applies Newey-West to a time-series of
coefficients, whereas the dependence is in the underlying data.
(iv) Z2 Statistic
The Z2-t (Z2-i) statistic is calculated using t-statistics from separate cross-
sectional (time-series) regressions for each time period (cross-sectional unit)
and is given by the expression:
t 1X T
Z2 ¼ , where t ¼ ^t t , (42.7)
seðtÞ T t¼1

se(t) is the standard error of the t-statistics based on their empirical distribu-
tion, and T is the number of time periods (cross-sectional units) in the sample.
Gow et al. (2010) suggest Z2 may suffer from cross-regression dependence
in the same way as the Fama-MacBeth approach does.
(v) One-Way Cluster-Robust Standard Errors
A number of studies in our survey use cluster-robust standard errors, with
clustering either along a cross-sectional dimension (e.g., analyst, firm, indus-
try, or country) or along a time-series dimension (e.g., year); we refer to
the former as CL-i and the latter as CL-t. Cluster-robust standard errors
42 Earnings Quality and Board Structure: Evidence from South East Asia 1191

(also referred to as Huber-White or Rogers standard errors) were proposed by

White (1980) as a generalization of the heteroskedasticity-robust standard
errors of White (1980). With observations grouped into G clusters of Ng
observations, for g in {1,. . .,G}, the covariance matrix is estimated using the
following expression:
   0 1  0 1 XG
^ B
V ^ ¼ XX ^ XX
B ^¼
,B
0 0
X g ug ug X g , (42.8)
g¼1

where Xg is the Ng  K matrix of regressors, and ug is the Ng-vector of

residuals for cluster g.
While one-way cluster-robust standard errors allow for correlation of unknown
form within cluster, it is assumed that errors are uncorrelated across clusters.
For example, clustering by time (firm) allows observations to be cross-
sectionally (serially) correlated but assumes independence over time
(across firms). While some studies consider both CL-i and CL-t separately,
separate consideration of CL-t and CL-i does not correct for both cross-
sectional and time-series dependence. Gow et al. (2010) find that t-statistics
for CL-t are inflated in the presence of time-series dependence and t-statistics
for CL-i are inflated in the presence of cross-sectional dependence. Thus, when
both forms of dependence are present, both CL-t and CL-i produce overstated
t-statistics.
(vi) Two-Way Cluster-Robust Standard Errors
An extension of cluster-robust standard errors is to allow for clustering along
more than one dimension. In contrast to one-way clustering, two-way cluster-
ing (CL-2) allows for both time-series and cross-sectional dependence. For
example, two-way clustering by firm and year allows for within-firm (time-
series) dependence and within-year (cross-sectional) dependence (e.g., the
observation for firm j in year t can be correlated with that for firm j in year
t + 1 and that for firm k in year t). To estimate two-way cluster-robust standard
errors, the expression in (A4) is evaluated using clusters along each dimension
(e.g., clustered by industry and clustered by year) to yield V1 and V2. Then the
same expression is calculated using the “intersection” clusters (in the example,
observations within an industry-year) to yield VI. The two-way cluster-robust
estimator V is calculated as V ¼ V1 + V2  VI. Standard econometric software
packages (e.g., Stata and SAS) contain routines for calculating one-way
cluster-robust standard errors, making it relatively straightforward to imple-
ment two-way cluster-robust standard errors. Gow et al. (2010) find that in the
presence of both cross-sectional and time-series dependence, the two-way
clustering method (by year and by firm) which allows for both cross-sectional
and time-series dependence produces well-specified test statistics. Johnston
and DiNardo (1997) and Greene (2000) are two econometric textbooks
that contain a detailed discussion of the econometrics issues relating to
panel data.
1192 K.-W. Lee

Acknowledgment I appreciate the research funding provided by Institute of Certified

Public Accountants of Singapore. I would like to dedicate this paper to my late father,
Yew-Ming Lee.

References
Ashbaugh, H., LaFond, R., & Mayhew, B. W. (2003). Do nonaudit services compromise auditor
independence? Further evidence. The Accounting Review, 78(3), 611–639.
Beasley, M. S. (1996). An empirical analysis of the relation between the board of director
composition and financial statement fraud. The Accounting Review, 71, 443–465.
Claessens, S., Djankov, S., & Lang, L. H. P. (2000). The separation of ownership and control in
East Asian corporations. Journal of Financial Economics, 58, 81–112.
Claessens, S., Djankov, S., Fan, J. P. H., & Lang, L. H. P. (2002). Disentangling the incentive and
entrenchment effects of large shareholdings. Journal of Finance, 57(2), 2741–2771.
Dechow, P. M., & Dichev, I. (2002). The quality of accruals and earnings: The role of accrual
estimation errors. The Accounting Review, 77, 35–59.
Dechow, P. M., Sloan, R. G., & Sweeney, A. P. (1996). Causes and consequences of earnings
manipulation: An analysis of firms subject to enforcement actions by the SEC. Contemporary
Accounting Research, 13, 1–36.
Fama, E., & Jensen, M. (1983). Agency problems and residual claims. Journal of Law and
Economics, 26, 327–349.
Fama, E., & MacBeth, J. (1973). Risk, return, and equilibrium: Empirical tests. The Journal of
Political Economy, 81(3), 607–636.
Fan, J., & Wong, T. J. (2002). Corporate ownership structure and the informativeness of account-
ing earnings in East Asia. Journal of Accounting & Economics, 33, 401–425.
Fan, J., & Wong, T. J. (2005). Do external auditors perform a corporate governance role in
emerging markets? Journal of Accounting Research, 43(1), 35–72.
Francis, J., Lafond, R., Olsson, P., & Schipper, K. (2005). The market pricing of accrual quality.
Journal of Accounting & Economics, 39, 295–327.
Gillette, A., Noe, T., & Rebello, M. (2003). Corporate board composition, protocols and voting
behavior: Experimental evidence. Journal of Finance, 58, 1997–2032.
Gow, I. D., Ormazabal, G., & Taylor, D. J. (2010). Correcting for cross-sectional and time series
dependence in accounting research. The Accounting Review, 85(2), 483–512.
Greene, W. (2000). Econometrics analysis. Upper Saddle River: Prentice-Hall.
Haw, I., Hu, B., Hwang, L., & Wu, W. (2004). Ultimate ownership, income management, and legal
and extra-legal institutions. Journal of Accounting Research, 42, 423–462.
Hermalin, B., & Weisbach, M. (1988). The determinants of board composition. Rand Journal of
Economics, 19, 589–606.
Johnson, S., Boone, P., Breach, A., & Friedman, E. (2000). Corporate governance in the Asian
financial crisis. Journal of Financial Economics, 58, 141–186.
Johnston, J., & DiNardo, J. (1997). Econometrics method. New York: Mc-Graw Hill.
Klein, A. (2002). Audit committee, board of director characteristics, and earnings management.
Journal of Accounting and Economics, 33(3), 375–400.
Kothari, S. P., Leone, A. J., & Wasley, C. E. (2005). Performance matched discretionary accruals.
Journal of Accounting and Economics, 39, 163–197.
La Porta, R., Lopez-de-Silanes, F., Shleifer, A., & Vishny, R. (1998). Law and finance. Journal of
Political Economy, 106, 1113–1155.
La Porta, R., Lopez-De-Silanes, F., & Shleifer, A. (1999). Corporate ownership around the world.
Journal of Finance, 54, 471–518.
Lang, M., Lins, K., & Miller, D. (2004). Concentrated control, analyst following, and valuation:
Do analysts matter most when investors are protected least? Journal of Accounting Research,
42(3), 589–623.
42 Earnings Quality and Board Structure: Evidence from South East Asia 1193

Lee, K. W. (2007). Corporate voluntary disclosure and the separation of cash flow rights from
control rights. Review of Quantitative Finance and Accounting, 28, 393–416.
Lee, K. W., Lev, B., & Yeo, H. H. (2007). Organizational structure and earnings management.
Journal of Accounting, Auditing and Finance, 22(2), 293–391.
Lee, K. W., Lev, B., & Yeo, H. H. (2008). Executive pay dispersion, corporate governance and
firm performance. Review of Quantitative Finance and Accounting, 30, 315–338.
Lee, C. F., Lee, K. W., & Yeo, H. H. (2009). Investor protection and convertible debt design.
Journal of Banking and Finance, 33(6), 985–995.
Leuz, C., Nanda, D., & Wysocki, P. (2003). Earnings management and investor protection: An
international comparison. Journal of Financial Economics, 69, 505–527.
Mehran, H. (1995). Executive compensation structure, ownership, and firm performance. Journal
of Financial Economics, 38(2), 163–184.
Morck, R., Shleifer, A., & Vishny, R. W. (1988). Management ownership and market valuation:
An empirical analysis. Journal of Financial Economics, 20, 293–315.
Newey, W., & West, D. (1987). A simple, positive semi-definite, heteroskedasticity and autocor-
relation consistent covariance matrix. Econometrica, 55(3), 703–708.
Peasnell, K., Pope, P., & Young, S. (2005). Board monitoring and earnings management:
Do outside directors influence abnormal accruals? Journal of Business Finance & Accounting,
32(7/8), 1311–1345.
Perry, T. (2000). Incentive compensation for outside directors and CEO turnover. Working paper,
Arizona State University.
Petersen, M. (2009). Estimating standard errors in finance panel data sets: Comparing approaches.
Review of Financial Studies, 22(1), 435–480.
Rosenstein, S., & Wyatt, J. (1990). Outside directors, board independence, and shareholder wealth.
Journal of Financial Economics, 26, 175–192.
Ryan, H., & Wiggins, R. (2004). Who is whose pocket? Director compensation, board indepen-
dence, barriers to effective monitoring. Journal of Financial Economics, 73, 497–524.
Shivdasani, A. (1993). Board composition, ownership structure, and hostile takeovers. Journal of
Accounting & Economics, 16, 167–188.
Shleifer, A., & Vishny, R. (1997). A survey of corporate governance. Journal of Finance,
52, 737–783.
Smith, C., & Watts, R. (1992). The investment opportunity set and corporate financing, dividend
and compensation policies. Journal of Financial Economics, 32, 263–292.
Warfield, T., Wild, J. J., & Wild, K. (1995). Managerial ownership, accounting choices, and
informativeness of earnings. Journal of Accounting and Economics, 20, 61–91.
Weisbach, M. (1988). Outside directors and CEO turnover. Journal of Financial Economics,
20, 431–460.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica, 48(4), 817–838.
Rationality and Heterogeneity of Survey
Forecasts of the Yen-Dollar Exchange 43
Rate: A Reexamination

Richard Cohen, Carl S. Bonham, and Shigeyuki Abe

Contents
43.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1196
43.2 Background: Testing Rationality in the Foreign Exchange Market . . . . . . . . . . . . . . . . . . . 1199
43.2.1 Why Test Rational Expectations with Disaggregated Survey
Forecast Data? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1200
43.2.2 Rational Reasons for the Failure of the Rational Expectations Hypothesis
Using Disaggregated Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1201
43.3 Description of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1202
43.4 Empirical Tests of Rationality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1203
43.4.1 Joint Tests of Unbiasedness and Weak Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 1204
43.4.2 Pretests for Rationality: The Stationarity of the Forecast Error . . . . . . . . . . . . . 1212
43.4.3 Univariate Tests for Unbiasedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218
43.4.4 Unbiasedness Tests Using Error Correction Models . . . . . . . . . . . . . . . . . . . . . . . . . 1222
43.4.5 Explicit Tests of Weak Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1228
43.5 Micro-homogeneity Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1230
43.5.1 Ito’s Heterogeneity Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1239
43.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1241
Appendix 1: Testing Micro-homogeneity with Survey Forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1244
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1245

R. Cohen
University of Hawaii Economic Research Organization and Economics, University of Hawaii at
Manoa, Honolulu, HI, USA
e-mail: afrc2@cbpp.uaa.alaska.edu
C.S. Bonham (*)
College of Business and Public Policy, University of Alaska Anchorage, Anchorage, AK, USA
e-mail: bonham@hawaii.edu
S. Abe
Faculty of Policy Studies, Doshisha University, Kyoto, Japan
e-mail: sabe@mail.doshisha.ac.jp

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1195
DOI 10.1007/978-1-4614-7750-1_43,
# Springer Science+Business Media New York 2015
1196 R. Cohen et al.

Abstract
This chapter examines the rationality and diversity of industry-level forecasts of
the yen-dollar exchange rate collected by the Japan Center for International
Finance. In several ways we update and extend the seminal work by Ito (1990,
American Economic Review 80, 434–449). We compare three specifications for
testing rationality: the “conventional” bivariate regression, the univariate regres-
sion of a forecast error on a constant and other information set variables, and an
error correction model (ECM). We find that the bivariate specification, while
producing consistent estimates, suffers from two defects: first, the conventional
restrictions are sufficient but not necessary for unbiasedness; second, the test has
low power. However, before we can apply the univariate specification, we must
conduct pretests for the stationarity of the forecast error. We find a unit root in
the 6-month horizon forecast error for all groups, thereby rejecting unbiasedness
and weak efficiency at the pretest stage. For the other two horizons, we find
much evidence in favor of unbiasedness but not weak efficiency. Our ECM
rejects unbiasedness for all forecasters at all horizons. We conjecture that these
results, too, occur because the restrictions test sufficiency, not necessity.
We extend the analysis of industry-level forecasts to a SUR-type structure
using an innovative GMM technique (Bonham and Cohen 2001, Journal of
Business & Economic Statistics, 19, 278–291) that allows for forecaster cross-
correlation due to the existence of common shocks and/or herd effects. Our
GMM tests of micro-homogeneity uniformly reject the hypothesis that fore-
casters exhibit similar rationality characteristics.

Keywords
Rational expectations • Unbiasedness • Weak efficiency • Micro-homogeneity •
Heterogeneity • Exchange rate • Survey forecasts • Aggregation bias • GMM •
SUR

43.1 Introduction

This chapter examines the rationality of industry-level survey forecasts of the

yen-dollar exchange rate collected by the Japan Center for International Finance
(JCIF). Tests of rationality take on additional significance when performed on asset
market prices, since rational expectations is a necessary condition for market
efficiency. In the foreign exchange market, tests of forward rate unbiasedness
simultaneously test a zero risk premium in the exchange rate; hence this joint
hypothesis is also called the risk-neutral efficient market hypothesis (RNEMH).
The practical significance of such a hypothesis is that if the forward rate is indeed an
unbiased predictor of the future spot rate, then exchange risk can be costlessly
hedged in the forward market. However, the RNEMH has been rejected nearly
universally. Since the risk premium is unobservable, insight into the reason for the
rejection of the RNEMH can be gained by separately testing for rationality using
survey data on expectations. Because forecasters cannot be assumed to have
43 Rationality and Heterogeneity of Survey Forecasts 1197

identical information sets, we must use individual survey forecasts to avoid the
aggregation bias inherent in the use of mean or median forecasts.
We use data from the same source as Ito (1990), the seminal study recognizing
the importance of using individual data to test rationality hypotheses about the
exchange rate. To achieve stationarity of the realizations and forecasts (which each
have a unit root), Ito (1990) followed the conventional specification at the time of
subtracting the current realization from each. These variables are then referred to as
being in “change” form. To test unbiasedness he regressed the future rate of
depreciation on the forecasted return and tested the joint restrictions that the
intercept equalled zero and the slope coefficient equalled one. At the industry
level he found approximately twice as many rejections (at the 1 % level) at the
longest horizon (6 months) than at the two shorter horizons (1 and 3 months).
We extend Ito’s analysis in two principal respects: the specification of unbiased-
ness tests and inference in tests for micro-homogeneity of forecasters. One problem
with the change specification of unbiasedness tests is that, since there is much more
variation in the change in the realization than in the forecast, there is a tendency to
under-reject the part of the joint hypothesis that the coefficient on the forecast
equals one. This is precisely what we would expect in tests of variables which are
near random walks.
Second, and more fundamentally, Ito’s (1990) bivariate (joint) regression test of
unbiasedness is actually a test of sufficiency, not necessity as well as sufficiency.
Following Holden and Peel (1990), the necessary and sufficient condition for
unbiasedness is a mean zero forecast error. This is tested in a univariate regression
by imposing a coefficient of unity on the forecast and testing the restriction that the
intercept equals zero. This critique applies whether or not the forecast and realiza-
tion are integrated in levels. However, when the realization and forecast are
integrated in levels, we must conduct a pretest to determine whether the forecast
error is stationary. If the forecast and realization are both integrated and
cointegrated, then a necessary and sufficient condition for unbiasedness is that
intercept and slope in the cointegrating regression (using levels of the realization
and forecast) are zero and one, respectively. We test this hypothesis using Liu and
Maddala’s (1992) method of imposing the (0, 1) vector, then testing the “restricted”
cointegrating residual for stationarity.1,2
Third, we use the result from Engle and Granger (1987) that cointegrated vari-
ables have an error correction representation. First, we employ the specification and
unbiasedness restrictions originally proposed by Hakkio and Rush (1989). How-
ever, the unbiasedness tests using the ECM specification produce more rejections
over industry groups and horizons than the univariate or bivariate specifications.

1
If in addition the residuals from the cointegrating regression are white noise, this supports a type
of weak efficiency.
2
Pretesting the forecast error for stationarity is a common practice in testing the RNMEH, but the
only study we know of that applies this practice to survey forecasts of exchange rates is Osterberg
(2000), and he does not test for a zero intercept in the cointegrating regression.
1198 R. Cohen et al.

We conjecture that one possible explanation for this apparent anomaly is that,
similar to the joint restrictions in the bivariate test, the ECM restrictions test
sufficient conditions for unbiasedness, while the univariate restriction only tests
a necessary and sufficient condition. Thus, the ECM has a tendency to over-reject.
We then respecify the ECM, so that only the necessary and sufficient conditions are
tested. We compare our results to those obtained using the sufficient conditions
represented by the joint restrictions as well as the necessary and sufficient condition
represented by the univariate restriction.
The second direction in which we extend Ito’s (1990) analysis has to do with
testing for differences among forecasters’ ability to produce rational predictions.3
We recognize, as does Ito, that differences among forecasters over time indicate
that at least some individuals form biased forecasts. (The converse does not
necessarily hold, since a failure to reject micro-homogeneity could conceivably
be due to the same degree of irrationality of each individual in the panel.) Ito’s
heterogeneity test is a single-equation test of deviations of individual forecasts from
the mean forecast, where the latter may or may not be unbiased. In contrast, we test
for differences in individual forecast performance using a micro-homogeneity test,
i.e., testing for equal coefficients across the system of individual univariate
rationality equations.
In our tests for micro-homogeneity, we expect cross-forecaster error correlation
due to the possibility of common macro shocks and/or herd effects in expectations.
To this end, we incorporate two innovations not previously used by investigators
studying survey data on exchange rate expectations. First, in our micro-
homogeneity tests, we use a GMM system with a variance-covariance matrix that
allows for cross-sectional as well as moving average and heteroscedastic errors.
Here we follow the widely used practice of modeling the individual regression
residuals as an MA process of order h-1, where h is the number of periods in the
forecast horizon. However, no other researchers have actually tested whether an
MA process of this length is required to model the cross-sectional behavior of
rational forecast errors. Thus, second, to investigate the nature of the actual MA
processes, we use Pesaran’s (2004) CD test to examine the statistical significance of
the cross-sectional dependence of forecast errors, both contemporaneous and
lagged.
The organization of the rest of the chapter is as follows: in Sect. 43.2 we review
some fundamental issues in testing rationality in the foreign exchange market. In
Sects. 43.3 and 43.4, we conduct various rationality tests on the JCIF data.
Section 43.5 contains our micro-homogeneity tests. Section 43.6 summarizes and
discusses areas for future research.

3
Market microstructure theories assume that there is a minimum amount of forecaster (as well as
cross-sectional forecast) diversity. Also, theories of exchange rate determination that depend upon
the interaction between chartists (or noise traders) and fundamentalists by definition require
a certain structure of forecaster heterogeneity.
43 Rationality and Heterogeneity of Survey Forecasts 1199

43.2 Background: Testing Rationality in the Foreign

Exchange Market

The Rational Expectations Hypothesis (REH) assumes that economic agents know
the true data-generating process (DGP) for the forecast variable. This implies that
the market’s subjective probability distribution of the variable is identical to the
objective probability distribution, conditional on a given information set, Ft.
Equating first moments of the market, Em ðstþh jFt Þ, and objective, Eðstþh jFt Þ,
distributions,

Em ðstþh jFt Þ ¼ Eðstþh jFt Þ, (43.1)

where the right-hand side can be shortened to Et(st+h).

It follows that the REH implies that forecast errors have both unconditional and
conditional means equal to zero. A forecast is unbiased if its forecast error has an
unconditional mean of zero. A forecast is efficient if its error has a conditional mean
of zero. The condition that forecast errors be serially uncorrelated is a subset of the
efficiency condition where the conditioning information set consists of past values
of the realization and current as well as past values of the forecast.4
In this chapter we focus on testing whether forecasters can form rational
expectations of future depreciation. If not, then at least part of the explanation for
the failure of the RNEMH is due to the failure of the REH. There are two related
interest parity conditions. Covered interest parity, an arbitrage condition, holds if
ft,h  st ¼ it  it , i.e., the forward premium is equal to the interest differential
between domestic and foreign risk-free assets. Uncovered interest parity holds if
st + h  set ¼ it  it . Because uncovered interest parity assumes both unbiased
expectations and risk neutrality, some authors view it as equivalent to the
RNEMH (see Phillips and Maynard 2001).
The ability to decompose deviations from UIP into time-varying risk premium
and systematic forecast error components also has implications for policymakers.
Consider first the possibility of a violation of the risk neutrality hypothesis.
According to the portfolio balance model, if a statistically significant time-varying
risk premium component is found, this means that it  it is time-varying, which in
turn implies that foreign and domestic bonds are not perfect substitutes; changes in
relative quantities (which are reflected in changes in current account balances) will
affect the interest rate differential. In this way, sterilized official intervention can
have significant effects on exchange rates. Second, consider the possibility of
a violation of the REH. If a statistically significant expectational error of the
destabilizing (e.g., “bandwagon”) type is found, and policymakers are more rational
than speculators, a policy of “leaning against the wind” could have a stabilizing

4
It is important to note that the result from one type of rationality test does not have implications
for the results from any other types of rationality tests. In this chapter we test for unbiasedness and
weak efficiency, leaving the more stringent tests of efficiency with respect to publicly available
information for future analysis.
1200 R. Cohen et al.

effect on exchange rate movements. (See Cavaglia et al. 1994.) More generally,
monetary models of the exchange rate (in which the UIP condition is embedded),
which assume model-consistent (i.e., rational) expectations with risk neutrality,
generally have not performed well empirically, especially in out-of-sample fore-
casting. (See, e.g., Bryant 1995.) One would like to be able to attribute the model
failure to some combination of a failure of the structural assumptions (including
risk neutrality) and a failure of the expectational assumption.

43.2.1 Why Test Rational Expectations with Disaggregated Survey

Forecast Data?

Beginning with Frankel and Froot (1987) and Froot and Frankel (1989), much of
the literature examining exchange rate rationality in general, and the decomposition
of deviations from the RNEMH in particular, has employed the representative agent
assumption to justify using the mean or median survey forecast as a proxy for the
market’s expectation. In both studies, Frankel and Froot found significant evidence
of irrationality. Subsequent research has found mixed results. Liu and Maddala
(1992, p. 366) articulate the mainstream justification for using aggregated forecasts
in tests of the REH. “Although . . .data on individuals are important to throw light
on how expectations are formed at the individual level, to analyze issues relating to
market efficiency, one has to resort to aggregates.” In fact, Muth’s (1961, p. 316)
original definition of rational expectations seemed to allow for the possibility
that rationality could be applied to an aggregate (e.g., mean or median) forecast.
‘. . . [E]xpectations of firms (or, more generally, the subjective probability distri-
bution of outcomes) tend to be distributed, for the same information set, about the
predictions of the theory (or the “objective” probability distribution of outcomes).’
(Emphasis added.)
However, if individual forecasters have different information sets, Muth’s def-
inition does not apply. To take the simplest example, the (current) mean forecast is
not in any forecaster’s information set, since all individuals’ forecasts must be made
before a mean can be calculated. Thus, current mean forecasts contain private
information (see MacDonald 1992) and therefore cannot be tested for rationality.5
Using the mean forecast may also result in inconsistent parameter estimates.
Figlewski and Wachtel (1983) were the first to show that, in the traditional bivariate

5
A large theoretical literature relaxes Muth’s assumption that all information relevant for forming
a rational forecast is publicly available. Instead, this literature examines how heterogeneous
individual expectations are mapped into an aggregate market expectation, and whether the latter
leads to market efficiency. (See, e.g., Figlewski 1978, 1982, 1984; Kirman 1992; Haltiwanger and
Waldman 1989.) Our paper focuses on individual rationality but allows for the possibility of
synergism by incorporating not only heteroscedasticity and autocorrelation consistent standard
errors in individual rationality tests but also cross-forecaster correlation in tests of micro-
homogeneity. The extreme informational requirement of the REH led Pesaran and Weale (2006)
to propose a weaker form of the REH that is based on the (weighted) average expectation using
only publicly available (i.e., common) information.
43 Rationality and Heterogeneity of Survey Forecasts 1201

unbiasedness equation, the presence of private information variables in the mean

forecast error sets up a correlation with the mean forecast. This inconsistency
occurs even if all individual forecasts are rational. In addition, Keane and Runkle
(1990) pointed out that, when some forecasters are irrational, using the mean
forecast may lead to false acceptance of the unbiasedness hypothesis, in the
unlikely event that offsetting individual biases allow parameters to be consistently
estimated. See also Bonham and Cohen (2001), who argue that, in the case of
cointegrated targets and predictions, inconsistency of estimates in rationality tests
using the mean forecast can be avoided if corresponding coefficients in the indi-
vidual rationality tests pass a test for micro-homogeneity.6 Nevertheless, until the
1990s, few researchers tested for the rationality of individual forecasts, even when
those data were available.

43.2.2 Rational Reasons for the Failure of the Rational Expectations

Hypothesis Using Disaggregated Data

Other than a failure to process available information efficiently, there are numerous
explanations for a rejection of the REH. One set of reasons relates to measurement
error in the individual forecast. Researchers have long recognized that forecasts of
economic variables collected from public opinion surveys should be less informed
than those sampled from industry participants. However, industry participants,
while relatively knowledgeable, may not be properly motivated to devote the
time and resources necessary to elicit their best responses. The opposite is also
possible.7 Having devoted substantial resources to produce a forecast of the price of
a widely traded asset, such as foreign exchange, forecasters may be reluctant to
reveal their true forecast before they have had a chance to trade for their own
account.8
Second, some forecasters may not have the symmetric quadratic loss function
embodied in typical measures of forecast accuracy, e.g., minimum mean
squared error. (See Zellner 1986; Stockman 1987; Batchelor and Peel 1998.)
In this case, the optimal forecast may not be the MSE. In one scenario, related to

6
The extent to which private information influences forecasts is more controversial in the foreign
exchange market than in the equity or bond markets. While Chionis and MacDonald (1997)
maintain that there is little or no private information in the foreign exchange market, Lyons
(2002) argues that order flow explains much of the variation in prices. To the extent that one agrees
with the market microstructure emphasis on the importance of the private information embodied in
dealer order flow, the Figlewski-Wachtel critique remains valid in the returns regression.
7
Elliott and Ito (1999) show that, although a random walk forecast frequently outperforms the JCIF
survey forecasts using an MSE criterion, survey forecasts generally outperform the random walk,
based on an excess profits criterion. This supports the contention that JCIF forecasters are properly
motivated to produce their best forecasts.
8
To mitigate the confidentiality problem in this case, the survey typically withholds individual
forecasts until the realization is known or (as with the JCIF) masks the individual forecast by only
reporting some aggregate forecast (at the industry and total level) to the public.
1202 R. Cohen et al.

the incentive aspect of the measurement error problem, forecasters may have
strategic incentives involving product differentiation.9
In addition to strategic behavior, another scenario in which forecasters may
deviate from the symmetric quadratic loss function is simply to maximize trading
profits. This requires predicting the direction of change, regardless of MSE.10
Third, despite their best efforts, forecasters may find it difficult to distinguish
between a temporary and permanent shift in the DGP. This difficulty underlies at
least three theories of rational forecast errors: the peso problem, learning about past
regime changes, and bubbles.
Below we conduct tests for structural change in estimated unbiasedness coeffi-
cients. When unbiasedness cannot be rejected, the structural change test may show
certain subperiods in which unbiasedness did not hold. In the obverse case, when
unbiasedness can be rejected, the structural change test may show certain sub-
periods in which unbiasedness cannot be rejected. Either situation would lend some
support to the theories attributing bias to the difficulty of distinguishing temporary
from permanent shifts.

43.3 Description of Data

Every 2 weeks, the JCIF in Tokyo conducts telephone surveys of yen/dollar exchange rate
expectations from 44 firms. The forecasts are for the future spot rate at horizons of
1 month, 3 months, and 6 months. Our data cover the period May 1985 to March 1996.
This data set has very few missing observations, making it close to a true panel. For
reporting purposes, the JCIF currently groups individual firms into four industry catego-
ries: (1) banks and brokers, (2) insurance and trading companies, (3) exporters, and (4) life
insurance companies and importers. On the day after the survey, the JCIF announces
overall and industry average forecasts. (For further details concerning the JCIF database,
see the descriptions in Ito (1990, 1994), Bryant (1995), and Elliott and Ito (1999).)
Figure 43.1 shows that, over the sample period (one of flexible exchange rates and
no capital controls), the yen appreciated dramatically relative to the dollar, from
a spot rate of approximately 270 yen/dollar in May 1985 to approximately
90 yen/dollar in March 1996. The path of appreciation was not steady, however.
In the first 2 years of the survey alone, the yen appreciated to about 140 per dollar.

9
Laster et al. (1999) called this practice “rational bias.” Prominent references in this growing
literature include Lamont (2002), Ehrbeck and Waldmann (1996), and Batchelor and Dua
(1990a, b, 1992). Because we have access only to forecasts at the industry average level, we
cannot test the strategic incentive hypotheses.
10
See Elliott and Ito (1999), Boothe and Glassman (1987), LeBaron (2000), Leitch and Tanner
(1991), Lai (1990), Goldberg and Frydman (1996), and Pilbeam (1995). This type of loss function
may appear to be relevant only for relatively liquid assets such as foreign exchange, but not for
macroeconomic flows. However, the directional goal is also used in models to predict business
cycle turning points. Also, trends in financial engineering may lead to the creation of derivative
contracts in macroeconomic variables, e.g., CPI futures.
43 Rationality and Heterogeneity of Survey Forecasts 1203

Fig. 43.1 Yen-dollar 260

exchange rate versus time
240

220

200

180

160

140

120

100
st
80
85/05/29 88/03/30 91/02/26 94/01/11 96/11/12

The initial rapid appreciation of the yen is generally attributed to the Plaza meeting in
September 1985, in which the Group of Five countries decided to let the dollar
depreciate, relative to the other currencies. At the Louvre meeting in February 1987,
the Group of Seven agreed to stabilize exchange rates by establishing soft target
zones. These meetings may well be interpreted as unanticipated regime changes,
since, as we will see below, forecasters generally underestimated the rapid appreci-
ation following the Plaza meeting, then overestimated the value of the yen following
the Louvre meeting. Thus, forecasts during these periods may have been subject to
peso and learning problems. The period of stabilization lasted until about 1990, when
yen appreciation resumed and continued through the end of the sample period.

43.4 Empirical Tests of Rationality

Early studies of the unbiasedness aspect of rationality regressed the level of the
realization on the level of the forecast, testing the joint hypothesis that the intercept
equalled zero and the slope equalled one.11 However, since many macroeconomic
variables have unit roots, and realizations and forecasts typically share a common
stochastic trend, a rational forecast will be integrated and cointegrated with the target
series. (See Granger 1991, pp. 69–70.) According to the modern theory of regressions

11
The efficiency aspect of rationality is sometimes tested by including additional variables in the
forecaster’s information set, with corresponding hypotheses of zero coefficients on these variables.
See, e.g., Keane and Runkle (1990) for a more recent study using the level specification and
Bonham and Cohen (1995) for a critique of Keane and Runkle’s integration accounting.
1204 R. Cohen et al.

with integrated processes (see, inter alia Banerjee et al. 1993), conventional OLS
estimation and inference produce a slope coefficient that is biased toward one and,
therefore, a test statistic that is biased toward accepting the null of unbiasedness. The
second generation studies of unbiasedness addressed this inference problem by
subtracting the current realization from the forecast as well as the future realization,
transforming the levels regression into a “changes” regression. In this specification of
stationary variables, unbiasedness was still tested using the same (0, 1) joint hypothesis
as in the levels regression. (Ito (1990) is an example of this methodology.) However, an
implication of Engle and Granger (1987) is that the levels regression is now interpreted
as a cointegrating regression, with conventional t-statistics following nonstandard
distributions which depend on nuisance parameters. After establishing that the reali-
zation and forecast are integrated and cointegrated, we perform two types of rationality
tests. The first is a “restricted cointegration” test due to Liu and Maddala (1992). This is
a cointegration test imposing the (0, 1) restriction on the levels regression.
It is significant that, if realization and forecast are cointegrated, Liu and
Maddala’s (1992) technique is equivalent to regressing a stationary forecast error
on a constant and then testing whether the coefficient equals zero (to test unbiased-
ness) and/or whether the residuals are white noise (to test a type of weak efficiency).
Pretests for unit roots in the realization, forecast, and forecast error are required for
at least three reasons. First, univariate tests of unbiasedness are invalid if the
forecast error is not stationary. Second, following Holden and Peel (1990), we
show below (in Sect. 43.4.1.1) that nonrejection of the joint test in the bivariate
regression is sufficient but not necessary for unbiasedness, since the joint test is also
an implicit test of weak efficiency with respect to the lagged forecast error. A zero
intercept in the (correctly specified) univariate test is a necessary as well as
sufficient condition for unbiasedness. Third, the Engle and Granger (1987) repre-
sentation theorem proves that a cointegrating regression such as the levels joint
regression (Eq. 43.2 below) has an error correction form that includes both
differenced variables and an error correction term in levels. Under the joint null,
the error correction term is the forecast error. While the change form of the bivariate
regression, is not, strictly speaking, misspecified (since the regressor subtracts st,
e
not st1 , from ste), the ECM specification may produce a better fit to the data and,
therefore, a more powerful test of the unbiasedness restrictions. We conduct such
tests using a form of the ECM due to Hakkio and Rush (1989).

43.4.1 Joint Tests of Unbiasedness and Weak Efficiency

43.4.1.1 The Lack of Necessity Critique

Many, perhaps most, empirical tests of the “unbiasedness” of survey forecasts are
conducted using the bivariate regression equation


stþh  st ¼ ai, h þ bi, h sei, t, h  st þ ei, t, h : (43.2)

It is typical for researchers to interpret their nonrejection of the joint

null (ai,h, bi,h) ¼ (0, 1) as a necessary condition for unbiasedness. However,
43 Rationality and Heterogeneity of Survey Forecasts 1205

Fig. 43.2 Unbiasedness with st+h − st

a 6¼ 0, b 6¼ 1

E(st+h − st)

45⬚ E(s ei,t,h−st)

0 s ei,t,h−st

Holden and Peel (1990) show that this result is a sufficient, though not a necessary,
condition for unbiasedness. The intuition for the lack of necessity comes
from interpreting the right-hand side of the bivariate unbiasedness regression
as a linear combination of two potentially unbiased forecasts: a constant
equal to the unconditional mean forecast plus a variable forecast, i.e.,
st + h  st ¼ (1  bi,h)  E(sei;t;h  st) + bi,h(sei;t;h  st) + ei,t,h. Then the intercept is
ai,h ¼ (1  bi,h)  E(sei;t;h  st). The necessary and sufficient condition for
unbiasedness is that the unconditional mean of the subjective expectation
E[sei;t;h  st] equal the unconditional mean for the objective expectation
E[st + h  st]. However, this equality can be satisfied without ai,h being equal to
zero, i.e., bi,h ¼ 1.
Figure 43.2 shows that an infinite number of ai,h, bi,h estimates are
consistent with unbiasedness. The only constraint is that the regression line inter-
sects the 45 ray from the origin where the sample mean of the forecast and target
are equal. Note that, in the case of differenced variables, this can occur at the origin,
so that ai,h ¼ 0, but bi,h is unrestricted (see Fig. 43.3). It is easy to see why
unbiasedness holds: in Figs. 43.2 and 43.3 the sum of all horizontal deviations
from the 45 line to the regression line, i.e., forecast errors, equal zero. However,
when ai,h 6¼ 0, and ai,h 6¼ (1  bi,h)  E(sei;t;h  st), there is bias regardless of the
value of bi,h. See Fig. 43.4, where the bias, E(st + h  sei;t;h ), implies systematic
underforecasts.
1206 R. Cohen et al.

Fig. 43.3 Unbiasedness with st+h − st

a ¼ 0, b 6¼ 1

E(st+h − st)
0

E(s ei,t,h−st)
45⬚
0 s ei,t,h−st

st+h − st

E(st+h − st)
bias

45⬚ E(s ei,t,h−st)

Fig. 43.4 Bias with a 6¼ 0, 0 s ei,t,h−st
b¼1
43 Rationality and Heterogeneity of Survey Forecasts 1207

To investigate the rationality implications of different values for ai,h and bi,h, we
follow Clements and Hendry (1998) and rewrite the forecast error in the bivariate
regression framework of Eq. 43.2 as
 

i, t, h ¼ stþh  sei, t, h ¼ ai, h þ bi, h  1 sei, t, h  st þ ei, t, h (43.3)

A special case of weak efficiency occurs when the forecast and forecast error are
uncorrelated, i.e.,
h
i
  
2
E i, t, h sei, t, h  st ¼ 0 ¼ ai, h E sei, t, h  st þ bi, h  1 E sei, t, h  st
h
i
þ E ei, t, h sei, t, h  st
(43.4)
Thus, satisfaction of the joint hypothesis (ai,h, bi,h) ¼ (0,1) is also sufficient for weak
efficiency with respect to the current forecast. However, it should be noted that Eq. 43.4
may still hold even if the joint hypothesis is rejected. Thus, satisfaction of the joint
hypothesis represents sufficient conditions for both unbiasedness and this type of weak
efficiency, but necessary conditions for neither.
If bi,h ¼ 1, then, whether or not ai,h ¼ 0, the variance of the forecast error equals
the variance of the bivariate regression residual, since then var(i,t,h) ¼
(bi, h  1)2var(sei;t;h  st) + var(ei,t,h) + 2(bi,h  1)cov[(sei;t;h  st),ei,t,h] ¼ var(ei,t,h).
Figure 43.4 illustrates this point. Mincer and Zarnowitz (1969) required only that
bi, h ¼ 1 in their definition of forecast efficiency. If in addition to bi,h ¼ 1, ai,h ¼ 0, then
the mean square forecast error also equals the variance of the forecast. Mincer and
Zarnowitz emphasized that as long as the loss function is symmetric, as is the case
with a minimum mean square error criterion, satisfaction of the joint hypothesis
implies optimality of forecasts.

43.4.1.2 Empirical Results of Joint Tests

Since Hansen and Hodrick (1980), researchers have recognized that, when data are
sampled more frequently than the forecast horizon (h), forecast errors may follow
an h-1 period moving average process. The typical procedure has been to use
a variance-covariance matrix which allows for generalized serial correlation.
Throughout this chapter, we use the Newey-West (1987) procedure, with the
number of lagged residuals set to h-1. To ensure a positive semi-definite VCV
matrix, we use a Bartlett window (see Hamilton 1994, pp. 281–84).
In Tables 43.1, 43.2, and 43.3 we report results for the joint unbiasedness tests.
We reject the joint hypothesis (ai,h, bi,h) ¼ (0, 1) at the 5 % significance level for all
groups except banks and brokers at the 1-month horizon (indicating the possible
role of inefficiency with respect to the current forecast), but only for the exporters at
the 3- and 6-month horizons.
Now consider the results of the separate tests of the joint hypothesis. The
significance of the ai,h’s in the joint regressions (Eq. 43.2) generally deteriorates
1208 R. Cohen et al.

Table 43.1 Joint unbiasedness tests (1-month forecasts)

Individual regressions
st + h  st ¼ ai,h + bi,h(sei;t;h  st) + ei,t,h for h ¼ 2 (43.2)
Degrees of freedom ¼ 260
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Life insurance and
brokers trading companies Export industries import companies
ai,h 0.003 0.004 0.007 0.006
t (NW) 1.123 1.428 2.732 1.903
p-value 0.262 0.153 0.006 0.057
bi,h 0.437 0.289 0.318 0.008
t (NW) 1.674 1.382 1.237 0.038
R2 0.014 0.008 0.007 0.000
H0 : bi,h ¼ 1, for i ¼ 1,2,3,4
w2 4.666 4.666 4.666 4.666
p-value 0.031 0.001 0.000 0.000
Unbiasedness tests: H0 : ai,h ¼ 0, bi,h ¼ 1, for i ¼ 1,2,3,4
w2(NW) 4.696 11.682 29.546 19.561
p-value 0.096 0.003 0.000 0.000
MH tests H0 : ai,h ¼ aj,h, bi,h ¼ bj,h for all i, j 6¼ i
w2(GMM) 9.689
p-value 0.138
See Appendix 1 for structure of GMM variance-covariance matrix

with horizon. There are only two rejections at the 5 % level for each of the two
shorter horizons. However, the ai,h’s are all rejected at the 6.7 % significance level
for the 6-month horizon. The test results for the bi,h’s follow the opposite pattern
with respect to horizon. The null that bi,h ¼ 1 is rejected for all groups at the
1-month horizon, but only for the exporters at the 3-month horizon. There are no
rejections at the 6-month horizon. Thus, it appears that the pattern of rejection of the
joint hypothesis is predominantly influenced by tests of whether the slope coeffi-
cient equals one. In particular, tests of the joint hypothesis at the 1-month horizon
are rejected due to failure of this type of weak efficiency, not simple unbiasedness.
For this reason, Mincer and Zarnowitz (1969) and Holden and Peel (1990)
suggest that, if one begins by testing the joint hypothesis, rejections in this first
stage should be followed by tests of the simple unbiasedness hypothesis in a second
stage. Only if unbiasedness is rejected in this second stage should one conclude that
forecasts are biased. For reasons described below (in Sect. 43.4.2), our treatment
eliminates the first stage, so that unbiasedness and weak efficiency are separately
assessed using the forecast error as the dependent variable.
Finding greater efficiency at the longer horizon is unusual, because forecasting
difficulty is usually thought to increase with horizon. However, the longer horizon
result may not be as conclusive as the bi,h statistics suggest. For all tests at all
horizons, in only one case can the null hypothesis that bi,h equals zero not be rejected.
Thus, for the longer two horizons (with just the one exception for exporters at the
43 Rationality and Heterogeneity of Survey Forecasts 1209

Table 43.2 Joint unbiasedness tests (3-month forecasts)

Individual regressions
st + h  st ¼ ai,h + bi,h(sei;t;h  st) + ei,t,h for h ¼ 6 (43.2)
Degrees of freedom ¼ 256
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Life insurance and
brokers trading companies Export industries import companies
ai,h 0.013 0.011 0.017 0.014
t (NW) 1.537 1.362 2.060 1.517
p-value 0.124 0.173 0.039 0.129
bi,h 0.521 0.611 0.082 0.484
t (NW) 1.268 1.868 0.215 1.231
R2 0.018 0.026 0.001 0.016
H0 : bi,h ¼ 1, for i ¼ 1, 2, 3, 4
w2 1.362 1.415 5.822 1.728
p-value 0.243 0.234 0.016 0.189
Unbiasedness tests: H0 : ai,h ¼ 0, bi,h ¼ 1, for i ¼ 1, 2, 3, 4
w2(NW) 2.946 2.691 11.156 3.023
p-value 0.229 0.260 0.004 0.221
MH tests H0 : ai,h ¼ aj,h, bi,h ¼ bj,h for all i, j 6¼ i
w2(GMM) 5.783
p-value 0.448
See Appendix 1 for structure of GMM variance-covariance matrix

3-month horizon), hypothesis testing cannot distinguish between the null hypotheses
that bi,h equals one or zero. Therefore, we cannot conclude that weak efficiency with
respect to the current forecast holds while unbiasedness may not. The failure to
precisely estimate the slope coefficient also produces R2s that are below 0.05 in all
regressions.12 The conclusion is that testing only the joint hypothesis has the potential
to obscure the difference in performance between the unbiasedness and weak effi-
ciency tests. This conclusion is reinforced by an examination of Figs. 43.5, 43.6, and
43.7, the scatterplots, and regression lines for the bivariate regressions.13 All three
scatterplots have a strong vertical orientation. With this type of data, it is easy to find
the vertical midpoint and test whether it is different from zero. Thus, (one-parameter)
tests of simple unbiasedness are feasible. However, it is difficult to fit a precisely

12
As we report in Sect. 43.5, this lack of power is at least consistent with the failure to reject micro-
homogeneity at all three horizons.
13
Note that, for illustrative purposes only, we compute the expectational variable as the four-group
average percentage change in the forecast. However, recall that, despite the failure to reject micro-
homogeneity at any horizon, the Figlewski-Wachtel critique implies that these parameter esti-
mates are inconsistent in the presence of private information. (See the last paragraph in this
subsection.)
1210 R. Cohen et al.

Table 43.3 Joint unbiasedness tests (6-month forecasts)

Individual regressions
st + h  st ¼ ai,h + bi,h(sei;t;h  st) + ei,t,h for h ¼ 12 (43.2)
Degrees of freedom ¼ 256
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Life insurance and
brokers trading companies Export industries import companies
ai,h 0.032 0.032 0.039 0.034
t (NW) 1.879 1.831 2.099 1.957
p-value 0.060 0.067 0.036 0.050
bi,h 0.413 0.822 0.460 0.399
t (NW) 0.761 1.529 0.911 0.168
R2 0.01 0.044 0.021 0.012
H0 : bi,h ¼ 1, for i ¼ 1, 2, 3, 4
w2 1.166 0.110 1.147 1.564
p-value 0.280 0.740 0.284 0.211
Unbiasedness tests: H0 : ai,h ¼ 0, bi,h ¼ 1, for i ¼ 1, 2, 3, 4
w2(NW) 4.332 3.5 7.899 5.006
p-value 0.115 0.174 0.019 0.082
MH tests H0 : ai,h ¼ aj,h, bi,h ¼ bj,h for all i, j 6¼ i
w2(GMM) 7.071
p-value 0.314
See Appendix 1 for structure of GMM variance-covariance matrix

estimated regression line to this scatter, because the small variation in the forecast
variable inflates the standard error of the slope coefficient. This explains why the
bi,h’s are so imprecisely estimated that the null hypotheses that bi,h ¼ 1 and 0 are
simultaneously not rejected. This also explains why the R2s are so low. Thus,
examination of the scatterplots also reveal why bivariate regressions are potentially
misleading about weak efficiency as well as simple unbiasedness. Therefore, in
contrast to both Mincer and Zarnowitz (1969) and Holden and Peel (1990), we prefer
to separate tests for unbiasedness from tests for (all types of) weak efficiency at the
initial stage. This obviates the need for a joint test. In the next section, we conduct
such tests, making use of cointegration between forecast and realization where
it exists.14
More fundamentally, the relatively vertical scatter of the regression observations
around the origin is consistent with an approximately unbiased forecast of a random

14
However, in the general case of biased and/or inefficient forecasts, Mincer and Zarnowitz (1969,
p. 11) also viewed the bivariate regression ‘as a method of correcting the forecasts . . . to improve
[their] accuracy . . . Theil (1966, p. 33) called it the “optimal linear correction.”’ That is, the
correction would involve (1) subtracting ai,h and then (2) multiplying by 1/bi,h. Graphically, this is
a translation of the regression line followed by a rotation, until the regression line coincides with
the 45 line.
43 Rationality and Heterogeneity of Survey Forecasts 1211

Fig. 43.5 Actual versus st+h − st

expected depreciation, 0.1
1-month-ahead forecast

0.05

−0.05

−0.1

−0.15
−0.15 −0.1 −0.05 0 0.05 0.1
s ei,t,h−st

st+h − st
0.2

0.15

0.1

0.05

−0.05

−0.1

−0.15

−0.2
Fig. 43.6 Actual versus −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
expected depreciation, s ei,t,h−st
3-month-ahead forecast
1212 R. Cohen et al.

Fig. 43.7 Actual versus st+h − st

expected depreciation: 0.3
6-month-ahead forecast

0.2

0.1

−0.1

−0.2

−0.3
−0.3 −0.2 −0.1 0 0.1 0.2 0.3
s ei,t,h−st

walk-in exchange rate levels.15 In Figs. 43.8, 43.9, and 43.10, we observe
a corresponding time series pattern of variation between the forecasts and realiza-
tions in return form. As Bryant lamented in reporting corresponding regressions
using a shorter sample from the JCIF, “the regression. . .is. . .not one to send home
proudly to grandmother” (Bryant 1995, p. 51). He drew the conclusion that “ana-
lysts should have little confidence in a model specification [e.g., uncovered interest
parity] setting [the average forecast] exactly equal to the next-period value of the
model. . .[M]odel-consistent expectations. . .presume a type of forward-looking
behavior [e.g., weak efficiency] that is not consistent with survey data on
expectations” (Bryant 1995, p. 40).

43.4.2 Pretests for Rationality: The Stationarity of the Forecast Error

To test the null hypothesis of a unit root, we estimate the augmented Dickey-Fuller
(1979) (ADF) regression
X p
Dytþ1 ¼ a þ byt þ gt þ yk Dytþ1k þ etþ1 (43.5)
k¼1

15
Other researchers (e.g., Bryant 1995) have found similar vertical scatters for regressions where
the independent variable, e.g., the forward premium/discount ft,h  st, the “exchange risk pre-
mium” ft,h  st+h, or the difference between domestic and foreign interest rates (i  i*), exhibits
little variation.
43 Rationality and Heterogeneity of Survey Forecasts 1213

Fig. 43.8 Actual and 0.1

expected depreciation,
1-month-ahead forecast 0.08

0.06

0.04

0.02

−0.02

−0.04

−0.06

−0.08 sei,t,h − st
st+h − st
−0.1
86/05/14 91/01/16 96/10/15

0.2

0.15

0.1

0.05

−0.05

−0.1

−0.15
sei,t,h − st
Fig. 43.9 Actual and st+h − st
expected depreciation, −0.2
3-month-ahead forecast 86/05/14 91/01/16 96/07/30

where y is the level and first difference of the spot exchange rate, the level and first
difference of each group forecast, the residual from the (unrestricted) cointegrating
regression, and the forecast error (i.e., the residual from the “restricted”
cointegrating equation). The number of lagged differences to include in Eq. 43.5
1214 R. Cohen et al.

Fig. 43.10 Actual and 0.25

expected depreciation,
6-month-ahead forecast 0.2

0.15

0.1

0.05

−0.05

−0.1

−0.15

−0.2 sei,t,h − st
st+h − st
−0.25
86/05/14 90/11/27 96/05/15

is chosen by adding lags until a Lagrange multiplier test fails to reject the null
hypothesis of no serial correlation (up to lag 12). We test the null hypothesis of
a unit root (i.e., b ¼ 0) with the ADF t and z tests. We also test the joint null
hypothesis of a unit root and no linear trend (i.e., b ¼ 0 and g ¼ 0).
As can be seen in Tables 43.4, 43.5, and 43.6, we fail to reject the null of
a unit root in the log of the spot rate in two of the three unit root tests (the
exception being the joint null), but we reject the unit root in the h th difference for
all three horizons. We conclude that the log of the spot rate is integrated of order
one. Similarly, we conclude that the log of the forecast of each spot rate is
integrated of order one. Thus, we can conduct cointegration tests on the
spot rate and each corresponding forecast. The null of a unit root in the
(unrestricted) residual in the “cointegrating regression” is rejected at the 10 %
level or less for all groups and horizons except group three (exporters) at the
6-month horizon. Thus, we can immediately reject unbiasedness for the latter
group and horizon. Next, since a stationary forecast error is a necessary condition
for unbiasedness, we test for unbiasedness (as well as) and weak efficiency in
levels using Liu and Maddala’s (1992) method of “restricted cointegration.” This
specification imposes the joint restriction ai,h ¼ 0, bi,h ¼ 1 on the bivariate
regression

stþh ¼ ai, h þ bi, h sei, t, h þ ei, t, h (43.6)

and tests whether the residual (the forecast error) is nonstationary. In a bivariate
regression, any cointegrating vector is unique. Therefore, if we find that the forecast
43 Rationality and Heterogeneity of Survey Forecasts 1215

Table 43.4 Unit root tests (1-month forecasts h ¼ 2)

Dyt ¼ a + byt1 + gt + ∑pk¼1 ytDytk + et for h ¼ 2 (43.5)
Lags ADF t test ADF z test Joint test
Log of spot rate (n ¼ 276) 0 2.828 9.660 6.820**
Hth difference log spot rate 12 4.306*** 167.473*** 9.311***
Group 1 Banks and brokers
Log of forecast 0 2.274* 5.072 6.327**
Hth difference log forecast 0 7.752*** 96.639*** 30.085***
Forecast error
Restricted CI eq. 1 11.325*** 249.389*** 64.676***
Unrestricted CI eq. 1 65.581***
Group 2 Insurance and trading companies
Log of forecast 0 2.735* 5.149 6.705***
Hth difference log forecast 0 7.895*** 94.986*** 31.252***
Forecast error
Restricted CI eq. 1 11.624*** 270.302*** 68.053***
Unrestricted CI eq. 1 11.750***
Group 3 Export industries
Log of forecast 1 2.372 4.806 5.045**
Hth difference log forecast 0 8.346*** 111.632*** 34.889***
Forecast error
Restricted CI eq. 1 10.324*** 211.475*** 53.757***
Unrestricted CI eq. 1 10.392***
Group 4 Life insurance and import companies
Log of forecast 0 2.726* 5.009 6.438**
Hth difference log forecast 1 5.216*** 52.911*** 3.630***
Forecast error
Restricted CI eq. 1 10.977*** 231.837*** 60.820***
Unrestricted CI eq. 1 10.979***
*
Rejection at 10 % level
**
Rejection at 5 % level
***
Rejection at 1 % level

errors are stationary, then the joint restriction is not rejected, and (0, 1) must be the
unique cointegrating vector.16 The advantage of the one-step restricted
cointegration is that if the joint hypothesis is true, then tests which impose this
cointegrating vector have greater power than those which estimate a cointegrating
vector. See, e.g., Maynard and Phillips (2001).
Note that the Holden and Peel (1990) critique does not apply in the I(1) case,
because the intercept cannot be an unbiased forecast of a nonstationary variable.
Thus, the cointegrating regression line of the level realization on the level forecast

16
It is also possible to estimate the cointegrating parameters and jointly test whether they are zero
and one. A variety of methods, such as those due to Saikkonen (1991) or Phillips and Hansen
(1990), exist that allow for inference in cointegrated bivariate regressions.
1216 R. Cohen et al.

Table 43.5 Unit root tests (3-month forecasts h ¼ 6)

Dyt ¼ a + byt  1 + gt + ∑pk¼1 ytDyt  k + et for h ¼ 6 (43.5)
Lags ADF t test ADF z test Joint test
Log of spot rate (n ¼ 276) 0 2.828 9.660 6.820**
Hth difference log spot rate 2 4.760*** 49.769*** 11.351***
Group 1 Banks and brokers
Log of forecast 0 2.840* 4.852 7.610***
Hth difference log forecast 0 5.092*** 48.707*** 12.990***
Forecast error
Restricted CI eq. 6 3.022** 29.429*** 4.673**
Unrestricted CI eq. 6 3.343*
Group 2 Insurance and trading companies
Log of forecast 0 2.778* 4.533 8.858***
Hth difference log forecast 0 6.514*** 71.931*** 21.588***
Forecast error
Restricted CI eq. 6 3.068** 31.038*** 4.956**
Unrestricted CI eq. 2 4.539***
Group 3 Export industries
Log of forecast 0 3.105** 4.549 9.090***
Hth difference log forecast 1 4.677*** 41.524*** 10.944***
Forecast error
Restricted CI eq. 6 3.317** 31.207*** 5.659**
Unrestricted CI eq. 5 5.115***
Group 4 Life insurance and import companies
Log of forecast 0 2.863* 4.400 8.161***
Hth difference log forecast 1 4.324*** 39.870*** 9.352***
Forecast error
Restricted CI eq. 5 4.825*** 118.586*** 11.679***
Unrestricted CI eq. 4 5.123***
*
Rejection at 10 % level
**
Rejection at 5 % level
***
Rejection at 1 % level

must have both a ¼ 0 and b ¼ 1 for unbiasedness to hold. This differs from
Fig. 43.3, the scatterplot in differences, where ai,h ¼ 0 but bi,h 6¼ 1. Intuitively,
the reason for the difference in results is that the scatterplot in levels must lie in the
first quadrant, i.e., no negative values of the forecast or realization.
At the 1-month horizon, the null of a unit root in the residual of the restricted
cointegrating regression (i.e., the forecast error) is rejected at the 1 % level for all
groups. We find nearly identical results at the 3-month horizon; the null of a unit
root in the forecast error is rejected at the 5 % level for all groups. Thus, for these
regressions we can conduct rationality tests by regressing the forecast error on
a constant (hypothesized equal to zero for unbiasedness) and other information set
variables (whose coefficients are hypothesized equal to zero for efficiency). (Recall
just above that we failed to reject the null of a unit root in the unrestricted residual
43 Rationality and Heterogeneity of Survey Forecasts 1217

Table 43.6 Unit root tests (6-month forecasts h ¼ 12)

Dyt ¼ a + byt  1 + gt + ∑pk¼1 ytDyt  k + et for h ¼ 12 (43.5)
Lags ADF t test ADF z test Joint test
Log of spot rate (n ¼ 276) 0 2.828 9.660 6.820**
Hth difference log spot rate 17 3.189** 26.210*** 5.500**
Group 1 Banks and brokers
Log of forecast 0 2.947** 4.254 9.131***
Hth difference log forecast 0 4.772*** 44.018*** 11.389***
Forecast error
Restricted CI eq. 1 2.373 13.577* 2.947
Unrestricted CI eq. 7 3.285**
Group 2 Insurance and trading companies
Log of forecast 0 2.933** 4.004 9.531***
Hth difference log forecast 0 6.007*** 64.923*** 18.044***
Forecast error
Restricted CI eq. 1 2.114 11.464* 2.399
Unrestricted CI eq. 1 2.684*
Group 3 Export industries
Log of forecast 0 3.246** 4.059 10.704***
Hth difference log forecast 12 4.961*** 44.532*** 12.331***
Forecast error
Restricted CI eq. 12 1.515 5.601 1.466
Unrestricted CI eq. 0 2.931
Group 4 Life insurance and import companies
Log of forecast 0 3.133** 4.196 9.549***
Hth difference log forecast 0 4.795*** 44.062*** 11.537***
Forecast error
Restricted CI eq. 2 2.508 14.535** 3.148
Unrestricted CI eq. 1 2.851*
*
Rejection at 10 % level
**
Rejection at 5 % level
***
Rejection at 1 % level

for the 6-month forecasts of exporters.) Now, in the case of the restricted residual,
the other three groups failed to reject a unit root at the 10 % level in two out of three
of the unit root tests.17 (See Figs. 43.11, 43.12, and 43.13.) Thus, in contrast to the
results for the two shorter horizons, at the 6-month horizon, the evidence is clearly
in favor of a unit root in the forecast error for all four groups. Therefore, we reject
the null of simple unbiasedness because a forecast error with a unit root cannot be
mean zero. In fact, given our finding of a unit root in the forecast errors, rationality
tests regressing the forecast error on a constant and/or other information set
variables would be invalid.

17
As expected, exporters failed to reject at the 10 % level in all three tests.
1218 R. Cohen et al.

Fig. 43.11 Actual versus st+h

expected exchange rate: 260
1-month-ahead forecast
240

220

200

180

160

140

120

100

80
80 100 120 140 160 180 200 220 240 260
sei,t,h

43.4.3 Univariate Tests for Unbiasedness

The unbiasedness equation is specified as

i, t, h ¼ stþh  sei, t, h ¼ ai, h þ ei, t, h , (43.7)

where i,t,h is the forecast error of individual i, for an h-period-ahead forecast made
at time t. The results are reported in Tables 43.7 and 43.8. For the 1-month horizon,
unbiasedness cannot be rejected at conventional significance levels for any group.
For the 3-month horizon, unbiasedness is rejected only for exporters (at a p-value of
0.03). As we saw in the previous subsection, rationality is rejected for all groups at
the 6-month horizon, due to nonstationary forecast errors.18
In these unbiasedness tests, as well as all others, it is possible that coefficient
estimates for the entire sample are not stable over subsamples. The lower panels of
Tables 43.7 and 43.8 contain results of the test for equality of intercepts in four
equal subperiods, each consisting of approximately 75 biweekly forecasts:

i, t, h ¼ stþh  sei, t, h ¼ ai, h, 1 þ ai, h, 2 þ ai, h, 3 þ ai, h, 4 þ ei, t, h : (43.8)

18
The direction of the bias for exporters is negative; that is, they systematically underestimate the
value of the yen, relative to the dollar. Ito (1990) found the same tendency using only the first two
years of survey data (1985–1987). He characterized this depreciation bias as a type of “wishful
thinking” on the part of exporters.
43 Rationality and Heterogeneity of Survey Forecasts 1219

Fig. 43.12 Actual versus st+h

expected exchange rate: 260
3-month-ahead forecast
240

220

200

180

160

140

120

100

80
80 100 120 140 160 180 200 220 240 260
sei,t,h

st+h
260

240

220

200

180

160

140

120

100

Fig. 43.13 Actual versus 80

80 100 120 140 160 180 200 220 240 260
expected exchange rate:
sei,t,h
6-month-ahead forecast
1220 R. Cohen et al.

Table 43.7 Simple unbiasedness tests on individuals (1-month forecasts)

st + h  sei;t;h ¼ ai,h + ei,t,h (43.7)
h ¼ 2, degrees of freedom ¼ 261
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Export Life insurance and
brokers trading companies industries import companies
ai,h 0.000 0.001 0.002 0.002
t (NW) 0.115 0.524 0.809 0.720
p-value 0.909 0.600 0.418 0.472
MH tests H0 : ai,h ¼ aj,h, for all i, j 6¼ i
w2(GMM) 41.643 p-value 0.000
st + h  sei;t;h ¼ ai,h,1 + ai,h,2 + ai,h,3 + ai,h,4 + ei,t,h (43.8)
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Export Life insurance and
1985:05:29–1988.03:16 brokers trading companies industries import companies
ai,h,1 0.007 0.005 0.013 0.003
p-value 0.213 0.353 0.017 0.573
1988:03:30–1991:01:16
ai,h,2 0.009 0.008 0.010 0.010
p-value 0.134 0.164 0.090 0.114
1991:01:29–1993:11:16
ai,h,3 0.002 0.001 0.004 0.001
p-value 0.598 0.810 0.374 0.757
1993:11:30–1996:10:15
ai,h,4 0.002 0.004 0.001 0.003
p-value 0.675 0.433 0.814 0.519
Structural break tests H0 : ai,h,1 ¼ ai,h,2 ¼ · · · ¼ ai,h,4
w2 4.245 3.267 8.425 2.946
p-value 0.236 0.352 0.038 0.400
See Appendix 1 for structure of GMM variance-covariance matrix

For both 1- and 3-month horizons, all four forecaster groups undervalued the yen
in the first and third subperiods. This is understandable, as both these subperiods
were characterized by overall yen appreciation. (See Fig. 43.1.) Evidently, fore-
casters underestimated the degree of appreciation. Exporters were the only group to
undervalue the yen in the last subperiod as well, although that was not one of overall
yen appreciation. This is another perspective on the “wishful thinking” of
exporters.19
The main difference between the two horizons is in the significance of the test
for structural breaks. For the 1-month horizon, the estimates of the individual break
dummies generally do not reach statistical significance, and the test for their

19
Ito (1994) conducted a similar analysis for the aggregate of all forecasters, but without an
explicit test for structural breaks.
43 Rationality and Heterogeneity of Survey Forecasts 1221

Table 43.8 Simple unbiasedness tests on individuals (3-month forecasts)

st+h  sei;t;h ¼ ai,h + ei,t,h (43.7)
h ¼ 6, degrees of freedom ¼ 257
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Export Life insurance and
brokers trading companies industries import companies
ai,h 0.01 0.008 0.019 0.01
t (NW) 1.151 0.929 2.165 1.121
p-value 0.25 0.353 0.03 0.262
MH tests H0 : ai,h ¼ aj,h, for all i, j 6¼ i
w2(GMM) 40.16 p-value 0.000
st+h  sei;t;h ¼ ai,h,1 + ai,h,2 + ai,h,3 + ai,h,4 + ei,t,h (43.8)
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and Export Life insurance and
1985:05:29–1988.02:24 brokers trading companies industries import companies
ai,h,1 0.040 0.039 0.057 0.039
p-value 0.005 0.007 0.000 0.008
1988:03:16–1990:12:11
ai,h,2 0.023 0.025 0.022 0.021
p-value 0.169 0.111 0.187 0.229
1990:12:25–1993:09:28
ai,h,3 0.020 0.018 0.029 0.020
p-value 0.064 0.125 0.023 0.094
1993:10:12–1996:0730
ai,h,4 0.000 0.001 0.013 0.001
p-value 0.994 0.941 0.466 0.970
Structural break test H0 : ai,h,1 ¼ ai,h,2 ¼ · · · ¼ ai,h,4
w2 9.319 9.925 13.291 7.987
p-value 0.025 0.019 0.004 0.046
See Appendix 1 for structure of GMM variance-covariance matrix

equality rejects only for the exporters. Thus, the exporters’ bias was not constant
throughout the sample. In contrast, for the 3-month horizon, the test for no struc-
tural breaks is rejected at the 5 % level for all groups, even though unbiasedness
itself is rejected for the full sample only for exporters. Even setting aside the bias
and variability of exporters’ forecasts, our structural break tests allow us to con-
clude that there is considerably more variation around roughly zero mean forecast
errors at the longer horizon. This probably reflects the additional uncertainty
inherent in longer-term forecasts.20

20
This is consistent with the finding of nonstationary forecast errors for all groups at the 6-month
horizon.
1222 R. Cohen et al.

43.4.4 Unbiasedness Tests Using Error Correction Models

As mentioned at the beginning of the previous subsection, the Error Correction

Model provides an alternate specification for representing the relationship between
cointegrated variables:



stþh  st ¼ ai, h st  gi, h sei, th, h þ bi, h sei, t, h  sei, th, h

 (43.9)
þ di ðlags of stþh  st Þ þ i lags of sei, t, h  sei, th, h þ ei, t, h

According to this specification of the ECM, the change in the spot rate is
a function of the change in the forecast, interpreted as a short-run effect, and the
current forecast error, interpreted as a long-run adjustment to past disequilibria. ai,h,
the coefficient of the error correction term, represents the fraction of the forecast
error observed at t-h that is corrected by time t. A negative coefficient indicates
a stabilizing adjustment of expectations. This formulation of the ECM has the
advantage that the misspecification (due to omitted variable bias) of the regression
of the differenced future spot rate on the differenced current forecast can be gauged
by the statistical significance of the error correction term.21
The regressors include the smallest number of lagged dependent variables
required such that we do not reject the hypothesis that the residuals are white
noise. We impose gi,h ¼ 1 when “restricted” cointegration of st + h and sei;t;h is not
rejected. Recall that 1- and 3-month forecast errors were found to be stationary, so it
was for these two horizons that estimation of the simple unbiasedness equation was
possible. Although it would be valid to estimate the ECM at the 6-month horizon
using the (unrestricted) stationary cointegrating residual (i.e., for all groups but
exporters), we elect not to, because the nonstationarity of the forecast error itself
implies a failure of the unbiasedness restrictions.22
Then, as first asserted by Hakkio and Rush (1989), the unbiasedness restriction is
represented by the joint hypothesis that  ai,h ¼ bi,h ¼ 1 and all d and  coefficients
equal zero.23 (The hypothesized coefficient on the error correction term of 1

21
Zacharatos and Sutcliffe (2002) note that the inclusion of the contemporaneous spot forecast
(in their paper, the forward rate) as a regressor assumes that the latter is weakly exogenous; that is,
deviations from unbiasedness are corrected only by movements in the realized spot rate. These
authors prefer a bivariate ECM specification, in which the change in the future spot rate and the
change in the contemporaneous forecast are functions of an error correction term and lags of the
dependent variables. However, Zivot (2000) points out that if the spot rate and forecast are
contemporaneously correlated, then our single-equation specification does not make any assump-
tions about the weak exogeneity of the forecast.
22
Our empirical specification of the ECM also includes an intercept. This will help us to determine
whether there are structural breaks in the ECM.
23
Since we include an intercept, we also test the restriction that the intercept equals zero – both
individually and as part of the joint unbiasedness hypothesis.
43 Rationality and Heterogeneity of Survey Forecasts 1223

reflects the unbiasedness requirement that the entire forecast error is corrected
within the forecast horizon h.) We also test unbiasedness without including lagged
dependent variables but incorporating robust standard errors which allow for
generalized serial correlation and heteroscedasticity. This allows comparison with
the univariate and bivariate unbiasedness equations.
First, we compare the ECM results to the joint unbiasedness restrictions in the
change regressions, using robust standard errors in both cases. Although the esti-
mated coefficient of the error correction term is generally negative, indicating a stable
error correction mechanism,24 the coefficient does not reach a 5 % significance level
in any of the regressions. Thus, there is little evidence that the error correction term
plays a significant role in the long-run dynamics of exchange rate changes. The ECM
test results are nearly identical to the joint unbiasedness test results in Table 43.9. In
both specifications, unbiasedness is rejected for three of four groups at the 1-month
horizon and not rejected for three of four groups at the 3-month horizon. However,
even though the EC term is not (individually) significant in the ECMs, it does provide
explanatory power, relative to the joint unbiasedness specification. The R2s in the
ECM, while never more than 0.044, still are greater than in the joint unbiasedness
specification, typically by factors of three to five (Tables 43.10, 43.11, and 43.12).
Second, we compare the ECM results to the univariate simple unbiasedness
regressions, again using robust standard errors in both cases. The ECM unbiased-
ness restrictions are rejected at a 5 % level more often than in the simple unbiased-
ness tests. Whereas the only rejection of simple unbiasedness at the shorter two
horizons is for exporters at the 3-month horizon, the ECM restrictions are rejected
for three out of four groups at the 1-month horizon as well as for exporters at the
3-month horizon.
While it is uncontroversial that, for testing unbiasedness, the ECM is preferred to
the conventional bivariate specification in returns, it is not at all clear that the ECM
is preferred to the simple univariate test of unbiasedness. Can the more decisive
rejections of unbiasedness using the ECM versus the simple univariate specification
be reconciled?25
One way to proceed is to determine whether the unbiasedness restrictions
imposed on the ECM are necessary as well as sufficient, as is the case for the
simple unbiasedness test, or just sufficient, as is the case for the bivariate unbiased-
ness test. Thus, it is possible that the stronger rejections of unbiasedness in the ECM
specification are due to the implicit test of weak efficiency with respect to the
current forecast. That is, the Holden and Peel (1990) critique applies to the Hakkio
and Rush (1989) test in Eq. 43.9, as well as the joint unbiasedness test in the returns
regression. Setting bi,h, the coefficient of the contemporaneous differenced forecast,
equal to one produces an ECM in which the dependent variable is the forecast error:

24
The only exception is for exporters at the 1-month horizon.
25
The standard errors in the univariate regression are about the same as those for the ECM.
(By definition, of course, the R2s for the univariate regression equal zero.)
1224 R. Cohen et al.

Table 43.9 Error correction models (1-month forecasts)

Group 1 Banks and brokers
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
With robust standard errors (R ¼ 0.0195)
2
Coeff w (n)
2
n p-value
Constant 0.002 0.377 1 0.539
ai,h ¼ 0 0.465 2.884 1 0.089
ai,h ¼ 1 0.465 3.813 1 0.051
bi,h ¼ 1 0.491 3.847 1 0.050
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 3.910 3 0.271
With whitened residuals (R2 ¼ 0.605) Coeff F(n,d f) n p-value
Constant 0.001 0.238 1 0.627
ai,h ¼ 1 0.025 14.696 1 0.000
bi,h ¼ 1 0.453 4.895 1 0.028
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 6.790 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 51.115 12 0.000
and all lags of realizations and
forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w2(n) n p-value
Constant 0.002 0.293 1 0.582
ai,h ¼ 1 0.991 0.015 1 0.903
Constant ¼ 0 and ai,h ¼ 1 0.294 2 0.863
Group 2 Insurance and trading companies
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
With robust standard errors (R2 ¼ 0.009) Coeff w2(n) n p-value
Constant 0.003 1.168 1 0.280
ai,h ¼ 0 0.262 1.111 1 0.292
ai,h ¼ 1 0.262 8.807 1 0.003
bi,h ¼ 1 0.278 10.703 1 0.001
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 10.965 3 0.012
With whitened residuals (R2 ¼ 0.596) Coeff F(n,d f) n p-value
Constant 0.002 0.563 1 0.454
ai,h ¼ 1 0.036 13.614 1 0.000
bi,h ¼ 1 0.207 12.639 1 0.001
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 6.792 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 54.557 11 0.000
and all lags of realizations and
forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.003 0.787 1 0.3751
ai,h ¼ 1 1.026 0.113 1 0.7368
Constant ¼ 0 and ai,h ¼ 1 1.052 2 0.591
43 Rationality and Heterogeneity of Survey Forecasts 1225

Table 43.10 Error correction models (1-month forecasts)

Group 3 Export industries
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
With robust standard errors (R ¼ 0.009)
2
Coeff w (n)
2
n p-value
Constant 0.006 4.202 1 0.040
ai,h ¼ 0 0.305 1.335 1 0.248
ai,h ¼ 1 0.305 24.402 1 0.000
bi,h ¼ 1 0.256 20.516 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 27.207 3 0.000
With whitened residuals (R2 ¼ 0.602) Coeff F(n, d f) n p-value
Constant 0.002 0.632 1 0.428
ai,h ¼ 1 0.107 18.043 1 0.000
bi,h ¼ 1 0.055 17.987 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 8.455 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 55.054 10 0.000
and all lags of realizations and
forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w2(n) n p-value
Constant 0.001 0.082 1 0.7749
ai,h ¼ 1 0.887 2.321 1 0.1277
Constant ¼ 0 and ai,h ¼ 1 2.578 2 0.276
Group 4 Life Insurance and import companies
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
With robust standard errors (R2 ¼ 0.003) Coeff w2(n) n p-value
Constant 0.004 1.734 1 0.188
ai,h ¼ 0 0.066 0.083 1 0.773
ai,h ¼ 1 0.066 16.501 1 0.000
bi,h ¼ 1 0.112 16.086 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 17.071 3 0.001
With whitened residuals (R2 ¼ 0.607) Coeff F(n, d f) n p-value
Constant 0.002 0.392 1 0.532
ai,h ¼ 1 0.026 20.268 1 0.000
bi,h ¼ 1 0.226 12.020 1 0.001
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 10.794 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 57.702 11 0.000
and all lags of realizations and
forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.003 1.254 1 0.2629
ai,h ¼ 1 0.949 0.481 1 0.4879
Constant ¼ 0 and ai,h ¼ 1 1.628 2 0.443
1226 R. Cohen et al.

Table 43.11 Error correction models (3-month forecasts)

Group 1 Banks and brokers
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
Tests with robust standard errors (R ¼ 0.036)
2
Coeff w (n)
2
n p-value
Constant 0.010 1.306 1 0.253
ai,h ¼ 0 0.377 0.590 1 0.443
ai,h ¼ 1 0.377 1.604 1 0.205
bi,h ¼ 1 0.501 1.268 1 0.260
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 2.348 3 0.503
Tests with whitened residuals (R2 ¼ 0.863) Coeff F(n, d f) n p-value
Constant 0.008 4.173 1 0.044
ai,h ¼ 1 0.233 56.755 1 0.000
bi,h ¼ 1 0.178 51.113 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 28.974 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 121.851 14 0.000
and all lags of realizations and forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.006 0.556 1 0.456
ai,h ¼ 1 0.889 0.896 1 0.344
Constant ¼ 0 and ai,h ¼ 1 1.330 2 0.514
Group 2 Insurance and trading companies
st+h  st ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h (43.9)
Tests with robust standard errors (R ¼ 0.044)
2
Coeff w (n)
2
n p-value
Constant 0.008 0.874 1 0.350
ai,h ¼ 0 0.556 2.061 1 0.151
ai,h ¼ 1 0.556 1.310 1 0.252
bi,h ¼ 1 0.663 0.965 1 0.326
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 1.833 3 0.608
Tests with whitened residuals (R2 ¼ 0.844) Coeff F(n, d f) n p-value
Constant 0.005 1.400 1 0.239
ai,h ¼ 1 0.080 31.425 1 0.000
bi,h ¼ 1 0.167 40.346 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 23.551 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 148.338 10 0.000
and all lags of realizations and forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.005 0.291 1 0.589
ai,h ¼ 1 0.897 0.773 1 0.379
Constant ¼ 0 and ai,h ¼ 1 0.945 2 0.623
43 Rationality and Heterogeneity of Survey Forecasts 1227

Table 43.12 Error correction models (3-month forecasts)

Group 3 Export industries
st+h  si,t,h ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h
Tests with robust standard errors (R2 ¼ 0.026) Coeff w2(n) n p-value
Constant 0.013 2.303 0 0.129
ai,h ¼ 0 0.253 0.393 1 0.531
ai,h ¼ 1 0.253 3.422 1 0.064
bi,h ¼ 1 0.411 2.102 1 0.147
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 7.663 3 0.054
Tests with whitened residuals (R2 ¼ 0.856) Coeff F(n, d f) n p-value
Constant 0.003 0.840 1 0.361
ai,h ¼ 1 0.006 29.512 1 0.000
bi,h ¼ 1 0.205 40.582 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 16.290 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 182.912 10 0.000
and all lags of realizations and forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.012 1.971 1 0.160
ai,h ¼ 1 0.775 3.791 1 0.052
Constant ¼ 0 and ai,h ¼ 1 6.337 2 0.042
Group 4 Life insurance and import companies
st+h  si,t,h ¼ ci + ai(st  gisei;th;h ) + bi(sei;t;h  sei;th;h ) + ei,h
Tests with robust standard errors (R2 ¼ 0.038) Coeff w2(n) n p-value
Constant 0.009 0.993 1 0.319
ai,h ¼ 0 0.478 1.250 1 0.264
ai,h ¼ 1 0.478 1.488 1 0.223
bi,h ¼ 1 0.604 0.919 1 0.338
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 2.451 3 0.484
Tests with whitened residuals (R2 ¼ 0.845) Coeff F(n, d f) n p-value
Constant 0.003 0.510 1 0.477
ai,h ¼ 1 0.050 32.000 1 0.000
bi,h ¼ 1 0.062 32.469 1 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 21.673 3 0.000
Constant ¼ 0 and ai,h ¼ 1 and bi,h ¼ 1 169.286 9 0.000
and all lags of realizations and forecasts ¼ 0
st+h  sei;t;h ¼ di + (1 + ai,h)(st  sei;th;h ) (43.10)
bi,h ¼ 1 imposed, robust standard errors Coeff w (n)
2
n p-value
Constant 0.006 0.455 1 0.500
ai,h ¼ 1 0.865 1.405 1 0.236
Constant ¼ 0 and ai,h ¼ 1 1.726 2 0.422
1228 R. Cohen et al.

 

st  sei, t, h ¼ 1 þ ai, h st  sei, th, h (43.10)

Thus, in the ECM the necessary and sufficient condition for unbiasedness
is that ai,h equals 1.26 Table 43.9 contains tests of this conjecture. Here the
joint hypothesis that the intercept equals zero and ai,h equals minus one produces
exactly the same results as in the simple unbiasedness tests.27 It is interesting that,
even when we can decouple the test for weak efficiency with respect to the current
forecast from the unbiasedness test, the test of unbiasedness using this
ECM specification still requires weak efficiency with respect to the current
forecast error.28

43.4.5 Explicit Tests of Weak Efficiency

The literature on rational expectations exhibits even less consensus as to the

definition of efficiency than it does for unbiasedness. In general, an efficient
forecast incorporates all available information – private as well as public. It follows
that there should be no relationship between forecast error and any information
variables known to the forecaster at the time of the forecast. Weak efficiency
commonly denotes the orthogonality of the forecast error with respect to functions
of the target and prediction. For example, there is no contemporaneous relationship
between forecast and forecast error which could be exploited to reduce the error.
Strong efficiency denotes orthogonality with respect to the remaining variables
in the information set. Below we perform two types of weak efficiency tests.
In the first type, we regress each group’s forecast error on three sets of weak
efficiency variables29:

26
Since we estimate the restricted ECM with an intercept, unbiasedness also requires the intercept
to be equal to zero.
27
Since the intercept in Eq. 43.10 is not significant in any regression, the simple hypothesis that ai,h
equals one also fares the same as the simple unbiasedness tests.
28
For purposes of comparison with both the bivariate joint and simple unbiasedness restrictions,
we have used the ECM results using the robust standard errors. In all cases testing the ECM
restrictions using F-statistics based on whitened residuals produces rejections of all restrictions,
simple and joint, except a zero intercept. Hakkio and Rush (1989) found similarly strong rejections
of Eq. 43.9, where the forecast was the forward rate.
29
Notice that the first two sets of weak efficiency variables include the mean forecast, rather than
the individual group forecast. Our intention is to allow a given group to incorporate information
from other groups’ forecasts via the prior mean forecast. This requires an extra lag in the
information set variables, relative to a contemporaneously available variable such as the realized
exchange rate depreciation.
43 Rationality and Heterogeneity of Survey Forecasts 1229

1. Single and cumulative lags of the mean forecast error (lagged one period):

X
hþ7

stþh  sei, t, h ¼ ai, h þ bi, tþhk stþhk  sem, tþhk, h þ ei, t, h (43.11)
k¼hþ1

2. Single and cumulative lags of mean expected depreciation (lagged one period):

X
hþ7

stþh  sei, t, h ¼ ai, h þ bi, tþhk sem, tþhk, h  stk þ ei, t, h (43.12)
k¼hþ1

3. Single and cumulative lags of actual depreciation:

X
hþ6
stþh  sei, t, h ¼ ai, h þ bi, tþhk ðstþhk  stk Þ þ ei, t, h (43.13)
k¼h

For each group and forecast horizon, we regress the forecast error on the
most recent seven lags of the information set variable, both singly and
cumulatively. We use a Wald test of the null hypothesis ai,h ¼ bi,t + h  k ¼ 0 and
report chi-square test statistics, with degrees of freedom equal to the number of
regressors excluding the intercept. If we were to perform only simple regressions
(i.e., on each lag individually), estimates of coefficients and tests of significance
could be biased toward rejection due to the omission of relevant variables. If we
were to perform only multivariate regressions, tests for joint significance could be
biased toward nonrejection due to the inclusion of irrelevant variables. It is also
possible that joint tests are significant but individual tests are not. This will be the
case when the linear combination of (relatively uncorrelated) regressors spans the
space of the dependent variable, but individual regressors do not.
In the only reported efficiency tests on JCIF data, Ito (1990) separately regressed
the forecast error (average, group, and individual firm) on a single lagged forecast
error, lagged forward premium, and lagged actual change. He found that, for the
51 biweekly forecasts between May 1985 and June 1987, rejections increased from
a relative few at the 1- or 3-month horizons to virtual unanimity at the 6-month
horizon. When he added a second lagged term for actual depreciation, rejections
increased “dramatically” for all horizons.
The second type of weak efficiency tests uses the Breusch (1978)-Godfrey
(1978) LM test for the null of no serial correlation of order k ¼ h or greater, up
to order k ¼ h + 6, in the residuals of the forecast error regression, Eq. 43.11.
Specifically, we estimate

X
h1
 Xhþ6
^e i, t, h ¼ ai, h þ bi, k stþhk  sei, tk, h þ fi, l^e i, tl, h þ i, t, h (43.14)
k¼1 l¼h
1230 R. Cohen et al.

and test the null hypothesis H0 : bi,h ¼ . . . ¼ bi,h + 6 ¼ 0 for h ¼ 2,6.30,31 Results for
all efficiency tests for the 1- and 3-month horizons are presented in Tables 43.13,
43.14, 43.15, 43.16, 43.17, 43.18, 43.19, and 43.20. (Recall that the nonstationarity
of the forecast errors at the 6-month horizon is an implicit rejection of weak
efficiency.) For each group, horizon, and variable, there are seven individual
tests, i.e., on a single lag, and six joint tests, i.e., on multiple lags. These 13 tests
are multiplied by four groups times two horizons times three weak efficiency
variables for a total of 312 efficiency tests.
Using approximately nine more years of data than Ito (1990), we find many
rejections. In some cases, nearly all single lag tests are rejected, yet few, if any, joint
tests are rejected. (See, e.g., expected depreciation at the 3-month horizon.) In other
cases, nearly all joint tests are rejected, but few individual tests. (See, e.g., actual
depreciation at the 3-month horizon.) Remarkably, all but one LM test for serial
correlation at a specified lag produces a rejection at less than a 10 % level, with
most at less than a 5 % level. Thus, it appears that the generality of the alternative
hypothesis in the LM test permits it to reject at a much greater rate than the
conventional weak efficiency tests, in which the variance-covariance matrix incor-
porates the Newey-West-Bartlett correction for heteroscedasticity and serial corre-
lation. Finally, unlike Ito (1990), we find no strong pattern between horizon length
and number of rejections.

43.5 Micro-homogeneity Tests

In addition to testing the rationality hypotheses at the individual level, we are

interested in the degree of heterogeneity of coefficients across forecasters. Demon-
strating that individual forecasters differ systematically in their forecasts (and
forecast-generating processes) has implications for the market microstructure
research program. As Frankel and Froot (1990, p. 182) noted, “the tremendous
volume of foreign exchange trading is another piece of evidence that reinforces
the idea of heterogeneous expectations, since it takes differences among market
participants to explain why they trade.”
Micro-homogeneity should have implications for rationality as well. Intuitively,
if all forecasters pass rationality tests, then their corresponding regression coeffi-
cients should be equal. However, the converse is not necessarily true: if all fore-
casters have equal regression coefficients, they will not satisfy rationality
conditions if they are all biased or inefficient to the same degree with respect to

30
This is a general test, not only because it allows for an alternative hypothesis of higher-order
serial correlation of specified order but also because it allows for serial correlation to be generated
by AR, MA, or ARMA processes.
31
We use the F-statistic because the w2 test statistics tend to over-reject, while the F-tests have
more appropriate significance levels (see Kiviet 1987).
43 Rationality and Heterogeneity of Survey Forecasts 1231

Table 43.13 Weak efficiency tests (1-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(st + h  p sem;tþhp;h ) + ei,t,h for h ¼ 2 (43.11)
i¼1 i¼2 i¼3 i¼4
Insurance and trading Life insurance and
Banks and brokers companies Export industries import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
2 0.132 0.895 0.139 0.709 1.871 0.171 0.334 0.563
3 0.914 0.361 1.971 0.160 0.027 0.869 0.186 0.667
4 0.160 0.689 0.006 0.938 1.634 0.201 0.714 0.398
5 0.450 0.502 0.050 0.823 1.749 0.186 1.180 0.277
6 0.046 0.831 0.104 0.747 0.686 0.408 0.188 0.665
7 0.002 0.967 0.282 0.595 0.069 0.793 0.001 0.970
8 0.091 0.763 0.436 0.509 0.022 0.883 0.300 0.584
Cum.
3 0.765 0.682 1.778 0.411 1.746 0.418 0.585 0.746
4 4.626 0.201 3.463 0.326 8.763 0.033 5.349 0.148
5 4.747 0.314 4.382 0.357 7.680 0.104 5.081 0.279
6 5.501 0.358 5.592 0.348 7.652 0.176 5.768 0.329
7 6.252 0.396 6.065 0.416 8.879 0.180 6.677 0.352
8 5.927 0.548 5.357 0.617 8.390 0.299 6.087 0.530
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
2 122.522 0.000 6
8 43.338 0.000 6
Cum.
3 136.830 0.000 9
8 201.935 0.000 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction for
serial correlation w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept (n ¼ 1 for single lag,
n ¼ max. lag 2 for cumulative lags)

the same variables. For the univariate unbiasedness regressions, the null of micro-
homogeneity is given by H0: aih ¼ ajh, for all i, j 6¼ i. Before testing for homoge-
neous intercepts in Eq. 43.7, we must specify the form for our GMM system
variance-covariance matrix. Keane and Runkle (1990) first accounted for cross-
sectional correlation (in price level forecasts) using a GMM estimator on pooled
data. Bonham and Cohen (2001) tested the pooling specification by replacing
Zellner’s (1962) SUR variance-covariance matrix with a GMM counterpart
that incorporates the Newey-West single-equation corrections (used in our
individual equation tests above) plus allowances for corresponding
cross-covariances, both contemporaneous and lagged. Bonham and Cohen (2001)
1232 R. Cohen et al.

Table 43.14 Weak efficiency tests (1-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(sem;tþhp;h st  p) + ei,t,h for h ¼ 2 (43.12)
i¼1 i¼2 i¼3 i¼4
Insurance and Export Life insurance and
Banks and brokers trading companies industries import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
2 2.325 0.020 5.641 0.018 3.658 0.056 7.011 0.008
3 4.482 0.106 3.519 0.061 3.379 0.066 5.877 0.015
4 3.162 0.075 2.580 0.108 2.805 0.094 4.911 0.027
5 3.956 0.047 2.993 0.084 3.102 0.078 7.467 0.006
6 6.368 0.012 4.830 0.028 5.952 0.015 9.766 0.002
7 8.769 0.003 6.786 0.009 7.755 0.005 12.502 0.000
8 5.451 0.020 4.114 0.043 4.417 0.036 7.564 0.006
Cum.
3 5.592 0.061 6.138 0.046 4.116 0.128 7.508 0.023
4 5.638 0.131 5.896 0.117 4.283 0.232 7.888 0.048
5 5.189 0.268 4.964 0.291 3.784 0.436 8.009 0.091
6 6.025 0.304 5.068 0.408 4.847 0.435 8.401 0.136
7 7.044 0.317 5.746 0.452 5.940 0.430 9.434 0.151
8 10.093 0.183 8.494 0.291 7.919 0.340 12.530 0.084
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
2 40.462 0.000 6
8 30.739 0.000 6
Cum.
3 42.047 0.000 6
8 46.124 0.004 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction
for serial correlation w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept (n ¼ 1 for single lag,
n ¼ max. lag 2 for cumulative lags)

constructed a Wald statistic for testing the micro-homogeneity of individual

forecaster regression coefficients in a system.32
Keane and Runkle (1990) provided some empirical support for their modeling of
cross-sectional correlations, noting that the average covariance between a pair of

32
Elliott and Ito (1999) used single-equation estimation that incorporated a White correction for
heteroscedasticity and a Newey-West correction for serial correlation. (See the discussion below
of Ito’s tests of forecaster heterogeneity.)
43 Rationality and Heterogeneity of Survey Forecasts 1233

Table 43.15 Weak efficiency tests (1-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(st + h  p  st  p) + ei,t,h for h ¼ 2 (43.13)
i¼1 i¼2 i¼3 i¼4
Insurance and Export Life insurance and
Banks and brokers trading companies industries import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
2 0.328 0.743 0.639 0.424 1.249 0.264 0.023 0.879
3 1.621 0.203 3.060 0.080 0.000 0.993 0.550 0.458
4 0.002 0.964 0.335 0.562 0.819 0.366 0.146 0.702
5 0.086 0.770 0.042 0.837 1.001 0.317 0.344 0.557
6 0.165 0.685 0.916 0.339 0.029 0.864 0.095 0.758
7 0.850 0.357 1.861 0.172 0.329 0.566 1.152 0.283
8 0.597 0.440 1.088 0.297 0.317 0.574 1.280 0.258
Cum.
3 1.978 0.372 3.169 0.205 1.940 0.379 1.132 0.568
4 3.304 0.347 3.501 0.321 5.567 0.135 3.318 0.345
5 3.781 0.436 4.248 0.373 5.806 0.214 3.598 0.463
6 3.651 0.601 4.646 0.461 5.756 0.331 3.819 0.576
7 4.493 0.610 5.609 0.468 6.608 0.359 5.040 0.539
8 5.619 0.585 6.907 0.439 7.907 0.341 6.521 0.480
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
2 150.698 0.000 6
8 45.652 0.000 6
Cum.
3 161.950 0.000 9
8 214.970 0.000 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction
for serial correlation
w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept (n ¼ 1 for single lag,
n ¼ max. lag 2 for cumulative lags)

forecasters is 58 % of the average forecast variance. In contrast, we use Pesaran’s

(2004) CD (cross-sectional dependence) test to check for lagged as well as
contemporaneous correlations of forecast errors among pairs of forecasters:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2T X N 1 X
N
CD ¼ ^ ,
r (43.15)
N ðN  1Þ i¼1 j¼iþ1 ij

where T is the number of time periods, N ¼ 4 is the number of individual forecasters,

^ ij is the sample correlation coefficient between forecasters i and j, i ¼
and r 6 j.
1234 R. Cohen et al.

Table 43.16 LM test for serial correlation (1-month forecasts)

H0 : bi,h ¼ . . . ¼ bi,h + 6 ¼ 0, for h ¼ 2
Xh1
 Xhþ6
in ^e i, t, h ¼ ai, h þ b
k¼1 i, k
stþhk  sei, tk, h þ f ^e
l¼h i, l i, tl, h
þ i, t, h ,
where e is generated from
st + h  sei;t;h ¼ ai,h + ∑ k¼1
h1
bi,k(st + h  k  sei;tk;h ) + ei,t,h
Cum. lags (k) 2 3 4 5 6 7 8
nk 219 205 192 179 166 153 144
i¼1
Banks and brokers
F(k, n  k) 29.415 18.339 14.264 11.180 9.699 7.922 6.640
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000
i¼2
Insurance and trading companies
F(k, n  k) 30.952 19.506 15.372 11.661 9.695 8.120 7.050
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000
i¼3
Export industries
F(k, n  k) 32.387 20.691 16.053 12.951 10.628 9.418 7.520
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000
i¼4
Life insurance and import companies
F(k, n  k) 29.694 18.606 14.596 11.093 9.586 9.154 7.937
p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.000

Under the null hypothesis of no cross-correlation, CD a N ð0; 1Þ.33 See Table 43.21
for CD test results. We tested for cross-correlation in forecast errors from lag zero up
to lags four and eight for the 1 and 3-month forecast horizons, respectively.
(The nonstationarity of the 6-month forecast error precludes using the CD test at
that horizon.) At the 1-month horizon, cross-correlations from lags zero to four are
each significant at the 5 % level. Since rational forecasts allow for (individual)
serial correlation of forecast errors at lags of h-1 or less, and h ¼ 2 for the 1-month
horizon, the cross-correlations at lags two through four indicate violations of weak
efficiency. Similarly, at the 3-month horizon, where h-1 ¼ 5, there is significant
cross-correlation at lag six.34 However, it should be noted that, for many
lags shorter than h, one cannot reject the null hypothesis that there are no
cross-correlated forecast errors.

33
Unlike Breusch and Pagan’s (1980) LM test for cross-sectional dependence, Pesaran’s (2004)
CD test is robust to multiple breaks in slope coefficients and error variances, as long as the
unconditional means of the variables are stationary and the residuals are symmetrically distributed.
34
There are three instances of statistically significant negative test statistics for lags greater than
h-1, none for lags less than or equal to h-1. Thus, some industries produce relatively high forecast
errors several periods after others produce relative low forecast errors, and this information is not
fully incorporated in some current forecasts.
43 Rationality and Heterogeneity of Survey Forecasts 1235

Table 43.17 Weak efficiency tests (3-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(st + h  p  sem;tþhp;h ) + ei,t,h for h ¼ 6 (43.11)
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and trading Life insurance and
brokers companies Export industries import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
6 0.667 0.414 0.954 0.329 4.493 0.034 1.719 0.190
7 0.052 0.820 0.071 0.789 1.434 0.231 0.268 0.605
8 0.006 0.940 0.010 0.921 0.382 0.537 0.001 0.976
9 0.055 0.814 0.043 0.836 0.140 0.708 0.060 0.806
10 0.264 0.607 0.278 0.598 0.001 0.980 0.432 0.511
11 0.299 0.585 0.381 0.537 0.020 0.888 0.598 0.439
12 0.172 0.678 0.336 0.562 0.011 0.918 0.633 0.426
Cum.
7 8.966 0.011 11.915 0.003 19.663 0.000 12.350 0.002
8 12.288 0.006 16.263 0.001 23.290 0.000 15.146 0.002
9 11.496 0.022 15.528 0.004 22.417 0.000 14.778 0.005
10 8.382 0.136 12.136 0.033 16.839 0.005 12.014 0.035
11 11.596 0.072 18.128 0.006 23.782 0.001 15.330 0.032
12 11.527 0.117 15.983 0.025 21.626 0.003 13.038 0.071
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
6 188.738 0.000 6
12 63.364 0.000 6
Cum.
7 217.574 0.000 9
12 229.567 0.000 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction
for serial correlation
w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept
(n ¼ 1 for single lag, n ¼ max. lag 2 for cumulative lags)

Nevertheless, in our micro-homogeneity tests, we follow Bonham and Cohen

(2001), allowing for an MA(h-1) residual process, both individually and among
pairs of forecast errors. (See the Appendix 1 for details.) By more accurately
describing the panel’s residual variance-covariance structure, we expect this sys-
tems approach to improve the consistency of our estimates. Consider first the four
bivariate regressions in Tables 43.1, 43.2, and 43.3. Recall that we rejected the joint
hypothesis (ai,h, bi,h) ¼ (0, 1) at the 5 % significance level for all groups at the
1-month horizon (indicating the possible role of inefficiency with respect to the
1236 R. Cohen et al.

Table 43.18 Weak efficiency tests (3-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(sem,t + h  p,h  st  p) + ei,t,h for h ¼ 6 (43.12)
i¼1 i¼2 i¼3 i¼4
Banks and brokers Insurance and trading Export industries Life insurance and
companies import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
6 3.457 0.063 2.947 0.086 3.470 0.062 3.681 0.055
7 4.241 0.039 3.834 0.050 4.390 0.036 4.370 0.037
8 5.748 0.017 5.177 0.023 5.410 0.020 6.053 0.014
9 6.073 0.014 5.843 0.016 5.968 0.015 6.474 0.011
10 8.128 0.004 7.868 0.005 7.845 0.005 8.521 0.004
11 8.511 0.004 8.004 0.005 8.308 0.004 8.429 0.004
12 6.275 0.012 6.691 0.010 6.635 0.010 6.079 0.014
Cum.
7 4.717 0.095 4.985 0.083 4.954 0.084 4.928 0.085
8 5.733 0.125 5.209 0.157 5.045 0.168 6.736 0.081
9 5.195 0.268 5.411 0.248 5.112 0.276 6.053 0.195
10 7.333 0.197 9.245 0.100 9.456 0.092 7.872 0.163
11 8.539 0.201 6.658 0.354 7.488 0.278 7.955 0.241
12 8.758 0.271 6.747 0.456 7.796 0.351 8.698 0.275
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
6 57.130 0.000 6
12 58.230 0.000 6
Cum.
7 63.917 0.000 9
12 126.560 0.000 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction
for serial correlation
w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept
(n ¼ 1 for single lag, n ¼ max. lag 2 for cumulative lags)

current forecast), but only for the exporters at the 3- and 6-month horizons.
However, there are no rejections of micro-homogeneity for any horizon.35
The micro-homogeneity test results are very different for both the 1- and 3-month
systems of univariate unbiasedness regressions in Tables 43.7 and 43.8. (Recall that

35
The nonrejection of micro-homogeneity in bivariate regressions does not, however, mean that one
can avoid aggregation bias by using the mean forecast. Even if the bivariate regressions were
correctly interpreted as joint tests of unbiasedness and weak efficiency with respect to the current
forecast, and even if the regressions had sufficient power to reject a false null, the micro-homogeneity
tests would be subject to additional econometric problems. According to the Figlewski-Wachtel
(1983) critique, successfully passing a pretest for micro-homogeneity does not ensure that estimated
coefficients from such consensus regressions will be consistent. See Sect. 43.2.1.
43 Rationality and Heterogeneity of Survey Forecasts 1237

Table 43.19 Weak efficiency tests (3-month forecasts)

st + h  sei;t;h ¼ ai,h + ∑p¼h
hþ6
bi,t + h  p(st + h  p  st  p) + ei,t,h for h ¼ 6 (43.13)
i¼1 i¼2 i¼3 i¼4
Banks and brokers Insurance and trading Export industries Life insurance and
companies import companies
Lags w2 p-value w2 p-value w2 p-value w2 p-value
Single
6 0.268 0.604 0.450 0.502 3.657 0.056 1.065 0.302
7 0.055 0.814 0.037 0.848 0.599 0.439 0.003 0.957
8 0.331 0.565 0.305 0.581 0.029 0.864 0.230 0.632
9 0.513 0.474 0.482 0.488 0.022 0.883 0.577 0.448
10 1.038 0.308 1.077 0.299 0.318 0.573 1.344 0.246
11 1.335 0.248 1.532 0.216 0.563 0.453 1.872 0.171
12 1.184 0.276 1.620 0.203 0.616 0.433 1.979 0.159
Cum.
7 6.766 0.034 8.767 0.012 15.683 0.000 10.052 0.007
8 8.752 0.033 11.784 0.008 18.330 0.000 11.162 0.011
9 8.654 0.070 11.588 0.021 18.929 0.001 11.309 0.023
10 9.421 0.093 12.890 0.024 19.146 0.002 12.275 0.031
11 9.972 0.126 13.137 0.041 19.597 0.003 13.003 0.043
12 8.581 0.284 11.823 0.107 17.670 0.014 11.431 0.121
Selected micro-homogeneity tests
H0 : ai,h ¼ aj,h, bi,t + hp ¼ bj,t + hp for all i, j 6¼ i
w2(GMM) p-value n
Single
6 151.889 0.000 6
12 66.313 0.000 6
Cum.
7 164.216 0.000 9
12 193.021 0.000 24
See Appendix 1 for structure of GMM VCV matrix incorporating Newey-West correction
for serial correlation
w2 statistics for mean forecast error regressions (p-value underneath)
Degrees of freedom (n) represent number of regressors, excluding intercept
(n ¼ 1 for single lag, n ¼ max. lag 2 for cumulative lags)

unbiasedness was rejected for all groups at the 6-month horizon due to the
nonstationarity of the forecast error.) Despite having only one failure of unbiasedness
at the 5 % level for the two shorter horizons, micro-homogeneity is rejected at a level
of virtually zero for both horizons. The rejection of micro-homogeneity at the
1-month horizon occurs despite the failure to reject unbiasedness for any of the
industry groups. We hypothesize that the consistent rejection of micro-homogeneity
regardless of the results of individual unbiasedness tests is the result of sufficient
variation in individual bias estimates as well as precision in these estimates.
According to these tests, aggregation of individual forecasts into a mean forecast is
invalid at all horizons.
1238 R. Cohen et al.

Table 43.20 LM test for serial correlation (3-month forecasts)

H0 : bi,h ¼ . . . ¼ bi,h + 6 ¼ 0, for h ¼ 6
Xh1
 Xhþ6
in ^e i, t, h ¼ ai, h þ b
k¼1 i, k
stþhk  sei, tk, h þ f ^e
l¼h i, l i, tl, h
þ i, t, h ,
where e is generated from
Xh1

stþh  sei, t, h ¼ ai, h þ k¼1
b i, k s tþhk  s e
i, tk , h þ ei,t,h
Cum. lags (k) 6 7 8 9 10 11 12
nk 126 117 108 99 94 89 84
i¼1
Banks and brokers
F(k, n  k) 3.452 2.856 3.023 2.951 2.599 2.652 2.921
p-value 0.003 0.009 0.004 0.004 0.008 0.006 0.002
i¼2
Insurance and trading companies
F(k, n  k) 3.499 2.850 3.408 2.907 2.492 2.584 2.341
p-value 0.003 0.009 0.002 0.004 0.011 0.007 0.012
i¼3
Export industries
F(k, n  k) 4.687 3.956 4.409 3.572 2.928 2.819 2.605
p-value 0.000 0.001 0.000 0.001 0.003 0.003 0.005
i¼4
Life insurance and import companies
F(k, n  k) 2.352 2.482 2.501 2.168 1.866 1.794 1.811
p-value 0.035 0.021 0.016 0.031 0.060 0.067 0.059

In addition to testing the weak efficiency hypothesis at the individual level, we

are interested in the degree of heterogeneity of coefficients across forecasters. Here
the null of micro-homogeneity is given by H0: fil ¼ fjl, for l ¼ h, . . . h + 6, for all i,
j 6¼ i. As explained in the section on efficiency tests, there are 312 tests (not 468, due
to a nonstationary forecast error for all four groups at the 6-month horizon)/four
groups ¼ 83 micro-homogeneity tests. The null hypothesis of equal coefficients is
H0 : ai,h ¼ aj,h, bi,t+hk ¼ bj,t+hk for all i, j 6¼ i. As with the micro-homogeneity tests
for unbiasedness, our GMM variance-covariance matrix accounts for serial corre-
lation of order h-1 or less, generalized heteroscedasticity, and cross-sectional
correlation or order h-1 or less. We report w2(n) statistics, where n is the number
of coefficient restrictions, with corresponding p-values. Rather than perform all
83 micro-homogeneity tests, we choose a sample consisting of the shortest and
longest lag for which there are corresponding individual and joint tests (i.e., for the
k ¼ h + 1st and k ¼ h + 6th lag). Thus, there are four tests (two individual and two
corresponding joint tests) times two horizons times three variables for a total of
24 tests. Every one of the micro-homogeneity tests are rejected at the 0 % level. As
pointed out by Bryant (1995), a finding of micro-heterogeneity in unbiasedness and
weak efficiency tests also casts doubt on the assumption of a rational representative
agent commonly used in macroeconomic and asset-pricing models (Table 43.22).
43 Rationality and Heterogeneity of Survey Forecasts 1239

Table 43.21 CD tests for st+h  sei;t;h ¼ ai,h + ei,t,h (43.7)

cross-sectional qffiffiffiffiffiffiffiffiffiffiffiffiXN1 XN a
CD ¼ NðN1 2T
Þ
^  N ð0; 1Þ
r
i¼1 j¼iþ1 ij (43.15)
Lag length CD p-value
3-month horizon
0 31.272 0.000
1 2.461 0.014
2 0.387 0.699
3 2.322 0.020
4 1.594 0.111
h1¼5 1.461 0.144
6 5.887 0.000
7 0.340 0.734
8 1.456 0.145
N ¼ 24, T ¼ 276, r ^ ij is the sample correlation coefficient between
forecasters i and j, i 6¼ j

43.5.1 Ito’s Heterogeneity Tests

In Table 43.23, we replicate Ito’s (1990) and Elliott and Ito’s (1999) test for
forecaster “heterogeneity.” This specification regresses the deviation of the indi-
vidual forecast from the cross-sectional average forecast on a constant. Algebrai-
cally, Ito’s regression can be derived from the individual forecast error regression
by subtracting the mean forecast error regression. Thus, because it simply replaces
the forecast error with the individual deviation from the mean forecast, it does not
suffer from aggregation bias (c.f. Figlewski and Wachtel (1983)) or pooling bias
(c.f. Zarnowitz 1985) (Table 43.24).36, 37
   
sei, t, h  sem, t, h ¼ ai, h  am þ ei, t, h  em, t (43.16)

As above, we use the Newey-West-Bartlett variance-covariance matrix.

One may view Ito’s “heterogeneity” tests as complementary to our micro-
homogeneity tests. On the one hand, one is not certain whether a single (or pair
of?) individual rejection(s) of, say, the null hypothesis of a zero mean deviation in
Ito’s test would result in a rejection of micro-homogeneity overall. On the other
hand, a rejection of micro-homogeneity does not tell us which groups are the most
significant violators of the null hypothesis. It turns out that Ito’s mean deviation test
produces rejections at a level of 6 % or less for all groups at all horizons except for

36
Recall that our group results are not entirely comparable to Ito’s (1990), since our data set, unlike
his, combines insurance companies and trading companies into one group and life insurance
companies and import-oriented companies into another group.
37
Chionis and MacDonald (1997) performed an Ito-type test on individual expectations data from
Consensus Forecasts of London.
1240 R. Cohen et al.

Table 43.22 Ito tests (1-month forecasts)

Individual regressions
sei;t;h  sem;t;h ¼ (ai,h  am) + (ei,t,h  em,t) for h ¼ 2
Degrees of freedom ¼ 263
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and trading Export Life insurance and import
brokers companies industries companies
ai,h 0 0.001 0.003 0.002
t (NW) 0.173 2.316 5.471 3.965
p-value 0.863 0.021 0 0
MH tests H0 : ai,h ¼ aj,h, for all i, j 6¼ I
w2(GMM) 40.946
p-value 0

Table 43.23 Ito tests (3-month forecasts)

Individual regressions
sei;t;h  sem;t;h ¼ (ai,h  am) + (ei,t,h  em,t) for h ¼ 6
Degrees of freedom ¼ 263
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and trading Export Life insurance and import
brokers companies industries companies
ai,h 0.002 0.003 0.008 0.002
t (NW) 2.307 3.986 5.903 1.883
p-value 0.021 0 0 0.06
MH tests H0 : ai,h ¼ aj, for all i, j 6¼ I
w2(GMM) 37.704
p-value 0

Table 43.24 Ito tests (6-month forecasts)

Individual regressions
sei;t;h  sem;t;h ¼ (ai,h  am) + (ei,t,h  em,t) for h ¼ 12
Degrees of freedom ¼ 263
i¼1 i¼2 i¼3 i¼4
Banks and Insurance and trading Export Life insurance and import
brokers companies industries companies
ai,h 0.004 0.003 0.01 0
t (NW) 3.52 2.34 4.549 0.392
p-value 0 0.019 0 0.695
MH tests H0 : ai,h ¼ aj,h, for all i, j 6¼ I
w2(GMM) 23.402
p-value 0.001
43 Rationality and Heterogeneity of Survey Forecasts 1241

banks and brokers at the 1-month horizon and life insurance and import companies
at the 6-month horizon.38 Since Ito’s regressions have a similar form (though not
a similar economic interpretation) to the tests for univariate unbiasedness in
Tables 43.7 and 43.8, it is not surprising that micro-homogeneity tests on the
four-equation system of Ito equations produce rejections at a level of virtually
zero for all three horizons.

43.6 Conclusions

In this chapter, we undertake a reexamination of the rationality and diversity of

JCIF forecasts of the yen-dollar exchange rate. In several ways we update and
extend the seminal paper by Ito (1990). In particular, we have attempted to explore
the nature of rationality tests on integrated variables. We show that tests based on
the “conventional” bivariate regression in change form, while correctly specified in
terms of integration accounting, have two major shortcomings. First, following
Holden and Peel (1990), they are misspecified as unbiasedness tests, because
rejection of the (0, 1) restriction on the slope and intercept is a sufficient, not
a necessary, condition for unbiasedness. Only a zero restriction on the intercept in
a regression of the forecast error on a constant is both necessary and sufficient for
unbiasedness. Second, tests using the bivariate specification suffer from a lack of
power. Yet, this is exactly what we would expect in an asset market whose price is
a near random walk: the forecasted change is nearly unrelated to (and varies much
less than) the actual change.
In contrast, we conduct pretests for rationality based on determining whether the
realization and forecast are each integrated and cointegrated. In this case, following
Liu and Maddala (1992), a “restricted” cointegration test, which imposes a (0, 1)
restriction on the cointegrating vector, is necessary for testing unbiasedness.
(We show that the Holden and Peel (1990) critique does not apply if the regressor
and regressand are cointegrated.) If a unit root in the restricted residual is rejected,
then the univariate test which regresses the forecast error on a constant is equivalent
to the restricted cointegration test. Testing this regression for white noise residuals
is one type of weak efficiency test. Testing other stationary regressors in the
information set for zero coefficients produces additional efficiency tests.
In the univariate specification, we find that, for each group, the ability to
produce unbiased forecasts deteriorates with horizon length: no group rejects
unbiasedness at the 1-month horizon, but all groups reject at the 6-month horizon,
because the forecast errors are nonstationary. Exporters consistently perform worse
than the other industry groups, with a tendency toward depreciation bias.

38
Elliott and Ito (1999), who have access to forecasts for the 42 individual firms in the survey, find
that, for virtually the same sample period as ours, the null hypothesis of a zero deviation from the
mean forecast is rejected at the 5 % level by 17 firms at the 1-month horizon, 13 firms for the
3-month horizon, and 12 firms for the 6-month horizon. These authors do not report results by
industry group.
1242 R. Cohen et al.

Using only 2 years of data, Ito (1990) found the same result for exporters,
which he described as a type of “wishful thinking.”
The unbiasedness results are almost entirely reversed when we test the
hypothesis using the conventional bivariate specification. That is, the joint
hypothesis of zero intercept and unit slope is rejected for all groups at the
1-month horizon, but only for exporters and the 3- and 6-month horizons. Thus,
in stark contrast to the univariate unbiasedness tests, as well as Ito’s (1990)
bivariate tests, forecast performance does not deteriorate with increases in
the horizon.
Also, since Engle and Granger (1987) have showed that cointegrated variables
have an error correction representation, we impose joint “unbiasedness” restrictions
first used by Hakkio and Rush (1989) on the ECM. However, we show that these
restrictions also represent sufficient, not necessary, conditions, so these tests could
tend to over-reject. We then develop and test restrictions which are both necessary
and sufficient conditions for unbiasedness. The test results confirm that the greater
rate of rejections of the joint “unbiasedness” restrictions in the ECM is caused by
the failure of the implicit restriction of weak efficiency with respect to the lagged
forecast. When we impose the restriction that the coefficient of the forecast equals
one, the ECM unbiasedness test results mimic those of the simple univariate
unbiasedness tests. For this data set, at least, it does not appear that an ECM
provides any value added over the simple unbiasedness test. Furthermore, since
the error correction term is not statistically significant in any regressions, it is
unclear whether the ECM provides any additional insight into the long-run adjust-
ment mechanism of exchange rate changes.
The failure of more general forms of weak efficiency is borne out by two types of
explicit tests for weak efficiency. In the first type, we regress the forecast error on
single and cumulative lags of mean forecast error, mean forecasted depreciation, and
actual depreciation. We find many rejections of weak efficiency. In the second type,
we use the Godfrey (1978) LM test for serial correlation of order h through h + 6 in the
residuals of the forecast error regression. Remarkably, all but one LM test at a specified
lag length produces a rejection at less than a 10 % level, with most at less than a 5 %
level. (As in the case of the univariate unbiasedness test, all weak efficiency tests at the
6-month horizon fail due to the nonstationarity of the forecast error.)
Whereas Ito (1990) and Elliott and Ito (1999) measured diversity as
a statistically significant deviation of an individual’s forecast from the cross-
sectional average forecast, we perform a separate test of micro-homogeneity for
each type of rationality test – unbiasedness as well as weak efficiency – that we first
conducted at the industry level. In order to conduct the systems estimation and
testing required for the micro-homogeneity test, our GMM estimation and inference
make use of an innovative variance-covariance matrix that extends the Keane and
Runkle (1990) counterpart from a pooled to an SUR-type structure. Our variance-
covariance matrix takes into account not only serial correlation and heterosce-
dasticity at the individual level (via a Newey-West-Bartlett correction) but also
forecaster cross-correlation up to h-1 lags. We document the statistical significance
of the cross-sectional correlation using Pesaran’s (2004) CD test.
43 Rationality and Heterogeneity of Survey Forecasts 1243

In the univariate unbiasedness tests, we find that, irrespective of the ability to

produce unbiased forecasts at a given horizon, micro-homogeneity is rejected at
virtually a 0 % level for all horizons. We find this result to be somewhat counter-
intuitive, in light of our prior belief that micro-homogeneity would be more likely
to obtain if there were no rejections of unbiasedness. Evidently, there is sufficient
variation in the estimated bias coefficient across groups and/or high precision of
these estimates to make the micro-homogeneity test quite sensitive. Micro-
homogeneity is also strongly rejected in the weak efficiency tests.
In contrast to the results with the univariate unbiasedness specification, micro-
homogeneity is not rejected at any horizon in the bivariate regressions. We conjec-
ture that the imprecise estimation of the slope coefficient makes it difficult to reject
joint hypotheses involving this coefficient.
In conclusion, we recommend that all rationality tests be undertaken using
simple univariate specifications at the outset (rather than only if the joint bivariate
test is rejected, as suggested by Mincer and Zarnowitz (1969) and Holden and Peel
(1990) and employed by Gavin (2003)). Before conducting such tests, one should
test the restricted cointegrated regression residuals, i.e., the forecast error, for
stationarity. Clearly, integration accounting and regression specification matter
for rationality testing.
While our rationality tests do not attempt to explain cross-sectional dispersion,
the widespread rejection of micro-homogeneity in different specifications of unbi-
asedness and weak efficiency tests39 provides more motivation for the classification
of forecasters into types (e.g., fundamentalist and chartist/noise traders) than for
simply assuming a representative agent (with rational expectations).
There are characteristics of forecasts other than rationality which are of intrinsic
interest. Given our various rejections of rational expectations, it is natural to
explore what expectational mechanism the forecasters use. Ito (1994) tested the
mean JCIF forecasts for extrapolative and regressive expectations, as well as
a mixture of the two.40 Cohen and Bonham (2006) extend this analysis using
individual forecast-generating processes and additional learning model specifica-
tions. And, much of the literature on survey forecasts has analyzed the accuracy of
predictions, typically ranking forecasters by MSE. One relatively unexplored issue
is the statistical significance of the ranking, regardless of loss function. However,
other loss functions, especially nonsymmetric ones, are also reasonable. For
example, Elliott and Ito (1999) have ranked individual JCIF forecasters using
a profitability criterion. As mentioned in Sect. 43.2.2, the loss function may
incorporate strategic considerations that result in “rational bias.” Such an explora-
tion would require more disaggregated data than the JCIF industry forecasts to
which we have access.

39
We put less weight on the results of the weaker tests for micro-homogeneity in the bivariate
regressions.
40
He also included regressors for adaptive expectations and the forward premium.
1244 R. Cohen et al.

Appendix 1: Testing Micro-homogeneity with Survey Forecasts

The null hypothesis of micro-homogeneity is that the slope and intercept coeffi-
cients in the equation of interest are equal across individuals. This chapter considers
the case of individual unbiasedness regressions such as Eq. 43.2 in the text, repeated
here for convenience,


stþh  st ¼ ai, h þ bi, h sei, t, h  st þ ei, t, h (43.17)

and tests H0 : a1 ¼ a2 ¼ . . . ¼ aN and b1 ¼ b2 ¼ . . . ¼ bN.

Stack all N individual regressions into the Seemingly Unrelated Regression system

S ¼ Fy þ e (43.18)

where S is the NT  1 stacked vector of realizations, st + h, and F is an NT  2N

block diagonal data matrix:
2 3
F1
F54 ⋱ 5: (43.19)
FN

Each Fi ¼ [i sei;t;h ] is a T  2 matrix of ones and individual i’s forecasts, y ¼ [a1 b1 . . .

aN bN]0 , and e is an NT  1 vector of stacked residuals. The vector of restrictions,
Ry ¼ r, corresponding to the null hypothesis of micro-homogeneity is normally
distributed, with Ry  r  N[0, R(F0 F)1F0 OF(F0 F)1R0 ], where R is the 2(N  1)
 2N matrix
2 3
1 0 1 0 ... 0
6 0 1 0 1 0 ⋮7
R¼6
4⋮ 0
7, (43.20)
⋱ ⋱ ⋱ 0 5
0 ... 0 1 0 1

and r is a 2(N  1)  1 vector of zeros. The corresponding Wald test statistic,

0 h   i

R^
y  r RðF0 FÞ1 F OF
0
^ F0 F 1 R0 R^y  r , is asymptotically distributed as
a chi-square random variable with degrees of freedom equal to the number of
restrictions, 2(N  1).
For most surveys, there are a large number of missing observations. Keane and
Runkle (1990), Davies and Lahiri (1995), Bonham and Cohen (1995, 2001), and to
the best of our knowledge all other papers which make use of pooled regressions in tests
of the REH have dealt with the missing observations using the same approach. The
pooled or individual regression is estimated by eliminating the missing data points in
both the forecasts and the realization. The regression residuals are then padded
with zeros in place of missing observations to allow for the calculation of own and
43 Rationality and Heterogeneity of Survey Forecasts 1245

cross-covariances. As a result, many individual variances and cross-covariances are

calculated with relatively few pairs of residuals. These individual cross-covariances are
then averaged. In Keane and Runkle (1990) and Bonham and Cohen (1995, 2001) the
assumption of 2(k + 1) second moments, which are common to all forecasters, is made
for analytical tractability and for increased reliability. In contrast to the forecasts from
the Survey of Professional Forecasters used in Keane and Runkle (1990) and Bonham
and Cohen (1995, 2001), the JCIF data set contains virtually no missing observations.
As a result, it is possible to estimate each individual’s variance-covariance matrix
(and cross-covariance matrix) rather than average over all individual variances and
cross-covariance pairs as in the aforementioned papers.
We assume that for each forecast group i,
 
E ei, t, h ei, t, h ¼ s2i, 0 for all i, t,
 
E ei, t, h ei, tþk ¼ s2i, k for all i, t, k such that 0 < k  h, (43.21)
 
E ei, t, h ei, tþk ¼ 0 for all i, t, k such that k > h,

Similarly, for each pair of forecasters i and j, we assume

 
E ei, t, h ej, t ¼ di, j ð0Þ 8i, j, t,
 
E ei, t, h ej, tþk ¼ di, j ðkÞ 8i, j, t, k such that k 6¼ 0, and  h  k  h: (43.22)
 
E ei, t, h ej, tþk ¼ 0 8i, j, t, k such that k > jhj:

Thus, each pair of forecasters has a different T  T cross-covariance matrix:

2 3
di, j ð0Þ di, j ð1Þ ... di, j ðhÞ 0
6 d ð 1Þ di, j ð0Þ di, j ð1Þ ... 7
6 i, j 0 7
6 7
Pi , j ¼ 6
6 ⋮ ⋱ ⋱ ⋱ ⋮ 7 7, (43.23)
6 7
4 ... di, j ð1Þ di , j ð 0Þ di, j ð1Þ 5
0 di, j ðhÞ ... di , j ð 1Þ di , j ð 0Þ
0
Finally, note that Pi,j 6¼ Pj,i, rather Pi;j ¼ Pj,i. The complete variance-covariance
matrix, denoted O, has dimension NT  NT, with matrices Qi on the main diagonal
and Pi,j off the diagonal.
The individual Qi, variance-covariances matrices are calculated using the
Newey and West (1987) heteroscedasticity-consistent, MA(j)-corrected form. The
Pi,j matrices are estimated in an analogous manner.

References
Banerjee, A., Dolado, J. J., Galbraith, J. W., & Hendry, D. F. (1993). Co-integration, error correction
and the econometric analysis of non-stationary data. Oxford: Oxford University Press.
Batchelor, R., & Dua, P. (1990a). Forecaster ideology, forecasting technique, and the accuracy of
economic forecasts. International Journal of Forecasting, 6, 3–10.
1246 R. Cohen et al.

Batchelor, R., & Dua, P. (1990b). Product differentiation in the economic forecasting industry.
International Journal of Forecasting, 6, 311–316.
Batchelor, R., & Dua, P. (1992). Conservatism and consensus-seeking among economic fore-
casters. Journal of Forecasting, 11, 169–181.
Batchelor, R., & Peel, D. A. (1998). Rationality testing under asymmetric loss. Economics Letters,
61, 49–54.
Bonham, C. S., & Cohen, R. H. (1995). Testing the rationality of price forecasts: Comment.
American Economic Review, 85(1), 284–289.
Bonham, C. S., & Cohen, R. H. (2001). To aggregate, pool, or neither: Testing the rational-
expectations hypothesis using survey data. Journal of Business & Economic Statistics, 19,
278–291.
Boothe, P., & Glassman, D. (1987). Comparing exchange rate forecasting models. International
Journal of Forecasting, 3, 65–79.
Breusch, T. S. (1978). Testing for autocorrelation in dynamic linear models. Australian Economic
Papers, 17, 334–355.
Breusch, T. S., & Pagan, A. R. (1980). The Lagrange multiplier test and its application to model
specifications in econometrics. Review of Economic Studies, 47, 239–253.
Bryant, R. C. (1995). The “exchange risk premium,” uncovered interest parity, and the treatment
of exchange rates in multicountry macroeconomic models. Technical report, The Brookings
Institution.
Cavaglia, S. M. F. G., Verschoor, W. F. C., & Wolff, C. C. P. (1994). On the biasedness of forward
foreign exchange rates: Irrationality or risk premia? Journal of Business, 67, 321–343.
Chionis, D., & MacDonald, R. (1997). Some tests of market microstructure hypotheses in the
foreign exchange market. Journal of Multinational Financial Management, 7, 203–229.
Clements, M. P., & Hendry, D. F. (1998). Forecasting economic time series. Cambridge:
Cambridge University Press.
Cohen, R. H., & Bonham, C. S. (2006). Specifying the forecast generating process for exchange
rate survey forecasts. Working paper, Department of Economics, University of Hawaii Manoa.
Davies, A., & Lahiri, K. (1995). A new framework for testing rationality and measuring aggregate
shocks using panel data. Journal of Econometrics, 68, 205–227.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series
with a unit root. Journal of the American Statistical Association, 74, 427–431.
Ehrbeck, T., & Waldmann, R. (1996). Why are professional forecasters biased? Agency versus
behavioral explanations. The Quarterly Journal of Economics, 111, 21–40.
Elliott, G., & Ito, T. (1999). Heterogeneous expectations and tests of efficiency in the yen/dollar
forward exchange rate market. Journal of Monetary Economics, 43, 435–456.
Engle, R., & Granger, C. W. S. (1987). Cointegration and error correction representation,
estimation and testing. Econometrica, 55, 251–276.
Figlewski, S. (1978). Market “efficiency” in a market with heterogeneous information. Journal of
Political Economy, 86, 581–597.
Figlewski, S. (1982). Information diversity and market behavior. Journal of Finance, 37, 87–102.
Figlewski, S. (1984). Information diversity and market behavior: A reply. Journal of Finance, 39,
299–302.
Figlewski, S., & Wachtel, P. (1983). Rational expectations, informational efficiency, and tests
using survey data: A reply. Review of Economics and Statistics, 65, 529–531.
Frankel, J. A., & Froot, K. A. (1987). Using survey data to test standard propositions regarding
exchange rate expectations. American Economic Review, 77, 133–153.
Frankel, J. A., & Froot, K. A. (1990). Chartists, fundamentalists, and trading in the foreign
exchange market. American Economic Review, 80, 181–185.
Froot, K. A., & Frankel, J. A. (1989). Forward discount bias: Is it an exchange risk premium?
Quarterly Journal of Economics, 104, 139–161.
Gavin, W. T. (2003) FOMC forecasts: Is all the information in the central tendency? Economic
Review, Federal Reserve Bank of St. Louis, May/June, 27–46.
43 Rationality and Heterogeneity of Survey Forecasts 1247

Godfrey, L. G. (1978). Testing for higher order serial correlation in regression equations when the
regressors include lagged dependent variables. Econometrica, 46, 1303–1310.
Goldberg, M. D., & Frydman, R. (1996). Imperfect knowledge and behaviour in the foreign
exchange market. Economic Journal, 106, 869–893.
Granger, C. W. J. (1991). Developments in the study of cointegrated economic variables. In
R. F. Engle & C. W. J. Granger (Eds.), Long run economic relationships: Readings in
Cointegration (pp. 65–80). Oxford: Oxford University Press.
Hakkio, C. S., & Rush, M. (1989). Market efficiency and cointegration: An application to the
sterling deutschemark exchange markets. Journal of International Money and Finance, 8,
75–88.
Haltiwanger, J. C., & Waldman, M. (1989). Rational expectations in the aggregate. Economic
Inquiry, 27, 619–636.
Hamilton, J. (1994). Time series analysis. Princeton: Princeton University Press.
Hansen, L. P., & Hodrick, R. J. (1980). Forward exchange rates as optimal predictors of future spot
rates: An econometric analysis. Journal of Political Economy, 88, 829–853.
Holden, K., & Peel, D. A. (1990). On testing for unbiasedness and efficiency of forecasts.
Manchester School, 58, 120–127.
Ito, T. (1990). Foreign exchange rate expectations: Micro survey data. American Economic
Review, 80, 434–449.
Ito, T. (1994). Short-run and long-run expectations of the yen/dollar exchange rate. Journal of the
Japanese and International Economies, 8, 119–143.
Keane, M. P., & Runkle, D. E. (1990). Testing the rationality of price forecasts: New evidence
from panel data. American Economic Review, 80, 714–735.
Kirman, A. P. (1992). Whom or what does the representative individual represent? Journal of
Economic Perspectives, 6, 117–136.
Kiviet, J. F. (1987). Testing linear econometric models. Ph.D. Thesis, University of Amsterdam.
Lai, K. (1990). An evaluation of survey exchange rate forecasts. Economics Letters, 32, 61–65.
Lamont, O. (2002). Macroeconomic forecast and microeconomic forecasters. Journal of Economic
Behavior and Organization, 48, 265–280.
Laster, D., Bennett, P., & Geoum, I. S. (1999). Rational bias in macroeconomic forecasts.
Quarterly Journal of Economics, 293–318.
LeBaron, B. (2000) Technical trading profitability in foreign exchange markets in the 1990’s.
Technical report, Brandeis U.
Leitch, G., & Tanner, J. E. (1991). Economic forecast evaluation: Profits versus the conventional
error measures. The American Economic Review, 81, 580–590.
Liu, P., & Maddala, G. S. (1992). Rationality of survey data and tests for market efficiency in the
foreign exchange markets. Journal of International Money and Finance, 11, 366–381.
Lyons, R. K. (2002). Foreign exchange: Macro puzzles, micro tools. Economic Review, Federal
Reserve Bank of San Francisco, 51–69.
MacDonald, R. (1992). Exchange rate survey data: A disaggregated G-7 perspective. The Man-
chester School, LX(Supplement), 47–62.
Maynard, A., & Phillips, P. C. B. (2001). Rethinking and old empirical puzzle: Econometric
evidence on the forward discount anomaly. Journal of Applied Econometrics, 16, 671–708.
Mincer, J., & Zarnowitz, V. (1969). The evaluation of economic forecasts. In Economic forecasts
and expectations. New York: National Bureau of Economic Research.
Muth, J. (1961). Rational expectations and the theory of price movements. Econometrica, 29,
315–335.
Newey, W., & West, K. (1987). A simple positive semi-definite heteroscedasticity and autocor-
relation consistent covariance matrix. Econometrica, 55, 703–708.
Osterberg, W. P. (2000). New results on the rationality of survey measures of exchange-rate
expectations. Federal Reserve Bank of Cleveland Economic Review, 36, 14–21.
Pesaran, H. M. (2004). General diagnostic tests for cross section dependence in panels. CESifo
working paper series 1229, CE-Sifo Group Munich.
1248 R. Cohen et al.

Pesaran, H. M., & Weale, M. (2006). Survey expectations. In G. E. C. W. J. Granger &

A. Timmermann (Eds.), Handbook of economic forecasting. Amsterdam: North-Holland.
Phillips, P., & Hansen, B. E. (1990). Statistical inference in instrumental variables regression with
I(1) processes. Review of Economic Studies, 57, 99–125.
Phillips, P. C. B., & Maynard, A. (2001). Rethinking an old empirical puzzle: Econometric
evidence on the forward discount anomaly. Journal of Applied Econometrics, 16, 671–708.
Pilbeam, K. (1995). Exchange rate models and exchange rate expectations: an empirical investi-
gation. Applied Economics, 27, 1009–1015.
Saikkonen, P. (1991). Asymptotically efficient estimation of cointegration regressions. Economet-
ric Theory, 7, 1–21.
Stockman, A. C. (1987). Economic theory and exchange rate forecasts. International Journal of
Forecasting, 3, 3–15.
Theil, H. (1966). Applied economic forecasting. Amsterdam: North-Holland.
Zacharatos, N., & Sutcliffe, C. (2002). Is the forward rate for the Greek Drachma unbiased?
A VECM analysis with both overlapping and non-overlapping data. Journal of Financial
Management and Analysis, 15, 27–37.
Zarnowitz, V. (1985). Rational expectations and macroeconomic forecasts. Journal of Business &
Economic Statistics, 3, 293–311.
Zellner, A (1962). An efficient method of estimating seemingly unrelated regressions and tests for
aggregation bias. Journal of the American Statistical Association, June, 348–368.
Zellner, A. (1986). Biased predictors, rationality and the evaluation of forecasts. Economics
Letters, 21, 45–48.
Zivot, E. (2000). Cointegration and forward and spot exchange rate regressions. Journal of
International Money and Finance, 19, 785–812.
Stochastic Volatility Structures and
Intraday Asset Price Dynamics 44
Gerard L. Gannon

Contents
44.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1250
44.2 Stochastic Volatility and GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1252
44.3 Serial Correlation in Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1255
44.4 Persistence, Co-Persistence, and Non-Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1258
44.4.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1260
44.4.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1260
44.5 Weighted GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1260
44.6 Empirical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1261
44.6.1 Index Futures, Market Index, and Stock Price Data . . . . . . . . . . . . . . . . . . . . . . . . . 1261
44.6.2 Estimates of the Autoregressive Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1263
44.6.3 Conditional Variance Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1264
44.6.4 Weighted GARCH Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1267
44.6.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1270
44.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1273
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1274
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1275

Abstract
The behavior of financial asset price data when observed intraday is quite
different from these same processes observed from day to day and longer
sampling intervals. Volatility estimates obtained from intraday observed data
can be badly distorted if anomalies and intraday trading patterns are not
accounted for in the estimation process.
In this paper I consider conditional volatility estimators as special cases of
a general stochastic volatility structure. The theoretical asymptotic distribution
of the measurement error process for these estimators is considered for particular

G.L. Gannon
Deakin University, Burwood, VIC, Australia
e-mail: gerard@deakin.edu.au; gleonon@yahoo.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1249
DOI 10.1007/978-1-4614-7750-1_44,
# Springer Science+Business Media New York 2015
1250 G.L. Gannon

features observed in intraday financial asset price processes. Specifically,

I consider the effects of (i) induced serial correlation in returns processes,
(ii) excess kurtosis in the underlying unconditional distribution of returns, (iii)
market anomalies such as market opening and closing effects, and (iv) failure to
account for intraday trading patterns.
These issues are considered with applications in option pricing/trading strat-
egies and the constant/dynamic hedging frameworks in mind. Empirical exam-
ples are provided from transactions data sampled into 5-, 15-, 30-, and 60-min
intervals for heavily capitalized stock market, market index, and index futures
price processes.

Keywords
ARCH • Asymptotic distribution • Autoregressive parameters • Conditional var-
iance estimates • Constant/dynamic hedging • Excess kurtosis • Index futures •
Intraday returns • Market anomalies • Maximum likelihood estimates • Mis-
specification • Mis-specified returns • Persistence • Serial correlation •
Stochastic volatility • Stock/futures • Unweighted GARCH • Volatility
co-persistence

44.1 Introduction

One issue considered in Nelson (1990a) is whether it is possible to formulate an

ARCH data generation process that is similar to the true process, in the sense that
the distribution of the sample paths generated by the ARCH structure and the
underlying diffusion process becomes “close” for increasingly finer discretizations
of the observation interval. Maximum likelihood estimates are difficult to obtain
from stochastic differential equations of time-varying volatility common in the
finance literature. If the results in Nelson hold for “real-time” data when ARCH
structures approximate a diffusion process, then these ARCH structures may be
usefully employed in option pricing equations. In this paper I consider the ARCH
structure as a special case of a general stochastic volatility structure. One advantage
of an ARCH structure over a general stochastic volatility structure lies in compu-
tational simplicity. In the ARCH structure, it is not necessary for the underlying
processes to be stationary or ergodic. The crucial assumption in an option pricing
context is that these assumed processes approach a diffusion limit. These assumed
diffusion limits have been derived for processes assumed to be observed from
day-to-day records. Given that market anomalies such as market opening and
market closing effects exist, any volatility structure based on observations sampled
on a daily basis will provide different volatility estimates. Evidence of these
intraday anomalies and effects on measures of constant volatility is reported in
Edwards (1988) and Duffie et al. (1990). Brown (1990) argues that the use of
intraday data in estimating volatility within an option pricing framework leads to
volatility estimates that are too low. This can be overcome by rescaling, assuming
the anomalies are accounted for.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1251

Better estimates of volatility may then be obtained by employing intraday

observations and allowance made for anomalies and trading activity within condi-
tional volatility equations. However, mis-specifications in either or both first- and
second-moment equations may mitigate against satisfying the conditions for an
approximate diffusion limit. Then it is important to investigate cases where the
diffusion limit is not attainable and identify the factors which help explain the
behavior of the process. If these factors can be accounted for in the estimation
process then these formulations can be successfully employed in options pricing
and trading strategies.
The specific concern in this paper is the effect on the asymptotic distribution
of the measurement error process and on parameter estimates, obtained from the
Generalized ARCH (GARCH(1,1)) equations for the conditional variance, as
the observation interval approaches transactions records (d!0). Three issues are
considered for cases where the diffusion limit may not be achieved at these
observation intervals. The first issue is the effect of mis-specifying the dynamics
of the first-moment-generating equation on resultant GARCH(1,1) parameter
estimates. The second issue is the effect on measures of persistence obtained
from the GARCH structure when increasing kurtosis is induced in the underly-
ing unconditional distribution as d!0. This leads to a third issue which is
concerned with evaluating effects of inclusion of weighting (mixing) variables
on parameter estimates obtained from these GARCH(1,1) equations. If these
mixing variables are important then standard, GARCH equation estimates will
be seriously distorted. These mixing variables may proxy the level of activity
within particular markets or account for common volatility of assets trading in
the same market.
Sampling the process too finely does result in induced positive or negative serial
correlation in return processes. The main distortion to the basis change is generated
from cash index return equations. However, the dominant factor distorting
unweighted GARCH estimates is induced excess kurtosis in unconditional distri-
butions of returns. Many small price changes are dominated by occasional large
price changes. This effect leads to large jumps in the underlying distribution
causing continuity assumptions for higher derivatives of the conditional variance
function to break down.
These observations do not directly address issues related to intraday market
trading activity and possible contemporaneous volatility effects transmitted to
and from underlying financial asset price processes. This effect was considered
within the context of a structural Simultaneous Volatility (SVL) model in Gannon
(1994). Further results for the SVL model are documented, along with results of
parameter estimates, in Gannon (2010). In this paper the intraday datasets on cash
index and futures price processes from Gannon (2010) are again employed to
check the effects on parameter estimates obtained from GARCH and weighted
GARCH models. A set of intraday sampled stock prices are also employed in
this paper. The relative importance of mis-specification of the second-moment
equation dynamics over mis-specification of first-moment equation dynamics is
the most important issue. If intraday trading effects are important, this has
1252 G.L. Gannon

implications for smoothness and continuity assumptions necessary in deriving

diffusion limit results for the unweighted GARCH structure.
If these effects are severe then an implied lower sampling boundary needs to be
imposed in order to obtain sensible results. This is because the measure of persis-
tence obtained from GARCH structures may approach the Integrated GARCH
(IGARCH) boundary and become explosive or conditional heteroskedasticity
may disappear. This instability can be observed when important intraday anomalies
such as market opening and closing effects are not accounted for within the
conditional variance specifications. Distortions to parameter estimates are most
obvious when conditional second-moment equations are mis-specified by failure to
adequately account for observed intraday trading patterns. These distortions can be
observed across a wide class of financial assets and markets.
If these financial asset price processes have active exchange traded options
contracts written on these “underlying assets,” it is important to study the intraday
behavior of these underlying financial asset price processes. If systematic features
of the data series can be identified then it is possible to account for these features in
the estimation process. Then intraday estimates of volatility obtained from condi-
tional variance equations, which incorporate structural effects in first and/or con-
ditional second-moment equations, can be usefully employed in option pricing
equations. Identifying intraday anomalies and linking trading activity to contem-
poraneous volatility effects mean the improved estimator can be employed within
a trading strategy. This can involve analysis of the optimal time to buy options
within the day to minimize premium cost. Alternatively, optimal buy or sell
straddle strategies based on comparison of estimated volatility estimates relative
to market implied volatility can be investigated. In this paper I focus on theoretical
results which can explain the empirically observed behavior of these estimators
when applied to intraday financial asset price processes. I start by summarizing
relevant results which are currently available for conditional variance structures as
the observation interval reduces to daily records (h!0). These results are modified
and extended in order to accommodate intraday observation intervals. The alterna-
tive first-moment-generating equations are described and the basis change defined
and discussed within the context of the co-persistence structure. I then focus on the
general GARCH structure and state some further results for specific cases of the
GARCH and weighted GARCH (GARCH-W) structure.

44.2 Stochastic Volatility and GARCH

Nelson and Foster (1994) derive and discuss properties for the ARCH process as the
observation interval reduces to daily records (h!0) when the underlying process is
driven by an assumed continuous diffusion process. Nelson and Foster (1991)
generalized a Markov process with two state variables, hXt and hs2t , only one of
which hXt is ever directly observable. The conditional variance hs2t is defined
conditional on the increments in hXt per unit time and conditional on an information
set hBt. Modifying the notation from h to d (to account for intraday discretely
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1253

observed data), and employing the notation d!0 to indicate reduction in the
observation interval from above, when their assumptions 2, 3, and 10 hold, when
d is small, dXt, j(ds2t )) is referred to as a near diffusion if for any T, 0  T  1,
(dXt, j (ds2t ))0tT ) (Xt, j(s2t ))0tT.
If we assume these data-generating processes are near diffusions, then the
general discrete time stochastic volatility structure, defined in Nelson and Foster
(1991), may be described using the following modified notation:

2 3 2 3 2 3
d Xðkþ1Þd d Xkd mðd Xkd ; skd Þ
6 7 6 7 6 7
4
5 ¼ 4  
5 þ d :4 5
j d sð2kþ1Þd j d s2kd lðd Xkd ; d skd Þ
2 2 31 2 3
d skd Lj, x ðd Xkd ; d skd Þ 2 d Z1, kd
6 76 7
þ d1=2 4 54 5
Lj, x ðd Xkd ; d skd Þ L2 ðd Xkd ; d skd Þ Z
d 2, kd
(44:1)

where (dZ1,kd,d Z2,kd)k¼0,1 is i.i.d. with mean zero and identity covariance matrix.
In Eq. 44.1 d is the size of the observation interval, X may describe the asset price
return, and s2 the volatility of the process. It is not necessary to assume the data-
generating processes are stationary or ergodic, but the crucial assumption is that the
data-generating processes are near diffusions.
In the ARCH specification, dZ2,kd is a function of dZ1,kd so that ds2kd(dhkd) can be
inferred from past values of the one observable process dXkd. This is not true for
a general stochastic volatility structure where there are two driving noise terms.
For the first-order Markov ARCH structure, a strictly increasing function of
∧2  
estimates d st d h∧
t of the conditional variance process dst (dht) is defined as f(s ),
2 2

∧  mean
and estimates of the conditional
∧ ∧

per∧unit
 of time of the increments in X and
f(s2) are defined as m x, s and k x, s . Estimates of ds2kd are updated by the
recursion:
   

∧2 ∧2 ∧ ∧
f d sðkþ1Þd ¼ s d skd þ d k d Xkd ; d skd

∧
  ∧2
 (44:2)
∧
þ d a d Xkd ; d skd g d Z1, kd ; d Xkd , d skd
1=2

∧ ∧
where k ð:Þ, a ð:Þ, m ð:Þ, and g(.) are continuous on bounded (j(s2), x) sets and
g(z1, x.s2) assumed continuous everywhere with the first three derivatives of g with
respect to z1 well defined and bounded. The function g(dZ1,kd,.) is normalized
  to
have mean zero and unit conditional variance. Nonzero drifts in f d s2kd are
∧
allowed for in the k ð:Þ term and non-unit conditional variances accounted  for
in the a(.) term. The second term on the right measures the change in f d s2kd
forecast by the ARCH structure while the last term measures the surprise change.
1254 G.L. Gannon

∧  
The GARCH(1,1)
2  structure
 2  is obtained by setting fðs2 Þ ¼ s2 , k ðx;sÞ ¼ o  ys2 ,
gðz1 Þ ¼ z1  1 =SD z1 , and a(x,s) ¼ a . s2 . SD(z21). The parameters o, y, and a
index a family of GARCH(1,1) structures. In a similar manner the EGARCH volatility
∧
structure can be defined. That is, by selectingf ð:Þ, k ð:Þ,gð:Þ, and a(.), various “filters”
can be defined within this framework.
Other conditions in Nelson and Foster (1991) define the rate at which the
normalized measurement error process dQt mean reverts relative to (dXt,ds2t ). It
becomes white noise as d!0 on a standard time scale but operates on a faster
time scale mean-reverting more rapidly. Asymptotically optimal choice of a(.)
and f(.) given g(.) can be considered with respect to minimizing the asymptotic
variance of the measurement error. This is considered on a faster time scale
(T + t) than T. The asymptotically optimal choice of g(.) depends upon the
assumed relationship between Z1 and Z2. In the ARCH structure, Z2 is a function
of Z1 so that the level driving ds2t can be recovered from shocks driving dX2t .
Without further structure in the equation specified for s2t , we are unable to
recover information about changes in s2t . Their discussion is strictly in terms
of constructing a sequence of optimal ARCH filters which minimize the asymp-
totic variance of the asymptotic distribution of the measurement errors. This
approach is not the same as choosing an ARCH structure that minimizes the
measurement error variance for each d.
The asymptotic distribution of the measurement error process, for large t and
small d,

∧2    
s
d Tþtd 1=2  d s2
1=2
 Q ; s2
;
d T d T d TX ¼ q; s2 ; x
Tþtd

 
with derivatives evaluated as f0 d s2T , j0 (ds2T), etc. and the notation simplified as
f0 and j0 is approximately normal with mean
h∧ 00
i h ∧
i
ð2s2 j0 Þ k d =j0  a2 j =2ðj0 Þ3  ld =f0 þ L2 f0 3 þ 2 as: md  m d :E½gz 
d1=2
a : E½ Z1 : g z 
(44:3)

and variance

ð2s2 f0 Þ : ða2 =f0  þ l2 =½j0 2  2 aL:CovðZ2 ; gÞ=½s0 ½j0 

d1=2 : (44:4)
a:E½Z1 :gz 

General results in Nelson and Foster (1991, 1994) for the GARCH(1,1)
structure are that GARCH(1,1) can be more accurately measured firstly
the less variable and the smaller is ds2t , second the thinner the tails of Z1, and
third the more the true data-generating mechanism resembles an ARCH
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1255

structure as opposed to a stochastic volatility structure. If the true data-

generating process is GARCH(1,1), then Corr (Z21, Z2) ¼ 1.
As d!0, the first result will generally hold, and the second can be
checked from the unconditional distribution of the returns process. The latter
result is the most difficult to evaluate. Now reconsider some assumptions
necessary to obtain these results and reasons these assumptions may not hold
when d!0.
(i) Mis-specification of the difference between the estimated and true drift in
h ∧ i
mean, d m t  d mt , is assumed fixed as d!0 so that effects of mis-specifying
this drift has an effect that vanishes at rate d1/2 and is negligible asymptot-
ically. These terms do not appear in the expression for the variance of the
asymptotic distribution of the measurement error. As d!0, the effect of
bid/ask bounce and order splitting in futures price processes and
non-trading-induced effects on market indices becomes more severe.
Mis-specification of the drift in the mean is not constant. Whether this effect
transfers to estimates of conditional variances is an empirical issue.
(ii) The conditional variance of the increments in ds2t involves the fourth moment of
dZ1,kd so that the influence of this fourth moment remains as the diffusion limit is
approached. Excess kurtosis is a feature of intraday financial price changes.
∧
(iii) Values of k d and ld are considered fixed as d!0 so that effects of mis-specifying
the drift in s2 has an effect that also vanishes at rate d1/2. As well, although these
drift terms enter the expression for the asymptotic bias of the measurement error,
these also do not appear in the expression for the asymptotic variance. The term
gz represents part of the “surprise” change in the recursion defined in Eq. 44.2 and
is directly linked to departures from normality observed in point (ii). These
departures from normality can be generated by extremes in Z1 induced by large
jumps in the underlying distribution. In this case first and second derivatives of f
may be discontinuous throughout the sample space as well. Then the expression
for the bias in this asymptotic distribution of the measurement errors may be
explosive.
(iv) The ARCH specification of the drift in mean and variance only enters the 0p(d1/2)
terms of the measurement error. Asymptotically, the differences in the condi-
tional variance specifications are more important, appearing in the 0p(d1/4) terms.
If the conditional variance specification is not correct then the measurement error
variance is affected. This is because matching the ARCH and true variance of the
variance cannot proceed.
I will consider these issues further by generalizing a theoretical framework in
which to address each in the above order. Firstly I consider the relationship
between serial correlation in returns on the market index, the index futures, and
basis change as d!0. Second, these effects are considered in the context of
the co-persistence structure for the basis change. Finally, I consider effects on
conditional variance parameter estimates when mixing and weighting variables
are included in the equations for the conditional variance.
1256 G.L. Gannon

44.3 Serial Correlation in Returns

Consider a first-order autoregressive process for the spot asset price return
(an AR(D) representation):

st ¼ r1 st1 þ et (44:5)

where st is the return on the asset (the difference in the natural logarithm of the spot
asset price levels or the difference in the spot asset price levels) between time t and
t1, r1 is the first-order serial correlation coefficient between s at time t and t1 and
et is an assumed homoskedastic disturbance term, s2e and 1 < r1 < 1. This equation
provides an approximation to the time-series behavior of the spot asset price return.
An alternative specification to Eq. 44.5 is a simple autoregressive equation for
the level of the spot asset price S (or natural logarithm of the levels) at time t and
t1, (an AR(L) representation)

St ¼ r1 St1 þt at (44:6)

where at is an assumed homoskedastic disturbance term, s2a , and the value of f1

may be greater or less than one.
The assumed bid/ask bounce in futures price changes is approximated by a MA
(1) process in Miller et al. (1994). For the index futures price, this specification is

f t ¼ at þ y1 at1 , (44:7)

where ft is the index futures price change and at is an assumed mean zero, serially
uncorrelated shock variable with a homoskedastic variance, s2a , and 1 < y1 < 0.
The basis change is defined as

bt ¼ f t  i t , (44:8)

where f and i are the index futures return and index portfolio return, respectively.
Whether shocks generated from Eqs. 44.5 or 44.6 generate differences in
parameter estimates and measures of persistence obtained from conditional
second-moment equations is an issue. Measures of persistence for the basis change
may be badly distorted from employing index futures and index portfolio changes.
Measures of persistence for index futures and index portfolios may be badly
distorted by employing observed price changes.
The simplest GARCH structure derived from Eq. 44.1 for the conditional
variance is the GARCH(1,1):

ht ¼ o þ a1 e2t1 þ b1 ht1 (44:9)

where ht(s2t ) is the conditional variance at time t and e2t1 are squared unconditional
shocks generated from any assumed first-moment equation and 0  a1, b1  1 and
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1257

a1 + b1  1. This parameterization is a parsimonious representation of an ARCH

(p0 ) process where a geometrically declining weighting pattern on lags of e2 is
imposed. This is easily seen by successive substitution for htj(j ¼ 1, . . . , J) as
J ! 1,

   
ht ¼ o 1 þ b1 þ    þ bJ1 þ a1 e2t1 þ b1 e2t2 þ    þ bJ1 e2tJ1 þ Remainder:
(44:10)

Now consider Eq. 44.6, with f1 fixed at 1, as representing one mis-specified

spot asset price process, Eq. 44.5 representing the mis-specified
differenced autoregressive process, and Eq. 44.7 representing the
mis-specified differenced moving average process. Taking expected values,
then the unconditional variance when Eq. 44.6 is the mis-specified representa-
tion and f1 set equal to one is

   
E s2t ¼ E a2t : (44:11)

When Eq. 44.7 is the moving average (MA) representation, the unconditional
variance relative to shocks generated via Eq. 44.11 is

    
E s2t MA ¼ E 1 þ y2 a2t , (44:12)

and when Eq. 44.5 is the autoregressive (AR) representation, the unconditional
variance relative to shocks generated via Eq. 44.11 is

 

  1  r1  2 
E s2t AR ¼ E at : (44:13)
1 þ r1

The conditional variance from a GARCH(1,1) structure for Eq. 44.11 can be
rewritten as

   
ht ðsÞ ¼ o 1 þ b1 þ    þ bJ1 þ a1 a2t1 þ b1 a2t2 þ    þ bJ1 a2tJ1
þ Remainder: (44:14)

If Eq. 44.7 is the representation, then, relative to the conditional variance

equation from Eq. 44.11,
     
ht ðsÞMA ¼ o 1 þ b1 þ    þ bJ1 þ a1 1 þ y2 a2t1 þ b 1 þ y2 a2t2
  
þ   bJ1 1 þ y2 a2tJ1 þ Remainder, (44:15)
1258 G.L. Gannon

and if Eq. 44.5 is the representation, then, relative to the conditional variance
equation from Eq. 44.11,
 
ht ðsÞAR ¼ o 1 þ b1 þ    þ bJ1
      
1  r1 2 1  r1 2 1  r1 2
þ a1 at1 þ b1 at2 þ   bJ1 atJ1 þ Remainder:
1 þ r1 1 þ r1 1 þ r1
(44:16)

If o, a1, and b1 were equivalent in Eqs. 44.14, 44.15, and 44.16, when the
conditional variance is driven by Eq. 44.7 with y1 negative, then ht(s)MA > ht(s),
and when the conditional variance is driven by Eq. 44.5 with r1 negative, then
ht(s)AR > ht(s) and with r1 positive then ht(s)AR < ht(s). However, given the scaling
factor in Eq. 44.15 relative to Eq. 44.16, the potential for distortions to GARCH
parameter estimates is greater when the underlying process is driven by Eq. 44.5
relative to Eq. 44.7.

44.4 Persistence, Co-Persistence, and Non-Normality

Now define et as shocks from any of the assumed first-moment equations

from Sect. 3 with the following simplified representation of a GARCH(1,1)
structure obtained from Eq. 44.1 where ht(s2t ) represents the conditional variance
and zt(dZ1 ,kd) the stochastic part:
pffiffiffi
et ¼ ht z t zt  NIDð0; 1Þ: (44:17)

In the univariate GARCH(1,1) structure, ht converges and is strictly stationary if

E[1n(b1 + a1z2ti)] < 0. Then ∑i¼1,kd1n(b1 + a1z2ti) is a random walk with negative
drift which diverges to 1 as the observation interval reduces.
Now consider the co-persistence structure in the context of the constant hedging
model. Defining the true processes for the differences in the natural logarithm of the
spot index price and the natural logarithm of the futures price as

it ¼ g1 xt þ it
(44:18)
f t ¼ g2 xt þ ft ,

the common “news” factor xt is IGARCH, in the co-persistence structure, while the
idiosyncratic parts are assumed jointly independent and independent of xt and not
IGARCH. The individual processes have infinite unconditional variance. If a linear
combination is not IGARCH, then the unconditional variance of the linear combi-
nation is finite and a constant hedge ratio (defined below) leads to substantial
reduction in portfolio risk.
A time-varying hedge ratio can lead to greater reduction in portfolio risk under
conditions discussed in Ghose and Kroner (1994) when the processes are
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1259

co-persistent in variance. From a practical perspective, account needs to be taken of

the rebalancing costs of portfolio adjustment.
A nonoptimal restricted linear combination is the basis change defined as the
difference between the change in the log of the index futures price and change in
the log of the spot index level. This implied portfolio is short 1 unit of the spot for
every unit long in the futures. For the futures and spot price processes reported in
McCurdy and Morgan (1987), the basis change is co-persistent in variance.
If there are “news factors” xf t 6¼ xi t, then the constant hedge ratio may not exist.
Define these processes as

it ¼ g1 xi t þ i t
(44:19)
f t ¼ g 2 xf t þ  f t

then the estimated constant hedge ratio which is short g units of the spot for every
1 unit long in the futures is

" # 2 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 3
∧ ∧ ∧
∧ 
∧ cov ðf; iÞ ∧ p var ½xf  var ½xi 
g¼ ∧
¼ 4g 2 4 ∧
5 þ var Z i 5 (44:20)
var ðiÞ jt var ½xi 
jt

∧
where r is the correlation between xft and xit.
When both xf and xi follow IGARCH processes and no common factor structure
exists, then the estimated constant hedge diverges. Ghose and Kroner (1994)
investigate this case.
When xft follows an IGARCH process but xit is weak GARCH, then the
estimated constant hedge ratio cannot be evaluated. There are two problems:
(a) The estimated sample variance of xf in Eq. 44.20 is infinite as T ! 1.

∧  pffiffiffi ∧   ∧
(b) p ¼ c ov xf = v a r xf var ½xi  so that there is no linear combination
jt
of xf and xi which can provide a stationary unconditional variance.
This last observation has a direct parallel from the literature for cointegration in
the means of two series.
If xf is an approximate I(1) process and xi is I(0), then there is no definable linear
combination of xf and xi.
When observing spot index and futures prices over successively finer intervals,
the co-persistence structure may not hold for at least two further reasons. This
argument relates directly to the horizon t+kd for the hedging strategy. This argu-
ment also relates directly to distortions possibly induced onto a dynamic hedging
strategy, as d!0.
Perverse behavior can be observed in spot index level changes as oversampling
becomes severe. The smoothing effect due to a large proportion of the portfolio
entering the non- and thin-trading group generates a smoothly evolving process
with short-lived shocks generated by irregular news effects.
1260 G.L. Gannon

General results assume that the z0ts are drawn from a continuous distribution.
When sampling futures price data at high frequency, then the discrete nature of the
price recording mechanism guarantees that there are discontinuities in return-
generating processes. The distribution of the z0ts can become extremely peaked due
to multiple small price changes and can have very long thin tails due to abrupt shifts
in the distribution. As d ! 0,  1 < E[1n(b1 + a1zti 2
)] < + 1 for a large range of
values for 0  a1, b1  1 and a1 + b1 < 1, and this depends on the distribution
induced by oversampling and resultant reduction in (a1 + b1). In the limit E[ (zt)4]/[E
(z2t )]2 ! 1. Then even-numbered higher moments of zt are unbounded as d!0. This
oversampling can lead to two extreme perverse effects generated by bid/ask bounce
or zero price changes. The effect depends upon liquidity in the respective markets.

44.4.1 Case 1

  
a1 ! 1, a1 þ b1 > 1 and E ln b1 þ a1 z2ti > 0:

The intuitive explanation for this result relies on oversampling (not

overdifferencing) in highly liquid markets. The oversampling approaches analysis
of transactions. At this level bid/ask bounce and order splitting require an appro-
priate model. Any arbitrary autoregressive model, for unconditional first moments,
generates unconditional shocks relating predominantly to behavior of the most
recent shock. These effects carry through to conditional squared innovations.

44.4.2 Case 2

Oversampling can produce many zero price changes in thin markets. In this latter
case as d!0 then a1 + b1 ! 0.
This explanation can apply to relatively illiquid futures (and spot asset) price
changes. That is, conditional heteroskedasticity disappears as oversampling
becomes severe.
The effect on the basis change when there is a relatively illiquid futures market
and oversampling may be badly distorted. As well, failure to account for anomalies
in conditional variance equations can severely distort estimates.

44.5 Weighted GARCH

∧   
Recall from Eq. 44.1 that k is a function defining the estimated drift in f d s2t so
that l is a function defining true drift in {j(ds2t )}. In the ARCH structure, the drift
∧ 00
in ht(s2t ) in the diffusion limit is represented by k =f0  a2 f =2ðf0 Þ3, whereas for the
stochastic differential equation defined from assumptions 2, 3, and 10 in Nelson and
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1261

Foster (1991, 1994), this diffusion limit is l/j0 + L2j00 /2 (j0 )3. The effect on the
expression for the bias in the asymptotic distribution of the measurement error
00 00
process can be explosive if derivatives in the terms a2 f =2ðf0 Þ3  L2 j =2ðj0 Þ3
cannot be evaluated because of discontinuities in the process. This can happen when
important intraday effects are neglected in the conditional variance equation speci-
fication. As well, the bias can diverge as d!0 if the dZ1,kd terms are badly distorted.
Occasional large jumps in the underlying distribution contribute large Op(1) move-
ments while the near diffusion components contribute small Op(d1/2) increments. When
sampling intraday financial data, there are often many small price changes which tend
to be dominated by occasional large shifts in the underlying distribution.
Failure to account for intraday effects (large shocks to the underlying distribu-
tion) can lead to a mixture of Op(1) and Op(d1/2) effects in the process. One
approach is to specify this mixed process as a jump diffusion. An alternative is to
account for these effects by incorporating activity measures in the specification of
conditional variance equations.
It follows that failure
to account

for these Op(1) effects can lead to an explosive
∧
measurement error d h t d ht . However, in empirical applications this measure-
ment error is unobservable since dht is unobservable. Failure to account for these
jumps in the underlying distribution imply that the unweighted GARCH(1,1)
structure cannot satisfy the necessary assumptions required to approximate
a diffusion limit.

44.6 Empirical Examples

The first issue is the effect of possible mis-specification of the first-moment

equation dynamics and resultant effect on estimates of persistence of individual
processes as d ! 0. If the effect is not important (mean irrelevance) then the focus
of attention is on the estimates from the co-persistence structure and implications
for a constant hedge ratio.
The second issue is the effect of inclusion of variables to proxy intraday activity
on measures of persistence. However, it is still important to consider possible
effects from mis-specifying the first-moment equation on parameter estimates
obtained from conditional variance equations as d ! 0. If there is strong condi-
tioning from measures of activity onto the market price processes, the conditioning
should be independent of specification of the dynamics of the first-moment
equation (mean irrelevance).

44.6.1 Index Futures, Market Index, and Stock Price Data

The data has been analyzed for the common trading hours for 1992 from the ASX
and SFE, i.e., from 10.00 a.m. to 12.30 p.m. and from 2.00 p.m. to 4.00 p.m.
1262 G.L. Gannon

This dataset was first employed in examples reported in Gannon (1994) and
a similar analysis undertaken in Gannon (2010) for the SVL models as is
undertaken in this paper for the GARCH and GARCH-W models. Further details
of the sampling and institutional rules operating in these markets are
briefly reported in the Appendix of this paper. Transactions for the Share Price
Index futures (SPI) were sampled from the nearest contract 3 months to expira-
tion. The last traded price on the SPI levels and stock price levels were then
recorded for each observation interval. The All Ordinaries Index (AOI) is the
recorded level at the end of each observation interval. During these common
trading hours, the daily average number of SPI futures contracts traded 3 months
to expiration was 816. As well, block trades were extremely rare in this series.
Transactions on the heavily capitalized stock prices were extremely dense within
the trading day.
These seven stocks were chosen from the four largest market capitalized group-
ings according to the ASX classification code in November 1991, i.e., general
industrial, banking, manufacturing, and mining. The largest capitalized stocks
were chosen from the first three categories as well as the four largest capitalized
mining stocks. This selection provides for a diversified portfolio of very actively
traded stocks which comprised 32.06 % of total company weights from the
300 stocks comprising the AOI.
All datasets were carefully edited in order to exclude periods where the
transaction capturing broke down. The incidence of this was rare. As well, lags
were generated and therefore the effects of overnight records removed. A natural
logarithmic transformation of the SPI and AOI prices is undertaken prior to
analysis.
Opening market activity for the SPI is heaviest during the first 40 min of
trading. Trade in the SPI commences at 9.50 a.m. but from 10.30 a.m. onwards
volume of trade tapers off until the lunchtime close. During the afternoon
session, there is a gradual increase in volume of trade towards daily market
close at 4.10 p.m. Excluding the market opening provides the familiar U-shaped
pattern of intraday trading volume observed on other futures markets. SPI price
volatility is highest during the market opening period with two apparent reverse
J-shaped patterns for the two daily trading sessions (small J effect in afternoon
session). The first and last 10 min of trade in the SPI are excluded from this
dataset.
Special features govern the sequence at which stocks open for trade on the
ASX. Individual stocks are allocated a random opening time to within plus or
minus 30 s of a fixed opening time. Four fixed opening times, separated by 3-min
intervals starting at 10.00 a.m., operated throughout 1992. Four alphabetically
ordered groups then separately opened within the first 10 min of trading time.
The last minute of trading on the ASX is also subject to a random closing time
between 3.59 p.m. and 4.00 p.m. The effect of these institutional procedures on
observed data series can be potentially severe.
Both of these activity effects in market opening prices of trading on the SFE, and
ASX should be accounted for in the estimation process.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1263

Table 44.1 Autoregressive parameter estimates for SPI, AOI, and basis change
Interval ft it bt
Full day
30 min 0.0324 0.0909 0.1699
(1.54) (4.33) (8.19)
15 min 0.0497 0.0953 0.1348
(3.45) (6.42) (9.15)
05 min 0.0441 0.3317 0.0137
(5.14) (41.2) (1.60)
Excluding market open
30 min 0.0299 0.1213 0.1947
(1.34) (5.48) (8.89)
15 min 0.0301 0.2531 0.2145
(1.91) (16.6) (13.9)
05 min 0.0532 0.3284 0.0968
(5.84) (38.2) (10.7)
Asymptotic t-statistics in brackets
ft ¼ Ft – Ft is the difference in the observed log level of the SPI
it ¼ It – I is the difference in the observed log level of the AOI
Equation 44.5, i.e., an autoregressive specification for the differences, an [AR(D)], is estimated for
both data series with one lag only for the autoregressive parameter
Dummy variables are included, for the first data series, in order to account for market opening
effects for the SPI, institutional features governing market opening on the ASX and therefore
effects transmitted to the basis change. Two separate dummy variables are included for the first
pair of 5-min intervals
This form for the basis change is bt ¼ ft – it
In the lower panel, results are reported from synchronized trading from 10.30 a.m. That is, the first
40 min and first 30 min of normal trade in the SPI and AOI, respectively, is excluded

44.6.2 Estimates of the Autoregressive Parameters

In Table 44.1 the first-order autoregressive parameter estimate is reported for the
observed differenced series for the SPI (f), AOI (i), and basis change (b). In the top
panel, these parameter estimates are from observations for the full (synchronized)
trading day. These equations include dummy variables to account for institutional
market opening effects. In the following tables of results, SING refers to singular-
ities in the estimation process.
For the SPI futures price process, low first-order negative serial correlation is
initially detected in the log of the price change. As the sampling interval is reduced,
low first-order positive serial correlation can be detected in the series. This feature
of the data accords with order splitting and non-trading-induced effects.
Low positive first-order serial correlation can be detected in differences of the
log of the market index and serial correlation increases. Positive serial correlation is
high at 15-min intervals for the opening excluded set. When sampling the market
index at 5-min intervals, substantial positive first-order serial correlation is detected
1264 G.L. Gannon

in the log of the price change process. Miller et al. (1994) demonstrate that thin
trading and non-trading in individual stocks induce a positive serial correlation in
the observed spot index price change process. However, the smoothing effect from
non-trading in individual stocks, averaging bid/ask bounce effects in heavily traded
stocks, and under-differencing the aggregate process also contribute.
Of interest is the reduction in serial correlation of the basis spread as d ! 0. As
the log futures price change moves into the order splitting/non-price change region,
positive serial correlation is induced in the log of the futures price change. This, in
part, helps offset the increasing positive serial correlation induced in the log of the
spot index.

44.6.3 Conditional Variance Estimates

Allowing for and excluding market open/closing effects in first-moment equations

make little difference to GARCH parameter estimates. This observation holds for
alternative specifications of the dynamics for the first-moment equation at 30- and
15-min intervals. However, mis-specifying the first-moment equation dynamics is
important in the conditional second-moment equations for the AOI at 5-min intervals.
The GARCH(1,1) estimates for the SPI, AOI, and basis are generated from
Eq. 44.6, i.e., an autoregressive specification of the levels AR(L) and Eq. 44.5 the
AR(D), respectively (Table 44.2).
Opening dummy variables are incorporated in both first-moment and conditional
second-moment equations for the full-day data series. Two dummy variables,
corresponding to the first- two 5-min intervals, are included in both first and condi-
tional second-moment equations for the full-day data series. Results in the lower
panel are obtained from excluding all trade in the SPI and AOI prior to 10.30 a.m.
At 5-min sampling intervals, a transient effect is introduced into the
unconditional distributions which carries through to the conditional variance
estimates. Values for a1 and b1 differ both within the same data series for
alternative first-moment equations and across data series for the same form of
first-moment equation. For one extreme case (with exclusion of market open-
ing), the GARCH parameter estimates for the AOI and therefore the basis
change do depend on the mis-specification of the first-moment equation.
From the first set of results (full day), it would appear as if the SPI is persistent in
variance, the AOI is not and neither is the basis change persistent in variance.
A second extreme case occurs at 30- and 15-min intervals (with exclusion
of market opening). In this case GARCH parameter estimates are almost identical
for alternative specifications of the dynamics of the first-moment equation. How-
ever, for the AOI the sum of the GARCH parameter estimates is near the IGARCH
boundary, and the same feature is observed in the basis change.
It would appear that, given anomalies are adequately accounted for,
mis-specification of the (weak) dynamic structure of price changes in these processes
is only relevant for estimating these GARCH equations for the AOI at 5-min intervals.
The important issue is the correct specification of the market opening effects.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1265

Table 44.2 GARCH estimates for the SPI (Ft), AOI (It), and basis change (Bt) for AR(L) and
AR(D) specifications
Ft It Bt
Full day
30 min
a1 0.0527 0.0517 0.1019 0.1002 0.0904 0.0741
(7.84) (7.58) (7.23) (6.91) (9.04) (7.70)
b1 0.9266 0.9284 0.0196 0.0227 0.0204 0.0251
(115) (111) (2.09) (2.24) (2.06) (2.31)
15 min
a1 0.0602 0.0592 0.1922 0.1514 0.1050 0.1000
(14.7) (14.5) (13.2) (11.4) (13.6) (13.6)
b1 0.9151 0.9167 0.0893 0.1100 0.0852 0.1186
(224) (225) (9.89) (10.9) (5.45) (7.17)
05 min
a1 0.0358 0.0362 0.3540 0.2338 0.1705 0.1685
(34.0) (34.6) (35.8) (27.9) (39.6) (38.5)
b1 0.9535 0.9530 0.2074 0.2428 0.4345 0.4469
(1075) (1087) (26.8) (29.3) (43.3) (45.6)
Excluding market open
30 min
a1 0.0448 0.0444 0.0668 0.0693 0.0684 0.0631
(8.31) (8.12) (8.71) (9.22) (11.5) (10.5)
b1 0.9473 0.9478 0.9074 0.9050 0.9114 0.9172
(155) (150) (92.9) (93.3) (114) (109)
15 min
a1 0.0385 0.0394 0.0867 0.0835 0.0310 0.0403
(14.8) (14.5) (12.9) (13.2) (12.9) (18.8)
b1 0.9551 0.9564 0.8832 0.8872 0.9499 0.9537
(382) (369) (113) (115) (232) (374)
05 min
a1 5.8E-5 0.0329 0.3326 0.2229 4.6E-5 0.0236
(33.7) (30.3) (37.5) (33.6) (27.4) (35.9)
b1 1.000 0.9638 0.4470 0.5795 1.000 0.9742
SING (906) (35.2) (46.4) SING (1353)
Asymptotis t-statistics in brackets

These preliminary results have been obtained from observations sampled for all
four futures contracts and a continuous series constructed for 1992. The AOI is not
affected by contract expiration. In Table 44.3, GARCH(1,1) estimates for the
separate SPI futures contracts, 3 months to expiration, and synchronously sampled
observations on the AOI are recorded. The full data series corresponding to
synchronized trading on the SFE and ASX is employed.
The same form of dummy variable set was imposed in both first and second-
moment equations. As well, a post-lunchtime dummy is included to account for the
1266 G.L. Gannon

Table 44.3 GARCH estimates for the SPI and AOI 3 months to expiration
MAR JUN SEP DEC
Ft It Ft It Ft It Ft It
30 min
a1 0.042 0.026 0.145 0.015 0.019 0.089 0.060 0.147
(2.37) (0.94) (3.95) (0.76) (2.96) (3.65) (3.24) (3.70)
b1 0.933 0.054 0.116 0.004 0.977 0.021 0.918 0.037
(35.6) (2.05) (0.56) (0.21) (118) (1.12) (35.0) (1.30)
15 min
a1 0.028 0.186 SING 0.107 0.037 0.183 0.036 0.195
(3.11) (5.85) (4.75) (4.96) (6.84) (6.08) (5.80)
b1 0.958 0.108 0.998 0.035 0.956 0.055 0.958 0.102
(72.6) (6.65) (3235) (2.11) (120) (3.75) (114) (3.67)
05 min
a1 0.051 0.240 SING 0.307 0.174 0.378 0.039 0.409
(7.86) (12.4) (14.8) (22.2) (17.0) (11.0) (19.3)
b1 0.881 0.231 0.999 0.118 0.649 0.161 0.956 0.259
(52.9) (12.1) (21170) (7.16) (38.5) (10.6) (247) (16.9)
GARCH(1,1) parameter estimates for the 3 months corresponding to expiration of the March,
June, September, and December contracts for 1992. Full-day data series are employed with
opening and post-lunchtime dummy variables in both first and conditional second-moment
equations. An AR(L) specification of the (log) mean equation is employed for these results

break in daily market trade at the SFE during 1992. The log of levels is specified for
the first-moment equations.
There is some instability within this set of parameter estimates. However,
a similar pattern emerges within the set of futures and the set of market index
estimates as was observed for the full-day series for 1992. The futures conditional
variance parameter estimates are close to the IGARCH boundary while the index
conditional variance estimates are not. This again implies that these processes
cannot be co-persistent in variance for these samples and observation intervals.
In order to obtain further insight into the, seemingly, perverse results for the
market index, a similar analysis was undertaken on seven of the largest market
capitalized stocks which comprised the AOI during 1992.
Relevant market opening and closing dummy variables were included in both
first and conditional second-moment equations accordingly. These effects were not
systematic in the first-moment equations and are not reported. The stock price
processes have not been transformed to natural logarithmic form for these estima-
tions. This is because the weighted levels of the stock prices are employed in
construction of the market index.
As the observation interval is reduced for these stock prices:
(i) The autoregressive parameter estimates for the price levels equation
converges to a unit root.
(ii) The first-order serial correlation coefficient for the price difference equation
moves progressively into the negative region.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1267

If these were the only stocks comprising the construction of the index, then we
might expect to see increasing negative serial correlation in the index. But this
observation would only apply if these processes were sampled from a continuous
process which was generated by a particular form of ARMA specification. Only in
the case of data observed from a continuous process could the results on temporal
aggregation of ARMA processes be applied.
These results should not be surprising as the combination of infrequent price
change and “price bounce” from bid/ask effects starts to dominate the time series.
As well, these bid/ask boundaries can shift up and down. When these processes are
aggregated, the effects of price bounce can cancel out. As well, the smoothing
effect of thinly and zero traded stocks within the observation intervals dampens and
offsets individual negative serial correlation observed in these heavily traded stocks
(Table 44.4).
Some of these autoregressive parameter estimates are quite high for the
AR(D) specifications. However, it is apparent that there is almost no difference in
GARCH(1,1) estimates from either the AR(L) or AR(D) specifications at each
observation interval for any stock price. In some instances estimation breaks
down at 5 min intervals.
If the conditional variances of these stock price movements contain common
news and announcement effects, then it should not be surprising that the
weighted aggregated process is not persistent in variance. This can happen
when news affects all stocks in the same market. As well, smoothing effects
from thin-traded stocks help dampen volatility shocks observed in heavily traded
stocks. These news and announcement effects may be irrelevant when observing
these same processes at daily market open to open or close to close. These news
and announcement effects may be due to private information filtering onto the
market prior to and following market open. It is during this period that overnight
information effects can be observed in both price volatility and volume. As well,
day traders and noise traders are setting positions. However, the ad hoc applica-
tion of dummy variables is not sufficient to capture the interaction between
volatility and volume. In the absence of specific measures of these news “vari-
ables,” the effects cannot be directly incorporated into a structural model.
However, these effects are often captured in the price volatility and reflected in
increased trading activity.

44.6.4 Weighted GARCH Estimates

Weighted GARCH estimates for the futures (log) price process with the accumu-
lated number of futures contracts traded within the interval t to t1 are reported in
Table 44.5. This choice ensures that volume measures are recorded within the
interval that actual prices define.
The weighting variable employed in the index (log) price process is the squared
shock from the futures price mean equation within the interval t to t1. There is no
natural “volume” of trade variable available for the market index. The form of mean
1268 G.L. Gannon

Table 44.4 Unconditional mean and GARCH(1,1) estimates: Australian stock prices
AR(L) AR(D) AR(L) AR(D)
Interval p1 r1 a1 b1 a1 b1
NAB
60 min 0.999 0.160 0.125 0.762 0.122 0.774
(17.2) (39.2) (16.4) (40.1)
30 min 0.999 0.135 0.086 0.862 0.081 0.876
(25.3) (123) (24.4) (135)
15 min 0.999 0.199 0.226 0.627 0.220 0.629
(22.5) (70.0) (22.2) (61.1)
05 min 1.00 0.287 2.4E-5 0.042 2.3E-5 0.049
(0.94) (0.28) (0.53) (0.24)
BHP
60 min 1.00 0.034 0.096 0.843 0.098 0.839
(9.26) (45.6) (9.14) (44.2)
30 min 1.00 0.087 0.067 0.898 0.066 0.899
(14.2) (129) (13.7) (128)
15 min 1.00 0.104 0.054 0.923 0.054 0.924
(20.5) (281) (20.2) (279)
05 min 1.00 0.123 0.132 0.795 0.130 0.799
(66.0) (361) (61.9) (363)
BTR
60 min 0.995 0.035 0.194 SING 0.196 SING
(9.68) (SING) (9.68) (SING)
30 min 0.998 0.109 0.232 0.294 0.231 0.297
(14.3) (7.36) (13.9) (7.55)
15 min 0.999 0.109 0.136 0.615 0.131 0.626
(17.6) (31.5) (17.5) (32.8)
05 min 1.00 0.074 0.076 0.831 0.074 0.835
(45.9) (231) (45.5) (233)
WMC
60 min 1.00 0.052 0.015 0.981 0.015 0.981
(6.58) (368) (6.54) (363)
30 min 1.00 0.016 0.272 0.234 0.265 0.251
(13.7) (5.39) (13.70 (5.85)
15 min 1.00 0.008 0.185 0.594 0.186 0.597
(19.6) (33.7) (19.3) (34.0)
05 min 1.00 0.065 0.013 0.003 SING SING
CRA
60 min 0.999 0.031 0.332 0.058 0.333 0.074
(13.0) (1.52) (12.5) (1.88)
30 min 0.999 0.008 0.231 0.456 0.231 0.458
(10.2) (5.89) (9.98) (5.69)
15 min 1.00 0.003 0.132 0.687 0.132 0.688
(24.0) (58.8) (23.8) (58.5)
(continued)
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1269

Table 44.4 (continued)

AR(L) AR(D) AR(L) AR(D)
Interval p1 r1 a1 b1 a1 b1
05 min 1.00 0.006 0.087 0.840 0.087 0.841
(84.1) (433) (80.1) (437)
MIM
60 min 0.998 0.015 0.016 0.979 0.037 0.934
(5.22) (227) (6.47) (94.4)
30 min 0.999 0.090 0.166 0.279 .160 0.281
(25.30) (123) (24.4) (135)
15 min 1.00 0.091 0.151 0.612 0.150 0.614
(23.9) (40.9) (23.5) (40.9)
05 min 1.00 0.084 0.001 0.841 0.001 0.814
(2.66) (9.86) (2.35) (9.46)
CML
60 min 1.00 0.033 0.088 0.867 0.085 0.873
(8.88) (52.1) (9.02) (56.2)
30 Min 1.00 0.045 0.037 0.953 0.039 0.950
(14.4) (52.1) (14.4) (269)
15 Min 1.00 0.056 0.196 0.647 0.189 0.658
(28.2) (62.4) (27.3) (64.6)
05 min 1.00 0.078 0.095 0.844 0.110 0.846
(74.6) (456) (75.90 (471)
Asymptotic t-statistics in brackets
Column 1 contains the observation interval
Column 2 contains the autoregressive parameter from an AR(L) specification of the price levels
equation
Column 3 contains the first-order autoregressive parameter estimate from an AR(D) specification
of the price change
Columns 4 and 5 contain the GARCH(1,1) parameter estimates from an AR(L) specification of the
price levels equation
Columns 6 and 7 contain the corresponding estimates from an AR(D) specification of the price changes

equation is the same as the generated corresponding results for Table 44.3, i.e., an
AR(L). However, the results are almost identical when alternative forms for the
mean equation are employed, i.e., AR(L) and AR(D).
The specifications generating the reported stock price estimates are augmented
to include a measure of trade activity within the observation interval. These
measures are the accumulated number of stocks traded in each individual stock
within the interval t to t1. The estimates are reported in Table 44.6. The condi-
tional variance parameter estimates are almost identical from the AR(L) and
AR(D) specifications of the mean equation. The logarithmic transformation has
not been taken for these stock prices.
Direct comparison of these GARCH parameter estimates (a1 and b1) with those
from the unweighted GARCH(1,1) estimates demonstrates the importance of this
measure of activity. The change in GARCH parameter estimates is striking.
1270 G.L. Gannon

Table 44.5 Weighted GARCH estimates for the SPI and AOI 3 months to expiration
March June September December
Ft It Ft It Ft It Ft It
30 min
a1 0.093 0.040 0.134 0.027 SING 0.079 0.030 0.007
(2.07) (1.51) (4.58) (1.10) (2.73) (1.01) (0.40)
b1 0.014 0.039 SING 0.068 SING 0.033 SING 0.025
(0.24) (1.62) (2.74) (1.91) (1.24)
15 min
a1 0.052 0.185 0.175 0.022 0.039 0.042 0.076 0.058
(1.84) (3.97) (6.67) (1.50) (2.47) (2.35) (2.89) (3.04)
b1 0.004 0.086 0.000 0.163 0.000 0.106 0.001 0.136
(0.19) (3.82) (0.00) (7.61) (0.00) (4.99) (0.01) (5.67)
05 min
a1 0.000 0.156 0.095 0.170 0.004 0.162 0.030 0.128
(4.04) (8.34) (7.79) (9.55) (32.1) (9.58) (8.32) (7.30)
b1 0.000 0.238 SING 0.232 0.000 0.270 0.000 0.327
(5.06) (11.9) (11.8) (0.12) (16.0) (7.79) (17.8)

The measures of persistence from these weighted estimates are never near the
IGARCH boundary. These effects are summarized in Table 44.7.
These results are generated from an AR(L) specification of the first-moment
equation. By adequately accounting for contemporaneous intraday market activity,
the time persistence of volatility shocks becomes less relevant. It follows that
deviations of the estimated conditional variance from the true (unobservable)
conditional variance are reduced (Table 44.8).

44.6.5 Discussion

In this section some empirical evidence is documented on the behavior of uncon-

ditional first and conditional second-moment effects for the market index, futures
contracts written on the market index, and for heavily traded stock prices. These
results are for Australian financial assets sampled on an intraday basis as the
observation interval approaches transactions time d ! 0.
The specific empirical findings were:
1. The autoregressive parameter estimate from a difference equation for the log of
the index futures is initially negative but moves into the positive region. This can
be attributed to bid/ask bounce being dominated by order splitting and
non-trading effects.
2. The autoregressive parameter estimate from a difference equation for the log of
the market index displays increasing positive serial correlation. This can be
attributed to non-trading smoothing effects in low capitalized stocks, averaging
of bid/ask bounce effects, and under-differencing the aggregated index.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1271

3. The basis change between the log futures price change and the log index level
change displayed two surprising effects:
(i) The autoregressive parameter for the basis change is initially negative, but
the strength of this effect weakens. This can be attributed to the log of the

Table 44.6 Weighted GARCH estimates: Australian stock prices

AR(L) AR(D)
Stock Interval a1 b1 X a1 b1 X
NAB 60 min 0.147 0.028 3.0E-9 0.144 0.027 3.0E-9
(11.7) (2.89) (36.9) (11.1) (2.35) (36.2)
30 min 0.168 0.143 2.8E-9 0.171 0.139 2.9E-9
(11.4) (12.9) (47.0) (11.6) (11.6) (46.8)
15 min 0.140 0.135 3.1E-9 0.131 0.142 3.1E-9
(15.7) (17.1) (114) (15.6) (16.80 (112)
BHP 60 min 0.125 0.053 9.4E-9 0.124 0.057 9.5E-9
(8.97) (2.94) (27.7) (8.49) (3.22) (28.2)
30 min 0.108 0.133 8.7E-9 0.095 0.137 8.9E-9
(10.8) (11.2) (32.8) (9.95) (12.8) (34.1)
15 min 0.132 0.110 9.9E-9 0.121 0.113 10E-9
(13.9) (13.9) (56.6) (13.2) (12.6) (58/8)
BTR 60 min 0.045 0.024 7.8E-9 0.042 0.011 8.0E-9
(3.06) (1.29) (19.0) (2.85) (0.64) (19.2)
30 min 0.116 0.051 8.0E-10 0.106 0.056 8E-10
(9.74) (3.94) (24.8) (9.46) (4.34) (25.3)
15 min 0.070 0.023 1.9E-9 0.058 0.026 1.9E-9
(14.0) (5.28) (38.80 (11.60 (5.45) (38.9)
WMC 60 min 0.110 0.040 1.9E-9 0.110 0.040 1.9E-9
(5.68) (2.18) (25.3) (5.67) (2.22) (25.5)
30 min 0.107 0.008 1.9E-9 0.108 0.009 1.9E-9
(8.35) (0.65) (30.8) (8.39) (0.76) (30.7)
15 min 0.078 0.053 1.8E-9 0.074 059 1.8E-9
(14.6) (6.47) (38.0) (14.2) (7.21) (38.1)
CRA 60 min 0.116 0.020 4.3E-8 0.116 0.015 4.3E-8
(7.04) (2.06) (26.8) (7.07) (1.69) (26.7)
30 min 0.056 0.023 5.4E-8 0.054 0.024 5.5E-8
(8.30) (4.95) (38.4) (8.11) (5.14) (38.4)
15 min 0.000 0.000 2.9E-7 0.000 0.000 2.9E-7
(0.0) (4.04) (207) (1.02) (6.88) (208)
MIM 60 min 0.074 SING 6.4E-10 0.077 0.000 6.E-10
(5.36) (21.7) (5.24) (0.00) (21.6)
30 min 0.071 0.074 6.1E-10 0.070 0.087 6.E-10
(5.84) (4.13) (24.6) (6.28) (4.90) (25.0)
15 min 0.040 0.068 8.7E-10 0.030 0.074 9.E-10
(6.67) (9.29) (40.0) (4.98) (9.68) (40.5)
(continued)
1272 G.L. Gannon

Table 44.6 (continued)

AR(L) AR(D)
Stock Interval a1 b1 X a1 b1 X
CML 60 min 0.125 SING 1.9E-8 0.122 SING 1.9E-8
(5.92) (17.8) (5.81) (17.8)
30 min 0.108 0.029 2.5E-8 0.101 0.034 2.6E-8
(8.65) (3.44) (30.8) (8.27) (4.16) (31.1)
15 min 0.056 0.000 4.0E-8 0.057 SING 4.4E-8
(16.3) (0.00) (56.2) (17.4) (59.2)
Asymptotic t-statistics in brackets
Columns 3 and 4 contain the weighted GARCH parameter estimates from an AR(L) specification
of the price levels equation with the volume parameter estimate in column 5
Columns 6–8 contain the corresponding estimates from an AR(D) specification of the price changes
Estimates are quite unstable at 5-min intervals

Table 44.7 Measures of persistence from GARCH(1,1) and weighted GARCH equations for the
SPI 3 months to expiration
G G-W G G-W G G-W G G-W
Interval March June September December
30 min 0.975 0.107 0.261 SING 0.996 SING 0.978 SING
15 min 0.986 0.056 SING 0.175 0.993 0.039 0.994 0.077
05 min 0.932 0.000 SING SING 0.823 0.004 0.995 0.030

futures price change behavior where the autoregressive parameter moved

into the positive region and price change became less frequent.
(ii) The unweighted log of the futures price change was close to the IGARCH
boundary. For the full-day data series, the log of the market index change
was not close to the IGARCH boundary and neither was the basis change.
When market opening trade was excluded, the log of the futures price
change, the log of the market index price change, and the basis change
were close to the IGARCH boundary although this effect dissipated as
d ! 0. However, volume of trade on both the SFE and ASX is heaviest
within this excluded interval. It follows that any conclusions concerning
the co-persistence structure would be misleading from this intraday data.
4. The autoregressive parameter from a levels equation for the stock prices con-
verges to a unit root. This can be attributed to small and zero price change
effects.
5. The autoregressive parameter from a difference equation for the stock prices
displays increasing tendency towards and into the negative serial correlation
region. This can be attributed to price bounce effects where the boundaries are
tight and relatively stable.
6. GARCH(1,1) or weighted GARCH conditional variance parameter estimates do
not depend on the specification of the dynamics of first-moment equation for
30- and 15-min intervals of these futures, market index, and stock price processes.
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1273

Table 44.8 Measures of persistence from GARCH(1,1) and weighted GARCH equations for
Australian stock price processes
NAB BHP BTR WMC
60 min 0.887 0.175 0.939 0.178 SING 0.069 0.996 0.150
30 min 0.948 0.311 0.965 0.241 0.526 0.167 0.506 0.115
15 min 0.853 0.275 0.977 0.242 0.751 0.093 0.779 0.131
CRA MIM CML
60 min 0.390 0.136 0.995 SING 0.995 SING
30 min 0.687 0.079 0.445 0.145 0.990 0.137
15 min 0.819 0.000 0.763 0.108 0.843 0.056
An AR(L) specification has been employed to generate these results
The measure of persistence is the calculated as a+b from the conditional variance equations
G represents persistence obtained from a GARCH(1,1) specification
G-W represents persistence obtained from a weighted GARCH specification
SING indicates that one or more of these GARCH parameters could not be evaluated due to
singularities

The GARCH(1,1) parameter estimates for the AOI at 5-min intervals were
different when alternative forms of first-moment equations were specified. There
was no perceivable difference in the other processes at this sampling frequency. It
would then appear that increasing positive serial correlation (smoothing) in the
observed returns process has a greater distorting effect on GARCH(1,1) parameter
estimates than increasing negative serial correlation (oscillation) in observed
returns processes. The most important effect is mis-specification of the conditional
variance equations from failure to adequately account for the interaction between
market activity and conditional variance (volatility).
Some implications of these results are:
7. When aggregating stock prices, which may be close to the IGARCH boundary,
persistence in variance can be low for the market index because (i) there is
common persistence present in heavily traded stock prices which when aggre-
gated do not display persistence in variance because heavily traded stock prices
react to market specific news instantaneously, (ii) the smoothing effect of less
heavily traded stocks dampens the volatility clustering which is often observed
in other financial assets such as exchange rates, and (iii) high volatility and
volume of trade effects within these markets following opening is better mea-
sured by employing relevant measures of activity than an ad hoc approach.
8. There is a strong and quantifiable relationship between activity in these markets
and volatility.

44.7 Conclusion

The behavior of financial asset price data observed intraday is quite different from
these data observed at longer sampling intervals such as day to day. Market
anomalies which distort intraday observed data mean that volatility estimates
1274 G.L. Gannon

obtained from data observed from day-to-day trades will provide different volatility
estimates. This latter feature then depends upon when the data is sampled within the
trading day. If these anomalies and intraday trading patterns are accounted for in the
estimation process, then better estimates of volatility are obtainable by employing
intraday observed data. However, this is dependent on a sampling interval that is
not so fine that these estimators break down.
The specific results indicate that serial correlation in returns processes can
distort parameter estimates obtainable from GARCH estimators. However,
induced excess kurtosis may be a more important factor in distortions to esti-
mates. The most important factor is mis-specification of the conditional variance
(GARCH) equation from omission of relevant variables which explain the
anomalies and trading patterns observed in intraday data. Measures of activity
do help explain systematic shifts in the underlying returns distribution and in
this way help explain “jumps” in the volatility process. This effect can be
observed in the likelihood function and in asymptotic standard errors of
weighting (mixing) variables. One feature is that the measure of volatility
persistence observed in unweighted univariate volatility estimators is reduced
substantially with inclusion of weighting variables.

Appendix

At the time the ASX data were collected, the exchange had just previously moved
from floor to screen trading with the six main capital city exchanges linked via
satellite and trade data streamed to trading houses and brokers instantaneously via
a signal G feed. The SFE maintained Pit trading for all futures and options on
futures contracts at the time.
Legal restrictions on third party use and development of interfaces meant the
ASX had a moratorium on such usage and development. The author was required to
obtain special permission from the ASX to capture trade data from a live feed from
broking house Burdett, Buckeridge, and Young (BBY). There was a further delay in
reporting results of research following the legal agreement obtained from the ASX.
Trade data for stock prices and volume of trade were then sampled into 5-min files
and subsequently into longer sampling interval files. The market index was refreshed
at 1-min intervals and the above sampling scheme repeated. Futures price trades were
supplied in two formats via feed: Pit (voice recorded) data and Chit data. Although
the Pit data provides an instantaneous record of trade data during the trading day,
some trades are lost during frantic periods of activity. The Chit records are of every
trade (price, volume, buyer, seller, time stamped to the nearest second, etc.).
The recorded Chits are placed in a wire basket on a carriageway and transferred up
the catwalk where recorders on computers enter details via a set of simplified
keystrokes. The average delay from trade to recording is around 30 s for the Chit
trades. These are then fed online to trading houses and brokers. At the end of the
trading day, these recorded trades are supplemented with a smaller set of records that
44 Stochastic Volatility Structures and Intraday Asset Price Dynamics 1275

were submitted to the catwalk late, e.g., morning trades that may have gone to lunch
in a brokers pocket and submitted during the afternoon session and also some late
submitted trades.
We created the intraday sampled files from both the Pit and Chit records.
However, we employed the Chit trades for analysis in this paper so as to have the
correct volume of trade details for each trading interval. All trades were reallocated
using the time stamps to the relevant time of trade, including trades not submitted
on time but supplied as an appendix to the trading day data. In this study the average
number of late Chits were not a high proportion of daily trades. These futures price
records were then sampled into relevant 5-min records and longer sampling frames
generated in the same manner as was employed for the stock prices.
For all series the first price and last price closest to the opening and closing nodes
for each sampling interval were recorded with volume of trade the accumulated
volume of trade within the interval defined by first and last trade.

References
Brown, S. (1990). Estimating volatility. In S. Figlewski et al. (Eds.), Financial options: From
theory to practice (pp. 516–537). Homewood: Business One Irwin.
Duffee, G., Kupiec, P., & White, A. P. (1990). A primer on program trading and stock price
volatility: A survey of the issues and evidence (Working Paper No. 109). FEDS Board of
Governors of the Federal Reserve System.
Edwards, F. R. (1988). Futures trading and cash market volatility: Stock index and interest rate
futures. Journal of Futures Markets, 8, 421–439.
Gannon, G. L. (1994). Simultaneous volatility effects in index futures. Review of Futures Markets,
13, 1027–1066.
Gannon, G. L. (2010). Simultaneous volatility transmissions and spillovers: Theory and evidence.
Review of Pacific Basin Financial Markets and Policies, 13, 127–156.
Ghose, D., & Kroner, K. F. (1994). Common persistence in conditional variances: Implications for
optimal hedging. Paper presented at the 1994 Meeting of the Australasian Econometric
Society.
McCurdy, T., & Morgan, I. G. (1987). Tests of the martingale hypothesis for foreign currency
futures with time varying volatility. International Journal of Forecasting, 3, 131–148.
Miller, M. H., Muthuswamy, J., & Whaley, R. E. (1994). Mean reversion of standard and poor’s
500 index basis changes: Arbitrage-induced or statistical illusion? Journal of Finance, 49,
479–513.
Nelson, D. B. (1990a). ARCH models as diffusion approximations. Journal of Econometrics, 45, 7–38.
Nelson, D. B. (1990b). Stationarity and persistence in the GARCH(1,1) model. Econometric
Reviews, 6, 318–334.
Nelson, D. B., & Foster, D. P. (1991). Estimating conditional variances with misspecified ARCH
models: Asymptotic theory. Graduate School of Business, University of Chicago, mimeo.
Nelson, D. B., & Foster, D. P. (1994). Asymptotic filtering theory for univariate ARCH models.
Econometrica, 62, 1–41.
Optimal Asset Allocation Under VaR
Criterion: Taiwan Stock Market 45
Ken Hung and Suresh Srivastava

Contents
45.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1278
45.2 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1279
45.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1281
45.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1288
Appendix 1: Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1288
Appendix 2: Optimal Portfolio Under a VaR Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1289
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1290

Abstract
Value at risk (VaR) measures the worst expected loss over a given time horizon
under normal market conditions at a specific level of confidence. These days,
VaR is the benchmark for measuring, monitoring, and controlling downside
financial risk. VaR is determined by the left tail of the cumulative probability
distribution of expected returns. Expected probability distribution can be
generated assuming normal distribution, historical simulation, or Monte
Carlo simulation. Further, a VaR-efficient frontier is constructed, and an
asset allocation model subject to a target VaR constraint is examined.
This paper examines the riskiness of the Taiwan stock market by determining
the VaR from the expected return distribution generated by historical simulation.
Our result indicates the cumulative probability distribution has a fatter left tail,
compared with the left tail of a normal distribution. This implies a riskier market.

K. Hung (*)
Division of International Banking & Finance Studies, Texas A&M International University,
Laredo, TX, USA
e-mail: ken.hung@tamiu.edu
S. Srivastava
University of Alaska Anchorage, Anchorage, AK, USA
e-mail: afscs@uaa.alaska.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1277
DOI 10.1007/978-1-4614-7750-1_45,
# Springer Science+Business Media New York 2015
1278 K. Hung and S. Srivastava

We also examined a two-sector asset allocation model subject to a target VaR

constraint. The VaR-efficient frontier of the TAIEX traded stocks recommended
mostly a corner portfolio.

Keywords
Value at risk • Asset allocation • Cumulative probability distribution • Normal
distribution • VaR-efficient frontier • Historical simulation • Expected return
distribution • Two-sector asset allocation model • Delta • Gamma • Corner
portfolio • TAIEX

45.1 Introduction

Risk is defined as the standard deviation of unexpected outcomes, also known as

volatility. Financial market risks are of four types: interest rate risk, exchange rate
risk, equity risk, and commodity risk. For a fixed-income portfolio, the linear
exposure to the interest rate movement is measured by duration. Second-order
exposure is measured by convexity. In the equity market, linear exposure to market
movement is measured by the systematic risk or beta coefficient. In the derivative
markets, the first-order sensitivity to the value of underlying asset is measured by
delta, and second-order exposure is measured by gamma. Innovations in the
financial markets have introduced complicated portfolio choices. Hence, it is
becoming more difficult for managers to get useful and practical tools of market
risk measurement. The simple linear considerations such as Basis Point Value, or
first- or second-order volatility, are inappropriate. They can’t accurately reflect risk
at the time of dramatic price fluctuation.
VaR (value at risk) has become a popular benchmark for the downside risk
measurement.1 VaR converts the risks of different financial products into one
common standard: potential loss, so it can estimate market risk for various kinds
of investment portfolio. VaR is used to estimate the market risk of financial assets.
Special concern of the market risk is the downside risk of portfolio values resulting
from the fluctuation of interests, exchange rates, stock prices, or commodity prices.
VaR is consistent in estimating the financial risk estimation. It indicates risk of
dollar loss of portfolio value. Now the risks exposure of different investment
portfolios (such as equity and fixed income) or different financial products
(such as interest rate swaps and common stock) have a common basis for direct
comparison. For decision makers, VaR is not only a statistical summary; it can also
be used as a management and risk control tool to decide capital adequacy,
asset allocation, synergy-based salary policy, and so on.

1
Extensive discussion of value at risk can be found in Basak and Shapiro (2001), Beder (1995),
Dowd (1998), Fong and Vasicek (1997), Hendricks (1996), Hoppe (1999), Jorion (1997, 1997),
Schachter (1998), Smithson and Minton (1996a, b), and Talmor (1996).
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1279

VaR is a concept widely accepted by dealers, investors, and legislative authorities.

J.P. Morgan has advocated VaR and incorporated it in RiskMetrics (Morgan 1996).
RiskMetrics contain most of the data and formulas used to estimate daily VaR,
including daily updated fluctuation estimations for hundreds of bonds, securities,
currencies, commodities, and financial derivatives. Regulatory authorities and central
bankers from various countries at Basel Committee meetings agreed to use VaR as the
risk-monitoring tool for the management of capital adequacy. VaR has also been
widely accepted and employed by securities corporations, investment banks, com-
mercial banks, retirement funds, and nonfinancial institutions. Risk managers have
employed VaR in ex-post evaluation, that is, to estimate and justify the current market
risk exposure.2 Confidence-based risk measure was first proposed by Roy (1952).
The inclusion of VaR into the asset allocation model means the inclusion of
downside risk into model constraints. Within the feasible scope of investment
portfolio that meets shortfall constraints, the optimal investment portfolio is
decided by maximum expected return. The definition of shortfall constraint is that
the probability of investment portfolio value dropping to a certain level is set as the
specific disaster probability. The asset allocation framework that takes VaR as one
of its constraints has increased the importance of VaR and has employed VaR as an
ex-ante control tool of market risk.

45.2 Value at Risk

One difficulty in estimating VaR is the choice of various VaR methods

and corresponding hypotheses. There are three major methods to estimate
VaR: variance-covariance analysis, historical simulation, and Monte Carlo simulation.
Variance-covariance analysis assumes that market returns for financial products are
normally distributed, and VaR can be determined from market return’s variance and
covariance. The normal distribution hypothesis of variance-covariance analysis makes
it easy to estimate VaR at different reliability and different holding period
(see Appendix 1). Its major disadvantage is that the return in the financial market is
usually not in normal distribution and has fat tails. This means the probability of
extreme loss is more frequent than estimated by variance-covariance analysis.
Historical simulation assumes the future market return of the investment port-
folio is identical to the past returns; hence, the attributes of current market can be
used to simulate the future market return (Hendricks 1996; Hull and White 1998).
Historical simulation approach does not suffer from the tail-bias problem, for it
does not assume normal distribution. It relies on actual market return distribution,
and the estimation reflects what happened during the past sample period. It has
another advantage over variance-covariance analysis: it can be used for nonlinear
products, such as commodity derivatives. However, the problem with historical

2
Institutional use of VaR can be found in Basel (1995, 1998a, b, c, 1999), Danielsson et al. (1998),
and Danielsson Hartmann and de Vries (1998).
1280 K. Hung and S. Srivastava

simulation is its sensitivity to sample data. Many scholars pointed out that if
October 1987 is included into the observation period, then it would make great
difference to the estimation of VaR. Another problem with historical simulation is
that the left tail of actual return distribution is at zero stock prices. In other words, it
would not be accurate to assume a zero probability of loss that is greater than the
past loss. Lastly, the estimation of historical simulation is more complicated than
that of variance-covariance analysis. The VaR needs to be reestimated every time
the level of reliability or holding period changes.
Monte Carlo simulation can be used to generate future return distribution
for a wide range of financial products. It is done in two steps. First, a stochastic
process is specified for each financial variable along with appropriate parameters.
Second, simulated prices are determined for each variable, and portfolio loss is
calculated. This process is repeated 1,000 times to produce a probability distribu-
tion of losses. Monte Carlo simulation is the most powerful tool for generating
the entire probability distribution function and can be used to calculate VaR
for a wide range of financial products. However, it is time consuming and
expensive to implement.
Modern investment portfolio theories try to achieve optimal asset allocation via
maximizing the risk premium per unit risk, also known as the Sharpe ratio (Elton
and Gruber 1995). Within the framework of mean-variance, market risk is defined
as the expected probable variance of investment portfolio. To estimate risk with
standard deviation implies investors pay the same attention to the probabilities of
negative and positive returns. Yet investors have different aversion to investment’s
downside risk than to capital appreciation. Some investors may use semi-variance
to estimate the downside risk of investment. However, semi-variance has not
become popular.
Campbell et al. (2001) have developed an asset allocation model that takes VaR
as one of its constraints. This model takes the maximum expected loss preset by
risk managers (VaR) as a constraint to maximize expected return. In other words,
the optimal investment portfolio deduced from this model meets the constraint of
VaR. This model is similar to the mean-variance model that generates
the Sharpe index. If the expected return is a normal distribution, then this model
is identical with mean-variance model. Details of this model are presented in
the Appendix.
Other researchers have examined four models to introduce VaR for ex-ante
asset allocation of optimal investment portfolio: mean-variance (MV) model,
mini-max (MM) model, scenario-based stochastic programming (SP) model, and
a model that combines stochastic programming and aggregation/convergence
(SP-A). The investment portfolio constructed using the SP-A model has a higher
return in all empirical and simulation tests. Robustness test indicates that VaR
strategy results in higher risk tolerance than risk assessment that takes severe loss
into consideration. Basak and Shapiro (2001) pointed out that the drawback of risk
management lies in its focus on loss probability instead of loss severity. Although
loss probability is a constant, when severe loss occurs, it has greater negative
consequence than non-VaR risk management.
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1281

Lucas and Klaassen (1998) pointed out the importance of correctly assessing the
fat-tailed nature of return distribution. If the real return is not in normal distribution,
the asset allocation under the hypothesis of normal distribution will result in
non-efficiency or non-feasibility. An excellent discussion of VaR and risk
measurements is presented by Jorion (2001).

45.3 Empirical Results

In July 1997, financial crisis broke out in Southeast Asian nations and then prolifer-
ated to other Asian regions and led to a series of economic problems. Taiwan had also
been attacked by financial crisis in late 1998. Domestic financial markets fluctuated.
Corporations and individuals greatly suffered. The proliferation of financial crisis
within or among countries makes it impossible for corporations and individuals to
ignore market risk. Risk measurement is the first thing to do before investment.
The Taiwan market general weighted stock index, individual weighted stock price
indexes, and interbank short loan interest rate used in this research paper are obtained
from the data base of AREMOS. We divide the historical period into two groups,
from 1980 to 1999 and from 1991 to 1999, so as to analyze the impact of violent stock
market fluctuation on VaR estimation, such as the New York stock market collapse in
October 1987 and Taiwan stock market dramatic uprising from 1988 to 1990. The
first period is rather long and can indicate the nature of dramatic stock fluctuation.
The second period is rather short and can reflect the change of stock market tendency.
This paper employs historical simulation to reproduce the daily fluctuations of
returns for electrical machinery, cement, food, pulp and paper, plastics and
petroleum, and textile and fiber stocks trading in the Taiwan stock market during
the periods of 1980–1999 and 1991–1999. We estimate their VaRs under reliability
levels of 95 %, 97.5 %, and 99 %. The expected return of investment portfolio in
1999 is the sum of annual mean returns of various stocks multiplied with their
respective weights. We use this expected return to estimate year 1999 optimal stock
holding proportion and analyze the impact of different historical simulation periods
on optimal asset allocation.
Table 45.1 presents the summary of TSE general weighted stock index
(daily data) and estimated VaR for periods 1980–1999 and 1991–1999. Table 45.2
presents cumulative probability distributions of TSE daily index return for the
1980–1999 period and daily returns under the assumption of normality. It shows
that at confidence level lower than 95.8 % (e.g., 90 %), the left-tail probability for

Table 45.1 Summary of Taiwan Stock Exchange (TSE) daily index

Period Mean return Standard deviation Kurtosis VaR*
1980–1999 0.06 % 1.67 % 2.735 0.0263
1991–1999 0.04 % 1.60 % 2.464 0.0249
Annualized return for the two periods is 23.64 % and 11.87 %, respectively. VaR* is the maximum
expected return loss for 1-day holding period at a reliability level of 95 %
1282 K. Hung and S. Srivastava

Table 45.2 Cumulative probability distribution of TSE daily index

Taiwan index Historical data Normal distribution

Return (%) Cumulative probability Cumulative probability

7 0 1.20946E-05

6 0.004723 0.000144816

5 0.01102 0.001236891

4 0.021165 0.007577989

3 0.038482 0.033571597

2.9 0.040581 0.0382868

2.8 0.042855 0.043528507

2.7 0.045828 0.049334718

2.5 0.054749 0.062791449

2.3 0.066993 0.078949736

2 0.082736 0.108825855

1 0.185062 0.262753248
0 0.477873 0.485257048
1 0.775407 0.712585369
2 0.905895 0.876745371
3 0.960994 0.960522967
4 0.982158 0.990731272
5 0.992304 0.998424485
6 0.995977 0.999807736
7 1 0.999983254
Period: 1980–1999

historical distribution is higher than the normal return probability (4.2 %). Hence,
under the normal distribution assumption, the VaR is overestimated, and this leads
to an overcautious investment decision. At confidence level higher than 95.8 %
(e.g., 97.5 %), the left-tail probability for historical distribution is lower than the
normal return probability (4.2 %). Hence, under the normal distribution assump-
tion, the VaR is underestimated, and this leads to an overactive investment decision.
Figure 45.1 is the graphical presentation of the data in Table 45.2. The solid blue
line represents cumulative probability distributions of TSE daily index return, and the
dashed red line represents the normal distribution. The bottom panel is the enlarged
view of the left tail. This graph also indicates that VaR estimated using extreme
values of historical distribution will lead to an overactive investment decision.
Table 45.3 reports annualized returns and standard deviations for TSE daily
index and six selected industries: cement, electrical machinery, food, pulp and
paper, plastics and petroleum, and textile and fiber. For the 1980–1999 period,
the food industry had the greatest risk with a standard deviation of 53.47 % and
15.13 % annual return, whereas the overall market had a standard deviation of
50.80 % with 23.22 % annual return. Textile and fiber was the least risky industry
with a standard deviation of 41.21 % and 12.78 % annual return. For the 1991–1999
period, the electrical machinery industry had the greatest risk with a standard
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1283

1
Cumulative Probability Empirical
0.8
Normal
0.6

0.4

0.2

0
−8% −6% −4% −2% 0% 2% 4% 6% 8%
return
0.3

0.25 Empirical

0.2 Normal
Probability

0.15

0.1

0.05

0
−8% −7% −6% −5% −4% −3% −2% −1%
return

Fig. 45.1 Cumulative probability distribution of Taiwan daily stock index. Period: 1980–1999.
Left tail of the cumulative probability of Taiwan weighted daily stock index. Period 1980–1999.
Lower panel shows a fat left tail

Table 45.3 TSE index and selected industry’s returns and standard deviations

Period 1980–1998 1991–1998

Annual return Standard deviation Annual return Standard deviation
Industry (%) (%) (%) (%)
TSE index 23.22 50.80 9.40 36.58
Cement 12.48 48.37 0.45 24.92
Electrical 19.49 46.76 26.12 42.26
machinery
Food 15.13 53.47 9.55 32.29
Pulp and paper 8.50 45.78 4.64 39.41
Plastics and 13.32 43.99 11.39 36.26
petroleum
Textile and fiber 12.78 41.21 7.88 36.62
1284 K. Hung and S. Srivastava

0.250
r(p, 1980–1998)

0.200
expected return

0.150

0.100

0.050

0.000
72.500 73.000 73.500 74.000 74.500 75.000 75.500 76.000 76.500

risk

0.250
r(p, 1980–1998)

0.200
expected return

0.150

0.100

0.050

0.000
76.500 77.000 77.500 78.000 78.500 79.000 79.500
risk

Fig. 45.2 VaR-efficient frontier. The upper figure refers to investment portfolio of electrical
machinery and plastics and petroleum stocks, and the lower figure refers to investment portfolio of
cement and food stocks. VaR is set at reliability level of 95 %. Expected return and VaR are
estimated from TSE industry indexes from 1980 to 1998

deviation of 42.267 % and 26.12 % annual return, whereas the overall market had
a standard deviation of 36.58 % with 9.40 % annual return. The cement industry
was the least risky industry with a standard deviation of 24.92 % and 0.45 % annual
return. Next we constructed a two-industry optimal portfolio subject to VaR
constraint. The optimal asset allocation for the two-industry portfolio is obtained
by maximizing S(p) (derivation discussed in Appendix). The resulting
VaR-efficient frontiers are plotted in Fig. 45.2. The upper panel in Fig. 45.2 refers
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1285

Table 45.4 Optimal asset allocation for the two-industry portfolio obtained by maximizing S(p)
at different level of confidence
Confidence level
Portfolio choices 95 % 97.5 % 99 %
Electrical machinery cement {1,0}; {1,0}a {1,0}; {1,0} {1,0}; {1,0}
Electrical machinery food {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
Electrical machinery pulp and paper {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
Electrical machinery plastics and {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
petroleum
Electrical machinery textile and fiber {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
Cement food {0,1}; {0,1} {0,1}; {0,1} {0,1}; {0,1}
Cement pulp and paper {0,1}; {0,1} {0,1}; {0,1} {0,1}; {0,1}
Cement plastics and petroleum {0.12, 0.88}; {0.04, 0.96}; {0.17, 0.83};
{0,1} {0,1} {0,1}
Cement textile and fiber {0.67, 0.33}; {0.6, 0.4}; {0,1} {0.15, 0.85};
{0,1} {0,1}
Food pulp and paper {1,0}; {1,0} {1,0}; {1,0} {0.98, 0.02};
{0,1}
Food plastics and petroleum {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
Food textile and fiber {1,0}; {1,0} {1,0}; {1,0} {1,0}; {1,0}
Pulp and paper plastics and petroleum {0,1}; {0,1} {0,1}; {0,1} {0,1}; {0,1}
Pulp and paper textile and fiber {0,1}; {0,1} {0,1}; {0,1} {0,1}; {0,1}
Plastics and petroleum textile and fiber {0.99, 0.01}; {0.89, 0.11}; {1,0}; {1,0}
{1,0} {1,0}
First set {x, y} refers to the historical simulation for period 1980–1998, and second set {x, y} refers
to the historical simulation for period 1991–1998
a
{1, 0} represents 100 % investment in electrical machinery industry and 0 % investment in
cement industry

to VaR-efficient portfolios of electrical machinery and plastics and petroleum

stocks, and the lower panel in Fig. 45.2 refers to VaR-efficient portfolios of cement
and food stocks. VaR is set at the 95 % confidence level. Table 45.4 presents
portfolio weights for different combinations of industry stocks at different levels of
confidence. Asset allocations for most of the industry combinations represent
corner solutions, i.e., 100 % investment in one industry. For example, when
electrical machinery stocks are combined with stocks from any other industry, the
optimal portfolio is 100 % investment in electrical machinery stocks, for both the
time periods and the three confidence levels. Asset allocation for cement stocks is
dominated by the other five industry stocks. Allocations of food stocks dominate
all other stock weights except electrical machinery. The general nature of
asset allocation is the same for both time periods and the confidence levels.
Suppose an investor selects the VaR constraint for maximizing S(p) at
a specified level of confidence (say 95 %) and the actual VaR(c, p*) is at a higher
level (97.5 %); then VaR(portfolio) will be greater that target VaR*. In this case
investors will have to invest a portion of the fund in T-bills (B > 0, defined in
Appendix). This will make investment VaR(c, p*) equal to the VaR* in the preset
1286 K. Hung and S. Srivastava

Table 45.5 Optimal allocation for two-industry portfolio (historical simulation, period
1980–1998)
Confidence Cement Food Portfolio VaR Lending, Cement Food Cash
level (%) (%) (%) (c, p*) B yuan VaR* (%) (%) (%)
95 100 0 31.12 0 31.12 100 0 0
97.5 100 0 42.57 121 31.12 87.9 0 12.1
99 100 0 53.92 216 31.12 78.4 0 21.6
95 0 100 27.49 0 27.49 0 100 0
97.5 0 100 40.22 139 27.49 0 86.1 13.9
99 0 100 56.27 267 27.49 0 73.3 26.7
95 12 88 27.52 0 27.52 12 88 0
97.5 4 96 41.99 154 27.52 3.4 81.2 15.4
99 17 83 53.42 246 27.52 12.8 62.6 24.6
Pulp and Textile Portfolio Pulp and Textile
Confidence paper and fiber VaR(c, Lending, paper and fiber Cash
level (%) (%) (%) p*) B yuan VaR* (%) (%) (%)
95 0 100 29.58 0 29.58 0 100 0
97.5 0 100 41.97 132 29.58 0 86.8 13.2
99 0 100 53.85 230 29.58 0 77 23
Textile Textile
Confidence Cement and fiber Portfolio Lending, Cement and fiber Cash
level (%) (%) (%) VaR(c, p*) B yuan VaR* (%) (%) (%)
95 67 33 25.86 0 25.86 12 88 0
97.5 60 40 41.97 172 25.86 49.7 33.1 17.2
99 15 85 52.51 256 25.86 11.2 63.2 25.6
Plastics Plastics
and Textile Portfolio and Textile
Confidence petroleum and fiber VaR Lending, petroleum and fiber Cash
level (%) (%) (%) (c, p*) B yuan VaR* (%) (%) (%)
95 99 1 28.43 0 28.43 99 1 0
97.5 89 11 41.35 139 28.43 76.6 9.5 13.9
99 100 0 55.50 253 28.43 74.7 0 25.3
Confidence Food Textile and Portfolio Lending, Food Textile and Cash
level (%) (%) fiber (%) VaR(c, p*) B yuan VaR* (%) fiber (%) (%)
95 100 0 27.49 0 27.49 100 0 0
97.5 100 0 40.22 121 27.49 87.9 0 12.1
99 100 0 56.27 267 27.49 73.3 0 26.7
Two-industry portfolio with initial investment of 1,000 yuan
Allocations for other industry combinations are available to interested readers

constraint. In opposite case, the VaR* constraint is specified at a higher level than
the portfolio VaR(c, p*); then investors will borrow money to invest in risky assets
(B < 0). Tables 45.5 and 45.6 list examples of investment in two-industry stocks
and T-bills for periods 1980–1999 and 1991–1999, respectively. In each case, the
target VaR* in the preset constraint is at a 95 % level of confidence, and 1,000 yuan
is invested in the portfolio. In the first panel of Table 45.5, 1,000 yuan is invested in
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1287

Table 45.6 Optimal allocation for two-industry portfolio (historical simulation, period
1991–1998)
Confidence Cement Food Portfolio VaR Lending, Cement Food Cash
level (%) (%) (%) (c, p*) B yuan VaR* (%) (%) (%)
95 100 0 30.55 0 30.55 100 0 0
97.5 100 0 42.57 128 30.55 87.2 0 12.8
99 100 0 53.92 221 30.55 77.9 0 22.1
95 0 100 24.57 0 24.57 0 100 0
97.5 0 100 34.68 117 24.57 0 88.3 11.7
99 0 100 48.41 238 24.57 0 76.2 23.8
95 0 100 27.19 0 27.19 0 100 0
97.5 0 100 35.95 100 27.19 0 90 10
99 0 100 48.59 213 27.19 0 78.7 21.3
Pulp and Textile Portfolio Pulp and Textile
Confidence paper and fiber VaR(c, Lending, paper and fiber Cash
level (%) (%) (%) p*) B yuan VaR* (%) (%) (%)
95 0 100 27.75 0 27.75 0 100 0%
97.5 0 100 38.98 124 27.75 0 87.6 12.4
99 0 100 49.06 212 27.75 0 78.8 21.2
Textile Textile
Confidence Cement and fiber Portfolio Lending, Cement and fiber Cash
level (%) (%) (%) VaR(c, p*) B yuan VaR* (%) (%) (%)
95 0 100 27.75 0 27.75 0 100 0
97.5 0 100 38.98 124 27.75 0 87.6 12.4
99 0 100 49.06 212 27.75 0 78.8 21.2
Plastics Textile Plastics Textile
and and and and
Confidence petroleum fiber Portfolio Lending, petroleum fiber Cash
level (%) (%) (%) VaR(c, p*) B yuan VaR* (%) (%) (%)
95 100 0 27.19 0 27.19 0 100 0
97.5 100 0 35.95 100 27.19 0 90 10
99 100 0 48.59 213 27.19 0 78.7 21.3
Confidence Food Textile and Portfolio Lending, Food Textile and Cash
level (%) (%) fiber (%) VaR(c, p*) B yuan VaR* (%) fiber (%) (%)
95 100 0 24.57 0 24.57 100 0 0
97.5 100 0 34.68 117 24.57 88.3 0 11.7
99 100 0 48.41 238 24.57 76.2 0 23.8
Two-industry portfolio with initial investment of 1,000 yuan
Allocations for other industry combinations are available to interested readers

electrical machinery stocks and 0 in cement stocks. These allocations are from
Table 45.4. Portfolio VaR(c, p*) is 31.12, 42.57, and 53.92 yuan at 95 %, 97.5 %,
and 99 % respectively. This leads to the lending of 0, 121, and 216 yuan to meet the
target VaR of 31.12. Thus, a 97.5 % portfolio consists of 87.9 % in electrical
machinery stocks, 0 in cement stocks, and 21.1 % in T-bills.
1288 K. Hung and S. Srivastava

45.4 Conclusion

TSE daily index was found to be riskier than the market risk under the assumption of
normal distribution for market returns. This resulted in the left tail of cumulative return
distribution being fatter and a higher value at risk, indicating an overactive investment
activity. For asset allocation model under a constrained VaR framework, most of the
optimal portfolios have predominant investment in stocks from one industry. Hence, it
will be inappropriate to comment on the optimal allocation of future investment
portfolio based on the past stock performance of this unique period studied.

Appendix 1: Value at Risk

Let W0 be the initial investment and R be the rate of return of a portfolio. The value
of the portfolio at the end of the target horizon will be W ¼ W0(1 + R). Let m and s
be the expected return and standard deviation of R. The lowest portfolio value at the
confidence level c is defined as W* ¼ W0 (1 + R*). The relative VaR is the dollar
loss relative to the mean:

VaRðmeanÞ ¼ EðW Þ
W  ¼
W 0 ðR
mÞ (45.1)

The absolute VaR is the dollar loss relative to zero:

VaRðzeroÞ ¼ W 0
W  ¼
W 0 R (45.2)

W* and R* are minimum value and cutoff return, respectively. In this paper we are
discussing absolute VaR. The general form of VaR can be derived from the probability
distribution of the future portfolio value f(w). For a given confidence level c, the worst
possible portfolio value W* is such that probability of exceeding W* is c:
ð1
c¼ f ðwÞdw (45.3)
W

The probability of a value lower than W*, p ¼ P(w  W*) is 1–c:

ð W

p ¼ Pðw  W Þ ¼ 1
c ¼ f ðwÞdw (45.4)

1

Typical confidence level c is 95 %. This computation of VaR does not require

estimation of variance-covariance matrix.
When portfolio returns are normally distributed, then distribution f(w) can be
translated into a standard normal distribution F(e), where e has mean zero and
standard deviation of one. VaR can be determined from the tables of the cumulative
standard normal distribution:
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1289

ð 1
c
VaR ¼ Nð1
cÞ ¼ ’ðeÞde (45.5)

1

and the cutoff return R* ¼
zs + m. This Appendix is based on Jorion (2001).

Details of normal distribution can be found in Johnson and Wichern (2007).

Appendix 2: Optimal Portfolio Under a VaR Constraint

We present an asset allocation model under a value-at-risk constraint. This model

sets maximum expected loss not to exceed the VaR for a selected investment
horizon, T at a given confidence level. Then asset proportions are allocated across
the portfolio such that the wealth at the end of investment horizon is maximized.
Suppose W0 is the investor’s initial wealth and B is the amount that the investor can
borrow (B > 0) or lend (B < 0) at the risk-free interest rate rf. Let n be the number of
risky assets, gi be the fraction invested in risky asset i, and P(i, t) be the price of
asset I at time t. Then the initial value of the portfolio
X
n
W0 þ B ¼ gi Pði; 0Þ (45.6)
i¼1

represents the budget constraint.

Let VaR* be the target VaR consistent with investor’s risk aversion and WT be
the wealth at the end of the holding period, T. The downside risk constraint can be
written as

PrfðW 0
W T Þ  VaR g  ð1
cÞ (45.7)

where Pr {.} denotes the expected probability conditioned on information available

at time, t ¼ 0, and c is the confidence level. Equation 45.2 can be written as

PrfW T  ðW 0
VaR Þg  ð1
cÞ (45.8)

Let rp be the total portfolio return at the end of the holding period and T then the
expected wealth at the end of holding period; T can be written as
 
EðW T Þ ¼ ðW 0 þ BÞ 1 þ rp
Bð1 þ rf Þ (45.9)

Investor’s constrained wealth maximizing objective can be written as

Max: EðW T Þ
(45.10)
s:t: PrfW T  ðW 0
VaR Þg  ð1
cÞ

Performance measure S(p) and borrowed amount can be deduced from Eq. 45.5.
Let p* be the maximizing portfolio and q(c, p) defines the quantile that corresponds
to probability (1
c) which can be obtained from portfolio return’s cumulative
density function. Maximizing portfolio p* is defined as
1290 K. Hung and S. Srivastava

p : max SðpÞ ¼ rp
rf
(45.11)
p
W 0 r f
W 0 qðc; pÞ

Initial wealth in the denominator of Eq. 45.6 is a scale constant and does not
affect the asset allocation. Let VaR(c, p) denote portfolio p’s VaR, and then the
denominator of Eq. 45.6 can be written as

Fðc; pÞ ¼ W0 rf
VaRðc; pÞ (45.12)

If we consider rf as the benchmark return, then F(c, p) represents potential for

portfolio losses at the confidence level c. Performance measure S(p) represents
Sharpe-like reward-risk ratio, and optimization problem becomes
Optimal portfolio:

rp
rf
p : max SðpÞ ¼
p Fðc; pÞ

Optimal portfolio allocation is independent of the initial wealth. It is also

independent of the target VaR*. Risk measure F(c, p*) depends on VaR(c, p*)
and not on VaR*. Investors first allocates the wealth among risky assets and then
decides borrowing or lending depending on the value of {VaR*
VaR(c, p*)}.
Borrowed amount B can be written as

W 0 ðVaR
VaRðc; p ÞÞ
B¼
F0 ðc, p0 Þ

If {VaR*
VaR(c, p*)} is positive, then there is an opportunity to increase the

portfolio return by borrowing at the risk-free rate and invest it in risky asset. If
{VaR*
VaR(c, p*)} is negative, then the portfolio risk needs to be reduced by
investing a portion of the initial wealth in the risk-free asset. In either case
the relative proration of funds invested in risky assets remains the same.
Since VaR(c, p*) depends on the choice of holding period, confidence level, VaR
estimation technique, and the assumption regarding the expected return distribution,
the borrowing (B > 0) or lending (B < 0) will also change. This Appendix is based on
Campbell et al. (2001).

References
Basak, S., & Shapiro, A. (2001). Value-at-risk based risk management: Optimal policies and asset
prices. The Review of Financial Studies, 14(2), 371–405.
Beder, T. S. (1995). VAR: Seductive but dangerous. Financial Analysts Journal, 51(5), 12–24.
Campbell, R., Huisman, R., & Koedijk, K. (2001). Optimal portfolio selection in a value-at-risk
framework. Journal of Banking and Finance, 25, 1789–1804.
Dowd, K. (1998). Beyond value at risk: The new science of risk management. New York: Wiley.
45 Optimal Asset Allocation Under VaR Criterion: Taiwan Stock Market 1291

Elton, E. J., & Gruber, M. J. (1995). Modern portfolio theory and investment analysis (5th ed.).
New York: Wiley.
Fong, G., & Vasicek, O. A. (1997). A multidimensional framework for risk analysis. Financial
Analysts Journal, 53, 51–57.
Hendricks, D. (1996). Evaluation of value-at-risk models using historical data. FRBNY Economic
Policy Review, 2, 39–69.
Hoppe, R. (1999). It’s time we buried value-at-risk. Risk Professional, 1. October page 18.
London.
Hull, J., & White, A. (1998). Incorporating volatility updating into the historical simulation
method for value-at-risk. Journal of Risk, 1, 5–19.
Johnson, R., & Wichern, D. (2007). Applied multivariate statistical analysis (6th ed.). New York:
Pearson Education.
Jorion, P. (1997). In defense of VAR. Derivatives Strategy, 2, 20–23.
Jorion, P. (2001). Value at risk (2nd ed.). New York: McGraw Hill.
Lucas, A., & Klaassen, P. (1998). Extreme returns, downside risk, and optimal asset allocation.
Journal of Portfolio Management, 25(1), 71–79.
Morgan, J. P. (1996). RiskMetricsTM – technical document. New York: Morgan Guaranty Trust
Company.
Roy, A. D. (1952). Safety first and the holding of assets. Econometrica, 20, 431–449.
Schachter, B. (1998). An irreverent guide to value at risk. Risks and Rewards, (1), 17–18.
Smithson, C., & Minton, L. (1996a). Value-at-risk (1): The debate on the use of VaR. Risk, 9,
25–27.
Smithson, C., & Minton, L. (1996b). Value-at-risk (2): The debate on the use of VaR. Risk, 9,
38–39.
Talmor, S. (1996). History repeat itself. The Banker, (1), 75–76.
Alternative Methods for Estimating Firm’s
Growth Rate 46
Ivan E. Brick, Hong-Yi Chen, and Cheng-Few Lee

Contents
46.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1294
46.2 The Discounted Cash Flow Model and the Gordon Growth Model . . . . . . . . . . . . . . . . . . 1295
46.3 Internal Growth Rate and Sustainable Growth Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . . 1300
46.4 Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1303
46.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1309
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1309

Abstract
The most common valuation model is the dividend growth model. The growth rate
is found by taking the product of the retention rate and the return on equity. What
is less well understood are the basic assumptions of this model. In this paper, we
demonstrate that the model makes strong assumptions regarding the financing mix
of the firm. In addition, we discuss several methods suggested in the literature on
estimating growth rates and analyze whether these approaches are consistent with
the use of using a constant discount rate to evaluate the firm’s assets and equity.

This chapter is a slightly revised version of Chapter 64 of Encyclopedia of Finance, 2nd Edition
and Brick et al. (2014).
I.E. Brick (*)
Department of Finance and Economics, Rutgers, The State University of New Jersey,
Newark/New Brunswick, NJ, USA
e-mail: ibrick@andromeda.rutgers.edu
H.-Y. Chen
Department of Finance, National Central University, Taoyuan, Taiwan
e-mail: fnhchen@ncu.edu.tw
C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: lee@business.rutgers.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1293
DOI 10.1007/978-1-4614-7750-1_46,
# Springer Science+Business Media New York 2015
1294 I.E. Brick et al.

The literature has also suggested estimating growth rate by using the average
percentage change method, compound-sum method, and/or regression methods.
We demonstrate that the average percentage change is very sensitive to extreme
observations. Moreover, on average, the regression method yields similar but
somewhat smaller estimates of the growth rate compared to the compound-sum
method. We also discussed the inferred method suggested by Gordon and Gordon
(1997) to estimate the growth rate. Advantages, disadvantages, and the interrela-
tionship among these estimation methods are also discussed in detail.

Keywords
Growth rate • Discount cash flow model • Internal growth rate • Sustainable
growth rate • Compound sum method

46.1 Introduction

One of the more highly used valuation models is that developed by Gordon and
Shapiro (1956) and Gordon (1962) known as the dividend growth model. In security
analysis and portfolio management, growth rate estimates of earnings, dividends, and
price per share are important factors in determining the value of an investment or
a firm. These publications demonstrate that the growth rate is found by taking the
product of the retention rate and the return on equity. What is less well understood are
the basic assumptions of this model. In this paper, we demonstrate that the model
makes strong assumptions regarding the financing mix of the firm.
In addition, we will also discuss several methods suggested in the literature on
estimating growth rates. We will analyze whether these approaches are consistent
with the use of using a constant discount rate to evaluate the firm’s assets and
equity. In particular, we will demonstrate that the underlying assumptions of the
internal growth rate model (whereby no external funds are used to finance growth)
is incompatible with the constant discount rate model of valuation.
The literature has also suggested estimating growth rate by taking the average of
percentage change of dividends over a sample period, taking the geometric average
of the change in dividends or using regression analysis to estimate the growth rate
(e.g., Lee et al. 2009; Lee et al. 2012; Lee et al. 2000; and Ross et al. 2010). Gordon
and Gordon (1997) suggest first using the Capital Asset Pricing Model (CAPM) to
determine the cost of equity of the firm and then using the dividend growth model to
infer the growth rate. Advantages, disadvantages, and the interrelationship among
these estimation methods are also discussed in detail.
This paper is organized as follows. In Sect. 46.2 we present the Gordon and
Shapiro model (1956). We discuss the inherent assumptions of the model and its
implied method to estimate the growth rate. Section 46.3 analyzes the internal
growth rate and sustainable growth rate models. Section 46.4 describes leading
statistical methods for estimating firm’s growth rates. We will also present the
46 Alternative Methods for Estimating Firm’s Growth Rate 1295

inferred method suggested by Gordon and Gordon (1997) to estimate the growth
rate. Concluding remarks appear in Sect. 46.5.

46.2 The Discounted Cash Flow Model and the Gordon

Growth Model

The traditional academic approach to evaluate a firm’s equity is based upon the
constant discount rate method. One approach uses the after-tax weighted average
cost of capital as a discount rate. This model is expressed as:
X
1
CFut
Value of Equity ¼  Debtt , (46:1)
t¼1 ð1 þ ATWACOCÞt

where CFut is the expected unlevered cash flow of the firm at time t and Debtt is the
market value of debt outstanding. ATWACOC equals L(1  t)Rd + (1  L)r where
L is the market value proportion of debt, t is the corporate tax rate, Rd is the cost of
debt and r is the cost of equity. The first term on the right hand side of Eq. 46.1 is the
value of the assets. Subtracting out the value of debt yields the value of equity. The
price per share is therefore the value of equity divided by the number of shares
outstanding. Alternatively, the value of equity can be directly found by discounting
the dividends per share by the cost of equity, or more formally:
X
1
dt
Value of Common Stock ðP0 Þ ¼ , (46:2)
t¼1 ð1 þ r Þt

where dt is the dividend per share at time t. Boudreaux and Long (1979), and
Chambers et al. (1982) demonstrate the equivalence of these two approaches
assuming that the level of that the level of debt is a constant percentage of the
value of the firm.1 Accordingly:
X
1
Xt
 Debtt
t¼1 ð1 þ ATWACOCÞt X
1
dt
¼ (46:3)
#of Shares Outstaning t¼1 ð1 þ r Þt

If we assume that dividends per share grow at a constant rate g, then Eq. 46.2 is
reduced to the basic dividend growth model2:
d1
P0 ¼ : (46:4)
ð r  gÞ

1
See Brick and Weaver (1984, 1997) concerning the magnitude of error in the valuation using
a constant discount rate when the firm does not maintain a constant market based leverage ratio.
2
Gordon and Shapiro’s (1956) model assume that dividends were paid continuously and hence
P0 ¼ d1/(r  g).
1296 I.E. Brick et al.

Gordon and Shapiro (1956) demonstrates that if b is the fraction of earnings

retained within the firm, and r is the rate of return the firm will earn on all new
investments, then g ¼ br. Let It denote the level of new investment at time t.
Because growth in earnings arises from the return on new investments, earnings can
be written as:

Et ¼ Et1 þ rI t1 , (46:5)

where Et is the earnings in period t.3 If the firm’s retention rate is constant and used
in new investment, then the earnings at time t is

Et ¼ Et1 þ rbEt1 ¼ Et1 ð1 þ rbÞ: (46:6)

Growth rate in earnings is the percentage change in earnings and can be

expressed as

Et  Et1 Et1 ð1 þ rbÞ  Et1

gE ¼ ¼ ¼ rb: (46:7)
Et1 Et1

If a constant proportion of earnings is assumed to be paid out each year, the

growth in earnings equals the growth in dividends, implying g ¼ br. It is worth-
while to examine the implication of this model for the growth in stock prices over
time. The growth in stock price is

Ptþ1  Pt
gP ¼ : (46:8)
Pt

Recognizing that Pt and Pt+1 can be defined by Eq. 46.4, while noting that dt+2 is
equal to dt+1(1 + br) then:

d tþ2 dtþ1
 d tþ2  d tþ1 dtþ1 ð1 þ br Þ  d tþ1
gP ¼ k  rb k  rb ¼ ¼ ¼ br: (46:9)
d tþ1 dtþ1 dtþ1
k  rb

Thus, under the assumption of a constant retention rate, for a one-period model,
dividends, earnings, and prices are all expected to grow at the same rate.
The relationship between the growth rate, g, the retention rate, b, and the return
on equity, r, can be expanded to a multi-period setting as the following numerical
example illustrates. In this example, we assume that the book value of the firm’s
assets equal the market value of the firm. We will assume that the growth rate of the
firm sales and assets is 4 % and the tax rate is equal to 40 %. The book value of the

3
Earnings in this model are defined using the cash-basis of accounting and not on an accrual basis.
46 Alternative Methods for Estimating Firm’s Growth Rate 1297

assets at time 0 is $50 and we assume a depreciation rate of 10 % per annum. The
amount of debt outstanding is $12.50 and amount of equity outstanding is $37.50.
We assume that the cost of debt, Rd, is 12 % and the cost of equity, r, is 25 %,
implying an ATWACOC of 20.55 %. The expected dividend at t ¼ 1, d1, must
satisfy Eq. 46.4. That is, 37.50 ¼ d1/(0.25  0.04).
The unlevered cash flow is defined as Sales less Costs (excluding the depreci-
ation expense) less Investment less the tax paid. Tax paid is defined as the tax rate
(which we assume to be 40 %) times Sales minus Costs minus the Depreciation
Expense. Recognizing that the value of the firm is given by CFu1/(ATWACOC  g),
if firm value is $50, g ¼ 4 % and ATWACOC is 20.55 %, then the expected
unlevered cash flow is at time 1 is $8.28. We assume that the asset turnover ratio
is 1.7. Hence, if assets at time 0 is $50, the expected sales at time 1 is $85. To obtain
the level of investment, note that the depreciation expense at time 1 is $5. If the
book value of assets equals $52, then the firm must invest $7. To obtain an expected
unlevered cash flow at t ¼ 1 of $8.28, the Gross Profit Margin is assumed to be
approximately 26.03 %, resulting in expected costs at time 1 of $62.88. The interest
expense at time 1, is the cost of debt times the amount of debt outstanding at time
zero, or $1.50. The Earnings Before Taxes (EBT) is defined as Sales – Costs –
Interest Expense – Depreciation Expense, which equals $15.63 at time 1. 40 % of
EBT is the taxes paid or $6.25 resulting in a net income (NI) of $9.38. ROE, which
equals Net Income/Book Value of Equity at the beginning of the period is 25 %.
Since the aggregate level of dividends at time 1 is $7.88, then the dividend payout
ratio (1  b) is 84 %. Note that b is therefore equal to 16 % and b  ROE ¼ 4%.4
Further note that the firm will increase its book value of equity via retention of
NI by $1.50 (RE in the table). In order to maintain a leverage ratio of 25 %, the firm
must increase the level of debt from time 0 to time 1 by $0.50. The entries for time
periods 2–5 follow the logical extension of the above discussion, and as shown in
the table, the retention rate b is 16 % and ROE ¼ 25 % for each period. Again the
product of b and ROE results in the expected growth rate of 4 %. Further note, that
g ¼ 4 % imply that sales, costs, book value of asset, depreciation, unlevered cash
flow, cash flow to stockholders, value of debt and value of equity to increase by 4 %
per annum.
Investors may use a one-period model in selecting stocks, but future profitability
of investment opportunities plays an important role in determining the value of the
firm and its EPS and dividend per share. The rate of return on new investments can
be expressed as a fraction, c (perhaps larger than 1), of the rate of return security
holders require (r):

k ¼ cr: (46:10)

4
Generally, practioners define ROE as the ratio of the Net Income to the end of year Stockholders
Equity. Here we are defining ROE as the ratio of the Net Income to the beginning of the year
Stockholders Equity. Brick et al. (2012) demonstrate that the practitioner’s definition is one of the
sources for the Bowman Paradox reported in the Organization Management literature.
1298 I.E. Brick et al.

Table 46.1 The book value of the firm’s assets equal the market value of the firm (growth rate
is 4 %)
0 1 2 3 4 5
Assets $50.00 $52.00 $54.08 $56.24 $58.49 $60.83
Debt $12.50 $13.00 $13.52 $14.06 $14.62 $15.21
Equity $37.50 $39.00 $40.56 $42.18 $43.87 $45.62
Rd 0.12 0.12 0.12 0.12 0.12 0.12
r 0.25 0.25 0.25 0.25 0.25 0.25
ATWACOC 0.2055 0.2055 0.2055 0.2055 0.2055 0.2055
Asset turnover 1.7 1.7 1.7 1.7 1.7
GPM 0.26029 0.26029 0.26029 0.26029 0.26029
Sales $85.00 $88.40 $91.94 $95.61 $99.44
Cost $62.88 $65.39 $68.01 $70.73 $73.55
Depreciation $5.00 $5.20 $5.41 $5.62 $5.85
Interest exp. $1.50 $1.56 $1.62 $1.69 $1.75
EBT $15.63 $16.25 $16.90 $17.58 $18.28
Tax $6.25 $6.50 $6.76 $7.03 $7.31
NI $9.38 $9.75 $10.14 $10.55 $10.97
DIV $7.88 $8.19 $8.52 $8.86 $9.21
New debt $.50 $0.52 $0.54 $0.56 $0.59
CFu $8.28 $8.61 $8.95 $9.31 $9.68
Firm value $50.00 $52.00 $54.08 $56.24 $58.49 $60.83
Investment $7.00 $7.28 $7.57 $7.87 $8.19
Vequity $37.50 $39.00 $40.56 $42.18 $43.87 $45.62
RE $1.50 $1.56 $1.62 $1.69 $1.75
ROE 0.25 0.25 0.25 0.25 0.25
1-b 0.84 0.84 0.84 0.84 0.84
g 0.04 0.04 0.04 0.04 0.04

Substituting this into the well-known relationship that r ¼ Pd10 þ g and

rearranging, we have

ð1  bÞE1
k¼ : (46:11)
ð1  cbÞP0

If a firm has no extraordinary investment opportunities (r ¼ k), then c ¼ 1 and

the rate of return that security holders require is simply the inverse of the stock’s
price to earnings ratio. In our example of Table 46.1, NI at time 1 is $9.38 and the
value of equity at time 0 is $37.50. The ratio of these two numbers (which is
equivalent to EPS/P) is ROE or 25 %.
On the other hand, if the firm has investment opportunities that are expected to
offer a return above that required by the firm’s stockholders (c > 1), the earnings to
price ratio at which the firm sells will be below the rate of return required by
investors. To illustrate consider the following example whereby market value of the
46 Alternative Methods for Estimating Firm’s Growth Rate 1299

Table 46.2 The market value of the firm and equity is greater than its book value
0 1 2 3 4 5
Assets $50.00 $52.00 $54.08 $56.24 $58.49 $60.83
Firm value $60.00 $62.40 $64.90 $67.49 $70.19 $73.00
Debt $12.50 $13.00 $13.52 $14.06 $14.62 $15.21
Equity $47.50 $49.40 $51.38 $53.43 $55.57 $57.79
Rd 0.12 0.12 0.12 0.12 0.12 0.12
r 0.25 0.25 0.25 0.25 0.25 0.25
ATWACOC 0.2129 0.2129 0.2129 0.2129 0.2129 0.2129
Asset turnover 1.7 1.7 1.7 1.7 1.7
GPM 0.3093 0.3093 0.3093 0.3093 0.3093
Sales $85.00 $88.40 $91.94 $95.61 $99.44
Cost $58.71 $61.06 $63.50 $66.04 $68.68
Depreciation $5.00 $5.20 $5.41 $5.62 $5.85
Interest exp. $1.50 $1.56 $1.62 $1.69 $1.75
EBT $19.79 $20.58 $21.41 $22.26 $23.15
Tax $7.92 $8.23 $8.56 $8.91 $9.26
NI $11.88 $12.35 $12.84 $13.36 $13.89
DIV $9.98 $10.37 $10.79 $11.22 $11.67
New debt $.50 $0.52 $0.54 $0.56 $0.59
CFu $10.38 $10.79 $11.22 $11.67 $12.14
Firm value $60.00 $62.40 $64.90 $67.49 $70.19 $73.00
Investment $7.40 $7.70 $8.00 $8.32 $8.66
Vequity $47.50 $49.40 $51.38 $53.43 $55.57 $57.79
RE $1.90 $1.98 $2.06 $2.14 $2.22
Market based ROE 0.25 0.25 0.25 0.25 0.25
1-b 0.84 0.84 0.84 0.84 0.84
g 0.04 0.04 0.04 0.04 0.04

firm and equity is greater than its book value. This example is depicted in
Table 46.2. The basic assumptions of the model is as follows: We will assume
that the growth rate of the firm sales and book value of the assets is 4 %. The book
value of the assets at time 0 is again $50 and we assume a depreciation rate of 10 %
per annum. However, note that the market value of the firm is $60. The entries for
Debt and Equity represent market values. The amount of debt outstanding is $12.50
and amount of equity outstanding is now $47.50. We assume that the cost of debt,
Rd, is 12 % and the cost of equity, r, is 25 %, implying an ATWACOC of 21.29 %.
For the valuation of the firm to be internally consistent, the unlevered cash flow at
time 1 is $10.38. Similarly, the value of equity to be internally consistent, the
expected dividends at t ¼ 1 is $9.98. Note that net income is $11.88 implying
a dividend payout ratio of 84 % and a retention rate of 16 %. The book value based
ROE, k, is found by taking the net income divided by the book value of equity. In
our example, implied book value of equity is $37.50. Hence, k ¼ 31.68 %, implying
1300 I.E. Brick et al.

that the book value ROE is greater than the cost of equity which is the required rate
of return. But g is given by the market value based ROE which is defined as Net
Income over market value of equity. That is r ¼ 25 %. Note again, br is 4 %.
An investor could predict next year’s dividends, the firm’s long-term growth
rate, and the rate of return stockholders require (perhaps using the CAPM to
estimate r) for holding the stock. Equation 46.4 could then be solved for the
theoretical price of the stock that could be compared with its present price. Stocks
that have theoretical prices above actual price are candidates for purchase; those
with theoretical prices below their actual price are candidates for sale or for
short sale.

46.3 Internal Growth Rate and Sustainable Growth Rate Models

The internal growth rate model assumes that the firm can only finance its growth by
its internal funds. Consequently, the cash to finance growth must come from only
retained earnings. Therefore, retained earnings can be expressed as

Retained Earnings ¼ Earnings  Dividends

¼ Profit Margin  Total Sales-Dividends
(46:12)
¼ pðS þ DSÞ  pðS þ DSÞð1  bÞ
¼ pbðS þ DSÞ,

where
p ¼ the profit margin on all sales;
S ¼ annual sales; and
DS ¼ the increase in sales during the year.
Because retained earnings is the only source of new funds, the use of cash
represented by the increase in assets must equal the retained earnings:

Uses of Cash ¼ Sources of Cash

Increases in Assets ¼ Retained earnings

DST ¼ pðS þ DSÞb

¼ pbS þ pbDS,

DS½T  pb ¼ pSb,

DS pb
¼ , (46:13)
S T  pb

where T ¼ the ratio of total assets to sales. If we divide both numerator and
denominator of Eq. 46.13 by T and make rearrange the terms, then we can show
that the internal growth rate is:
46 Alternative Methods for Estimating Firm’s Growth Rate 1301

DS pb=T b  ROA
g¼ ¼ ¼ , (46:14)
S 1  pb=T 1  b  ROA

where ROA is the return on assets. The internal growth rate is the maximum growth
rate that can be achieved without debt or equity kind of external financing. But note
this assumption of not issuing new debt or common stock to finance growth is
inconsistent with the basic assumption of the constant discount rate models that the
firm maintains a constant market based leverage ratio. Hence, this model cannot be
used to estimate the growth rate and be employed by the Gordon Growth Model.
Higgins (1977, 1981, 2008) has developed a sustainable growth rate under
assumption that firms can generate new funds by using retained earnings or issuing
debt, but not issuing new shares of common stock. Growth and its management
present special problems in financial planning. From a financial perspective, growth
is not always a blessing. Rapid growth can put considerable strain on a company’s
resources, and unless management is aware of this effect and takes active steps to
control it, rapid growth can lead to bankruptcy. Assuming a company is not raising
new equity, the cash to finance growth must come from retained earnings and new
borrowings. Further, because the company wants to maintain a target debt-to-equity
ratio equal to L, each dollar added to the owners’ equity enables it to increase its
indebtedness by $L. Since the owners’ equity will rise by an amount equal to
retained earnings, the new borrowing can be written as:

New Borrowings ¼ Retained Earnings  Target Debt-to-Equity Ratio

¼ pbðS þ DSÞL:

The use of cash represented by the increase in assets must equal the two sources
of cash (retained earnings and new borrowings)5:

Uses of Cash ¼ Sources of Cash

Increases in Assets ¼ Retained Earnings þ New Borrowing


DST ¼ pbðS þ DSÞ þ pbðS þ DS L
¼ pbð1 þ LÞS þ pbð1 þ LÞDS

DS½T  pbð1 þ LÞ ¼ pbð1 þ LÞS

DS pbð1 þ LÞ
g¼ ¼ : (46:15)
S T  pbð1 þ LÞ

5
Increased in Assets is the net increase in assets. The total investment should also include the
depreciation expense as can be seen in our examples delineated in Tables 46.1 and 46.2. But
depreciation expense is also a source of funding. Hence, it is netted out in the relationship between
increases in assets and retained earnings and new borrowings.
1302 I.E. Brick et al.

In Eq. 46.15 the DS/S or g is the firm’s sustainable growth rate assuming no
infusion of new equity. Therefore, a company’s growth rate in sales must equal the
indicated combination of four ratios, p, b, L, and T. In addition, if the company’s
growth rate differs from g, one or more of the ratios must change. For example,
suppose a company grows at a rate in excess of g, then it must either use its assets
more efficiently, or it must alter its financial policies. Efficiency is represented by the
profit margin and asset-to-sales ratio. It therefore would need to increase its profit
margin (p) or decrease its asset-to-sales ratio (T) in order to increase efficiency. Financial
policies are represented by payout or leverage ratios. In this case, a decrease in its payout
ratio (1-b) or an increase in its leverage (L) would be necessary to alter its financial
policies to accommodate a different growth rate. It should be noted that increasing
efficiency is not always possible and altering financial policies are not always wise.
If we divide both numerator and denominator of Eq. 46.15 by T and rearrange
the terms, then we can show that the sustainable growth rate can be shown as

DS pbð1 þ LÞ=T b  ROE

g¼ ¼ ¼ : (46:16)
S 1  pbð1 þ LÞ=T 1  b  ROE

Please note that, in the framework of internal growth rate and sustainable growth
rate presented above, the source of cash are taken from the end of period values of
assets and assumed that the required financing occurs at the end of the period.
However, Ross et al. (2010) show that if the source of cash is from the beginning of
the period, the relationship between the use and the source of cash can be expressed
for the internal growth rate model as DST ¼ pSb and for the sustainable growth rate
model, DST ¼ pbS + pbSL . Such relationship will result an internal growth rate of
b  ROA and a sustainable growth rate of b  ROE. For example, Table 46.3
assumes identical assumptions to that of Table 46.1, but now we will assume
a growth rate of 4.1667 % and use total asset, total equity, and total debt from the
beginning of the period balance sheet to calculate the net income. Recall that ROE
is the net income divided by stockholders’ equity at the beginning of the period.
Note that the product of ROE and b will yield 4.1667 %.
Note that the intent of the Higgins’ sustainable growth rate allows only internal
source and external debt financing. Chen et al. (2013) incorporate Higgins (1977)
and Lee et al. (2011) frameworks, allowing company use both external debt and
equity, and derive a generalized sustainable growth rate as

b  ROE l  Dn  P=E
gðtÞ ¼ þ , (46:17)
1  b  ROE 1  b  ROE

where
l ¼ degree of market imperfection;
Dn ¼ number of shares of new equity issued;
P ¼ price per share of new equity; and
E ¼ total equity:
46 Alternative Methods for Estimating Firm’s Growth Rate 1303

Table 46.3 The book value of the firm’s assets equal the market value of the firm (sustainable
growth rate is 4.1667 %)
0 1 2 3 4 5
Assets $50.00 $52.08 $54.25 $56.51 $58.87 $61.32
Value $50.00 $52.08 $54.25 $56.51 $58.87 $61.32
Debt $12.50 $13.02 $13.56 $14.13 $14.72 $15.33
Equity $37.50 $39.06 $40.69 $42.39 $44.15 $45.99
R 0.12 0.12 0.12 0.12 0.12 0.12
Re 0.25 0.25 0.25 0.25 0.25 0.25
ATWACOC 0.2055 0.2055 0.2055 0.2055 0.2055 0.2055
Asset turnover 1.7 1.7 1.7 1.7 1.7
GPM 0.26029 0.26029 0.26029 0.26029 0.26029
Sales $85.00 $88.54 $92.23 $96.07 $100.08
Cost $62.88 $65.49 $68.22 $71.07 $74.03
Depreciation $5.00 $5.21 $5.43 $5.65 $5.89
Interest exp. $1.50 $1.56 $1.63 $1.70 $1.77
EBT $15.63 $16.28 $16.95 $17.66 $18.40
Tax $6.25 $6.51 $6.78 $7.06 $7.36
NI $9.38 $9.77 $10.17 $10.60 $11.04
DIV $7.81 $8.14 $8.48 $8.83 $9.20
New debt $7.60 $7.92 $8.25 $8.59 $8.95
CFu $8.19 $8.53 $8.89 $9.26 $9.64
Value $50.00 $52.08 $54.25 $56.51 $58.87 $61.32
Investment $7.08 $7.38 $7.69 $8.01 $8.34
RE $1.56 $1.63 $1.70 $1.77 $1.84
ROE 0.25 0.25 0.25 0.25 0.25
(1-b) 0.833333 0.833333 0.833333 0.833333 0.833333
g 0.041667 0.041667 0.041667 0.041667 0.041667

Comparing Eq. 46.17, the generalized sustainable growth rate has an additional
positive term, 1lDnp=E
ð1DÞROE , when the new equity issue is taken into account.
Therefore, Chen et al. (2013) show that Higgins’ (1977) sustainable growth rate
is underestimated because of the omission of the source of the growth related to new
equity issue.

46.4 Statistical Methods

Instead of relying on financial ratios to estimate firm’s growth rates, one may use
statistical methods to determine firm’s growth rates. A simple growth rate can be
estimated by calculating the percentage change in earnings over a time period, and
taking the arithmetic average. For instance, the growth rate in earnings over one
period can be expressed as:
1304 I.E. Brick et al.

Et  Et1
gt ¼ : (46:18)
Et1

The arithmetic average is given by

1X n
g¼ g: (46:19)
n t¼1 t

A more accurate estimate can be obtained by solving for the compounded

growth rate:

X t ¼ X 0 ð 1 þ gÞ t , (46:20)

or
 1=t
Xt
g¼  1, (46:21)
X0

where
X0 ¼ measure in the current period (measure can be sales, earnings, or
dividends); and
Xt ¼ measure in period t.
This method is called the discrete compound sum method of growth-rate esti-
mation. For this approach to be consistent with the dividend growth model, the
duration of each period (e.g., quarterly or yearly) must be consistent with the
compounding period used in the dividend growth model.
Another method of estimating the growth rate uses the continuous compounding
process. The concept of continuous compounding process can be expressed math-
ematically as

Xt ¼ X0 egt : (46:22)

Equation 46.21 describes a discrete compounding process and Eq. 46.22

describes a continuous compounding process. The relationship between
Eqs. 46.21 and 46.22 can be illustrated by using an intermediate expression such as:

 g mt
X t ¼ X0 1 þ , (46:23)
m

where m is the frequency of compounding in each year. If m ¼ 4, Eq. 46.23 implies

a quarterly compounding process; if m ¼ 365, it describes a daily process; and if
m approaches infinity, it describes a continuous compounding process. Thus
Eq. 46.22 can be derived from Eq. 46.23 based upon the definition
46 Alternative Methods for Estimating Firm’s Growth Rate 1305

 
1 m
lim 1 þ ¼ e: (46:24)
m!1 m

Then the continuous analog for Eq. 46.20 can be rewritten as

  m
g mt 1 ð g Þgt
lim Xt ¼ lim X0 1 þ ¼ X0 lim 1 þ ¼ X0 egt : (46:25)
m!1 m!1 m m!1 m=g

Therefore, the growth rate estimated by continuous compound-sum method can

be expressed by

1 Xt
g ¼ ln : (46:26)
t X0

If you estimate the growth rate via Eq. 46.26, you are implicitly assuming the
dividends are growing continuously and therefore the dividend growth model. In
this case, according to Gordon and Shapiro’s (1956) model, P0 ¼ d0/(r  g).
To use all the information available to the security analysts, two regression
equations can be employed. These equations can be derived from Eqs. 46.20 and
46.22 by taking the logarithm (ln) on both sides of equation:

lnXt ¼ lnX0 þ tlnð1 þ gÞ: (46:27)

If Eq. 46.27 can be used to estimate the growth rate, then the antilog of the
regression slope estimate would equal the growth rate. For the continuous
compounding process,

lnXt ¼ lnX0 þ gt: (46:28)

Both Eqs. 46.27 and 46.28 indicate that Xn is linearly related to t; and the growth
rate can be estimated by the ordinary least square (OLS) regression. For example,
growth rates for EPS and DPS can be obtained from an OLS regression by using
 
EPSt
ln ¼ a0 þ a1 T þ e1t , (46:29)
EPS0

and
 
DPSt
ln ¼ b0 þ b1 T þ e2t , (46:30)
DPS0

where EPSt and DPSt are earnings per share and dividends per share, respectively,
in period t, and T is the time indicators (i.e., T ¼ 1, 2, . . ., n). We denote a^1 and b^1 as
1306 I.E. Brick et al.

Table 46.4 Dividend behavior of firms Pepsico and Wal-Mart in dividends per share (DPS)
Year T PEP WMT Year T PEP WMT
1981 1 3.61 1.73 1996 16 0.72 1.33
1982 2 2.4 2.5 1997 17 0.98 1.56
1983 3 3.01 1.82 1998 18 1.35 1.98
1984 4 2.19 1.4 1999 19 1.4 1.25
1985 5 4.51 1.91 2000 20 1.51 1.41
1986 6 1.75 1.16 2001 21 1.51 1.49
1987 7 2.30 1.59 2002 22 1.89 1.81
1988 8 2.90 1.11 2003 23 2.07 2.03
1989 9 3.40 1.48 2004 24 2.45 2.41
1990 10 1.37 1.9 2005 25 2.43 2.68
1991 11 1.35 1.14 2006 26 3.42 2.92
1992 12 1.61 1.4 2007 27 3.48 3.17
1993 13 1.96 1.74 2008 28 3.26 3.36
1994 14 2.22 1.02 2009 29 3.81 3.73
1995 15 2.00 1.17 2010 30 3.97 4.2

the estimated coefficients for Eqs. 46.29 and 46.30. The

 estimated growth rates for
EPS and DPS, therefore, are expða^1 Þ  1 and exp b^1  1 in terms of discrete
compounding process and a^1 and b^1 in terms of continuous compounding process.6
Table 46.4 provides dividends per share of Pepsico and Wal-Mart during the
period from 1981 to 2010. Using the data in Table 46.4 for companies Pepsico and
Wal-Mart, we can estimate the growth rates for their respective dividend streams.
Table 46.5 presents the estimated the growth rates for Pepsico and Wal-Mart by
arithmetic average method, geometric average method, compound-sum method,
and the regression method in terms of discrete and continuous compounding
processes. Graphs of the regression equations for Pepsico and Wal-Mart are
shown in Fig. 46.1.
The slope of the regression for Pepsico shows an estimated coefficient for the
intercept is 0.56. The estimated intercept for Wal-Mart is 7.04. The estimated
growth rates for Pepsico and Wal-Mart, therefore, are 0.56 % and 7.29 % in
terms of discrete compounding process. Figure 46.1 also shows the true DPS and
predicted DPS for Pepsico and Wal-Mart. We find that the regression method, to
some extent, can estimate the growth rate for Wal-Mart more precisely than for
Pepsico. Comparing to the geometric average method, the regression method yields
a similar value of the estimated growth rate for Wal-Mart, while not for Pepsico.
There are some complications to be aware of when employing the arithmetic
average, the geometric average, and regression model in estimating the growth rate.
The arithmetic average is quite sensitive to extreme values. The arithmetic average,
therefore, has an upward bias that increases directly with the variability of the data.

6
If the earnings (or dividend) process follows Eq. 46.27, we can get same results from the
non-restricted model as Eqs. 46.29 and 46.30.
46 Alternative Methods for Estimating Firm’s Growth Rate 1307

Table 46.5 Estimated dividend growth rates for Pepsico and Wal-Mart
Pepsico (%) Wal-Mart (%)
Arithmetic average 4.64 8.99
Geometric average 0.99 5.45
Compound-sum method 0.99 5.30
Regression method (continuous) 0.56 7.04
Regression method (discrete) 0.56 7.29

Pepsico
2
1.8
1.6
1.4
1.2
DPS

1
0.8
0.6 True DPS
0.4 Predicted DPS
0.2
0
0 5 10 15 20 25 30
Time

DPSt
ln = − 0.6236 + 0.1947 T + εt
DPS0 (0.0056) (0.0113)

Wal-Mart
1.4
1.2
1
0.8
DPS

0.6 True DPS

Predicted DPS
0.4
0.2
0
0 10 20 30
Time
DPSt
Fig. 46.1 Regression models ln = − 0.9900 + 0.0704 T + εt
DPS0 (0.1286) (0.0075)
for Pepsico and Wal-Mart

Consider the following situation. Dividends in years 1, 2 and 3 are $2, $4 and $2.
The arithmetic average of growth rate is 25 % but the true growth rate is 0 %. The
difference in the two average techniques will be greater when the variability of the
data is larger. Therefore, it is not surprising that we find differences in the estimated
growth rates using arithmetic average and geometric average methods for Pepsico
and Wal-Mart in Table 46.5.
1308 I.E. Brick et al.

Table 46.6 Estimated dividend growth rates for 50 randomly selected companies
Firms with positive growth Firms with negative growth
50 Firms (%) (35 firms) (%) (15 firms) (%)
Arithmetic average 4.95 7.27 0.47
Geometric average 0.93 3.00 3.88
Compound-sum 0.83 2.91 4.02
method
Regression method 0.66 2.32 3.22
(continuous)
Regression method 0.71 2.37 3.15
(discrete)

The regression method uses more available information than the geometric
average, discrete compounding and continuous compounding methods in that it
takes into account the observed growth rates between the first and last period of the
sample. A null hypothesis test can be used to determine whether the growth rate
obtained from the regression method is statistically significantly different from zero
or not. However, logarithms cannot be taken with zero or negative numbers. Under
this circumstance the arithmetic average will be a better alternative.
We further randomly select 50 companies from S&P 500 index firms, which paid
dividends during 1981–2010, to estimate their dividend growth rates by arithmetic
average method, geometric average method, compound-sum method, and the
regression method in terms of discrete and continuous compounding processes.
Table 46.6 shows averages of estimated dividend growth rates for 50 random
companies by different methods. As we discussed before, the arithmetic average
is sensitive to extreme values and has an upward bias. We, therefore, find a larger
average of the estimated dividend growth rate using the arithmetic average method.
We also find that on average, the geometric, and compound sum methods yield
relatively smaller growth rate estimates as compared to the estimates obtained using
the regression methods to estimate growth rate. However, it appears that estimates
obtained using the geometric, compound sum and regression methods are very
similar.
Finally, Gordon and Gordon (1997) suggest that one can infer the growth rate
using the dividend growth model. In particular, the practitioner can use regression
analysis to calculate the beta of the stock and use the CAPM to estimate the cost of
equity. Since

d 0 ð 1 þ gÞ
P0 ¼ (46:31)
ðr  gÞ

and the price of the stock is given by the market, the cost of equity is obtained using
the CAPM, and d0 and the current dividend is known, one can infer the growth rate
using Eq. 46.31. If the inferred growth rate is less than the practitioner’s estimate,
then the recommendation will be to buy the stock. On the other hand, if the inferred
46 Alternative Methods for Estimating Firm’s Growth Rate 1309

growth is greater than the practitioner’s estimate, the recommendation will be to

sell the stock. However, it should be noted that the explanatory power of the CAPM
to explain the relationship between stock returns and risk has been extensively
questioned in the literature. See for example, Fama and French (1992).

46.5 Conclusion

The most common valuation model is the dividend growth model. The growth rate
is found by taking the product of the retention rate and the return on equity. What is
less well understood are the basic assumptions of this model. In this paper, we
demonstrate that the model makes strong assumptions regarding the financing mix
of the firm. In addition, we discuss several methods suggested in the literature on
estimating growth rates and analyze whether these approaches are consistent with
the use of using a constant discount rate to evaluate the firm’s assets and equity. In
particular, we demonstrate that the underlying assumptions of the internal growth
rate model (whereby no external funds are used to finance growth) are incompatible
with the constant discount rate model of valuation. The literature has also suggested
estimating growth rate by using the average percentage change method, compound-
sum method, and/or regression methods. We demonstrate that the average percent-
age change is very sensitive to extreme observations. Moreover, on average, the
regression method yields similar but somewhat smaller estimates of the growth rate
compared to the compound-sum method. We also discussed the inferred method
suggested by Gordon and Gordon (1997) to estimate the growth rate. Advantages,
disadvantages, and the interrelationship among these estimation methods are also
discussed in detail. Choosing an appropriate method to estimate firm’s growth rate
can yield a more precise estimation and be helpful for the security analysis and
valuation. However, all of these methods use historical information to obtain
growth estimates. To the extent that the future may differ from the past, will
ultimately determine the efficacy of any of these methods.

References
Boudreaux, K. J., & Long, R. W. (1979). The weighted average cost of capital as a cutoff rate:
A further analysis. Financial Management, 8, 7–14.
Brick, I. E., & Weaver, D. G. (1984). A comparison of capital budgeting techniques in identifying
profitable investments. Financial Management, 13, 29–39.
Brick, I. E., & Weaver, D. G. (1997). Calculating the cost of capital of an unlevered firm for use in
project evaluation. Review of Quantitative Finance and Accounting, 9, 111–129.
Brick, I. E., Palmon, O., & Venezia, I. (2012). The risk-return (Bowman) paradox and accounting
measurements. In I. Venezia & Z. Wiener (Eds.), Bridging the GAAP: Recent advances in
finance and accounting. London: World Scientific Publishing Company.
Brick, I. E., Chen H. Y., Hsieh C. H., & Lee C. F. (2014, forthcoming). A comparison of
alternative models for estimating firm’s growth rate. Review of Quantitative Finance and
Accounting.
1310 I.E. Brick et al.

Chambers, D. R., Harris, R. S., & Pringle, J. J. (1982). Treatment of financing mix in analyzing
investment opportunities. Financial Management, 11, 24–41.
Chen, H. Y., Gupta, M. C., Lee, A. C., & Lee, C. F. (2013). Sustainable growth rate, optimal
growth rate, and optimal payout ratio: A joint optimization approach. Journal of Banking and
Finance, 37, 1205–1222.
Fama, E. F., & French, K. R. (1997). The cross-section of expected stock returns. The Journal of
Finance, 47, 427–465.
Gordon, J., & Gordon, M. (1997). The finite horizon expected return model. The Financial
Analysts Journal, 53, 52–61.
Gordon, M., & Shaprio, E. (1956). Capital equipment analysis: The required rate of profit.
Management Science, 3, 102–110.
Higgins, R. C. (1977). How much growth can a firm afford? Financial Management, 6, 7–16.
Higgins, R. C. (1981). Sustainable growth under inflation. Financial Management, 10, 36–40.
Higgins, R. C. (2008). Analysis for financial management (9th ed.). New York: McGraw-Hill.
Lee, C. F., & Lee, A. C. (2013). Encyclopedia of finance (2nd ed.). New York: Springer.
Lee, C. F., Lee, J. C., & Lee, A. C. (2000). Statistics for business and financial economics
(2nd ed.). Singapore: World Scientific Publishing Company.
Lee, A. C., Lee, J. C., & Lee, C. F. (2009). Financial analysis, planning and forecasting: Theory
and application (2nd ed.). Singapore: World Scientific Publishing Company.
Lee, C. F., Finnerty, J. E., Lee, A. C., Lee, J., & Wort, D. H. (2012a). Security analysis, portfolio
management, and financial derivatives (2nd ed.). New York: Springer.
Lee, C. F., Gupta, M. C., Chen, H. Y., & Lee, A. C. (2012b). Optimal payout ratio under
uncertainty and the flexibility hypothesis: Theory and empirical evidence. Journal of Corpo-
rate Finance, 17, 483–501.
Ross, S. A., Westerfield, R. W., & Jaffe, J. (2010). Corporate finance (9th ed.). New York:
McGraw-Hill/Irwin.
Econometric Measures of Liquidity
47
Jieun Lee

Contents
47.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1312
47.2 Low-Frequency Liquidity Proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1313
47.2.1 The Roll Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1313
47.2.2 Effective Tick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1314
47.2.3 Amihud (2002) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1315
47.2.4 Lesmond, Ogden, and Trzcinka (LOT 1999) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1316
47.3 High-Frequency Liquidity Proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1318
47.3.1 Percent Quoted Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1318
47.3.2 Percent Effective Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1318
47.4 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1319
47.4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1319
47.4.2 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1319
47.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1321
Appendix 1: Solution to LOT (1990) Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1321
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1322

Abstract
A security is liquid to the extent that an investor can trade significant quantities of
the security quickly, at or near the current market price, and bearing low transac-
tion costs. As such, liquidity is a multidimensional concept. In this chapter,
I review several widely used econometrics or statistics-based measures that
researchers have developed to capture one or more dimensions of a security’s
liquidity (i.e., limited dependent variable model (Lesmond, D. A. et al. Review of
Financial Studies, 12(5), 1113–1141, 1999) and autocovariance of price changes
(Roll, R., Journal of Finance, 39, 1127–1139, 1984). These alternative proxies
have been designed to be estimated using either low-frequency or high-frequency

J. Lee
Economic Research Institute, Bank of Korea, Seoul, South Korea
e-mail: jelee@bok.or.kr

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1311
DOI 10.1007/978-1-4614-7750-1_99,
# Springer Science+Business Media New York 2015
1312 J. Lee

data, so I discuss four liquidity proxies that are estimated using low-frequency data
and two proxies that require high-frequency data. Low-frequency measures permit
the study of liquidity over relatively long time horizons; however, they do not
reflect actual trading processes. To overcome this limitation, high-frequency
liquidity proxies are often used as benchmarks to determine the best
low-frequency proxy. In this chapter, I find that estimates from the effective tick
measure perform best among the four low-frequency measures tested.

Keywords
Liquidity • Transaction costs • Bid-ask spread • Price impact • Percent effective
spread • Market model • Limited dependent variable model • Tobin’s model •
Log likelihood function • Autocovariance • Correlation analysis

47.1 Introduction

A security is liquid to the extent that an investor can trade significant quantities of
the security quickly, at or near the current market price, and bearing low transaction
costs. A security’s liquidity is an important characteristic variable, relevant in asset
pricing studies, studies of market efficiency, and even corporate finance. In the asset
pricing literature, researchers have considered whether liquidity is a priced risk
factor (e.g., Amihud and Mendelson 1986; Brennan and Subrahmanyam 1996;
Amihud 2002; Pastor and Stambaugh 2003). In corporate finance, researchers
have found that liquidity is related to capital structure, mergers and acquisitions,
and corporate governance (e.g., Lipson 2003; Lipson and Mortal 2007, 2009;
Bharath 2009; Chung et al. 2010).
In these and many other studies, researchers have chosen from a variety of
liquidity measures that have been developed. In turn, the variety of available
liquidity measures reflects the multidimensional aspect of liquidity. Note that the
definition of liquidity given above features four dimensions of liquidity: trading
quantity, trading speed, price impact, and trading cost. Some extant measures focus
on a single dimension of liquidity, while others encompass several dimensions.
For instance, the bid-ask spread measure in Amihud and Mendelson (1986), the
estimator of the effective spread in Roll (1984), and the effective tick estimator in
Goyenko et al. (2009) relate to the trading cost dimension. The turnover measure of
Datar et al. (1998) captures the trading quantity dimension. The measures in
Amihud (2002) and Pastor and Stambaugh (2003) are relevant to price impact.
The number of zero trading volume days in Liu (2006) emphasizes trading speed.
Finally, and different from the others, the measure in Lesmond et al. (1999) encom-
passes several dimensions of liquidity.
Among the available measures, this chapter focuses on six liquidity proxies,
including four that are commonly estimated using low-frequency data (i.e., daily
closing prices) and two that are commonly estimated using high-frequency
data (i.e., intraday trades and quotes). The low-frequency measures are in
Roll (1984), Goyenko et al. (2009), Lesmond et al. (1999), and Amihud (2002).
47 Econometric Measures of Liquidity 1313

The high-frequency measures are the percent quoted spread and the percent
effective spread. The low-frequency proxies are advantageous because they are
more amenable to the study of liquidity over relatively long time horizons and
across countries. However, they are limited because they do not directly reflect
actual trading processes, while the high-frequency measures do. Thus, high-
frequency liquidity proxies are often used as benchmarks to determine the best
low-frequency proxy. This is not a universal criterion, however, because each
measure captures a different dimension of liquidity and may lead to different
results in specific cross-sectional or time-series applications.
The remainder of this chapter is organized as follows. In Sects. 47.2 and 47.3,
I introduce and briefly discuss each of the low-frequency and high-frequency
liquidity measures, respectively. Section 47.4 provides an empirical analysis of
these liquidity measures, including the aforementioned test of the best
low-frequency measure. Section 47.5 concludes.

47.2 Low-Frequency Liquidity Proxies

Below I describe four widely used measures of liquidity: the Roll (1984) measure;
effective tick; the Amihud (2002) measure; and the Lesmond et al. (1999) measure.

47.2.1 The Roll Measure

Roll (1984) develops a measure of the effective bid-ask spread. He assumes that the
true value of a stock follows a random walk and that Pt, the observed closing price
on day t, is equal to the stock’s true value plus or minus half of the effective spread.
He also assumes that a security trades at either the bid price or the ask price, with
equal frequency. This relationship can be expressed as follows:

s
Pt ¼ Pt þ Qt
2


þ1 with probability 1=2 ðbuyer initiatedÞ
Qt  IID
1 with probability 1=2 ðseller initiatedÞ

where Qt is an order-type indicator variable, indicating whether the transaction at

time t is at the ask (buyer-initiated) or at the bid (seller-initiated) price. His
assumption that Pt is the fundamental value of the security implies that E
[Qt] ¼ 0; hence, Pr(Qt ¼ 1) ¼ Pr(Qt ¼ 1) ¼ 1/2. Also, there are no changes in
the fundamental value of the security (i.e., over a short horizon).
It follows that the process for price changes DPt is
 s  s
DPt ¼ DPt þ Qt  Qðt1Þ ¼ Qt  Qðt1Þ :
2 2
1314 J. Lee

Under the assumption that Qt is IID, the variance and covariance of DPt can be
easily calculated:
s2
var½DPt  ¼ :
2
s2
cov½DPt ; DPt1  ¼  :
hs 4 i
s
cov½DPt ; DPt1  ¼ cov ðQt  Qt1 Þ, ðQt1  Qt2 Þ
2 2
s2
¼ ½covðQt  Qt1 Þ, covðQt1  Qt2 Þ
4
s2
¼ ½covðQt ; Qt1 Þ  covðQt1 ; Qt1 Þ
4
þcovðQt1 ; Qt2 Þ  covðQt ; Qt2 Þ


s2 s2 1 2 1 2 s2
¼ ½varðQt1 Þ ¼  ð1  0Þ þ ð1  0Þ ¼ 
4 4 2 2 4

cov½DPt ; DPk1  ¼ 0, k > 1:

Solving for S yields Roll’s effective spread estimator:

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
S¼2 Cov DPt; ; DPt1 :

Roll’s measure is simple and intuitive: If P* is fixed so that prices take only two
values, bid or ask, and if the current price is the bid, then the change between
current price and previous price must be either 0 or –s and the price change between
current price and next price must be either 0 or s. Analogous possible price changes
apply when the current price is the ask.
The Roll measure S is generally estimated using daily data on price changes.
Roll (1984) and others have found that for some individual stocks, the
autocovariance that defines S is positive, rather than negative, so that S is undefined.
In this case, researchers generally choose one of three solutions: (1) treat the
observation as missing, (2) set the Roll spread estimate to zero, or (3) multiply
the covariance by negative one, calculate S, and multiply this estimate by negative
one to produce a negative spread estimate. In my empirical analysis to follow, I find
that results are insensitive to the alternative solutions, so I only report results of
setting S to zero when the observed autocovariance is positive.

47.2.2 Effective Tick

Goyenko et al. (2009) and Holden (2009) develop an effective tick measure that is
based on price clustering and changes in tick size. Below I describe the effective
tick measure in Goyenko et al. (2009), which is elegant in its simplicity.
47 Econometric Measures of Liquidity 1315

Consider four possible bid-ask spreads for a stock: $1/8, $1/4, $1/2, and $1. If the
spread is $1/4, the authors assume that bid and ask prices are associated with only
even quarters. Thus, if an odd-eighth transaction price shows up, it is instead
inferred that the spread is $1/8. The number of quotes that occur at $1/8 spread is
given by N1. The number of quotes at odd-quarter fractions ($1/8 and $1/4) is N2.
The number of quotes at odd-half ($1/8, $1/4, and $1/2) is N3. Finally, the number
of whole dollar quotes is given by N4. The following is the proportion of each price
fraction observed during the day:

Ni
Fi ¼ XI for i ¼ 1, . . . , I:
i¼1
Ni

Next, suppose that the unconstrained probability of the effective ith estimated
spread is

2Fi i ¼ 1
U ¼ 2Fi  Fi1 i ¼ 2, . . . , i
Fi  Fi1 i ¼ I:

The effective tick is a simply probability-weighted average of effective spread

size divided by average price in a given time interval:
XI
a  Si
Effective Tickit ¼ i¼1
,
P

where the probability a is constrained to be nonnegative and to be no more than

1 minus the probability of a finer spread, S is the spread, and P is the average price
in the time interval.
To obtain estimates of effective tick in this chapter, I must deal with changes in
minimum tick size that were instituted over time in the US equity markets. For
NYSE, AMEX, and NASDAQ stocks from 1/93 to 5/97, I used a fractional grid
accounting for price increments as small as $1/8. For NYSE and AMEX
(NASDAQ) stocks from 6/97 to 1/01 (6/97 to 3/01), I used a minimum tick size
increment of $1/16. Thereafter, I used a decimal grid for all stocks.

47.2.3 Amihud (2002)

The measures in Amihud (2002) and Pastor and Stambaugh (2003) both purport to
capture the price impact dimension of liquidity. Goyenko et al. (2009) show that the
Amihud (2002) measure performs well in measuring price impact while the Pastor
and Stambaugh (2003) measure is dominated by other measures. Pastor and
Stambaugh (2003, p. 679) also caution against their measure as a liquidity measure
for individual stocks, reporting large sampling errors in individual estimates.
1316 J. Lee

Referring also to Hasbrouck (2009), I do not discuss the Pastor and Stambaugh
(2003) measure. The Amihud (2002) measure is a representative proxy for price
impact, i.e., the daily price response associated with one dollar of trading volume:

jRetit j
Amiit ¼ ,
Volit

where Retit is the stock i’s return on day t and Volit is the stock i’s dollar volume on
day t. The average is calculated over all positive-volume days, since the ratio is
undefined for zero volume days.

47.2.4 Lesmond, Ogden, and Trzcinka (LOT 1999)

The Lesmond et al. (LOT 1999) liquidity measure is based on the idea that an
informed trader observing a mispriced stock will execute a trade only if the
difference between the current market price and the true price exceeds the
trader’s transaction costs; otherwise, no trade will occur. Therefore, they
argue that a stock with high transaction costs will have less frequent price
movements and more zero returns than a stock with low transaction costs.
Based on this relationship, they develop a measure of the marginal trader’s
effective transaction costs for an individual stock. Their measure utilizes the
limited dependent variable regression model of Tobin (1958) and Rosett (1959)
applied to the “market model.”

47.2.4.1 Market Model

The basic market model is a regression of the return, Rit on security i and period t, on
the contemporaneous market return, Rmt:

Rit ¼ ai þ bi Rmt þ eit (47.1)

The market model implies that a security’s return reflects the effect of new
information on the value of the stock, which can be divided into two components:
contemporaneous market-wide information (biRmt) and firm-specific information eit.
In an ideal market without frictions such as transaction costs, new information
will be immediately reflected into the security’s price, so Rit is the true return on
security i.

47.2.4.2 Relationship Between Observed and True Returns

In the presence of transaction costs, investors will trade only when the marginal
profits exceed the marginal transaction costs. In this context, transaction costs
would include various dimensions such as bid-ask spread, commissions, and price
impact, as well as taxes or short-selling costs, because investors make trading
decisions after considering overall transaction costs. Transaction costs inhibit
informative trades and therefore drive a wedge between observed returns and true
47 Econometric Measures of Liquidity 1317

Observed Return (Rit)

Region3

True Return (R*it )

Region1
Region2

Fig. 47.1 This figure illustrates the relationship between the observed return on a stock in the
presence of transaction costs that inhibit trading, Rit, and its true return in the absence of transaction
costs, R*it, where the latter reflects the true effects of new market-wide or firm-specific information.
The relationship can be divided into three regions: (1) Region 1, where the value of new information
is negative and exceeds transaction costs; (2) Region 2, where the transaction costs exceed the value
of new information regardless of the direction of the value of information; and (3) Region 3, where
the value of new information is positive and exceeds transaction costs

returns. In the presence of transaction costs, Lesmond et al. (1999) propose the
following relationship between observed and true returns:

Rit ¼ Rit  mSi , (47.2)

where mSi is the spread adjustment for security i, Rit is the observed return, and Rit is
true return.
Specifically, the relationship is as follows:

Rit ¼ Rit  a1i if Rit < a1i

Rit ¼ 0 if a1i < Rit < a2i (47.3)
Rit ¼ Rit  a2i if Rit > a2i

where a1i < 0 and a2i > 0. a1i is the transaction costs for the marginal investor when
information has a negative shock (selling), and a2i is the transaction costs for the
marginal investor when information has a positive shock (buying). Consequently,
the difference between a1i and a2i is a measure of round-trip transaction costs.
If the true return exceeds transaction costs, the marginal investor will continu-
ously trade, and the market price will respond until, for the next trade, marginal
profit is equal to marginal transaction costs. If the transaction costs are greater than
true returns, then the marginal investors will not trade, price will not move, and
consequently the zero returns will occur. Therefore, in this model the frequency of
zero returns is a simple alternative measure of transaction costs. The relationship
between observed returns and true returns is illustrated in Fig. 47.1.
1318 J. Lee

47.3 High-Frequency Liquidity Proxies

Next I describe two well-known spread proxies that can be estimated using high-
frequency data. These are the percent quoted spread and percent effective spread.

47.3.1 Percent Quoted Spread

The ask (bid) quotation is the price at which shares can be purchased (sold) with
immediacy. The difference, known as the percent quoted spread, is the cost of
a round-trip transaction and is generally expressed as a proportion of the average of
the bid and ask prices:

Askit  Bidit
Percent quoted spreadit ¼ :
Mit

In the empirical analysis in the next section, I estimate percent quoted spreads
using high-frequency data. Following convention, for each stock and trading day,
I find the highest bid and lowest ask prices over all venues at every point during the
day, denoting these “inside” ask and bid prices as Askit and Bidit, respectively. Mit
is then the average of, or midpoint between, Askit and Bidit. I then calculate the
average percent quoted spread for a stock and day as the time-weighted average of
all spreads observed for that stock during the day. Finally, percent quoted spread
for each stock is calculated by averaging the daily estimates across all trading days
within a given month.

47.3.2 Percent Effective Spread

Some trades occur within the range of inside bid and ask quotes, as when simulta-
neous buy and sell market orders are simply crossed. Thus, the inside bid-ask spread
may overestimate the realized amount of this component of transaction costs.
Hasbrouck’s (2009) measure of percent effective spread attempts to adjust for
this bias. For a given stock, percent effective spread is computed for all trades
relative to the prevailing quote midpoint:
 
Pit  Mit
Percent effective spreadit ¼ 2Dit ,
Mit

where, for stock i, Dit is the buy-sell indicator variable which takes a value of 1 (–1)
for buyer-initiated (seller-initiated) trades, Pit is the transaction price, and Mit is the
midpoint of the most recently posted bid and ask quotes. The average percent
effective spread for each day is a trade-weighted average across all trades during
the day. The monthly percent effective spread for each security is calculated by
averaging across all trading days within a given month.
47 Econometric Measures of Liquidity 1319

47.4 Empirical Analysis

47.4.1 Data

I estimate liquidity measures for NYSE, AMEX, and NASDAQ common stocks
over the years 1993–2008. I estimate the low-frequency measures using daily data
from the Center for Research in Security Prices (CRSP) database. I estimate the
high-frequency measures using the New York Stock Exchange Trades and Auto-
mated Quotes (TAQ) database. TAQ data is available only since 1993, which is
therefore the binding constraint in terms of staring year. In order to be included in
the sample, a stock must have at least 60 days of past return data. I discard
certificates, American Depositary Receipts (ADRs), shares of beneficial interest,
units, companies incorporated outside United States, American Trust components
closed-end funds, preferred stocks, and Real Estate Investment Trusts (REITs).
Regarding estimating the high-frequency measures, I determine the highest bid
and lowest ask across all quoting venues at every point during the day (NBBO
quotes) and then follow filters referring to Huang and Stoll (1997) and Brownless
and Gallo (2006). To reduce errors and outliers, I remove (1) quotes if either the bid
or ask price is negative; (2) quotes if either the bid or ask size is negative; (3) quotes
if bid-ask spread is greater than $5 or negative; (4) the quotes if transaction price is
negative; (5) quotes before-the-open and after-the-close trades and quotes;
(6) quotes if the bid, ask, or trade price differ by more than 20 % from the previous
quote or trade price; (7) quotes originating in market other than the primary
exchange because regional quotes tend to closely follow the quotes posted by the
primary exchange; and (8) %effective spread/%quoted spread>4.0.

47.4.2 Empirical Results

Table 47.1 reports correlations among the various liquidity estimates. In this table,
observations are pooled across all stocks and all months. All correlations are
reliably positive and substantial in magnitude, ranging from 0.382 to 0.971. The
two high-frequency measures, percent effective spread and percent quoted spread,
are very highly correlated (0.971). Using percent effective spread as our high-
frequency “benchmark,” its correlations with the low-frequency measures are 0.742
(Roll), 0.757 (effective tick), 0.621 (Amihud), and 0.586 (LOT). Based on the
aforementioned criterion, these results indicate that the effective tick and Roll
measures are the “best” low-frequency measures, as they have the highest correla-
tion with percent effective spread.
Table 47.2 presents the time-series means of monthly correlations of percent
effective spread with each of the low-frequency measures for the full sample period
as well as subperiods 1993–2000 (pre-decimalization) and 2001–2008 (post-
decimalization). For three of the four low frequencies measured, the correlation
with percent effective spread is higher in the first subperiod than the second
subperiod, which may reflect differential effects of decimalization on the various
1320 J. Lee

Table 47.1 Pooled correlations

ES QS Roll Eff. tick Ami LOT
ES 1
QS 0.971 1
Roll 0.742 0.744 1
Eff. tick 0.757 0.760 0.638 1
Ami 0.621 0.631 0.512 0.472 1
LOT 0.586 0.562 0.710 0.541 0.382 1
This table presents correlations among the liquidity estimates based on the pooled sample of
monthly time-series and cross-sectional observations. Observation can be dropped if there are
fewer than 60 days observations for the firm or if liquidity estimates are missing

Table 47.2 Average cross-sectional correlations with percent effective spread, monthly
estimates
Roll Eff. tick Ami LOT
1993–2008 0.662 0.685 0.684 0.561
1993–2000 0.748 0.754 0.662 0.670
2001–2008 0.576 0.615 0.706 0.452
For each month, I estimate the cross-sectional correlation between the liquidity proxies from the
low-frequency data and percent effective spread from TAQ. This table presents the average cross-
sectional correlations across all months. A stock is excluded only if it trades for less than 60 days
prior to an observation or if liquidity estimates are missing

Table 47.3 Summary statistics for stock-by-stock time-series correlations

N Roll Eff. tick Ami LOT
Full 13,374 0.319 0.578 0.603 0.308
NYSE 3,137 0.180 0.694 0.690 0.201
AMEX 1,597 0.231 0.517 0.643 0.310
NASDAQ 9,870 0.353 0.534 0.567 0.335
For each stock, I estimate the time-series correlation between the estimated liquidity measure and
percent effective spread from TAQ. The table presents the average time-series correlation across
all stocks. Observations are dropped if there are fewer than 60 days observations for the firm or if
a spread estimate is missing

dimensions of liquidity. For the full period as well as the first subperiod, effective
tick has the correlation with percent effective spread, while for the second period
the Amihud measure has the highest correlation with percent effective spread.
Table 47.3 shows stock-by-stock time-series correlations between the high-
frequency measure percent effective spread and each of the low-frequency
measures, using the full-period data but also breaking the observations down by
the exchange on which a stock trades. For the full sample as well as every exchange,
the Amihud and effective tick estimates have relatively high correlations with
percent effective spread, while the correlations are relatively low for the Roll and
LOT measures.
47 Econometric Measures of Liquidity 1321

Overall, based on the suggested criteria of correlation with a high-frequency

measure, the effective tick measure is best among the low-frequency measures
tested, as it exhibits higher correlations with percent effective spread based on time-
series, cross-sectional, and pooled tests. Again, though, I caution that, since each
liquidity measure captures only part of the multidimensional nature of liquidity, it is
difficult to judge which measure is best.

47.5 Conclusions

This chapter discusses several popular measures of liquidity that are based on
econometric approaches and compares them via correlation analysis. Among the
four low-frequency liquidity proxies, I find that the effective tick measure is
generally more highly correlated with the high-frequency measure (i.e., percent
effective spread). Thus, by this criterion the effective tick measure is the
“best” low-frequency measure of liquidity. However, since each liquidity measure
captures only part of the multidimensional nature of liquidity, it is difficult to judge
which measure is best. Consequently, from among the available measures of
liquidity, a researcher should choose the measure that is consistent with their
research purpose or perhaps consider employing several of them.

Appendix 1: Solution to LOT (1990) Model

To estimate transaction costs based on their model in Eq. 47.3, Lesmond

et al. (1999) introduce the limited dependent variable regression model of Tobin
(1958) and Rosett (1959). Tobin’s model specifies that data are available for the
explanatory variable, x, for all the observation while data are only partly observable
for the dependent variable, y, and for the other unobservable region, the information
is given whether or not data are above a certain threshold.
Considering this aspect of Tobin’s model, the limited dependent variable model is
an appropriate econometric method for the LOT model because a nonzero observed
return occurs only when marginal profit exceeds marginal transaction costs.
Assuming that market model is correct in the presence of transaction costs,
Lesmond et al. (1999) estimate transaction costs on the basis of Eqs. 47.1 and 47.3.
The equation system is

Rit ¼ bit Rmt þ eit ,

where

Rit ¼ Rit  a1i if Rit < a1i

Rit ¼ 0 if a1i < Rit < a2i (47.4)
Rit ¼ Rit  a2i if Rit > a2i
1322 J. Lee

The solution to this limited dependent regression variable model requires

a likelihood function to be maximized with respect to a1i, a2i, bi, and si.

Y Rit þ a1i  b Rmt 

Lða1i , a2i , bi , si =Rit , Rmt Þ ¼ Ø i

1
s i
Y
Rit þ a2i  b Rmt  
Rit þ a1i  bi Rmt

F i
F
2
si si
Y Rit þ a2i  b Rmt 
Ø i
,
3
s i

(47.5)

where Ø refers to the standard normal density function and F refers to the
cumulative normal distribution. The product is over the Region 1, 2, and 3 of
observations for which R*it < a1i, a1i < R*it < a2i, and R*it > a2i, respectively. The
log likelihood function is
2 3
6 1 7 1
log L ¼ S1 log4 5 S1 ðRit þ a1i  bi Rmt Þ2 þ S2 log½F2  F1 
1 2si 2
ð2psi Þ
2

2 2 3

6 1 7
þ S3 log6 7  1 S ðR þ a2i  bi Rmt Þ2 :
4 1 5 2si 2 3 it
ð2psi 2 Þ2
(47.6)

Given Eq. 47.6, a1i, a2i, bi, and si can be estimated. The difference between a2i
and a1i is the proxy of a round-trip transaction cost in the LOT model.

References
Amihud, Y. (2002). Illiquidity and stock returns: Cross-section and time-series effects. Journal of
Financial Markets, 5, 31–56.
Amihud, Y., & Mendelson, H. (1986). Asset pricing and the bid-ask spread. Journal of Financial
Economics, 17(2), 223–249.
Bharath, S. T., Pasquariello, P., et al. (2009). Does asymmetric information drive capital structure
decisions? Review of Financial Studies, 22(8), 3211–3243.
Brennan, M. J., & Subrahmanyam, A. (1996). Market microstructure and asset pricing: on the
compensation for illiquidity in stock returns. Journal of Financial Economics, 41, 441–464.
Brennan, M., Chordia, T., et al. (1998). Alternative factor specifications, security characteristics, and
the cross-section of expected stock returns. Journal of Financial Economics, 49(3), 345–373.
Brownlees, C. T., & Gallo, G. M. (2006). Financial econometric analysis at ultra-high frequency:
Data handling concerns. Computational Statistics & Data Analysis, 51, 2232–2245.
47 Econometric Measures of Liquidity 1323

Chung, K. H., Elder, J., & Kim, J. (2010). Corporate governance and liquidity. Journal of
Financial and Quantitative Analysis, 45(02), 265–291.
Datar, V. T., Naik, N. Y., & Radcliffe, R. (1998). Liquidity and stock returns: An alternative test.
Journal of Financial Markets, 1, 203–219.
Goyenko, R. Y., Holden, C. W., et al. (2009). Do liquidity measures measure liquidity? Journal of
Financial Economics, 92(2), 153–181.
Hasbrouck, J. (2009). Trading costs and returns for US equities: Estimating effective costs from
daily data. Journal of Finance, 64(3), 1445–1477.
Holden, C. W. (2009). New low-frequency spread measures. Journal of Financial Markets, 12(4),
778–813.
Huang, R. D., & Stoll, H. R. (1997). The components of the bid-ask spread: A general approach.
Review of Financial Studies, 10(4), 995–1034.
Lesmond, D. A., Ogden, J. P., et al. (1999). A new estimate of transaction costs. Review of
Financial Studies, 12(5), 1113–1141.
Lipson, M. L. (2003). Market microstructure and corporate finance. Journal of Corporate Finance,
9(4), 377–384.
Lipson, M. L., & Mortal, S. (2007). Liquidity and firm characteristics: Evidence from mergers and
acquisitions. Journal of Financial Markets, 10(4), 342–361.
Lipson, M. L., & Mortal, S. (2009). Liquidity and capital structure. Journal of Financial Markets,
12(4), 611–644.
Liu, W. (2006). A liquidity-augmented capital asset pricing model. Journal of Financial Econom-
ics, 82, 631–671.
Maddala, G. S. (1986). Limited-dependent and qualitative variables in econometrics. Cambridge:
Cambridge University Press.
Pástor, L., & Stambaugh, R. (2003). Liquidity risk and expected stock returns. Journal of Political
Economy, 111, 642–685.
Roll, R. (1984). A simple implicit measure of the effective bid-ask spread in an efficient market.
Journal of Finance, 39, 1127–1139.
Rosett, R. N. (1959). A statistical model of friction in economics. Econometrica: Journal of the
Econometric Society, 27, 263–267.
Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica:
Journal of the Econometric Society, 26, 24–36.
A Quasi-Maximum Likelihood
Estimation Strategy for Value-at-Risk 48
Forecasting: Application to Equity
Index Futures Markets

Oscar Carchano, Young Shin (Aaron) Kim, Edward W. Sun,

Svetlozar T. Rachev, and Frank J. Fabozzi

Contents
48.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1326
48.2 ARMA-GARCH Model with Normal and Tempered Stable Innovations . . . . . . . . . . . . 1328
48.3 VaR for the ARMA-GARCH Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1329
48.3.1 VaR and Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1329
48.4 Introduction of Trading Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1332
48.4.1 Different Variants of Trading Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1332
48.4.2 Lagged Relative Change of Trading Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1334
48.4.3 Lagged Trading Volume or Forecasting Contemporaneous
Trading Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1334
48.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1337
Appendix: VaR on the CTS Random Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1337
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1339

O. Carchano
Department of Financial Economics, University of Valencia, Valencia, Spain
e-mail: oscar.carchano@uv.es
Y.S.A. Kim
College of Business, Stony Brook University, Stony Brook, NY, USA
e-mail: aaron.kim@stonybrook.edu
E.W. Sun
KEDGE Business School and BEM Management School, Bordeaux, France
e-mail: edward.sun@kedgebs.com
S.T. Rachev
Department of Applied Mathematics and Statistics, College of Business, Stony Brook University,
SUNY, Stony Brook, NY, USA
FinAnalytica, Inc, New York, NY, USA
e-mail: svetlozar.rachev@stonybrook.edu
F.J. Fabozzi (*)
EDHEC Business School, EDHEC Risk Institute, Nice, France
e-mail: frank.fabozzi@edhec.edu; fabozzi321@aol.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1325
DOI 10.1007/978-1-4614-7750-1_48,
# Springer Science+Business Media New York 2015
1326 O. Carchano et al.

Abstract
We present the first empirical evidence for the validity of the ARMA-GARCH
model with tempered stable innovations to estimate 1-day-ahead value at risk
in futures markets for the S&P 500, DAX, and Nikkei. We also provide
empirical support that GARCH models based on normal innovations
appear not to be as well suited as infinitely divisible models for predicting
financial crashes. The results are compared with the predictions based on
data in the cash market. We also provide the first empirical evidence on
how adding trading volume to the GARCH model improves its forecasting
ability.
In our empirical analysis, we forecast 1 % value at risk in both spot and
futures markets using normal and tempered stable GARCH models following
a quasi-maximum likelihood estimation strategy. In order to determine
the accuracy of forecasting for each specific model, backtesting using Kupiec’s
proportion of failures test is applied. For each market, the model with
a lower number of violations is preferred. Our empirical result indicates the
usefulness of classical tempered stable distributions for market risk management
and asset pricing.

Keywords
Infinitely divisible models • Tempered stable distribution • GARCH models •
Value at risk • Kupiec’s proportion of failures test • Quasi-maximum likelihood
estimation strategy

48.1 Introduction

Predicting future financial market volatility is crucial for risk management of

financial institutions. The empirical evidence suggests that a suitable market
risk model must be capable of handling the idiosyncratic features of volatility,
that is, daily returns time variant amplitude and volatility clustering. There is a
well-developed literature in financial econometrics that demonstrates how
autoregressive conditional heteroskedastic (ARCH) and generalized ARCH
(GARCH) models – developed by Engle (1982) and Bollerslev (1986),
respectively – can be employed to explain the clustering effect of volatility1.
Moreover, the selected model should consider the stylized fact that asset return
distributions are not normally distributed, but instead have been shown to exhibit
patterns of leptokurtosis and skewness.

1
For a description of ARCH and GARCH modeling, see Chap. 8 in Rachev et al. (2007). The
chapter of the same reference describes ARCH and GARCH modeling with infinite variance
innovations. Engle et al. (2008) provide the basics of ARCH and GARCH modeling with
applications to finance.
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1327

Taking a different tact than the ARCH/GARCH with normal innovations

approach for dealing with the idiosyncratic features of volatility, Kim
et al. (2010) formulate an alternative model based on subclasses of the infinitely
divisible (ID) distributions. More specifically, for the S&P 500 return, they empir-
ically investigate five subclasses of the ID distribution, comparing their results to
that obtained using GARCH models based on innovations that are assumed to
follow a normal distribution. They conclude that, due to their failure to focus on
the distribution in the tails, GARCH models based on the normal innovations may
not be as well suited as ID models for predicting financial crashes.
Because of its popularity, most empirical studies have examined value at risk
(VaR) as a risk measure. These studies have focused on stock indices. For example,
Kim et al. (2011), and Asai and McAleer (2009) examine the S&P 500, DAX
30, and Nikkei 225 stock indices, respectively. A few researchers have studied this
risk measure for stock index futures contracts: Huang and Lin (2004) (Taiwan stock
index futures) and Tang and Shieh (2006) (S&P 500, Nasdaq 100, and Dow Jones
stock index futures). As far as we know, there are no empirical studies comparing
VaR spot and futures indices. For this reason, we compare the predictive perfor-
mance of 1-day-ahead VaR forecasts in these two markets.
We then introduce trading volume into the model, particularly, within the
GARCH framework. There are several studies that relate trading volume and
market volatility for equities and equity futures markets. Studies by Epps and
Epps (1976), Smirlock and Starks (1985), and Schwert (1989) document
a positive relation between volume and market volatility. Evidence that supports
the same relation for futures is provided by Clark (1973), Tauchen and Pitts (1983),
Garcia et al. (1986), Ragunathan and Peker (1997), and Gwilym et al. (1999).
Collectively, these studies clearly support the theoretical prediction of a positive
and contemporaneous relationship between trading volume and volatility. This
result is a common empirical finding for most financial assets, as Karpoff (1987)
showed when he summarized the results of several studies on the positive relation
between price changes and trading volume for commodity futures, currency futures,
common stocks, and stock indices.
Foster (1995) concluded that not only is trading volume important in determining
the rate of information (i.e., any news that affects the market), but also lagged volume
plays a role. Although contemporary trading volume is positively related to volatility,
lagged trading volume presents a negative relationship. Empirically, investigating
daily data for several indices such as the S&P 500 futures contract, Wang and Yau
(2000) observe that there is indeed a negative link between lagged trading volume
and intraday price volatility. This means that an increase in trading volume today
(as a measure of liquidity) will imply a reduction in price volatility tomorrow. In their
study of five currency futures contracts, Fung and Patterson (1999) do in fact find
a negative relationship between return volatility and past trading volume. In their
view, the reversal behavior of volatility with trading volume is generally consistent
with the overreaction hypothesis (see Conrad et al. 1994) and supports the sequential
information hypothesis (see Copeland 1976), which explains the relationship
between return volatility and trading volume.
1328 O. Carchano et al.

Despite the considerable amount of research in this area, there are no studies that
use trading volume in an effort to improve the capability of models to forecast
1-day-ahead VaR. Typically, in a VaR context, trading volume is only employed as
a proxy for “liquidity risk” – the risk associated with trying to close out a position.
In this paper, in contrast to prior studies, we analyze the impact of introducing
trading volume on the ability to enhance performance in forecasting VaR 1 day
ahead. We empirically test whether the introduction of trading volume will
reduce the number of violations (i.e., the number of times when the observed loss
exceeds the estimated one) in the spot and futures equity markets in the USA,
Germany, and Japan.
The remainder of this paper is organized as follows. ARMA-GARCH models
with normal and tempered stable innovations are reviewed in Sect. 48.2. In
Sect. 48.3, we discuss parameter estimation of the ARMA-GARCH models and
forecasting daily return distributions. VaR values and backtesting of the ARMA-
GARCH models are also reported in Sect. 48.2, along with a comparison of the
results for (1) the spot and futures markets and (2) the normal and tempered stable
innovations. Trading volume is introduced into the ARMA-GARCH model with
tempered stable innovations in Sect. 48.4. VaR and backtesting of the ARMA-
GARCH with different variants of trading volume are presented and compared to
the results for models with and without trading volume. We summarize our
principal findings in Sect. 48.5.

48.2 ARMA-GARCH Model with Normal and Tempered

Stable Innovations

In this section, we provide a review of the ARMA-GARCH models with normal

and tempered stable innovations. For a more detailed discussion, see Kim
et al. (2011).
Let (St)t
0 be the asset price process and (yt)t
0 be the return process of (St)t
0
defined by yt ¼ log StS1
t
. The ARMA(1,1)-GARCH(1,1) model is

yt ¼ ayt1 þ bst1 et1 þ st et þ ct
: (48.1)
s2t ¼ a0 þ a1 s2 t1 e2 t1 þ b1 s2 t1

where e0 ¼ 0 and a sequence (et)tcN ¼ 0 of independent and identically distributed

(iid) real random variables. The innovation et is assumed to follow the standard
normal distribution. This ARMA(1,1)-GARCH(1,1) model is referred to as the
“normal-ARMA-GARCH model.”
If the ets are assumed to be tempered stable innovations, then we obtain
a new ARMA(1,1)-GARCH(1,1) model. In this paper, we will consider the
standard classical tempered stable (denoted by stdCTS) distributions. This
ARMA(1,1)-GARCH(1,1) model is defined as follows: CTS-ARMA-GARCH
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1329

model, et  stdCTS(a, l+, l_). This distribution does not have a closed-form
solution for its probability density function. Instead, it is defined by its character-
istic function as follows: Let a 2 (0,2)\{1},C, l+, l_ > 0, and m ∈ ℝ. Then
a random variable X is said to follow the classical tempered stable (CTS) distribu-
tion if the characteristic function of X is given by
 
fx ðuÞ ¼ fCTS u : a, C, lþ , l, m
  
¼ exp ium  iuCT ð1  aÞ la1 þ l
a1
(48.2)
 a a a a

þ CGðaÞ ðlþ  iuÞ  lþ þ ðl  iuÞ  l ,

and we denote X  CTS(a, C, l+, l_, m).

The cumulants of X are defined by

1 ∂n  
Cn ðXÞ ¼ log E eiuX ju ¼ 0, n ¼ 1, 2, 3, . . . : :
i ∂u
n n

For the tempered stable distribution, we have E[X] ¼ c1(X) ¼ m. The cumulants
of the tempered stable distribution for n ¼ 2, 3, . . . are
 n an 
cn ðXÞ ¼ CGðn  aÞ lanþ þ ð1Þ l :

By substituting the appropriate value for the two parameters m and C into the
three tempered stable distributions, we can obtain tempered stable distributions
with zero mean and unit variance. That is, X  CTS(a, C, l+, l, 0) has zero mean
and unit variance by substituting
  
a2 1
C ¼ Gð2  aÞ la2 þ þl : (48.3)

The random variable X is referred to as the standard CTS distribution with

parameters (a, l+, l_) and denoted by X  stdCTS(a, l+, l_).

48.3 VaR for the ARMA-GARCH Model

In this section, we discuss VaR for the ARMA-GARCH model with normal and
tempered stable innovations.

48.3.1 VaR and Backtesting

The definition of VaR for a significance level  is

VaR ðXÞ ¼ inffx 2 ℝjPðX  xÞ > g:

1330 O. Carchano et al.

If we take the ARMA-GARCH model described in Sect. 48.2, we can define

VaR for the information until time t with significance level  as2
     
VaRt,  ytþ1 ¼ inf x 2 ℝPt ytþ1  x >  ,

where Pt(A) is the conditional probability of a given event A for the information
until time t.
Two models are considered: normal-ARMA(1,1)-GARCH(1,1) and
stdCTS-ARMA(1,1)-GARCH(1,1). For both models, the parameters have been
estimated for the time series between December 14, 2004 and December
31, 2008. For each daily estimation, we worked with 10 years of historical daily
performance for the S&P 500, DAX 30, and Nikkei 225 spot and futures indices.
More specifically, we used daily returns calculated based on the closing price of
those indices. In the case of futures indices, we constructed a unique continuous
time series using the different maturities of each futures index following the
methodology proposed by Carchano and Pardo (2009).3 Then, we computed
VaRs for both models.
The maximum likelihood estimation method (MLE) is employed to
estimate parameters of the normal-ARMA(1,1)-GARCH(1,1) model. For the CTS
distribution, the parameters are estimated as follows4:
1. Estimate parameters a0, a1, b1, a, b, c with normal innovations by the
MLE. Volatility clustering is captured by the GARCH model.
2. Extract residuals using those parameters. The residual distribution still presents
fat tail and skewness.
3. Fit the parameters of the innovation distribution (CTS) to the extracted residuals
using MLE. The fat tailed and skewed features of the residual distribution are
captured.
In order to determine the accuracy of VaR for the two models, backtesting using
Kupiec’s proportion of failures test (Kupiec 1995) is applied. We first calculate the
number of violations. Then, we compare the number of violations with the
conventional number of exceedances at a given significance level. In Table 48.1 the
number of violations and p-values for Kupiec’s backtest for the three stock indices
over the 41-year periods are reported. Finally, we sum up the number of violations
and their related p-values for 1 % VaRs for the normal and CTS-ARMA-GARCH
models.

2
VaR on the CTS distribution is described in the Appendix.
3
Thus, the last trading day of the front contract is chosen as the rollover date. Then, the return of
the day after the rollover date is calculated as the quotient between the closing price of the
following contract and the previous closing price of such contract. By doing so, all the returns
are taken from the same maturity.
4
A quasi-MLE strategy is followed because the ARMA-GARCH CTS model has too many
parameters. If all the parameters are estimated at once, then the GARCH parameters go to zero.
This strategy is also followed in Kim et al. (2009, 2010, 2011). For a discussion of the quasi- MLE
methodology, see Rachev et al. (2007, pp. 292–293) or Verbeek (2004, pp. 182–184).
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1331

Table 48.1 Normal-ARMA-GARCH versus CTS-ARMA-GARCH

1 year (255 days)
Dec. 14, 2004 Dec. 16, 2005 Dec. 21, 2006 Dec. 28, 2007
 Dec. 15, 2005  Dec. 20, 2006  Dec. 27, 2007  Dec. 31, 2008
N(p-value) N(p-value) N(p-value) N(p-value)
Model S&P 500 spot
Normal-ARMA- 1(0.2660) 3(0.7829) 8(0.0061) 10(0.0004)
GARCH
CTS-ARMA-GARCH 0 2(0.7190) 6(0.0646) 4(0.3995)
S&P 500 futures
Normal-ARMA- 3(0.7829) 3(0.7829) 7(0.0211) 9(0.0016)
GARCH
CTS-ARMA-GARCH 1(0.2660) 3(0.7829) 4(0.3995) 5(0.1729)
DAX 30 spot
Normal-ARMA- 4(0.3995) 4(0.3995) 3(0.7829) 6(0.0646)
GARCH
CTS-ARMA-GARCH 4(0.3995) 4(0.3995) 3(0.7829) 4(0.3995)
DAX 30 futures
Normal-ARMA- 3(0.7829) 5(0.1729) 6(0.0646) 6(0.0646)
GARCH
CTS-ARMA-GARCH 3(0.7829) 4(0.3995) 6(0.0646) 3(0.7829)
Nikkei 225 spot
Normal-ARMA- 2(0.7190) 4(0.3995) 5(0.1729) 5(0.1729)
GARCH
CTS-ARMA-GARCH 1(0.2660) 3(0.7829) 4(0.3995) 5(0.1729)
Nikkei 225 futures
Normal-ARMA- 2(0.7190) 2(0.7190) 7(0.0211) 5(0.1729)
GARCH
CTS-ARMA-GARCH 5(0.1729) 5(0.1729) 6(0.0646) 6(0.0646)
The number of violations (N) and p-values of Kupiec’s proportion of failures test for
the S&P 500, DAX 30, and Nikkei 225 spot and futures indices data has been shown.
Normal-ARMA-GARCH and CTS-ARMA-GARCH compared

Based on Table 48.1, we conclude the following for the three stock indices. First,
a comparison of the normal and tempered stable models indicates that there are no
cases using the tempered stable model at the 5 % significance level, whereas the normal
model is rejected five times. This evidence is consistent with the findings of Kim
et al. (2011). Second, a comparison of the spot and futures indices indicates that spot
data provide less than or the same number of violations than futures data. One potential
explanation is that futures markets are more volatile, particularly, when the market
falls.5 This overreaction to bad news could cause the larger number of violations.

5
We compared the spot and futures series when the markets discount bad news (negative returns).
We find that for the three stock indices, futures volatility is significantly greater than spot volatility
at a 5 % significance level. Moreover, for all three stock indices, the minimum return and the 1 %
percentile return are also lower for futures data than spot data.
1332 O. Carchano et al.

48.4 Introduction of Trading Volume

In the previous section, we showed the usefulness of the tempered stable model for
stock index futures. Motivated by the vast literature linking trading volume and
volatility, for the first time we investigate whether the introduction of trading
volume in the CTS model could improve its ability to forecast 1-day-ahead VaR.
Let (St)t
0 be the asset price process and (yt)t
0 be the return process of (St)t
0
defined by yt ¼ log StS1
t
. We propose the following ARMA(1,1)-GARCH(1,1) with
trading volume model:


yt ¼ ayt1 þ bst1 et1 þ st et þ c
(48.4)
s2t ¼ a0 þ a1 s2 t1 e2 t1 þ b1 s2 t1 þ g1 Volt1 ,

where e0 ¼ 0 and a sequence (et)tcN ¼ 0 of iid real random variables. The innovation
et is assumed to be the tempered stable innovation. We will consider the standard
classical tempered stable distributions. This new ARMA(1,1)-GARCH(1,1)-V
model is defined as follows:

CTS-ARMA-GARCH-V model : et  stdCTSða, lþ , l Þ:

The inclusion of lagged volume as an independent variable along with lagged

volatility into the model may cause a problem of multicollinearity. In order to
determine the seriousness of the problem, we calculated the model without volume,
extracted the GARCH series, and determined the degree of collinearity between
both variables. The most recommended measure in the literature is to calculate the
condition index following Belsley et al. (1980) and observe if the index exceeds
20, in which case collinearity is considered to be grave. In our case, the calculated
value was 4.9268, 3.2589, and 4.5569 for the S&P, DAX, and Nikkei, respectively.
Therefore, we concluded that collinearity is a minor problem.
Moreover, the ARMA-GARCH model is only affected in the GARCH frame-
work, particularly the equation coefficients (i.e., the volume variable can appear
insignificant when it is indeed significant), but not the numerical estimation of the
variance; neither is the forecast power of the global model. As our objective is to
forecast the VaR, we believe that the multicollinearity problem can be ignored
because the results will not be affected.

48.4.1 Different Variants of Trading Volume

For the S&P 500 cash and futures markets, we test the following versions of trading
volume in order to determine which one would be the most appropriate:
• Lagged trading volume in levels: V(t  1)
• Logarithm of lagged trading volume: log [V(t  1)]
• Relative change of lagged trading volume: log [V(t  1)/V(t  2)]
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1333

Table 48.2 CTS-ARMA-GARCH with lagged volume

1 year (255 days)
Dec. 14, 2004 Dec. 16, 2005 Dec. 21, 2006 Dec. 28, 2007
 Dec. 15, 2005  Dec. 20, 2006  Dec. 27, 2007  Dec. 31, 2008
N(p-value) N(p-value) N(p-value) N(p-value)
Model S&P 500 spot
CTS-ARMA-GARCH- 16(0.0000) 10(0.0004) 23(0.0000) 26(0.0000)
V(t  1)
CTS-ARMA-GARCH- 1(0.2660) 4(0.3995) 10(0.0004) 6(0.0646)
log [V(t  1)]
CTS-ARMA-GARCH- 0 2(0.7190) 6(0.0646) 5(0.1729)
ln [V(t  1)/V(t  2)]
S&P 500 futures
CTS-ARMA-GARCH- 1(0.2660) 11(0.0001) 4(0.3995) 6(0.0646)
V(t  1)
CTS-ARMA-GARCH- 1(0.2660) 3(0.7829) 4(0.3995) 5(0.1729)
log [V(t  1)]
CTS-ARMA-GARCH- 1(0.2660) 3(0.7829) 3(0.7829) 5(0.1729)
ln [V(t  1)/V(t  2)]
CTS-ARMA-GARCH- 0 1(0.2660) 15(0.0000) 3(0.7829)
V$(t  1)
CTS-ARMA-GARCH- 3(0.7829) 3(0.7829) 4(0.3995) 5(0.1729)
log [V$(t  1)]
CTS-ARMA-GARCH- 2(0.7190) 0 8(0.0061) 8(0.0061)
ln [V$(t  1)/V$(t  2)]
The number of violations (N) and p-values of Kupiec’s proportion of failures test for the
S&P 500 spot and futures indices with the different variants of volume using the stdCTS-
ARMA(1,1)-GARCH(1,1) model has been shown. V(t  1), log [V(t  1)], and ln [V(t  1)/
V(t  2)] stand for levels, logarithm, and relative change of the lagged trading volume, respec-
tively. V$(t  1), log [V$(t  1)], and ln [V$(t  1)/V$(t  2)] stand for levels, logarithm, and
relative change of the lagged trading volume in dollars, respectively

The spot series trading volume is in dollars; for the futures series, the trading value
is in number of contracts. We can calculate the volume of the futures market in dollars
too. The tick value of the S&P 500 futures contract is 0.1 index points or $25.
Multiplying the number of contracts by the price and finally by $250 (the contract’s
multiple), we obtain the trading volume series for the futures contract in dollars. Thus,
for the futures contract we get three new versions of trading volume to test:
• Lagged trading volume in dollars: V$(t  1)
• Logarithm of trading volume in dollars: log [V$(t  1)]
• Relative change of lagged trading volume in dollars: log [V$(t  1)/V$(t  2)]
By doing that, we can determine which series (in dollars or in contracts) seems to
be more useful for the futures index.
In Table 48.2 we report the number of violations and p-values of Kupiec’s backtest
for the different versions of the CTS-ARMA-GARCH-V model for the S&P 500 spot
1334 O. Carchano et al.

and futures indices. We count the number of violations and the corresponding
p-values for 1 %-VaRs of both markets. From Table 48.2, we conclude the following:
• The model with the lagged trading volume in level is rejected at the 1 %
significance level in all 4 years for the S&P 500 spot, and for the second period
(2005–2006) for the S&P 500 futures.
• The logarithm of trading volume in the model is rejected at the 5 % significance
level for the spot market for the third period (2006–2007), but it is not rejected in
any period for the futures market.
• The relative change of the lagged volume is not rejected at the 5 % significance
level in any period in either market. Of the three versions of trading volume tests,
this version seems to be the most useful for both spot and futures markets.
• The results for trading volume in contracts and the trading volume in dollars in the
futures market indicate that the former is rejected at the 1 % significance level only
for the lagged trading volume in level in the second period (2005–2006). Trading
volume in dollars is rejected three times, for the lagged trading volume in levels
for the third period (2006–2007) and for the lagged relative trading volume change
in the last two periods (2006–2007 and 2007–2008). These findings suggest that
the trading volume in contracts is the preferred measure.

48.4.2 Lagged Relative Change of Trading Volume

As we have just seen, the variant of trading volume that seems more useful
for forecasting 1-day-ahead VaR using CTS-ARMA-GARCH is the relative change
of trading volume. Next, we compare the original CTS-ARMA-GARCH model
with the new CTS-ARMA-GARCH-V model where V is the lagged relative change
of trading volume. Table 48.3 shows the number of violations and p-values of
Kupiec’s backtest for the two models for the three stock indices and both markets.
We sum up the number of violations and the corresponding p-values for 1 % VaRs for
each case.
Our conclusions from Table 48.3 are as follows. For the spot markets, the
introduction of trading volume does not mean a reduction in the number of
violations in any period for any index. However, for the futures markets, the
numbers of violations are the same or lower for the model with trading
volume than with the original model. Thus, by introducing trading volume, we
get a slightly more conservative model, increasing the VaR forecasted for futures
equity markets.

48.4.3 Lagged Trading Volume or Forecasting Contemporaneous

Trading Volume

Although there is some evidence which supports the relationship between lagged
trading volume and volatility, the literature is not as extensive as the studies
that establish a strong link between volatility and contemporaneous trading
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1335

Table 48.3 CTS-ARMA-GARCH versus CTS-ARMA-GARCH-V

1 year (255 days)
Dec. 14, 2004 Dec. 16, 2005 Dec. 21, 2006 Dec. 28, 2007
 Dec. 15, 2005  Dec. 20, 2006  Dec. 27, 2007  Dec. 31, 2008
N(p-value) N(p-value) N(p-value) N(p-value)
Model S&P 500 spot
CTS-ARMA-GARCH 0 2(0.7190) 6(0.0646) 4(0.3995)
CTS-ARMA-GARCH- 0 2(0.7190) 6(0.0646) 5(0.1729)
ln [V(t  1)/V(t  2)]
S&P 500 futures
CTS-ARMA-GARCH 1(0.2660) 3(0.7829) 4(0.3995) 5(0.1729)
CTS-ARMA-GARCH- 1(0.2660) 3(0.7829) 3(0.7829) 5(0.1729)
ln [V(t  1)/V(t  2)]
DAX 30 spot
CTS-ARMA-GARCH 4(0.3995) 4(0.3995) 3(0.7829) 4(0.3995)
CTS-ARMA-GARCH- 5(0.1729) 4(0.3995) 3(0.7829) 5(0.1729)
ln [V(t  1)/V(t  2)]
DAX 30 futures
CTS-ARMA-GARCH 3(0.7829) 4(0.3995) 6(0.0646) 3(0.7829)
CTS-ARMA-GARCH- 3(0.7829) 4(0.3995) 5(0.1729) 3(0.7829)
ln [V(t  1)/V(t  2)]
Nikkei 225 spot
CTS-ARMA-GARCH 1(0.2660) 3(0.7829) 4(0.3995) 5(0.1729)
CTS-ARMA-GARCH- 3(0.7829) 3(0.7829) 4(0.3995) 5(0.1729)
ln [V(t  1)/V(t  2)]
Nikkei 225 futures
CTS-ARMA-GARCH 5(0.1729) 5(0.1729) 6(0.0646) 6(0.0646)
CTS-ARMA-GARCH- 4(0.3995) 4(0.3995) 6(0.0646) 6(0.0646)
ln [V(t  1)/V(t  2)]
The number of violations (N) and p-values of Kupiec’s proportion of failures test for the S&P
500, DAX 30, and Nikkei 225 spot and futures indices has been reported. CTS-ARMA-GARCH
and CTS-ARMA-GARCH with lagged relative change of trading volume compared

volume. As there are countless ways to try to forecast trading volume, we begin by
introducing contemporaneous trading volume relative change in the model as
a benchmark to assess whether it is worthwhile to forecast trading volume.
In Table 48.4 we show the number of violations and p-values of Kupiec’s
backtest for the CTS-ARMA-GARCH with contemporaneous and lagged
relative change of trading volume for the three stock indices for both markets.
We count the number of violations and the corresponding p-values for 1 % VaRs for
the six indices.
Our conclusions based on the results reported in Table 48.4 are as follows.
First, with the exception of the S&P 500 futures, the introduction of the contem-
poraneous relative change of trading volume in the model is rejected at the 1 %
significance level for the last period analyzed (2007–2008). In the case of the S&P
1336 O. Carchano et al.

Table 48.4 CTS-ARMA-GARCH with contemporaneous volume versus CTS-ARMA-GARCH

with lagged volume
1 year (255 days)
Dec. 14, 2004 Dec. 16, 2005 Dec. 21, 2006 Dec. 28, 2007
 Dec. 15, 2005  Dec. 20, 2006  Dec. 27, 2007  Dec. 31, 2008
N(p-value) N(p-value) N(p-value) N(p-value)
Model S&P 500 spot
CTS-ARMA-GARCH- 0 3(0.7829) 3(0.7829) 8(0.0061)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 0 5(0.1729) 6(0.0646) 5(0.1729)
ln [V(t  1)/V(t  2)]
S&P 500 futures
CTS-ARMA-GARCH- 0 1(0.2660) 7(0.0211) 6(0.0646)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 1(0.2660) 3(0.7829) 3(0.7829) 5(0.1729)
ln [V(t  1)/V(t  2)]
DAX 30 spot
CTS-ARMA-GARCH- 0 1(0.2660) 3(0.7829) 11(0.0001)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 5(0.1729) 4(0.3995) 3(0.7829) 5(0.1729)
ln [V(t  1)/V(t  2)]
DAX 30 futures
CTS-ARMA-GARCH- 1(0.2660) 1(0.2660) 2(0.7190) 8(0.0061)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 3(0.7829) 4(0.3995) 5(0.1729) 3(0.7829)
ln [V(t  1)/V(t  2)]
Nikkei 225 spot
CTS-ARMA-GARCH- 3(0.7829) 5(0.1729) 7(0.0211) 8(0.0061)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 3(0.7829) 3(0.7829) 4(0.3995) 5(0.1729)
ln [V(t  1)/V(t  2)]
Nikkei 225 futures
CTS-ARMA-GARCH- 1(0.2660) 1(0.2660) 3(0.7829) 11(0.0001)
ln [V(t)/V(t  1)]
CTS-ARMA-GARCH- 4(0.3995) 4(0.3995) 6(0.0646) 6(0.0646)
ln [V(t  1)/V(t  2)]
The number of violations (N) and p-values of Kupiec’s proportion of failures test for the S&P
500, DAX 30, and Nikkei 225 spot and futures indices has been shown. CTS-ARMA-GARCH
with the relative change of trading volume and CTS-ARMA-GARCH with lagged relative change
of trading volume compared

500 futures, it is rejected at the significance level of 5 % for the third period
(2006–2007). Second, the model with lagged relative change of trading volume is
not rejected for any stock index or market. It seems to be more robust than
contemporaneous trading volume (although, in general, there are fewer violations
when using it).
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1337

Our results suggest that it is not worth making an effort to predict contemporane-
ous trading volume because the forecasts will be flawed and two variables
would have to be predicted (VaR and contemporaneous trading volume). Equiva-
lently, the lagged trading volume relative change appears to be more robust because it
is not rejected in any case, although it provides a poor improvement to the model.

48.5 Conclusions

Based on an empirical analysis of spot and futures trading for the S&P 500, DAX
30, and Nikkei 225 stock indices, in this paper we provide empirical evidence
about the usefulness of using classical tempered stable distributions for predicting
1-day-ahead VaR. Unlike prior studies that investigated CTS models in the
cash equity markets, we analyzed their suitability for both spot markets and
futures markets. We find in both markets the CTS models perform better in forecast-
ing 1-day-ahead VaR than models that assume innovations follow the normal law.
Second, we introduced trading volume into the CTS model. Our empirical
evidence suggests that lagged trading volume relative change provides a slightly
more conservative model (i.e., reduces the number of violations) to predict
1-day-ahead VaR for stock index futures contracts. We cannot state the same
for the cash market because the results are mixed depending on the index. After
that, we introduced contemporaneous trading volume to try to improve the forecast-
ing ability of the model, but in the end, it did not seem to be worth the effort. That is,
trading volume appeared not to offer enough information to improve forecasts.
Finally, we compared the number of violations of the estimated VaR in the spot
and futures equity markets. For the CTS model without volume, in general, we find
fewer violations in the spot indices than in the equivalent futures contracts.
In contrast, our results suggest that the number of violations in futures markets
is less in the case of the CTS model with trading volume in comparison to the
CTS model that ignores trading volume. But if we contrast spot and futures equity
markets, violations are still greater for futures than in spot markets. A possible
reason is that futures markets demonstrate extra volatility or an overreaction when
the market falls with respect to their corresponding spot markets.

Appendix: VaR on the CTS Random Variable

Let X be a CTS random variable. Since the CTS random variable is continuous and
infinitely divisible, we obtain VaR(X) ¼ FX(), where the cumulative distribu-
tion function FX of X is provided by the following proposition.

Proposition Let X be an infinitely divisible random variable and fx(u) be the

characteristic function of X. If there is a r > 0 such that jfx(Z)j < 1 for all the
complex z with ℑ(z) ¼ r, then
1338 O. Carchano et al.

ð / 
exr ixu fX ðu þ irÞ
FX ð x Þ ¼ ℜ e du , for x 2 ℝ (48.5)
p 0 r  ui

where ℜ(z) is the real part of a complex number z.

Proof By the definition of the cumulative density function, we have

ðx
FX ð x Þ ¼ Pð X  x Þ ¼ f XðtÞdt,
1

where fX(x) is the density function of X. The probability density function fX(t) can be
obtained from the characteristic function fX by the complex inverse formula (see
Doetsch 1970); that is,
ð 1þir
1
f X ðt Þ ¼ eitz fX ðzÞdz,
2p 1þir
and we have
ðx ð 1þip
1
FX ð x Þ ¼ eitz fX ðzÞdzdt
1 2p1þip
ð 1þip ð x
1
¼ eitz dtfX ðzÞdz:
2p 1þip 1

Note that if r > 0, then

   it 
 
lim eitðaþirÞ  ¼ lim e ðaþirÞ  ¼ lim ert ¼ 0, a 2 ℝ,
t!1 t!1 t!1

and hence
ðx
1  itz x 1
eitz dt ¼  e 1
¼  eixz
1 iz iz

where z ∈ ℂ with ℑ(z) ¼ r Thus, we have

ð
1 1þip 1 ixz
FX ðxÞ ¼  e fX ðzÞdz
2p 1þip iz
ð
1 1 1
¼ eixðuþirÞ fx ðu þ irÞdu
2p 1 iðu þ irÞ
ð
exr 1 ixu fX ðu þ irÞ
¼ e du:
2p 1 r  iu
48 A Quasi-Maximum Likelihood Estimation Strategy for Value-at-Risk Forecasting 1339

Let
fX ðu þ irÞ
gr ð uÞ ¼ :
r  iu

Then we can show that gr ðuÞ ¼ gr ðuÞ with u ∈ ℝ and hence we have
ð1
ð 1 
eixu gr ðuÞdu ¼ 2ℜ eixu gr ðuÞdu :
1 0

Therefore, we obtain Eq. 48.5.

References
Asai, M., & McAleer, M. (2009). The structure of dynamic correlations in multivariate stochastic
volatility models. Journal of Econometrics, 150, 182–192.
Belsley, D., Kuh, E., & Welsch, R. (1980). Regression diagnostics: Identifying influential data and
sources of collinearity. New York: Wiley.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroscedasticity. Journal of
Econometrics, 31, 307–327.
Carchano, O., & Pardo, A. (2009). Rolling over stock index futures contracts. Journal of Futures
Markets, 29, 684–694.
Clark, P. K. (1973). A subordinated stochastic process model with finite variance for speculative
prices. Econometrica, 41, 135–155.
Conrad, J. S., Hameed, A., & Niden, C. (1994). Volume and autocovariances in short-horizon
individual security returns. Journal of Finance, 49, 1305–1329.
Copeland, T. E. (1976). A model of asset trading under the assumption of sequential information
arrival. Journal of Finance, 34, 1149–1168.
Doetsch, G. (1970). Introduction to the theory and application of the Laplace transformation.
New York: Springer.
Engle, R. (1982). Autoregressive conditional heteroskedasticity with estimates of the variance
with estimates of the variance of United Kingdom inflation. Econometrica, 50, 987–1007.
Engle, R., Focardi, S. M., & Fabozzi, F. J. (2008). ARCH/GARCH Models in applied financial
econometrics. In F. J. Fabozzi (Ed.), Handbook of finance. Hoboken: Wiley. vol III.
Epps, T. W., & Epps, L. M. (1976). The stochastic dependence of security price changes and
transaction volumes: Implications for the mixture-of-distribution hypothesis. Econometrica,
44, 305–321.
Foster, F. J. (1995). Volume-volatility relationships for crude oil futures markets. Journal of
Futures Markets, 15, 929–951.
Fung, H. G., & Patterson, G. A. (1999). The dynamic relationship of volatility, volume, and market
depth in currency futures markets. Journal of International Financial Markets Institutions and
Money, 17, 33–59.
Garcia, P., Leuthold, M. R., & Zapata, H. (1986). Lead-lag relationships between trading volume
and price variability: New evidence. Journal of Futures Markets, 6, 1–10.
Gwilym, O., MacMillan, D., & Speight, A. (1999). The intraday relationship between volume and
volatility in LIFFE futures markets. Applied Financial Economics, 9, 593–604.
Huang, C. Y., & Lin, J. B. (2004). Value-at-risk analysis for Taiwan stock index futures: Fat tails
and conditional asymmetries in return innovations. Review of Quantitative Finance and
Accounting, 22, 79–95.
1340 O. Carchano et al.

Karpoff, J. M. (1987). The relation between price changes and trading volume: A survey. Journal
of Financial and Quantitative Analysis, 22, 109–126.
Kim, Y. S., Rachev, S. T., Chung, D. M., & Bianchi, M. L. (2009). The modified tempered stable
distribution, GARCH-models and option pricing. Probability and Mathematical Statistics, 29,
91–117.
Kim, Y. S., Rachev, S. T., Bianchi, M. L., & Fabozzi, F. J. (2010). Tempered stable and tempered
infinitely divisible GARCH models. Journal of Banking and Finance, 34, 2096–2109.
Kim, Y. S., Rachev, S. T., Bianchi, M. L., Mitov, I., & Fabozzi, F. J. (2011). Time series analysis
for financial market meltdowns. Journal of Banking and Finance, 35, 1879–1891.
Kupiec, P. (1995). Techniques for verifying the accuracy of risk measurement models. Journal of
Derivatives, 6, 6–24.
Ragunathan, V., & Peker, A. (1997). Price variability, trading volume and market depth: Evidence
from the Australian futures market. Applied Financial Economics, 7, 447–454.
Schwert, G. W. (1989). Why does stock market volatility change over time? Journal of Finance,
44, 1115–1153.
Smirlock, M., & Starks, L. (1985). A further examination of stock price changes and transactions
volume. Journal of Financial Research, 8, 217–225.
Tang, L. T., & Shieh, J. S. (2006). Long memory in stock index futures markets: A value-at-risk
approach. Physica A: Statistical Mechanics and its Applications, 366, 437–448.
Tauchen, G., & Pitts, M. (1983). The price variability-volume relationship on speculative markets.
Econometrica, 51, 485–505.
Verbeek, M. (2004). A guide to modern econometrics. Hoboken: Wiley.
Wang, G., & Yau, J. (2000). Trading volume, bid-ask spread, and price volatility in futures
markets. Journal of Futures Markets, 20, 943–970.
Computer Technology for Financial
Service 49
Fang-Pang Lin, Cheng-Few Lee, and Huimin Chung

Contents
49.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1342
49.1.1 Information Technology (IT) for Financial Services . . . . . . . . . . . . . . . . . . . . . . . . 1342
49.1.2 Competitiveness Through IT Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1343
49.2 Performance Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1346
49.2.1 High-End Computing Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1346
49.2.2 Compute Intensive IT Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1349
49.2.3 Data-Intensive IT Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1352
49.3 Distributed and Parallel Financial Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1354
49.3.1 Financial Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1355
49.3.2 Monte Carlo Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1357
49.3.3 Distribution and Parallelism Based on Random Number Generation . . . . . . . 1360
49.4 Case Study and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1367
49.4.1 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1367
49.4.2 Grid Platforms Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1369
49.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1375
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1377

F.-P. Lin (*)

National Center for High Performance Computing, Hsinchu, Taiwan
e-mail: fplin@nchc.narl.org.tw
C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: cflee@business.rutgers.edu
H. Chung
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: chunghui@mail.nctu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1341
DOI 10.1007/978-1-4614-7750-1_49,
# Springer Science+Business Media New York 2015
1342 F.-P. Lin et al.

Abstract
Securities trading is one of the few business activities where a few seconds processing
delay can cost a company big fortune. The growing competition in the market
exacerbates the situation and pushes further towards instantaneous trading even in
split second. The key lies on the performance of the underlying information system.
Following the computing evolution in financial services, it was a centralized process
to begin with and gradually decentralized into a distribution of actual application logic
across service networks. Financial services have tradition of doing most of its heavy-
duty financial analysis in overnight batch cycles. However, in securities trading it
cannot satisfy the need due to its ad hoc nature and requirement of fast response. New
computing paradigms, such grid and cloud computing, aiming at scalable and
virtually standardized distributed computing resources, are well suited to the chal-
lenge posed by the capital market practices. Both consolidate computing resources by
introducing a layer of middleware to orchestrate the use of geographically distributed
powerful computers and large storages via fast networks. It is nontrivial to harvest the
most of the resources from this kind of architecture. Wiener process plays a central
role in modern financial modeling. Its scaled random walk feature, in essence, allows
millions of financial simulation to be conducted simultaneously. The sheer scale can
only be tackled via grid or cloud computing. In this study the core computing
competence for financial services is examined. Grid and cloud computing will be
briefly described. How the underlying algorithm for financial analysis can take
advantage of grid environment is chosen and presented. One of the most popular
practiced algorithms Monte Carlo simulation is used in our case study for option
pricing and risk management. The various distributed computational platforms are
carefully chosen to demonstrate the performance issue for financial services.

Keywords
Financial service • Grid and cloud computing • Monte Carlo simulation • Option
pricing • Risk management • Cyberinfrastructure • Random number generation •
High-end comptuing • Financial simulation • Information technology

49.1 Introduction

49.1.1 Information Technology (IT) for Financial Services

The finance services industry involves a broad range of organizations such as banks,
credit card companies, insurance companies, consumer finance companies, stock
brokerages, investment funds, and some government-sponsored enterprises. The
industry represents a significant share of the global market. Information technology
(IT) in the financial service industry is considered as an indispensable tool for
productivity as well as competitiveness in the market. The IT spending in financial
service industry grows constantly across different industry verticals (banking,
insurance, and securities and investments). The impact directly from the use of
advanced IT brings on financial services industry on the rise.
49 Computer Technology for Financial Service 1343

The structure of the industry has changed significantly in the last two decades as
companies, which are not traditionally viewed as financial service providers, have
taken advantage of opportunities created by technology to enter the market. New
technology-based services keep emerging. These changes are direct result of the
interaction of technology with the industrial environment, such as economic atmo-
sphere, societal pressures, and the legal/regulatory environment in which the
financial service industry operates. The effects of IT on the internal operations,
the structure, and the types of services offered by the financial service industry have
been particularly profound (Phillips et al. 1984; Hauswald and Marquez 2003;
Griffiths and Remenyi 2003). IT technology has been and continues to be both
a motivator and facilitator of change in the financial service industry, which
ultimately leads to competitiveness of the industry. The change is in particular
radical after 1991 when the World Wide Web was invented by Tim Berners-Lee
and his group for information sharing in the community of high energy physics. It
was later introduced to the rest of the world, which subsequently changed the face
of how people doing business today.
Informational considerations have long been recognized to determine not only the
degree of competition but also the pricing and profitability of financial services and
instruments. Recent technological progress has dramatically affected the production
and availability of information, thereby changing the nature of competition in such
informationally sensitive markets. Hauswald and Marquez (2003) investigate how
advances in information technology (IT) affect competition in the financial services
industry, particularly credit, insurance, and securities markets. Two aspects of
improvement in IT are focused: better processing and easier dissemination of infor-
mation. In other words, two dimensions of technology progress that affect competi-
tion in financial services can be defined as advances in the ability to process and
evaluate information and in the ease of obtaining information generated by compet-
itors. While better technology may result in improved information processing, it
might also lead to low cost or even free access to information through, for example,
informational spillovers. They show that in the context of credit screening, better
access to information decreases interest rates and the returns from screening. On the
other hand, an improved ability to process information increases interest rates and
bank profits. Hence predictions regarding financial claims’ pricing hinge on the
overall effect ascribed to technological progress. Their results conclude that in
general financial market informational asymmetries drive profitability.
The viewpoint of Hauswald and Marquez is adopted in this work. Assuming
competitors in the dynamics of financial market possess similar capacity, the infor-
mational asymmetries can be created sometimes only between seconds and now are
possible to be achieved through the outperformance of underlying IT platforms.

49.1.2 Competitiveness Through IT Performance

Following the computing evolution in financial services, it was a centralized pro-

cess to begin with and gradually decentralized into a distribution of practical
1344 F.-P. Lin et al.

cluster computing grid computing cloud computing big data

Search Volume index Google Trends

0
2004 2005 2006 2007 2008 2009 2010 2011 2012
News reference volume

Fig. 49.1 The trend history from Google Trend according to global Search Volume and global
News Reference Volume, in which the alphabetic letters represent the specific events that relate to
each curve

trading application logic across service networks. Financial services have tradition
of doing most of its heavy lifting financial analysis in overnight batch cycles.
However, in securities trading it cannot satisfy the need due to its ad hoc nature
and requirement of fast response.
New computing paradigms, grid computing and cloud computing were subse-
quently emerged in the last decade. The grid computing was initially incorporated
into the core context of a well-referenced Atkins’ report of National Science Board
of the United States, namely, “Revolutionizing Science and Engineering Through
Cyberinfrastructure” (Atkins et al. 2003), which lays down a visionary path for
future IT platform development of the world. One may observe this trend from
statistics from Google Trend regarding the global Search Volume and global News
Reference Volume of key phrases of “cluster computing,” “grid computing,”
“cloud computing,” and “Big Data” (Fig. 49.1), which represents four main stream
computing paradigms in high-end quantitative analysis.
Cluster computing is a group of coupled computers that work closely together so
that in many respects they can be viewed as though they are a single computer.
They are connected with high-speed local area networks and the purpose is usually
to gain more compute cycles with better cost performance and higher availability.
The grid computing aims at virtualizing scalable geographically distributed com-
puting and observatory resources to maximize compute cycles and data transaction
rates with minimum cost. Cloud computing is more of recent development owing to
the similar technology used in global information services providers, such as
Google and Amazon. The cloud is referred to as a subset of Internet if to be
explained in a simplest fashion. Within the cloud the computers also talk with
servers instead of communicating with each other similarly to that of peer-to-peer
computing (Milojicic et al. 2002). There are no definitive definitions for the above
49 Computer Technology for Financial Service 1345

terminology. However, people tend to view clusters as one of foundational com-

ponents of grids, or grids as a meta-cluster on wide area networks. This is also
known as horizontal integration. The cloud virtualizes further the compute, store,
and network in a utility sense and provides an interface between users and grids. We
refer to Foster et al. (2008) and Yelick et al. (2011) for a comparison. This
perspective considers grids as a backbone of cyberinfrastructure to support the
clouds. Similarly, in early days of development of grids, there is a so-called
“@HOME” style PC grids (Korpela et al. 2001), which are exactly working on at
least ten of thousands of PCs, in which owners of PCs donate their CPU times when
their machines are in idle. The PC grids can be specifically categorized as clouds.
Figure 49.1 shows that there is a gradual drop in the curve of search volume for grid
computing and cluster computing, and many surges, grows but a recent quick drop on
cloud computing since its introduction in mid-2007. The new rising technology is Big
Data (Bughin et al. 2010), which implies a paradigm shift from compute centric,
network centric gradually to data-centric computing. However, the size of the search
volume strongly relates to the degree of maturity of each computing paradigm. This is
obvious in cluster computing. Clusters are the major market products, either in
supercomputers from big vendors, such as IBM, HP, SGI, and NEC, or from aggre-
gation of PCs in university research laboratories. Figure 49.1 also implies constant
market need for high-end computing. The performance and security issues are funda-
mental to general distributed and parallel computing, which also remain as a challenge
to cluster, grid, cloud, and Big Data (Lauret et al. 2010; Ghoshal et al. 2011;
Ramakrishnan et al. 2011). Performance models in compute-based grid, which is
also cloud-like, environment, are adopted in this work. The general definitions of grid
and cloud computing will be introduced and briefly compared. To tackle the core
performance issue, grids are chosen to demonstrate how fundamental financial calcu-
lations can be improved, hence leverage the financial service.
Grid computing, following by cloud computing as shown in Fig. 49.1, has been
matured to serve as a production environment for finance services in recent years.
Grid computing is well suited to the challenge posed by the capital market prac-
tices. In this study the core computing competence for financial services will be
examined and how underlying algorithms for financial analysis can take advantage
of grid environment scrutinized. One of the most popular practiced algorithms is
Monte Carlo simulation (MCS), and it will be specifically used in our case study for
calculations of option pricing and for value at risk (VaR) in risk management.
Three grid platforms are carefully chosen to exploit the performance issue for
financial services. The first one is traditional grid platform with heterogeneous and
distributed resources. Usually digital packets are connected via optical fibers.
For long distance, depending on network traffics, it will produce approximately
150-300 microseconds (mm) latency across the Pacific Ocean. This is the
physical constrain of light speed when traveling through the fiber channels. There-
fore, even in split-second packets can still travel to anywhere in the world. The
Pacific Rim Applications and Grid Middleware Assembly (PRAGMA) grid is
a typical example, which linked with 14 countries and 36 sites. The system
is highly heterogeneous. The computer nodes mounted to PRAGMA grid range
1346 F.-P. Lin et al.

from usual PC clusters to high-end supercomputers. The second one is a special

Linux, or DRBL, PC cluster. It converts system into a homogenous Linux system
and exploits the compute cycles of the cluster. The intention is to provide dynamic
and flexible resources to cope better with uncertainty of the traders’ cycle demand.
Finally, PC grid is chosen to demonstrate finance services that can be effectively
conducted through a cloud-based computing. The usefulness of PC grid is based on
the fact that 90 % of CPUs time of PCs were in idled status.

49.2 Performance Enhancement

In this section two types of grid systems, compute intensive and data intensive,
respectively, are introduced. The classification of the types is based on various grid
applications. Traditionally, the grid systems provide a general platform to harvest or to
scavenge, if used only in idle status, compute cycles for a collection of resources
across boundaries of institutional administration. In real world most applications are in
fact data centric. For example, in a trading center, it collects tick-by-tick volume data
from all related financial markets and is driven by informational flows, hence typical
data centric. However, as noted in Sect. 49.3.2.1, the core competence still lies on the
performance enhancement of the IT system. The following two subsections will give
more details of compute intensive as well as data-intensive grid systems by a survey of
current development of grids specifically for financial services. In some cases, e.g.,
high-frequency data with real-time analysis, two systems have to work together to get
better performance. Our emphasis will be more on compute intensive grid system.

49.2.1 High-End Computing Technology

49.2.1.1 Definitions of High-End Computing

Grid was coined by Ian Foster (Foster and Kessleman 2004) who gave the essence
of the definitions as below:
The sharing that we are concerned with is not primarily file exchange but rather direct
access to computers, software, data, and other resources, as is required by a range of
collaborative problem solving and resource-brokering strategies emerging in industry,
science, and engineering. This sharing is, necessarily, highly controlled, with resource
providers and consumers defining clearly and carefully just what is shared, who is allowed
to share, and the conditions under which sharing occurs. A set of individuals and/or
institutions defined by such sharing rules form what we call a virtual organization.

The definition is centered on the concept of virtual organization, but it is not

explicit enough to explain what the grid is. Foster then provides additional checklist
as below to safeguard the possible logic pitfalls of the definition. Hereby, grid is
a system that:
1. Coordinates resources that are not subject to centralized control.
A grid integrates and coordinates resources and users that live within different
control domains – for example, the user’s desktop vs. central computing, different
49 Computer Technology for Financial Service 1347

administrative units of the same company, or different companies – and addresses

the issues of security, policy, payment, membership, and so forth that arise in these
settings. Otherwise, we are dealing with a local management system.
2. Using standard, open, general-purpose protocols and interfaces.
A grid is built from multipurpose protocols and interfaces that address such
fundamental issues as authentication, authorization, resource discovery, and
resource access. As I discuss further below, it is important that these protocols
and interfaces be standard and open. Otherwise, we are dealing with an appli-
cation specific system.
3. To deliver nontrivial qualities of service.
A grid allows its constituent resources to be used in a coordinated fashion to
deliver various qualities of service, relating, for example, to response time,
throughput, availability, and security, and/or co-allocation of multiple resource
types to meet complex user demands, so that the utility of the combined system
is significantly greater than that of the sum of its parts.
The definition of grid thus far is well accepted and has been stably used up to
now. The virtual organization (VO) has strong implication of community driven
and collaborative sharing of distributed resources. The advance of development of
optical fiber network in recent years plays a critical role of why grids can be
a reality. It is also the reason why now the computing paradigm shifts to distrib-
uted/grid computing.
Additionally, perhaps the most generally useful definition is that a grid consists
of shared heterogeneous computing and data resources networked across admin-
istrative boundaries. Given such a definition, a grid can be thought of as both an
access method and a platform, with grid middleware being the critical software that
enables grid operation and ease of use.
The term “cloud computing” has been used to refer to different concepts,
models, and services over the last few years. The definition for cloud computing
provided by the National Institute of Standards and Technology (NIST) is well
received in the IT community, which defines cloud computing as a model for
enabling convenient, on-demand network access to a shared pool of configurable
computing resources (e.g., networks, servers, storage, applications, and services)
that can be rapidly provisioned and released with minimal management effort
service provider interaction (Mell and Grance 2011). The model gains popularity
in the industry for its emphasis on pay-as-you-go and elasticity, the ability to
quickly expand and collapse the utilized service as demand requires. Thus new
approaches to distributed computing and data analysis have also emerged in
conjunction with the growth of cloud computing. These include models like
MapReduce (Dean and Ghemawat 2004) and scalable key-value stores like Big
Table (Chang et al. 2006).
From the high-end computing perspective, cloud computing technology allows
users to have the ability to get on-demand access to resources to replace or
supplement existing systems, as well as the ability to control the software environ-
ment. Yet the core competence still lies on the performance of financial calculation
and further of the transactions of financial processes. This work will focus on the
1348 F.-P. Lin et al.

core competence in financial calculation based on grid environment. Grid comput-

ing technology will be used to explain how the core financial calculations can be
significantly accelerated in various distributed and parallel computing environ-
ments. The calculation models in this work can be easily migrated to pure cloud
environments.

49.2.1.2 Essence of IT Technology

To realize the above goal, it needs to handle technically interoperability of middleware
that is capable of communicating between heterogeneous computer systems across
institutional boundaries. The movement of grid began in 1996 by Ian Foster and
Kessleman (2004). Before their development, another branch of high-performance
computing that focuses on connecting geographically distributed supercomputers to
achieve one single grand task had been developed by Smarr and Catlett (1992). They
coined such a methodology as metacomputing and their query has been how we can
have infinite computing power under the physical limit, such as Moore’s Law.
However, it remains to be less useful because its limit goal on pursuing top performance
without noticing practical use in real world. The idea lives on and generates many tools
dedicated to high-performance/high-throughput computing, such as Condor (Litzkow
et al. 1988), Legion (Grimshaw and Wulf 1997), and UNICORE (Almond and Snelling
1999). Condor, as suggested by the name of the project, is devised to scavenge a large
cluster of idle workstations. Legion is closer to the development of worldwide virtual
computer. The goal of UNICORE is even much simpler and practical. It was developed
when Germany government decided to consolidate their five national supercomputer
centers into a virtual one to reduce the management cost and needed a software tool to
integrate them, hence the UNICORE. These tools were successful under their devel-
opment scope. However they fail to meet the first and the second items in Foster’s
checklist in the previous section.
The emergency of grids follows the similar path as that of Condor and Legion at the
first place, in which its development aims at resources sharing in high-performance
computing. However, its vision in open standards and the concept of virtual organiza-
tion allows its development go far beyond merely cluster supercomputers together. It
gives a broader view of resources sharing, in which it is not only limited to the sizable
computing cycles and storage space to be shared but also extended virtually to
calculable machines that are able to hook up to the Internet, such as sensors and sensor
loggers, storage servers, and computers. Since 1996, Foster and his team have been
developing software tools to achieve the purpose. Their software Globus Toolkit
(Foster and Kessleman 2004) is now a de facto middleware for grids. However, the
ambitious development is still considered insufficient to meet the ever-growing com-
plexity of grid systems.
As mentioned earlier that grid is based on open specifications and standards, they
allow all stakeholders within the virtual organization/grid to communicate with each
other with ease and enable ones more to focus on integrated value creation activities.
The open specifications and standards are made by the community of Open Grid
Forum (OGF), which plays as a standard body and made, discussed, and announced
new standards during regular OGF meetings. Grid Specifications and Standards
49 Computer Technology for Financial Service 1349

include architecture, scheduling, resource management, system configuration, data,

data movement, security, and grid security infrastructure. In 2004, OGF announced
Globus Toolkit version, which adopts both the open standard of grid, Open Grid
Services Architecture (OGSA), and the more widely adopted World Wide Web
standard, web services resource framework (WSRF), which ultimately enable grids
to tackle issues of both scalability and complexity of very large grid systems.

49.2.2 Compute Intensive IT Systems

The recent development of computational finance based on grids is hereby scruti-

nized and remarks given. Our major interest is to see if the split-second perfor-
mance is well justified under the grid architecture. Also, real-time issue with real
market parametric data should be used as input for practical simulation. In addition,
issues of intersystem, interdisciplinary and geographically distribution of resources,
and the degree of virtualization are crucial to the success of such a grid. The chosen
projects are reviewed and discussed as follows:
1. PicsouGrid
This is a French grid project for financial service. It provides a general
framework for computation finance and targets on applications of option trading,
option pricing, Monte Carlo simulation, aggregation of statistics, etc. (Stokes-
Rees et al. 2007). The key for this development is the implementation of the
middleware ProActive. ProActive is an in-house Java library for distributed
computing developed by INRIA Sophia Antipolis, France. It provides transpar-
ent asynchronous distributed method calls and is implemented on top of Java
RMI. It is also used in commercial applications. It also provides fault tolerance
mechanism. The architecture is shown in Fig. 49.2, which is very similar to most
of grid applications apart from the software stack used. The option pricing was
tested in an approximately 894 CPUs. The underlying computer systems are
heterogeneous. The system is used for metacomputing. As a result, the system
has to specifically design to orchestrate and to synchronize and re-synchronize
the whole distributed processes for one calculation. Once the grid system
requires synchronization between processes, which implies stronger coupling
of algorithm of interest, the performance will be seriously affected. There is no
software treatment to solve such problems and should be tackled by physical
infrastructure, e.g., optical fiber network with Layer 2 light path.
2. FinGrid
FinGrid stands for Financial Information Grid. Its study includes components
of bootstrapping, sentimental analysis, and multi-scale analysis, which focuses
on information integration and analysis, e.g., data mining. It takes advantage of
the huge collection of numerical and textual data simultaneously to emphasize
the study of societal issues (Amad et al. 2004; Ahmad et al. 2005; Gillam
et al. 2005). The architecture of FinGrid is shown in Fig. 49.3. It is a typical 3-
tier system, in which the first tier facilitates the client in sending a request to one
of the services: Text Processing Service or Time Series Service; the second tier
1350 F.-P. Lin et al.

option pricing request

Worker reserve
MC simulation packet ProActive workers
Sub-
heartbeat monitor Server
MC result

ProActive
Server Sub- ProActive Worker
Client
Server

ProActive
JavaSpace
virtual shared
DB memory (to v3)

Fig. 49.2 Architecture of PicsouGrid for option pricing based on Monte Carlo simulation
(Stokes-Rees et al. 2007)

2.Distribute Slave 0
Main CoG Java SSL
Cluster the tasks
1.Send service Data Provider
at location X
request Text Slave 1 (Reuters News
Processing Globus Triarch SSL
Client Source)
Service
4.Notify the user Slave 2
about the results Time Series “Machines” “Machines”
Service 3.Receive Data Provider (FinGrid) (Reuters)
Slave 3 at location Y
results (Numerical
Data)

Fig. 49.3 The architecture of Financial Information Grid (FinGrid)

facilitates the execution of parallel tasks in the main cluster and is distributed to
a set of slave machines (nodes), and the third tier comprises the connection of the
slave machines to the data providers. This work focuses on small scale and
dedicated grid system. It pumps in real and live numerical and textual data from
say Reuters and performs real-time sophisticated data mining analysis. This is
a good prototype for financial grid. However, it will encounter similar problem
as that of PicsouGrid if it is to scale up. The model is more successful in
automatically combining real data and the analysis.
3. IBM Japan collaborates with life insurance company and adopts PC grids
concept to scavenge more compute cycles (Tanaka 2003):
In this work an integrated risk management system (see Fig. 49.4) is modi-
fied, in which the future scenarios of red circle of Fig. 49.4 are send via grid
middleware to a cluster of PCs. According to the size of the given PCs, the
49 Computer Technology for Financial Service 1351

Current portfolio Present value of portfolio

PV of each asset

Market price of risk

Risk premium adjustment rate

Basic equations
Future scenarios
(Stochastic
differential eqs.)

Scenarios
Default-free
interest rates
(domestic. foreign)
1 2 3 N

Hazard rate Future valuation

Price 1 Price 2 Price 3 Price N

Stock price

Exchange rate
Distribution of future value of
portfolio (individual assets)

Return valuation

Risk valuation

Fig. 49.4 Architecture of Integrated Risk Management System (Tanaka 2003)

number scenarios are then divided in a work balanced manner for each PC. This
is the most typical use of compute intensive grid systems and a good practice for
production system. However, the key issues that discussed in the above two
cases cannot be answered in this study. Similar architecture can also be found in
EGrid (Leto et al. 2005).
4. UK e-Science developed a grid service discovery in the financial market sector
focusing on integration of different knowledge flows (Bell and Ludwig 2005).
From application’s viewpoint, business and technical architecture of financial
service applications may be segmented by product, process, or geographic con-
cerns. Segmented inventories make inter silo reuse difficult. The service integration
model is adopted and a loosely coupled inventory – containing differing explicit
1352 F.-P. Lin et al.

SEDI4G
COMPONENT
REPOSTORY
1
SCR
(Apache JNDI or
MDS)

WEB START CLIENT

2

SEDI4G
SEDI4G CLIENT DISCOVERY SEDI4G MATCHING
4 3
VIEW SCV CONTROL SERVICE SMAS
SERVICE SDCS

5
Ontology

Grid Grid Grid Grid

Service Service Service Service
Description Description Description Description
1 2 3 n

6
Described in

Grid Grid Grid Grid

Service Service Service Service
1 2 3 n

Fig. 49.5 The semantic discovery for Grid Services Architecture (SEDI4G) (Bell and Ludwig
2005)

capability knowledge. Three use cases were specifically chosen in this work to
explore the use of semantic searching:
Use case 1 – Searching for trades executed with a particular counterparty
Use case 2 – Valuing a portfolio of interest rate derivative products
Use case 3 – Valuing an option-based product
The use cases were chosen to provide examples of three distinct patterns of use –
aggregation, standard selection, and multiple selection. The architecture (see
Fig. 49.5) is bound specifically with the user cases. The advantage for grid in this
case is that it can be easily tailored into specific user need to integrate different
applications, which is a crucial strength of using grid.

49.2.3 Data-Intensive IT Systems

Grid in financial services from the perspective of web services towards financial
services industry. The perspective is more on transactional side. Once the bottle-
neck of compute cycle is solved, the data-centric nature will play the key role again.
49 Computer Technology for Financial Service 1353

The knowledge flows back to the customized business logic should provide the best
path for users to access the live data of interest. There is no strong focus of
development on this data-intensive grid system. Even in FinGrid, which claims in
streaming live data for real-time analysis, the data issue remains part of compute
grids. However, the need for dynamic data management is obvious as mentioned in
Amad et al. (2004). Hereby, we like to introduce and implement a dynamic data
management software Ring Buffer Network Bus (RBNB) DataTurbine to serve
such a purpose.
RBNB DataTurbine was used recently to support global environmental obser-
vatory network, which involves linking with ten of thousand of sensors and is able
to obtain the observed data online. It meets grid/cyberinfrastructure
(CI) requirements with regard to data acquisition, instrument management, and
state-of-health monitoring including reliable data capture and transport, persistent
monitoring of numerous data channels, automated processing, event detection and
analysis, integration across heterogeneous resources and systems, real-time tasking
and remote operations, and secure access to system resources. To that end, stream-
ing data middleware provides the framework for application development and
integration.
Use cases of RBNB DataTurbine include adaptive sampling rates, failure detec-
tion and correction, quality assurance, and simple observation (see Tilak
et al. 2007). Real-time data access can be used to generate interest and buy-in
from various stakeholders. Real-time streaming data is a natural model for many
applications in observing systems, in particular event detection and pattern recog-
nition. Many of these applications involve filters over data values, or more gener-
ally, functions over sliding temporal windows. The RBNB DataTurbine
middleware provides a modular, scalable, robust environment while providing
security, configuration management, routing, and data archival services. The
RBNB DataTurbine system acts as an intermediary between dissimilar data mon-
itoring and analysis devices and applications. As shown in Fig. 49.6, a modular
architecture is used, in which a source or “feeder” program is a Java application that
acquires data from an external live data sources and feeds it into the RBNB server.
Additional modules display and manipulate data fetched from the RBNB server.
This allows flexible configuration where RBNB serves as a coupling between
relatively simple and “single purpose” suppliers of data and consumers of data,
both of which are presented a logical grouping of physical data sources. RBNB
supports the modular addition of new sources and sinks with a clear separation of
design, coding, and testing (ref. Fig. 49.6). From the perspective of distributed
systems, the RBNB DataTurbine is a “black box” from which applications and
devices send data and receive data. RBNB DataTurbine handles all data manage-
ment operations between data sources and sinks, including reliable transport,
routing, scheduling, and security. RBNB accomplishes this through the innovative
use of memory and file-based ring buffers combined with flexible network objects.
Ring buffers are a programmer-configurable mixture of memory and disk, allowing
system tuning to meet application-dependent data management requirements. Net-
work bus elements perform data stream multiplexing and routing. These elements
1354 F.-P. Lin et al.

Client
Live Data
DAV
Source
TCP/IP

Client

Live Data Tomcat

Source USB Feeder HTTP

HTTP
Client
TCP/IP

Live Data
Source
API
RBNB
Client
Feeder

Live Data API

API
Source

Client

Plugin

Plugin

Fig. 49.6 RBNB DataTurbine use scenario for collaborative applications

combine to support seamless real-time data archiving and distribution over existing
local and wide area networks. Ring buffers also connect directly to client applica-
tions to provide streaming-related services including data stream subscription,
capture, rewind, and replay. This presents clients with a simple, uniform interface
to real-time and historical (playback) data.

49.3 Distributed and Parallel Financial Simulation

In the previous sections, we address issues of incorporating IT technology for

financial competitiveness and derive that the core lies on the performance of IT
platform, providing the competitors in the market have similar capacity and are
49 Computer Technology for Financial Service 1355

equally informed. Grid technology, as the leading IT development in high-

performance computing, is introduced as the cutting-edge IT platform to meet our
goal. Many companies have adopted similar technology of grids with success as
mentioned in Sect. 49.1. There are also increasing research interests, which result in
the work discussed in Sect. 49.2.3. Better performance, however, cannot be
achieved by merely using a single architecture as observed in the cases of
Sect. 49.2.3. The architecture obviously has to be specifically chosen for the
analysis of interest. Simultaneously, the analysis procedures have to be tailored
into the chosen architecture for performance fine-tune.
In this section, we will introduce and discuss analysis procedures of financial
simulation and how to tailor the analysis procedures into grid architectures by
distribution and parallelism. The popular calculations for option pricing and for
value at risk (VaR) in trading practice are used to serve the purpose. The calculation
is based on Monte Carlo simulation, which is chosen not only because it is a well-
received approach due to the absence of straightforward closed-form solutions for
many financial models but also a numerical method intrinsically suited to mass
distribution and mass parallelism. The success of Monte Carlo simulation lies on
the quality of random number generator, which will be discussed in details at the
end of the section.

49.3.1 Financial Simulation

There are wide variety of sophisticated financial models developed, to name a few,
ranging from analysis in time series, fractals, nonlinear dynamics, and agent-based
modeling to applications in optional pricing, portfolio management, and market risk
measure, etc. (Schmidt 2005), in which option pricing and VaR calculations of
market risk measure can be considered crucial and one of the most practiced
activities in market trading.

49.3.1.1 Option Pricing

An option is an agreement between two parties to buy or sell an asset at a certain
time in the future for a certain price. There are two types of options:
Call Option: A call option is a contract that gives the right to its holder (i.e., buyer)
without creating an obligation, to buy a prespecified underlying asset at
a predetermined price. Usually this right is created for a specific time period,
e.g., 6 months or more. If the option can be exercised only at its expiration (i.e.,
the underlying asset can be purchased only at the end of the life of the option),
the option is referred to as a European-style Call Option (or European Call). If it
can be exercised any date before its maturity, the option is referred to as an
American-style Call Option (or American Call).
Put Option: A put option is a contract that gives its holder the right without creating
the obligation to sell a prespecified underlying asset at a predetermined price. If
the option can be exercised only at its expiration (i.e., the underlying asset can be
sold only at the end of the life of the option), the option is referred to as
1356 F.-P. Lin et al.

a European-style Put Option (or European Put). If it can be exercised any date
before its maturity, the option is referred to as an American-style Put Option
(or American Put).
To price options in computational finance, we use the following notation: K is
the strike price; T is the time to maturity of the option; St is the stock price at time t;
r is the risk-free interest rate; m is the drift rate of the underlying asset (a measure of
the average rate of growth of the asset price); s is the volatility of the stock; and
V denotes the option value. Here is an example to illustrate the concept of option
pricing. Suppose an investor enters into a call option contract to buy a stock at price
K after 3 months. After 3 months, the stock price is St. If St > K then one
can exercise one’s option by buying the stock at price K and by immediately
selling in the market to make a profit of ST
K. On the other hand, if he
ST
K to be buy the stock. Hence, we see that a call option to buy the stock at
time T at price K will get payoff (ST
K)+, where (ST
K)+  max(ST
K, 0)
(Schmidt 2005; Hull 2003).

49.3.1.2 Market Risk Measurement Based on VaR

Market risks are the prospect of financial losses or gains, due to unexpected changes
in market prices and rates. Evaluating the exposure to such risks is nowadays of
primary concern to risk managers in financial institutions. Until the late 1980s
market risk was estimated through gap and duration analysis (interest rates),
portfolio theory (securities), sensitivity analysis (derivatives), or scenarios analysis.
However, all these methods could be either applied only to very specific assets or
relied on subjective reasoning.
Since the early 1990s a commonly used market risk estimation methodology has
been the value at risk (VaR). A VaR measure is the highest possible loss L incurred
from holding the current portfolio over a certain period of time at a given confi-
dence level (Dowd 2002):

PðL > VaRÞ  1
c (49.1)

where c is the confidence level, typically 95 %, 97.5 %, or 99 %, and P is

cumulative distribution function. By convention, L ¼
DX(t), where DX(t) is the
relative change (return) in portfolio value over the time horizon t. Hence, large
values of L correspond to large losses (or large negative returns).
The VaR figure has two important characteristics: (1) it provides a common
consistent measure of risk across different positions and risk factors and (2) it takes
into account the correlations or dependencies between different risk factors.
Because of its intuitive appeal and simplicity, it is no surprise that in a few years
value at risk has become the standard risk measure used around the world.
However, VaR has a few deficiencies, among them the non-subadditivity –
a sum of VaR’s two portfolios can be smaller than the VaR of the combined
portfolio. To cope with these shortcomings, Artzner et al. proposed an alternative
measure that satisfies the assumptions of a coherent risk measure. The expected
49 Computer Technology for Financial Service 1357

shortfall (ES), also called expected tail loss (ETL) or conditional VaR, is the
expected value of the losses in excess of VaR:

ES ¼ EðLjL > VaRÞ (49.2)

It is interesting to note that although new to the finance industry – expected

shortfall has been familiar to insurance practitioners for a long time. It is very
similar to the mean excess function which is used to characterize claim size
distribution; see (Cizek et al. 2011).
The essence of the VaR and ES computations is estimation of low quantiles in the
portfolio return distributions. Hence, the performance of market risk measurement
methods depends on the quality of distribution assumptions on the underlying risk
factors. Many of the concepts in theoretical and empirical finance developed over
the past decades, including the classical portfolio theory, the Black-Scholes-Merton
option pricing model, and even the RiskMetrics variance-covariance approach to
VaR rest upon the assumption that asset returns follow a normal distribution. The
assumption is not justified by real market data. Our interest is more on the calcula-
tion side. For interested readers we refer further to (Weron 2004).

49.3.2 Monte Carlo Simulations

49.3.2.1 Monte Carlo and Quasi-Monte Carlo Methods

In general, Monte Carlo (MC) and quasi-Monte Carlo (QMC) methods are applied
to estimate the integral of function f(x) over [0, 1]d unit hypercube where d is the
dimension of the hypercube:
ð
I¼ f ðxÞdx (49.3)
½0;1d

In MC methods, I is estimated by evaluating f(x) at N independent points

randomly chosen from a uniform random distribution over [0, 1]d and then evalu-
ating average

XN
^I ¼ 1 f ðxi Þ (49.4)
N i¼1

From the law of large numbers, ^I ! I as N ! 1. The standard deviation is

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u 1 X N
t ðf ðx i Þ
I Þ2 (49.5)
N
1 i¼1

Therefore, the error of MC methods is proportional to N
1/2.

1358 F.-P. Lin et al.

QMC methods compute the above integral based on low-discrepancy

(LD) sequences. The elements in a LD sequence are “uniformly” chosen from [0, 1]d
rather than “randomly.” The discrepancy is a measure to evaluate the uniformity of
points over [0, 1]d. Let {qn} be a sequence in [0, 1]d; the discrepancy D*N of qn is defined
as follows, using Niederreiter’s notation (Niederreiter 1992):
 
sup AðB; qn Þ 
DN ðqn Þ ¼ 
d
vd ðBÞ (49.6)
B 2 ½0; 1Þ N

where B is a subcube of [0, 1]d containing the origin, A(B, qn) is the number of
points in qn that fall into B, and Vd(B) is the d-dimensional Lebesgue measure of B.
The elements of qn are said uniformly distributed if its discrepancy D*N ! 0 as
N ! 1. From the theory of uniform distribution sequences (Kuipers and
Niederreiter 1974), the estimate of the integral using a uniformly distributed
XN
sequence {qn} is ^I ¼ N1 f ðqn Þ, as N ! 1 then ^I ! I . The integration error
n¼1
bound is given by the Koksman-Hlawka inequality:
 
 1X N 
 
I
f ðqn Þ  V ðf ÞDN ðqn Þ (49.7)
 N n¼1 

where V(f) is the variation of the function in the sense of Hardy and Krause (see
Kuipers and Niederreiter 1974), which is assumed to be finite.
The inequality suggests a smaller error can be obtained by using sequences
with smaller discrepancy. The discrepancy of many uniformly distributed
sequences satisfies O((log N)d/N). These sequences are called low-discrepancy
(LD) sequences (Chen et al. 2006). Inequality (49.7) shows that the estimates
using a LD sequence satisfy the deterministic error bound O((log N)d/N).

49.3.2.2 Monte Carlo Simulations for Option Pricing

Under the risk-neutral measure, the price of a fairly valued European call option is
the expectation of the payoff E[e
rT(ST
K)+]. In order to compute the expecta-
tion, Black and Scholes (1973) modeled the stochastic process generating the price
of a non-dividend-paying stock as geometric Brownian motion:

dSt ¼ mSt dt þ sSt dW t (49.8)

where W is a standard Wiener process, also known as Brownian motion. Under the
risk-neutral measure, the drift m is set to m ¼ r.
To simulate the path followed by S, suppose the life of the option has been
divided into n short intervals of length Dt(Dt ¼ T/n), the updating of the stock price
at t + Dt from t is (Hull 2003):
pffiffiffiffiffi
StþDt
St ¼ rSt Dt þ sSt Z Dt (49.9)
49 Computer Technology for Financial Service 1359

where Z is a standard random variable, i.e., Z(0,1). This enables the value of SDt to
be calculated from initial value St at time Dt, the value at time 2Dt to be calculated
from SDt, and so on. Hence, a completed path for S has been constructed.
In practice, in order to avoid discretization errors, it is usual to simulate lnS
rather than S. From It ^o’s lemma, the process followed by of Eq. 49.9 is (Bratley and
Fox 1988)
 
s2
dlnS ¼ r
dt þ sdz (49.10)
2

so that
 
s2 pffiffiffiffiffi
lnStþDt
lnSt ¼ r
dt þ sZ Dt (49.11)
2

or equivalently
  
s2 pffiffiffiffiffi
StþDt ¼ St exp r
dt þ sZ Dt (49.12)
2

Substituting independent samples Zi,   , Zn from the normal distribution into

(Eq. 49.12) yields independent samples ST(i), i ¼ 1,   , n, of the stock price at
expiry time T. Hence, the option value is given by

1X n
1X n h i
V¼ Vi ¼ e
rT max ST ðiÞ
K, 0 (49.13)
n i¼1 n i¼1

The QMC simulations follow the same steps as the MC simulations, except that
the pseudorandom numbers are replaced by LD sequences. The basic LD sequences
known in literature are Halton (1960), Sobol (1967), and Faure (1982). Niederreiter
(1992) proposed a general principle of generating LD sequences. In finance, several
examples have shown that the Sobol sequence is superior to others. For example,
Galanti and Jung (1997) observed that the Sobol sequence outperforms the Faure
sequence, and the Faure marginally outperforms the Halton sequence. In this
research, we use Sobol sequence in our experiments. The generator used for
generating the Sobol sequence comes from the modified algorithm 659 of Joe and
Kuo (2003).

49.3.2.3 Monte Carlo Bootstrap for VaR

Monte Carlo simulation is applicable with virtually any model of changes in risk
factors and any mechanism for determining a portfolio’s value in each market
scenario. But revaluing a portfolio in each scenario can present a substantial
computational burden, and this motivates research into ways of improving the
efficiency of Monte Carlo methods for VaR.
1360 F.-P. Lin et al.

The bootstrap (Efon 1981; Efron and Tibshirani 1986) is a simple and straight-
forward method for calculating approximated biases, standard deviations, confi-
dence intervals, and so forth, in almost any nonparametric estimation problem.
Method is a keyword here, since little is known about the bootstrap’s theoretical
basis, except that (a) it is closely related to the jackknife in statistic inferring;
(b) under reasonable condition, it gives asymptotically correct results; and (c) for
some simple problems which can be analyzed completely, for example, ordinary
linear regression, the bootstrap automatically produces standard solutions.
The bootstrap method is straightforward. Suppose we observe returns Xi ¼ xi,
i ¼ 1, 2,   , n, where the Xi are independent and identically distributed (iid) according
to some unknown probability distribution F. The Xi may be real valued and
two-dimensional or take values in a more complicated space. A given parameter y(F),
perhaps the mean, median, correlation, and so forth, is to be estimated, and we agree
to use the estimate ^y ¼ y F ^ , where F^ is the empirical distribution function putting
mass 1/n at each observed value xi. We wish to assign some measure of accuracy to ^y.
Let s(F) be some measure
 of accuracy that we would use if F were known, for
example, sðFÞ ¼ SDF ^y , the standard deviation ^
 of y when X1, X2,   , Xn  F .
(idd)

The bootstrap estimate of accuracy s ^¼s F ^ is the nonparametric maximum

likelihood estimate of s(F). In order to calculate s ^ it is usually necessary to employ
numerical methods. (a) A bootstrap sample X*1, X*2,   , X*n is drawn from F, ^ in which
each Xi independently takes value xj with probability 1/n, j ¼ 1, 2,   , n. In
*

other words, X*1, X*2,   , X*n is an independent sample of size n drawn with replace-
ment from the set of observations {x1,x2,   , xn}. (b) This gives a bootstrap empirical
distribution function F ^  , the empirical distribution

 of the n values X*1, X*2,   , X*n,

and a corresponding bootstrap value ^y ¼ y F ^ . (c) Steps (a) and (b) are
repeated, independently, in a large number of times, say N, giving bootstrap values
^ 1 , y
y ^ 2 ,   , y
^ N . (d) The value of s ^ is approximated, in the case where s(F) is the

standard deviation by the sample standard deviation of the ^y values, where
Xn j
^¼
m ^y
j¼1
(49.14)
N
and
Xn
j 2
^y
m
^
j¼1
^2 ¼
s (49.15)
N
1

49.3.3 Distribution and Parallelism Based on Random Number

Generation

Financial variables, such as prices and returns, are random time-dependent

variables. Wiener process plays the central role in modeling. As shown in
49 Computer Technology for Financial Service 1361

Eqs. 49.8 and 49.9 for approximating the underlying prices St+Dt, or the boot-
strap samples of return X*i , the solution methods involve basic market parame-
ters, drift m, volatility s, and risk-free interest rate r, current underlying price
S or return X, strike price K, and Wiener process, which isprelatedffiffiffiffiffi to time to
maturity Dt and standard random variable Z, i.e., DW ¼ Z Dt . Monte Carlo
methods simulate this nature of the Brownian motion directly. It follows Wiener
process and approximates the standard random variable Z by introducing pseudo
iid random number into each Wiener process. When the simulation number is
large enough, e.g., if n in Eq. 49.13 is large enough, the mean value will
approach the exact solution. The large number for n also implies the perfor-
mance problems are the key problems for Monte Carlo methods. One the other
hand, the iid property of the random number Z shows possible solution to tackle
the performance problem through mass distribution and/or parallelism. The
solution method centers on the random number generation.
The techniques of random number generation can be developed in a simple form
through the approximation of a d-dimensional integral, e.g., (Eq. 49.3). Mass
distribution and parallelism required solutions of for large dimension. However,
most modern techniques in random number generation have limitations. In this
study, both tradition pseudorandom number generation and high-dimensional
low-discrepancy random number generator are considered.
Following Sect. 49.3.2.1 better solution can be achieved by making use of Sobol
sequences, which were proposed by Sobol (1967). A computer implementation
in Fortran 77 was subsequently given by Bratley and Fox (1988) as Algorithm 659.
Other implementations are available as C, Fortran 77, or Fortran 90 routines in the
popular Numerical Recipes collection of software. However, as given, all these
implementations have a fairly heavy restriction on the maximum value of
d allowed. For Algorithm 659, Sobol sequences may be generated to approximate
integrals in up to 40 dimensions, while the Numerical Recipes routines allow the
generation of Sobol sequences to approximate integrals in up to six dimensions
only. The FinDer software of Paskov and Traub (1995) provides an implementation
of Sobol sequences up to 370 dimensions, but it is licensed software. As computers
become more powerful, there is an expectation that it should be possible to
approximate integrals in higher and higher dimensions. Integrals in hundreds of
variables arise in applications such as mathematical finance (e.g., see Paskov and
Traub (1995)). Also, as new methods become available for these integrals, one
might wish to compare these new methods with Sobol sequences. Thus, it would be
desirable to extend these existing implementations such as Algorithm 659 so they
may be used for higher-dimensional integrals. We remark that Sobol sequences are
now considered to be examples of (t, d)-sequences in base 2. The general theory of
these low-discrepancy (t, d)-sequences in base b is discussed in detail in
Niederreiter (1992). The generation of Sobol sequences is clearly explained in
Bratley and Fox (1988). We review the main points so as to show what extra data
would be required to allow Algorithm 659 to generate Sobol sequences to approx-
imate integrals in more than 40 dimensions. To generate the j th component of the
1362 F.-P. Lin et al.

points in a Sobol sequence, we need to choose a primitive polynomial of some

degree sj in the field ℤ2, that is, a polynomial of the form

xsj þ a1, j xsj
1 þ    þ asj
1, j x þ 1, (49.16)

where the coefficients a1, j _  _asj
1, j are either 0 or 1.

We use these coefficients to define a sequence {m1, j, m2, j,   } of positive
integers by the recurrence relation

mk, j ¼ a1, j mk
1, j 22 a2, j mk
2, j   

(49.17)
2sj
1 asj
1, j mk
sj þ1, j 2sj asj , j mk
sj , j mk
sj , j

for k
sj + 1, where is the bit-by-bit exclusive-OR operator. The initial values
m1, j , m2, j ,   , msj , j can be chosen freely provided that each Mk,j, 1  k  sj is odd
and less than 2k. The “direction numbers” {v1,j,v2,j,   } are defined by

mk, j
v1, j  (49.18)
2k

Then xi,j, the j th component of the ith point in a Sobol sequence, is given by

x i , j ¼ b1 v 1 , j b2 v 2 , j    (49.19)

Where bl is the lth bit from the right when i is written in binary, that is, (  b2b1)2
is the binary representation of i. In practice, a more efficient Gray code implemen-
tation proposed by Antonov and Saleev (1979) is used; see Bratley and Fox (1988)
for details. We then see that the implementation in Bratley and Fox (1988) may be
used to generate Sobol sequences to approximate integrals in more than 40 dimen-
sions by providing more data in the form of primitive polynomials and direction
numbers (or equivalently, values of m1, j , m2, j ,   , msj , j ). When generating such
Sobol sequences, we need to ensure that the primitive polynomials used to generate
each component are different and that the initial values of the mk,j’s are chosen
differently for any two primitive polynomials of the same degree. The error bounds
for Sobol sequences given in Sobol (1967) indicate we should use primitive poly-
nomials of as low a degree as possible. We discuss how additional primitive
polynomials may be obtained in the next section. After these primitive polynomials
have been found, we need to decide upon the initial values of the mk,j for 1ksj.
As explained above, all we require is that they be odd and that mk,j <2k. Thus, we
could just choose them randomly, subject to these two constraints. However, Sobol
and Levitan (1976) introduced an extra uniformity condition known as Property A.
Geometrically, if the cube [0, 1]d is divided up by the planes xj ¼ 1/2 into 2d equally
sized subcubes, then a sequence of points belonging to [0, 1]d possesses Property
A if, after dividing the sequence into consecutive blocks of 2d points, each one of
49 Computer Technology for Financial Service 1363

the points in any block belongs to a different subcube. Property A is not that useful
to have for large d because of the computational time required to approximate an
integral using 2d points. Also, Property A is not enough to ensure that there are no
bad correlations between pairs of dimensions. Nevertheless, Property A would
seem a reasonable criterion to use in deciding upon a choice of the initial mk,j.
The numerical results for Sobol sequences given in Sect. 49.4 suggest that the
direction numbers obtained here are indeed reasonable. Other ways of obtaining the
direction numbers are also possible. For example, in Cheng and Druzdzel (2000),
the initial direction numbers are obtained by an interesting technique of minimizing
a measure of uniformity in two dimensions. This technique may alleviate the
problem of bad correlations between pairs of dimensions that was mentioned
above. Sobol (1967) showed that a Sobol sequence used to approximate
a d-dimensional integral possesses Property A if and only if

detðV d Þ ¼ 1ðmod2Þ, (49.20)

where Vd is the d  d binary matrix defined by

2 3
v1, 1, 1 v2, 1, 1 vd , 1, 1
6 v1, 2, 1 v2, 2, 1 
vd , 2, 1 7
Vd ¼ 6
4
7 (49.21)
⋮ ⋱ ⋮ 5
v1, d, 1 v2, d, 1  vd, d, 1

With vk,j,1 denoting the first bit after the binary point of vk,j,. The primitive
polynomials and direction numbers used in Algorithm 659 are taken from Sobol
and Levitan (1976), and a subset of this data may be found in Sobol (1967). Though
it is mentioned in Sobol (1967) that Property A is satisfied for d  16, that is,
det(Vd) ¼ 1 (mod 2) for all d  16, our calculations showed that Property A is
actually satisfied for d  16. As a result, we change the values of the mk,j for 21 
j  40, but keep the primitive polynomials. For j
41, we obtain additional
primitive polynomials. The number of primitive polynomials of degree
s is f(2s
1)/s, where f is Euler’s totient function. Including the special case for
j ¼ 1 when all the Mk,j are 1, this allows us to approximate integrals in up to
dimension d ¼ 1,111 if we use all the primitive polynomials of degree 13 or less.
We then choose values of the Mk,j so that we can generate Sobol sequences
satisfying Property A in dimensions d up to 1,111. This is done by generating
some values randomly, but these are subsequently modified so that the condition
det(Vd) ¼ 1 (mod 2) is satisfied for all d up to 1,111. This process involves
evaluating values of the vk,j,1’s to obtain the matrix Vd and then evaluating the
determinant of Vd. A more detailed discussion of this strategy is given in the next
section. It is not difficult to produce values to generate Sobol’s points for approx-
imating integrals in even higher dimensions.
The following figures are the two-dimensional plots of high-dimensional Sobol
sequences of Joe and Kuo with d ¼ 1,000. It is compared with pseudorandom number
1364 F.-P. Lin et al.

Comparison of Randomness
Pseudo Random Number
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Comparison of Randomness
Sobol Seq dimension 1:2
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 49.7 Pseudorandom number plot comparing with quasi-random number of Sobol for
dimensions 1 and 2

generation. The number of sampling points is 3,000. In Fig. 49.7 pseudorandom

number is plotted in comparison with that of quasi-random number of Sobol. The
leading dimensions 1 and 2 of Sobol sequences are used. The improvement is
immense. In order to understand more of the nature of Sobol sequences, we chose
prime dimensional numbers 499, 503, 991, and 997, respectively, as suggested by Joe
49 Computer Technology for Financial Service 1365

Comparison of Randomness
Sobol Seq dimension 498:501
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Comparison of Randomness
Sobol Seq dimension 990-997
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 49.8 Comparison of adjacent dimensions in quasi-random number Sobol sequence. The
dimensions are chosen according to prime numbers. There is a high discrepancy found in higher
dimensions of Sobol sequence modified by Joe and Kuo (2003)

and Kuo. The results are plotted in Figs. 49.8 and 49.9. It is found that there are
stronger correlations between Sobol sequences of nonadjacent dimensions in the
fashion of the dimensional comparison of their randomness. Larger numbers of
sampling points, e.g., 10,000, are also tested and the patterns persist. It implied the
1366 F.-P. Lin et al.

Comparison of Randomness
Sobol Seq dimension 2:498
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Comparison of Randomness
Sobol Seq dimension 2:990
1

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 49.9 Comparison of nonadjacent dimensions. High discrepancy is found in their correlations
and forms clusters of islands in the distribution

violation of idd assumption and may incur problems in mass distribution and paral-
lelism, in which each process the random number is generated independently without
knowing what other processes are doing. The dependency may deteriorate the quality
of randomness. Nevertheless, in our numerical experiments, there are no significant
differences found thus far (Figs. 49.10 and 49.11).
49 Computer Technology for Financial Service 1367

1.2
Distribution in One Vector

0.8
Probability

0.6

0.4

0.2

0
0 100 200 300 400 500 600 700 800 900 1000
Number of Coordinates

Fig. 49.10 The distribution of probability in directional vector vi,j of Sobol sequences at
i ¼ 3,000 with j ∈ {1,   ,1000}. The mean of the distribution is 0.491357, which approaches
the mean of the normal distribution 0.5

49.4 Case Study and Discussions

49.4.1 Case Study

49.4.1.1 Asian Options and Rainbow Options

To demonstrate what the grid computing can contribute to the financial service in
a significant manner, two kinds of popular options, Asian options and rainbow
options, are chosen for Monte Carlo pricing model. Asian options have payoffs that
depend on the average price of the underlying asset such as stocks, commodities, or
financial indices. However, there is no exact closed-form formula existed for these
popular options. Rainbow options, also known as basket options, are referred to as
an entire class of options which consist of more than one underlying asset. Rainbow
options usually call or put on the best or worst of the underlying assets, or options
which pay the best or worst of the assets. They are excellent tools for hedging risk
of multiple assets. The rainbow options are therefore used for our bootstrap
calculations of VaR.

49.4.1.2 Parallelization, Distribution, and Message Passing

Interface (MPI)
MPI is a library specification for message passing, proposed and developed as
a standard by a broadly based committee of vendors, developers, and users (Snir
et al. 1996). MPI was designed for high performance on both massively parallel
1368 F.-P. Lin et al.

Convergence History of Mean Value of Asian Option Pricing

14.5
Psudo RNG
14 Quasi RNG

13.5
Mean of Asian Option Pricing

13

12.5

12

11.5

11

10.5

10
0 200 400 600 800 1000 1200 1400 1600
Number of Simulations

Convergence History of Mean Value of Asian Option Pricing

14.5
QRNG dim 1
QRNG dim 997
14

13.5
Mean of Asian Option Pricing

13

12.5

12

11.5

11

10.5

10
0 200 400 600 800 1000 1200 1400 1600
Number of Simulations

Fig. 49.11 The convergence history of the mean value of Asian option pricing with risk-free
interest rate r ¼ 0.1, underlying asset spot price S ¼ 100, strike price X ¼ 100, duration to maturity
T ¼ 1, and volatility s ¼ 0.3: The comparison is based on a single dimension of the extended high-
dimensional Sobol sequences. The quasi-random number generator (QRNG) outperforms pseu-
dorandom number generator. The test also is conducted to compare the convergence history
between different dimensions in Sobol sequences and found that all perform consistently as
shown in the right figure, in which the low dimension and high dimension are chosen for the
comparison
49 Computer Technology for Financial Service 1369

machines and on distributed clusters. The MPI standard is nowadays widely

accepted and used in the community of high-performance computing.
The basic MPI functions are point-to-point pair-wise message passing for send
and for receive. Collective communications are also provided for ease of use as well
as better performance. These communication methods, when used in supercom-
puters, do facilitate the parallelization of numerical methods that require both heavy
compute cycles and stronger dependency between parallel processes.
Recent development of supercomputer, affected by the popularity of cluster
computing in PCs, tends to be designed hierarchically scalable. Further extension
of clusters of supercomputers can be regarded as initial concept of grids (see Sect.
49.2). The use of MPI is straightforward in this kind of hardware architecture and
interlink of networks. There is always an obvious physical limitation in this
architecture, which is also proportional to the limitation of the investment of
governmental research funding. People tend to use mass distribution of computers,
mostly PCs, which linked loosely in the Internet cloud. The terminology cloud is
often used in networking community to show that in the Internet there is no specific
network path from one computer to another. MPI working in such an environment
is expected to be inefficient and unstable, e.g., high network latency induced packet
lost in long-distance real-time communication. In the following sections three
specific platforms, including local clusters, geographically distributed large clus-
ters, and PC grids with ten of thousand of PCs connected in the cloud, will be used
for the financial calculations to demonstrate benefits in using grids.

49.4.1.3 Empirical Study for Data Grid System

In order to demonstrate the usefulness of grid system, in particular in data-intensive
application, the real market data are used, including daily from iShares (Morgan
Stanley Capital International) MSCI Taiwan Index (ETF) and Taiwan Stock
Exchange Center (TSEC) weighted index, extracted specifically from May
31, 2005, to May 31, 2008, and 30 days tick-by-tick trading data from Taiwan
Futures Exchange Center (TAIFEX).

49.4.2 Grid Platforms Tests

The various grid platforms are carefully chosen to demonstrate the performance
issue in finance services, which include a small diskless remote boot Linux (DBRL)
PC clusters, large-scale and geographically widely distributed test-bed the Pacific
Rim Applications and Grid Middleware Assembly (PRAGMA) compute grid, and
a densely distributed at-home style PC grid, which resembles the clouding
computing.

49.4.2.1 Diskless Remote Boot Linux (DRBL) Cluster

DRBL is an in-house program of National Center for High-Performance Computing
(NCHC) and was an original software product developed by Steven Shiau and his
group under the auspice of National Knowledge Innovation Grid (KING) of
1370 F.-P. Lin et al.

OS Image
00101010
01101011
111000

Duplicate OS Images
OS Image OS Image OS Image
OS Image OS Image
00101010 00101010 00101010
00101010 00101010
01101011 01101011 01101011
01101011 01101011
111000 111000 111000
111000 111000

Fig. 49.12 The schematic of DRBL system: DRBL duplicates image files of operational system
via network to the clients, in which the clients’ original operational systems are untouched.
Therefore, the clients are temporarily turned into dedicated compute resources, which also provide
additional security to financial data

Taiwan. It was developed initially as a centralized system management tool aiming

at small and median size of PC clusters. It is nowadays recognized internationally as
one of the most advanced mass backup solutions. Here DRBL is used as an
alternative scavenger for compute cycles of spare clusters. When needed, it con-
verts systems of PC clusters into an aggregated and homogenous Linux system,
simultaneously with a mass backup of the original systems, and recovered back the
original systems once the need was satisfied. The most popular use is to convert
a PC classroom into a compute Linux cluster. In such a case, compute cycles of the
clusters can be fully exploited. In a grid environment, this is a perfect case to
resources scaleup when in contingent need and once the situation relieved resources
will be released correspondingly. Such a dynamic feature can be beneficial for the
financial services.
The schematic of DRBL system can be shown in Fig. 49.12, where DRBL
duplicates image files of an operational system, e.g., Linux kernel, via network to
49 Computer Technology for Financial Service 1371

Table 49.1 Comparison of performance between DRBL-based PC platform with 32 nodes,

FORMOSA II of NCHC with a batch job of 32 nodes and IBM Cluster 1350 with a batch job of
32 nodes. The PCs are 20 XEON 2.6 GHz and 4GB RAM
DRBL cluster FORMOSA II IBM cluster 1350
Asian option pricing (AOP) 81 49 24
Rainbow option pricing (ROP) 329 197 98
VaR calculation (based on ROP 322 194 97
with bootstrap)
Unit: micro-secs per Monte Carlo path. Averaged from 1000,000 Monte Carlo paths

Table 49.2 Comparison of speedup ratios based on the calculations in Table 49.1
DRBL cluster FORMOSA II IBM cluster 1350
Asian option pricing (AOP) 28.68 29.46 30.16
Rainbow option pricing (ROP) 27.31 28.59 29.34
VaR calculation (based on ROP 27.20 28.10 29.10
with bootstrap)
Unit: speedup ratio: CPU(nonparallel single node)/CPU(parallel single node)

the clients. The clients’ original operational systems are not used. The clients are
therefore temporarily turned into dedicated compute resources, which also provide
additional security to financial data.
The test case here involves Monte Carlo simulations on Asian option pricing,
on rainbow option pricing and on a bootstrap VaR calculation, respectively
(see Sect. 49.3.2). The market parameters are given as risk-free interest rate
r ¼ 0.1, underlying asset spot price S ¼ 100, strike price X ¼ 100, duration
to maturity T ¼ 125, and volatility s ¼ 0.3. For the rainbow options a linear
weighted combination of 4 underlying assets is assumed with underlying prices of
Si ¼ 100, 110, 120, and 130 and the corresponding weightings of volatility si ¼ 0.3,
0.4, 0.5, and 0.6. The correlation matrix is taken to be
0 1
0:5 0:4 0:5
rij ¼ @ 0:4 0:3 0:4 A
0:5 0:4 0:6

The calculation of the VaR uses the same 4-dimensional rainbow options with
additional expectation of return 0.07, 0.08, 0.09, and 0.10, respectively. They are
calculated in DRBL cluster as well as benchmarked with two cluster-based super-
computers in NCHC. The results are shown in Tables 49.1 and 49.2.
In Table 49.1, instead of giving a total wall clock time of the calculation with
some given numbers of Monte Carlo simulations or paths, a more useful averaged
single Monte Carlo simulation based on 1,000,000 simulations is used to demon-
strate the performance when different system architectures are used. The results
show that the traditional big irons, i.e., supercomputers, still outperform the cluster.
1372 F.-P. Lin et al.

Yet, considering there is no extra cost invested in the computing resources and still
obtains compute cycle in a sizable manner, the approach is appealing to be further
developed in to a fully operational production system.

49.4.2.2 Pacific Rim Applications and Grid Middleware Assembly

(PRAGMA) Grid
The Pacific Rim Applications and Grid Middleware Assembly (PRAGMA),
founded in 2002, is an open international organization that focuses on a variety of
practical issues of building international scientific collaborations in a number of
application areas. PRAGMA grid is established by the resources and data comput-
ing working group as a global grid test-bed for benchmarking the interoperability of
grid middleware and the usability and productivity of grids. The PRAGMA grid
consists of physical resources as well as system administration supports from
29 institutions across 5 continents and 14 countries. It is an instantiation of
a useful, interoperable, and consistently available grid system that is neither
dictated by the needs of a single science domain nor funded by a single national
agency. It does not have uniform yet robust infrastructure management and sup-
ports a wide range of scientific applications. The software stack of the system is
shown in (Fig. 49.13).
The PRAGMA grid successfully tackles the issues of distance and time
zone differences among sites, lack of infrastructure tools for heterogeneous global
grid, nonuniform system and network environments, and diverse application
requirements. For more technical details both in theory and practice we refer to
(Abramson et al. 2006)
Following the similar test case in Sect. 49.4.2.2, but extended the platform with
a collection of clusters across institute boundaries, the common job submission is
executed via a homogeneous middleware Globus Toolkit. We demonstrate
the usefulness of the platform by grouping compute resources across
national boundaries and still achieve good performance. The results are shown in
Tables 49.3 and 49.4.

49.4.2.3 At-Home Style PC Grid

With the continued penetration of personal computers and the remarkable improve-
ment of CPU processing speed, 80–90 % of most PCs’ processing power is
untapped, according to a study. This does not mean that many PCs remain turned
off, but that the capacity of the CPU, the brain of the PC, is not fully utilized. In case
the CPU is more extensively used when a task requiring an enormous number of
operations, such as three-dimensional graphical processing, is assigned, it sits idle
most of the time during word processing and Internet browsing because CPU
processing speeds are much faster than the speeds of input from the keyboard or
the communication line.
This fact led to the idea of virtually gathering the power of idle CPUs to use as
a computer resource. In other words, this means networking numerous computers to
make them work like a single high-performance computer and assigning complex
processing tasks to it. The assigned task will be divided into a myriad of small tasks
49 Computer Technology for Financial Service 1373

Applications
TDDFT QM/MD CCAM mpiBLAST iGAP ...

Ninf-G Nimrod/G Mpich-g2 Gfarm SCMSWeb MOGAS

Application Middleware Infrastructure Middleware

Globus (required)

SGE PBS LSF SQMS ...

Local job scheduler (require one)

Fig. 49.13 Software stack developed in the PRAGMA grid

Table 49.3 Comparison of performance between Group A, which consists of 13 nodes from
NCHC and 15 from UCSD, and Group B, which consists of 122 nodes collectively from UCSD,
AIST, NCHC, and Osaka University. The details of resources are referred to (http://pragma-goc.
rocksclusters.org/pragma-doc/resources.html)
Group A Group B
Asian option pricing (AOP) 68 56
Rainbow option pricing (ROP) 278 234
VaR calculation (based on ROP with bootstrap) 271 229
Unit: seconds/per Monte Carlo path. Averaging from 1000,000 Monte Carlo paths

Table 49.4 Comparison of speedup ratios based on the calculations in Table 49.3
Group A Group B
Asian option pricing (AOP) 25.24 107.58
Rainbow option pricing (ROP) 25.81 110.34
VaR calculation (based on ROP with bootstrap) 25.63 110.09
Unit: seconds/per Monte Carlo path. CPU(nonparallel single node)/CPU(parallel single node)

and allocated to individual computers on the grid-like network. Even with increas-
ingly faster CPUs, the power of PCs is not comparable to those of supercomputers,
but in a networked environment where individual PCs simply process complex task
in parallel, PCs can deliver surprisingly high performance. This is the core concept
of CP grid computing, and it came into a reality several years ago (Table 49.5).
The commoditization and the increased processing speed of PCs lead the growth
of idle CPU power. This will facilitate the construction of PC grid computing
system along with improvement of communication environment by broadband
connectivity (Chen et al. 2006). In practice, a PC grid platform Korea@Home
(or K@H) is used in our study. Its architecture is shown in Fig. 49.14. It is based on
MS Windows. Asian option pricing is used to demonstrate the performance of the
1374 F.-P. Lin et al.

Table 49.5 Summary of the case for the PC grid calculations

Asian option pricing Statistics
Number of Monte Carlo path 1,000,000  10.000
The running period (1) (wall clock time) 28 h 51 m (104,911 s)
Number of jobs 10,000
CPU time per job 28  30 s
Total running time (2) (wall clock time) 4 days 14 h 31 m (397,919 s)
Speedup ratio ðð21ÞÞ 3.79

Secure/Intelligent
P2P Agent Technology Agent Agent
Cooperation

Cooperation Cooperation

Distribution Distribution Agent

Agent

Distribution
Distribution
Internet

Subscription

Distribution
job request

result
transmission [6T Application Filed]
[Resource
[Large-scale Application] Application Provider
Management] Korea@Home Server

Fig. 49.14 The architecture of Korea@Home, a specific @Home style PC grid used in our case
study (Jun-Weon Yoon 2008)

system, in which the number of iterations is taken to be 1,000,000. The duration to

maturity is further divided by 10,000 periods, which push the system to run on the
mass parallel system of K@H. To tackle this scale, or even larger scale for all kinds
of possible scenarios in real trading practice, an off-line distributed and parallel
approach is adopted. The total number of Monte Carlo simulation is 1,000,000 
10,000. It was divided into 10,000 jobs and each job consists of 1,000,000 Monte
Carlo simulations. The market parameters are given as above.
The results demonstrated here are not as good as expected (see Table 49.1). It shows
that the speedup ratio is only 3.79. In this test, K@H further divides the jobs into ten
groups. Each group was send and run in a sequential fashion, which causes the low
speedup ratio. However, if one looks into the executed CPU time for each job, our
assumption is still valid. We simulated the result in a small cluster with the
same scenario and Monte Carlo paths. The result shows 90 % speedup can be
easily achieved.
49 Computer Technology for Financial Service 1375

49.4.2.4 RBNB Data Grid

RBNB DataTurbine in market data streaming is implemented here (see Fig. 49.15),
in which the real data from iShares MSCI Taiwan Index (ETF) and TSEC weighted
index from May 31, 2005, to May 31, 2008, and 30 days tick-by-tick trading data
from Taiwan Futures Exchange Center (TAIFEX) are used and open in different
file-based online channels from the RBNB DataTurbine. The purpose is to dynam-
ically manage the high-frequency market data and connect the data with analysis
applications on the fly. The system implemented up-to-date large-scale dynamic
and static data management.

49.5 Conclusions

Securities trading is one of the few business activities where a few seconds
processing delay can cost a company big fortune. The growing competitive in the
market exacerbates the situation and pushes further towards instantaneous trading
even in split second. The key lies on the performance of the underlying information
system. Following the computing evolution in financial services, it was
a centralized process to begin with and gradually decentralized into a distribution
of actual application logic across service networks. Financial services have tradi-
tion of doing most of its heavy lifting financial analysis in overnight batch cycles.
However, in securities trading it cannot satisfy the need due to its ad hoc nature and
requirement of immediate response. A new computing paradigm, grid computing,
aiming at virtualizing scale-up distributed computing resources, is well suited to the
challenge posed by the capital market practices.
In this study we revisit the theoretical background of how performance will
affect the market competition. The core concept lies on information asymmetry.
Due to the advance of IT, even in split second, it will be a matter of win or lose in
real market practice. After establishing the motivation, we review recent grid
development specifically used for finance service. Monte Carlo simulations are
chosen not only because of its popularity in real world but also because of its nature
so-called “fine grain” or mass parallelism approach. The success of Monte Carlo
simulations lies on better random number generators. The well-recognized Sobol
sequences as a quasi-random number generator are carefully studied to ensure the
quality of Monte Carlo simulations when employed for mass parallelism. Then
some popular basic option pricing models, collectively Asian option pricing,
rainbow option pricing, and VaR calculation with constant market parameters,
are introduced as drivers to introduce more details of grids for better finance
service. Finally, we test various grid platforms, based on the methodology of
mass parallelism and mass distribution, with the drivers. The real market data are
also used, but at this stage they are only used to demonstrate the dynamic data
management, in which grids can offer better.
During this study, we encountered system architect Koschnick from Z€urcher
Kantonalbank of Switzerland (Koschnick 2008). Coincidentally, the system they
plan to migrate from big irons is the similar system to that of DRBL with additional
1376 F.-P. Lin et al.

Fig. 49.15 RBNB DataTurbine streaming open for data channels of iShares MSCI Taiwan Index
(ETF) and TSEC weighted index from May 31, 2005, to May 31, 2008, and 30 days tick-by-tick
trading data from Taiwan Futures Exchange Center (TAIFEX) (Real data plot in collaboration
with Strandell et al. 2007)
49 Computer Technology for Financial Service 1377

virtual local area networks (VLAN) for security. The system is used for overnight
batch job as well as real-time trading practice. It is confirmed that for the years to
come, financial services providers will adopt more grid or grid-based technology to
enhance their competitiveness.
Our future work will be following the current work, continuously using the
current grid platforms and extending them to the use high-frequency real market
data. Along the track of this development, we will also develop sophisticated Monte
Carlo-based option pricing and risk management based on tick-by-tick daily market
information.

References
Abramson, D., et al. (2006). Deploying scientific applications to the PRAGMA grid testbed:
strategies and lessons. In Sixth IEEE international symposium on cluster computing and the
grid, Singapore, pp. 241–248.
Ahmad, K., Gillam, L., & Cheng, D. (2005). Textual and quantitative analysis: Towards a new,
e-mediated social science. In Proceedings of the 1st international conference on e-Social
science, Manchester, pp. 22–24.
Almond, J., & Snelling, D. (1999). UNICORE: Uniform access to supercomputing as an element
of electronic commerce. Future Generation Computer Systems, 15(5), 539–548.
Amad, K., et al. (2004). Financial information grid –an ESRC e-Social social science pilot.
In Proceedings of the third UK e-Science programme all-hands meeting, Nottingham.
Antonov, I. A., & Saleev, V.M. (1979). An economic method of computing LP[tau]-sequences.
USSR Computational Mathematics and Mathematical Physics, 19(1), 252–256.
Atkins, D.E., et al. (2003). Revolutionizing science and engineering through cyberinfrastructure.
Report of the National Science Foundation Blue-Ribbon Advisory Panel on Cyberinfras-
tructure. National Science Foundation.
Bell, D., & Ludwig, A. S. (2005). Grid service discovery in the financial markets sector. In 7th
international conference on information technology interfaces, Cavtat, Croatia.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81, 637–654.
Bratley, P., & Fox, B. L. (1988). Algorithm 659: Implementing Sobol’s quasirandom sequence
generator. ACM Transactions on Mathematical Software, 14, 88–100.
Bughin, J., Chui, M., & Manyika, J. (2010). Cloud, big data and smart assets: Ten tech-enabled
business trends to watch. McKinsey Quarterly August, 7–8.
Chang, F., Dean, J., Ghemawat, S., Hsieh, W.C., Wallach, D.A., Burrows, M., Chandra, M., Fikes, A.,
Gruber, R.E. (2006). Big table: A distributed storage system for structured data. In OSDI 06:
Proceedings of the 7th USENIX Symposium on Operating Systems Design and Implementation,
Seattle, WA, p. 15-15.
Cheng, J., & Druzdzel, M. J. (2000). AIS-BN: An Adaptive importance sampling algorithm for
evidential reasoning in large bayesian networks. Journal of Artificial Intelligence Research,
13, 155–188.
Chen, G., Thulasiraman, P., & Thulasiram, R.K. (2006). Distributed quasi-monte carlo algorithm
for option pricing on HNOWs using mpC. In Proceedings of the 39th annual simulation
symposium (ANSS’06), Huntsville.
Cizek, P., Härdle, W., & Weron, R. (2011). Statistical tools for finance and insurance. Berlin:
Springer.
Dean, J., & Ghemawat, S. (2004). Map reduce: Simpli_ed data processing on large clusters.
In OSDI 04, San Francisco, CA, pp. 137–150.
Dowd, K. (2002). Measuring market risk. Chichester/England/New York: Wiley.
1378 F.-P. Lin et al.

Efon, B. (1981). Censored data and the bootstrap. Journal of the American Statistical Association,
76(374), 312–319.
Efron, B., & Tibshirani, R. (1986). Bootstrap methods for standard errors, confidence intervals,
and other measures of statistical accuracy. Statistical Science, 1(1), 54–75.
Faure, H. (1982). Discrepance de Suites Associees a un Systeme de Numeration (en Dimension s).
Acta Arithmetica, 41, 337–351.
Foster, I., & Kessleman, C. (2004). The grid: Blueprint for a new computing infrastructure
(2nd ed.). Amsterdam: Morgan Kaufmann.
Foster, I., Zhao, Y., Raicu, I., & Lu, S. (2008). Cloud computing and grid computing 360-degree
compared. In Grid computing environments workshop, Austin, TX, pp. 1–10.
Galanti, S., & Jung, A. (1997). Low-discrepancy sequences: Monte Carlo simulation of option
prices. Journal of Derivatives, 5(1), 63–83.
Ghoshal, D., Canon, R. S., & Ramakrishnan, L. (2011). I/O performance of virtualized cloud
environments. In The second international workshop on data intensive computing in the
clouds, Seattle, WA, pp 71–80.
Gillam, L., Ahmad, K., & Dear, G. (2005). Grid-enabling social scientists: The Fingrid infra-
structure. In Proceedings of the 1st international conference on e-social science,
Manchester, pp. 22–24.
Griffiths, P., & Remenyi, D. (2003). Information technology in financial services:
A model for value creation. Electronic Journal of Information Systems Evaluation,
6(2), 107–116.
Grimshaw, A. S., & Wulf, W. A. (1997). The legion vision of a worldwide virtual computer.
Communications of the ACM, 40(1), 39–45.
Halton, J. H. (1960). On the efficiency of certain quasirandom sequences of points in evaluating
multidimensional integrals. Numerische Mathematik, 2, 84–90.
Hauswald, R., & Marquez, R. (2003). Information technology and financial services competition.
The Review of Financial Studies, 16(3), 921–948.
Hull, J. C. (2003). Options, futures and other derivatives (5th ed.). Upper Saddle River:
Prentice Hall.
Joe, S., & Kuo, F. Y. (2003). Remark on algorithm 659: Implementing Sobol’s quasirandom
sequence generator. ACM Transactions on Mathematical Software, 29(1), 49–57.
Korpela, E., et al. (2001). SETI@home-massively distributed computing for SETI. Computing in
Science and Engineering, 3(1), 78–83.
Koschnick, T. (2008). Cluster & security: Separating application environments. In Int’l
Supercomputing conference and exhibition, Dresden, pp. 17–20.
Kuipers, L., & Niederreiter, H. (1974). Uniform distribution of sequence. New York: Wiley.
Lauret, J., Walker, M., Goasguen, M., & Hajdu, L. (2010). From grid to cloud, the STAR
experience. In SciDAC 2010 proceedings, Chattanooga, TN.
Leto, A., et al. (2005). EGRID Project: Experience report on the implementation of a grid
infrastructure for the analysis of financial data. In Proceedings of the international conference
on information technology: Coding and computing, Las Vegas, NV.
Litzkow, M. J., Livny, M., & Mutka, M. W. (1988). Condor-a hunter of idle workstations. In 8th
international conference on distributed computing systems, proceeding, San Jose, CA,
pp. 104–111.
Mell, P., Grance, T. (2011). The NIST definition of cloud computing (pp. 800-145). Special
Publication.
Milojicic, S. D., et al. (2002). Peer-to-peer computing (HL laboratories research report).
Niederreiter, H. (1992). Random number generation and Quasi-Monte Carlo methods. In CBMS-
NSF regional conference series in appl. math. Philadelphia, PA.
Paskov, S. H., & Traub, J. F. (1995). Faster valuation of financial derivatives. Journal of Portfolio
Management, 22(1), 113–120.
Phillips, A., et al. (1984). Effects of information technology on financial services systems.
Washington, D. C.: U.S. Congress, Office of Technology Assessment, OTA-CIT-202.
49 Computer Technology for Financial Service 1379

Ramakrishnan, L., Canon, R. S., Muriki, K., Sakrejda, I., & Wright, N. J. (2011). Evaluating
interconnect and virtualization performance for high performance computing. In Proceedings
of 2nd international workshop on performance modeling, benchmarking and simulation of high
performance computing systems, Seattle, WA, pp. 1–2.
Schmidt, A. B. (2005). Quantitative finance for physicists: An introduction. Academic Press
Advanced Finance Series, Elsevier, Oxford.
Smarr, L., & Catlett, C. (1992). Metacomputing. Communications of the ACM archive,
35(6), 44–52.
Snir, M., et al. (1996). MPI: The complete reference. The MIT Press, Cambridge, Massachusetts.
Sobol, I. M. (1967). On the distribution of points in a cube and the approximate evaluation of
integers. U.S.S.R. Computational Mathematics and Mathematical Physics, 7(4), 86–112.
Sobol, I. M., & Levitan, Y. L. (1976). The production of points uniformly distributed in a
multidimensional cube. Tech. Rep. 40, Institute of Applied Mathematics, USSR Academy of
Sciences (in Russian).
Stokes-Rees, I., et al. (2007). Multi-cluster parallel job submission: Experiences with Monte Carlo
simulations for computational finance on Grid5000 (IRIS PicsouGrid Project Report).
Strandell, E., Tilak, S., Chou, H. M., Wang, Y. T., Lin, F. P., Arzberger, P. Fountain, T. (2007).
Data management at kenting’s underwater ecological observatory. In Proceedings of the third
International Conference on Intelligent Sensors, Sensor Networks and Information Processing,
Melbourne, Australia.
Tanaka, S. (2003). Joint development project applies grid computing technology to financial risk
management. Nikkei Computer, special section on grid computing.
Tilak, S., Hubbard, P. Miller, M., & Fountain, T. (2007). The ring buffer network bus (RBNB) data
turbine streaming data middleware for environmental observing systems. In Proceedings of
third IEEE international conference on e-science and grid computing, Bangalore, India,
pp. 125–133.
Weron, R. (2004). Handbook of computational statistics. Berlin: Springer.
Yelick, K., Coghlan, S., Draney, B., Canon, R. S. (2011). The magellan report on cloud computing
for science. ASCR, DOE. Washington, DC.
Yoon, J. (2008) Korea@Home (PC Grid project in Korea). GridAsia 2008: Singapore.
Long-Run Stock Return and the Statistical
Inference 50
Yanzhi Wang

Contents
50.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1382
50.2 Long-Run Return Estimation in Event-Time Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1383
50.2.1 Return Estimations: CAR, BHAR, and Earnings
Announcement Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1384
50.2.2 Benchmark Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1386
50.2.3 Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1387
50.3 Long-Run Return Estimation in Calendar-Time Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 1392
50.3.1 Return Estimations: Mean Monthly Return and Factor Models . . . . . . . . . . . . . 1393
50.3.2 Conditional Market Model and Ibbotson’s RATS . . . . . . . . . . . . . . . . . . . . . . . . . . . 1394
50.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1396
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1396

Abstract
This article introduces the long-run stock return methodologies and their statis-
tical inference. The long-run stock return is usually computed by using a holding
strategy more than 1 year but up to 5 years. Two categories of long-run return
methods are illustrated in this article: the event-time approach and calendar-time
approach. The event-time approach includes cumulative abnormal return,
buy-and-hold abnormal return, and abnormal returns around earnings announce-
ments. In former two methods, it is recommended to apply the empirical
distribution (from the bootstrapping method) to examine the statistical inference,
whereas the last one uses classical t-test. In addition, the benchmark selections in
the long-run return literature are introduced. Moreover, the calendar-time
approach contains mean monthly abnormal return, factor models, and Ibbotson’s

Y. Wang
Yuan Ze University, Taiwan
Department of Finance, College of Management, National Taiwan University, Taipei, Taiwan
e-mail: yzwang@ntu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1381
DOI 10.1007/978-1-4614-7750-1_50,
# Springer Science+Business Media New York 2015
1382 Y. Wang

RATS, which could be tested by time-series volatility. Generally, calendar-time

approach is more prevailing due to its robustness, yet event-time method is still
popular for its ease of implementation in the real world.

Keywords
Long-run stock return • Buy-and-hold return • Factor model • Event time •
Calendar time • Cumulative abnormal • Return • Ibbotson’s RATS • Conditional
market model • Bootstrap • Zero-investment portfolio

50.1 Introduction

The long-run stock return has been an important facet of the stock performance
of firms since the 1990s. Ritter (1991) starts the line of long-run return studies
by investigating the initial public offering (IPO) cases. He finds that the post-IPO
stock performance is poor in the long run. His paper is then followed by many
scholars who study other important corporate events and asset pricing anomalies.1
Fama (1998) reviews these papers which engage in long-run event studies. Hence,
the long-run return method is nowadays a standard way for the stock performance
of firms.
I review the methodologies about the long-run stock abnormal return in this
article. Since the mid-1990s, some papers started to be aware of the properties of
long-term stock performance and suggested various ways to calculate the long-run
abnormal return. Therefore, Barber and Lyon (1997), Kothari and Warner (1997),
and Lyon et al. (1999) review and compare long-run return methods in the late
1990s. Based on these methodology papers, I update recent developments about
the new methods that are not mentioned in these two papers, such as the earnings
announcement abnormal return applications (La Porta et al. 1997), conditional
market model method (Eberhart et al. 2004; Petkova and Zhang 2005), Ibbotson’s
RATS (Ibbotson 1975; Agrawal et al. 1992; Peyer and Vermaelen 2009), and zero-
investment portfolio method (Daniel and Titman 1997; Eberhart et al. 2004). I also
make a clearer categorization on these long-run stock return methodologies given
that the long-run return methodologies nowadays are much more mature than in the
late 1990s.
Two main categories of the long-run stock performance illustrated in this paper
are event-time and calendar-time approaches. The event-time approach includes
cumulative abnormal return (CAR), buy-and-hold abnormal return (BHAR),
rebalanced buy-and-hold abnormal return (RBHAR), and abnormal returns around

1
For corporate events, papers have studied seasoned equity offerings, mergers, dividend initiations
and omissions, quarterly earnings announcements, share repurchases, proxy flights, stock splits
and spinoffs, and other corporate events for their long-run stock performance. For asset pricing
anomalies, papers investigate value premium, momentum profit, research and development profit,
accrual effect, asset growth, net share issuance, and other anomalies in terms of the long-term
impact.
50 Long-Run Stock Return and the Statistical Inference 1383

earnings announcements. The calendar-time approach contains mean monthly

abnormal return (MMAR), Ibbotson’s RATS, and factor models. The factor
model has various specifications, in which Fama and French (1993) three-factor
model and Carhart (1997) four-factor model are most prevailing two.
In addition, to estimate the abnormal return, the benchmark is vital because
a misspecification of the matching procedure yields an incorrect statistical infer-
ence and leads to the statistical type I or type II error. For event-time approach,
either a matching firm or a matching portfolio can control for the firm characteristic
effect (Daniel and Titman 1997). Yet, the more relevant problems are how
we should select the weighting scheme and what the matching criterion should
be. On one hand, the selection of the weighting scheme determines the degree of
the abnormal return and involves in different magnitude of transaction costs
from the rebalancing problem. On the other hand, different matching methods
may result in a superior or an inferior statistical inference, particularly the skewness
of the long-run return estimation.
For MMAR of the calendar-time approach, the benchmark selection problem is
similar to what we may face in the event-time approach. Yet, the family of factor
models has a simpler benchmark problem because the only issue is the factor model
specification (e.g., a three-factor or a four-factor model). The same situation applies
the Ibbotson’s RATS that we only pay attention to the market model setting,
namely, adding more independent variables in the market model. Nevertheless,
few complicated modifications for the factor models are employed in the long-run
return studies, such as conditional market model and the zero-investment
portfolio method in factor model. The former deals with the time-varying system-
atic risk loadings in the market model, whereas the later is to combine the
factor model method with the matching methods in the event-time approach.
In general, the robustness of calendar-time approach raises its popularity in recent
finance studies.
The rest of this paper is organized as follows. Section 50.2 introduces the long-
run return estimation in event-time approach. Section 50.3 states the calendar-time
method applied in the long-run stock return estimation. Finally, Sect. 50.4
concludes this review paper.

50.2 Long-Run Return Estimation in Event-Time Approach

To estimate the long-run stock return, it is nature to identify an event to track the
stock performance of a firm. Figure 50.1a shows a general time line for the long-run
return estimation. For any specific event and its event day noted by day 0, we can
start a holding strategy by purchasing the stock from event day 0 to 251th day (for
1-year return), to 503th day (for 2-year return), 755th day (for 3-year return), to
1,007th day (for 4-year return), or to 1,259th day (for 5-year return). Papers usually
compute the long-run stock return up to 5 years (Ritter 1991). As indicated in
Fig. 50.1a, an accounting reporting lag is required between the event date and the
previous fiscal year-end with at least 4 months.
1384 Y. Wang

a Fiscal Fiscal Fiscal

Year Year Event Day Year
End -1 End 0 End +1

Stock Returns (one to five years)

Reporting lag varies from 4

to 15 months

Return
b Fiscal Formation Fiscal
Fiscal
Year Year Day Year
End -1 End 0 (July 1st ) End +1

Stock Returns (one to five years)

Reporting lag varies from 6

to 17 months

Fig. 50.1 (a) time line of general event study. (b) time line of asset pricing study

For the asset pricing study, we usually do not have a specific event day for
tracking the compounding daily returns. As suggested in Fama and French (1992,
1993), a general investment strategy starts the return formation from 1 July of each
year by exerting the accounting information in the previous fiscal year-end with at
least 6 months reporting lag. Figure 50.1b shows the time line for asset pricing
studies. Because asset pricing papers start from 1st July, the long-run returns are
estimated based on compounded monthly returns. Basically, 1-year return includes
12 monthly returns, while 5-year return is computed upon 60 monthly returns.

50.2.1 Return Estimations: CAR, BHAR, and Earnings

Announcement Returns

The cumulative abnormal return (CAR) is estimated as follows. For a given bench-
mark E(Ri,t), the abnormal return ARi,t ¼ Ri,t – E(Ri,t). The average of estimations
for CAR is

X
N
ARi, t
X
T
i¼1
CAR ¼ , (50.1)
t¼1
N
50 Long-Run Stock Return and the Statistical Inference 1385

in which the Nt is for number of observations at time t and T for the holding period
(252 for a year if the return is computed based on daily returns; 12 for a year if the
return is computed based on monthly returns).
The buy-and-hold abnormal return is defined as
" # " #
Y
T   Y
T   
1 þ Ri, t 1 1 þ E Ri , t 1
X
N
t¼1
X
N
t¼1
BHAR ¼  : (50.2)
i¼1
N i¼1
N

Thus, we can use classical t-test or other methods (will be illustrated in later
section) to carry out the statistical inference for CAR and BHAR. One modification
on BHAR is the rebalanced buy-and-hold abnormal return (RBHAR), which comes
from the combination of rebalanced return (Rreb) and BHAR. Given the rebalanced
return as
0 1
XN

T B
Ri , t C
Y B C
R ¼ B i¼1 C
B1 þ N C  1,
reb
(50.3)
t¼1 @ A

the RBHAR is to calculate buy-and-hold return for a certain period (e.g., a year) and
rebalance the portfolio equally for every specific period. Taking the 1-year
rebalanced BHAR as the example, we should compute 1-year BHAR for each
event year (i.e., first to fifth event year) and obtain the average BHAR for every
event year. Finally, we obtain compounding return for this average rebalanced
BHAR with yearly rebalancing. Because BHAR has inflated compounding return,
which results in many outliers in the long-run return estimation, RBHAR with
rebalancing every year is able to reduce the impact from extreme values (Ikenberry
et al. 1995; Chan et al. 2010). In general, RBHAR could be described as

8 " # 9
> Y
T   >
>
> 1 þ Ri, j  1>
>
Y>
year5 < X
N
j¼1
>
=
RBHAR ¼ 1þ
>
year1 >
N >
>
>> i¼1 >
>
: ;
8 " # 9
> Y
T    >
>
> 1 þ E Ri , j  1>
>
Y>
year5 < X
N
j¼1
>
=
 1þ (50.4)
>
year1 >
N >
>
>
> i¼1 >
>
: ;

Note that T stands for 252 days as an event year, and year 1 to year 5 represent
the first event year to the fifth event year.
1386 Y. Wang

Distinct from CAR and BHAR, the quarterly earnings announcement return is
the estimate of the long-term stock performance by successive short-term quarterly
announcement returns (La Porta et al. 1997; Denis and Sarin 2001; Chan
et al. 2004). In other words, we can use the 2-day or 3-day abnormal returns
around quarterly announcement dates, and the successive earnings surprises should
be aligned with the long-run stock abnormal return. The benchmark selection
determines the long-run abnormal return estimation because of its statistical
property (will be discussed in next section), so the results may be changed
when benchmark settings are altered. This sensitive outcome driven from the
benchmark problem leads to an inconclusive long-run abnormal return. Yet, the
short-run return is not sensitive to the selection of benchmarks. For example, given
10 % and 20 % of market index returns, the 3-day market index returns are
expected to be 0.12 % and 0.24 % only. When an event occurs, the 3-day
announcement return could be generally above 1 %, which significantly exceeds
any selected benchmark returns. Therefore, if a corporate event is followed by
profitability improvements that are not observed by investors, then there should be
successive earnings surprises following the event date, and the short-term
announcement abnormal returns around the earnings announcement dates
should be positive. To capture the long-run abnormal return, papers usually study
12–20 quarters (for 3–5 years) for the quarterly earnings announcement abnormal
return. Generally, the quarterly earnings announcement abnormal returns capture
about 25–40 % of the long-run stock return of a firm (Bernard and Thomson 1989;
Sloan 1996).

50.2.2 Benchmark Problem

Long-run stock return has some issues regarding the calculation and testing when
we select the benchmarks for the expected return of the firm. Namely, the conven-
tional methodologies for the long-run stock return might be biased if the chosen
matching procedure is inappropriate. First, some long-term return measures have
the rebalancing and new-listing biases problems. Second, long-run stock return is
positive skewed, thus the traditional t-test is inappropriate for the long-run return.
Although I will review the skewness-adjusted t-statistics and empirical p-value in
the next section, matching firm method is another way to alleviate the skewness
concern in the long-run return studies.
The first benchmark for the expected return is the CRSP equal-weighted
index return, which includes whole stocks in CRSP database and computes the
simple average of stock returns. However, this approach involves rebalancing
problem, which ignores the transaction costs from broker’s fee and tax to
the government. To maintain the weights on stocks equally, investors must sell
profitable stocks and buy stocks with loss. This rebalancing leads to huge transac-
tion costs that are not considered in the CRSP equal-weighted index return, making
the abnormal return underestimated. In fact, CAR per se also has this rebalancing
problem because of its average in cross section.
50 Long-Run Stock Return and the Statistical Inference 1387

The second benchmark is the CRSP value-weighted index return, which also
comes from CRSP database. This index return uses firm value (equal to the price
timing shares outstanding) as the weighted scheme to compute the weighted
average of stock returns. Most importantly, it is a rebalance-free benchmark and
accordingly has no transaction cost during the holding period (except the beginning
and end). Therefore, recent papers tend to use CRSP value-weighted index return
instead of CRSP equal-weighted index return as the benchmark return.
The third benchmark is the reference portfolio return. Before constructing the
reference portfolio, we may need matching criterions. As the important pricing
factors, size and book-to-market (BM) ratio are the two most important in deter-
mining the expected return (Fama and French 1992, 1993, 1996; Lakonishok
et al. 1994; Daniel and Titman 1997). In particular, we have to determine
the matching pool, which includes stocks that are irrelevant to the sample firm.
We form 50 size and book-to-market portfolios (ten size portfolios and five book-
to-market portfolios in each size decile) where the size and book-to-market cutoff
points are obtained from stock in NYSE exchange. We then are able to compute
either equal-weighted or value-weighted portfolio return as the expected return.
The last one is the matching firm method. Because the long-run stock return is
positive skewed, we may take the long-run return of the matching firm as the
expected return. The skewed return of the sample firm and skewed return of the
matching firm offset each other and make the abnormal return symmetric (Barber
and Lyon,1997). In addition, matching firm method avoids the new-listing bias and
rebalancing problem. In general, the matching criterions of the matching firm are
similar to the reference portfolio. Within each reference portfolio, a matching firm
could be selected by minimizing the book-to-market difference between the sample
firm and the matching firm (Ikenberry et al. 1995). Sometimes, papers use few
matching firms but not single matching firm as the benchmark to avoid few outlier
impacts (e.g., Lee 1997) or use different matching variable (e.g., Ikenberry and
Ramnath 2002; Eberhart et al. 2004). Generally, various matching methods under
size and book-to-market effect controls do not largely change the results.

50.2.3 Statistical Inference

The most important statistical problem for the long-run abnormal return is the
skewness. The minimum loss of a long-term stock investment is 100 % while
the maximum potential gain approaches infinite. Thus, the distribution of the long-
run stock return is positive skewed. If we test the long-run stock return by a standard
normal distribution, then we tend to reject the null hypothesis (that suggests
no abnormal return) for negative returns and accept the null for positive returns.
This misspecification leads to a type I error in the distributional left tail but causes
a type II error in the distributional right tail.
To solve the skewness problem, Barber and Lyon (1997) suggest the matching
firm method because abnormal stock return is the return difference between the
sample firm and matching firm, making the skewness from matching firm and
1388 Y. Wang

sample firm offset by each other. In addition, there are two ways to alleviate the
testing problem under the skewness, one is the skewness-adjusted t-statistic and the
other is the empirical p-value, suggested by Ikenberry et al. (1995) and Lyon
et al. (1999).
Given an abnormal return ARi, the skewness-adjusted t-statistic illustrated by
Lyon et al. (1999) is tested as follows:

pffiffiffiffi 1 1
tsa ¼ N ðSÞ þ ^g S2 þ ^g ,
3 6N

where

X
N  3
ARi  AR
AR i¼1
S¼ , and ^g ¼ : (50.5)
sðARi Þ NsðARi Þ3

Note that the S is the conventional t-statistic and ^g is the coefficient for the
skewness adjustment.
The second suggested statistical inference method is the empirical p-value. This
approach uses bootstrapping method to construct an empirical distribution with
general long-term return features. We use bootstrapping method to select the
pseudo-sample firm. Thus, we are able to compare sample firm and pseudo-sample
firm as the base of the empirical p-value. In fact, it is possible that we may face
moment conditions (higher than third moment condition) in the long-term return
estimation, and the skewness-adjusted t-statistic is not enough to capture the return
characteristic. Also, the strong cross-sectional correlations among sample observa-
tions in BHARs can lead to poorly specified test statistics (Fama 1998; Lyon
et al. 1999; Brav 2000). Under the empirical distribution, we can examine the
statistical inference without a parametric distribution but are able to capture more
unknown statistical features.
As mentioned above, the empirical p-value is generated from the empirical
distribution from bootstrapping, and the empirical distribution well controls the
skewness and time-dependent properties of the long-run stock returns (Ikenberry,
et al. 1995; and Lyon et al. 1999; Chan et al. 2004). In addition, this empirical
p-value also solves the statistical inference problem in RBHAR because we may
have too few observations in times series for computing standard deviation.2
To construct the empirical distribution, we need to find pseudo-sample firms that
share similar firm characteristics as the sample firm but do not have the interested
corporate events. Next, we construct 25 size and book-to-market portfolios (five
size portfolios and five book-to-market portfolios in each size quintile) from all

2
For example, if we compute a 5-year RBHAR, then we have five averages of BHARs in five event
years only.
50 Long-Run Stock Return and the Statistical Inference 1389

nonevent firms. The nonevent firm selection criterions are similar to the matching
firm selection. Then, we randomly sample one pseudo-sample firm out of the
corresponding portfolio for each sample firm. For example, if an event firm is
with the second size quintile and third book-to-market quintile, then we
randomly choose a pseudo firm that is also with the second size quintile and third
book-to-market quintile. Hence, we are able to obtain N pseudo firms for N sample
firms. Upon these pseudo-sample firms, we calculate the long-run return by CAR,
BHAR, or RBHAR method. Finally, we repeat the sampling and long-run return
estimation for 1,000 times and obtain 1,000 averages of long-run returns for pseudo-
sample firms. The empirical distribution is plotted according to the frequency
distribution diagram of those 1,000 average returns of pseudo-sample firms.
If the long-run return is larger than the (1a) percentile of the 1,000 average
pseudo-sample firm returns, then we obtain the p-value as a for testing the
positive average abnormal return. Similarly, if the long-run return is smaller than
a percentile of the 1,000 average pseudo-sample firm returns, then we obtain the
p-value as a for testing the negative average abnormal return.
Figures 50.2a–50.2d are empirical distributions for 1-year to 4-year long-run
returns upon size/book-to-market controlled pseudo-sample firms of US repurchase
firms during 1980–2008. I collect the repurchase data from SDC database as the
example for the empirical distribution construction. For those empirical distribu-
tions, it is obvious that 4-year return has more outliers than 1-year return. Moreover,
the 4-year return figure has lower kurtosis and is more positive skewed. Obviously,
the shape of long-run returns does not obey normal distribution, and the empirical
p-value is more relevant to the long-run abnormal return testing.
I also show the specification (statistical size) for different long-run return methods
in Table 50.1, which is obtained from Table 5 of Barber and Lyon (1997) and Table 3
of Lyon et al. (1999). They show the percentage of 1,000 random samplings from
200 firms rejecting the null hypothesis that suggests no abnormal return in terms of
CAR, BHAR, and RBHAR with different benchmark and testing methods. First, the
CAR with matching portfolio as the benchmark has type I error in left tail when
measuring the abnormal return in 5 years. Second, the matching firm method yields
good specification no matter how we focus on size control, BM control, or
a combination control for both size and BM ratio. Third, skewness-adjusted t-statistics
has type I error, implying that skewness is not the only statistical feature that we
should address. Forth, empirical p-value method performs well even when adopting
the reference portfolio as the benchmark, at least for 10 % significance level.
Figure 50.3 shows the testing power of alternative tests by using BHAR as the
primary method, and this figure is originally plotted in Fig. 1 of Lyon et al. (1999).
The empirical distribution performs better in testing power than classical t-test,
no matter what the standard empirical distribution or the bootstrapped skewness-
adjusted t-statistic is employed.
In sum, in the event-time approach, the BHAR is suggested. Matching firm
is a better matching method than other approaches. In statistical testing, the
empirical p-value could be the best way due to its well statistical size control and
testing power.
1390 Y. Wang

a 70
60
50
40
30
20
10
0
0.08
0.09
0.1
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.19
0.2
0.21
0.22
0.23
0.24
0.25
0.26
b 40
35
30
25
20
15
10
5
0
0.29
0.3
0.31
0.32
0.33
0.33
0.34
0.35
0.36
0.37
0.37
0.38
0.39
0.4
0.41
0.41
0.42
0.43
0.44
0.45
c 35
30
25
20
15
10
5
0
0.54
0.54
0.55
0.56
0.57
0.57
0.58
0.59
0.59
0.6
0.61
0.61
0.62
0.63
0.64
0.64
0.65
0.66
0.66
0.67
0.68
0.68

d 30
25
20
15
10
5
0
0.77

0.78

0.79

0.79

0.80

0.81

0.82

0.83

0.83

0.84

0.85

0.86

0.87

0.87

0.88

0.89

0.90

0.91

0.91

Fig. 50.2 (a) One-year return distribution of size/BM controlled pseudo-sample firms. (b)
Two-year return distribution of size/BM controlled pseudo-sample firms. (c) Three-year return
distribution of size/BM controlled pseudo-sample firms. (d) Four-year return distribution of size/
BM controlled pseudo-sample firms
50 Long-Run Stock Return and the Statistical Inference 1391

Table 50.1 Specification (size) for alternative statistical tests

A: Specification (size) for CAR with different benchmarks
Two-tailed theoretical significant level (%) 1 5 10
Theoretical cumulative density function (%) 0.5 99.5 2.5 97.5 5.0 95.0
Description of return benchmark Mean Skew
Panel C: 60-month CARs
Size deciles 0 2.4* 0.6 8.0* 1.2 14.7* 3.45 1.11
Book-to-market deciles 0.1 0.7 1.9 4.4* 2.6 7.6* 1.47 1.24
Fifty size/book-to-market portfolio 0.2 1.3* 0.9 5.5* 2.2 10.0* 2.10 1.21
Equally weighted market index 0.0 5.5* 0.2 17.3* 0.5 25.1* 6.27 1.11
Size-matched control firm 0.6 0.3 2.1 2.2 5.2 4.3 0.59 0.14
Book-to-market-matched control firm 0.4 0.8 2.9 3.1 5.2 5.4 0.00 0.01
Size-/book-to-market-matched control firm 0.2 0.4 2.4 2.3 4.3 4.3 0.63 0.07
Fama-French three-factor model a 0.5 0.3 2.1 2.3 4.9 5.1 0.94 1.76
B: Specification (size) for BHAR with different benchmarks
Statistic Benchmark Two-tailed theoretical significance level
1% 5% 10 %
Theoretical cumulative
density function (%)
0.5 99.5 2.5 97.5 5 95
t-Statistic Rebalanced size/book-to- 11.7* 0.0 23.7* 0.0 33.2* 0.2
market portfolio
t-Statistic Buy-and-hold size/book-to- 2.4* 0.0 6.1* 0.5 10.5* 1.6
market portfolio
Skewness-adjusted Buy-and-hold size/book-to- 1.4* 0.4 4.4* 1.7 8.2* 4.8
t-statistic market portfolio
t-Statistic Size/book-to-market control 0.1 0.1 3.0 1.9 5.4 3.9
firm
Bootstrapped Buy-and-hold size/book-to- 0.6 1.2* 2.2 3.1 5.0 5.7
skewness-adjusted market portfolio
t-statistic
Empirical p-value Buy-and-hold size/book-to- 0.2 1.5* 2.7 3.7* 4.9 6.3
market portfolio
This table is from Table 5 of Barber and Lyon (1997, p. 363) and Table 3 of Lyon
et al. (1999, p. 179). The numbers presented represent the percentage of 1,000 random samples
of 200 firms that reject the null hypothesis of 5-year CAR (Panel A) and buy-and-hold abnormal
return (Panel B) at the theoretical significance levels of 1 %, 5 %, or 10 % in favor of the
alternative hypothesis of a significantly negative abnormal return (i.e., calculated p-value is less
than 0.5 % at the 1 % significance level) or a significantly positive abnormal return (calculated
p-value is greater than 99.5 % at the 1 % significance level). The alternative statistics and
benchmarks are described in detail in the main text. * indicates the percentage is significantly
different from the theoretical significance level at the 5 % (Panel A) and 1 % level (Panel B),
one-sided binomial test statistic
1392 Y. Wang

100

empirical
Percentage of Samples Rejecting Null

p value
80
bootstrapped
skewness-adj.
60 t statistic

40

t statistic,
control firm
20

0
−20 −15 −10 −5 0 5 10 15 20
Induced Level of Abnormal Return (%)

Fig. 50.3 Power of alternative tests in random ample. This figure is from Fig. 1 of Lyon
et al. (1999, p. 180). The percentage of 1,000 random samples of 200 firms rejecting the null
hypothesis of no annual buy-and-hold abnormal return at various induced levels of abnormal
return (horizontal axis) based on control firm method, bootstrapped skewness-adjusted t-statistic,
and empirical p-values

50.3 Long-Run Return Estimation in Calendar-Time Approach

A potential problem in the event-time return estimations is the cross-sectional

dependence. For example, buy-and-hold return is computed over a long horizon;
it is possible that many sample firms’ returns overlap with each other, making
strong cross-sectional correlations among long-horizon returns. This cross-
sectional dependence is even more profound when we have repeating events by
the same firm, such as repurchase, SEO, merger, and stock splits. In addition,
buy-and-hold returns also enlarge the long-run abnormal return because of the
inflated returns stemming from the compounding effect. Therefore, the long-run
return results might disappear if we apply other methodologies to compute the
long-run abnormal return (Mitchell and Stafford 2000). Fama (1998) also
documents that the long-run stock return should be examined by the value-weighted
factor model since the buy-and-hold return usually uses an equal-weighted
scheme that is related to the ignored transaction costs. Although Loughran and
Ritter (2000) suggest that the value-weighted factor model is the least powerful test
for long-run returns, the calendar-time method could be always a robust check for
our long-run return estimation.
50 Long-Run Stock Return and the Statistical Inference 1393

To estimate the abnormal return in calendar time, we need to form a monthly

portfolio for each calendar month. The portfolio return could be generated by an
equal-weighted, value-weighted, or a log-value-weighted method (Ikenberry
et al. 2000). Upon these monthly returns in calendar time, we are able to carry
out the mean monthly abnormal return or factor model analysis.

50.3.1 Return Estimations: Mean Monthly Return and Factor Models

The first method in the calendar-time approach is the mean monthly abnormal
return (MMAR). We can start from a 5-year holding strategy on a corporate event.
For any given calendar month (e.g., June 2002), we need to include a stock if it had
the corporate event in the past 60 months (e.g., looking backward at a period of May
2002–June 1997). We then need to form a monthly portfolio for this specific calendar
month, and the portfolio return could be computed upon equal-weighted, value-
weighted or log-value-weighted scheme. Next, we repeat the abovementioned step
for all calendar months throughout the sample period, and then the mean monthly
return is the time-series average of monthly portfolio returns.
Similar to the benchmark problem in event-time approach, we need to select the
CRSP market index return, reference portfolio, or the matching firm as the bench-
mark for the MMAR. For any benchmark E(Ri), the MMAR is

X
T  
RPt  RM
t
t¼1
MMAR ¼ ,
T
where

X
Nt X
Nt
 
wi, t Ri, t wi, t E Ri, t
i¼1 i¼1
RPt ¼ t ¼
, and RM : (50.6)
Nt Nt
T is for calendar month in this setting; RPt is the monthly portfolio return of
sample firms; RM t is the monthly portfolio return of benchmarks; and Nt is the
number of observations in each calendar month. The statistical inference can be
either classical t-statistic or the Newey and West (1987) estimation.
As suggested by Fama (1998), Mitchell and Stafford (2000), and Schultz (2003),
factor model is a robust method in estimating the long-run stock abnormal return.
The standard Fama and French three-factor and Carhart four-factor models could be
described as
 
RPt  r f , t ¼ a þ b r m, t  r f , t þ sSMBt þ hHMLt þ et , (50.7)
 
RPt  r f , t ¼ a þ b r m, t  r f , t þ sSMBt þ hHMLt þ mMOMENTUMt þ et ,
(50.8)
1394 Y. Wang

where RPt is the sample firm portfolio return for each calendar month, and could
be obtained from the equal-weighted, value-weighted, or log-value-weighted
average. This average return in calendar is similar to what we compute in the
MMAR. rf is the risk-free rate, usually the short-term Treasury bill rate; rm is
usually computed as CRSP value-weighted index return; SMB is small firm
portfolio return minus big-firm portfolio return; HML is the high book-to-market
portfolio return minus low book-to-market portfolio return; MOMENTUM is
the winner portfolio return minus loser portfolio return where winner and
loser portfolios are identified by past 1-year return. SMB, HML, and MOMENTUM
are applied to control size and book-to-market and momentum effects,
respectively (Fama and French 1992, 1993; Jegadeesh and Titman 1993;
Lakonishok et al. 1994; Carhart 1997). The abnormal return is the regression
intercept and can be tested based on the t-values or Newey and West (1987)
estimator.
Next, I introduce a modification of the factor model analysis: the zero-
investment portfolio method. Daniel and Titman (1997) and Eberhart et al.
(2004) study the long-run return by using the zero-investment portfolio approach
to control for both risk and firm characteristic effects. To form the factor
model under a zero-investment portfolio strategy, we have to buy sample stocks
and short-sell matching stocks. Taking Carhart (1997) four-factor model as the
example, we have
 P   
RPt  RM
t ¼ a a
M
þ ðbP  bM Þ r m, t  r f , t þ ðsP  sM ÞSMBt
þ ðhP  hM ÞHMLt þ ðmP  mM ÞMOMENTUMt þ et , (50.9)

and we use (aP–aM) as the abnormal return controlling for both risks and
firm characteristics. It is also the hedging portfolio return controlled for the
common risk factors. The matching firm selection criterions can apply the steps
in benchmark problem section. For other modifications of the factor model analysis,
Eberhart et al. (2004) provide more examples in their Table 3.

50.3.2 Conditional Market Model and Ibbotson’s RATS

One major challenge to the standard market model is that the risk loadings
are assumed to be unchanged. To estimate the factor loadings, we usually need
long time series to obtain the estimated risk loadings, and the fixed risk
loading over time is naturally assumed in the OLS analysis. Yet, the magnitude
of the risk of a firm could be changed; in particular many corporate events
change the risk of the firm (e.g., R&D increases could be followed by risk increases,
and share repurchase could be followed by risk decreases). Accordingly it is
needed to introduce the conditional market model to address the time-varying
market model.
50 Long-Run Stock Return and the Statistical Inference 1395

There are at least two ways to address the time-varying risks in the regression
model as the conditional market model. To simplify the problem, I use the CAPM
as the first example. First, the systematic risk could change along with some
firm characteristics and macroeconomic variables. Petkova and Zhang (2005) use
the following regression analysis to estimate abnormal return:

 
RPt  r f , t ¼ a þ ðb0 þ b1 DIV t þ b2 DEFt þ b3 TERMt þ b4 TBt Þ r m, t  r f , t þ et :
(50.10)
(b0 + b1DIVt + b2DEFt + b3TERMt + b4TBt) is the bt that accommodates
to the time-varying risk loading. They assume that the risk changes with the
dividend yield (DIV), the default spread (DEF), the term spread (TERM), and the
short-term Treasury bill rate (TB). Again, the abnormal return is the intercept a.
The second method is the rolling regression, as suggested by Petkova and Zhang
(2005) and Eberhart et al. (2004). If we have a sample period from January 1990,
then we can use the portfolio returns in the first 60 months (i.e., January
1990–December 1994) to carry out the Carhart (1997) four-factor regression.
We substitute the estimated factor loadings from these 60 monthly returns into
the equity premiums in 61th month (i.e., January 1995) and then obtain the
expected portfolio return for 61th month. Thus, the abnormal return for 61th
month is from the portfolio return of the sample firm minus the expected portfolio
return. Next, we need to repeat the abovementioned steps for every month by
rolling return windows. Finally, we estimate the abnormal return as the average
of abnormal returns across time and use the time-series volatility to test the
statistical significance.
The final method relating to the time-varying risk is the Ibbotson (1975) RATS
though it is not under the family of the factor model analysis. The original setting of
Ibbotson RATS is designed for the long-run return estimation, yet recent papers
use this method combining the factor model analysis to measure the long-run
abnormal return (e.g., Peyer and Vermaelen 2009). Based on Carhart (1997)
four-factor model, we regress the security excess return on the Carhart (1997)
four factors for each month in the event time. Given a 60-month holding
strategy, we have to carry out this regression for 1st month to 60th month following
the corporate event date. Then, the abnormal return for month t is the intercept of
this four-factor regression:
 
Ri, t  r f , t ¼ at þ bt r m, t  r f , t þ st SMBt þ ht HMLt þ mt MOMENTUMt þ et :
(50.11)
The regression analysis is similar to what I introduce in Eq. 50.8; however,
the regression is examined every event month t. For every event month or event
year, we can obtain the average abnormal return as the average of the intercepts
(at), which is obtained from a model allowing time-varying risks.
1396 Y. Wang

50.4 Conclusion

The long-run return studies have been investigated for many corporate events and
asset pricing studies in the past two decades. I introduce the long-run stock return
methodologies and their statistical inference adopted in recent papers. Two cate-
gories of long-run return methods are illustrated: the event-time approach and
calendar-time approach. Under the event-time category, we have methods includ-
ing cumulative abnormal return, buy-and-hold abnormal return, and abnormal
returns around earnings announcements. Although the event-time approach is
able to be implemented as an investment strategy in real world, it also raises
more benchmark and statistical inference problems. Under the calendar-time cate-
gory, we have mean monthly abnormal return, factor models, and Ibbotson’s
RATS. Generally, calendar-time approach is more popular due to its robustness
and variety. For any long-run return study, I may suggest that combining works on
those methodologies could be necessary.

References
Agrawal, A., Jaffe, J., & Mandelker, G. (1992). The post-merge performance of acquiring firms:
A re-examination of an anomaly. Journal of Finance, 47, 1605–1621.
Barber, B., & Lyon, J. (1997). Detecting long-run abnormal stock returns: The empirical power
and specification of test statistics. Journal of Financial Economics, 43, 341–372.
Bernard, V., & Thomas, J. (1989). Post-earnings-announcement drift: Delayed price response or
risk premium. Journal of Accounting Research, 27, 1–36.
Brav, A. (2000). Inference in long-horizon event studies: A Bayesian approach with application to
initial public offerings. Journal of Finance, 55, 1979–2016.
Carhart, M. (1997). On persistence in mutual fund performance. Journal of Finance, 52, 57–82.
Chan, K., Ikenberry, D., & Lee, I. (2004). Economic sources of gain in share repurchases. Journal
of Financial and Quantitative Analysis, 39, 461–479.
Chan, K., Ikenberry, D., Lee, I., & Wang, Y. (2010). Share repurchases as a potential tool to
mislead investors. Journal of Corporate Finance, 16, 137–158.
Daniel, K., & Titman, S. (1997). Evidence on the characteristics of cross sectional variation in
stock returns. Journal of Finance, 52, 1–33.
Denis, D., & Sarin, A. (2001). Is the market surprised by poor earnings realizations following
seasoned equity offerings? Journal of Financial and Quantitative Analysis, 36(2), 169–193.
Eberhart, A., Maxwell, W., & Siddique, A. (2004). An examination of long-term abnormal stock
returns and operating performance following R&D increases. Journal of Finance, 59(2),
623–650.
Fama, E., & French, K. (1992). The cross-section of expected returns. Journal of Finance, 47,
427–466.
Fama, E., & French, K. (1993). Common risk factors in the returns on stocks and bonds. Journal of
Financial Economics, 33, 3–56.
Fama, E., & French, K. (1996). Multifactor explanations of asset pricing anomalies. Journal of
Finance, 51, 55–84.
Fama, E. F., (1998). Market efficient, long-term returns, and behavioral finance. Journal of
Financial Economics, 49, 283–306.
Ibbotson, R. (1975). Price performance of common stock new issues. Journal of Financial
Economics, 2, 235–272.
50 Long-Run Stock Return and the Statistical Inference 1397

Ikenberry, D., & Ramnath, S. (2002). Underreaction to self-selected news events: The case of
stock splits. Review of Financial Studies, 15, 489–526.
Ikenberry, D., Lakonishok, J., & Vermaelen, T. (1995). Market underreaction to open market share
repurchases. Journal of Financial Economics, 39, 181–208.
Ikenberry, D., Lakonishok, J., & Vermaelen, T. (2000). Stock repurchases in Canada: Performance
and strategic trading. Journal of Finance, 55, 2373–2397.
Jegadeesh, N., & Titman, S. (1993). Returns to buying winners and selling losers: Implications for
stock market efficiency. Journal of Finance, 48, 65–91.
Kothari, S. P., & Warner, J. (1997). Measuring long-horizon security price performance. Journal
of Financial Economics, 43, 301–339.
La Porta, R., Lakonishok, J., Shleifer, A., & Vishny, R. (1997). Good news for value stocks:
Further evidence on market efficiency. Journal of Finance, 52(2), 859–874.
Lakonishok, J., Shleifer, A., & Vishny, R. (1994). Contrarian investment, extrapolation, and risk.
Journal of Finance, 49, 1541–1578.
Lee, I. (1997). Do firms knowingly sell overvalued equity? Journal of Finance, 52, 1439–1466.
Loughran, T., & Ritter, J. (2000). Uniformly least powerful tests of market efficiency. Journal of
Financial Economics, 55, 361–389.
Lyon, J., Barber, B., & Tsai, C. L. (1999). Improved methods for tests of long-run abnormal stock
returns. Journal of Finance, 54, 165–201.
Mitchell, M., & Stafford, E. (2000). Managerial decisions and long-term stock price performance.
Journal of Business, 73, 287–329.
Newey, W., & West, K. (1987). A simple, positive semi-definite, heteroskedasticity and autocor-
relation consistent covariance matrix. Econometrica, 55, 703–708.
Petkova, R., & Zhang, L. (2005). Is value riskier than growth. Journal of Financial Economics, 78,
187–202.
Peyer, U., & Vermaelen, T. (2009). The nature and persistence of buyback anomalies. Review of
Financial Studies, 22, 1693–1745.
Ritter, J. (1991). The long-run performance of initial public offerings. Journal of Finance, 46,
1–27.
Schultz, P. (2003). Pseudo market timing and the long-run underperformance of IPOs. Journal of
Finance, 58, 483–517.
Sloan, R. (1996). Do stock prices fully reflect information in accruals and cash flows about future
earnings? Accounting Review, 71, 289–315.
Value-at-Risk Estimation via
a Semi-parametric Approach: Evidence 51
from the Stock Markets

Cheng-Few Lee and Jung-Bin Su

Contents
51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1400
51.2 Empirical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1403
51.2.1 Parametric Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1403
51.2.2 Semi-parametric Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1406
51.3 Evaluation Methods of Model-Based VaR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1406
51.3.1 Log-Likelihood Ratio Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1407
51.3.2 Binary Loss Function or Failure Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1408
51.3.3 Quadratic Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1408
51.3.4 The Unconditional Coverage Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1408
51.3.5 Unexpected Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409
51.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409
51.4.1 Data Preliminary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409
51.4.2 Estimation Results for Alternate VaR Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1410
51.4.3 The Results of VaR Performance Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1413
51.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1422
Appendix 1: The Left-Tailed Quantiles of the Standardized SGT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1423
Appendix 2: The Procedure of Parametric VaR Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1423
Appendix 3: The Procedure of Semi-parametric VaR Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1428
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1429

C.-F. Lee
Department of Finance and Economics, Rutgers Business School, Rutgers, The State University of
New Jersey, Piscataway, NJ, USA
Graduate Institute of Finance, National Chiao Tung University, Hsinchu, Taiwan
e-mail: lee@business.rutgers.edu; cflee@business.rutgers.edu
J.-B. Su (*)
Department of Finance, China University of Science and Technology, Nankang, Taipei, Taiwan
e-mail: jungbinsu@cc.cust.edu.tw; jungbinsu@gmail.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1399
DOI 10.1007/978-1-4614-7750-1_51,
# Springer Science+Business Media New York 2015
1400 C.-F. Lee and J.-B. Su

Abstract
This study utilizes the parametric approach (GARCH-based models) and the
semi-parametric approach of Hull and White (Journal of Risk 1: 5–19, 1998)
(HW-based models) to estimate the Value-at-Risk (VaR) through the accuracy
evaluation of accuracy for the eight stock indices in Europe and Asia stock markets.
The measure of accuracy includes the unconditional coverage test by Kupiec
(Journal of Derivatives 3: 73–84, 1995) as well as two loss functions, quadratic
loss function, and unexpected loss. As to the parametric approach, the parameters of
generalized autoregressive conditional heteroskedasticity (GARCH) model are
estimated by the method of maximum likelihood and the quantiles of asymmetric
distribution like skewed generalized student’s t (SGT) can be solved by composite
trapezoid rule. Sequentially, the VaR is evaluated by the framework proposed by
Jorion (Value at Risk: the new benchmark for managing financial risk. New York:
McGraw-Hill, 2000). Turning to the semi-parametric approach of Hull and White
(Journal of Risk 1: 5–19, 1998), before performing the traditional historical
simulation, the raw return series is scaled by a volatility ratio where the volatility
is estimated by the same procedure of parametric approach. Empirical results show
that the kind of VaR approaches is more influential than that of return distribution
settings on VaR estimate. Moreover, under the same return distributional setting,
the HW-based models have the better VaR forecasting performance as compared
with the GARCH-based models. Furthermore, irrespective of whether the GARCH-
based model or HW-based model is employed, the SGT has the best VaR forecast-
ing performance followed by student’s t, while the normal owns the worst VaR
forecasting performance. In addition, all models tend to underestimate the real
market risk in most cases, but the non-normal distributions (student’s t and SGT)
and the semi-parametric approach try to reverse the trend of underestimating.

Keywords
Value-at-Risk • Semi-parametric approach • Parametric approach • Generalized
autoregressive conditional heteroskedasticity • Skewed generalized student’s t •
Composite trapezoid rule • Method of maximum likelihood • Unconditional
coverage test • Loss function

51.1 Introduction

Over the last two decades, a number of global and national financial disasters have
occurred due to failures in risk management procedures. For instance, US Savings and
Loan crisis of 1989–1991, Japanese asset price bubble collapse of 1990, Black
Wednesday of 1992–1993, 1994 economic crisis in Mexico, 1997 Asian Financial
Crisis, 1998 Russian financial crisis, financial crisis of 2007–2010, followed by the
late 2000s recession, and the 2010 European sovereign debt crisis. The crises caused
many enterprises to be liquidated and many countries to face near depressions in
their economies. These painful experiences once again underline the importance
of accurately measuring financial risks and implementing sound risk management
51 Valuet-isk Estimation via a Semi-rametric Approach 1401

policies. Hence, Value-at-Risk (VaR) is a widely used risk measure of the risk of loss
on a specific portfolio of financial assets because it is an attempt to summarize the
total risk with a single number. For example, if a portfolio of stocks has a 1-day 99 %
VaR of US$1,000, there is a 1 % probability that the portfolio will fall in value by
more than US$1,000 over a 1-day period. In other words, we are 99 % certain that we
will not lose more than US$1,000 in the next 1 day, where 1 day is the time horizon,
99 % is the confidence level, and the US$1,000 is the VaR of the portfolio.
VaR estimates are currently based on either of three main approaches: the
historical simulation, the parametric method, and the Monte Carlo simulation.
The Monte Carlo simulation is a class of computational algorithms that rely on
repeated random sampling to compute their results. That is, this approach allows for
an infinite number of possible scenarios you are exposing yourself to huge model
risks in determining the likelihood of any given path. In addition, as you had more
and more variables that could possibly alter your return paths, model complexity
and model risks also increase in scale. Like historical simulation, however, this
methodology removes any assumption of normality and thus, if modeled accu-
rately, probably would give the most accurate measure of the portfolio’s true
VaR. Besides, little research such as Vlaar (2000) had applied this approach to
estimate the VaR. The parametric method is also known as variance/covariance
approach. This method is popular because the only variables you need to do the
calculation are the mean and standard deviation of the portfolio, indicating the
simplicity of the calculations. The parametric method assumes that the returns of
the portfolios are normally distributed and serially independent. In practice, this
assumption of return normality has proven to be extremely risky. Indeed, this was
the biggest mistake that LTCM made gravely underestimating their portfolio
risks. Another weakness with this method is the stability of the standard deviation
through time as well as the stability of the variance/covariance matrix in your
portfolio. However, it is easy to depict how correlations have changed over
time particularly in emerging markets and through contagion in times of
financial crisis. Additionally, numerous studies focused on the parametric approach
of generalized autoregressive conditional heteroskedasticity (GARCH) family
variance specifications (i.e., risk metrics, asymmetric power ARCH (APARCH),
exponential GARCH (EGARCH), threshold GARCH (TGARCH), integrated
GARCH (IGARCH), and fractional IGARCH (FIGARCH)) to estimate the VaR
(see Vlaar (2000), Giot and Laurent (2003a, b), Gencay et al. (2003), Cabedo and
Moya (2003), Angelidis et al. (2004), Huang and Lin (2004), Hartz et al. (2006), So
and Yu (2006), Sadeghi and Shavvalpour (2006), Bali and Theodossiou (2007),
Bhattacharyya et al. (2008), Lee et al. (2008), Lu et al. (2009), Lee and Su (2011),
and so on). Lately, in the empirical study of parametric VaR approach, several
researches have utilized the other type of volatility specifications besides GARCH
family such as the ARJI-GARCH-based model (hereafter ARJI) of Chan and
Maheu (2002) which combines the GARCH specification of volatility and
autoregressive jump intensity (ARJI) in jump intensity (see Su and Hung (2011)
Chang et al. (2011), and so on). Moreover, the other types of long memory volatility
specifications such as fractional integrated APARCH (FIAPARCH) and hyperbolic
1402 C.-F. Lee and J.-B. Su

GARCH (HYGARCH) besides FIGARCH mentioned above are also used to estimate
the VaR (see Aloui and Mabrouk (2010), Degiannakis et al. (2012)). As to distribution
setting, some special distributions like the Weibull distribution (see Gebizlioglu
et al. (2011)), the asymmetric Laplace distribution (see Chen et al. (2012)), and the
Pearson type-IV distribution (see Stavroyiannis et al. (2012)) are also employed to
estimate VaR.
The historical simulation assumes that the past will exactly replicate the future.
The VaR calculation of this approach is literally ranking all of your past historical
returns in terms of lowest to highest and computing with a predetermined confi-
dence rate what your lowest return historically has been.1 In addition, several
studies such as Vlaar (2000), Gencay et al. (2003), Cabedo and Moya (2003), and
Lu et al. (2009) had applied this approach to estimate the VaR. Even though it is
relatively easy to implement, there is a couple of shortcomings of this approach, and
first of all is that it imposes a restriction on the estimation assuming asset returns are
independent and identically distributed (iid) which is not the case. From empirical
evidence, it is known that asset returns are clearly not independent as they exhibit
volatility clustering.2 Therefore, it can be unrealistic to assume iid asset returns.
Second restriction relates to time. Historical simulation applies equal weight to all
returns of the whole period, and this is inconsistent with the nature where there is
diminishing predictability of data that are further away from the present.
These two shortcomings of historical simulation lead this paper to use the
approach proposed by Hull and White (1998) (hereafter, HW method) as
a representative of the semi-parametric approach. This semi-parametric approach
combines the abovementioned parametric approach of GARCH-based variance
specification with the weighted historical simulation. The weighted historical
simulation applies decreasing weights to returns that are further away from
the present, which overcomes the inconsistency of historical simulation with
diminishing predictability of data that are further away from the present. Hence,
this study utilizes the parametric approach (GARCH-N, GARCH-T, and GARCH-
SGT models) and the semi-parametric approach of Hull and White (1998) (HW-N,
HW-T, and HW-SGT models), totaling six models, to estimate the VaR for the
eight stock indices in Europe and Asia stock markets, then uses three accuracy
measures: one likelihood ratio test (the unconditional coverage test (LRuc)
of Kupiec (1995)) and two loss functions (the average quadratic loss function
(AQLF) of Lopez (1999) and the unexpected loss (UL)) to compare the forecasting
ability of the aforementioned models in terms of VaR.
Our results show that the kind of VaR approaches is more influential than that
of return distribution settings on VaR estimate. Moreover, under the same return
distributional setting, the HW-based models have the better VaR forecasting

1
This means if you had 200 past returns and you wanted to know with 99 % confidence what’s the
worst you can do, you would go to the 2nd data point on your ranked series and know that 99 % of
the time you will do no worse than this amount.
2
Large changes tend to be followed by large changes, of either sign, and small changes tend to be
followed by small changes.
51 Valuet-isk Estimation via a Semi-rametric Approach 1403

performance as compared with the GARCH-based models. Furthermore, irrespective

of whether the GARCH-based model or HW-based model is employed, the skewed
generalized student’s t (SGT) has the best VaR forecasting performance followed
by student’s t, while the normal owns the worst VaR forecasting performance.
In addition, all models tend to underestimate the real market risk in most cases, but
the non-normal distributions (student’s t and SGT) and the semi-parametric approach
try to reverse the trend of underestimating.
The remainder of this paper is organized as follows. Section 51.2 describes
the methodology of two dissimilar VaR approaches (the parametric and semi-
parametric approaches) and the VaR calculations using these approaches.
Section 51.3 provides criteria for evaluating risk management, and Sect. 51.4
reports on and analyzes the empirical results of the out-of-sample VaR forecasting
performance. The final section makes some concluding remarks.

51.2 Empirical Methodology

In this paper, there are two approaches of calculating VaR to be introduced, that is, the
parametric method and the semi-parametric approach. Here, we use the GARCH(1,1)
model with three conditional distributions, namely, the normal, student’s t, and SGT
distributions, to estimate the corresponding volatility in terms of different stock
indices then employ the framework of Jorion (2000) to evaluate the VaR of paramet-
ric approach whereas utilizing the weighting scheme of volatility proposed by Hull
and White (1998) (hereafter, HW method) which is a straightforward extension of
traditional historical simulation to calculate the VaR of semi-parametric VaR.

51.2.1 Parametric Method

Many time series data of financial assets appear to exhibit autocorrelated and vola-
tility clustering. Bollerslev et al. (1992) showed that the GARCH(1,1) specification
works well in most applied situations. Furthermore, the unconditional distribution of
those returns displays leptokurtosis and a moderate amount of skewness. Hence,
this study thus considers the applicability of the GARCH(1,1) model with three
conditional distributions, namely, the normal, student’s t, and SGT distributions, to
estimate the corresponding volatility in terms of different stock indices and use the
GARCH model as an official delegate of the VaR model.

51.2.1.1 GARCH Model with Normal Distribution

Let rt ¼ (ln Pt  ln Pt  1)  100, where Pt denotes the stock price and rt denotes the
continuously compounded daily returns of the underlying assets on time t. The GARCH
(1,1) model with SGT distribution (GARCH-SGT) can be expressed as follows:

rt ¼ m þ et , et ¼ et st ,  et  IID SGTð0; 1; k; l; nÞ (51.1)

1404 C.-F. Lee and J.-B. Su

s2t ¼ o þ ae2t1 þ bs2t1 (51.2)

where et is the current error and m and st2 are the conditional mean and
variance of return, respectively. Moreover, the variance parameters o, a, and b
are the parameters to be estimated and obey the constraints o, a, b > 0 and
a + b < 1. IID denotes that the standardized errors et are independent and
identically distributed. Since et is drawn from the standard normal distribution,
the probability density function for et is
 2
1 e
f ðet Þ ¼ pffiffiffiffiffiffi exp  t , (51.3)
2p 2

and the log-likelihood function of GARCH-N model thus can be written as

 
LðcÞ ¼ ln f ðrt jOt1 ; cÞ ¼ 0:5 ln 2p þ ln s2t þ ðrt  mÞ2 =s2t (51.4)
where c ¼ [m, o, a, b] is the vector of parameters to be estimated and Ot  1 denotes
the information set of all observed returns up to time t  1. Under the framework of
the parametric techniques (Jorion 2000), the 1-day-ahead VaR based on GARCH-N
model can be calculated as

VaRN ^ tþ1jt
tþ1jt ¼ m þ Fc ðet Þ  s (51.5)

where Fc(et) is the left-tailed quantile at c% for the standardized normal distribu-
^ tþ1jt is the one-step-ahead forecasts of the standard deviation of the returns
tion. s
conditional on all information upon the time t.

51.2.1.2 GARCH Model with Student’s t Distribution

Since the characteristics of many financial data are non-normal, the student’s t
distribution is most commonly employed to capture the fat-tailed properties of their
empirical distributions. Moreover, Bollerslev (1986) argued that using the student’s
t distribution as the conditional distribution for GARCH model is more satisfactory
since it exhibits thicker tail and larger kurtosis than normal distribution. Under
the same specifications of mean and variance equation as the GARCH-N model,
the probability density function for the standardized student’s t distribution can be
represented as follows:

 nþ1
Gð0:5ðn þ 1ÞÞ e2t 2

f ð et Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 þ (51.6)
Gð0:5 nÞ pðn  2Þ n2

where G(•) is the gamma function and n is the shape parameter. Hence, the
log-likelihood function of the GARCH-T model can be expressed as
51 Valuet-isk Estimation via a Semi-rametric Approach 1405

" #
Gð0:5ðn þ 1ÞÞ
LðcÞ ¼ ln f ðrt jOt1 ; cÞ ¼ ln pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Gð0:5nÞ pðn  2Þ
"   # (51.7)
nþ1 rt  m 2 1
ln st  ln 1 þ ð n  2Þ
2 st

where c ¼ [m, o, a, b, n] is the vector of parameters to be estimated. The 1-day-ahead

VaR based on GARCH-T model can be obtained as

^ tþ1jt
VaRTtþ1jt ¼ m þ Fc ðet ; nÞ  s (51.8)

where Fc(et ; n) denotes the left-tailed quantile at c% for standardized student’s

t distribution with shape parameter n.

51.2.1.3 GARCH Model with Skewed Generalized Student’s

t Distribution
This study also employs the SGT distribution of Theodossiou (1998) which allows
return innovation to follow a flexible treatment of both skewness and excess kurtosis in
the conditional distribution of returns. Under the same specifications of mean and
variance as the GARCH-N model, the probability density function for the standardized
SGT distribution is derived by Lee and Su (2011) and can be represented as follows:

 nþ1
j e t þ dj k k
f ð et Þ ¼ C 1 þ (51.9)
½1 þ sign ðet þ dÞlk yk

1
12 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where y ¼ Sð1lÞ B k1 ; kn 2 B k3 ; n2
k , SðlÞ ¼ 1 þ 3l2  4A2 l2 ,

1 n
0:5 3 n2
0:5 k
1 n
1
A ¼ B k2 ; n1
k B k;k B k; k , d ¼ S2lA
ðlÞ, C ¼ 2y B k ; k

where k, n, and l are scaling parameters and C and y are normalizing constants
ensuring that f(•) is a proper p.d.f. The parameters k and n control the height and tails
of density with constraints k > 0 and n > 2, respectively. The skewness parameter l
controls the rate of descent of the density around the mode of et with 1 < l < 1. In
the case of positive (resp. negative) skewness, the density function skews toward the
right (resp. left). Sign is the sign function, and B(•) is the beta function. The
parameter n has the degrees of freedom interpretation in case l ¼ 0 and
k ¼ 2. The log-likelihood function of the GARCH-SGT model thus can be written as
  k
nþ1 rt  m 
LðcÞ ¼ ln f ðrt jOt1 ; cÞ ¼ ln C  lnst  
ln 1 þ  þ d
k st
   k  (51.10)
rt  m k
1 þ sign þd l y
st
1406 C.-F. Lee and J.-B. Su

where c ¼ [m, o, a, b, k, l, n] is the vector of parameters to be estimated, and Ot  1

denotes the information set of all observed returns up to time t  1.
Under the framework of the parametric techniques (Jorion 2000), the 1-day-ahead
VaR based on GARCH-SGT model can be calculated as

VaRSGT ^ tþ1jt
tþ1jt ¼ m þ Fc ðet ; k; l; nÞ  s (51.11)

where Fc(et; k, l, n) denotes the left-tailed quantile at c% for standardized SGT

distribution with shape parameters k, l, and n and can be evaluated by a numerical
integral method (composite trapezoid rule).3 Particularly, the SGT distribution
generates the student’s t distribution for l ¼ 0 and k ¼ 2. Moreover, the SGT
distribution generates the normal distribution for l ¼ 0, k ¼ 2, and n ¼ 1.

51.2.2 Semi-parametric Method

In this paper, we use the approach proposed by Hull and White (1998) (hereafter, HW
method) as a representative of the semi-parametric approach. The HW method is
a straightforward extension of traditional historical simulation. Instead of using the
actual historical percentage changes in market variables for the purposes of calculat-
ing VaR, we use historical changes that have been adjusted to reflect the ratio of the
current daily volatility to the daily volatility at the time of the observation and assume
that the variance of each market variable during the period covered by the historical
data is monitored using a GARCH model. The methodology is explained in the
following three steps: First, use a raw return series, {r1, r2, r3,......, rt¼T}, to fit the
GARCH(1,1) models with alternative distributions expressed as in Sect. 51.2.1. Thus,
a series of daily volatility estimates, {s1, s2, s3,......, st¼T}, are obtained where T is
the number of estimated samples. Second, the modified return series are obtained by
the raw return series multiplied by the ratio of the current daily volatility to the daily
volatility at the time of the observation, sT/si. That is, the modified return series are
expressed as {r1*, r2*, r3*,......, rt¼T*}, where ri* ¼ ri(sT/si). Finally, sort the returns
ascendingly to achieve the empirical distribution. Thus, VaR is the percentile that
corresponds to the specified confidence level.
The HW-GARCH-SGT model (simply called HW-SGT) implies that the
standardized residual return of the GARCH-SGT model is applied by the HW
approach to estimate the VaR so are HW-N and HW-T models.

51.3 Evaluation Methods of Model-Based VaR

Many financial institutions have been required to hold capital against their market risk
exposure, while the market risk capital requirements are based on the VaR estimates

3
See Faires and Burden (2003) for more details.
51 Valuet-isk Estimation via a Semi-rametric Approach 1407

generated by the financial institutions’ own risk management models. Explicitly, the
accuracy of these VaR estimates is of concern to both financial institutions and their
regulators. Hence, model accuracy is important to all VaR model users. To compare
the forecasting ability of the aforementioned models in terms of VaR, this study
considers three accuracy measures: the unconditional coverage test of Kupiec (1995)
which are quite standard in the literatures. Moreover, the quadratic loss function and
the unexpected loss are introduced and used for determining the accuracy of model-
based VaR measurements.

51.3.1 Log-Likelihood Ratio Test

Before performance competition of alternative VaR models, with regard to two

models, we can use the log-likelihood ratio test to compare which one model has the
better matching ability of between actual data and empirical model. This can be
regarded as the preliminary analysis. The log-likelihood ratio test is a statistical test
used to compare the fit of two models, one of which, the null model, is a special
case of the other, the alternative model. The test is based on the likelihood ratio,
which expresses how many times more likely the data are under one model than the
other. This log-likelihood ratio can then be used to compute a p-value, or compared
to a critical value, to decide whether to reject the null model in favor of the
alternative model.
The log-likelihood ratio test LRN (LRT), used to test the null hypothesis that
log-returns are normally (student’s t) distributed against the alternative hypothesis,
is given by

LRN or LRT ¼ 2ðLRr  LRu Þ  w2 ðmÞ (51.12)

where LRr and LRu are, respectively, the maximum value of the log-likelihood
values under the null hypothesis of the restricted model and the alternative
hypothesis of the unrestricted model and m is the number of the restricted
parameters in the restricted model. For example, LRN for GARCH-SGT model
could be used to test the null hypothesis that log-returns are normally distributed
against the alternative hypothesis that they are SGT distributed. The null hypoth-
esis for testing normality is H0:k ¼ 2, l ¼ 0 and n!1, and the alternative
hypothesis is H1:k∊R+, n > 2 and |l| < 1. Restate, LRN ¼ 2(LRr  LRu) 
w2(3) where LRr and LRu are, respectively, the maximum value of the
log-likelihood values under the null hypothesis of restricted model (GARCH-N
model) and the alternative hypothesis of unrestricted model (GARCH-SGT
model) and m is the number of the restricted parameters in the restricted model
(k ¼ 2, l ¼ 0 and n!1) and equal to 3 in this case. At the same inference, LRN
for GARCH-T model follows the w2(1) distribution with one degree of freedom.
Moreover, LRT for GARCH-SGT model follows the w2(2) distribution with two
degrees of freedom.
1408 C.-F. Lee and J.-B. Su

51.3.2 Binary Loss Function or Failure Rate

If the predicted VaR is not able to cover the realized loss, this is termed a violation.
A binary loss function (BLF) is merely the reflection of the LR test of unconditional
coverage test and gives a penalty of one to each exception of the VaR. The BLF for
long position can be defined as follows:

1 if rtþ1 < VaRtþ1jt ,
BLtþ1 ¼ (51.13)
0 if rtþ1  VaRtþ1jt :

where BLt+1 represents the 1-day-ahead BLF for long position. If a VaR model truly
provides the level of coverage defined by its confidence level, then the average
binary loss function (ABLF) or the failure rate over the full sample will equal c for
the (1  c)th percentile VaR.

51.3.3 Quadratic Loss Function

The quadratic loss function (QLF) of Lopez (1999) penalizes violations differently
from the binary loss function and pays attention to the magnitude of the violation.
The QLF for long position can be expressed as


2
1 þ rtþ1  VaRtþ1jt if rtþ1 < VaRtþ1jt ,
QLtþ1 ¼ (51.14)
0 if rtþ1  VaRtþ1jt :
where QLt+1 represents the 1-day-ahead QLF for long position. The quadratic term
in Eq. 51.14 ensures that large violations are penalized more than the small
violations which provides a more powerful measure of model accuracy than the
binary loss function.

51.3.4 The Unconditional Coverage Test

Kupiec (1995) proposes the unconditional coverage test which is a likelihood

ratio test for testing the model accuracy which is identical to a test of the null
p) equals the specified model
hypothesis that the probability of failure for each trial (^
probability (p). The likelihood ratio test statistics is given by

LRuc ¼ 2lnðpn1 ð1  pÞn0 p ^ Þn0 Þ  w2 ð1Þ

^ n1 ð1  p (51.15)

^ ¼ n0nþn1 1 is the maximum likelihood estimate of p, n1 denotes a Bernoulli

where p
random variable representing the total number of VaR violations, and n0 + n1
51 Valuet-isk Estimation via a Semi-rametric Approach 1409

represents the full sample size. The LRuc test can be employed to test whether the
sample point estimate is statistically consistent with the VaR model’s prescribed
confidence level or not.

51.3.5 Unexpected Loss

The unexpected loss (UL) will equal the average magnitude of the violation over
the full sample. The magnitude of the violation for long position is given by

rtþ1  VaRtþ1jt if rtþ1 < VaRtþ1jt ,
Ltþ1 ¼ (51.16)
0 if rtþ1  VaRtþ1jt :

where Lt+1 is the 1-day-ahead magnitude of the violation for long position.

51.4 Empirical Results

The study data comprises daily prices of the following eight stock indices: the
Austria ATX (6/29/1999–8/10/2009), the Belgium Brussels (10/19/1999–
8/10/2009), the France CAC40 (10/22/1999–8/10/2009) and the Switzerland
Swiss (9/8/1999–8/10/2009) in Europe, the India Bombay (7/8/1999–8/10/2009),
the Malaysia KLSE (6/23/1999–8/10/2009), the South Korea KOSPI (6/21/1999–
8/10/2009), and the Singapore STRAITS (8/24/1999–8/10/2009) in Asia, where the
numbers in parentheses are the start and end dates for our sample. Daily closing
spot prices for the study period, totaling 2,500 observations, were obtained from
http://finance.yahoo.com. The stock returns are defined as the first difference in the
logarithms of daily stock prices then multiplied by 100.

51.4.1 Data Preliminary Analysis

Table 51.1 summarizes the basic statistical characteristics of return series for both
the estimation and forecast periods. Notably, the average daily returns are all
negative (resp. positive) for forecast (resp. estimation) period and very small
compared with the variable standard deviation, indicating high volatility. Except
the Brussels of estimation period and the CAC40, Swiss and Bombay of forecast
period, all returns series almost exhibit negative skewness for both the estimation
and forecast periods. The excess kurtosis all significantly exceeds zero at the 1 %
level, indicating a leptokurtic characteristic. Furthermore, J-B normality test statis-
tics are all significant at the 1 % level and thus reject the hypothesis of normality
and confirm that neither return series is normally distributed. Moreover, the
Ljung-Box Q2(20) statistics for the squared returns are all significant at the 1 %
1410 C.-F. Lee and J.-B. Su

Table 51.1 Descriptive statistics of daily return

Mean Std. dev. Max. Min. Skewness Kurtosis J-B Q2(20)
Panel A. Estimation period (2,000 observations)
ATX 0.0675 0.9680 4.6719 7.7676 0.6673c 4.4547c 1,802.17c 547.23c
Brussels 0.0179 1.1577 9.3339 5.6102 0.2607c 6.0567c 3,079.66c 1,479.84c
CAC40 0.0090 1.4012 7.0022 7.6780 0.0987a 2.9924c 749.48c 2,270.73c
Swiss 0.0086 1.1602 6.4872 5.7803 0.0530 4.5084c 1,694.78c 1,985.51c
Bombay 0.0648 1.5379 7.9310 11.8091 0.5632c 4.2350c 1,600.43c 707.26c
KLSE 0.0243 0.9842 5.8504 6.3422 0.3765c 6.2537c 3,306.35c 5,54.16c
KOSPI 0.0397 1.8705 7.6971 12.8046 0.4671c 3.6010c 1,153.39c 365.29c
STRAITS 0.0239 1.1282 4.9052 9.0949 0.5864c 4.8254c 2,055.03c 239.53c
Panel B. Forecast period (500 observations)
ATX 0.1352 2.6532 12.0210 10.2526 0.0360 2.4225c 122.37c 735.66c
Brussels 0.1195 1.9792 9.2212 8.3192 0.0888 3.1545c 207.97c 581.42c
CAC40 0.0907 2.1564 10.5945 9.4715 0.2209b 4.4068c 408.65c 353.15c
Swiss 0.0704 1.8221 10.7876 8.1077 0.2427b 4.4101c 410.10c 502.59c
Bombay 0.0101 2.6043 15.9899 11.6044 0.2529b 3.4248c 249.70c 57.24c
KLSE 0.0166 1.2040 4.2586 9.9785 1.1163c 9.8218c 2,113.64c 17.87c
KOSPI 0.0327 2.1622 11.2843 11.1720 0.4177c 4.3977c 417.47c 343.50c
STRAITS 0.0594 2.0317 7.5305 9.2155 0.1183 2.3827c 119.45c 219.52c
Notes: 1. a, b, and c denote significantly at the 10 %, 5 %, and 1 % levels, respectively. 2. J-B statistics
are based on Jarque and Bera (1987) and are asymptotically chi-squared distributed with 2 degrees of
freedom. 3. Q2(20) statistics are asymptotically chi-squared distributed with 20 degrees of freedom

level and thus indicate that the return series exhibit linear dependence and strong
ARCH effects. Therefore, the preliminary analysis of the data suggests the use of
a GARCH model to capture the fat-tails and time-varying volatility found in these
stock indices return series.
Descriptive graphs (levels of spot prices, density of the daily returns against
normal distribution) for each stock index are illustrated in Fig. 51.1a–h. As shown
in Fig. 51.1, all stock indices have experienced a severe slide in price levels and
display pictures of volatile bear markets for forecast period. Moreover, comparing
density graphs against the normal distribution shows that each return distribution of
data employed exhibits non-normal characteristics. This provides evidence in favor
of some of skewed, leptokurtic, and fat-tailed return distributions. These results are
in line with those of Table 51.1.

51.4.2 Estimation Results for Alternate VaR Models

This section estimates the GARCH(1,1) model with alternative distributions

(normal, student’s t, and SGT) for performing VaR analysis. For each data series,
three GARCH models are estimated with a sample of 2,000 daily returns, and the
estimation period is then rolled forwards by adding one new day and dropping the
51 Valuet-isk Estimation via a Semi-rametric Approach 1411

a ATX stock index : overall-sample daily return density for overall

period period(ATX)
5000 1.00 0.6
CLOSE Mean 0.02702
4500 0.75
0.5 Std Error 1.47034
4000 0.50 Skewness −0.38145
3500 0.25 0.4 Exc Kurtosis 9.57690
3000 0.00 0.3
2500 −0.25
0.2
2000 −0.50
1500 −0.75 0.1
1000 −1.00 0.0
−15 −10 −5
99
00
01
02
03
04
05
06
07
08
09
0 5 10 15
19
20
20
20
20
20
20
20
20
20
20 Date
b Brussel stock index : overall-sample daily return density for overall
period period(Brussel)
5000 1.00 0.6
CLOSE Mean −0.00951
4500 0.75
0.5 Std Error 1.36291
4000 0.50 Skewness 0.00652
3500 0.25 0.4
Exc Kurtosis 6.24403
3000 0.00 0.3
2500 −0.25 0.2
2000 −0.50
0.1
1500 −0.75
1000 −1.00 0.0
−10 −5 0 5 10 15
00
01
02
03
04
05
06
07
08
09
20
20
20
20
20
20
20
20
20
20

Date
c CAC40 stock index : overall-sample daily return density for overall
period period(CAC40)
7000 1.00 0.45
CLOSE Mean −0.01090
0.75 0.40
6000 Std Error 1.58137
0.50 0.35
Skewness 0.02402
0.25 0.30 Exc Kurtosis 5.01027
5000 0.25
0.00
0.20
4000 −0.25 0.15
−0.50 0.10
3000
−0.75 0.05
2000 −1.00 0.00
−10 −5 0 5 10 15
00
01
02
03
04
05
06
07
08
09
20
20
20
20
20
20
20
20
20
20

Date
d Swiss stock index : overall-sample
daily return density for overall
period
10000 1.00 period(Swiss)
CLOSE 0.6
9000 0.75 Mean −0.00717
0.50 0.5 Std Error 1.31938
8000 Skewness 0.06587
0.25 0.4
7000 Exc Kurtosis 5.91237
0.00 0.3
6000
−0.25
0.2
5000 −0.50
0.1
4000 −0.75
−1.00 0.0
3000
−10 −5 0 5 10 15
00
01
02
03
04
05
06
07
08
09
20
20
20
20
20
20
20
20
20
20

Date

Fig. 51.1 (continued)

1412 C.-F. Lee and J.-B. Su

e Bombay stock index : overall-sample daily return density for overall

period period(Bombay)
22500 1.00 0.40
20000 CLOSE 0.75 Mean 0.04980
0.35
Std Error 1.80202
17500 0.50 0.30
Skewness −0.15519
15000 0.25 0.25
Exc Kurtosis 5.59973
12500 0.00 0.20
10000 −0.25 0.15
7500 −0.50 0.10
5000 −0.75 0.05
2500 −1.00 0.00
−20 −10 0 10 20 30
20 9
00

20 1
02

20 3
04

20 5
20 6
07

20 8
09
9

0
0

0
19

20

20

20

20
Date
f KLSE stock index : overall-sample daily return density for overall
period period(KLSE)
1600 1.00 0.7
CLOSE 0.75 Mean 0.01615
1400 0.6
0.50 Std Error 1.03179
1200 0.5 Skewness −0.62305
0.25 Exc Kurtosis 7.87111
0.4
1000 0.00
0.3
−0.25
800 0.2
−0.50
600 −0.75 0.1
400 −1.00 0.0
−15 −10 −5 0 5 10
99
00
01
02
03
04
05
06
07
08
09
19
20
20
20
20
20
20
20
20
20
20

Date
g KOSPI stock index : overall-sample daily return density for overall
period period(KOSPI)
2200 1.00 0.35
2000 CLOSE 0.75 Mean 0.02528
0.30 Std Error 1.93219
1800 0.50
0.25 Skewness −0.46118
1600 0.25 Exc Kurtosis 3.94058
1400 0.00 0.20
1200
−0.25 0.15
1000
800 −0.50 0.10
600 −0.75 0.05
400 −1.00 0.00
99
00
01
02
03
04
05
06
07
08
09

−20 −10 0 10 20 30
19
20
20
20
20
20
20
20
20
20
20

Date
h STRAITS stock index : overall-sample daily return density for overall
period period(STRAITS)
4000 1.00 0.6
CLOSE 0.75 Mean 0.00730
3500 0.5 Std Error 1.35780
0.50
3000 Skewness −0.39368
0.25 0.4
Exc Kurtosis 5.34473
2500 0.00 0.3
−0.25
2000 0.2
−0.50
1500 −0.75 0.1
1000 −1.00
0.0
99
00
01
02
03
04
05
06
07
08
09

−10 −5 0 5 10 15
19
20
20
20
20
20
20
20
20
20
20

Date

Fig. 51.1 The stock index in level and daily return density (versus normal) for whole sample (a) ATX,
(b) Brussels, (c) CAC40, (d) Swiss, (e) Bombay, (f) KLSE, (g) KOSPI, (h) STRAITS stock indices
51 Valuet-isk Estimation via a Semi-rametric Approach 1413

most distant day. In this procedure, according to the theory of Sect. 51.2, the
out-of-sample VaR is computed for the next 500 days.
Table 51.2 lists the estimation results4 of the GARCH-N, GARCH-T, and
GARCH-SGT models for the ATX, Brussels, CAC40, and Swiss stock indices in
Europe and the Bombay, KLSE, STRAITS, and KOSPI stock indices in Asia during
the first in-sample period. The variance coefficients o, a, and b are all positive and
significant almost at the 1 % level. Furthermore, the sums of parameters a and b for
these three models are less than one thus ensuring that the conditions for stationary
covariance hold. As to the fat-tail parameters in student’s t distribution, the fat-tail
parameter (n) ranges from 4.9906 (KLSE) to 14.9758 (CAC40) for GARCH-T
model. All these shape parameters are all significant at 1 % level and obey the
constraint n > 2 and thereby implying that the distribution of returns has larger,
thicker tails than the normal distribution. Turning to the shape parameters in SGT
distribution, the fat-tail parameter (n) ranges from 4.9846 (KLSE) to 21.4744
(KOSPI), and the fat-tail parameter (k) is between 1.5399 (KOSPI) and 2.3917
(Bombay). The skewness parameter (l) ranges from 0.1560 (Bombay) to
0.0044(KLSE). Moreover, these three coefficients are almost significant at the
1 % level and thereby these negative skewness parameters imply that the distribu-
tion of returns has a leftward tail. Therefore, both fat-tails and skewness cannot be
ignored in modeling these stock indices returns. The Ljung-Box Q2(20) statistics
for the squared returns are all not significant at the 10 % level and thus indicate that
serial correlation does not exist in standard residuals, confirming that the GARCH
(1,1) specification in these models is sufficient to correct the serial correlation of
these eight return series in the conditional variance equation.
Moreover, as shown in Table 51.2, the LRN statistics for both GARCH-T and
GARCH-SGT models are all significant at the 1 % level, indicating that the null
hypothesis of normality for either stock index is rejected. These results thus imply that
both the student’s t and SGT distributions closely approximate the empirical return
series as compared with the normal distribution. Furthermore, except for ATX and
KLSE stock indices, the LRT statistics of GARCH-SGT model are all significant,
implying that the SGT distribution more closely approximates the empirical return
series than the student’s t does. To sum up, the SGT distribution closely approximates
the empirical return series followed by student’s t and normal distributions.

51.4.3 The Results of VaR Performance Assessment

In this paper, we utilize the parametric approach (GARCH-N, GARCH-T, and

GARCH-SGT models) and the semi-parametric approach (HW-N, HW-T, and
HW-SGT models), totaling six models, to estimate the VaR.; thereafter, it was
compared with the observed return, and both results were recorded. This section

4
The parameters are estimated by QMLE (quasi-maximum likelihood estimation; QMLE) and the
BFGS optimization algorithm, using the econometric package of WinRATS 6.1.
1414 C.-F. Lee and J.-B. Su

Table 51.2 Estimation results for alternative models (estimation period)

ATX Brussels CAC40 Swiss
Panel A. GARCH(1,1) with normal distribution
m 0.1007c(0.0202) 0.0760c(0.0177) 0.0537b(0.0227) 0.0488c(0.0187)
o c
0.0612 (0.0124) 0.0239c(0.0032) 0.0143c(0.0029) 0.0243c(0.0057)
a c
0.1146 (0.0158) 0.1470c(0.0090) 0.0799c(0.0100) 0.1184c(0.0136)
b c
0.8209 (0.0227) 0.8354c(0.0034) 0.9132c(0.0105) 0.8632c(0.0150)
2
Q (20) 16.147 14.731 22.333 19.883
LL 2648.73 2649.75 3156.40 2748.86
Panel B. GARCH(1,1) with student’s t distribution
m 0.0998c(0.0178) 0.0775c(0.0162) 0.0622c(0.0216) 0.0589c(0.0167)
o c
0.0623 (0.0160) 0.0179c(0.0049) 0.0118c(0.0043) 0.0177c(0.0052)
a c
0.0986 (0.0188) 0.1319c(0.0177) 0.0785c(0.0110) 0.1078c(0.0151)
b c
0.8324 (0.0292) 0.8560c(0.0180) 0.9166c(0.0109) 0.8799c(0.0153)
c
n 7.3393 (1.0609) 9.4946c(1.7035) 14.9758c(4.0264) 9.6205c(1.6621)
2
Q (20) 17.334 19.676 21.719 18.712
LL 2610.15 2626.31 3146.96 2724.27
LRN 77.16c 46.88c 18.88c 49.18c
Panel C. GARCH(1,1) with SGT distribution
m 0.0875c(0.0177) 0.0691c(0.0158) 0.0525b(0.0217) 0.0479c(0.0175)
o c
0.0626 (0.0173) 0.0175c(0.0044) 0.0115c(0.0043) 0.0172c(0.0048)
a c
0.0952 (0.0189) 0.1277c(0.0163) 0.0774c(0.0103) 0.1086c(0.0148)
b 0.8343c(0.0323) 0.8590c(0.0161) 0.9170c(0.0106) 0.8787c(0.0150)
n 7.8001c(2.4847) 6.9261c(1.7169) 13.3004b(6.3162) 7.8987c(2.1709)
l 0.0660 (0.0290) 0.1019c(0.0346)
b
0.1175c(0.0335) 0.1175c(0.0323)
k 1.9710c(0.2290) 2.3745c(0.2782) 2.1219c(0.2220) 2.2601c(0.2519)
2
Q (20) 17.779 20.509 21.803 18.791
LL 2608.01 2621.51 3140.58 2717.94
LRN (LRT) 81.44c(4.28) 56.48c(9.6c) 31.64c(12.76c) 61.84c(12.66c)
Bombay KLSE KOSPI STRAITS
Panel A. GARCH(1,1) with normal distribution
m 0.1427c(0.0262) 0.0453c(0.0164) 0.1212c(0.0318) 0.0623c(0.0196)
o c
0.0906 (0.0201) 0.0077c(0.0029) 0.0214c(0.0082) 0.0143c(0.0042)
a c
0.1438 (0.0167) 0.0998c(0.0174) 0.0799c(0.0146) 0.1031c(0.0134)
b c
0.8189 (0.0206) 0.8989c(0.0165) 0.9177c(0.0141) 0.8938c(0.0123)
2
Q (20) 19.954 27.905 11.214 15.574
LL 3453.18 2490.05 3843.01 2895.36
Panel B. GARCH(1,1) with student’s t distribution
m 0.1583c(0.0250) 0.0351b(0.0142) 0.1377c(0.0302) 0.0652c(0.0192)
o c
0.0863 (0.0222) 0.0116b(0.0050) 0.0163b(0.0078) 0.0135c(0.0049)
a c
0.1417 (0.0198) 0.1116c(0.0254) 0.0639c(0.0128) 0.0806c(0.0136)
b c
0.8225 (0.0234) 0.8848c(0.0248) 0.9332c(0.0127) 0.9119c(0.0139)
c
n 8.3410 (1.3143) 4.9906c(0.5782) 7.2792c(1.1365) 6.7643c(0.9319)
(continued)
51 Valuet-isk Estimation via a Semi-rametric Approach 1415

Table 51.2 (continued)

Bombay KLSE KOSPI STRAITS
Q2(20) 19.980 24.477 11.304 16.554
LL 3420.18 2412.82 3801.67 2838.34
LRN 66.0c 154.46c 82.68c 114.04c
Panel C. GARCH(1,1) with SGT distribution
m 0.1266c(0.0261) 0.0341b(0.0147) 0.1021c(0.0285) 0.0516c(0.0185)
o c
0.0836 (0.0201) 0.0116b(0.0049) 0.0167b(0.0077) 0.0132c(0.0045)
a c
0.1350 (0.0196) 0.1117c(0.0242) 0.0613c(0.0133) 0.0785c(0.0135)
b c
0.8282 (0.0228) 0.8847c(0.0240) 0.9345c(0.0135) 0.9138c(0.0136)
c
n 6.2282 (1.3602) 4.9846c(1.0922) 21.4744(15.8310) 6.2641c(1.4297)
l 0.1560 (0.0303) 0.0044(0.0281)
c
0.1006c(0.0266) 0.0745b(0.0296)
k 2.3917c(0.2801) 2.0016c(0.2450) 1.5399c(0.1513) 2.1194c(0.2256)
2
Q (20) 21.167 24.455 11.067 16.595
LL 3408.46 2412.81 3791.66 2835.52
LRN (LRT) 89.44c(23.44c) 154.48c(0.02) 102.7c(20.02c) 119.68c(5.64a)
Notes: 1.a, b andc denote significantly at the 10 %, 5 %, and 1 % levels, respectively. 2. Numbers in
parentheses are standard errors. 3. LL indicates the log-likelihood value. 4. The critical value of the
LRN test statistics at the 1 % significance level is 6.635 for GARCH-T and 11.345 for GARCH-SGT
model. 5. The critical value of the LRT test statistics at the 10 %, 5 %, and 1 % significance level is
4.605, 5.991, and 9.210 for GARCH-SGT model, respectively. 6. Q2(20) statistics are asymptotically
chi-squared distributed with 20 degrees of freedom

then uses three accuracy measures: one likelihood ratio test (the unconditional
coverage test (LRuc) of Kupiec (1995)) and two loss functions (the average qua-
dratic loss function (AQLF) of Lopez (1999) and the unexpected loss (UL)) to
compare the forecasting ability of the aforementioned models in terms of VaR.
Figure 51.2 graphically illustrates the long VaR forecasts of the GARCH-N,
GARCH-T, and GARCH-SGT models at alternate levels (95 %, 99 %, and 99.5 %)
for all stock indices. Tables 51.3, 51.4 and 51.5 provide the failure rates and the results
of the prior three accuracy evaluation tests (LRuc, AQLF, and UL) for the aforemen-
tioned six models at the 95 %, 99 %, and 99.5 % confidence levels, respectively. As
observed in Tables 51.3, 51.4 and 51.5, we find that, except for a few cases at the 99 %
and 99.5 % confidence levels, all models tend to underestimate real market risk
because the empirical failure rate is higher than the theoretical failure rate in most
cases. The abovementioned exceptional cases emerge at the GARCH-SGT model
of 99 % level (CAC40); both the GARCH-T and GARCH-SGT models of 99.5 %
level (KLSE and STRAITS); the HW-N (KLSE), HW-T (KLSE), and HW-SGT
(KLSE and STRAITS) models of 99 % level; and the HW-N (STRAITS), HW-T
(ATX and STRAITS), and HW-SGT (ATX, KLSE, and STRAITS) models of 99.5 %
level, where the stock indices in parentheses behind the models are the exceptional
cases. Moreover, the empirical failure rate of the above exceptional cases is
lower than the theoretical failure rate, indicating that the non-normal distributions
(student’s t and SGT) and the semi-parametric approach try to reverse the trend of
underestimating real market risk, especially at the 99.5 % level.
1416 C.-F. Lee and J.-B. Su

ATX index: VaR at Brussels index: VaR at

a alternate levels
b
15 15 alternate levels
15 15

10 10 10 10

5 5 5 5

0 0 0 0
Return

Return

Return

Return
-5 −5 −5 −5

−10 −10 −10 −10

−15 −15 −15 −15

−20 −20 −20 −20

2008 2009 2008 2009
Date Date
CAC40 index: VaR at Swiss index: VaR at
c alternate levels d alternate levels
15 15 15 15

10 10 10 10

5 5 5 5

0 0
Return

Return
0 0
Return

Return

−5 −5 −5 −5

−10 −10 −10 −10

−15 −15
−15 −15

−20 −20
−20 −20
2008 2009
2008 2009
Date
Date
f KLSE index: VaR at
e Bombay index: VaR at alternate levels
alternate levels 5.0 5.0
15 15
2.5 2.5
10 10
0.0 0.0
5 5 −2.5 −2.5

−5.0 −5.0
Return

Return

0 0
Return

Return

−7.5 −7.5
−5 −5
−10.0 −10.0
−10 −10
−12.5 −12.5
−15 −15 −15.0 −15.0

−20 −20 −17.5 −17.5

2008 2009 2008 2009
Date Date

Fig. 51.2 (continued)

51 Valuet-isk Estimation via a Semi-rametric Approach 1417

g KOSPI index: VaR at h STRAITS index: VaR at

alternate levels alternate levels
15 15 15 15

10 10 10 10

5 5 5 5

0 0 0 0

Return

Return
Return

Return
−5 −5 −5 −5

−10 −10 −10 −10

−15 −15 −15 −15

−20 −20 −20 −20

2008 2009 2008 2009
Date Date

Return GARCH_SGT 95% GARCH_SGT 99% GARCH_SGT 995%

GARCH_N 95% GARCH_N 99% GARCH_N 995%
GARCH_T 95% GARCH_T 99% GARCH_T 995%

Fig. 51.2 Long VaR forecasts at alternative level for the normal, student’s t, and SGT distribu-
tion. (a) ATX, (b) Brussels, (c) CAC40, (d) Swiss, (e) Bombay, (f) KLSE, (g) KOSPI,
(h) STRAITS stock indices

As to the back-testing, the back-testing is a specific type of historical

testing that determines the performance of the strategy if it had actually been
employed during the past periods and market conditions. In this paper, the
unconditional coverage tests (LRuc) proposed by Kupiec (1995) is employed to
test whether the unconditional coverage rate is statistically consistent with the
VaR model’s prescribed confidence level and thus is applied as the back-testing
to measure the accuracy performance of these six VaR models. To interpret
the result of accepting back-testing in Tables 51.3, 51.4 and 51.5, there is
an illustration in the following. In Table 51.3, the VaR estimates based on
GARCH-N, GARCH-T, and GARCH-SGT models, respectively, have a total
of 2 (KLSE and STRAITS), 2 (KLSE and STRAITS), and 5 (Brussels, Bombay,
KLSE, STRAITS, and KOSPI) acceptances for the LRuc test when applying to all
stock indices returns under 95 % confidence level, where the stock indices in
parentheses behind the number are the acceptance cases. For 99 % confidence
level, Table 51.4 shows that the GARCH-N, GARCH-T, and GARCH-SGT
models pass the LRuc tests with a total of 1, 5, and 8 stock indices, respectively;
for 99.5 % confidence level, Table 51.5 gives that the GARCH-N, GARCH-T,
and GARCH-SGT models pass the LRuc tests with a total of 3, 7, and 7 stock
indices, respectively. Hence, under all confidence levels, there is a total of 6, 14,
and 20 acceptances for GARCH-N, GARCH-T, and GARCH-SGT models
(the parametric approach), respectively. On the contrary, for 95 % confidence
1418 C.-F. Lee and J.-B. Su

Table 51.3 Out-of-sample long VaR performance at the 95 % confidence level

GARCH HW-GARCH
Failure rate (LRuc) AQLF UL Failure rate (LRuc) AQLF UL
Panel A. ATX
N 0.0960(17.75) 0.30360 –0.10151 0.0960(17.75) 0.30642 –0.10308
T 0.0960(17.75) 0.3 2058 –0.10865 0.0920(15.04) 0.30342 –0.10384
SGT 0.0900(13.75) 0.29877 –0.10212 0.0980(19.18) 0.32263 –0.10858
Panel B. Brussels
N 0.0780(7.10) 0.20246 –0.07099 0.0680(3.08*) 0.18374 –0.06673
T 0.0720(4.51) 0.19494 –0.07020 0.0620(1.41*) 0.16978 –0.06268
SGT 0.0660(2.45*) 0.18178 –0.06602 0.0620(1.41*) 0.17067 –0.06262
Panel C. CAC40
N 0.0740(5.31) 0.21446 –0.06485 0.0700(3.76*) 0.20928 –0.05844
T 0.0800(8.07) 0.22598 –0.06887 0.0720(4.51) 0.19510 –0.05617
SGT 0.0760(6.18) 0.21146 –0.06281 0.0700(3.76*) 0.19390 –0.05600
Panel D. Swiss
N 0.0760(6.18) 0.18567 –0.06181 0.0620(1.41*) 0.15104 –0.05088
T 0.0740(5.31) 0.18729 –0.06345 0.0600(0.99*) 0.14222 –0.04694
SGT 0.0740(5.31) 0.17506 –0.05698 0.0560(0.36*) 0.13928 –0.04697
Panel E. Bombay
N 0.0780(7.10) 0.34428 –0.10214 0.0800(8.07) 0.33643 –0.09967
T 0.0820(9.11) 0.35999 –0.10617 0.0780(7.10) 0.33182 –0.09639
SGT 0.0700(3.76*) 0.31488 –0.09366 0.0800(8.07) 0.33629 –0.09824

Panel F. KLSE
N 0.0560(0.36*) 0.20084 –0.04595 0.0740(5.31) 0.23255 –0.05623
T 0.0600(0.99*) 0.21827 –0.05152 0.0700(3.76*) 0.23146 –0.05276
SGT 0.0580(0.64*) 0.21375 –0.05007 0.0640(1.90*) 0.22378 –0.05306
Panel G. STRAITS
N 0.0580(0.64*) 0.21980 –0.06233 0.0640(1.90*) 0.24194 –0.06675
T 0.0640(1.90*) 0.25978 –0.07222 0.0620(1.41*) 0.23921 –0.06511
SGT 0.0620(1.41*) 0.24670 –0.06736 0.0560(0.36*) 0.23035 –0.06428
Panel H. KOSPI
N 0.0740(5.31) 0.27949 –0.08826 0.0600(0.99*) 0.24267 –0.07816
T 0.0760(6.18) 0.32360 –0.09624 0.0620(1.41*) 0.25234 –0.07522
SGT 0.0672(2.82*) 0.27616 –0.08236 0.0620(1.41*) 0.24475 –0.07531
Notes: 1. *Indicates that the model passes the unconditional coverage test at the 5 % significance
level and the critical value of the LRuc test statistics at the 5 % significance level is 3.84. 2. The red
(resp. blue) font represents the lowest (resp. highest) AQLF and unexpected loss when the
predictive accuracies of three different innovations with the same VaR method are compared.
3. The delete-line font represents the lowest AQLF and unexpected loss when the predictive
accuracies of two different VaR methods with the same innovation are compared. 4. The model
acronyms stand for the following methods: HW-GARCH non-parametric method proposed by Hull
and White (1998), GARCH parametric method of GARCH model, N the standard normal distri-
bution, T the standardized student’s t distribution, SGT the standardized SGT distribution proposed
by Theodossiou (1998)
51 Valuet-isk Estimation via a Semi-rametric Approach 1419

Table 51.4 Out-of-sample long VaR performance at the 99 % confidence level

GARCH HW-GARCH
Failure rate (LRuc) AQLF UL Failure rate (LRuc) AQLF UL
Panel A. ATX
N 0.0300(13.16) 0.20577 –0.02704 0.0160(1.53*) 0.06164 –0.01770
T 0.0200(3.91) 0.19791 –0.01986 0.0160(1.53*) 0.04977 –0.01551
SGT 0.0160(1.53*) 0.17702 –0.01725 0.0160(1.53*) 0.05279 –0.01626

Panel B. Brussels
N 0.0260(8.97) 0.13491 –0.02538 0.0180(2.61*) 0.03685 –0.01726
T 0.0180(2.61*) 0.11962 –0.01947 0.0160(1.53*) 0.03465 –0.01559
SGT 0.0160(1.53*) 0.11033 –0.01711 0.0160(1.53*) 0.03536 –0.01596

Panel C. CAC40
N 0.0200(3.91) 0.14794 –0.01913 0.0160(1.53*) 0.07472 –0.01808
T 0.0100(0.00*) 0.14325 –0.01494 0.0100(0.00*) 0.05628 –0.01436
SGT 0.0060(0.94*) 0.13144 –0.01354 0.0140(0.71*) 0.05992 –0.01466

Panel D. Swiss
N 0.0260(8.97) 0.12387 –0.01967 0.0180(2.61*) 0.03938 –0.01353
T 0.0160(1.53*) 0.11943 –0.01516 0.0140(0.71*) 0.03528 –0.01341
SGT 0.0140(0.71*) 0.10151 –0.01244 0.0140(0.71*) 0.03512 –0.01338

Panel E. Bombay
N 0.0300(13.16) 0.23811 –0.03748 0.0120(0.18*) 0.04961 –0.01751
T 0.0220(5.41) 0.22706 –0.02822 0.0120(0.18*) 0.04539 –0.01586
SGT 0.0180(2.61*) 0.19232 –0.02152 0.0120(0.18*) 0.04504 –0.01617

Panel F. KLSE
N 0.0160(1.53*) 0.16522 –0.01892 0.0080(0.21*) 0.07681 –0.01341
T 0.0100(0.00*) 0.16891 –0.01585 0.0060(0.94*) 0.08114 –0.01425
SGT 0.0100(0.00*) 0.16278 –0.01551 0.0060(0.94*) 0.07965 –0.01384

Panel G. STRAITS
N 0.0240(7.11) 0.16096 –0.01848 0.0120(0.18*) 0.09107 –0.01763
T 0.0100(0.00*) 0.17606 –0.01568 0.0100(0.00*) 0.07477 –0.01403
SGT 0.0100(0.00*) 0.16258 –0.01406 0.0080(0.21*) 0.07278 –0.01361

Panel H. KOSPI
N 0.0220(5.41) 0.18050 –0.02675 0.0200(3.91) 0.03799 –0.01639
T 0.0200(3.91) 0.20379 –0.02199 0.0180(2.61*) 0.04722 –0.01942
SGT 0.0163(1.68*) 0.15465 –0.01563 0.0180(2.61*) 0.04487 –0.01941
Note: 1. *indicates that the model passes the unconditional coverage test at the 5 % significance
level and the critical value of the LRuc test statistics at the 5 % significance level is 3.84. 2. The red
(resp. blue) font represents the lowest (resp. highest) AQLF and unexpected loss when the
predictive accuracies of three different innovations with the same VaR method are compared.
3. The delete-line font represents the lowest AQLF and unexpected loss when the predictive
accuracies of two different VaR methods with the same innovation are compared. 4. The model
acronyms stand for the following methods: HW-GARCH non-parametric method proposed by
Hull and White (1998), GARCH parametric method of GARCH model, N the standard normal
distribution, T the standardized student’s t distribution, SGT the standardized SGT distribution
proposed by Theodossiou (1998)
1420 C.-F. Lee and J.-B. Su

Table 51.5 Out-of-sample long VaR performance at the 99.5 % confidence level
GARCH HW-GARCH
Failure rate (LRuc) AQLF UL Failure rate (LRuc) AQLF UL
Panel A. ATX
N 0.0960(17.75) 0.30360 –0.10151 0.0960(17.75) 0.30642 –0.10308
T 0.0960(17.75) 0.3 2058 –0.10865 0.0920(15.04) 0.30342 –0.10384
SGT 0.0900(13.75) 0.29877 –0.10212 0.0980(19.18) 0.32263 –0.10858
Panel B. Brussels
N 0.0780(7.10) 0.20246 –0.07099 0.0680(3.08*) 0.18374 –0.06673
T 0.0720(4.51) 0.19494 –0.07020 0.0620(1.41*) 0.16978 –0.06268
SGT 0.0660(2.45*) 0.18178 –0.06602 0.0620(1.41*) 0.17067 –0.06262
Panel C. CAC40
N 0.0740(5.31) 0.21446 –0.06485 0.0700(3.76*) 0.20928 –0.05844
T 0.0800(8.07) 0.22598 –0.06887 0.0720(4.51) 0.19510 –0.05617
SGT 0.0760(6.18) 0.21146 –0.06281 0.0700(3.76*) 0.19390 –0.05600
Panel D. Swiss
N 0.0760(6.18) 0.18567 –0.06181 0.0620(1.41*) 0.15104 –0.05088
T 0.0740(5.31) 0.18729 –0.06345 0.0600(0.99*) 0.14222 –0.04694
SGT 0.0740(5.31) 0.17506 –0.05698 0.0560(0.36*) 0.13928 –0.04697
Panel E. Bombay
N 0.0780(7.10) 0.34428 –0.10214 0.0800(8.07) 0.33643 –0.09967
T 0.0820(9.11) 0.35999 –0.10617 0.0780(7.10) 0.33182 –0.09639
SGT 0.0700(3.76*) 0.31488 –0.09366 0.0800(8.07) 0.33629 –0.09824

Panel F. KLSE
N 0.0560(0.36*) 0.20084 –0.04595 0.0740(5.31) 0.23255 –0.05623
T 0.0600(0.99*) 0.21827 –0.05152 0.0700(3.76*) 0.23146 –0.05276
SGT 0.0580(0.64*) 0.21375 –0.05007 0.0640(1.90*) 0.22378 –0.05306
Panel G. STRAITS
N 0.0580(0.64*) 0.21980 –0.06233 0.0640(1.90*) 0.24194 –0.06675
T 0.0640(1.90*) 0.25978 –0.07222 0.0620(1.41*) 0.23921 –0.06511
SGT 0.0620(1.41*) 0.24670 –0.06736 0.0560(0.36*) 0.23035 –0.06428
Panel H. KOSPI
N 0.0740(5.31) 0.27949 –0.08826 0.0600(0.99*) 0.24267 –0.07816
T 0.0760(6.18) 0.32360 –0.09624 0.0620(1.41*) 0.25234 –0.07522
SGT 0.0672(2.82*) 0.27616 –0.08236 0.0620(1.41*) 0.24475 –0.07531
Note: 1.*Indicates that the model passes the unconditional coverage test at the 5 % significance
level and the critical value of the LRuc test statistics at the 5 % significance level is 3.84. 2. The red
(resp. blue) font represents the lowest (resp. highest) AQLF and unexpected loss when the
predictive accuracies of three different innovations with the same VaR method are compared.
3. The delete-line font represents the lowest AQLF and unexpected loss when the predictive
accuracies of two different VaR methods with the same innovation are compared. 4. The model
acronyms stand for the following methods: HW-GARCH non-parametric method proposed by Hull
and White (1998), GARCH parametric method of GARCH model, N the standard normal distri-
bution, T the standardized student’s t distribution, SGT the standardized SGT distribution proposed
by Theodossiou (1998)
51 Valuet-isk Estimation via a Semi-rametric Approach 1421

level, Table 51.3 describes that the HW-N, HW-T, and HW-SGT models pass the
LRuc tests with a total of 5, 5, and 6 stock indices, respectively. Moreover, for
99 % confidence level, Table 51.4 depicts that the HW-N, HW-T, and HW-SGT
models pass the LRuc tests with a total of 7, 8, and 8 stock indices, respectively;
for 99.5 % confidence level, Table 51.5 illustrates that the HW-N, HW-T, and
HW-SGT models pass the LRuc tests with a total of 7, 8, and 8 stock indices,
respectively. Hence, under all confidence levels, there is a total of 19, 21, and
22 acceptances for HW-N, HW-T, and HW-SGT models (the semi-parametric
approach), respectively.
From the abovementioned results, we can find the following two important
phenomena: First, under the same return distributional setting, the number of
acceptance of the HW-based models is greater or equal than those of the
GARCH-based models, irrespective of whether the case of individual level
(95 %, 99 %, or 99.5 %) or all levels (95 %, 99 %, and 99.5 %) is considered.
For example, with regard to all levels, the number of acceptance of the HW-N
model (19) is greater than those of the GARCH-N models (6). These results reveal
that the HW-based models (semi-parametric approach) have the better VaR fore-
casting performance as compared with GARCH-based models (parametric
approach). Second, the number of acceptance of the SGT distribution is the greatest
followed by the student’s t and normal distributions, irrespective of whether the
GARCH-based model (parametric) or HW-based model (semi-parametric
approach) is employed. For instance, with regard to all levels, the number of
acceptance of the GARCH-SGT model (20) is the greatest followed by the
GARCH-T model (14) and GARCH-N model (6). These results indicate that
the SGT has the best VaR forecasting performance followed by student’s t while
the normal owns the worst VaR forecasting performance.
Turning to the other two accuracy measures (i.e., AQLF and UL), the two loss
functions (the average quadratic loss function (AQLF) and the unexpected loss
(UL)) reflect the magnitude of the violation which occur as the observed return
exceeds the VaR estimation. The smaller the AQLF and UL are generated,
the better the forecasting performance of the models is. As observed in
Tables 51.3, 51.4 and 51.5, we can also find the following two important phe-
nomena which are similar as those of the back-testing as was mentioned above:
First, under the same return distributional setting, the AQLF and UL generated by
the HW-based models are smaller than those generated by the GARCH-based
models, irrespective of whether the 95 %, 99 %, or 99.5 % level is considered.
These results reveal that the HW-based models (semi-parametric approach)
significantly have the better VaR forecasting performance as compared with
GARCH-based models (parametric approach), which is in line with the results
of the back-testing. Second, for all confidence levels, the GARCH-SGT
model yields the lowest AQLF and UL for most of the stock indices.
Moreover, for most of the stock indices, the GARCH-N model produces the
highest AQLF and UL for both 99 % and 99.5 % levels, while the GARCH-T
model gives the highest AQLF and UL for 95 % level. These results indicate
that the GARCH-SGT model significantly owns the best out-of-sample VaR
1422 C.-F. Lee and J.-B. Su

performance, while the GARCH-N model appears to have the worst out-of-
sample VaR performance. On the contrary, for all confidence levels, the HW-N
model bears the highest AQLF and UL for most of the stock indices, while the
HW-SGT model gives the lowest AQLF and UL for half of the stock indices,
indicating that the HW-N model significantly owns the worst out-of-sample VaR
performance, while the HW-SGT model appears to bear the highest out-of-
sample VaR performance. Consequently, it seems reasonable to conclude that
the SGT has the best VaR forecasting performance followed by student’s t, while
the normal owns the worst VaR forecasting performance, which appears to be
consistent with the results of back-testing.
To sum up, according to the three accuracy measures, the HW-based models
(semi-parametric approach) have the better VaR forecasting performance as com-
pared with GARCH-based models (parametric approach), and the SGT has the best
VaR forecasting performance followed by student’s t, while the normal owns the
worst VaR forecasting performance. In addition, the kind of VaR approach is more
influential than that of return distribution setting on VaR estimate.

51.5 Conclusion

This study utilizes the parametric approach (GARCH-N, GARCH-T, and GARCH-
SGT models) and the semi-parametric approach of Hull and White (1998) (HW-N,
HW-T, and HW-SGT models), totaling six models, to estimate the VaR for the
eight stock indices in Europe and Asia stock markets, then uses three accuracy
measures: one likelihood ratio test (the unconditional coverage test (LRuc) of
Kupiec (1995)) and two loss functions (the average quadratic loss function
(AQLF) of Lopez (1999) and the unexpected loss (UL)) to compare the forecasting
ability of the aforementioned models in terms of VaR.
The empirical findings can be summarized as follows. First, according to the
results of the log-likelihood ratio test, the SGT distribution closely approximates
the empirical return series followed by student’s t and normal distributions.
Second, in terms of the failure rate, all models tend to underestimate the real market
risk in most cases, but the non-normal distributions (student’s t and SGT) and
the semi-parametric approach try to reverse the trend of underestimating real
market risk, especially at the 99.5 % level. Third, the kind of VaR approaches is
more influential than that of return distribution settings on VaR estimate.
Moreover, under the same return distributional setting, the HW-based models
(semi-parametric approach) have the better VaR forecasting performance as com-
pared with the GARCH-based models (parametric approach). Finally, irrespective
of whether the GARCH-based model (parametric) or HW-based model (semi-
parametric approach) is employed, the SGT has the best VaR forecasting perfor-
mance followed by student’s t, while the normal owns the worst VaR forecasting
performance.
51 Valuet-isk Estimation via a Semi-rametric Approach 1423

Appendix 1: The Left-Tailed Quantiles of the Standardized SGT

The standardized SGT distribution was derived by Lee and Su (2011) and expressed
as follows:
 nþ1
j e t þ dj k k
f ð et Þ ¼ C 1 þ k k (51.17)
½1 þ signðet þ dÞl y

1
12 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where y ¼ Sð1lÞ B k1 ; kn 2 B k3 ; n2
k , SðlÞ ¼ 1 þ 3l2  4A2 l2 ,

1 n
0:5 3 n2
0:5 k
1 n
1
A ¼ B k2 ; n1
k B k;k B k; k , d ¼ S2lA
ðlÞ, C ¼ 2y B k ; k

where k, n, and l are scaling parameters and C and y are normalizing constants
ensuring that f(•) is a proper p.d.f. The parameters k and n control the height and tails
of density with constraints k > 0 and n > 2, respectively. The skewness parameter l
controls the rate of descent of the density around the mode of et with  1 < l < 1. In
the case of positive (resp. negative) skewness, the density function skews toward the
right (resp. left). Sign is the sign function, and B(•) is the beta function. The parameter
n has the degrees of freedom interpretation in case l ¼ 0 and k ¼ 2. Particularly, the
SGT distribution generates the student’s t distribution for l ¼ 0 and k ¼ 2. Moreover,
the SGT distribution generates the normal distribution for l ¼ 0, k ¼ 2, and n ¼ 1.
As observed from Table 51.2, the shape parameters in SGT distribution, the
fat-tail parameter (n) ranges from 4.9846 (KLSE) to 21.4744 (KOSPI), and
the fat-tail parameter (k) is between 1.5399 (KOSPI) and 2.3917 (Bombay). The
skewness parameter (l) ranges from 0.1560 (Bombay) to 0.0044 (KLSE).
Therefore, the left-tailed quantiles of the SGT distribution with various combinations
of shape parameters (0.15  l  0.05; 1.0  k  2.0; n ¼ 10) at alternate levels are
obtained by the composite trapezoid rule and are listed in Table 51.6. Moreover,
Fig. 51.3 depicts the left-tailed quantiles surface of SGT (versus normal) distribution
with various combinations of shape parameters (0.25  l  0.25; 0.8  k  2.0;
n ¼ 10 and 20) at 10 %, 5 %, 1 %, and 0.5 % levels. Notably, Fc(et;k ¼ 2, l ¼ 0, n ¼ 1)
where c ¼ 0.1, 0.05, 0.01, and 0.005 in Fig. 51.3 represents the left-tailed quantiles of
normal distribution at 10 %, 5 %, 1 %, and 0.5 % levels, which is 1.28155,
1.64486, 2.32638, and 2.57613, respectively.

Appendix 2: The Procedure of Parametric VaR Approach

The parametric method is very popular because the only variables you need to do
the calculation are the mean and standard deviation of the portfolio, indicating the
simplicity of the calculations. Moreover, from the literatures’ review mentioned
above, numerous studies focused on the parametric approach of the GARCH family
 1424

Table 51.6 The left-tailed quantiles of SGT distribution with n ¼ 10 and various combinations (k, l) at alternate levels
k\l 0.15 0.10 0.05 0.00 0.05
Panel A. 10 % level
1.0 1.6848(1.1536) 1.6443(1.1359) 1.6003(1.1165) 1.5532(1.0958) 1.5033(1.0739)
1.1 1.7005(1.1831) 1.6615(1.1659) 1.6193(1.1472) 1.5742(1.1273) 1.5266(1.1062)
1.2 1.7103(1.2066) 1.6729(1.1901) 1.6324(1.1721) 1.5893(1.1530) 1.5439(1.1329)
1.3 1.7160(1.2257) 1.6800(1.2098) 1.6413(1.1927) 1.6001(1.1744) 1.5568(1.1552)
1.4 1.7189(1.2414) 1.6843(1.2261) 1.6471(1.2097) 1.6078(1.1923) 1.5665(1.1740)
1.5 1.7196(1.2544) 1.6863(1.2398) 1.6508(1.2241) 1.6132(1.2075) 1.5738(1.1901)
1.6 1.7189(1.2652) 1.6869(1.2512) 1.6528(1.2363) 1.6168(1.2204) 1.5792(1.2039)
1.7 1.7171(1.2744) 1.6864(1.2610) 1.6536(1.2467) 1.6191(1.2316) 1.5831(1.2159)
1.8 1.7147(1.2822) 1.6850(1.2694) 1.6536(1.2557) 1.6205(1.2413) 1.5860(1.2264)
1.9 1.7116(1.2888) 1.6831(1.2766) 1.6528(1.2635) 1.6211(1.2498) 1.5881(1.2356)
2.0 1.7083(1.2946) 1.6807(1.2828) 1.6516(1.2704) 1.6211(1.2573) 1.5894(1.2438)
Panel B. 5 % level
1.0 1.6848(1.7252) 1.6443(1.6851) 1.6003(1.6418) 1.5532(1.5955) 1.5033(1.5468)
1.1 1.7005(1.7349) 1.6615(1.6966) 1.6193(1.6553) 1.5742(1.6114) 1.5266(1.5652)
1.2 1.7103(1.7397) 1.6729(1.7031) 1.6324(1.6639) 1.5893(1.6222) 1.5439(1.5785)
1.3 1.7160(1.7413) 1.6800(1.7064) 1.6413(1.6690) 1.6001(1.6294) 1.5568(1.5880)
1.4 1.7189(1.7406) 1.6843(1.7073) 1.6471(1.6716) 1.6078(1.6340) 1.5665(1.5947)
1.5 1.7196(1.7385) 1.6863(1.7065) 1.6508(1.6726) 1.6132(1.6368) 1.5738(1.5995)
1.6 1.7189(1.7353) 1.6869(1.7047) 1.6528(1.6723) 1.6168(1.6382) 1.5792(1.6027)
1.7 1.7171(1.7318) 1.6864(1.7021) 1.6536(1.6711) 1.6191(1.6387) 1.5831(1.6048)
1.8 1.7147(1.7270) 1.6850(1.6990) 1.6536(1.6694) 1.6205(1.6383) 1.5860(1.6061)
1.9 1.7116(1.7224) 1.6831(1.6955) 1.6528(1.6671) 1.6211(1.6375) 1.5881(1.6067)
2.0 1.7083(1.7177) 1.6807(1.6918) 1.6516(1.6646) 1.6211(1.6362) 1.5894(1.6068)
C.-F. Lee and J.-B. Su
 51

Panel C. 1 % level
1.0 3.1972(3.1305) 3.0964(3.0358) 2.9871(2.9336) 2.8702(2.8247) 2.7467(2.7100)
1.1 3.1281(3.0522) 3.0323(2.9630) 2.9289(2.8673) 2.8188(2.7657) 2.7029(2.6589)
1.2 3.0623(2.9797) 2.9711(2.8956) 2.8732(2.8057) 2.7693(2.7106) 2.6602(2.6109)
1.3 3.0005(2.9132) 2.9136(2.8336) 2.8207(2.7489) 2.7224(2.6596) 2.6195(2.5661)
1.4 2.9430(2.8522) 2.8600(2.7768) 2.7717(2.6968) 2.6785(2.6126) 2.5811(2.5247)
1.5 2.8896(2.7963) 2.8103(2.7247) 2.7261(2.6489) 2.6376(2.5693) 2.5452(2.4864)
1.6 2.8402(2.7451) 2.7643(2.6769) 2.6839(2.6049) 2.5995(2.5295) 2.5117(2.4511)
1.7 2.7945(2.6980) 2.7216(2.6329) 2.6447(2.5644) 2.5641(2.4927) 2.4804(2.4184)
1.8 2.7522(2.6547) 2.6821(2.5924) 2.6083(2.5270) 2.5312(2.4588) 2.4512(2.3881)
1.9 2.7129(2.6147) 2.6454(2.5550) 2.5745(2.4925) 2.5006(2.4274) 2.4240(2.3600)
2.0 2.6764(2.5777) 2.6113(2.5204) 2.5431(2.4606) 2.4720(2.3983) 2.3986(2.3339)
Panel D. 0.5 % level
1.0 3.9229(3.7705) 3.7941(3.6512) 3.6542(3.5224) 3.5043(3.3851) 3.3459(3.2405)
1.1 3.8012(3.6396) 3.6789(3.5277) 3.5470(3.4077) 3.4064(3.2802) 3.2583(3.1464)
1.2 3.6897(3.5224) 3.5736(3.4173) 3.4490(3.3050) 3.3168(3.1862) 3.1780(3.0618)
Valuet-isk Estimation via a Semi-rametric Approach

1.3 3.5882(3.4176) 3.4778(3.3185) 3.3599(3.2131) 3.2352(3.1020) 3.1047(2.9859)

1.4 3.4961(3.3235) 3.3910(3.2299) 3.2791(3.1307) 3.1612(3.0263) 3.0380(2.9175)
1.5 3.4124(3.2389) 3.3121(3.1502) 3.2057(3.0564) 3.0938(2.9581) 2.9773(2.8557)
1.6 3.3363(3.1625) 3.2403(3.0782) 3.1389(2.9893) 3.0325(2.8963) 2.9219(2.7997)
1.7 3.2669(3.0932) 3.1749(3.0129) 3.0779(2.9284) 2.9765(2.8403) 2.8712(2.7488)
1.8 3.2034(3.0301) 3.1150(2.9534) 3.0221(2.8730) 2.9252(2.7891) 2.8248(2.7023)
1.9 3.1452(2.9724) 3.0601(2.8991) 2.9710(2.8223) 2.8781(2.7424) 2.7820(2.6597)
2.0 3.0916(2.9196) 3.0097(2.8492) 2.9239(2.7757) 2.8348(2.6994) 2.7426(2.6206)
Note: 1. k and l denote the shape parameter and skewness parameter of SGT distribution, respectively. 2. Numbers in parentheses are quantiles with n ¼ 20.
3. Numbers in table are obtained using the composite Simpson’s rule with WinRATS 6.1 packages
1425
1426 C.-F. Lee and J.-B. Su

a Quantiles of SGT distribution with various combinations (κ,λ) at 10%

−0.8 Fα=0.10(ε;κ,λ,n=20)

LHS Critical Value −1

−1.2

−1.4

−1.6
Fα=0.10(ε;κ,λ,n=∞)
−1.8
2
1.5 0.4
Fα=0.10(ε;κ,λ,n=10)
0.2
1 0
−0.2
Kappa(κ) 0.5 −0.4 lambda(λ)

b Quantiles of SGT distribution with various combinations (κ,λ) at 5%

−1

−1.2
LHS Critical Value

−1.4

−1.6

−1.8

−2 Fα=0.05(ε;κ,λ,n=∞)
2 Fα=0.05(ε;κ,λ,n=10)
1.5 Fα=0.05(ε;κ,λ,n=20) 0.4
0.2
1 0
−0.2
Kappa(κ) 0.5 −0.4 lambda(λ)

c Quantiles of SGT distribution with various combinations (κ,λ) at 1%

F∝=0.01(ε;κ,λ,n=20)
−2

−2.5
LHS Critical Value

−3

−3.5 Fα=0.01(ε;κ,λ,n=10)
F (ε;κ=2,λ=0,n=∞)
α=0.01

−4
2
1.5 0.4
0.2
1 0
−0.2
Kappa(κ) 0.5 −0.4 lambda(λ)

Fig. 51.3 (continued)

51 Valuet-isk Estimation via a Semi-rametric Approach 1427

d Quantiles of SGT distribution with various combinations (κ,λ) at 0.5%

−2

−2.5
LHS Critical Value

−3

−3.5
F (ε;κ,λ,n=20)
α=0.005 F(ε;κ=2,λ=0,n=∞)
−4 α=0.005

−4.5
2 F(ε;κ,λ,n=10)
α=0.005
1.5 0.4
0.2
1 0
−0.2
Kappa(κ) 0.5 −0.4
lambda(λ)

Fig. 51.3 The left-tailed quantiles of SGT distribution with n ¼ 10, 20, and various combinations
(k, l). (a) 10 %, (b) 5 %, (c) 1 %, (d) 0.5 % confidence levels

variance specifications to estimate the VaR. Furthermore, numerous time series

data of financial assets appear to exhibit autocorrelated and volatility clustering,
and the unconditional distribution of those returns displays leptokurtosis and
a moderate amount of skewness. This study thus considers the applicability of the
GARCH(1,1) model with three conditional distributions (the normal, student’s t,
and SGT distributions) to estimate the corresponding volatility in terms of different
stock indices, then employs the framework of Jorion (2000) to evaluate the VaR of
parametric approach. We take an example of the GARCH-SGT model. The meth-
odology of parametric VaR approach is based on a rolling window procedure. The
window size is fixed at 2,000 observations. More specifically, the procedure is
conducted in the following manner:
Step 1: For each data series, using the econometric package of WinRATS 6.1, the
parameters are estimated with a sample of 2,000 daily returns by quasi-maximum
likelihood estimation (QMLE) of log-likelihood function such as Eq. 51.10 and by
the BFGS optimization algorithm. Thus, with c ¼ [m, o, a, b, k, l, n], the vector
of parameters is estimated. The empirical results of GARCH-SGT model are listed
in Table 51.2 for all stock indices surveyed in this paper. As to the empirical
results of GARCH-N and GARCH-T models, they are also provided by the
same approach.
Step 2: Based on the framework of the parametric techniques (Jorion 2000), the
1-day-ahead VaR based on GARCH-SGT model can be calculated by Eq. 51.11.
Then the one-step-ahead VaR forecasts are compared with the observed returns, and
the comparative results are recorded for subsequent evaluation using statistical tests.
Step 3: The estimation period is then rolled forwards by adding one new day and
dropping the most distant day. By replicating step 1 and step 2, the vector of
1428 C.-F. Lee and J.-B. Su

parameters is estimated, and then the 1-day-ahead VaR can be calculated for the
next 500 days.
Step 4: For the out-sample period (500 days), via the comparable results between
the one-step-ahead VaR forecasts and the observed returns, the 1-day-ahead
BLF, QLF, and UL can be calculated by using Eqs. 51.13, 51.14 and 51.16. On
the other hand, the unconditional coverage test, LRuc, is evaluated by employing
Eq. 51.15. Thereafter, with regard to the GARCH-based models with alternate
distributions (GARCH-N, GARCH-T, and GARCH-SGT), the unconditional
coverage test (LRuc) and three loss functions (failure rate, AQLF, and UL) are
obtained and are reported in the left panel of Tables 51.3, 51.4 and 51.5 for 95 %,
99 %, and 99.5 % levels.

Appendix 3: The Procedure of Semi-parametric VaR Approach

In this paper, we use the approach proposed by Hull and White (1998) as
a representative of the semi-parametric approach. This method mainly couples
a weighting scheme of volatility with the traditional historical simulation. Hence,
it can be regarded as a straightforward extension of traditional historical simulation.
The weighting scheme of volatility is expressed as follows. Instead of using the
actual historical percentage changes in market variables for the purposes of calcu-
lating VaR, we use historical changes that have been adjusted to reflect the ratio of
the current daily volatility to the daily volatility at the time of the observation and
assume that the variance of each market variable during the period covered by the
historical data is monitored using a GARCH-based models. We take an example of
the HW-SGT model. This methodology is explained in the following five steps:
Step 1: For each data series, using the econometric package of WinRATS 6.1, the
parameters are estimated with a sample of 2,000 daily returns by quasi-
maximum likelihood estimation (QMLE) of log-likelihood function such as
Eq. 51.10 and by the BFGS optimization algorithm. Thus, with c ¼ [m, o, a,
b, k, l, n], the vector of parameters is estimated. This step is the same as the first
step of parametric approach. Consequently, a series of daily volatility estimates,
{s1, s2, s3,......, st¼T}, are obtained where T is the number of estimated samples
and equals 2,000 in this study.
Step 2: The modified return series are obtained by the raw return series multiplied
by the ratio of the current daily volatility to the daily volatility at the time of the
observation, sT/si. That is, the modified return series are expressed as {r1*, r2*,
r3*,......, rt¼T*}, where ri* ¼ ri(sT/si).
Step 3: Resort this modified return series ascendingly to achieve the empirical
distribution. Thus, VaR is the percentile that corresponds to the specified
confidence level. Then the one-step-ahead VaR forecasts are compared with
the observed returns, and the comparative results are recorded for subsequent
evaluation using statistical tests.
Step 4: The estimation period is then rolled forwards by adding one new day and
dropping the most distant day. By replicating steps 1–3, the vector of parameters
51 Valuet-isk Estimation via a Semi-rametric Approach 1429

is estimated, and then the 1-day-ahead VaR can be calculated for the next
500 days. This step is the same as the third step of parametric approach.
Step 5: For the out-sample period (500 days), via the comparable results between
the one-step-ahead VaR forecasts and the observed returns, the 1-day-ahead
BLF, QLF, and UL can be calculated by using Eqs. 51.13, 51.14, and 51.16.
On the other hand, the unconditional coverage test, LRuc, is evaluated by
employing Eq. 51.15. Thereafter, with regard to HW-based models with
alternate distributions (HW-N, HW-T, and HW-SGT), the unconditional
coverage test (LRuc) and three loss functions (failure rate, AQLF, and UL)
are obtained and are reported in the right panel of Tables 51.3, 51.4 and 51.5
for 95 %, 99 %, and 99.5 % levels.

References
Aloui, C., & Mabrouk, S. (2010). Value-at-risk estimations of energy commodities via long-
memory, asymmetry and fat-tailed GARCH models. Energy Policy, 38, 2326–2339.
Angelidis, T., Benos, A., & Degiannakis, S. (2004). The use of GARCH models in VaR estimation.
Statistical Methodology, 1, 105–128.
Bali, T. G., & Theodossiou, P. (2007). A conditional-SGT-VaR approach with alternative GARCH
models. Annals of Operations Research, 151, 241–267.
Bhattacharyya, M., Chaudhary, A., & Yadav, G. (2008). Conditional VaR estimation using
Pearson’s type IV distribution. European Journal of Operational Research, 191, 386–397.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of
Econometrics, 31, 307–327.
Bollerslev, T., Chou, R. Y., & Kroner, K. F. (1992). ARCH modeling in finance: A review of the
theory and empirical evidence. Journal of Econometrics, 52, 5–59.
Cabedo, J. D., & Moya, I. (2003). Estimating oil price ‘Value at Risk’ using the historical
simulation approach. Energy Economics, 25, 239–253.
Chan, W. H., & Maheu, J. M. (2002). Conditional jump dynamics in stock market returns. Journal
of Business Economics Statistics, 20, 377–389.
Chang, T. H., Su, H. M., & Chiu, C. L. (2011). Value-at-risk estimation with the optimal dynamic
biofuel portfolio. Energy Economics, 33, 264–272.
Chen, Q., Gerlach, R., & Lu, Z. (2012). Bayesian value-at-risk and expected shortfall forecasting via
the asymmetric Laplace distribution. Computational Statistics and Data Analysis, 56, 3498–3516.
Degiannakis, S., Floros, C., & Dent, P. (2012). Forecasting value-at-risk and expected shortfall
using fractionally integrated models of conditional volatility: International evidence. Interna-
tional Review of Financial Analysis, 27, 21–33.
Faires, J. D., & Burden, R. (2003). Numerical methods (3rd ed.). Pacific Grove: Tomson Learning.
Gebizlioglu, O. L., Şenoğlu, B., & Kantar, Y. M. (2011). Comparison of certain value-at-risk
estimation methods for the two-parameter Weibull loss distribution. Journal of Computational
and Applied Mathematics, 235, 3304–3314.
Gençay, R., Selçuk, F., & Ulug€ulyaˇgci, A. (2003). High volatility, thick tails and extreme value
theory in value-at-risk estimation. Insurance: Mathematics and Economics, 33, 337–356.
Giot, P., & Laurent, S. (2003a). Value-at-risk for long and short trading positions. Journal of
Applied Econometrics, 18, 641–664.
Giot, P., & Laurent, S. (2003b). Market risk in commodity markets: A VaR approach. Energy
Economics, 25, 435–457.
Hartz, C., Mittnik, S., & Paolellad, M. (2006). Accurate value-at-risk forecasting based on the
normal-GARCH model. Computational Statistics & Data Analysis, 51, 2295–2312.
1430 C.-F. Lee and J.-B. Su

Huang, Y. C., & Lin, B. J. (2004). Value-at-risk analysis for Taiwan stock index futures: Fat tails
and conditional asymmetries in return innovations. Review of Quantitative Finance and
Accounting, 22, 79–95.
Hull, J., & White, A. (1998). Incorporating volatility updating into the historical simulation
method for value-at-risk. Journal of Risk, 1, 5–19.
Jarque, C. M., & Bera, A. K. (1987). A test for normality of observations and regression residuals.
International Statistics Review, 55, 163–172.
Jorion, P. (2000). Value at risk: The new benchmark for managing financial risk. New York:
McGraw-Hill.
Kupiec, P. (1995). Techniques for verifying the accuracy of risk management models. The Journal
of Derivatives, 3, 73–84.
Lee, C. F., & Su, J. B. (2011). Alternative statistical distributions for estimating value-at-risk:
Theory and evidence. Review of Quantitative Finance and Accounting, 39, 309–331 .
Lee, M. C., Su, J. B., & Liu, H. C. (2008). Value-at-risk in US stock indices with skewed
generalized error distribution. Applied Financial Economics Letters, 4, 425–431.
Lopez, J. A. (1999). Methods for evaluating value-at-risk estimates. Federal Reserve Bank of San
Francisco Economic Review, 2, 3–17.
Lu, C., Wu, S. C., & Ho, L. C. (2009). Applying VaR to REITs: A comparison of alternative
methods. Review of Financial Economics, 18, 97–102.
Sadeghi, M., & Shavvalpour, S. (2006). Energy risk management and value at risk modeling.
Energy Policy, 34, 3367–3373.
So, M. K. P., & Yu, P. L. H. (2006). Empirical analysis of GARCH models in value at risk
estimation. International Financial Markets, Institutions & Money, 16, 180–197.
Stavroyiannis, S., Makris, I., Nikolaidis, V., & Zarangas, L. (2012). Econometric modeling and
value-at-risk using the Pearson type-IV distribution. International Review of Financial Anal-
ysis, 22, 10–17.
Su, J. B., & Hung, J. C. (2011). Empirical analysis of jump dynamics, heavy-tails and skewness on
value-at-risk estimation. Economic Modelling, 28, 1117–1130.
Theodossiou, P. (1998). Financial data and the skewed generalized t distribution. Management
Science, 44, 1650–1661.
Vlaar, P. J. G. (2000). Value at risk models for Dutch bond portfolios. Journal of Banking &
Finance, 24, 1131–1154.
Modeling Multiple Asset Returns by
a Time-Varying t Copula Model 52
Long Kang

Contents
52.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1432
52.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1434
52.3 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1436
52.3.1 Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1436
52.3.2 Modeling Marginal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1437
52.3.3 Modeling Dependence Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1438
52.3.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1440
52.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1441
52.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1442
52.5.1 Marginal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1442
52.5.2 Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1444
52.5.3 Time-Varying Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1444
52.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1448
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1448

Abstract
We illustrate a framework to model joint distributions of multiple asset returns
using a time-varying Student’s t copula model. We model marginal distributions
of individual asset returns by a variant of GARCH models and then use
a Student’s t copula to connect all the margins. To build a time-varying structure
for the correlation matrix of t copula, we employ a dynamic conditional corre-
lation (DCC) specification. We illustrate the two-stage estimation procedures for
the model and apply the model to 45 major US stocks returns selected from nine

L. Kang
Department of Finance, Antai College of Economics and Management, Shanghai Jiao Tong
University, Shanghai, China
The Options Clearing Corporation and Center for Applied Economics and Policy Research,
Indiana University, Bloomington, IN, USA
e-mail: lkang@indiana.edu; kanglong@gmail.com

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1431
DOI 10.1007/978-1-4614-7750-1_52,
# Springer Science+Business Media New York 2015
1432 L. Kang

sectors. As it is quite challenging to find a copula function with very flexible

parameter structure to account for difference dependence features among all
pairs of random variables, our time-varying t copula model tends to be a good
working tool to model multiple asset returns for risk management and
asset allocation purposes. Our model can capture time-varying conditional
correlation and some degree of tail dependence, while it also has limitations
of featuring symmetric dependence and inability of generating high tail
dependence when being used to model a large number of asset returns.

Keywords
Student’s t copula • GARCH models • Asset returns • US stocks • Maximum
likelihood • Two-stage estimation • Tail dependence • Exceedance correlation •
Dynamic conditional correlation • Asymmetric dependence

52.1 Introduction

There have been a large number of applications of copula theory in financial

modeling. The popularity of copula mainly results from its capability of
decomposing joint distributions of random variables into marginal distributions
of individual variables and the copula which links the margins. Then the task of
finding a proper joint distribution becomes to find a copula form which features
a proper dependence structure given that marginal distributions of individual vari-
ables are properly specified. Among many copula functions, Student’s t copula is
a good choice, though not perfect, for modeling multivariate financial data as an
alternative to a normal copula, especially for a very large number of assets.
The t copula models are very useful tools to describe joint distributions of multiple
assets for risk management and asset allocation purposes. In this chapter,
we illustrate how to model the joint distribution of multiple asset returns under a
Copula-GARCH framework. In particular, we show how we can build and estimate
a time-varying t copula model for a large number of asset returns and how well
the time-varying t copula accounts for some dependence features of real data.
There are still two challenging issues when applying copula theory to multiple
time series. The first is how to choose a copula that best describes the data. Different
copulas feature different dependence structure between random variables. Some
copulas may fit one particular aspect of the data very well but do not have a very
good overall fit, while others may have the opposite performance. What criteria to
use when we choose from copula candidates is a major question remaining to be
fully addressed. Secondly, how to build a multivariate copula which is sufficiently
flexible to simultaneously account for the dependence structure for each pair of
random variables in joint distributions is still quite challenging. We hope to shed
some light on those issues by working through our time-varying t copula model.
Under a Copula-GARCH framework, we first model each asset return with
a variant of GARCH specification. Based on different properties of asset returns,
we choose a proper GARCH specification to formulate conditional distributions of
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1433

each return. Then, we choose a proper copula function to link marginal distributions
of each return to form the joint distribution. As in marginal distributions of each
return, the copula parameters can also be specified as being dependent on previous
observations to make the copula structure time varying for a better fit of data. In this
chapter, we have an AR(1) process for the conditional mean and a GJR-GARCH
(1,1) specification for the conditional volatility for each return. We employ
a Student’s t copula with a time-varying correlation matrix (by a DCC specification)
to link marginal distributions. Usually the specified multivariate model contains
a huge number of parameters, and the estimation by maximum likelihood estimator
(MLE) can be quite challenging. Therefore, we pursue a two-stage procedure,
where all the GARCH models for each return are estimated individually first and
copula parameters are estimated in the second stage with estimated cumulative
distribution functions from the first stage.
We apply our model to modeling log returns of 45 major US stocks selected from
nine sectors with a time span ranging from January 3, 2000 to November 29, 2011.
Our estimation results show that AR(1) and GJR-GARCH(1,1) can reasonably well
capture empirical properties of individual returns. The stock returns possess fat tails
and leverage effects. We plot the estimated conditional volatility on selected stocks
and volatility spikes which happened during the “Internet Bubbles” in the early
2000s and the financial crisis in 2008. We estimate a DCC specification for the time-
varying t copula and also a normal copula for comparison purposes. The parameter
estimates for time-varying t copula are statistically significant, which indicates a
significant time-varying property of the dependence structure. The time-varying
t copula yields significantly higher log-likelihood than normal copula. This improve-
ment of data fitness results from flexibility of t copula (relative to normal copula) and
its time-varying correlation structure.
We plot the time-varying correlation parameter for selected pairs of stocks under
the time-varying t copula model. The correlation parameters fluctuate around
certain averages, and they spike during the 2008 crisis for some pairs. For
45 asset returns, the estimated degree-of-freedom (DoF) parameter of the t copula
is around 25. Together with the estimated correlation matrix of the t copula, this
DoF leads to quite low values of tail dependence coefficients (TDCs). This may
indicate the limitation of t copulas in capturing possibly large tail dependence
behavior for some asset pairs when being used to model a large number of asset
returns. Nevertheless, the time-varying Student’s t copula model has a relatively
flexible parameter structure to account for the dependence among multiple asset
returns and is a very effective tool to model the dynamics of a large number of asset
returns in practice.
This chapter is organized as follows. Section 52.2 gives a short literature review
on recent applications of copulas to modeling financial time series. Section 52.3
introduces our copula model where we introduce copula theory, Copula-GARCH
framework, and estimation procedures. In particular, we elaborate on how to
construct and estimate a time-varying t copula model. Section 52.4 documents
the data source and descriptive statistics for the data set we use. Section 52.5 reports
estimation results and Sect. 52.6 concludes.
1434 L. Kang

52.2 Literature Review

Copula-GARCH models were previously proposed by Jondeau and Rockinger

(2002) and Patton (2004, 2006a).1 To measure time-varying conditional depen-
dence between time series, the former authors use copula functions with time-
varying parameters as functions of predetermined variables and model marginal
distributions with an autoregressive version of Hansen’s (1994) GARCH-type
model with time-varying skewness and kurtosis. They show for many market
indices, dependency increases after large movements and for some cases it
increases after extreme downturns. Patton (2006a) applies the Copula-GARCH
model to modeling the conditional dependence between exchange rates. He
finds that mark-dollar and yen-dollar exchange rates are more correlated
during depreciation against dollar than during appreciation periods. By a similar
approach, Patton (2004) models the asymmetric dependence between “large cap”
and “small cap” indices and examines the economic and statistical significance of
the asymmetries for asset allocations in an out-of-sample setting. As in above
literature, copulas are mostly used in capturing asymmetric dependence and tail
dependence between times series. Among copula candidates, Gumbel’s copula
features higher dependence (correlation) at upper side with positive upper tail
dependence, and rotated Gumbel’s copula features higher dependence
(correlation) at lower side with positive lower tail dependence. Hu (2006) studies
the dependence structure between a number of pairs of major market indices by
a mixed copula approach. Her copula is constructed by a weighted sum of three
copulas–normal, Gumbel’s, and rotated Gumbel’s copulas. Jondeau and Rockinger
(2006) model the bivariate dependence between major stock indices by a Student’s
t copula where the parameters are assumed to be modeled by a two-state Markov
process.
The task of flexibly modeling dependence structure becomes more challenging
for n-dimensional distributions. Tsafack and Garcia (2011) build up a complex
multivariate copula to model four international assets (two international
equities and two bonds). In his model, he assumes that the copula form has
a regime-switching setup where in one regime he uses an n-dimensional normal
copula and in the other he uses a mixed copula of which each copula component
features the dependence structure of two pairs of variables. Savu and Trede (2010)
develop a hierarchical Archimedean copula which renders more flexible parameters
to characterize dependency between each pair of variables. In their model,
each pair of closely related random variables is modeled by a copula of a
particular Archimedean class, and then these pairs are nested by copulas as
well. The nice property of Archimedean family easily leads to the validity of the

1
Alternative approaches are also developed, such as in Ang and Bekaert (2002), Goeij and
Marquering (2004), and Lee and Long (2009), to address non-normal joint distributions of asset
returns.
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1435

joint distribution constructed by this hierarchical structure. (Trivedi and

Zimmer 2006) apply trivariate hierarchical Archimedean copulas to model
sample selection and treatment effects with applications to the family health-care
demand.
Statistical goodness-of-fit tests can provide some guidance for selecting copula
models. Chen et al. (2004) propose two simple goodness-of-fit tests for multivariate
copula models, both of which are based on multivariate probability integral trans-
form and kernel density estimation. One test is consistent but requires the estima-
tion of the multivariate density function and hence is suitable for a small number of
random variables, while the other may not be consistent but requires only kernel
estimation of a univariate density function and hence is suitable for a large number
of assets. Berg and Bakken (2006) propose a consistent goodness-of-fit test for
copulas based on the probability integral transform, and they incorporate in their
test a weighting functionality which can increase influence of some specific areas of
copulas.
Due to their parameter structure, the estimation of Copula-GARCH models
also suffers from “the curse of dimensionality”.2 The exact maximum likelihood
estimator (MLE) works in theory.3 In practice, however, as the number of time
series being modeled increases, the numerical optimization problem in MLE will
become formidable. Joe and Xu (1996) propose a two-stage procedure, where in the
first stage only parameters in marginal distributions are estimated by MLE and then
the copula parameters are estimated by MLE in the second stage. This two-stage
method is called inference for the margins (IFM) method. Joe (1997) shows that
under regular conditions the IFM estimator is consistent and has the property of
asymptotic normality and Patton (2006b) also shows similar estimator properties
for the two-stage method. Instead of estimating parametric marginal distributions in
the IFM method, we can estimate the margins by using empirical distributions,
which can avoid the problem of mis-specifying marginal distributions. This method
is called canonical maximum likelihood (CML) method by Cherubini et al. (2004).
Hu (2006) uses this method and she names it as a semi-parametric method. Based
on Genest et al. (1995), she shows that CML estimator is consistent and has
asymptotical normality. Moreover, copula models can also be estimated under
a nonparametric framework. Deheuvels (1981) introduces the notion of empirical
copula and shows that the empirical copula converges uniformly to the underlying
true copula. Finally, Xu (2004) shows how the copula models can be estimated
with a Bayesian approach. The author shows how a Bayesian approach can be
used to account for estimation uncertainty in portfolio optimization based on
a Copula-GARCH model, and she proposes to use a Bayesian MCMC algorithm
to jointly estimate the copula models.

2
For a detailed survey on the estimation of Copula-GARCH model, see Chap. 5 of Cherubini
et al. (2004).
3
See Hamilton (1994) and Greene (2003) for more details on maximum likelihood estimation.
1436 L. Kang

52.3 The Model

52.3.1 Copula

We introduce our Copula-GARCH model framework by first introducing the

concept of copula. A copula is a multivariate distribution function with uniform
marginal distributions as its arguments, and its functional form links all the margins
to form a joint distribution of multiple random variables.4 Copula theory is mainly
based on the work of Sklar (1959), and we state the Sklar’s theorem for continuous
marginal distributions as follows.

Theorem 52.1 Let F1(x1), . . ., Fn(xn) be given marginal distribution functions and
continuous in x1,. . . , xn, respectively. Let H be the joint distribution of (x1,. . . , xn).
Then there exists a unique copula C such that
n
H ðx1 ; . . . ; xn Þ ¼ CðF1 ðx1 Þ, . . . , Fn ðxn ÞÞ, 8ð x 1 ; . . . ; x n Þ 2 ℝ : (52.1)
Conversely, if we let F1(x1), . . ., Fn(xn) be continuous marginal distribution
functions and C be a copula, then the function H defined by Eq. 52.1 is a joint
distribution function with marginal distributions F1(x1), . . ., Fn(xn).
The above theory allows us to decompose a multivariate distribution function
into marginal distributions of each random variable and the copula form linking the
margins. Conversely, it also implies that to construct a multivariate distribution, we
can first find a proper marginal distribution for each random variable and then
obtain a proper copula form to link the margins. Depending on which dependence
measure used, the copula function mainly, not exclusively, governs the dependence
structure between individual variables. Hence, after specifying marginal distribu-
tions of each variable, the task of building a multivariate distribution solely
becomes to choose a proper copula form which best describes the dependence
structure between variables.
Differentiating Eq. 52.1 with respect to (x1,. . . , xn) leads to the joint density
function of random variables in terms of copula density. It is given as
Y
n
n
hðx1 ; . . . ; xn Þ ¼ cððF1 ðx1 Þ, . . . , Fn ðxn ÞÞ f i ðxi Þ, 8ðx1 ; . . . ; xn Þ 2 ℝ , (52.2)
i¼1

where c(F1(x1), . . ., Fn(xn)) is the copula density and fi(xi) is the density function for
variable i. Equation 52.2 implies that the log-likelihood of the joint density can be
decomposed into components which only involve each marginal density and
a component which involves copula parameters. It provides a convenient structure
for a two-stage estimation, which will be illustrated in details in the following
sections.

4
See Nelsen (1998) and Joe (1997) for a formal treatment of copula theory, and Bouye
et al. (2000), Cherubini et al. (2004), and Embrechts et al. (2002) for applications of copula theory
in finance.
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1437

To better fit the data, we usually assume the moments of distributions of random
variables are time varying and depend on past variables. Therefore, the distribution
of random variables at time t becomes a conditional one, and then the above copula
theory needs to be extended to a conditional case. It is given as follows.5

Theorem 52.2 Let Ot1 be the information set up to time t, and let F1(x1,tjOt1),
. . ., Fn(xn,tjOt1) be continuous marginal distribution functions conditional on
Ot1. Let H be the joint distribution of (x1,. . . , xn) conditional on Ot1. Then
there exists a unique copula C such that
    n
H x1 , . . . , xn jOt1 Þ ¼ C F1 x1 jOt1 Þ, . . . , Fn xn jOt1 ÞjOt1 Þ, 8ðx1 ; . . . ; xn Þ 2 ℝ :
(52.3)

Conversely, if we let F1(x1,tjOt1), . . ., Fn(xn,tjOt1) be continuous conditional

marginal distribution functions and C be a copula, then the function H defined by
Eq. 52.3 is a conditional joint distribution function with conditional marginal
distributions F1(x1,tjOt1), . . ., Fn(xn,tjOt1).
It is worth noting that for the above theorem to hold, the information set Ot1 has
to be the same for the copulas and all the marginal distributions. If different
information sets are used, the conditional copula form on the right side of
Eq. 52.3 may not be a valid distribution. Generally, the same information set
used may not be relevant for each marginal distributions and the copula. For
example, the marginal distributions or the copula may be only conditional on
a subset of the universally used information set. At the very beginning of estimation
of the conditional distributions, however, we should use the same information set
based on which we can test for insignificant explanatory variables so as to stick to
a relevant subset for each marginal distribution or the copula.

52.3.2 Modeling Marginal Distributions

Before building a copula model, we need to find a proper specification for marginal
distributions of individual asset returns, as mis-specified marginal distributions
automatically lead to a mis-specified joint distribution. Let xi,t be asset i return at
time t, and its conditional mean and variance are modeled as follows:

xi, t ¼ a0, i þ a1, i xi, t1 þ ei, t , (52.4)

pffiffiffiffiffiffiffi
ei, t ¼ hi , t  i , t , (52.5)
 
hi, t ¼ b0, i þ b1, i hi, t1 þ b2, i e2i, t1 þ b3, i e2i, t1 1 ei, t1 < 0 : (52.6)

5
See Patton (2004).
1438 L. Kang

As shown in Eqs. 52.4, 52.5 and 52.6, we model the conditional mean as an
AR(1) process and the conditional variance as a GJR(1,1) specification.6
We have parameter restrictions as b0,i > 0, b1,i  0, b2,i  0, b2,i + b3,i  0, and
b1, i þ b2, i þ 12 b3, i < 1 . 1(ei,t1 < 0) is an indicator function, which equals one
when ei,t1 < 0 and zero otherwise. We believe that our model specifications can
capture the features of the individual stock returns reasonably well. It is worth
noting that Eqs. 52.4, 52.5 and 52.6 can include more exogenous variables to better
describe the data. Alternative GARCH specifications can be used to describe the
time-varying conditional volatility. We assume i,t is i.i.d. across time and follows
a Student’s t distribution with DoF vi.
Alternatively, to model the conditional higher moments of the series, we can
follow Hansen (1994) and Jondeau and Rockinger (2003) who assume a skewed
t distribution for the innovation terms of GARCH specifications and find that the
skewed t distribution fits financial time series better than normal distribution.
Accordingly, we can assume i,t  Skewed T(i,tjvi,t,li,t) with zero mean and unitary
variance where vi,t is DoF parameter and li,t is skewness parameter. The two
parameters are time varying and depend on lagged values of explanatory variables
in a nonlinear form. For illustration purposes, however, we will only use Student’s
t distribution for i,t in this chapter.

52.3.3 Modeling Dependence Structure

Normal copula and Student’s t copula are two copula functions from elliptical
families, which are frequently used in modeling joint distributions of random
variables. In this chapter, we also estimate a normal copula model for comparison
purposes. Let F1 denote the inverse of the standard normal distribution F and F∑,n
be n-dimensional normal distribution with correlation matrix ∑. Hence, the
n-dimensional normal copula is
 
Cðu; SÞ ¼ FS, N F1 ðu1 Þ, . . . , F1 ðun Þ , (52.7)

and its density form is

 
fS, n F1 ðu1 Þ, . . . , F1 ðun Þ
cðu; SÞ ¼ Y , (52.8)
n  
f F1 ðui Þ
i¼1

where f and f∑,n are the probability density functions (pdfs) of F and F∑,n,
respectively. It can be shown via Sklar’s theorem that normal copula generates
standard joint normal distribution if and only if the margins are standard normal.

6
See Glosten et al. (1993).
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1439

On the other hand, let Tv1 be the inverse of standard Student’s t distribution Tv
with DoF parameter7 v > 2 and TR,v be n-dimensional Student’s t distribution with
correlation matrix R and DoF parameter v. Then n-dimensional Student’s t copula is
 
Cðu; R; nÞ ¼ T R, n T 1 1
n ðu1 Þ, . . . , T n ðun Þ , (52.9)

and its density function is

 
tR, n T 1 1
v ðu1 Þ, . . . , T v ðun Þ
cðu1 ; . . . ; un Þ ¼ Y ,
n  1 
t n T v ð ui Þ
i¼1

where tv and tR,v are the pdfs of Tv and TR,v, respectively.

Borrowing from the dynamic conditional correlation (DCC) structure of
multivariate GARCH models, we can specify a time-varying parameter structure
in the t copula as follows.8 For a t copula, the time-varying correlation matrix is
governed by
 
Qt ¼ ð1  a  bÞS þ a Bt1 B0t1 þ bQt1 , (52.10)

where S is the unconditional covariance matrix of Bt ¼ (Tn1(u1,t), . . ., Tn1(un,t))0

and a and b are nonnegative and satisfy the condition a + b < 1. We assign
Q0 ¼ S and the dynamics of Qt is given by Eq. 52.10. Let qi,j,t be the i,j element
of the matrix Qt, and the i,j element of the conditional correlation matrix Rt can be
calculated as

qi, j, t
ri, j, t ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : (52.11)
qi, i, t qj, j, t

Moreover, the specification of Eq. 52.10 guarantees that the conditional

correlation matrix Rt is positive definite.

Proposition 52.1 In Eqs. 52.10 and 52.11, if

(a) a  0 and b  0,
(b) a + b < 1,
(c) All eigenvalues of S are strictly positive, then the correlation matrix Rt is
positive definite.

7
In contrast to the previous standardized Student’s t distribution, the standard Student’s
t distribution here has variance as v/(v2).
8
Please see Engle and Sheppard (2001) and Engle (2002) for details on the multivariate
DCC-GARCH models.
1440 L. Kang

0
Proof First, (a) and (b) guarantee the system for BtBt is stationary and S exists. With
Q0 ¼ S, (c) guarantees Q0 is positive definite. With (a) to (c), Qt is the sum of
a positive definite matrix, a positive semi-definite matrix, and a positive definite
matrix both with nonnegative coefficients and then is positive definite for all t.
Based on the proposition Eq. 52.1 in Engle and Sheppard (2001), we prove that Rt
is positive definite.

52.3.4 Estimation

We illustrate the estimation procedure by writing out the log-likelihoods for

observations. Let Y ¼ {y,g1, . . .,gn} be the set of parameters in the joint distribution
where y is the set of parameters in the copula and gt is the set of parameters in
marginal distributions for asset i. Then the conditional cumulative distribution
function (cdf) of n asset returns at time t is given as
     
F x1, t , . . . , xn, t Xt1 , Y ¼ C u1, t , . . . , un, t Xt1 , y (52.12)

where Xt1 is a vector of previous observations, C(|Xt1,y) is the conditional copula,

and ui,t ¼ Fi(xi,t|Xt1,gi) is the conditional cdf of the margins. Differentiating both
sides with respect to x1t,. . . . , xn,t leads to the density function as
     Y
n   
f x1, t , . . . , xn, t Xt1 , Y ¼ c u1, t , . . . , un, t Xt1 , y f i xi, t Xt1 , gi ,
i¼1
(52.13)
where c(| Xt1,y) is the density of the conditional copula and fi(xi,t| Xt1,gi) is the
conditional density of the margins. Accordingly, the log-likelihood of the sample is
given by

X
T   
LðYÞ ¼ log f x1, t , . . . , xn, t Xt1 , Y : (52.14)
t¼1

With Eq. 52.13, the log-likelihood can be written as

X
T    XT X
n   
Lðy; g1 ; . . . ; gn Þ ¼ log c u1, t , . . . , un, t Xt1 , y þ f i xi, t Xt1 , gi :
t¼1 t¼1 i¼1
(52.15)

From Eq. 52.15, we observe that the copula and marginal distributions are
additively separate. Therefore, we can estimate the model by a two-stage MLE
procedure. In the first stage, the marginal distribution parameters for each asset are
estimated by MLE, and then with estimated cdf of each asset, we estimate the
copula parameters by MLE. Based on Joe (1997) and Patton (2006b), this two-stage
estimator is consistent and asymptotically normal.
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1441

With our model specifications, we first estimate the univariate GJR-GARCH

(1,1) with an AR(1) conditional mean and Student’s t distribution by MLE. In the
second stage, we need to estimate the parameters for the constant normal copula
and the time-varying Student’s t copula. Let xt ¼ (F 1(u1,t), . . ., F 1(un,t))0 , and
^ which maximizes the
we can analytically derive the correlation matrix estimator S
log-likelihood of the normal copula density as

XT
^¼1
S xt x0t : (52.16)
T t¼1

As there is no analytical solution for MLE of Student’s t copula, the numerical

maximization problem is quite challenging. Following Chen et al. (2004), however,
with Bt ¼ (T1 1 0
n (u1,t), . . ., Tn (un,t)) , we can calculate the sample covariance matrix
of Bt as S, which is a function of DoF parameter v. By setting Q0¼Sˆ, we can express
ˆ
Qt and Rt for all t in terms of a, b, and v using Eq. 52.10. Then we can estimate a, b,
and v by maximizing the log-likelihood of t copula density. In the following
sections, we apply our estimation procedure to the joint distribution of 45 selected
major US stock returns.

52.4 Data

We apply our model to modeling log returns of 45 major US stocks from nine
sectors: Consumer Discretionary, Consumer Staples, Energy, Financials, Health
Care, Industrials, Technology, Materials, and Utilities. Table 52.1 shows stock
symbols and company names of the selected 45 companies. We select five major
companies from each sector to form the stock group. The time span ranges from
January 3, 2000 to November 29, 2011 with 2990 observations. We download data
from yahoo finance (http://finance.yahoo.com/). The log returns are calculated from
daily close stock prices adjusted for dividends and splits.
To save space, we only plot and calculate descriptive statistics of nine stocks
with each from one sector. Figure 52.1 plots the log returns of those nine selected
stocks, and there are two periods of volatility clusterings due to “Internet Bubbles”
in the early 2000s and the financial crisis in 2008, respectively. We observe that
during the financial crisis in 2008, major banks, such as Citigroup, incurred huge
negative and positive daily returns. Table 52.2 shows the calculated mean, standard
deviation, skewness, and kurtosis for the nine stocks. The average returns for
the nine stocks are close to zero. Major banks, represented by Citigroup, have
significantly higher volatility. Most of the stocks are slightly positively skewed, and
only two have slight negative skewness. All the stocks have kurtosis greater than
three indicating fat tails, and again major banks have significantly fatter tails.
All the descriptive statistics indicate that the data property of individual returns
needs to be captured by a variant of GARCH specification.
1442 L. Kang

Table 52.1 Symbols and names of 45 selected stocks from nine sectors
Sector Consumer discretionary Consumer Staples Energy
Stock symbol MCD: McDonald’s WMT: Wal-Mart Stores XOM: Exxon Mobil Corp.
Inc.
HD: Home Depot PG: Procter & Gamble CVX: Chevron Corp.
Co.
DIS: Walt Disney Co. KO: Coca-Cola Co. COP:
CONOCOPHILLIPS.
TGT: Target WAG: Walgreen Co. DVN: Devon
Energy Corp.
LOW: Lowe’s MO: Altria Group Inc. SLB: Schlumberger
Limited
Sector Financials Health Care Industrials
Stock symbol C: Citigroup Inc. JNJ: Johnson & GE: General Electric Co.
Johnson
BAC: Bank of PFE: Pfizer Inc. UNP: Union Pacific Corp.
America Corp.
JPM: JPMorgan Chase & ABT: Abbott UTX: United
Co. Laboratories Technologies Corp.
USB: U.S. Bancorp MRK: Merck & MMM: 3 M Co.
Co. Inc.
WFC: Wells Fargo & Co. AMGN: Amgen Inc. BA: Boeing Co.
Sector Technology Materials Utilities
Stock symbol T: AT&T Inc. NEM: Newmont EXC: Exelon Corp.
Mining Corp.
MSFT: Microsoft Corp. DD: E.I. DuPont de FE:
Nemours & Co. FirstEnergy Corp.
IBM: International Business DOW: Dow Chemical Co. PPL: PPL
Machines Corp. Corporation
CSCO: Cisco Systems Inc. FCX: Freeport-McMoRan D: Dominion
Copper & Gold Inc. Resources, Inc.
HPQ: Hewlett-Packard Co. PX: Praxair Inc. DUK: Duke
Energy Corp.

52.5 Empirical Results

52.5.1 Marginal Distributions

We briefly report estimation results for marginal distributions of 45 stock returns.

For convenience, we only show the estimates and standard errors (in brackets) for
nine selected stocks with each from one sector in Table 52.3. The star indicates
statistical significance at a 5 % level. Consistent with our observations in Table 52.2,
all the nine stocks have low values of DoF indicating fat tails. The parameter b3,i is
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1443

MCD WMT
0.1 0.15
0.05 0.1
Log returns

Log returns
0 0.05
−0.05 0

−0.1 −0.05
−0.15 −0.1
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)
XOM C
0.2 0.5

0.1

Log returns
Log returns

0 0

−0.1

−0.2 −0.5
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)
JNJ GE
0.2 0.2

0.1 0.1
Log returns

Log returns

0 0

−0.1 −0.1

−0.2 −0.2
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)
T NEM
0.2 0.6

0.1 0.4
Log returns

Log returns

0 0.2

−0.1 0
−0.2 −0.2
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)
EXC
0.2

0.1
Log returns

−0.1

−0.2
00 02 04 06 08 10 12
Date (in year)

Fig. 52.1 The log returns of the nine of our 45 selected stocks with each from one sector have
been plotted
1444 L. Kang

Table 52.2 Descriptive statistics (mean, standard deviation, skewness, and kurtosis) for the nine
of our 45 selected stocks with each from each sector
Stock
symbol MCD WMT XOM C JNJ GE T NEM EXC
Mean 3.73E- 7.21E- 3.15E- 8.05E- 1.96E- 2.89E- 1.23E- 3.65E- 4.48E-
04 06 04 04 04 04 05 04 04
Std. dev. 0.017 0.017 0.017 0.037 0.013 0.022 0.019 0.027 0.018
Skewness 0.21 0.13 0.02 0.48 0.53 0.04 0.12 0.34 0.05
Kurtosis 8.25 7.72 12.52 35.55 17.83 9.99 8.68 8.22 10.58
# obs. 2,990 2,990 2,990 2,990 2,990 2,990 2,990 2,990 2,990

statistically significant at a 5 % level for eight of the nine stocks indicating

significant leverage effects for stock returns. The parameters in conditional mean
are statistically significant for some stocks and not for others. In Fig. 52.2, we
plot estimated conditional volatility for the stocks MCD, WMT, XOM, and
C. Consistent with Fig. 52.1, we observe MCD and WMT have significant high
volatility in the early 2000s and 2008, while XOM and C have their volatility hikes
mainly in 2008 with C, representing Citigroup, having the highest conditional
volatility during the 2008 crisis.

52.5.2 Copulas

We report estimation results for the time-varying t copula parameters in Table 52.4.
All the three parameters a, b, and v are statistically significant. The estimate a is
close to zero and the estimate for b is close to one. The estimate for v is about 25. As
our estimation is carried out on the joint distribution of 45 stock returns, the
estimate for v shed some light on how much Student’s t copula can capture tail
dependence when used to fit a relatively large number of variables. We also report
the log-likelihood for time-varying Student’s t copula and normal copula in
Table 52.4. As the correlation matrix in normal copula is estimated by its sample
correlation, we did not report it here. We find that time-varying t copula has
significantly higher log-likelihood than normal copula, which results from the
more flexible parameter structure of t copula and the time-varying parameter
structure.

52.5.3 Time-Varying Dependence

Our time-varying t copula features a time-varying dependence structure among all

the variables. The DoF parameter, together with the correlation parameters, governs
the tail dependence behavior of multiple variables. We plot the estimated
 52

Table 52.3 The GARCH estimation results of individual stock returns for the nine of our selected 45 stocks with each from one sector. Values in brackets are
standard errors. The star indicates the statistical significance at a 5 % level
Stock
symbol MCD WMT XOM C JNJ GE T NEM EXC
Conditional mean
a0,i 6.69E-04* 1.77E-05 6.50E-04* 1.03E-06 1.17E-04 1.93E-06 2.73E-04 3.31E-04 7.10E-04*
(2.25E-04) (2.11E-04) (2.41E-04) (2.42E-04) (1.59E-04) (2.32E-04) (2.29E-04) (3.87E-04) (2.36E-04)
a1,i 0.030 0.037* 0.076* 0.003 0.004 0.015 0.009 0.056* 0.019
(0.018) (0.018) (0.019) (0.018) (0.018) (0.018) (0.018) (0.018) (0.019)
Conditional variance
b0,i 1.69E-06* 8.22E-07* 5.97E-06* 1.96E-06* 1.87E-06* 1.26E-06* 1.37E-06* 3.11E-06* 3.91E-06*
(5.21E-07) (3.50E-07) (1.28E-06) (5.53E-07) (4.40E-07) (3.93E-07) (4.47E-07) (1.31E-06) (9.82E-07)
b1,i 0.947* 0.952* 0.901* 0.908* 0.905* 0.948* 0.943* 0.959* 0.897*
(0.007) (0.007) (0.013) (0.009) (0.011) (0.007) (0.007) (0.007) (0.012)
b2,i 0.024* 0.026* 0.025* 0.045* 0.025* 0.017* 0.027* 0.037* 0.066*
(0.008) (0.008) (0.012) (0.011) (0.011) (0.007) (0.008) (0.009) (0.014)
b3,i 0.046* 0.038* 0.092* 0.094* 0.127* 0.067* 0.051* 0.001 0.046*
(0.012) (0.013) (0.018) (0.017) (0.021) (0.012) (0.012) (0.011) (0.018)
Modeling Multiple Asset Returns by a Time-Varying t Copula Model

Degree of freedom
vi 6.187* (0.67) 7.002* (0.92) 9.496* (1.59) 6.505* (0.75) 5.796* (0.62) 6.622* (0.66) 8.403* (1.20) 7.831* (1.09) 9.247* (1.28)
1445
1446 L. Kang

x 10−3 MCD x 10−3 WMT

2 2

Conditional Volatility
Conditional Volatility

1.5 1.5

1 1

0.5 0.5

0 0
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)
x 10−3 XOM C
6 0.06
Conditional Volatility

Conditional Volatility
5 0.05
4 0.04
3 0.03
2 0.02
1 0.01
0
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date (in year) Date (in year)

Fig. 52.2 The estimated time-varying conditional volatility for four selected stocks has been
plotted

Table 52.4 The estimates and standard errors for time-varying Student’s t copula. Values in
brackets are standard errors. The star indicates the statistical significance at a 5 % level. We also
report the log-likelihood for time-varying t copula and normal copula
Time-varying t copula Normal copula
Parameter estimates
a 0.0031* (0.0002)
b 0.984* (0.0016)
v 25.81* (1.03)
Log-likelihood of copula component
Log-likelihood 40,445.44 38,096.36

conditional correlation parameters of t copula for four selected pairs of stock returns
in Fig. 52.3. For those four pairs, the conditional correlation parameter fluctuates
around certain positive averages. The two pairs, MCD-WMT and NEM-EXC,
experienced apparent correlation spikes during the 2008 financial crisis. Moreover,
Fig. 52.4 shows the estimated TDCs for the four pairs. We find that with the DoF
around 25, the TDCs for those pairs of stock returns are very low, though some pairs
do exhibit TDC spikes during the 2008 crisis. The low values of TDCs indicate
possible limitations of t copula to account for tail dependence when being used to
model a large number of variables.
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1447

XOM-C MCD-WMT
0.5 0.45
Conditional correlation

Conditional correlation
0.45
0.4
0.4
0.35 0.35
0.3
0.3
0.25
0.2 0.25
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date in year Date in year
C-JNJ NEM-EXC
0.45 0.25
Conditional correlation

Conditional correlation
0.4 0.2
0.35
0.15
0.3
0.25 0.1

0.2 0.05
00 02 04 06 08 10 12 00 02 04 06 08 10 12
Date in year Date in year

Fig. 52.3 The estimated time-varying correlation parameters in t copula for four selected pairs of
stock returns have been plotted

x 10−3 x 10−3
Tail Dependence Coefficient

Tail Dependence Coefficient

XOM-C MCD-WMT
4 3.5
3
3
2.5
2 2
1.5
1
1
0 0.5
00 01 02 03 04 05 06 07 08 09 10 11 12 00 01 02 03 04 05 06 07 08 09 10 11 12
Date in year Date in year
x 10−3 x 10−4
Tail Dependence Coefficient
Tail Dependence Coefficient

C-JNJ NEM-EXC
2.5 6

2 5
4
1.5
3
1
2
0.5 1
0 0
00 01 02 03 04 05 06 07 08 09 10 11 12 00 01 02 03 04 05 06 07 08 09 10 11 12
Date in year Date in year

Fig. 52.4 The time-varying tail dependence coefficient (TDC) for the four selected pairs of stock
returns has been plotted
1448 L. Kang

52.6 Conclusion

We illustrate an effective approach (Copula-GARCH models) to model the dynam-

ics of a large number of multiple asset returns by constructing a time-varying
Student’s t copula model. Under a general Copula-GARCH framework, we specify
a proper GARCH model for individual asset returns and use a copula to link the
margins to build the joint distribution of returns. We apply our time-varying
Student’s t copula model to 45 major US stock returns, where each stock return is
modeled by an AR(1) and GJR-GARCH(1,1) specification and a Student’s t copula
with a DCC dependence structure is used to link all the returns. We illustrate how
the model can be effectively estimated by a two-stage MLE procedure, and our
estimation results show time-varying t copula model has significant better fitness of
data than normal copula models.
As it is quite challenging to find a copula function with very flexible parameter
structure to account for difference dependence features among all pairs of random
variables, our time-varying t copula model tends to be a good working tool to model
multiple asset returns for risk management and asset allocation purposes. Our
model can capture time-varying conditional correlation and some degree of tail
dependence, while it also has limitations of featuring symmetric dependence and
inability of generating high tail dependence when being used to model a large
number of asset returns. Nevertheless, we hope that this chapter provides
researchers and financial practitioners with a good introduction on the Copula-
GARCH models and a detailed illustration on constructing joint distributions of
multiple asset returns using a time-varying Student’s t copula model.

References
Ang, A., & Bekaert, G. (2002). International asset allocation with regime shifts. Review of
Financial Studies, 15(4), 1137–1187.
Berg, D., & Bakken, H. (2006). A copula goodness-of-fit test based on the probability integral
transform. Working Paper at http://www.danielberg.no/publications/Btest.pdf
Bouyé, E., Valdo D., Ashkan N., Gaël R., & Thierry, R. (2000). Copulas for finance – A reading
guide and some applications. Available at SSRN http://ssrn.com/abstract¼1032533. March.
Chen, X., Yanqin F., & Andrew, J. P. (2004). Simple tests for models of dependence between
multiple financial time series, with applications to U.S. Equity returns and exchange rates
(Discussion Paper 483). London School of Economics, Revised July 2004.
Cherubini, U., Luciano, E., & Vecchiato, W. (2004). Copula methods in finance (Wiley series in
financial engineering). Chichester: Wiley.
De Goeij, P., & Marquering, W. (2004). Modeling the conditional covariance between stock
and bond returns: A multivariate GARCH approach. Journal of Financial Econometrics, 2(4),
531–564.
Deheuvels, P. (1981). A Nonparametric test for independence. Paris: Universite de Paris,
Institut de Statistique.
Engle, R. (2002). Dynamic conditional correlation: A simple class of multivariate generalized
autoregressive conditional heteroskedasticity models. Journal of Business & Economic Statis-
tics, 20(3), 339–350.
52 Modeling Multiple Asset Returns by a Time-Varying t Copula Model 1449

Engle, R. F., & Sheppard, K. (2001). Theoretical and empirical properties of dynamic conditional
correlation multivariate GARCH (NBER Working Paper No. 8554). Cambridge, MA: National
Bureau of Economic Research.
Genest, C., Ghoudi, K., & Rivest, L. (1995). A semiparametric estimation procedure of depen-
dence parameters in multivariate families of distributions. Biometrika, 82, 543–552.
Glosten, L. R., Ravi, J., & Runkle, D. E. (1993). On the relation between the expected value and
the volatility of the nominal excess return on stocks. Journal of Finance, 48(5), 1779–1801.
Greene, W. H. (2003). Econometric analysis. Upper Saddle River: Prentice Hall.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Hansen, B. E. (1994). Autoregressive conditional density estimation. International Economic
Review, 35(3), 705–730.
Hu, L. (2006). Dependence patterns across financial markets: A mixed copula approach. Applied
Financial Economics, 16, 717–729.
Joe, H. (1997). Multivariate models and dependence concepts. London: Chapman and Hall.
Joe, H., & Xu, J. J. (1996). The estimation method of inference functions for margins for
multivariate models (Technical Report 166). Vancouver: Department of Statistics, University
of British Columbia.
Jondeau, E., & Rockinger, M. (2002). Conditional dependency of financial series: the copula-
GARCH model (FAME Research Paper 69). Geneva: International Center for Financial Asset
Management and Engineering.
Jondeau, E., & Rockinger, M. (2003). Conditional volatility, skewness, and kurtosis: Existence,
persistence, and comovements. Journal of Economic Dynamics and Control, 27(10),
1699–1737.
Jondeau, E., & Rockinger, M. (2006). The copula-GARCH model of conditional dependencies:
An international stock market application. Journal of International Money and Finance,
25, 827–853.
Lee, T.-H., & Long, X. (2009). Copula-based multivariate GARCH model with uncorrelated
dependent errors. Journal of Econometrics, 150, 207–218.
Nelson, R. B. (1998). An introduction to copula. New York: Springer.
Patton, A. J. (2004). On the out-of-sample importance of skewness and asymmetric dependence
for asset allocation. Journal of Financial Econometrics, 2(1), 130–168.
Patton, A. J. (2006a). Estimation of multivariate models for time series of possibly different
lengths. Journal of Applied Econometrics, 21(2), 147–173.
Patton, A. J. (2006b). Modelling asymmetric exchange rate dependence. International Economic
Review, 47(2), 527–556.
Savu, C., & Trede, M. (2010). Hierarchies of Archimedean copulas. Quantitative Finance,
10, 295–304.
Sklar, A. (1959). Fonctions de repartition an dimensions et leurs marges. Paris: Publications de
l’lnstitut de statistique de l’Universite de Paris.
Trivedi, P. K., & Zimmer, D. M. (2006). Using trivariate copulas to model sample selection
and treatment effects: Application to family health care demand. Journal of Business and
Economic Statistics, 24(1), 63.
Tsafack, G., & Garcia, R. (2011). Dependence structure and extreme comovements in interna-
tional equity and bond markets. Journal of Banking and Finance, 35, 1954–1970.
Xu, Y. (2004). Incorporating estimation risk in copula-based portfolio optimization. Miami:
Department of Economics, University of Miami.
Internet Bubble Examination with
Mean-Variance Ratio 53
Zhidong D. Bai, Yongchang C. Hui, and Wing-Keung Wong

Contents
53.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1452
53.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1454
53.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1455
53.4 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1458
53.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1461
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1463

Abstract
To evaluate the performance of the prospects X and Y, financial professionals are
interested in testing the equality of their Sharpe ratios (SRs), the ratios of the
excess expected returns to their standard deviations. Bai et al. (Statistics and
Probability Letters 81, 1078–1085, 2011d) have developed the mean-variance-
ratio (MVR) statistic to test the equality of their MVRs, the ratios of the excess
expected returns to its variances. They have also provided theoretical reasoning
to use MVR and proved that their proposed statistic is uniformly most powerful
unbiased. Rejecting the null hypothesis infers that X will have either smaller

Z.D. Bai (*)

KLAS MOE & School of Mathematics and Statistics, Northeast Normal University,
Changchun, China
Department of Statistics and Applied Probability, National University of Singapore,
Singapore, Singapore
e-mail: baizd@nenu.edu.cn
Y.C. Hui
School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, China
e-mail: huiyc180@mail.xjtu.edu.cn
W.-K. Wong
Department of Economics, Hong Kong Baptist University, Kowloon, Hong Kong
e-mail: awong@hkbu.edu.hk

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1451
DOI 10.1007/978-1-4614-7750-1_53,
# Springer Science+Business Media New York 2015
1452 Z.D. Bai et al.

variance or larger excess mean return or both leading to the conclusion that X is
the better investment. In this paper, we illustrate the superiority of the MVR test
over the traditional SR test by applying both tests to analyze the performance of
the S&P 500 index and the NASDAQ 100 index after the bursting of the Internet
bubble in the 2000s. Our findings show that while the traditional SR test
concludes the two indices being analyzed to be indistinguishable in their per-
formance, the MVR test statistic shows that the NASDAQ 100 index
underperformed the S&P 500 index, which is the real situation after the bursting
of the Internet bubble in the 2000s. This shows the superiority of the MVR test
statistic in revealing short-term performance and, in turn, enables investors to
make better decisions in their investments.

Keywords
Mean-variance ratio • Sharpe ratio • Hypothesis testing • Uniformly most
powerful unbiased test • Internet bubble • Fund management

53.1 Introduction

Internet stocks obtained huge gains in the late 1990s, followed by huge losses from
early 2000. In just 2 years from 1998 to early March 2000, prices of Internet stocks
rose by sixfold and outperformed the S&P 500 by 482 %. Technology stocks
generally showed a similar trend based on the fact that NASDAQ 100 index
quadrupled in value over the same period and outperformed the S&P 500 index
by 268 %. On the other hand, NASDAQ 100 index dropped by 64.28 % in value
during the Internet bubble crash and underperformed the S&P 500 index by
173.87 %.
The spectacular rise and fall of Internet stocks in the late 1990s has stimulated
research into the causes of the Internet stock bubble. Theories had been developed
to explain the Internet bubble. For example, Baker and Stein (2004) develop
a model of market sentiment with irrationally overconfident investors and short-
sale constraints. Ofek and Richardson (2003) provide circumstantial evidence that
Internet stocks attract mostly retail investors who are more prone to be
overconfident about their ability to predict future stock prices than institutional
investors. Perkins and Perkins (1999) suggest that during the Internet boom,
investors were confidently betting on the continued rise of Internet stocks because
they knew that high demand and limited equity float implies substantial upside
returns. Moreover, Ofek and Richardson (2003) provide indirect evidence that
Internet stock prices were supported by a combination of factors such as limited
float, short-sale constraints, and aggressive trend chased by retail investors, whereas
Statman (2002) shows that this asymmetric payoff must have made Internet stocks
appear to be an extremely attractive gamble for risk seekers. On the other hand,
Fong et al. (2008) use stochastic dominance methodology (Fong et al. 2005; Broll
et al. 2006; Chan et al. 2012; Lean et al. 2012) to identify dominant types of risk
preferences in the Internet bull and bear markets. They conclude that investor risk
53 Internet Bubble Examination with Mean-Variance Ratio 1453

preferences (Wong and Li 1999; Wong and Chan 2008) have changed over this
cycle, and the change is related to utility theory (Wong 2007; Sriboonchitta
et al. 2009) and behavioral finance (Lam et al. 2010, 2012).
In this paper, we apply both the mean-variance ratio (MVR) test and the Sharpe
ratio (SR) test to examine the performance of the NASDAQ 100 index and the S&P
500 index during the bursting of the Internet bubble in the 2000s. The tests are
relied on the theory of the mean-variance (MV) portfolio optimization (Markowitz
1952; Bai et al. 2009a, b). The Markowitz efficient frontier also provides the basis
for many important financial economics advances, including the Sharpe-Lintner
capital asset pricing model (CAPM, Sharpe 1964; Lintner 1965) and the well-
known optimal one-fund theorem (Tobin 1958). Originally motivated by the MV
analysis, the optimal one-fund theorem, and the CAPM model, the Sharpe ratio,
the ratio of the excess expected return to its volatility or standard deviation, is
one of the most commonly used statistics in the MV framework. The SR is now
widely used in many different areas in Finance and Economics, from the evalua-
tion of portfolio performance to market efficiency tests (see, e.g., Ofek and
Richardson 2003).
Jobson and Korkie (1981) develop a SR statistic to test for the equality of two
SRs. The test statistic has been modified and improved by Cadsby (1986) and
Memmel (2003). Lo (2002) carries out a more thorough study of the statistical
property of the SR estimator. Using standard econometric methods with several
different sets of assumptions imposed on the statistical behavior of the returns
series, Lo derives the asymptotic statistical distribution for the SR estimator and
shows that confidence intervals, standard errors, and hypothesis tests can be com-
puted for the estimated SRs in much the same way as regression coefficients such as
portfolio alphas and betas are computed.
The SR test statistic developed by Jobson and Korkie (1981) and others provides
a formal statistical comparison of performance among portfolios. One deficiency of
the SR statistic is that it has only an asymptotic distribution. Hence, the SR test has
its statistical properties only for large samples, but not for small samples. Never-
theless, the performance of assets is often compared by using small samples,
especially when markets undergo substantial changes resulting from changes in
short-term factors and momentum. Under these circumstances, it is more meaning-
ful to use limited data to predict the assets’ future performance. In addition, it is not
meaningful to measure SRs for extended periods when the means and standard
deviations of the underlying assets are found empirically to be nonstationary and/or
to possess structural breaks. For small samples, the main difficulty in developing
the SR test is that it is impossible to obtain a uniformly most powerful unbiased
(UMPU) test to check for the equality of SRs. To circumvent this problem, Bai
et al. (2011d) propose to use an alternative statistic, the MVR tests to compare
performance of assets. They also discuss the evaluation of the performance of assets
for small samples by providing a theoretical framework and then invoking both
one-sided and two-sided UMPU MVR tests. Moreover, Bai et al. (2012) further
extend the MVR statistics to compare the performance of prospects after the effect
of the background risk has been mitigated.
1454 Z.D. Bai et al.

Applying the traditional SR test, we fail to reject the possibility of having any
significant difference between the performance of the S&P 500 index and the
NASDAQ 100 index during the bursting of the Internet bubble in the 2000s. This
finding implies that the two indices being analyzed could be indistinguishable in
their performance during the period under the study. However, we conjecture that
this conclusion is most likely to be inaccurate as the lack of sensitivity of the SR test
in analyzing small samples. Thus, we propose to use the MVR test in the analysis.
As expected, the MVR test shows that the MVR of the weekly return on S&P
500 index is different from that on the NASDAQ 100 index. We conclude that the
NASDAQ 100 index underperformed the S&P 500 index during the period under
the study. The proposed MVR test can discern the performance of the two indices
and hence is more informative than tests using the SR statistics for investors to
decide on their investments.
The rest of the paper is organized as follows: Section 53.2 discusses the
data while Sect. 53.3 provides the theoretical framework and discusses the
theory for both one-sided and two-sided MVR tests. In Sect. 53.4, we demonstrate
the superiority of the MVR tests over the traditional SR tests by applying both tests
to analyze the performance of the S&P 500 index and the NASDAQ 100 index
during the bursting of the Internet bubble in the 2000s. This is followed by
Sect. 53.4 which summarizes our conclusions and shares our insights.

53.2 Data

The data used in this study consists of weekly returns on two stock indices: the S&P
500 and the NASDAQ 100 index. We use the S&P 500 index to represent
non-technology or “old economy” firms. Our proxy for the Internet and technology
sectors is the NASDAQ 100 index. Firms represented in the NASDAQ 100 include
those in the computer hardware and software, telecommunications, and biotech-
nology sectors. The NASDAQ 100 index is value weighted.
Our sample period is from January 1, 2000 to December 31, 2002, to study
the effect of the crash in the Internet bubble. Before 2000, there is a clear upward
trend in technology stock prices emerging from around that period and this
period spans a period of intense IPO and secondary market activities for Internet
stocks. Schultz and Zaman (2001) report that 321 Internet firms went public
between January 1999 and March 2000, accounting for 76 % of all new Internet
issues since the first wave of Internet IPOs began in 1996. Ofek and Richardson
(2003) find that the extraordinary high valuations of Internet stocks between the
early 1998 and February 2000 were accompanied by very high trading volume and
liquidity. The unusually high volatility of technology stocks is only partially
explained by the rise in the overall market volatility. Our interest centers on the
bear market from January 1, 2000 to December 31, 2002. All data for this study
are from datastream.
53 Internet Bubble Examination with Mean-Variance Ratio 1455

53.3 Methodology

Let Xi and Yi (i ¼ 1, 2,   , n) be independent excess returns drawn from the

corresponding normal distributions N(m, s2) and N(
, t2) with joint density p(x, y)
such that

 
mX 1 X 2
X 1 X 2
pðx; yÞ ¼ k  exp xi  2 xi þ 2 yi  2 yi (53.1)
s 2 2s t 2t

   
n=2 n=2 2
n
2
Where k ¼ ð2ps2 Þ ð2pt2 Þ exp  nm
2s 2 exp  2t 2

To evaluate the performance of the prospects X and Y, financial professionals are

interested in testing the hypotheses

m
m

H 0 :  versus H 1 : > (53.2)
s t s t

to compare the performance of their corresponding SRs, ms and
t , the ratios of the
excess expected returns to their standard deviations.
If the hypothesis H0 is rejected, it infers that X is the better investment
prospect with larger SR because X has either larger excess mean return or
smaller standard deviation or both. Jobson and Korkie (1981) and Memmel
(2003) develop test statistics to test the hypotheses in Eq. 53.2 for large
samples but their tests would not be appropriate for testing small samples as the
distribution of their test statistics is only valid asymptotically but not valid for
small samples. However, it is especially relevant in investment decisions to test
the hypotheses in Eq. 53.2 for small samples to provide useful investment infor-
mation to investors. Furthermore, as it is impossible to obtain any UMPU test
statistic to test the inequality of the SRs in Eq. 53.2 for small samples, Bai
et al. (2011d) propose to use the following hypothesis to test for the inequality of
the MVRs:

m
m

H 01 :  versus H11 : > : (53.3)
s2 t 2 s2 t 2

In addition, they develop the UMPU test statistic to test the above hypotheses.
Rejecting the hypothesis H0 infers that X will have either smaller variance or larger
excess mean return or both leading to the conclusion that X is the better investment.
As sometimes investors conduct the two-sided test to compare the MVRs, the
following hypotheses are included in our study:
1456 Z.D. Bai et al.

m
m

H 02 : ¼ 2 versus H12 : 6¼ 2 : (53.4)
s 2 t s 2 t

One may argue that the MVR test is that SR test is scale invariant, whereas the
MV ratio test is not. To support the MVR test to be an acceptable alternative test
statistic, Bai et al. (2011d) show the theoretical justification for the use of the MVR
test statistic in the following remark:
Remark 53.1 One may think that the MVR can be less favorable than the SR as the
former is not scale invariant while the latter is. However, in some financial
processes, the mean change in a short period of time is proportional to its variance
change. For example, many financial processes can be characterized by the fol-
lowing diffusion process for stock prices formulated as

dY t ¼ mP ðY t Þdt þ sðY t ÞdW Pt ,

where mP is an N-dimensional function, s is an N  N matrix and WPt is an

N-dimensional standard Brownian motion under the objective probability measure
P. Under this model, the conditional mean of the increment dYt given Yt is mP(Yt)dt
and the covariance matrix is s(Yt)sT(Yt)dt. When N ¼ 1, the SR will be close to
0 while the MVR will be independent of dt. Thus, when the time period dt is small,
the MVR will be advantageous over the SR.
To further support for the use of MVR, Bai et al. (2011d) document the MVR in
the context of Markowitz MV optimization theory as follows: suppose that there is
p-branch of assets S ¼ (s1,   , sp)T whose returns are denoted by r ¼ (r1,  ∙, rp)T
with mean m ¼ (m1,   , mp)T and covariance matrix S ¼ (sij). In addition, we
suppose that investors will invest capital C on the p-branch of securities S such that
they solve for their optimal investment plans c ¼ (c1,  ∙, cp)T to allocate their
investable wealth on the p-branch of securities to obtain maximize return subject at
a given level of risk.
The above maximization problem can be formulated as the following optimiza-
tion problem:

max R ¼ cT m, subject to cT Sc  s20 (53.5)

where s20 is a given risk level. We call R satisfying Eq. 53.5 the optimal return and
c be its corresponding allocation plan. One could easily extend the separation
theorem and the mutual fund theorem to obtain the solution of Eq. 53.51 from the
following lemma:

1
We note that Bai et al. (2009a, b, 2011c) have also used the same framework as in 53.5.
53 Internet Bubble Examination with Mean-Variance Ratio 1457

Lemma 53.1 For the optimization setting displayed in Eq. 53.5, the optimal return,
R, and its corresponding investment plan, c, are obtained as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R ¼ s0 mT S1 m

and
s0
c ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S1 m: (53.6)
m S1 m
T

From Lemma 53.1, the investment plan, c, is proportional to the MVR when S
is a diagonal matrix. Hence, when the asset is concluded as superior in performance
utilizing the MVR test, its corresponding weight could then be computed based on
the corresponding MVR test value. Thus, another advantage of using the MVR test
over the SR test is that it not only allows investors to compare the performance of
different assets, but it also provides investors with information of the assets weight.
The MVR test enables investors to compute the corresponding allocation for the
assets. On the other hand, as the SR is not proportional to the weight of the
corresponding asset, an asset with the highest SR would not infer that one should
put highest weight on this asset as compared with our MVR. In this sense, the test
proposed by Bai et al. (2011d) is superior to the SR test.
Bai et al. (2011d) have also developed both one-sided UMPU test and two-sided
UMPU test of equality of the MVRs in comparing the performances of different
prospects with hypotheses stated in Eqs. 53.3 and 53.4, respectively. We first state
the one-sided UMPU test for the MVRs as follows:
Theorem 53.1 Let Xi and Yi (i ¼ 1, 2,   , n) be independent random variables with
joint distribution function defined in Eq. 53.1. For the hypotheses setup in Eq. 53.3,
there exists a UMPU level-a test with the critical function f(u, t) such that

1, when u  C0 ðtÞ
fðu; tÞ (53.7)
0, when u < C0 ðtÞ
where C0 is determined by
ð1
f n, t ðuÞ du ¼ K 1 ; (53.8)
C0

with

 
n1
1 !n  1  1
2
u 2
2 ðt1  uÞ 2
f n, t ðuÞ ¼ t2  t3  ,
n n
ð
K 1 ¼ a f n, t ðuÞ du;
O

in which
1458 Z.D. Bai et al.

X
n X
n X
n
U¼ Xi , T1 ¼ Xi þ Yi,
i¼1 i¼1 i¼1
Xn Xn
T2 ¼ X2i , T3 ¼ Y 2i , T ¼ ðT 1 ; T 2 ; T 3 Þ;
i¼1 i¼1

pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi

with O ¼ fujmaxð nt2 , t1  nt3 Þ  u  minð nt2 , t1 þ nt3 Þg to be the
support of the joint density function of (U, T).
We call the statistic U in Theorem 53.1 the one-sided MVR test statistic or
simply the MVR test statistic for the hypotheses setup in Eq. 53.3 if no confusion
arises. In addition, Bai et al. (2011d) have introduced the two-sided UMPU test
statistic as stated in the following theorem to test for the equality of the MVRs listed
in Eq. 53.4:
Theorem 53.2 Let Xi and Yi (i ¼ 1, 2,   , n) be independent random variables with
joint distribution function defined in Eq. 53.1. Then, for the hypotheses setup in Eq.
53.4, there exists a UMPU level-a test with critical function
 
1, when u  C1 ðtÞ or  C2 ðt 
fðu; tÞ ¼ (53.9)
0, when C1 ðtÞ < u < C2 ðt
in which C1 and C2 satisfy
8 ð
>
>
C2
>
< f n, t ðuÞ du ¼ K 2
ð C2
C1
, (53.10)
>
>
>
:

uf n, t ðuÞ du ¼ K 3
C1

where
ð
K 2 ¼ ð 1  aÞ f n, t ðuÞ du,
ðO
K 3 ¼ ð 1  aÞ u f n, t ðuÞ du:
O

The terms fn;t (u), Ti (i ¼ 1, 2, 3) and T are defined in Theorem 53.1.

We call the statistic U in Theorem 53.2 the two-sided MVR test statistic or
simply the MVR test statistic for the hypotheses setup in Eq. 53.4 if no confusion
arises. To obtain the critical values C1 and C2 for the test, readers may refer to
Bai et al. (2011d, 2012).

53.4 Illustration

In this section, we demonstrate the superiority of the MVR tests over the traditional
SR tests by illustrating the applicability of the MVR tests to examine the Internet
bubble during January 2000 and December 2002. For simplicity, we only
53 Internet Bubble Examination with Mean-Variance Ratio 1459

demonstrate the two-sided UMPU test.2 The data for this study consists of weekly
returns on two stock indices: the S&P 500 and the NASDAQ 100 index. The sample
period covers from January 2000 to December 2002 in which the data from the first
week of November 2000 to the last week of January 2001 (3 months) are used to
compute the MVR in January 2001, while the data from the first week of December
2000 to the last week of February 2001 are used to compute the MVR in February
2001, and so on. However, if the period used to compute the SRs is too short, the
result would not be meaningful as discussed in our previous sections. Thus, we
utilize a longer period from the first week of February 2000 to the last week of
January 2001 (12 months) to compute the SR ratio in January 2001, from the first
week of March 2000 to the last week of February 2001 to compute the SR ratio in
February 2001, and so on.
Let X with mean mX and variance s2X be the weekly return on S&P 500 while
Y with mean mY and variance s2Y be the weekly return on the NASDAQ 100 index.
We test the following hypotheses:

mX mY mX mY
H0 : ¼ versus H1 : 6¼ : (53.11)
s2X s2Y s2X s2Y

To test the hypotheses in Eq. 53.11, we first compute the values of the test
function U for the MVR statistic shown in Eq. 53.9, then compute the critical values
C1 and C2 under the test level of 5 % for the pair of indices and display the values in
Table 53.1.
For comparison, we also compute the corresponding SR statistic developed by
Jobson and Korkie (1981) and Memmel (2003) such that

^Y m
s ^X  s^X m
^Y
z¼ pffiffiffi , (53.12)
^y
which follows standard normal distribution asymptotically with

 
1 1 2 2 1 2 2 mX mY 2
y¼ 2sX sY  2sX sY sX, Y þ mX sY þ mY sX 
2 2
s
T 2 2 sX sY X, Y

to test for the equality of the SRs for the funds by setting the following hypotheses
such that

mX mY mX mY
H 0 : ¼ versus H1 : 6¼ : (53.13)
sX sY sX sY

2
The results of the one-sided test which draw a similar conclusion are available on request.
1460 Z.D. Bai et al.

Table 53.1 The results of the mean-variance ratio test and Sharpe ratio test for NASDAQ and
S&P 500, from January 2001 to December 2002
MVR test SR test
Date month/year U C1 C2 Z
01/2001 0.0556 0.1812 0.1267 1.0906
02/2001 0.0636 0.1843 0.1216 1.8765
03/2001 0.1291 0.2291 0.0643 1.1787
04/2001 0.0633 0.2465 0.1633 0.9590
05/2001 0.0212 0.1937 0.2049 0.8313
06/2001 0.0537 0.1478 0.1983 0.8075
07/2001 0.0421 0.1399 0.1132 0.6422
08/2001 0.1062 0.1815 0.0886 0.6816
09/2001 0.1623* 0.1665 0.2728 1.0125
10/2001 0.1106 0.3507 0.1742 0.5931
11/2001 0.0051 0.2386 0.2825 0.1898
12/2001 0.1190 0.0165 0.2041 0.1573
01/2002 0.0316 0.0744 0.1389 0.0157
02/2002 0.0067 0.1389 0.1013 0.0512
03/2002 0.0216 0.1349 0.0853 0.1219
04/2002 0.0444 0.1739 0.0848 0.1885
05/2002 0.0588 0.1766 0.1094 0.0446
06/2002 0.1477 0.2246 0.0267 0.3408
07/2002 0.2167* 0.0101 0.0578 0.0984
08/2002 0.1526* 0.0452 0.1242 0.1024
09/2002 0.2121* 0.0218 0.0551 0.6304
10/2002 0.0416 0.1249 0.2344 0.0361
11/2002 0.0218 0.1056 0.2150 0.0008
12/2002 0.1265 0.0015 0.2417 0.3908
Note: The MVR test statistic U is defined in Eq. 53.9 and its critical values C1 and C2 are defined in
Eqs. 53.10, respectively. The SR test statistic Z is defined in Eq. 53.12. The level is a ¼ 0.05, and
“*” means significant at levels 5 %. Here, the sample size of the MVR test is 3 months, while the
sample size of the SR test 12 months. Recall that  z0.025 1.96

Instead of using a 2-month data to compute the values of our proposed statistic,
we use the overlapping 12-month data to compute the SR statistic. The results are
also reported in Table 53.1.
The limitation of applying the SR test is that it would usually conclude indis-
tinguishable performances between the indices, which may not be the situation in
reality. In this aspect, looking for a statistic to evaluate the difference between
indices for short periods is essential. The situation in reality is that the Internet
stocks registered large gains in the late 1990s, followed by large losses from 2000.
As we mentioned before, the NASDAQ 100 index comprises 100 of the largest
domestic and international technology firms including those in the computer hard-
ware and software, telecommunications, and biotechnology sectors, while the S&P
53 Internet Bubble Examination with Mean-Variance Ratio 1461

5000 NASDAQ
S&P 500
4000
Index

3000

2000

1000
01/03/2000
03/13/2000
05/22/2000
07/31/2000
10/09/2000
12/18/2000
02/26/2001
05/07/2001
07/16/2001
10/01/2001
12/10/2001
02/19/2002
04/29/2002
07/08/2002
09/16/2002
11/25/2002
02/03/2003
04/14/2003
06/23/2003
09/02/2003
11/10/2003
Fig. 53.1 Weekly indices of NASDAQ and S&P 500 from January 3, 2000 to December 31, 2003

500 index represents non-technology or “old economy” firms. After the bursting of
the Internet bubble in the 2000s, as shown in Fig. 53.1, the NASDAQ 100 declined
much more and underperformed the S&P 500. From Table 53.1, we find that the
MVR test statistic does not disappoint us in that it does pick up significant
differences in performances between the S&P 500 and the NASDAQ 100 index
in September 2001, July 2002, August 2002, and September 2002, but SR test does
not conclude any distinguishable performances between the indices. Further to
say, from Table 53.1, we observe that m ^X > m ^ Y in September 2001, July 2002,
August 2002, and September 2002. This infers that the MVR test statistics can
detect the real situation that the NASDAQ 100 index underperformed the S&P
500 index, but the traditional SR test cannot detect any difference. Thus, we
conclude that investors could be able to profiteer from the Internet bubble if they
apply the MVR test.

53.5 Concluding Remarks

In this paper, we employ the MVR test statistics developed by Bai et al. (2011d) to
examine the performances between the S&P 500 index and the NASDAQ
100 index during Internet bubble from January 2000 to December 2002. We
illustrate the superiority of the MVR test over the traditional SR test by applying
both tests to analyze the performance of the S&P 500 index and the NASDAQ
100 index after the bursting of the Internet bubble in the 2000s. Our findings show
that while the traditional SR test concludes the two indices being analyzed to be
indistinguishable in their performance, the MVR test statistic shows that the
NASDAQ 100 index underperformed the S&P 500 index, which is the real situation
1462 Z.D. Bai et al.

after the bursting of the Internet bubble in the 2000s. This shows the superiority of
the MVR test statistic in revealing short-term performance and, in turn, enables the
investors to make better decisions about their investments.
There are two basic approaches to the problem of portfolio selection under
uncertainty. One approach is based on the concept of utility theory (Gasbarro
et al. 2007; Wong et al. 2006, 2008). Several stochastic dominance (SD) test statistics
have been developed; see, for example, Bai et al. (2011a) and the references therein
for more information. This approach offers a mathematically rigorous treatment for
portfolio selection, but it is not popular among investors since investors would have
to specify their utility functions and choose a distributional assumption for the returns
before making their investment decisions.
The other approach is the mean-risk (MR) analysis that has been discussed in
this paper. In this approach, the portfolio choice is made with respect to two
measures – the expected portfolio mean return and portfolio risk. A portfolio is
preferred if it has higher expected return and smaller risk. These are convenient
computational recipes and they provide geometric interpretations for the trade-off
between the two measures. A disadvantage of the latter approach is that it is derived
by assuming the Von Neumann-Morgenstern quadratic utility function and that
returns are normally distributed (Hanoch and Levy 1969). Thus, it cannot capture
the richness of the former approach. Among the MR analyses, the most popular
measure is the SR introduced by Sharpe (1966). As the SR requires strong assump-
tions that the returns of assets being analyzed have to be iid, various measures for
MR analysis have been developed to improve the SR, including the Sortino ratio
(Sortino and van der Meer 1991), the conditional SR (Agarwal and Naik 2004), the
modified SR (Gregoriou and Gueyie 2003), value at risk (Ma and Wong 2010),
expected shortfall (Chen 2008), and the mixed Sharpe ratio (Wong et al. 2012).
However, most of the empirical studies, see, for example, Eling and Schuhmacher
(2007), find that the conclusions drawn by using these ratios are basically the same
as that drawn by the SR. Nonetheless, Leung and Wong (2008) have developed
a multiple SR statistic and find that the results drawn from the multiple Sharpe ratio
statistic can be different from its counterpart pair-wise SR statistic comparison,
indicating that there are some relationships among the assets that have not being
revealed using the pair-wise SR statistics. The MVR test could be the right
candidate to reveal these relationships.
One may claim that the limitation of the MVR test statistic is that it can only
draw conclusion for investors with quadratic utility functions and for normal-
distributed assets. Wong (2006), Wong and Ma (2008), and others have shown
that the conclusion drawn from the MR comparison is equivalent to the comparison
of expected utility maximization for any risk-averse investor, not necessarily with
only quadratic utility function, and for assets with any distribution, not necessarily
normal distribution, if the assets being examined belong to the same location-scale
family. In addition, one can also apply the results from Li and Wong (1999) and
Egozcue and Wong (2010) to generalize the result so that it will be valid for any
risk-averse investor and for portfolios with any distribution if the portfolios
being examined belong to the same convex combinations of (same or different)
53 Internet Bubble Examination with Mean-Variance Ratio 1463

location-scale families. The location-scale family can be very large, containing

normal distributions as well as t-distributions, gamma distributions, etc. The stock
returns could be expressed as convex combinations of normal distributions,
t-distributions, and other location-scale families; see, for example, Wong and
Bian (2000) and the references therein for more information. Thus, the conclusions
drawn from the MVR test statistics are valid for most of the stationary data
including most, if not all, of the returns of different portfolios.
Last, we note that to improve the effectiveness of applying the MVR test in
evaluating financial assets performance, one may incorporate other techniques/
approaches/models, for example, fundamental analysis (Wong and Chan 2004),
technical analysis (Wong et al. 2001, 2003), behavioral finance (Matsumura
et al. 1990), prospect theory (Broll et al. 2010; Egozcue et al. 2011), and advanced
econometrics (Wong and Miller 1990; Bai et al. 2010, 2011b), to measure the
performance of different financial assets and assist investors to make wiser decisions.

Acknowledgment We would like to thank the editor C.-F. Lee for his substantive comments that
have significantly improved this manuscript. The third author would also like to thank Professors
Robert B. Miller and Howard E. Thompson for their continuous guidance and encouragement. The
research is partially supported by grants from North East Normal University, National University
of Singapore, Hong Kong Baptist University and the Research Grants Council of Hong Kong. The
first author thanks the financial support from NSF China grant 11171057, Program for Changjiang
Scholars and Innovative Research Team in University, and the Fundamental Research Funds for
the Central Universities and NUS grant R-155-000-141-112.

References
Agarwal, V., & Naik, N. Y. (2004). Risk and portfolios decisions involving hedge funds. Review of
Financial Studies, 17, 63–98.
Bai, Z. D., Liu, H. X., & Wong, W. K. (2009a). Enhancement of the applicability of Markowitz’s
portfolio optimization by utilizing random matrix theory. Mathematical Finance, 19, 639–667.
Bai, Z. D., Liu, H. X., & Wong, W. K. (2009b). On the markowitz mean-variance analysis of self-
financing portfolios. Risk and Decision Analysis, 1, 35–42.
Bai, Z. D., Wong, W. K., & Zhang, B. Z. (2010). Multivariate linear and non-linear causality tests.
Mathematics and Computers in Simulation, 81, 5–17.
Bai, Z. D., Li, H., Liu, H. X., & Wong, W. K. (2011a). Test statistics for prospect and markowitz
stochastic dominances with applications. Econometrics Journal, 14, 278–303.
Bai, Z. D., Li, H., Wong, W. K., & Zhang, B. Z. (2011b). Multivariate causality tests with
simulation and application. Statistics and Probability Letters, 81, 1063–1071.
Bai, Z. D., Liu, H. X., & Wong, W. K. (2011c). Asymptotic properties of eigenmatrices of a large
sample covariance matrix. Annals of Applied Probability, 21, 1994–2015.
Bai, Z. D., Wang, K. Y., & Wong, W. K. (2011d). Mean-variance ratio test, a complement to
coefficient of variation test and Sharpe ratio test. Statistics and Probability Letters,
81, 1078–1085.
Bai, Z. D., Hui, Y. C., Wong, W. K., & Zitikis, R. (2012). Evaluating prospect performance:
Making a case for a non-asymptotic UMPU test. Journal of Financial Econometrics, 10(4),
703–732.
Baker, M., & Stein, J. C. (2004). Market liquidity as a sentiment indicator. Journal of Financial
Markets, 7, 271–300.
1464 Z.D. Bai et al.

Broll, U., Wahl, J. E., & Wong, W. K. (2006). Elasticity of risk aversion and international trade.
Economics Letters, 91, 126–130.
Broll, U., Egozcue, M., Wong, W. K., & Zitikis, R. (2010). Prospect theory, indifference curves,
and hedging risks. Applied Mathematics Research Express, 2010, 142–153.
Cadsby, C. B. (1986). Performance hypothesis testing with the Sharpe and Treynor measures:
A comment. Journal of Finance, 41, 1175–1176.
Chan, C. Y., de Peretti, C., Qiao, Z., & Wong, W. K. (2012). Empirical test of the efficiency of UK
covered warrants market: Stochastic dominance and likelihood ratio test approach. Journal of
Empirical Finance, 19, 162–174.
Chen, S. X. (2008). Nonparametric estimation of expected shortfall. Journal of Financial Econo-
metrics, 6, 87–107.
Egozcue, M., & Wong, W. K. (2010). Gains from diversification on convex combinations:
A majorization and stochastic dominance approach. European Journal of Operational
Research, 200, 893–900.
Egozcue, M., Fuentes Garcı́a, L., Wong, W. K., & Zitikis, R. (2011). Do investors like to
diversify? A study of Markowitz preferences. European Journal of Operational Research,
215, 188–193.
Eling, M., & Schuhmacher, F. (2007). Does the choice of performance measure influence the
evaluation of hedge funds? Journal of Banking and Finance, 31, 2632–2647.
Fong, W. M., Wong, W. K., & Lean, H. H. (2005). International momentum strategies:
A stochastic dominance approach. Journal of Financial Markets, 8, 89–109.
Fong, W. M., Lean, H. H., & Wong, W. K. (2008). Stochastic dominance and behavior towards
risk: The market for internet stocks. Journal of Economic Behavior and Organization,
68, 194–208.
Gasbarro, D., Wong, W. K., & Zumwalt, J. K. (2007). Stochastic dominance analysis of iShares.
European Journal of Finance, 13, 89–101.
Gregoriou, G. N., & Gueyie, J. P. (2003). Risk-adjusted performance of funds of hedge funds using
a modified Sharpe ratio. Journal of Wealth Management, 6, 77–83.
Hanoch, G., & Levy, H. (1969). The efficiency analysis of choices involving risk. Review of
Economic Studies, 36, 335–346.
Jobson, J. D., & Korkie, B. (1981). Performance hypothesis testing with the Sharpe and Treynor
measures. Journal of Finance, 36, 889–908.
Lam, K., Liu, T., & Wong, W. K. (2010). A pseudo-Bayesian model in financial decision making
with implications to market volatility, under- and overreaction. European Journal of Opera-
tional Research, 203, 166–175.
Lam, K., Liu, T., & Wong, W. K. (2012). A new pseudo Bayesian model with implications to
financial anomalies and investors’ behaviors. Journal of Behavioral Finance, 13(2), 93–107.
Lean, H. H., Phoon, K. F., & Wong, W. K. (2012). Stochastic dominance analysis of CTA funds.
Review of Quantitative Finance and Accounting, doi:10.1007/s11156-012-0284-1.
Leung, P. L., & Wong, W. K. (2008). On testing the equality of the multiple Sharpe ratios, with
application on the evaluation of iShares. Journal of Risk, 10, 1–16.
Li, C. K., & Wong, W. K. (1999). Extension of stochastic dominance theory to random variables.
RAIRO Recherche Opérationnelle, 33, 509–524.
Lintner, J. (1965). The valuation of risky assets and the selection of risky investment in stock
portfolios and capital budgets. Review of Economics and Statistics, 47, 13–37.
Lo, A. (2002). The statistics of Sharpe ratios. Financial Analysis Journal, 58, 36–52.
Ma, C., & Wong, W. K. (2010). Stochastic dominance and risk measure: A decision-theoretic
foundation for VaR and C-VaR. European Journal of Operational Research, 207, 927–935.
Markowitz, H. M. (1952). Portfolio selection. Journal of Finance, 7, 77–91.
Matsumura, E. M., Tsui, K. W., & Wong, W. K. (1990). An extended multinomial-Dirichlet model
for error bounds for dollar-unit sampling. Contemporary Accounting Research, 6, 485–500.
Memmel, C. (2003). Performance hypothesis testing with the Sharpe ratio. Finance Letters,
1, 21–23.
53 Internet Bubble Examination with Mean-Variance Ratio 1465

Ofek, E., & Richardson, M. (2003). A survey of market efficiency in the internet sector. Journal of
Finance, 58, 1113–1138.
Perkins, A. B., & Perkins, M. C. (1999). The internet bubble: Inside the overvalued world of high-
tech stocks. New York: Harper Business.
Schultz, P., & Zaman, M. (2001). Do the individuals closest to internet firms believe they are
overvalued? Journal of Financial Economics, 59, 347–381.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Journal of Finance, 19, 425–442.
Sharpe, W. F. (1966). Mutual funds performance. Journal of Business, 39, 119–138.
Sortino, F. A., & van der Meer, R. (1991). Downside risk. Journal of Portfolio Management,
17, 27–31.
Sriboonchitta, S., Wong, W. K., Dhompongsa, D., & Nguyen, H. T. (2009). Stochastic dominance
and applications to finance, risk and economics. Boca Raton: Chapman and Hall/CRC.
Statman, M. (2002). Lottery players/stock traders. Journal of Financial Planning, 14–21.
Tobin, J. (1958). Liquidity preference as behavior towards risk. Review of Economic Studies,
25, 65–86.
Wong, W. K. (2006). Stochastic dominance theory for location-scale family. Journal of Applied
Mathematics and Decision Sciences, 2006, 1–10.
Wong, W. K. (2007). Stochastic dominance and mean-variance measures of profit and loss
for business planning and investment. European Journal of Operational Research,
182, 829–843.
Wong, W. K., & Bian, G. (2000). Robust Bayesian inference in asset pricing estimation. Journal of
Applied Mathematics & Decision Sciences, 4, 65–82.
Wong, W. K., & Chan, R. (2004). The estimation of the cost of capital and its reliability.
Quantitative Finance, 4, 365–372.
Wong, W. K., & Chan, R. (2008). Markowitz and prospect stochastic dominances. Annals of
Finance, 4, 105–129.
Wong, W. K., & Li, C. K. (1999). A note on convex stochastic dominance theory. Economics
Letters, 62, 293–300.
Wong, W. K., & Ma, C. (2008). Preferences over location-scale family. Economic Theory,
37, 119–146.
Wong, W. K., & Miller, R. B. (1990). Analysis of ARIMA-noise models with repeated time series.
Journal of Business and Economic Statistics, 8, 243–250.
Wong, W. K., Chew, B. K., & Sikorski, D. (2001). Can P/E ratio and bond yield be used to beat
stock markets? Multinational Finance Journal, 5, 59–86.
Wong, W. K., Manzur, M., & Chew, B. K. (2003). How rewarding is technical analysis? Evidence
from Singapore stock market. Applied Financial Economics, 13, 543–551.
Wong, W. K., Thompson, H. E., Wei, S., & Chow, Y. F. (2006). Do winners perform better than
losers? A stochastic dominance approach. Advances in Quantitative Analysis of Finance and
Accounting, 4, 219–254.
Wong, W. K., Phoon, K. F., & Lean, H. H. (2008). Stochastic dominance analysis of Asian hedge
funds. Pacific-Basin Finance Journal, 16, 204–223.
Wong, W. K., Wright, J. A., Yam, S. C. P., & Yung, S. P. (2012). A mixed Sharpe ratio. Risk and
Decision Analysis, 3(1–2), 37–65.
Quantile Regression in Risk Calibration
54
Shih-Kang Chao, Wolfgang Karl Härdle, and Weining Wang

Contents
54.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1468
54.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1472
54.2.1 Constructing Partial Linear Model (PLM) for CoVaR . . . . . . . . . . . . . . . . . . . . . . . 1472
54.2.2 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1475
54.2.3 Risk Contribution Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1476
54.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1478
54.3.1 CoVaR Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1478
54.3.2 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1480
54.3.3 Global Risk Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1482
54.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1484
Appendix 1: Local Linear Quantile Regression (LLQR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1484
Appendix 2: Confidence Band for Nonparametric Quantile Estimator . . . . . . . . . . . . . . . . . . . . . . 1485
Appendix 3: PLM Model Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1487
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1488

Abstract
Financial risk control has always been challenging and becomes now an even
harder problem as joint extreme events occur more frequently. For decision
makers and government regulators, it is therefore important to obtain accurate

€
The financial support from the Deutsche Forschungsgemeinschaft via SFB 649 “Okonomisches
Risiko,” Humboldt-Universität zu Berlin is gratefully acknowledged
S.-K. Chao (*) • W. Wang
Ladislaus von Bortkiewicz Chair of Statistics, C.A.S.E. – Center for Applied Statistics and
Economics, Humboldt–Universität zu Berlin, Berlin, Berlin, Germany
e-mail: shih-kang.chao@cms.hu-berlin.de; wangwein@cms.hu-berlin.de
W.K. Härdle
Ladislaus von Bortkiewicz Chair of Statistics, C.A.S.E. – Center for Applied Statistics and
Economics, Humboldt–Universität zu Berlin, Berlin, Berlin, Germany
Lee Kong Chian School of Business, Singapore Management University, Singapore, Singapore
e-mail: haerdle@wiwi.hu-berlin.de

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1467
DOI 10.1007/978-1-4614-7750-1_54,
# Springer Science+Business Media New York 2015
1468 S.-K. Chao et al.

information on the interdependency of risk factors. Given a stressful situation for

one market participant, one likes to measure how this stress affects other factors.
The CoVaR (Conditional VaR) framework has been developed for this purpose.
The basic technical elements of CoVaR estimation are two levels of quantile
regression: one on market risk factors; another on individual risk factor.
Tests on the functional form of the two-level quantile regression reject the
linearity. A flexible semiparametric modeling framework for CoVaR is pro-
posed. A partial linear model (PLM) is analyzed. In applying the technology to
stock data covering the crisis period, the PLM outperforms in the crisis time,
with the justification of the backtesting procedures. Moreover, using the data on
global stock markets indices, the analysis on marginal contribution of risk
(MCR) defined as the local first order derivative of the quantile curve sheds
some light on the source of the global market risk.

Keywords
CoVaR • Value-at-Risk • Quantile regression • Locally linear quantile regression •
Partial linear model • Semiparametric model

54.1 Introduction

Sufficiently accurate risk measures are needed not only in crisis times. In the last
two decades, the world has gone through several financial turmoils, and the
financial market is getting riskier and the scale of loss soars. Beside marginal
extremes that can shock even a well-diversified portfolio, the focus of intensified
research in the recent years has been on understanding the interdependence of risk
factors and their conditional structure.
The most popular risk measure is the Value-at-Risk (VaR), which is defined as the
t-quantile of the return distribution at time t + d conditioned on the information set F t:

VaRttþd ¼ inf fx 2 ℝ : PðXtþd  xjF t Þ  tg:

def
(54.1)

Here Xt denotes the asset return and t is taking values such as 0.05, 0.01 or 0.001
to reflect negative extreme risk.
Extracting information in economic variables to predict VaR brings quantile
regression into play here, since VaR is the quantile of the conditional asset return
distribution. Engle and Manganelli (2004) propose the nonlinear Conditional
Autoregressive Value-at-Risk (CaViaR) model, which uses (lag) VaR and lag returns.
Chernozhukov and Umantsev (2001) propose linear and quadratic time series models
for VaR prediction. Kuan et al. (2009) propose the Conditional AutoRegressive
Expectile (CARE) model, and argue that expectiles are more sensitive to the scale
of losses. These studies and many others apply quantile regression in a prespecified,
often linear functional form. In a more nonparametric context, Cai and Wang (2008)
estimate the conditioned cdf by a double kernel local linear estimator and find the
54 Quantile Regression in Risk Calibration 1469

quantile by inverting the cdf. Schaumburg (2011) uses the same technique together
with extreme value theory for VaR prediction. Taylor (2008) proposes Exponentially
Weighted Quantile Regression (EWQR) for estimating VaR time series.
The aforementioned studies focus mainly on the VaR estimation for single assets
and do not directly take into account the escalated spillover effect in crisis periods.
This risk of joint tail events of asset returns has been identified and studied. Further,
Brunnermeier and Pedersen (2008) show that the negative feedback effect of a “loss
spiral” and a “margin spiral” leads to the joint depreciation of assets prices. It is
therefore important to develop risk measures which can quantify the contagion effects
of negative extreme event.
Acharya et al. (2010) propose the concept of marginal expected shortfall (MES),
which measures the contribution of individual assets to the portfolio expected shortfall.
Via an equilibrium argument, the MES is shown to be a predictor to a financial
institution’s risk contribution. Brownlees and Engle (2012) demonstrate that the
MES can be written as a function of volatility, correlation, and expectation conditional
on tail events. Huang et al. (2012) propose the distress insurance premium (DIP),
a measure similar to MES but computed under the risk-neutral probability. This
measure can therefore be viewed as the market insurance premium against the event
that the portfolio loss exceeds a low level. Adams et al. (2012) construct financial
indices on return of insurance companies, commercial banks, investment banks, and
hedge funds, and use a linear model for the VaRs of the four financial indices to forecast
the state-dependent sensitivity VaR (SDSVaR). The risk measures proposed above
have some shortcomings though: The computation of DIP is demanding since this
involves the simulation of rare events. MES suffers from the scarcity of data because it
conditions on a rare event.
In Adrian and Brunnermeier (2011) (henceforth AB), the CoVaR concept of
conditional VaR is proposed, which controls the effect of the negative extreme event
of some systemically risky financial institutions. Formally, let C(Xi,t) be some event of
a asset i return Xi,t at time t and take Xj,t as another asset return (e.g., the market index).
The CoVaRtj|i,t is defined as the t-quantile of the conditional probability distribution:
n
  o


P Xj, t  CoVaRtjji, t
C Xi, t , Mt ¼ t, (54.2)

where Mt is a vector of market variables defined inoSect. 54.2.1. The standard

  n
CoVaR approach is to set C Xi, t ¼ Xi, t ¼ VaRtXi, t . In AB, Xj,t is the weekly
return which is constructed from a vast data set comprised of all publicly traded
commercial banks, broker dealers, insurance companies, and real estate companies
in the USA. Further, AB propose DCoVaR (measure of marinal risk contribution) as
the difference between CoVaRtjji1, t and CoVaRtjji2, t, where t1 ¼ 0.5 is associated with
the normal state and t2 ¼ 0.05 is associated with the financial distress state.
The formulation of this conditional risk measure has several advantages. First, the
cloning property: After dividing a systemically risky firm into several clones, the
value of CoVaR conditioned on the entire firm does not differ from the one condi-
tioned on one of the clones. Second, the conservativeness. The CoVaR value is more
1470 S.-K. Chao et al.

conservative than VaR, because it conditions on an extreme event. Third, CoVaR is

endogenously generated and adapted to the varying environment of the market.
The recipe of AB for CoVaR construction is as follows: In the first step, one predicts
the VaR of an individual asset Xi, t through a linear model on market variables:
Xi, t ¼ ai þ gΤi Mt1 þ ei, t , (54.3)
where gΤi means the transpose of gi and Mt is a vector of the state variables
(see Sect. 54.2.1). This model is estimated with quantile regression of Koenker
and Bassett (1978) to get the coefficients ð^a i ; ^g i Þ with F1
ei, t ðtjMt1 Þ ¼ 0. The VaR of
asset i is predicted by

d i, t ¼ ^a i þ ^g Τ Mt1 :
VaR (54.4)
i

In the second step, one models the asset j return as a linear function of asset
return i and market variables Mt:

Xj, t ¼ ajji þ bjji Xi, t þ gΤjji Mt1 þ ej, t , (54.5)

 
Again one employs quantile regression and obtains coefficients ^a jji ; b^jji ; ^g jji .
The CoVaR is finally calculated as:

d AB ¼ ^a jji þ b^jji VaR

CoVaR d i, t þ ^g Τ Mt1 :
jji (54.6)
jji, t

In Eq. 54.5, the variable Xi,t influences the return Xj,t in a linear fashion.
However, the linear parametric model may not be flexible enough to capture the
tail dependence between i and j. The linearity of the conditioned quantile curves of
Xj on Xi is challenged by the confidence bands of the nonparametric quantile curves,
as shown in Fig. 54.1. The left tail quantile from linear parametric quantile
regression (red) lies well outside the confidence band (gray dashed curve) of Hardle
and Song (2010). This motivates empirically that a linear model is not flexible
enough for the CoVaR question at hand.
Nonparametric models can be used to account for the nonlinear structure of the
conditional quantile, but the challenge for using such models is the curse of dimension-
ality, as the quantile regression in CoVaR modeling often involves many variables. Thus,
we resort to semiparametric partial linear model (PLM) which preserves some flexibility
of the nonparametric model while suffers little from the curse of dimensionality.
As an illustration, the VaR/CoVaR of Goldman Sachs (GS) returns are shown,
given the returns of Citigroup (C) and S&P500 (SP). S&P500 index return is used
as a proxy for the market portfolio return.
Choosing market variables is crucial for the VaR/CoVaR estimation. For the
variables representing market states, we follow the most popular choices such as
VIX, short-term liquidity spread, etc. In particular, the variable we use for real
estate companies is the Dow Jones U.S. real estate index. The daily data date from
August 4, 2006 to August 4, 2011.
54 Quantile Regression in Risk Calibration 1471

0.0
0.0

−0.5
−0.5

0.0 0.5 0.0 0.5

Fig. 54.1 Goldman Sachs (GS) and Citigroup (C) weekly returns 0.05(left) and 0.1(right)
quantile functions. The y-axis is GS daily returns and the x-axis is the C daily returns.
The blue curve are the locally linear quantile regression curves (see Appendix 1). The locally
linear quantile regression bandwidth are 0.1026 and 0.0942. The red lines are the linear parametric
quantile regression line. The antique white dashed curves are the asymptotic confidence band
(see Appendix 2) with significance level 0.05. The sample size N ¼ 546

To see if the estimated VaRs/CoVaRs are accurate, we utilize the backtesting

procedures described in Berkowitz et al. (2011). We compare three (Co)VaR estimat-
ing methods in this study: VaR computed by linear quantile regression on market
variables; CoVaR; PLM CoVaR proposed here. The VaR is one-sided interval predic-
tion, the violations (the asset return exceeds estimated VaR/CoVaR) should happen
unpredictably if the VaR algorithm is accurate. In other words, the null hypothesis is
that the series of violations of VaR is a martingale difference, given all the past
information. Furthermore, if the time series is autocorrelated, we can reject the null
hypothesis of martingale difference right away; therefore, autocorrelation tests can be
utilized in this context. The Ljung-Box test is not the most appropriate approach here
since it has a too strong null hypothesis (i.i.d. sequence). Thus, we additionally apply
the Lobato test. The CaViaR test, which is inspired by the CaViaR model, is proposed
and shown to have the best overall performance by Berkowitz et al. (2011) among other
alternative tests with an exclusive desk-level data set. To illustrate the VaR/CoVaR
performances in the crisis time, we separately apply the CaViaR test to the violations of
the whole sample period and to the financial crisis period.
The results show that during the financial crisis period from mid-2008 to mid-2009,
the PLM CoVaR of GS given C performs better than that constructed from the
technique of AB and the PLM CoVaR given SP. In particular, these results suggest
that with appropriate modeling techniques (accounting for nonlinearity), the CoVaR of
GS calculated from conditioning on C reflects some structurally risk which is not
reflected from conditioning on market returns such as SP during financial crisis.
In contrast to DCoVaR, we use a mathematically more intuitive way to analyze
the marginal effect by taking the first order derivative of the quantile function.
1472 S.-K. Chao et al.

We call it “marginal contribution of risk” (MCR). Bae et al. (2003) and many others
have pointed out the phenomenon of financial contagion across national borders. This
motivates us to consider the stock indices of a few developed markets and explore
their risk contribution to the global stock market. MCR results show that when the
global market condition varies, the source of global market risk can be different. To
be more specific, when the global market return is bad, the risk contribution from the
USA is the largest. On the other hand, during financially stable periods, Hong Kong
and Japan are more significant risk contributors than the USA to the global market.
This study is organized as follows: Sect. 54.2 introduces the construction and
the estimation of the PLM model of CoVaR. The backtesting methods and our
risk contribution measure are also introduced in this section. Section 54.3 presents
the Goldman Sachs CoVaR time series and the backtesting procedure results.
Section 54.4 presents the conclusion and possible further studies. Appendices
describe the detailed estimation and statistical inference procedures used in
this study.

54.2 Methodology

Quantile regression is a well-established technique to estimate the conditional quantile

function. Koenker and Bassett (1978) focus on the linear functional form. An extension
of linear quantile regression is the PLM quantile regression. A partial linear model for
the dynamics of assets return quantile is constructed in this section. The construction is
justified by a linearity test based on a conservative uniform confidence band proposed
in Hardle and Song (2010). For more details on semiparametric modeling and PLM, we
refer to Härdle et al. (2004) and Härdle et al. (2000).
The backtesting procedure is done via the CaViaR test. Finally, the methodology
of MCR is introduced, which is an intuitive marginal risk contribution measure. We
will apply the method to a data set of global market indices in developed countries.

54.2.1 Constructing Partial Linear Model (PLM) for CoVaR

Recall how the CoVaR is constructed:

d t ¼ ^a i þ ^g i Mt1 ,
VaRi,
AB
d
CoVaR a jji þ b^jji VaR
jji, t ¼ ^
d i, t þ ^gΤ Mt1 :
jji
 
where ð^a i ; ^g i Þ and ^a jji ; b^jji ; ^g jji are estimated from a linear model using standard
linear quantile regression.
We have motivated the need for more general functional forms for the quantile
curve. We therefore relax the model to a non- or semiparametric model. The market
54 Quantile Regression in Risk Calibration 1473

variable Mt is multidimensional, and the data frequency here is daily. The following
key variables are entering our analysis:
1. VIX: Measuring the model-free implied volatility of the market. This index is
known as the “fear gauge” of investors. The historical data can be found in the
Chicago Board Options Exchange’s website.
2. Short-term liquidity spread: Measuring short-term liquidity risk by the differ-
ence between the 3-month treasury repo rate and the 3-month treasury bill rate.
The repo data is from the Bloomberg database and the treasury bill rate data is
from the Federal Reserve Board H.15.
3. The daily change in the 3-month treasury bill rate: AB find that the changes have
better explanatory power than the levels for the negative tail behavior of asset returns.
4. The change in the slope of the yield curve: The slope is defined by the difference
of the 10-year treasury rate and the 3-month treasury bill rate.
5. The change in the credit spread between 10 years BAA-rated bonds and the
10 years treasury rate.
6. The daily Dow Jones U.S. Real Estate index returns: The index reflects the
information of lease rates, vacancies, property development, and transactions of
real estates in the USA.
7. The daily S&P500 index returns: The approximate of the theoretical market
portfolio returns.
The variables 3, 4, 5 are from the Federal Reserve Board H.15 and the data of
6 and 7 are from Yahoo Finance.
First we conduct a statistical check of the linearity between GS return and the
market variables using the confidence band as constructed in Appendix 2. As shown
in Fig. 54.2, except for some ignorable outsiders, the linear quantile regression line
lies in the LLQR asymptotic confidence band.
On the other hand, there is nonlinearity between two individual assets Xi and Xj.
To illustrate this, we regress Xj on Mt, and then take the residuals and regress them
on Xi. Again the Xj,t is GS daily return and Xi is C daily return. The result is shown
in Fig. 54.3. The linear QR line (red) lies well outside the LLQR confidence band
(magenta) when the C return is negative. The linear quantile regression line is fairly
flat. The risk of using a linear model is obvious in this figure: The linear regression
can “average out” the humped relation of the underlying structure (blue), and
therefore imply a model risk in estimation.
Based on the results of the linearity tests above, we construct a PLM model:

Xi, t ¼ ai þ gΤi Mt1 þ ei, t , (54.7)

Τ  
Xj, t ¼ ^a jji þb^jji Mt1 þ ljji Xi, t þ ej, t , (54.8)

where Xi,t, Xj,t are asset returns of i, j firms. Mt is a vector of market variables at time
t as introduced before. If i ¼ S&P500, Mt is set to consist of the first 6 market
variables only. Notice the variable Xi,t enter the Eq. 54.8 nonlinearly.
1474 S.-K. Chao et al.

0.2
0.0
0.0

−0.4
−0.3

0.1 0.3 0.5 0.7 −1.5 −1.0 −0.5 0.0 0.5

VIX Liquidity Spread

0.2 0.2

0.0
0.0

−0.3 −0.2
−0.5 0.0 0.5 0.00 0.01 0.02 0.03 0.04
Change in yields of 3 mon. TB Slope of yield curve

0.2 0.2

0.0 0.0

−0.3 −0.3
−0.001 0.001 0.003 −0.05 0.00 0.05 0.10
Credit Spread S&P500 Index Returns

0.2

0.0

−0.2

−0.2 −0.1 0.0 0.1 0.2

DJUSRE Index Returns

Fig. 54.2 The scatter plots of GS daily returns to the seven market variables with the LLQR
curves. The bandwidths are selected by the method described in Appendix 1. The LLQR band-
widths are 0.1101, 0.1668, 0.2449, 0.0053, 0.0088, 0.0295 and 0.0569. The data period is from
August 4, 2006, to August 4, 2011. N ¼ 1260. t ¼ 0.05

Applying the algorithm of Koenker and Bassett (1978) to Eq. n54.7 and the o
process described in Appendix 3 to Eq. 54.8, we get f^a i ; ^g i g and ^a jji , b^i , ^l ðÞ
 
with F1 1
ei, t ðtjMt1 Þ ¼ 0 for Eq. 54.7 and Fei, t tjM t1 , X i, t ¼ 0 for Eq. 54.8. Finally,
we estimate the PLM CoVaRjji,t by
54 Quantile Regression in Risk Calibration 1475

0.2

0.0

−0.2

−0.4

−0.6

−0.8
−0.15 −0.10 −0.05 0.00 0.05 0.10

Fig. 54.3 The nonparametric part l^GSjC ðÞ of the PLM estimation. The y-axis is the GS daily
returns. The x-axis is the C daily returns. The blue curve is the LLQR quantile curve. The red line
is the linear parametric quantile line. The magenta dashed curves are the asymptotic confidence
band with significance level 0.05. The data is from June 25, 2008, to December 23, 2009.
378 observations. Bandwidth ¼ 0.1255. t ¼ 0.05

d i, t ¼ ^a i þ ^g Τi Mt1 ,
VaR (54.9)

 
^Τ
a^jji þ bej Mt1 þ ^l jji VaR
PLM
d d i, t :
CoVaR jji, t ¼ e (54.10)

54.2.2 Backtesting

The goal of the backtesting procedure is to check if the VaR/CoVaR is accurate

enough so that managerial decisions can be made based on them. The VaR forecast
is a (one-sided) interval forecast. If the VaR algorithm is correct, then the violations
should be unpredictable, after using all the past information. Formally, if we define
the violation time series as

1, dt ;
if Xt < VaR
It ¼ t
0, otherwise:

where VaRdt can be replaced by CoVaR d t in the case of CoVaR. It should form
t t
a sequence of martingale difference.
There is a large literature on martingale difference tests. We adopt
Ljung-Box test, Lobato test, and the CaViaR test. The Ljung-Box test and Lobato
test aim to check whether the time series is autocorrelated. If the time series is
autocorrelated, then we reject of course the hypothesis that the time series is a
martingale difference.
^ k be the estimated autocorrelation of lag k of the sequence
Particularly, let r
of violation {It} and n be the length of the time series. The Ljung-Box test
statistics is
1476 S.-K. Chao et al.

Xm
^ 2k
r L
LBðmÞ ¼ nðn þ 2Þ ! wðmÞ, (54.11)
k¼1
nk

as n ! 1.
This test is too strong though in the sense that the asymptotic distribution is
derived based on the i.i.d. assumption. A modified Box-Pierce test is proposed by
Lobato et al. (2001), who also consider the test of no autocorrelation, but their test
is more robust to the correlation of higher (greater than the first) moments.
(Autocorrelation in higher moments does not contradict with the martingale differ-
ence hypothesis.) The test statistics is given by
Xm
^ 2k L
r
Lð m Þ ¼ n ! wðmÞ,
^v
k¼1 kk

as n ! 1, where
Xnk  2
1
n ðyi  yÞ2 yiþk  y
^v kk ¼ i¼1
n Xn o2 :
2
1
N ð
i¼1 i
y  y Þ

The CaViaR test, proposed by Berkowitz et al. (2011), is based on the idea that if
the sequence of violation is a martingale difference, there ought to be no correlation
between any function of the past variables and the current violation. One way to test
this uncorrelatedness is through a linear model. The model is

I t ¼ a þ b1 I t1 þ b2 VaRt þ ut ,

where VaRt can be replaced by CoVaRt in the case of conditional VaR. The residual
ut follows a Logistic distribution since It is binary. We get the estimates of the
 Τ
coefficients b^1 ; b^2 . Therefore, the null hypothesis is b^1 ¼ b^2 ¼ 0 . This
hypothesis can be tested by Wald’s test.
We set m ¼ 1 or 5 for the Ljung-Box and Lobato tests. For the CaViaR test, two
data periods are considered separately. The first is the overall data from August
4, 2006, to August 4, 2011. The second is the data from August 4, 2008, to August
4, 2009, the period when the financial market reached its bottom. By separately
testing the two periods, we can gain more insights into the PLM model.

54.2.3 Risk Contribution Measure

The risk contribution of one firm to the market is one of the top concerns among
central bankers. The regulator can restrict the risky behaviors of the financial
institution with high-risk contribution to the market, and reduce the
institution’s incentive to take more risk. AB propose the idea of DCoVaR, which
is defined by
54 Quantile Regression in Risk Calibration 1477

DCoVaRtjji, t ¼ CoVaRtjji, t  CoVaR0:5

jji, t : (54.12)

t
where CoVaRjji,t is defined as in the introduction. j, i represent the financial system
and an individual asset. t ¼ 0.5 corresponds to the normal state of the individual
asset i. This is essentially a sensitivity measure quantifying the effect to the
financial system from the occurrence of a tail event of asset Xi.
In this study, we adopt a mathematically intuitive way to measure the marginal
effect by searching the first order derivative of the quantile function. Because the
spillover effect from stock market to stock market has already got much attention, it
is important to investigate the risk contribution of a local market to the global stock
market. The estimation is conducted as follows:
First, one estimates the following model nonparametrically:

Xj, t ¼ f 0:05
j ðX t Þ þ ej , (54.13)

The quantile function fj 0.05() is estimated with local linear quantile regression
with t ¼ 0.05, described with more details in Appendix 1. Xj is the weekly return of
the stock index of an individual country and X is the weekly return of the global
stock market.
Second, with f^j ðÞ, we compute the “marginal contribution of risk” (MCR) of
0:05

institution j by


@ ^0:05 ðxÞ


f
MCRtj ¼
j
1 , (54.14)
@x

x ¼ F^x ðtk Þ
1
where F^ ðtk Þ is a consistent estimator of the tk quantile of the global market return,
and it can be estimated by regressing Xt on the time trend. We put k ¼ 1, 2 with
t1 ¼ 0.5 and t2 ¼ 0.05. The quantity Eq. 54.14 is similar to the MES proposed by
Acharya et al. (2010) in the sense that the conditioned event belongs to the
information set of the market return, but we reformulate it in the VaR framework
instead of the expected shortfall framework.
There are some properties of the MCR to be described further. First, tk deter-
mines the condition of the global stock market. This allows us to explore the risk
contribution from the index j to the global market, given different global market
status. Second, the higher the value of MCR, the more risk factor j imposes on the
market in terms of risk. Third, since the function fj 0.05() is estimated by LLQR, the
quantile curve is locally linear, and therefore, the local first order derivative is
straightforward to compute.
We choose indices j ¼ S&P500, NIKKEI225, FTSE100, DAX30, CAC40, Hang
Seng as the approximate of the market returns of each developed country or market.
The global market is approximated by the MSCI World (developed countries)
market index. The data is weekly from April 11, 2004, to April 11, 2011, and
t ¼ 0.05.
1478 S.-K. Chao et al.

54.3 Results

54.3.1 CoVaR Estimation

The estimation results of VaR/CoVaR are shown in this section. We compute three
types of VaR/CoVaR of GS, with a moving window size of 126 business days and
t ¼ 0.05.
First, the VaR of GS is estimated using linear quantile regression:

d GS, t ¼ ^a GS þ^g ΤGS Mt1 ,

VaR (54.15)

Mt 2 ℝ is introduced in Sect. 54.2.1.

7

Second, the CoVaR of GS given C returns is estimated:

d C, t ¼ ^a C þ^g ΤC Mt1 ;
VaR (54.16)

d AB
CoVaR a GSjC þ b^GSjC VaR
GSjC, t ¼ ^
d C, t þ ^gΤ Mt1 :
GSjC (54.17)

If the SP replaces C, the estimates are generated from

e
d SP, t ¼ ^a SP þ ^gΤ M
VaR SP t1 ; (54.18)

d AB
CoVaR a GSjSP þ b^GSjSP VaR
GSjSP, t ¼ ^
d SP, t þ ^gΤ M e
GSjSP t1 , (54.19)

e t 2 ℝ6 is the vector of market variables without the market portfolio return.

where M
Third, the PLM CoVaR is generated:

d C, t ¼ ^a C þ ^gΤ Mt1 ;
VaR (54.20)
C

 
^Τ
d PLM ¼ e
CoVaR GSjC, t a^GSjC þ beGSjC Mt1 þ ^l GSjC VaR
d C, t : (54.21)

If SP replaces C:

e
d SP, t ¼ ^a SP þ ^gΤ M
VaR SP t1 ; (54.22)
 
^Τ
d PLM ¼ e
CoVaR GSjSP, t a^GSjSP þ beGSjSP M
e t1 þ ^ d SP, t :
l GSjSP VaR (54.23)

The coefficients in Eqs. 54.15–54.20, and 54.22 are estimated from the linear
quantile regression and those in Eqs. 54.21 and 54.23 are estimated from the
method described in Appendix 3.
d GS, t sequence. The VaR forecasts (red) seem to form
Figure 54.4 shows the VaR
a lower cover of the GS returns (blue). This suggests that the market variables Mt
54 Quantile Regression in Risk Calibration 1479

Fig. 54.4 The VaRd GS, t . The 0.2

d GS, t and
red line is the VaR
blue stars are daily returns of
GS. The dark green curve is
the median smoother of the
d GS, t curve with h ¼ 2.75.
VaR 0.0
t ¼ 0.05. The window size is
252 days

−0.2

2007 2008 2009 2010 2011

0.2

0.0

−0.2

2007 2008 2009 2010 2011

Fig. 54.5 The CoVaR of GS given the VaR of C. The gray dots mark the daily returns of GS. The
d PLM . The dark blue curve is the median LLQR smoother of
light green dashed curve is the CoVaR GSjC, t
d AB . The purple
the light green dashed curve with h ¼ 3.19. The cyan dashed curve is the CoVaR GSjC, t
curve is the median LLQR smoother of the cyan dashed curve with h ¼ 3.90. The red curve is the
d GS, t . t ¼ 0.05. The moving window size is 126 days
VaR

have some predictive power for the left tail quantile of the GS return distribution.
Figure 54.5 shows the sequences CoVaR d AB d PLM
GSjSP, t (cyan) and CoVaRGSjC, t (light
green). As the time series of the estimates is too volatile, we smooth it further by the
median LLQR. The two estimates are similar as the market state is stable, but
during the period of financial instability (from mid-2008 to mid-2009), the two
estimates have different behavior. The performances of these estimates are
evaluated by backtesting procedure in Sect. 54.3.2.
Table 54.1 shows the summary statistics of the VaR/CoVaR estimates. The first
three rows show the summary statistics of VaR d GS, t , VaR
d C, t , and VaR
d SP, t . The
d GS, t has lower mean and higher standard deviation than the other two. Partic-
VaR
ularly during 2008–2009, the standard deviation of the GS VaR is twice as much as
the other two. The mean and standard deviation of the VaR d C, t and VaR d SP, t
are rather similar. The last four rows show the summary statistics of CoVaRd PLM ,
GSjC, t
1480 S.-K. Chao et al.

Table 54.1 VaR/CoVaR summary statistics. The overall period is from August 4, 2006, to
August 4, 2011. The crisis period is from August 4, 2008, to August 4, 2009. The numbers in
the table are scaled up by 102
mean-overall sd-overall mean-crisis sd-crisis
d GS, t
VaR 3.66 3.08 7.43 4.76
d C, t
VaR 2.63 1.67 4.62 2.25
d SP, t
VaR 2.09 1.57 3.88 2.24
d PLM
CoVaR 4.26 3.84 8.79 5.97
GSjC, t
d AB
CoVaR 4.60 4.30 10.36 6.32
GSjC, t
d PLM
CoVaR 3.86 3.30 8.20 4.69
GSjSP, t
d AB
CoVaR 5.81 4.56 12.65 5.56
GSjSP, t

d AB , CoVaR
CoVaR d PLM , and CoVaR d AB
GSjC, t GSjSP, t GSjSP, t . This shows that the CoVaR
obtaining from the AB model has smaller mean and greater standard deviation
than the CoVaR obtaining from PLM model.
Figure 54.6 shows the bandwidth sequence of the nonparametric part of the
PLM estimation. The bandwidth varies with time. Before mid-2007, the bandwidth
sequence is stably jumping around 0.2. After that the sequence becomes very
volatile. This may have something to do with the rising systemic risk.

54.3.2 Backtesting

For the evaluation of the CoVaR models, we resort to the backtesting procedure
described in Sect. 54.2.2. In order to perform the backtesting procedure,
the sequences {It} (defined in Sect. 54.2.2) have to be computed for all
VaR/CoVaR estimates. Figure 54.7 shows the timings of the violations
{t : It ¼ 1} of CoVaRd PLM , CoVaR d AB d
GSjC, t GSjC, t and VaR GS, t . This figure shows the total
number of violations of PLM CoVaR and CoVaR is similar, while VaR d GS, t has more
d
violations than the both. The VaR GS, t has a few clusters of violations in both
financial stable and unstable periods. This may result from the failure VaR d GS, t to
d PLM
adapt for the negative shocks. The violations of CoVaRGSjC, t are more evenly
distributed. The violations of CoVaR d AB
GSjC, t have large clusters during financially
stable period, while the violation during financial crisis period is meager. This
d AB tend to overreact, as it is slack during the stable
contrast suggests that CoVaR GSjC, t
period but is too tight during the unstable period.
Figure 54.8 shows the timings of the violations {t : It ¼ 1} of CoVaR d PLM ,
GSjSP, t
d
CoVaR AB d d PLM
GSjSP, t , and VaR GS, t . The overall number of violations of CoVaRGSjSP, t is
more than that of VaR d GS, t , and it has many clusters. CoVaR d PLM
GSjSP, t behaves
d PLM
differently from CoVaRGSjC, t. The SP may not be more informative than C, though
54 Quantile Regression in Risk Calibration 1481

0.4

0.1

2007 2008 2009 2010 2011

^
Fig. 54.6 LLQR bandwidth in the daily estimation of CoVaRPLM
GSjC, t . The average bandwidth is
0.24

Fig. 54.7 The timings of

violations {t : It ¼ 1}. The top
circles are the violations of
d PLM , totally
the CoVaR GSjC, t
95 violations. The middle
squares are the violations of
d AB , totally
CoVaR GSjC, t
98 violations. The bottom
stars are the violations of
d GS, t , totally
VaR
109 violations. Overall data
N ¼ 1260 2007 2008 2009 2010 2011

the efficient market hypothesis suggests so. The violation of CoVaR d AB

GSjSP, t is fewer
than the other two measures, and the clustering is not significant.
The backtesting procedure is performed separately for each sequence of {It}. The null
hypothesis is that each sequence of {It} forms a series of martingale difference. Six
different tests are applied for each {It}: Ljung-Box tests with lags 1 and 5, Lobato test with
lags 1 and 5, and finally the CaViaR test with two data periods: overall and crisis period.
The result is shown in Table 54.2. First, in Panel 1 of Table 54.2, the VaR d GS, t is
rejected by the LB(5) test and the two CaViaR tests. This shows that a linear
quantile regression on the seven market variables may not give accurate estimates,
in the sense that the violation {It} of VaR d GS, t does not form a martingale sequence.
Next we turn to the CoVaRd AB and d PLM . In Panel 2, the low p-values of
CoVaR
GSjSP, t GSjSP, t
the two CaViaR tests show that both the AB model and PLM model conditioned on
SP are rejected, though the p-value of the AB model almost reaches the 5 %
significant level. In particular, the CoVaR d PLM
GSjSP, t is rejected by the L(5) and LB
(5) tests. Both the parametric and semiparametric models fail with this choice of
variable. This suggests that the market return does not provide enough information
in risk measurement.
We therefore need more informative variables. Panel 3 of Table 54.2 illustrates
this by using C daily returns, which may contain information not revealed in the
1482 S.-K. Chao et al.

Fig. 54.8 The timings of

violations {t : It ¼ 1}. The top
circles are the violations of
d PLM , totally
CoVaR GSjSP, t
123 violations. The middle
squares are the violations of
d AB
CoVaR GSjSP, t , totally
39 violations. The bottom
stars are the violations of
d GS, t , totally
VaR
109 violations. Overall data
N ¼ 1,260

2007 2008 2009 2010 2011

Table 54.2 Goldman Sachs VaR/CoVaR backtesting p-values. The overall period is from August
4, 2006, to August 4, 2011. The crisis period is from August 4, 2008, to August 4, 2009. LB(1) and
LB(5) are the Ljung-Box tests of lags 1 and 5. L(1) and L(5) are the Lobato tests of lags 1 and
5. CaViaR-overall and CaViaR-crisis are two CaViaR tests described in Sect. 2.2 applied on the
two data periods
Measure LB(1) LB(5) L(1) L(5) CaViaR-overall CaViaR-crisis
Panel 1
d GS, t
VaR 0.3449 0.0253* 0.3931 0.1310 1.265  106*** 0.0024**
Panel 2
d AB
CoVaR 0.0869 0.2059 0.2684 0.6586 8.716  107*** 0.0424*
GSjSP, t
d PLM
CoVaR 0.0518 0.0006*** 0.0999 0.0117* 2.2  1016*** 0.0019**
GSjSP, t
Panel 3
d AB
CoVaR 0.0489* 0.2143 0.1201 0.4335 3.378  109*** 0.0001***
GSjC, t
d PLM
CoVaR 0.8109 0.0251* 0.8162 0.2306 2.946  109*** 0.0535
GSjC, t
* ** ***
, and denote significance at the 5 %, 1 % and 0.1 % levels

d AB
market and improve the performance of the estimates. The CoVaR GSjC, t is rejected
by the two CaViaR tests and the LB(1) test with 0.1 % and 5 % significant level.
d PLM is not rejected by the CaViaR-crisis test. This implies that
However, CoVaR GSjC, t
the nonparametric part in the PLM model captures the nonlinear effect of C returns
to GS returns, which can lead to better risk-measuring performance.

54.3.3 Global Risk Contribution

In this section, we present the MCR (defined in Sect. 54.2.3), which measures the
marginal risk contribution of risk factors. We choose t1 ¼ 0.5, associated to the
54 Quantile Regression in Risk Calibration 1483

Fig. 54.9 The M C Rtj 1 ,

t ¼ 0.5. j:CAC, FTSE, DAX,
0.6
Hang Seng, S&P500 and
NIKKEI225. The global
market return is approximated 0.4
by MSCI World

0.2

0.0

−0.2
2005 2006 2007 2008 2009 2010 2011

1.5

1.0

0.5

Fig. 54.10 The M CRtj 2 ,

t ¼ 0.05. j:CAC, FTSE, 0.0
DAX, Hang Seng, S&P500
and NIKKEI225. The global
market return is approximated
by MSCI World 2005 2006 2007 2008 2009 2010 2011

normal (median) state and t2 ¼ 0.05, associated to an negative extreme state.

Figure 54.9 shows the M C Rtj 1 from local markets j to the global market. When
the MSCI World is at its normal state, the Hang Seng index in normal times
contributes the most risk to the MSCI World at all times. The NIKKEI225 places
second; the contribution from S&P500 varies most with the time; the risk contri-
bution from DAX30 is nearly zero. The contributions from CAC40 and FTSE100
are negative.
Assuming that the MSCI World is at its bad state (t2 ¼ 0.05), the M C Rtj 2 differs
from M C Rtj 1 , see Fig. 54.10. One sees that the S&P500 imposes more pressure on
the world economy than the other countries, especially during the financial crisis of
2008 and 2009. The contribution from Hang Seng is no longer of the same
significance. The three European markets are relatively stable.
This analysis suggests that the risk contribution from individual stock market
varies a lot with the state of global economy.
1484 S.-K. Chao et al.

54.4 Conclusion

In this study, we construct a PLM model for the CoVaR, and we compare it to the
AB model by backtesting. Results show that PLM CoVaR is preferable, especially
during a crisis period. The study of the MCR reveals the fact that the risk from each
country can vary with the state of global economy.
As an illustration, we only study the Goldman Sachs conditional VaR with
Citigroup and S&P500 as conditioned risk sources. In practice, we need to choose
variables. In Hautsch et al. (2011), the Least Absolute Shrinkage and Selection
Operator (LASSO) techniques are used to determine the most relevant systemic risk
sources from a pool of financial institutions. A VAR (Vector Autoregression) model
may be also suitable for capturing the asset dynamics, but the estimation may be
more involved. We may include other firm-specific variables such as corporate
bond yields as these variables can bear other information which is not included in
the stock returns or stock indices.

Appendix 1: Local Linear Quantile Regression (LLQR)

Let {(Xi, Yi)}in ¼ 1  ℝ2 be independently and identically distributed (i.i.d.) bivariate

random variables. Denote by FY|x(u) the conditional cumulative distribution func-
tion (cdf) and l(x) ¼ F1 Y|x (t) the conditional quantile curve to level t, given
observations {(xi, yi)}ni ¼ 1, one may write this as

yi ¼ lðxi Þ þ ei ,
1
with Fe|x (t) ¼ 0. A locally linear kernel quantile estimator (LLQR) is estimated as
^
l ðx 0 Þ ¼ ^
a 0 from:
  Xn x  x 
^a 0 ; ^b 0 ¼ argmin
i 0
K (54.24)
fa0 , b0 g i¼1 h

rt fyi  a0  b0 ðxi  x0 Þg, (54.25)

where h is the bandwidth, K(·) is a kernel, and rt(·) is the check function given by
 
rt ðuÞ ¼ t  1fu<0g u (54.26)
Figure 54.11 illustrates the check functions. Different loss functions give differ-
ent estimates. u2 corresponds to the conditional mean. rt(u) corresponds to the
conditional tth quantile.
It is shown by Fan et al. (1994) that the locally linear kernel estimator is asymptot-
ically efficient in a minimax sense. It also possesses good finite sampling
property which is adaptive to a variety of empirical density g(x) and has good boundary
property.
54 Quantile Regression in Risk Calibration 1485

Fig. 54.11 This figure

presents the check function. 1.5
The dotted line is u2. The
dashed and solid lines are
check functions rt(u) with
t ¼ 0.5 and 0.9 respectively

1.0

0.5

0.0

−2 −1 0 1 2

Next, we describe the method to compute the bandwidths. The approach used
here follows Yu and Jones (1998). The bandwidth is chosen by
h  2 i1=5
ht ¼ hmean tð1  tÞ’ F1 ðtÞ , (54.27)

where hmean is the locally linear mean regression bandwidth, which can be com-
puted by the algorithm described in Ruppert and Wand (1995) or Ruppert
et al. (1995). ’(·) and F(·) are the pdf and cdf of the standard normal distribution.
Since we discuss the case for VaR, t is usually small. ht needs to be enlarged to
allow for more smoothing (usually taking 1.5ht or 2ht).
The approach is acceptable but not so flexible, because it is based on assuming the
quantile functions are parallel. A more flexible approach was developed by Spokoiny
et al. (2011). In order to stabilize the bandwidth choice, we first regress yi on the rank
of the corresponding xi and then rescale the resulted estimated values to the original
x space. Carroll and Hardle (1989) show that this local bandwidth estimator and the
global bandwidth estimator are asymptotically equivalent.

Appendix 2: Confidence Band for Nonparametric Quantile

Estimator

The uniform confidence band of the quantile estimator is based on the Theorem 2.2
and Corollary 2.1 presented in Hardle and Song (2010). The details are as follows.
1486 S.-K. Chao et al.

Let {(Xi,Yi)}in ¼ 1 be as in Appendix 1. Define Kh(u) ¼ h1 K(u/h) and similar to

Eq. 54.25 let ln(x) and l(x) are zeros (w.r.t. y) of the functions:

X
n
e n ðy; xÞ def
H ¼ n1 K h ðx  Xi Þrt ðY i  yÞ;
i¼1

ð
e ðy; xÞ def
H ¼ f ðx; yÞrt ðy  yÞdy,
ℝ

where rt(·) is the check function defined as Eq. 54.26.

Ð A
Theorem 1 Let h ¼ nd, 15 < d < 13 , l(K) ¼ 2
 A K (u)du, where K(·) is
supported
Ð A on [A, A]. J ¼ [0, 1]. Define c1(K) ¼ {K (A) + K2(A)}/2l(K), c2(K) ¼
2
0 2
 A{K (u)} du/2l(K) and

8 1=2
logfc ðKÞg
> 1=2
< ð2d lognÞ þ ð2d logn
> þ 12flogd þ loglogng,
1
p1=2

dn ¼ if c1 ðK > 0;
>
: ð2d lognÞ1=2 þ ð2d logn1=2 logfc2 ðKÞgg, otherwise:
>
2p

Then
" ( ) #
jln ðxÞ  lðxÞj
P ð2d lognÞ1=2 sup r ðxÞ  dn <z
x2J lðK Þ1=2
! expf2expðzÞg,

as n ! 1, with

r ðxÞ ¼ ðnhÞ1=2 f flðxÞjxgff X ðxÞ=tð1  tÞg1=2 ,

where fX(·) is the marginal pdf for X and f(·|x) is the conditional pdf of Y on X ¼ x.
The corollary followed by the theorem explicitly indicates how a uniform
confidence interval can be constructed.
Corollary 1 An approximate (1  a)  100 % confidence band is

n o1=2 1  
n 1=2 o
ln ðnhÞ1=2 tð1  tÞlðK Þ=f^X ðtÞ f^ lðt
t g  dn þ cðaÞð2d logn ,

where c(a) ¼ log 2  log|log(1  a)| and f^X ðtÞ, f^flðtÞjtg are consistent estimates
for fX(t), f{l(t)|t}.
54 Quantile Regression in Risk Calibration 1487

0.5
0.5

0.0 0.0

0.0 0.5 0.0 0.5

Fig. 54.12 GS and C weekly returns 0.90(left) and 0.95(right) quantile functions. The y-axis is
GS daily returns and the x-axis is the C daily returns. The blue curves are the LLQR curves (see
Appendix 1). The LLQR bandwidths are 0.0942 and 0.1026. The red lines are the linear parametric
quantile regression line. The antique white curves are the asymptotic confidence band (see
Appendix 2) with significance level 0.05. N ¼ 546

Figure 54.1 is done by the techniques introduced in Appendices 1 and 2. Another

illustration with right tail quantiles is in Fig. 54.12. We plot the LLQR curve for 0.9
and 0.95 quantile. Both the two linear quantile regression lines lie outside the LLQR
confidence band as the Citigroup returns are positive.

Appendix 3: PLM Model Estimation

For the PLM estimation, we adopt the algorithm described in Song et al. (2012).
Given data {(Xt, Yn)}Tn ¼ 1 bivariate and {Mt}Tn ¼ 1 multivariate random variables.
The PLM is:

Y t ¼ a þ bΤ Mt1 þ lðXt Þ þ et :

Let an denote an increasing sequence of positive integers and set bn ¼ a1 n .

For given n, dividing the interval [0, 1] into an subintervals Int, t ¼ 1,. . ., an with
equal length bn. On each Int, l(·) can approximately be taken as a constant.
The PLM estimation procedure is:
1. Inside each partition Int, a linear quantile regression is performed to get b^i , then
their weighted mean gives b. ^ Formally, let rt(·) be the check function defined as
Eq. 54.26, l1 , . . . , lan are constants,
1488 S.-K. Chao et al.

( )
X
n X
an
 
b^ ¼ argmin min rt Xj, t  a  b Mt1  lm 1 Xi, t 2 I nt
T
b l1 , ..., lan
t¼1 m¼1

2. Computing the LLQR nonparametric quantile estimates of l(·) as outlined in

n Τ
on
Appendix 1 from Xi, t , Xj, t  ^a  b^ Mt1 .
t¼1

References
Acharya, V. V., Pedersen, L. H., Philippon, T., & Richardson, M. (2010). Measuring systemic risk
(Working paper 10-02). Federal Reserve Bank of Cleveland.
Adams, Z., F€uss, R., & Gropp, R. (2012). Spillover effects among financial institutions: A state-
dependent sensitivity value-at-risk approach, SAFE Working Paper Series, No. 20, http://dx.
doi.org/10.2139/ssrn.2267853.
Adrian, T., & Brunnermeier, M. K. (2011). CoVaR (Staff reports 348). Federal Reserve Bank of
New York.
Bae, K.-H., Karolyi, G. A., & Stulz, R. M. (2003). A new approach to measuring financial
contagion. Review of Financial Studies, 16(3), 717–763.
Berkowitz, J., Christoffersen, P., & Pelletier, D. (2011). Evaluating value-at-risk models with
desk-level data management science, 57, 2213–2227.
Brownlees, C. T., & Engle, R. F. (2012). Volatility, correlation and tails for systemic risk
measurement. Available at SSRN: http://ssrn.com/abstract=1611229 or http://dx.doi.org/
10.2139/ssrn.1611229.
Brunnermeier, M., & Pedersen, L. H. (2008). Market liquidity and funding liquidity. Review of
Financial Studies, 22(2201), 2238.
Cai, Z., & Wang, X. (2008). Nonparametric estimation of conditional VaR and expected shortfall.
Journal of Econometrics, 147(120), 130.
Carroll, R., & Hardle, W. (1989). Symmetrized nearest neighbor regression estimates. Statistics &
Probability Letters, 7(315), 318.
Chernozhukov, V., & Umantsev, L. (2001). Conditional value-at-risk: Aspects of modeling and
estimation. Empirical Economics, 26(271), 292.
Engle, R., & Manganelli, S. (2004). CAViaR: Conditional autoregressive value at risk by regres-
sion quantiles. Journal of Business & Economic Statistics, 22(367), 381.
Fan, J., Hu, T.-C., & Truong, Y. K. (1994). Robust nonparametric function estimation. Scandina-
vian Journal of Statistics, 21, 433–446.
Härdle, W., & Song, S. (2010). Confidence bands in quantile regression. Econometric Theory, 26,
1180–1200.
Härdle, W., Liang, H., & Gao, J. (2000). Partially linear models. Heidelberg: Physica-Verlag.
Härdle, W., Muller, M., Sperlicli, S., & Werwatz, A. (2004). Nonparametric and semiparametric
models. Berlin: Springer.
Hautsch, N., Schaumburg, J., & Schienle, M. (2011). Financial network systemic risk contribu-
tions (Discussion paper 2011-072). CRC 649, Humboldt-Universitat zu Berlin.
Huang, X., Zhou, H., & Zhu, H. (2012). Systemic risk contributions Journal of Financial Services
Research, 42, 55–83. Springer US.
Koenker, R., & Bassett, G. S. (1978). Regression quantiles. Econometrica, 46, 33.
Kuan, C.-M., Yeh, J.-H., & Hsu, Y.-C. (2009). Assessing value at risk with CARE, the conditional
autoregressive expectile models. Journal of Econometrics, 150, 261–270.
Lobato, I., Nankervis, J. C., & Savin, N. (2001). Testing for autocorrelation using a modified
Box-Pierce Q test. International Economic Review, 42(1), 187–205.
54 Quantile Regression in Risk Calibration 1489

Ruppert, D., & Wand, M. P. (1995). Multivariate locally weighted least, squares regression.
Annals of Statistics, 23(1346), 1370.
Ruppert, D., Sheatlier, S. J., & Wand, M. P. (1995). An effective bandwidth selector for
local least squares regression. Journal of the American Statistical Association, 90(1257),
1270.
Schaumburg, J. (2011). Predicting extreme VaR: Nonparametric quantile regression with refine-
ments from extreme value theory (Discussion Paper 2010-009). CRC 649, Humboldt-
Universitat zu Berlin.
Song, S., Ritov, Y., & Härdle, W. (2012). Partial linear quantile regression and bootstrap
confidence bands. Journal of Multivariate Analysis, 107, 244–262.
Spokoiny, V., Wang, W., & Härdle, W. (2013). Local quantile regression. Journal of
Statistical Planning and Inference, 143(7), 1109–1129. doi:10.1016/j.jspi.2013.03.008,
with discussions.
Taylor, J. W. (2008). Using exponentially weighted quantile regression to estimate value at risk
and expected shortfall. Journal of Financial Econometrics, 6, 382–406.
Yu, K., & Jones, M. C. (1998). Local linear quantile regression. Journal of the American Statistical
Association, 98(228), 237.
Strike Prices of Options for Overconfident
Executives 55
Oded Palmon and Itzhak Venezia

Contents
55.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1492
55.2 Overconfidence and the Optimal Exercise Prices of Executive
Incentive Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1494
55.3 The Simulation Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1497
55.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1498
55.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1503
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1504
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1506

Abstract
We explore via simulations the impacts of managerial overconfidence on the
optimal strike prices of executive incentive options. Although it has been shown
that, optimally, managerial incentive options should be awarded in-the-money,
in practice most firms award them at-the-money. We show that the optimal strike
prices of options granted to overconfident executive are directly related to their
overconfidence level and that this bias brings the optimal strike prices closer to
the institutionally prevalent at-the-money prices. Our results thus support the
viability of the common practice of awarding managers with at-the-money
incentive options. We also show that overoptimistic CEOs receive lower com-
pensation than their realistic counterparts and that the stockholders benefit from

O. Palmon (*)
Department of Finance and Economics, Rutgers Business School – Newark and New Brunswick,
Piscataway, NJ, USA
e-mail: palmon@business.rutgers.edu; palmon@rbs.rutgers.edu
I. Venezia
School of Business, The Hebrew University, Jerusalem, Israel
Bocconi University, Milan, Italy
e-mail: msvenez@mscc.huji.ac.il

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1491
DOI 10.1007/978-1-4614-7750-1_55,
# Springer Science+Business Media New York 2015
1492 O. Palmon and I. Venezia

their managers bias. The combined welfare of the firm’s stakeholders is, how-
ever, positively related to managerial overconfidence.
The Monte Carlo simulation procedure described in Sect. 55.3 uses
a Mathematica program to find the optimal effort by managers and the optimal
(for stockholders) contract parameters. An expanded discussion of the simula-
tions, including the choice of the functional forms and the calibration of the
parameters, is provided in Appendix 1.

Keywords
Overconfidence • Managerial effort • Incentive options • Strike price • Simula-
tions • Behavioral finance • Executive compensation schemes • Mathematica
optimization • Risk aversion • Effort aversion

55.1 Introduction

The optimal structure of executive compensation has intrigued academic researchers

as well as practitioners for a long time. Most principal-agent models dealing with this
issue yield rather complex payment schedules, making it quite challenging to test
their predictions. In practice there is a widespread use of simple compensation
schemes such as linear or piecewise linear (stock option) contracts. An important
question that arises in this case is what are the optimal parameters for these simple
schemes? In particular what are the optimal strike prices for incentive option
schemes? Unfortunately, this important issue has received only little attention.
Institutional and tax factors could be to blame for this neglect. Before the 2006
changes in the US tax rules, the “intrinsic value” of executive options was taxed, and
this discouraged firms from granting their executives in-the-money options. Granting
out-of-the-money options seemed unfair and there is no empirical or theoretical
evidence for advantages to such practice (Mahajan 2002). Indeed, only a very
small fraction of firms used such strike prices.1 The virtual monopoly of at-the-money
strike prices, their institutional appeal, or some other unknown factors might have
discouraged academics from studying the merits and demerits of this practice.
In their landmark paper, Hall and Murphy (2000, 2002) attribute the pervasive-
ness of granting at-the-money options to their property of being the most sensitive
to changes in the stock price. Palmon et al. (2008), however, have shown that
issuing the most sensitive options is not necessarily optimal when managers are risk
and effort averse.2 Within a model explicitly considering the choice of the contract
parameters by stockholders and the resulting effort chosen by risk-averse and
effort-averse managers, they show for a wide range of parameters that well describe

1
According to Mahajan (2002), less than 1 % of firms used out-of-the-money strike prices.
Furthermore, in his study firms did not benefit from awarding such options to their managers.
2
Hall and Murphy (2000) did not show that at-the-money strike prices are optimal, just that they
possess the highest sensitivity to stock prices. They did not assume effort aversion by managers
either.
55 Strike Prices of Options for Overconfident Executives 1493

managers’ risk and effort aversion, that in-the-money options provide are optimal.
Such options provide managers a better risk-return trade-off and ultimately consti-
tute a better form of compensation than either out-of-the-money or at-the-money
options.3 Palmon et al. further argue that the asymmetric tax treatment of options
under the old (prior to 2006) tax system, which penalized the issuance of
in-the-money options, may have driven firms to use at-the-money options.
Whereas most studies of the issue of optimal incentive contracts assumed
that managers as well as stockholders are rational, there exists extensive literature
that documents that managers often are overconfident. Of the few studies that
explore the effect of cognitive biases on managerial compensation, none however
explores the effect of overconfidence on the optimal strike prices for the incentive
options. Gervais et al. (2011) investigate the optimal form of managerial
compensation under overconfidence but define overconfidence in the sense of
too-high-precision-of-estimates (calibration), and the managers in their model
exert effort to obtain better information on the investment parameters. There is an
abundant literature however that indicates the pervasiveness of overconfidence in
the optimism or “better than average” sense rather than in the calibration interpre-
tation (see, e.g., Malmendier and Tate 2005a, b, 2008; Roll 1986; Suntheim 2012).4
Oyer and Schaefer (2005) and Bergman and Jenter (2007) also consider the effect of
optimism, and other sentiments, on managerial compensation, but they do not
consider the effect of these sentiments on the optimal strike prices or on the
managers’ effort.
In this paper we investigate the hitherto unexplored question of the effect of
overconfidence on the optimal strike prices for risk-averse and effort-averse man-
agers. We show that overconfidence leads to higher optimal strike prices of
managerial incentive schemes, and that awarding overconfident CEOs at-the-
money options mitigates the stockholders’ vs. managers’ agency problem, leading
to higher managers’ productivity. Our results thus provide support for the viability
of the ubiquitous yet seemingly unoptimal practice of awarding CEOs with
at-the-money incentive options.
Whereas the main focus of the paper is the interaction between overconfidence
and the strike prices of managerial incentive options, it also sheds light on the effect
of overconfidence on the firm’s stakeholders (stockholders and managers). We
predict, as empirically shown by Otto (2011), that overoptimistic CEOs receive
lower compensation than their realistic counterparts. However, the stockholders
benefit from their managers bias since they pay less and enjoy the productivity of
the higher effort the overconfident manager exerts. We construct a measure of the
combined welfare of managers and stockholders and demonstrate that it is

3
Dittman et al. (2010) found that for a range of parameterizations, a principal-agent model with
loss-averse agents generates convex compensation contract but did not investigate the parameters
of the options to be used in the compensation package. Recently, however, Dittman and Yu (2011)
found that in-the-money options are optimal.
4
Glaser and Weber (2007) note that only overconfidence in the better than average sense affects
trading.
1494 O. Palmon and I. Venezia

positively related to managerial overconfidence, a result helping explain the per-

sistence of this bias.5
The paper is constructed as follows. In Sect. 55.2 we present the model. In
Sect. 55.3 we explain the simulation method, and in Sect. 55.4 we present the
simulations’ results. Section 55.5 concludes.

55.2 Overconfidence and the Optimal Exercise Prices of

Executive Incentive Options

We consider a one-period Holmstrom (1979)-type model where a risk-neutral firm

employs an overconfident, risk-averse, and effort-averse manager.6 The cash
flows, X, of the firm depend on the manager’s effort and on exogenous stochastic
factors. The manager is assumed to provide some effort which is the minimum
necessary to run the firm and hence may be considered observable, but can provide
also unobservable extra effort. The more extra effort the manager exerts, the higher
will be the expected cash flows. Because stockholders cannot observe managers’
extra effort, managerial compensation may depend on the firm’s cash flows (which
depend on effort), but cannot be determined directly based on extra effort.
We assume that the cash flows of the firm, X, are lognormally distributed with
the following distribution function:
n o  pffiffiffiffiffiffi
f ðXÞ ¼ exp 0:5f½logðXÞ  mðYÞ=sg2 = Xs 2p (55.1)

where Y denotes the managerial extra effort (a managerial choice variable) and
m(Y) and s denote, respectively, the mean and the standard deviation of the
underlying normal distribution of the natural logarithm of X. We assume that
managerial effort increases cash flows and that overconfident managers
overestimate the impact of their effort on cash flows. Formally, we use the follow-
ing specification:

mðYÞ ¼ Lnðm0 þ 500lYÞ  s2 =2, (55.2)

where l denotes the degree of overconfidence. We assume that stockholders have

realistic expectations, which are represented by l ¼ 1, and that managers use
l > 1 to form their expectations. Thus, f(X) can be written as f(X, l), where
f(X, l ¼ 1) represents the realistic cash flow distribution, while f(X, l > 1)

5
Palmon and Venezia (2012) explore the effect of managerial overconfidence on the firm’s
stockholders and show that overconfidence may improve welfare. However, that study does not
investigate the optimal strike price of managerial incentive options.
6
In our model we assume symmetry of information between the manager and the firm regarding
the distribution of cash flows of the firm except for the different view of the effect of the manager’s
effort on cash flows.
55 Strike Prices of Options for Overconfident Executives 1495

represents the cash flow distribution as viewed by overconfident managers. For

notation brevity, we suppress the l in f(X, l). By the known properties of the
the mean and variance of X equal e½mðYÞþ0:5s  and
2

hlognormal2 distribution,
i
e½2mðYÞþs  es  1 , respectively. Thus, it follows from Eq. 55.2 that a person
2

with a l overconfidence measure believes that the mean of the cash flows X is
emðYÞþ0:5s ¼ m0 þ 500lY and that their coefficient of variation is approximately s.7
2

Since managers and stockholders differ in their perception of the distributions of cash
flows, one must be careful in their use. In what follows we refer to the distribution of
cash flows as seen by stockholders as the realistic distribution, and will make a
special note whenever the manager’s overconfident beliefs are used.
Except for her overconfidence, the manager is assumed to be rational and to
choose her extra effort so as to maximize the expected value of the following utility
function which exhibits constant relative risk aversion (CRRA) with respect to
compensation:

1 1 1g
UðI; YÞ ¼ NYb þ I (55.3)
1g 1g

In Eq. 55.3, I denotes the manager’s monetary income, g denotes the constant
relative risk aversion measure, N is a scaling constant representing the importance
of effort relative to monetary income in the manager’s preferences, and the positive
parameter b is related to the convexity of the disutility of effort.
Since stockholders cannot observe the manager’s extra effort, they propose
compensation schemes that depend on the observed cash flows, but not on
Y. Stockholders, which we assume to be risk neutral, strive to make the compen-
sation performance sensitive in order to better align the manager’s incentives with
their own. Stockholders offer the manager a compensation package that includes
two components: a fixed wage (W) that she will receive regardless of her extra
effort and of the resulting cash flows and options with a strike price (K) for
a fraction (s) of the equity of the firm. We assume that stockholders offer the
contract that maximizes the value of their equity.
The following timeline of decisions is assumed. At the beginning of the period,
the firm chooses the parameters of the compensation contract (K, W, and s) and
offers this contract to the manager. Observing the contract parameters, and taking
into account the effects of her endeavors on firm cash flows and hence on her
compensation, the manager determines the extra-effort level Y that maximizes her
expected utility. At the end of the period, X is revealed, and the firm distributes the
cash flows to the manager and to the stockholders and then dissolves. The priority of
payments is as follows. The firm first pays the wages or only part of them if the cash
flows do not suffice. If the cash flows exceed the wage, W, but not (K + W), then the

h 2 i
7
More precisely the square of the coefficient of variation is es  1 which can be approximated
by s2 since for any small z, ez1 is close to z.
1496 O. Palmon and I. Venezia

managers just receive their fixed wage. The managers are paid the value of the
options s(X  K  W), in addition to W if X exceeds K + W. The manager therefore
receives the cash flows I(X) defined by
8 9 8 9
<X = <X  W =
IðxÞ ¼ W when WXWþK (55.4)
: ; : ;
W þ sðX  W  KÞ WþKX

The shareholders get the residual cash flows.

In the above cash flow formula, the first range covers the case where cash flows
do not suffice to pay the entire wage. The second range covers the case where the
options expire out-of-the-money, and the manager gets the promised wage. The
third range represents cash flows that are large enough so that the options expire
in-the-money. In addition to the wage, the manager receives a proportion, s, of the
value of the firm above the threshold value of K. The expected utility of the
manager E{U[I(X),Y]} which governs her behavior, and her expected compensa-
tion E[I(X)], can be obtained by integrating her utility U[I(X),Y] given in Eq. 55.3
and her compensation I(X), given in Eq. 55.4, respectively. We note that the
manager chooses the effort level so as to maximize the expected utility using her
perception of the distribution of the firm’s final cash flows, while stockholders
choose the parameters of the compensation contract using the realistic cash flow
distribution to calculate the expected cash flows and managerial compensation.
Shareholders receive all cash flows that are not received by the manager. Since
stockholders are risk neutral and rational, stockholders’ equity value (SEV) is the
expected value of these payments, using the realistic distribution function, and
hence,8
ð1
SEV ¼ EðCashflowsÞ  E½IðXÞ ¼ Xf ðXÞdX  E½IðXÞ (55.5)
0

While the derivation of the optimal contract for any set of exogenous parameters
is conceptually straightforward, unfortunately, closed form solutions cannot be
obtained in our integrative model. Hence, following Hall and Murphy (2000), we
resort to simulations to evaluate the optimal contracts and analyze their properties.
In addition, we cannot use the Black-Scholes model to evaluate the executive stock
options since this model takes the values of the underlying asset as given, whereas
a crucial aspect of the managerial incentive scheme of our model is that managerial
extra effort and firm value are endogenously determined. We therefore introduce
a model that simultaneously simulates the manager’s optimal extra-effort level as
well as the expected values of the executive stock options and shareholders’ equity

8
Discounting the cash flows by an appropriate risk-adjusted discount rate would yield a linear
transformation of equity values. To simplify the presentation, and as is common in the literature,
we abstract from that.
55 Strike Prices of Options for Overconfident Executives 1497

for each compensation package. We check the robustness of our results by using
alternative parameters for the manager’s utility function and the distribution func-
tions of the cash flows.

55.3 The Simulation Procedures

We assume that managers have external employment opportunities and that stock-
holders offer managerial compensation packages that provide the managers with
a comparable expected utility.9 Without loss of generality (i.e., by an appropriate
definition of the wage units), we assume that these external employment opportu-
nities provide the manager an expected utility that equals the level of utility that is
obtained from a fixed compensation of 100 in the absence of any extra effort. Thus,
in all the simulations, we set the manager’s expected utility to correspond to the
level obtained from a fixed compensation of 100 (wage ¼ 100 and no option grants)
and no extra effort (which is the optimal extra-effort choice when no options are
granted).10 We then search over a grid of strike prices (using four-digit accuracy)
and find for each strike price the percentage of options that should be awarded so
that the manager’s expected utility equals the expected utility target when the
manager chooses the optimal extra-effort level. We identify the strike price that
is associated with the highest equity level and refer to this contract as the optimal
contract for the given set of parameters.
In calibrating the other parameters for the simulations, we try to approximately
conform to Hall and Murphy (2000) and Hall and Liebman (1998); to studies that
simulate decisions with effort aversion, such as Bitler et al. (2005); and to studies that
explore the effect of overconfidence on corporate decisions, such as Malmendier and
Tate (2005a, b, 2008).11 Accordingly, we set the parameters in our base case as
follows. The coefficient of variation, s, equals 0.3, and thus, the standard deviation is
0.3E(X). Since the expected cash flows serve as numeraire, the volatility is deter-
mined solely by the coefficient of variation. In our base case, we set the managerial
wage to equal 50.12 The expected cash flows as viewed by an overconfident manager
with an overconfidence measure of l are E(X) ¼ 45,000 + 500lY (i.e., m0 ¼ 45,000).
The risk aversion and effort aversion parameters are g ¼ 4 and b ¼ 3, respectively.

9
See Appendix 1 for more details.
10
When the manager is overconfident, this expected utility is calculated according to the man-
ager’s expectations.
11
See Appendix 1 for the explanation for the calibration of our model. To be on the safe side and in
stride with explanations for the risk premium puzzle, we use higher values for the risk aversion
parameter.
12
It should be noted that although the wage level in our base case equals half of the fixed
compensation that corresponds to the utility target, it equals only about 11 % of the expected
compensation under the optimal contract when managers are realistic. When managers are
overconfident, a wage of 50 consists of less than 11 % of total compensation according to the
manager’s expectations but more than 11 % according to the realistic expectations.
1498 O. Palmon and I. Venezia

We consider overconfidence levels between l ¼ 1 (no overconfidence) and l ¼ 2.5 in

0.5 increments.
We examine the robustness of the results to deviations from the base case
combination of parameters by simulating with several alternative sets of exogenous
parameters. We repeat the analysis for many alternative sets of the exogenous
parameters: the manager’s risk and extra-effort aversion, g and b, as well as the
volatility measure of cash flows, s.13

55.4 Results and Discussion

In Table 55.1 we present the impact of overconfidence on the strike price that
stockholders choose to offer: the moneyness, the percentage of the firm given as
options, the effort choice of managers, the stockholders’ equity value, and the
expected managerial compensation. The expected compensation is calculated
both under the realistic distribution and under the subjective distribution of the
manager.
One observes from Table 55.1 that the strike price, the options’ moneyness, the
optimal managerial effort, the value of the stockholders’ equity, and the expected
compensation according to the managers’ expectations are directly related to
overconfidence. The optimal strike price (in thousands of dollars; strike prices
will be denoted in thousands of dollars in the rest of the study) for a rational
manager is 40.71, with a 0.60 moneyness (which can be described as deep-in-the-
money), but it rises to 63.74 with a 0.89 moneyness (closer to at-the-money) when
l ¼ 2.5.14 Managers also work harder the more overconfident they are
(Y increases from around 47 when they are realistic to around 52 when
l ¼ 2.5). Consequently, in order to hold the managers’ expected utility fixed,
their subjective expected monetary compensation must increase with
overconfidence to compensate for the extra risk resulting from the higher strike
price and for the additional effort they exert. The expected compensation the
stockholders perceive they pay according to the realistic expectation, however, is
inversely related to the overconfidence measure as they take advantage of
managers’ unrealistic expectations. The SEVs of the optimal contracts increase
as managerial overconfidence increases (see column 5, the SEV rises from 68,099
when l ¼ 1 to 70,958 when l ¼ 2.5, an increase of about 6 %). That is, the
stockholders benefit from the managers overestimating their powers.
This analysis suggests that stockholders are able to induce overconfident man-
agers to exert higher effort levels even though the objective contract parameters
they offer them are less favorable (the managers work harder but receive lower

13
Because of scaling there is no need to conduct robustness checks for the expected cash flows.
14
The moneyness measure depends on the strike price and the value of equity, which in turn
depends on effort. Thus, the moneyness measure varies with overconfidence because effort varies
with overconfidence, even when the strike price remains constant.
 55

Table 55.1 Stockholders’ equity values (SEV), strike prices (K), moneyness (K/S), effort levels (Y), and other parameters of interest. Base case:
g ¼ 4 b ¼ 3, s ¼ 0.3, wage ¼ 50, N ¼ 4,000
Expected Sum of SEV and
Percentage of Stockholders compensation, Expected realistic
Strike price Moneyness firm given as Equity Value realistic compensation, expected
Overconfidence (K) in 1000s (K/S) options, (s) Effort (Y) (SEV) valuation managers valuation compensation
measure, l 1 2 3 4 5 6 7 8
1 40.71 0.5944 1.40% 47.083 68099.42 441.89 441.89 68541.31
1.5 48.16 0.6891 1.35% 49.867 69580.04 353.41 513.34 69933.45
Strike Prices of Options for Overconfident Executives

2 55.88 0.7910 1.35% 51.393 70417.74 278.65 598.49 70696.39

2.5 63.74 0.8962 1.38% 52.351 70958.63 217.12 698.06 71175.75
1499
1500 O. Palmon and I. Venezia

expected compensation). In particular the optimal strike prices of the options that
stockholders award overconfident managers are increasing with their
overconfidence. While realistic managers estimate that there is a substantial prob-
ability that options with an at-the-money strike price will be worthless regardless of
their effort, overconfident managers may believe that their efforts will enhance the
values of such options making them valuable.
In practice executive options usually are provided with at-the-money strike
prices. When overconfidence or some other behavioral biases are not present,
theory has shown (see, e.g., Dittmann et al. 2010; Dittmann and Yu 2011; Palmon
et al. 2008), contrary to Hall and Murphy, that at-the-money prices are not optimal.
Hall and Murphy argue that at-the-money prices are optimal because they provide
maximum sensitivity to stock prices, but their argument does not hold when the
managers are risk averse and effort averse. Managers must be adequately compen-
sated for their efforts and for risk taking, and a balance must be reached between
their efforts, risk taking, and their pay. As Palmon et al. have shown, the optimal
balance is reached by issuing in-the-money options which do not necessarily
provide maximum sensitivity to stock prices. If managers are overconfident, how-
ever, that makes them more amenable for stock price sensitivity, and hence, they
will prefer higher strike prices which are closer to the at-the-money options usually
awarded in practice.
We also note in Table 55.1 that managerial overconfidence increases stock-
holders’ equity value. Given that the compensation is determined so as to equate
the manager’s expected utility to the target expected utility, it follows from
Table 55.1 that consistent with the results of Palmon and Venezia (2012), the
total welfare of both the managers and the stockholders improves with increased
managerial overconfidence. The fixed-level expected utility of the manager is
determined according to their subjective, overoptimistic perception. However,
when evaluated according to the realistic view, expected managerial compensa-
tion falls with overconfidence. We note that, nonetheless, the difference between
the monetary expected compensations according to the overoptimistic and real-
istic expectations is smaller than the monetary gains to stockholders from
overconfidence, so that the sum of realistic compensation and SEV rises with
overconfidence (see column 8). Thus, also in terms of realistic monetary values,
the welfare of the stakeholders (stockholders and managers) increases with
overconfidence.
In Table 55.2 we provide sensitivity analysis examining the effect of each of the
parameters on the behavior of stockholders and managers. We present the results of
only one or two changes in each of the exogenous parameters, but we conduct many
other simulations, and all provide the same qualitative results.15 In all the panels,
higher overconfidence measure is associated with higher strike prices, moneyness
levels, optimal managerial effort, value of the stockholders’ equity, and expected
compensation according to the managers’ expectations. They also are associated

15
The results of these simulations can be obtained from the authors upon request.
 55

Table 55.2 Stockholders’ equity values (SEV), strike prices (K), moneyness (K/S), effort levels (Y), and other parameters of interest. Departures from the
base case
Expected
compensation, Sum of SEV
Percentage of Stockholders’ realistic Expected and realistic
Strike price Moneyness firm given as Equity Value, stockholders’ compensation, expected
Overconfidence (K) in 1000s (K/S) options, (s) Effort (Y) SEV valuation managers’ valuation compensation
measure, l 1 2 3 4 5 6 7 8
Panel A: wage ¼ 25
1 31.56 0.4490 1.25 % 50.63 69,807.58 508.76 508.76 70,316.34
1.5 37.49 0.5252 1.14 % 52.83 71,000.72 413.11 563.28 71,413.83
2 43.57 0.6053 1.08 % 54.01 71,670.56 334.41 622.54 72,004.97
2.5 49.74 0.6875 1.04 % 54.75 72,106.28 269.37 688.74 72,375.65
Panel B: wage ¼ 75
1 51.70 0.7859 1.63 % 41.71 65,518.91 336.51 336.51 65,855.42
1.5 60.61 0.8965 1.70 % 45.36 67,411.64 269.97 420.45 67,681.61
2 69.91 1.0186 1.82 % 47.42 68,498.86 211.05 522.88 68,709.91
2.5 79.42 1.1462 2.01 % 48.72 69,195.89 166.20 653.44 69,362.09
Strike Prices of Options for Overconfident Executives

Panel C: s ¼ 0.2
1 48.97 0.6935 1.80 % 51.33 70,225.15 441.49 441.49 70,666.64
1.5 58.39 0.8138 1.83 % 53.60 71,489.22 312.64 542.64 71,801.86
2 68.10 0.9412 2.02 % 54.82 72,198.16 209.92 692.03 72,408.08
2.5 77.94 1.0716 2.35 % 55.56 72,642.57 137.08 912.80 72,779.65
Panel D: s ¼ 0.4
1 34.02 0.5090 1.19 % 43.77 66,438.75 444.93 444.93 66,883.68
1.5 40.00 0.5850 1.12 % 46.86 68,050.65 377.20 503.87 68,427.85
2 46.22 0.6675 1.09 % 48.59 68,975.32 317.47 566.82 69,292.79
2.5 52.58 0.7534 1.07 % 49.69 69,577.77 266.95 635.35 69,844.72
1501

(continued)
Table 55.2 (continued)
1502

Expected
compensation, Sum of SEV
Percentage of Stockholders’ realistic Expected and realistic
Strike price Moneyness firm given as Equity Value, stockholders’ compensation, expected
Overconfidence (K) in 1000s (K/S) options, (s) Effort (Y) SEV valuation managers’ valuation compensation
measure, l 1 2 3 4 5 6 7 8
Panel E: g ¼ 3
1 45.70 0.6801 2.09 % 44.49 66,732.83 513.41 513.41 67,246.24
1.5 53.94 0.7839 2.14 % 47.71 68,445.70 411.53 639.26 68,857.23
2 62.52 0.8969 2.28 % 49.52 69,438.59 319.32 795.50 69,757.91
2.5 71.28 1.0142 2.49 % 50.66 70,086.17 243.48 992.82 70,329.65
Panel F: g ¼ 5
1 36.91 0.5325 1.08 % 48.73 68,964.71 400.77 400.77 69,365.48
1.5 43.74 0.6198 1.01 % 51.24 70,295.63 323.30 449.43 70,618.93
2 50.78 0.7127 0.97 % 52.59 71,038.19 258.80 504.00 71,296.99
2.5 57.95 0.8085 0.96 % 53.44 71,515.08 206.66 567.94 71,721.74
Panel G: Ν ¼ 2,000
1 44.54 0.5886 1.36 % 61.44 75,242.70 477.96 477.96 75,720.66
1.5 54.10 0.7013 1.31 % 64.39 76,830.62 364.92 565.18 77,195.55
2 63.94 0.8204 1.32 % 65.98 77,714.30 274.85 672.50 77,989.15
2.5 73.93 0.9426 1.37 % 66.97 78,277.56 205.29 804.84 78,482.85
Panel H: Ν ¼ 6,000
1 38.87 0.5977 1.41 % 40.15 64,656.11 421.35 421.35 65,077.47
1.5 45.30 0.6826 1.36 % 42.83 66,067.86 346.87 486.08 66,414.74
2 51.97 0.7745 1.35 % 44.31 66,875.94 279.15 557.17 67,155.09
2.5 58.80 0.8701 1.38 % 45.25 67,401.85 223.48 643.22 67,625.33
O. Palmon and I. Venezia
55 Strike Prices of Options for Overconfident Executives 1503

with lower expected managerial compensation according to the realistic expecta-

tion. This indicates that the qualitative results obtained from the base case prevail
also for a host of other parameters.
In panels A and B, we examine the effect of the fixed wages on the results.
A higher wage level implies a lower value for the option component of the compen-
sation. Imposing the use of lower-valued options, stockholders choose options that
are more responsive to cash flow changes. This is obtained by an increase in the
ownership percentage and in the strike price. In panels C and D, we set the coefficient
of variation, s, which equals 0.3 in the base case, to 0.2 and 0.4, respectively. We
observe in panel D that facing a higher coefficient of variation, managers prefer
a compensation that is less sensitive to firm cash flows, which is achieved by selecting
a contract specifying a smaller ownership fraction and a lower strike price. Finally, in
panels E and F, we observe the effects of varying risk aversion and in panels G and
H those of varying effort aversion. Overall, higher risk aversion levels are associated
with less risky compensations as they induce optimal contracts with a smaller option
ownership percentage and a lower strike price. Higher effort aversion results in lower
SEV and also in lower total monetary welfare.
The above sensitivity analysis shows that the effect of changing the parameters
quite conforms to intuition, adding to the robustness of our results. We also note
that regardless of the parameters considered, the effect of overconfidence on the
qualitative behavior is the same as that observed from the base case. In particular,
the higher the overconfidence, the higher the optimal strike prices and the closer
they are to the at-the-money levels.

55.5 Conclusion

Our study suggests an explanation for the puzzling questions of why most incentive
stock options are issued with at-the-money strike prices. This practice seems
arbitrary and beyond its institutional appeal and its expired tax advantages; its
main theoretical backing is that it provides the highest sensitivity to stock price.
Several studies however have shown that in many cases it is inferior to awarding
in-the-money options. Our analysis demonstrates that the optimal strike prices of
incentive stock options when managers are overconfident are higher than the
corresponding strike prices when managers are realistic, and are closer to the
at-the-money strike prices awarded in practice. This makes at-the-money options
more attractive to overconfident managers, and hence, given the ubiquity of
overconfident managers, it provides support for the popularity of awarding such
options. We also show that overoptimistic CEOs receive lower compensation than
their realistic counterparts and that the stockholders benefit from their managers’
bias. The combined welfare of the firm’s stakeholders however is positively related
to managerial overconfidence, hence providing support to the survival of manage-
rial overconfidence.
Assef and Santos (2005) interpret the strike price as an intermediate instrument
(between wages and stocks) in the incentive schemes for managers. Similarly one
1504 O. Palmon and I. Venezia

can interpret an in-the-money strike price as an intermediate instrument between

a stock (zero strike price) and an at-the-money option. Since in practice, because of
institutional reasons or inertia, firms are constrained to choose options with
at-the-money strike price, they achieve their instrumental in-the-money strike
price by choosing an appropriate weight of options relative to stock grants in
their compensation contract. According to such an interpretation and from our
results showing that higher overconfidence implies higher strike prices, it follows
that the observable weight of options in the compensation contract may serve as
a proxy for an unobservable degree of confidence.

Appendix 1

In this appendix we expand on the simulations we conduct. These simulations are

intended to identify the contracts that yield the highest stockholders’ equity value
subject to manager’s incentive compatibility and participation constraints. That is,
the managers choose their effort optimally, and their resulting expected utility
equals a predetermined level representing their alternative opportunities. Because
we are studying the impact of overconfidence on the strike price, our calculations
focus on the trade-off between the strike price and the fraction of the company that
is awarded as options. For simplicity, we consider contracts that include only
a fixed wage and options.
The first step in our simulation is the selection of the appropriate distribution of
the company’s cash flow as a function of managerial effort and the manager’s utility
as a function of managerial effort and compensation. In accordance with conven-
tional assumptions in the options literature, we assume that the firms’ cash flows, X,
are lognormally distributed with the distribution function (55.1) where Y denotes
the managerial extra effort (a managerial choice variable) and m(Y) and s denote,
respectively, the mean and the standard deviation of the underlying normal distri-
bution of the natural logarithm of X. We assume that managerial effort increases
cash flows, that overconfident managers overestimate the impact of their effort on
cash flows, and that the impact of effort on the mean of the natural logarithm of X is
presented in Eq. 55.2.
We refer to Hek (1999) and Bitler et al. (2005) for the choice of the parameters
and the shape of the manager’s utility function that depends also on leisure.16
We start with a base case of parameters and repeat the analysis for a large set of
parameters around the base case. We chose the base case so that these parameters
and the deviations around them that we also analyze cover the equivalent param-
eters used in similar studies. These simulations help verify that our results are
robust to the choice of parameter values. They also are used to examine to what

16
We found additional estimates of effort disutility (leisure utility) in the following papers: Dowell
(1985), Kiker and Mendes de Oliveira (1990), and Prasch (2001). These estimates varied in the
functional form as well as in the level of effort aversion.
55 Strike Prices of Options for Overconfident Executives 1505

extent the effects of changes in the parameter values on the outcomes coincide with
economic intuition.
The parameter m0 serves as a numeraire for the other cash flows related param-
eters, and is chosen, without loss of generality, to equal 45,000. That, in the absence
of managerial extra effort, the expected value of the company’s cash flows is
45,000. Since the expected cash flows serve as numeraire, the ratio of the standard
deviation of the cash flows per share to their expected value is a surrogate for the
standard deviation of stock returns. Since Hall and Murphy (2000) used a standard
deviation of 0.3, we chose this value also for our base case coefficient of variation.
The appropriate measure of risk aversion is harder to agree upon. Early estimates
of risk aversion put this variable at around two (see, e.g., Mehra and Prescott 1985),
but they are based on aggregate data and not on CEO compensation data.17 In our
study, in line with more advanced econometric methods (see, e.g., Campbell
et al. 1996), we prefer using a base case risk aversion measure of four, slightly
higher than the measure of three suggested by Malmendier and Tate (2008) and
Hall and Liebman (1998). Our simulations (see, e.g., Glasserman 2003) and
sensitivity analysis, of course, cover these parameters as well.
The next step in the simulation process is to identify, for each overconfidence
level, the executive options’ strike price that is optimal for stockholders. All the
simulations were conducted using Mathematica. Because it is not possible to
express the equity value as an explicit function of the strike price, we search for
the optimal strike price by calculating the equity values that are associated with
a set of discrete strike prices. Our search was facilitated by assuming that the
stockholders know the manager’s reservation expected utility. We assume that
reservation utility to equal the utility obtained from a fixed salary of 100 with no
extra effort.
For any given wage, the strike price and the fraction of the company awarded to
the manager (which is a continuous variable representing the number of options
the manager receives; we will henceforth use the latter expression) determine the
value of the options to the managers and their cost to the stockholders. For each
strike price and number of options, we then find the effort that the manager chooses
to apply in order to maximize his/her expected utility Eq. 55.3. For each given
strike price, the stockholders, well aware of the managers’ reactions, will offer them
the number of options that yield their reservation utility. We calculate the value of
the stockholders’ equity for each strike price (in thousands of dollars, using two
digits beyond the decimal point) and identify the strike price that yields a maximum
for stockholders’ equity.
For each set of parameters for the cash flow distribution function and the
managerial utility function, as well as for the several values of fixed salary
(50 for the base case, 25 and 75 for the presented robustness simulations), we
obtain the optimal effort, stockholders’ equity value, and the expected managerial

17
Similar estimates are provided in other contexts by Carpenter (2000), Constantinides
et al. (2002), Epstein and Zin (1991), Friend and Blume (1975), and Levy (1994).
1506 O. Palmon and I. Venezia

compensation according to the manager’s overconfident view and the stockholders’

realistic expectations. We repeat these simulations for several values of the
overconfidence measure l. Presented here however are just four such values:
1 for the realistic expectations and 1.5, 2, and 2.5 for increasing levels of
overconfidence. We used quite a few simulations but choose to present a subset
of the results as all showed the same qualitative results. In addition to verifying the
robustness of the results to the choice of the parameter values, they also help
examine to what extent the effects of changes in the parameter values on the
outcomes coincide with economic intuition.

Acknowledgment We thank Darius Palia, Orly Sade, and seminar participants at Rutgers
University and the Universitat Pompeu Fabra for helpful comments and suggestions. The financial
support of The Sanger Family Chair for Banking and Risk Management, The Galanter Fund, The
Mordecai Zagagi Fund, the Whitcomb Center for Research in Financial Services, and The School
of Accounting, the Hebrew University are gratefully acknowledged.

References
Assef, J. G., & Santos, M. (2005). Stock options and managerial optimal contracts. Economic
Theory, 26, 813–837.
Bergman, N. K., & Jenter, D. (2007). Employee sentiment and stock option compensation. Journal
of Financial Economics, 84, 667–712.
Bitler, M. P., Moskowitz, T. J., & Vissing-Jorgensen, A. (2005). Testing agency theory with
entrepreneur extra-effort and wealth. Journal of Finance, 60(2), 539–576.
Campbell, J., Lo, A., & MacKinlay, C. (1996). The econometrics of financial markets. New Jersey,
Princeton: Princeton University Press.
Carpenter, J. N. (2000). Does option compensation increase managerial risk appetite. Journal of
Finance, 55, 2311–2331.
Constantinides, G. M., Donaldson, J. M., & Mehra, R. (2002). Junior can’t borrow: A new
perspective on the equity premium puzzle. Quarterly Journal of Economics, 117, 269–296.
Dittmann, I., & Yu, K. C. (2011). How important are risk-taking incentives in executive compen-
sation? (Working Paper). Rotterdam: Erasmus University.
Dittmann, I., Maug, E., & Spalt, O. (2010). Sticks or carrots: Optimal CEO compensation when
managers are loss-averse. Journal of Finance, 65, 2015–2050.
Dowell, R. (1985). Risk preference and the work-leisure trade-off. Economic-Inquiry, 23,
691–701.
Epstein, L. G., & Zin, S. E. (1991). Risk aversion, and the temporal behavior of consumption and
asset returns: An empirical analysis. Journal of Political Economy, 99, 263–86.
Friend, I., & Blume, M. (1975). The demand for risky assets. American Economic Review, 65,
900–927.
Gervais, S., Heaton, J. B., & Odean, T. (2011). Overconfidence, compensation contracts, and
capital budgeting. Journal of Finance, 66, 1735–1777.
Glaser, M., & Weber, M. (2007). Overconfidence and trading volume. The Geneva Risk and
Insurance Review, 32, 1–36.
Glasserman, P. (2003). Monte Carlo methods in financial engineering. Berlin: Springer.
Hall, B., & Liebman, J. B. (1998). Are CEOs really paid like bureaucrats? Quarterly. Journal of
Economics, 113, 653–691.
Hall, B., & Murphy, K. J. (2000). Optimal exercise prices for executive stock options. American
Economic Review, 90, 209–14.
55 Strike Prices of Options for Overconfident Executives 1507

Hall, B. J., & Murphy, K. J. (2002). Stock options for undiversified executives. Journal of
Accounting and Economics, 33, 3–42.
Hek, P. A. (1999). Endogenous technological change with leisure-dependent utility. Economic
Theory, 14, 669–84.
Holmstrom, B. (1979). Moral hazard and observability. Bell Journal of Economics, 10, 74–91.
Kiker, B. F., & Mendes-de-Oliveira, M. (1990). Estimation and valuation of non-leisure time.
Oxford Bulletin of Economics and Statistics, 52, 115–141.
Levy, H. (1994). Absolute and relative risk aversion: An experimental study. Journal of Risk and
Uncertainty, 8, 289–307.
Mahajan, S. (2002). Role of out-of-money options in executive compensation (Working Paper).
New York: NYU, Stern School of Business.
Malmendier, U., & Tate, G. (2005a). Does overconfidence affect corporate investment? CEO
overconfidence measures revisited. European Financial Management, 11, 649–659.
Malmendier, U., & Tate, G. (2005b). CEO overconfidence and corporate investment. Journal of
Finance, 60, 2661–2700.
Malmendier, U., & Tate, G. (2008). Who makes acquisitions? CEO overconfidence and the
market’s reaction. Journal of Financial Economics, 89, 20–43.
Mehra, R., & Prescott, E. C. (1985). The equity premium: A puzzle. Journal of Monetary
Economics, 15, 145–161.
Otto, C. A. (2011). CEO optimism and incentive compensation (Working Paper). London:
Business School.
Oyer, P., & Schaefer, S. (2005). Why do some firms give stock options to all employees? An
empirical examinations of alternative theories. Journal of Financial Economics, 76, 99–133.
Palmon, O., & Venezia, I. (2012, forthcoming). A rationale for hiring irrationally overconfident
managers. In Encyclopedia of finance (2nd ed.). Berlin: Springer.
Palmon, O., Chen, R. R., Bar Yosef, S., & Venezia, I. (2008). Optimal strike prices of stock options
for effort averse executives. Journal of Banking and Finance, 32, 229–239.
Prasch, R. E. (2001). Work, leisure, and the labor supply curve: A reply to David Spencer. Journal
of Economic Issues, 35, 1001–1007.
Roll, R. (1986). The hubris hypothesis of corporate takeovers. Journal of Business, 59, 197–216.
Suntheim, F. (2012). CEO overconfidence in banking (Working Paper). Milano: Bocconi
University.
Density and Conditional
Distribution-Based Specification Analysis 56
Diep Duong and Norman R. Swanson

Contents
56.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1510
56.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1513
56.2.1 Diffusion Models in Finance and Financial Econometrics . . . . . . . . . . . . . . . . . . 1513
56.2.2 Overview on Specification Tests and Model Selection . . . . . . . . . . . . . . . . . . . . . . 1519
56.3 Consistent Distribution-Based Specification Tests and Predictive
Density-Type Model Selection for Diffusion Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1521
56.3.1 One-Factor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1521
56.3.2 Multifactor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1532
56.3.3 Model Simulation and Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1537
56.3.4 Bootstrap Critical Value Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1547
56.4 Summary of Empirical Applications of the Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1553
56.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1558
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1558

Abstract
The technique of using densities and conditional distributions to carry out
consistent specification testing and model selection amongst multiple diffusion
processes has received considerable attention from both financial theoreticians
and empirical econometricians over the last two decades. In this chapter, we
discuss advances to this literature introduced by Corradi and Swanson
(J Econom 124:117–148, 2005), who compare the cumulative distribution
(marginal or joint) implied by a hypothesized null model with corresponding
empirical distributions of observed data. We also outline and expand upon

D. Duong (*)
Department of Business and Economics, Utica College, Utica, NY, USA
e-mail: dnduong@utica.edu
N.R. Swanson
Department of Economics, Rutgers, The State University of New Jersey, New Brunswick, NJ, USA
e-mail: nswanson@econ.rutgers.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1509
DOI 10.1007/978-1-4614-7750-1_56,
# Springer Science+Business Media New York 2015
1510 D. Duong and N.R. Swanson

further testing results from Bhardwaj et al. (J Bus Econ Stat 26:176–193, 2008)
and Corradi and Swanson (J Econom 161:304–324, 2011). In particular, para-
metric specification tests in the spirit of the conditional Kolmogorov test of
Andrews (Econometrica 65:1097–1128, 1997) that rely on block bootstrap
resampling methods in order to construct test critical values are first discussed.
Thereafter, extensions due to Bhardwaj et al. (J Bus Econ Stat 26:176–193,
2008) for cases where the functional form of the conditional density is unknown
are introduced, and related continuous time simulation methods are introduced.
Finally, we broaden our discussion from single process specification testing to
multiple process model selection by discussing how to construct predictive
densities and how to compare the accuracy of predictive densities derived
from alternative (possibly misspecified) diffusion models. In particular, we
generalize simulation steps outlined in Cai and Swanson (J Empir Financ
18:743–764, 2011) to multifactor models where the number of latent variables
is larger than three. We finish the chapter with an empirical illustration of model
selection amongst alternative short-term interest rate models.

Keywords
Multifactor diffusion process • Specification test • Out-of-sample forecasts •
Conditional distribution • Model selection • Block bootstrap • Jump process

56.1 Introduction

The last three decades have provided a unique opportunity to observe numerous
interesting developments in finance, financial econometrics, and statistics. For
example, although starting as a narrow subfield, financial econometrics has recently
transformed itself into an important discipline, equipping financial economic
researchers and industry practitioners with immensely helpful tools for estimation,
testing, and forecasting. One of these developments has involved the development
of “state-of-the-art” consistent specification tests for continuous time models,
including not only the geometric Brownian motion process used to describe the
dynamics of asset returns (Merton (1973)) but also a myriad of other diffusion
models used in finance, such as the Ornstein-Uhlenbeck process introduced by
Vasicek (1977); the constant elastic volatility process applied by Beckers (1980);
the square root process due to Cox et al. (1985); the so-called CKLS model by Chan
et al. (1992); various three-factor models proposed Chen (1996); stochastic vola-
tility processes such as generalized CIR of Andersen and Lund (1997); and the
generic class of affine jump diffusion processes discussed in Duffle et al. (2000).1
The plethora of available diffusion models allow decision makers to be flexible
when choosing a specification to be subsequently used in contexts ranging from
equity and option pricing, to term structure modeling and risk management.

1
For complete details, see Sect. 56.2.2.
56 Density and Conditional Distribution-Based Specification Analysis 1511

Moreover, the use of high-frequency data when estimating such model, in contin-
uous time contexts, allows investors to continuously update their dynamic trading
strategies in real time.2 However, for statisticians and econometricians, the vast
number of available models has important implications for formalizing model
selection and specification testing methods. This has led to several key papers
that have recently been published in the area of parametric and nonparametric
specification testing. Most of the papers focus on the ongoing “search” for correct
Markov and stationary models that “fit” historical data and associated dynamics. In
this literature, it is important to note that correct specification of a joint distribution
is not the same as that of a conditional distribution, and hence the recent focus on
conditional distributions, given that most models have an interpretation as condi-
tional models. In summary, the key issue in the construction of model selection and
specification tests of conditional distributions is the fact that knowledge of the
transition density (or conditional distribution) in general cannot be inferred from
knowledge of the drift and variance terms of a diffusion model. If the functional
form of the density is available parametrically, though, one can test the hypothesis
of correct specification of a diffusion via the probability integral transform
approach of Diebold et al. (1998); the cross-spectrum approach of Hong (2001),
Hong and Li (2005), and Hong et al. (2007); the martingalization-type Kolmogorov
test of Bai (2003); or the normality transformation approaches of Bontemps and
Meddahi (2005) and Duan (2003). Furthermore, if the transition density is
unknown, one can construct a nonparametric test by comparing a kernel density
estimator of the actual and simulated data, for example, as in Altissimo and Mele
(2009) and Thompson (2008), or by comparing the conditional distribution of the
simulated and the historical data, as in Bhardwaj et al. (2008). One can also use the
methods of Aı̈t-Sahalia (2002) and Aı̈t-Sahalia et al. (2009), in which they compare
closed form approximations of conditional densities under the null, using data-
driven kernel density estimates.
For clarity and ease of presentation, we categorize the above literature into two
areas. The first area, initiated by the seminal work of Aı̈t-Sahalia (1996) and later
followed by Pritsker (1998) and Jiang (1998), breaks new ground in the continuous
time specification testing literature by comparing marginal densities implied by
hypothesized null models with nonparametric estimates thereof. These sorts of
tests examine one-factor specifications. The second area of testing, as initiated in
Corradi and Swanson (2005), does not look at densities. Instead, they compare
cumulative distributions (marginal, joint, or conditional) implied by a hypothesized
null model with corresponding empirical distributions. A natural extension of these
sorts of tests involves model selection amongst alternative predictive densities
associated with competing models. While Corradi and Swanson (2005) focus on
cases where the functional form of the conditional density is known, Bhardwaj
et al. (2008) use simulation methods to examine testing in cases where the func-
tional form of the conditional density is unknown. Corradi and Swanson (2011) and

2
For further discussion, see Duong and Swanson (2010, 2011).
1512 D. Duong and N.R. Swanson

Cai and Swanson (2011) take the analysis of Bhardwaj et al. (2008) on Step further
and focus on the comparison of out-of-sample predictive accuracy of possibly
misspecified diffusion models, when the conditional distribution is not known in
closed form (i.e., they “choose” amongst competing models based on predictive
density model performance). The “best” model is selected by constructing tests that
compare both predictive densities and predictive conditional confidence intervals
associated with alternative models.
In this chapter, we primarily focus our attention on the second area of the model
selection and testing literature.3 One feature of all the tests that we shall discuss is
that, given that they are based on the comparison of CDFs, they obtain parametric
rates. Moreover, the tests can be used to evaluate single and multiple factor and
dimensional models, regardless of whether or not the functional form of the
conditional distribution is known.
In addition to discussing simple diffusion process specification tests of Corradi
and Swanson (2005), we discuss tests discussed in Bhardwaj et al. (2008) and
Corradi and Swanson (2011) and provide some generalizations and additional
results. In particular, parametric specification tests in the spirit of the conditional
Kolmogorov test of Andrews (1997) that rely on block bootstrap resampling
methods in order to construct test critical values are first discussed. Thereafter,
extensions due to Bhardwaj et al. (2008) for cases where the functional form of the
conditional density is unknown are introduced, and related continuous time simu-
lation methods are introduced. Finally, we broaden our discussion from single
dimensional specification testing to multiple dimensional selection by discussing
how to construct predictive densities and how to compare the accuracy of predictive
densities derived from alternative (possibly misspecified) diffusion models as in
Corradi and Swanson (2011). In addition, we generalize simulation and testing
procedures introduced in Cai and Swanson (2011) to more complicated multifactor
and multidimensional models where the number of latent variables is larger than
three. These final tests can be thought of as continuous time generalizations of the
discrete time “reality check” test statistics of White (2000), which are widely used
in empirical finance (see, e.g., Sullivan et al. (1999, 2001)). We finish the chapter
with an empirical illustration of model selection amongst alternative short-term
interest rate models, drawing on Bhardwaj et al. (2008), Corradi and Swanson
(2011) and Cai and Swanson (2011).
Of the final note is that the test statistics discussed here are implemented via use
of simple bootstrap methods for critical value simulation. We use the bootstrap
because the covariance kernels of the (Gaussian) asymptotic limiting distributions
of the test statistics are shown to contain terms deriving from both the contribution
of recursive parameter estimation error (PEE) and the time dependence of data.
Asymptotic critical value thus cannot be tabulated in a usual way. Several methods
can easily be implemented in this context. First one can use block bootstrapping
procedures, as discussed below. Second one can use the conditional p-value

3
For a recent survey on results in the first area of this literature, see Aı̈t-Sahalia (2007).
56 Density and Conditional Distribution-Based Specification Analysis 1513

approach of Corradi and Swanson (2002) which extends the work of Hansen (1996)
and Inoue (2001) to the case of nonvanishing parameter estimation error. Third is
the subsampling method of Politis et al. (1999), which has clear efficiency “costs,”
but is easy to implement. Use of the latter two methods yields simulated
(or subsample based) critical values that diverge at rate equivalent to the block
size length under the alternative. This is the main drawback to their use in our
context. We therefore focus on use of a block bootstrap that mimics the contribution
of parameter estimation error in a recursive setting and in the context of time series
data. In general, use of the block bootstrap approach is made feasible by
establishing consistency and asymptotic normality of both simulated generalized
method of moments (SGMM) and nonparametric simulated quasi-maximum like-
lihood (NPSQML) estimators of (possibly misspecified) diffusion models, in
a recursive setting, and by establishing the first-order validity of their bootstrap
analogs.
The rest of the paper is organized as follows. In Sect. 56.2, we present our
setup and discuss various diffusion models used in finance and financial economet-
rics. Section 56.3 outlines the specification testing hypotheses, presents the
cumulative distribution-based test statistics for one-factor and multiple-
factor models, discusses relevant procedures for simulation and estimation, and
outlines bootstrap techniques that can be used for critical value tabulation.
In Sect. 56.4, we present a small empirical illustration. Section 56.5 summarizes
and concludes.

56.2 Setup

56.2.1 Diffusion Models in Finance and Financial Econometrics

For the past two decades, continuous time models have taken center stage in the
field of financial econometrics, particularly in the context of structural modeling,
option pricing, risk management, and volatility forecasting. One key advantage of
continuous time models is that they allow financial econometricians to use the full
information set that is available. With the availability of high-frequency data and
current computation capability, one can update information, model estimates, and
predictions in milliseconds. In this section, we will summarize some of the standard
models that have been used in asset pricing as well as term structure modeling.
Generally, assume that financial asset returns follow Ito-semimartingale processes
with jumps, which are the solution to the following stochastic differential equation
system:
ðt ð
Xðt
Þ ¼ bðXðs
Þ, y0 Þds
l0 t yfðyÞdy
0 Y
ðt X
Jt (56:1)
þ sðXðs
Þ, y0 ÞdW ðsÞ þ yj ,
0 j¼1
1514 D. Duong and N.R. Swanson

where X(t
) is a cadlag process (right continuous with left limit) for t 2 ℜ+ and is an
N-dimensional vector of variables, W(t) is an N-dimensional Brownian motion, b(·)
is N-dimensional function of X(t
), and s(·) is an N  N matrix-valued function of
X(t
), where y0 is an unknown true parameter. Jt is a Poisson process with intensity
parameter l0, l0 finite, and the N-dimensional jump size, yj, is i.i.d. with marginal
distribution given by f. Both Jt and yj are assumed
Ð to be independent of the driving
Brownian motion, W(t).4 Also, note that Yyf(y)dy denotes the mean jump size,
hereafter denoted by m0. Over a unit time interval, there are on average l0 jumps, so
that over the time span [0, t], there are on average l0t jumps. The dynamics of X(t
)
is then given by
 
dXðtÞ ¼ bðXðt
Þ, y0 Þ
l0 my, 0 dt
ð
(56:2)
þ sðXðt
Þ, y0 ÞdW ðtÞ þ ypðdy; dtÞ,
Y

where p(dy, dt) is a random Poisson measure giving point mass at y if a jump occurs
in the interval dt and b(·), s(·) are the “drift” and “volatility” functions defining the
parametric specification of the model. Hereafter, the same (or similar) notation is
used throughout when models are specified.
Through not an exhaustive list, we review some popular models. Models are
presented with the “true” parameters.

56.2.1.1 Diffusion Models Without Jumps

Geometric Brownian Motion (Log Normal Model)
In this setup, b(X(t
), y0) ¼ b0X(t) and s(X(t
), y0) ¼ s0X(t).

dXðtÞ ¼ b0 XðtÞdt þ s0 XðtÞdW ðtÞ,

where b0 and s0 are constants and W(t) is a one-dimensional standard Brownian

motion. (Below, other constants such as a0, b0, l0, g0, d0, 0, k0, and O0 are also
used in model specifications.)
This model is popular in the asset pricing literature. For example, one can model
equity prices according to this process, especially in the Black-Scholes option setup
or in structured corporate finance.5 The main drawback of this model is that the
return process (log(price)) has constant volatility and is not time varying. However,
it is widely used as a convenient “first” econometric model.
Vasicek (1977) and Ornstein-Uhlenbeck Process: The process is used to model
asset prices, specifically in term structure modeling, and the specification is

dXðtÞ ¼ ða0 þ b0 XðtÞÞdt þ s0 dW ðtÞ,

Hereafter, X(t
) denotes the cadlag, while Xt denotes discrete skeleton for t ¼ 1, 2, . . . .
4

5
See Black and Scholes (1973) for details.
56 Density and Conditional Distribution-Based Specification Analysis 1515

where W(t) is a standard Brownian motion and a0, b0, and s0 are constants. b0 is
negative to ensure the mean reversion of X(t).
Cox et al. (1985) use the following square root process to model the term
structure of interest rates:

pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kða0
XðtÞÞdt þ s0 XðtÞdW ðtÞ,

where W(t) is a standard Brownian motion, a0 is the long-run mean of X(t), k

measures the speed of mean reversion, and s0 is a standard deviation parameter and
is assumed to be fixed. Also, non-negativity of the process is imposed, as 2kb0 > s20.
Wong (1964) points out that in the CIR model, X(t) with the dynamics evolving
according to
pffiffiffiffiffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ ðða0
l0 Þ
XðtÞÞdt þ a0 XðtÞdW ðtÞ, (56:3)

a0 > 0 and a0
l0 > 0,

belongs to the linear exponential (or Pearson) family with a closed form cumulative
distribution. a0 and l0 are fixed parameters of the model.
The constant elasticity of variance or CEV model is specified as follows:

dXðtÞ ¼ a0 XðtÞdt þ s0 XðtÞb0 =2 dW ðtÞ,

where W(t) is a standard Brownian motion and a0, s0, and b0 are fixed constants.
Of note is that the interpretation of this model depends on b0, i.e., in the case of
stock prices, if b0 ¼ 2, then the price process X(t) follows a lognormal diffusion; if
b0 < 2, then the model captures exactly the leverage effect as price and volatility
are inversely correlated.
Amongst other authors, Beckers (1980) used this CEV model for stocks, Marsh
and Rosenfeld (1983) apply a CEV parametrization to interest rates, and Emanuel
and Macbeth (1982) utilize this setup for option pricing.
The generalized constant elasticity of variance model is defined as follows:
 
dXðtÞ ¼ a0 XðtÞ
ð1
b0 Þ þ l0 XðtÞ dt þ s0 XðtÞb0 =2 dW ðtÞ,

where the notation follows the CEV case. l0 is another parameter of the model. This
process nests log diffusion when b0 ¼ 2 and nests square root diffusion when
b0 ¼ 1.
Brennan and Schwartz (1979) and Courtadon (1982) analyze the model:

dXðtÞ ¼ ða0 þ b0 XðtÞÞdt þ s0 XðtÞ2 dW ðtÞ,

where a0, b0, s0 are fixed constants and W(t) is a standard Brownian motion.
1516 D. Duong and N.R. Swanson

Duffie and Kan (1996) study the specification:

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ ða0 þ XðtÞÞdt þ b0 þ g0 XðtÞdW ðtÞ,

where W(t) is a standard Brownian motion and a0, b0, and g0 are fixed parameters.
Aїt-Sahalia (1996) looks at a general case with general drift and CEV diffusion:

 
dXðtÞ ¼ a0 þ b0 XðtÞ þ g0 XðtÞ2 þ 0 =XðtÞ dt þ s0 XðtÞb0 =2 dW ðtÞ:

In the above expression, a0, b0, g0, 0, s0, and b0 are fixed constants and W(t) is
again a standard Brownian motion.

56.2.1.2 Diffusion Models with Jumps

For term structure modeling in empirical finance, the most widely studied class of
models is the family of affine processes, including diffusion processes that incor-
porate jumps.
Affine Jumps Diffusion Model: X(t
) is defined to follow an affine jump diffusion if

pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ k0 ða0  XðtÞÞdt þ O0 DðtÞdW ðtÞ þ dJ ðtÞ,

where X(t
) is an N-dimensional vector of variables of interest and is a cadlag

process, W(t) is an N-dimensional independent standard Brownian motion, k0 and
O0 are square N  N matrices, a0 is a fixed long-run mean, and D(t) is a diagonal
matrix with ith diagonal element given by

dii ðtÞ ¼ y0i þ d00i XðtÞ:

In the above expressions, y0i and d0i 0 are constants. The jump intensity is
assumed to be a positive, affine function of X(t), and the jump size distribution
is assumed to be determined by its conditional characteristic function. The attrac-
tive feature of this class of affine jump diffusions is that, as shown in Duffie
et al. (2000), it has an exponential affine structure that can be derived in closed
form, i.e.,
 
FðXðtÞÞ ¼ exp aðtÞ þ bðtÞ0 XðtÞ ,

where the functions a(t) and b(t) can be derived from Riccati equations.6
Given a known characteristic function, one can use either GMM to estimate the

6
For details, see Singleton (2006), p. 102.
56 Density and Conditional Distribution-Based Specification Analysis 1517

parameters of this jump diffusion, or one can used quasi-maximum likelihood

(QML), once the first two moments are obtained. In the univariate case without
jumps, as a special case, this corresponds to the above general CIR model with
jumps.

56.2.1.3 Multifactor and Stochastic Volatility Model

Multifactor models have been widely used in the literature, particularly in
option pricing, term structure, and asset pricing. One general setup has (X(t),
V(t))0 ¼ (X(t), V1(t), . . ., Vd(t))0 where only the first element, the diffusion
process Xt, is observed while V(t) ¼ (V1(t), . . ., Vd(t))dx1
0 is latent. In addition,
X(t) can be dependent on V(t). For instance, in empirical finance, the most
well-known class of the multifactor models is the stochastic volatility model
expressed as

! ! !
dXðtÞ b1 ðXðtÞ, y0 Þ s11 ðV ðtÞ, y0 Þ
¼ dt þ dW 1 ðtÞ
dV ðtÞ b2 ðV ðtÞ, y0 Þ 0
! (56:4)
s12 ðV ðtÞ, y0 Þ
þ dW 2 ðtÞ,
s22 ðV ðtÞ, y0 Þ

where W1(t)11 and W2(t)11 are independent standard Brownian motions and V(t)
is latent volatility process. b1(·) is a function of X(t) and b2(·), s11(·), s22(·), and
s22(·) are general functions of V(t), such that system of Eq. 56.4 is well defined.
Popular specifications are the square root model of Heston (1993), the GARCH
diffusion model of Nelson (1990), lognormal model of Hull and White (1987), and
the eigenfunction models of Meddahi (2001). Note that in this stochastic volatility
case, the dimension of volatility is d ¼ 1. More general setup can involve d driving
Brownian motions in V(t) equation.
As an example, Andersen and Lund (1997) study the generalized CIR model
with stochastic volatility, specifically

pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kx0 ðx0
XðtÞÞdt þ V ðtÞdW 1 ðtÞ,
pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kv0 ðv0
V ðtÞÞdt þ sv0 V ðtÞdW 2 ðtÞ,

where X(t) and V(t) are price and volatility processes, respectively, kx0, kv0 > 0 to
ensure stationarity, x0 is the long-run mean of (log) price process, and v0 and sv0 are
constants. W1(t) and W2(t) are scalar Brownian motions. However, W1(t) and W2(t)
are correlated such that dW1(t)dW2(t) ¼ rdt where the correlation r is some
constant r 2 [
1, 1]. Finally, note that volatility is a square root diffusion process,
which requires that kv0 v0 > s2v0 .
1518 D. Duong and N.R. Swanson

Stochastic Volatility Model with Jumps (SVJ): A standard specification is

 pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kx0 x0
XðtÞÞdt þ V ðtÞdW 1 ðtÞ þ J u dqu
J d dqd ,
pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kv0 ðv0
V ðtÞÞdt þ sv0 V ðtÞdW 2 ðtÞ,

where qu and qd are Poisson processes with jump intensity parameters lu and ld,
respectively, and are independent of the Brownian motions W1(t) and W2(t).
In particular, lu is the probability of a jump-up, Pr(dqu (t) ¼ 1) ¼ lu, and ld is
the probability of a jump-down, Pr(dqd(t) ¼ 1) ¼ ld. Ju and Jd are jump-up  
and jump-down sizes and have exponential distributions: f ðJ u Þ ¼ B1u exp
JBuu
 
and f ðJ d Þ ¼ B1d exp
JBdd , where Bu, Bd > 0 are the jump magnitudes, which are
the means of the jumps, Ju and Jd.
Three-Factor Model (CHEN): The three-factor model combines various features
of the above models, by considering a version of the oft examined three-factor
model due to Chan et al. (1992), which is discussed in detail in Dai and Singleton
(2000). In particular,
pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kx0 ðyðtÞ
XðtÞÞdt þ pVffiffiffiffiffiffiffiffi
ðtÞdW ffi 1 ðtÞ,
dV ðtÞ ¼ kv0 ðv
V ðtÞÞdt þ sv0 V ð t Þ
pffiffiffiffiffiffiffiffidW 2 ðtÞ, (56:5)
 
dyðtÞ ¼ ky0 yðtÞ
yðtÞ dt þ sy0 yðtÞdW 3 ðtÞ,

where W1(t), W2(t) W3(t) are independent Brownian motions and V and y are the
stochastic volatility and stochastic mean of X(t), respectively. kx0, kv0, ky0, v0 , y0 ,
sv0, sy0 are constants. As discussed above, non-negativity for V(t) and y(t) requires
that 2kv0 v0 > s2v0 and 2ky0 y0 > s2y0 .
Three-Factor Jump Diffusion Model (CHENJ): Andersen et al. (2004) extend the
three-factor Chen (1996) model by incorporating jumps in the short rate process,
hence improving the ability of the model to capture the effect of outliers and to
address the finding by Piazzesi (2004, 2005) that violent discontinuous movements
in underlying measures may arise from monetary policy regime changes. The
model is defined as follows:
pffiffiffiffiffiffiffiffiffi
dXðtÞ ¼ kx0 ðyðtÞ
XðtÞÞdt þ V ðtÞdW 1 ðtÞ þ J u dqu
J d dqd ,
 pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kv0 v0
V ðtÞÞdt þ sv0 V ðtÞdW 2 ðtÞ, (56:6)
 pffiffiffiffiffiffiffiffi
dyðtÞ ¼ ky0 y0
yðtÞÞdt þ sy0 yðtÞdW 3 ðtÞ

where all parameters are similar as in Eq. 56.5; W1(t), W2(t), and W3(t) are
independent Brownian motions; and qu and qd are Poisson processes
with jump intensities lu0 and ld0, respectively, and are independent of the
Brownian motions Wr(t), Wv(t), and Wy(t). In particular, lu0 is the
56 Density and Conditional Distribution-Based Specification Analysis 1519

probability of a jump-up, Pr(dqu (t) ¼ 1) ¼ lu0, and ld0 is the probability of a jump-
down, Pr(dqd(t) ¼ 1) ¼ ld0. Ju and Jd are  jump-up
 and jump-down sizes and have
exponential distributions f ðJ u Þ ¼ B1u0 exp
BJu0u and f ðJ d Þ ¼ B1d0 exp
BJd0d , where
Bu0, Bd0 > 0 are the jump magnitudes, which are the means of the jumps Ju and Jd.

56.2.2 Overview on Specification Tests and Model Selection

The focus in this chapter is specification testing and model selection. The “tools”
used in this literature have been long established. Several key classical contribu-
tions include the Kolmogorov-Smirnov test (see, e.g., Kolmogorov (1933) and
Smirnov (1939)), various results on empirical processes (see, e.g., Andrews
(1993) and the discussion in Chap. 19 of van der Vaart (1998) on the contributions
of Glivenko, Cantelli, Doob, Donsker, and others), the probability integral trans-
form (see, e.g., Rosenblatt (1952)), and the Kullback–Leibler information criterion
(see, e.g., White (1982) and Vuong (1989)). For illustration, the empirical distri-
bution mentioned above is crucial in our discussion of predictive densities because
it is useful in estimation, testing, and model evaluation. Let Yt is a variable of
interest with distribution F and parameter y0. The theory of empirical distributions
provides a result that
1 X T
pffiffiffi ð1fY t  ug
Fðujy0 ÞÞ
T t¼1

satisfies a central limit theorem (with a parametric rate) if T is large (i.e., asymp-
totically). In the above expression, 1{Yt  u} is the indicator function which takes
value 1 if Yt  u and 0 otherwise. In the case where there is parameter estimation
error, we can use more general results in Chap. 19 of van der Vaart (1998). Define
ð
1X T
PT ð f Þ ¼ f ðY i Þ and Pð f Þ ¼ fdP,
T i¼1

where P is a probability measure associated with F. Here, Pn( f ) converges to P( f )

almost surely for all the measurable functions f for which P( f ) is defined. Suppose
one wants to test the null hypothesis that P belongs to a certain family
fPy0 : y0 2 Yg , where y0 is unknown; it is natural to use a measure of the
discrepancy between Pn and Py^ for a reasonable estimator y^t of y0. In particular,
1  
if y^t converges to y0 at a root-T rate, pffiffiffi PT
Pyt^ has been shown to satisfy
T
7
a central limit theorem.
With regard to dynamic misspecification and parameter estimation error, the
approach discussed for the class of tests in this chapter allows for the construction

7
See Theorem 19.23 in van der Vaart (1998) for details.
1520 D. Duong and N.R. Swanson

of statistics that admit for dynamic misspecification under both hypotheses. This
differs from other classes of tests such as the framework used by Diebold
et al. (1998), Hong (2001), and Bai (2003) in which correction dynamic specifica-
tion under the null hypothesis is assumed. In particular, DGT use the probability
ð Yt
integral transform to show that Ft ðY t jℑt
1 ; y0 Þ ¼ f t ðyjℑt
1 ; y0 Þdy is identically

1
and independently distributed as a uniform random variable on [0; 1], where Ft(·)
and ft(·) are a parametric distribution and density with underlying parameter y0, Yt is
again our random variable of interest, and ℑt is the information set containing all
“relevant” past information.  They
thus suggest using the difference between the

empirical distribution of Ft Y t ℑt
1 ; y^t : and the 45 line as a measure of “goodness
of fit,” where y^t is some estimator of y0. This approach has been shown to be very
useful for financial risk management (see, e.g., Diebold et al. (1999)), as well as for
macroeconomic forecasting (see, e.g., Diebold et al. (1998) and Clements and
Smith (2000, 2002)). Similarly, Bai (2003) develops aKolmogorov-type
 test of
^
Ft(Yt|ℑt
1, y0) on the basis of the discrepancy between Ft Y t ℑt
1 ; y t : and the CDF
of a uniform on [0; 1]. As the test involves estimator y^t , the limiting distribution
reflects the contribution of parameter estimation error and is not nuisance
parameter-free. To overcome this problem, Bai (2003) proposes a novel approach
based on a martingalization argument to construct a modified Kolmogorov test
which has a nuisance parameter-free limiting distribution. This test has power
against violations of uniformity but not against violations of independence. Hong
(2001) proposes another related interesting test, based on the generalized spectrum,
which has power against both uniformity and independence violations, for the case
in which the contribution of parameter estimation error vanishes asymptotically. If
the null is rejected, Hong (2001) also proposes a test for uniformity robust to
nonindependence, which is based on the comparison between a kernel density
estimator and the uniform density. Two features differentiate the tests surveyed in
this chapter from the tests outlined in the other papers mentioned above. First, the
tests discussed here assume strict stationarity. Second, they allow for dynamic
misspecification under the null hypothesis. The second feature allows us to obtain
asymptotically valid critical values even when the conditioning information set
does not contain all of the relevant past history. More precisely, assume that we are
interested in testing for correct specification, given a particular information set
which may or may not contain all of the relevant past information. This is important
when a Kolmogorov test is constructed, as one is generally faced with the problem
of defining ℑt
1. If enough history is not included, then there may be dynamic
misspecification. Additionally, finding out how much information (e.g., how many
lags) to include may involve pre-testing, hence leading to a form of sequential test
bias. By allowing for dynamic misspecification, such pre-testing is not required.
Also note that critical values derived under correct specification given ℑt
1 are not
in general valid in the case of correct specification given a subset of ℑt
1. Consider
the following example. Assume that we are interested in testing whether
56 Density and Conditional Distribution-Based Specification Analysis 1521

the conditional distribution of Yt|Yt
1 follows normal distribution N(a1Yt
1, s1).

Suppose also that in actual fact the “relevant” information set has ℑt
1
including both Yt
1 and Yt
2, so that the true conditional model is
Ytjℑt
1 ¼ YtjYt
1, Yt
2 ¼ N(a1Yt
1 + a2Yt
2, s2). In this case, correct specification
holds with respect to the information contained in Xt
1; but there is dynamic misspeci-
fication with respect to Yt
1 and Yt
2. Even without taking account of parameter
estimation error, the critical values obtained assuming correct dynamic specification
are invalid, thus leading to invalid inference. Stated differently, tests that are designed
to have power against both uniformity and independence violations (i.e., tests that
assume correct dynamic specification under the null) will reject an inference which is
incorrect, at least in the sense that the “normality” assumption is not false. In summary,
if one is interested in the particular problem of testing for correct specification for
a given information set, then the approach of tests in this chapter is appropriate.

56.3 Consistent Distribution-Based Specification Tests

and Predictive Density-Type Model Selection
for Diffusion Processes

56.3.1 One-Factor Models

In this section, we outline the setup for the general class of one-factor jump
diffusion specifications. All analyses carry through to the more complicated case
of multifactor stochastic volatility models which we will elaborate upon in the next
subsection. In the presentation of the tests, we follow a view that all candidate
models, either single or multiple dimensional ones, are approximations of reality
and can thus be misspecified. The issue of correct specification (or misspecification)
of a single model and the model selection test for choosing amongst multiple
competing models allow for this feature.
To begin, fix the time interval [0, T] and consider a given single one-factor
candidate model the same as Eq. 56.1, with the true parameters y0, l0, m0 to be
replaced by its pseudo true analogs y{, l, m, respectively, and 0  t  T:
ðt ð ðt X
 {
   Jt
Xðt
Þ ¼ b Xðs
Þ, y ds
lt yfðyÞdy þ s Xðs
Þ, y{ dW ðsÞ þ yj ,
0 Y 0 j¼1
or
   
dXðt
Þ ¼ b Xðt
Þ, y{
lm dt: ð
  (56:7)
þ s Xðt
Þ, y{ dW ðtÞ þ ypðdy; dtÞ,
Y

where variables are defined the same as in Eqs. 56.1 and 56.2. Note that as the above
model is the one-factor version of Eqs. 56.1 and 56.2 where the dimension of X(t
)
is 1  1, W(t) is a one-dimensional standard Brownian motion and jump size, and yj
1522 D. Duong and N.R. Swanson

is one-dimensional variable for all j. Also note both Jt and yj are assumed to be
independent of the driving Brownian motion.
If the single model is correctly specified, then b(X(t
), y{) ¼ b0(X(t
), y0),
s(X(t
), y{) ¼ s0(X(t
), y0), l ¼ l0, m ¼ m0, and f ¼ f0 where b0(X(t
), y0),
s0(X(t
), y0), l0, m0, f0 are unknown and belong to the true specification.
Now consider a different case (not a single model) where m candidate models are
involved. For model k with 1  k  m, denote its corresponding specification to be
(bk(X(t
), y{k ), sk(X(t
), y{k ), lk, mk, fk). Two scenarios immediate arise. Firstly, if
the model k is correctly specified, then bk(X(t
), y{k ) ¼ b0(X(t
), y0), sk(X(t
),
y{k ) ¼ s0(X(t
), y0), lk ¼ l0, mk ¼ m0, and fk ¼ f0 which are similar to the case of
a single model. In the second scenario, all the models are likely to be misspecified
and modelers are faced with the choice of selecting the “best” one. This type of
problem is well fitted into the class of accuracy assessment tests initiated earlier by
Diebold and Mariano (1995) or White (2000).
The tests discussed hereafter are Kolmogorov-type tests based on the construc-
tion of cumulative distribution functions (CDFs). In a few cases, the CDF is known
in closed form. For instance, for the simplified version of the CIR model as in
Eq. 56.3, X(t) belongs to the linear exponential (or Pearson) family with the gamma
CDF of the form8
ð u  
2ð1
a=lÞ
1
l   
exp
x= l2 dx
2
Fðu; a; lÞ ¼ 0 (56:8)
Gð2ð1
a=lÞÞ
Ð1
where G(x) ¼ 0 txexp(
t)dt, and a, l are constants.
Furthermore, if we look at the pure diffusion process without jumps
   
dXðtÞ ¼ b XðtÞ, y{ dt þ s XðtÞ, y{ dW ðtÞ, (56:9)

where b(·) and s ¼ s(·) are drift and volatility functions, it is known that the
stationary density, say f(x, y{), associated with the invariant probability measure
can be expressed explicitly as9
  ðx  { !
 { c y{ 2b u; y
f x; y ¼  {  exp   du ,
s x; y
2 s2 u; y{
{ {
Ð u c(y{) is a constant ensuring that f integrates to one. The CDF, say F(u,y )
where
¼ f(x,y ) dx, can then be obtained using available numerical integration
procedures.
However, in most cases, it is impossible to derive the CDFs in closed form. To
obtain a CDF in such cases, a more general approach is to use simulation. Instead of

8
See Wong (1964) for details.
9
See Karlin and Taylor (1981) for details.
56 Density and Conditional Distribution-Based Specification Analysis 1523

estimating the CDF directly, simulation techniques estimate the CDF indirectly
utilizing its generated sample paths and the theory of empirical distributions. The
specification of a specific diffusion process will dictate the sample paths and
thereby corresponding test outcomes.
Note that in the historical context, many early papers in this literature are probability
density based. For example, in a seminal paper, Ait-Sahalia (1996) compares the
marginal densities implied by hypothesized null models with nonparametric estimates
thereof. Following the same framework of correct specification tests, Corradi and
Swanson (2005) and Bhardwaj et al. (2008), however, do not look at densities. Instead,
they compare the cumulative distribution (marginal or joint) implied by a hypothesized
null model with the corresponding empirical distribution. While Corradi and Swanson
(2005) focus on the known unconditional distribution, Bhardwaj et al. (2008) look at
the conditional simulated distributions. Corradi and Swanson (2011) make extensions
to multiple models in the context of out-of-sample accuracy assessment tests. This
approach is somewhat novel to this continuous time model testing literature.
Now suppose we observe a discrete sample path X1, X2, . . ., XT (also referred as
skeletons).10 The corresponding hypotheses can be set up as follows:
Hypothesis 1 Unconditional Distribution Specification Test of a Single Model
H0 : F(u,y{) ¼ F0(u,y0), for all u, a.s.
HA : Pr(F(u,y{)
F0(u,y0) 6¼ 0) > 0, for some u 2 U, with nonzero Lebesgue
measure.
where F0(u, y0) is the true cumulativedistribution
 implied by the above density,
 { y{
i.e., F0(u,y0) ¼ Pr(Xt  u). F u; y ¼ Pr Xt  u is the cumulative distribution of
{
the proposed model. Xyt is a skeleton implied by model (56.7).
Hypothesis 2 Conditional Distribution Specification Test of a Single Model
H0 : Ft(u|Xt, y{) ¼ F0,t(u|Xt, y0), for all u, a.s.
HA : Pr(Ft(u|Xt, y{)
F0,t(u|Xt, y0) 6¼ 0) > 0, for some u 2 U, with nonzero
Lebesgue measure.
   { { 

where Ft ujXt ; y{ ¼ Pr Xytþt  u Xyt ¼ Xt is t-step ahead conditional
distributions and t ¼ 1, . . . , T
t. F0,t(u|Xt, y0) is t-step ahead true conditional
distributions.
Hypothesis 3 Predictive Density Test for Choosing Amongst Multiple
Competing Models
The null hypothesis is that no model can outperform model 1 which is the
benchmark model.11

10
As mentioned earlier, we follow Corradi and Swanson (2005) by using notation X(·) when
defining continuous time processes and Xt for a skeleton.
11
See White (2000) for a discussion of a discrete time series analog to this case, whereby point
rather than density-based loss is considered; Corradi and Swanson (2007b) for an extension of
White (2000) that allows for parameter estimation error; and Corradi and Swanson (2006) for an
extension of Corradi and Swanson (2007b) that allows for the comparison of conditional distri-
butions and densities in a discrete time series context.
1524 D. Duong and N.R. Swanson

H0:
!
max EX F y
{ ð u2 Þ
F y
{ ð u1 Þ
k¼2, ..., m X1,1 tþt ðXt Þ X 1 ðXt Þ
!2 1, tþt

ðF0 ðu2 jXt Þ
F0 ðu1 jXt ÞÞ
!

EX F y
{ ðu2 Þ
F y
{ ð u1 Þ
Xk,k tþt ðXt Þ X k ðXt Þ
!2k, tþt

ðF0 ðu2 jXt Þ
F0 ðu1 jXt ÞÞ :

HA: negation of H00

where

   { 
yk y{
F { ð uÞ ¼ Ftk { t
u Xt ; yk ¼ Py{ Xk, tþt  u Xt ¼ Xt ,
k
y
Xk,k tþt ðXt Þ k

which is the conditional distribution of Xt+t, given Xt, and evaluated at u under the
y{
probability law generated by model k. Xk,k tþt ðXt Þ with 1  t  T
t is the
skeleton implied by model k, parameter y{k , and initial value Xt. Analogously, define
Ft0 ðujXt ; y0 Þ ¼ Pty0 ðXtþt  ujXt Þ to be the “true” conditional distribution.
Note that the three hypotheses expressed above apply exactly the same to
the case of multifactor diffusions. Now, before moving to the statistics
description section, we briefly explain the intuitions in facilitating construction of
the tests.
In the first case (Hypothesis 1), Corradi and Swanson (2005) construct a
Kolmogorov-type test based on comparison of the empirical distribution and the
unconditional CDF implied by the specification of the drift, variance, and jumps.
Specifically, one can look at the scaled difference between

 {  ðu
   
F u; y{ ¼ Pr Xyt  u ¼ f x; y{ dx

and estimator of the true F0(u|Xt, y0), the empirical distribution of Xt defined as

1X T
1fXt  ug,
T t¼1

where 1{Yt  u} is indicator function which takes value 1 if Yt  u and 0 otherwise.

56 Density and Conditional Distribution-Based Specification Analysis 1525

Similarly for the second case of conditional distribution (Hypothesis 2), the test
statistic VT can be a measure of the distance between the t-step ahead conditional
{ {
distribution of Xytþt , given Xyt ¼ Xt , as
   { { 

Ft u Xt ; y{ ¼ Pr Xytþt  u Xyt ¼ Xt ,

to an estimator of the true F0,t(u|Xt,y0), the conditional empirical distribution of Xt+t

conditional on the initial value Xt defined as

1 X T
t
1fXtþt  ug,
T
t t¼1

In the third case (Hypothesis 3), model accuracy is measured in terms of a

distributional analog of mean square error. As is commonplace in the out-of-sample
evaluation literature, the sample of T observations is divided into two subsamples,
such that T ¼ R + P, where only the last P observations are used for predictive
evaluation. A t-step ahead prediction error under model k is 1{u1  Xt+t  u2}

(Ftk(u2|Xt, y{k )
Ftk(u1|Xt, y{k )) where 2  k  m and similarly for model 1 by
replacing index k with index 1. Suppose we can simulate P
t paths of t-step ahead
skeleton12 using Xt as starting values where t ¼ R, . . ., R + P
t, from which we
can construct a sample of P
t prediction errors. Then, these prediction errors can
be used to construct a test statistic for model comparison. In particular, model 1 is
defined to be more accurate than model k if
0      !2 1
{ {
Ft1 t
u2 X t ; y1
F1 u1 X t ; y1
F@   A

Ft1 ðu2 jXt ; y0 Þ
Ft1 ðu1 jXt ; y0 Þ
0      !2 1
t t { t t {
Fk u 2 X t ; y 1
Fk u 1 X t ; y k
< E@   A,

Ft0 ðu2 jXt ; y0 Þ
Ft0 ðu1 jXt ; y0 Þ

where E(·) is an expectation operator and E(1{u1  Xt+t  u2}jXt)

¼ Ft0(u2jXt,y0)
Ft0(u1jXt,y0). Concretely, model k is worse than model 1 if on
average t-step ahead prediction errors under model k is larger than that of model 1.
Finally, it is important to point out some main features characterized by all the
three test statistics. Processes X(t) hereafter are required to satisfy the regular
conditions, i.e., assumptions A1–A8 in Corradi and Swanson (2011). Regarding
{
model estimation (in Sect. 56.3.3), y{ and yk are unobserved and need to be
estimated. While Corradi and Swanson (2005) and Bhardwaj et al. (2008) utilize
(recursive) simulated generalized method of moments (SGMM), Corradi and
Swanson (2011) make extension to (recursive) nonparametric simulated

12
See Sect. 56.3.3.1 for model simulation details.
1526 D. Duong and N.R. Swanson

quasi-maximum likelihood (NPSQML). For the unknown distribution and condi-

tional distribution, it will be pointed out in Sect. 56.3.3.2 that F(u,y{), Ft(u|Xt, y{),
and F y{ ðuÞ can be replaced by their simulated counterparts using the (recursive)
Xk,k tþt ðXt Þ
SGMM and NPSQML parameter estimators. In addition, test statistics converge to
functional of Gaussian processes with covariance kernels that reflect time depen-
dence of the data and the contribution of parameter estimation error (PEE). Limiting
distributions are not nuisance parameter-free, and critical values thereby cannot be
tabulated by the standard approach. All the tests discussed in this chapter rely on the
bootstrap procedures for obtaining the asymptotically valid critical values, which we
will describe in Sect. 56.3.4.

56.3.1.1 Unconditional Distribution Tests

For one-factor diffusions, we outline the construction of unconditional test
statistics in the context where CDF is known in closed form. In order to test the
Hypothesis 1, consider the following statistic:
ð
V 2T , N, h ¼ V 2T , N, h ðuÞpðuÞ,
U

where

T   
1 X
V T , N, h ¼ pffiffiffi 1fXt  ug
F u; y^T , N , h :
T t¼1
ð
In the above expression, U is a compact interval and pðuÞdu ¼ 1, 1fXt  ug is
U
again the indicator function which returns value 1 if Xt  u and 0 otherwise. Further,
as defined in Sect. 56.3.3, y^T , N, h hereafter is a simulated estimator where T is sample
size and h is the discretization interval used in simulation. In addition, with the abuse
of notation, N is a generic notation throughout this chapter, i.e., N ¼ L, the length of
each simulation path for (recursive) SGMM, and N ¼ M, the number of random
draws (simulated paths) for (recursive) NPQML estimator.13 Also note in our nota-
tion that as the above test is in sample specification test, the estimator and the
statistics are constructed using the entire sample, i.e., y^T , N, h .
It has been shown in Corradi and Swanson (2005) that under regular conditions
and if the estimator is estimated by SGMM, the above statistics converges
to a functional of Gaussian process.14 In particular, pick the choice T, N ! 1,
h ! 0, T/N ! 0, and Th2 ! 0.

13
M is often chosen to coincide with S, the number of simulated paths used when simulating
distributions.
14
For details and the proof, see Theorem 1 in Corradi and Swanson (2005).
56 Density and Conditional Distribution-Based Specification Analysis 1527

Under the null,

ð
V 2T , N, h ! Z 2 ðuÞpðuÞ,
U

where Z is a Gaussian process with covariance kernel. Hence, the limiting distri-
2
bution of VT,N,h is a functional of a Gaussian process with a covariance kernel that
reflects both PEE and the time series nature of the data. As y^T , N, h is root-T
consistent, PEE does not disappear in the asymptotic covariance kernel.
Under HA, there exists an e > 0 such that
 
1 2
lim Pr V > e ¼ 1:
T!1 T T, N, h

For the asymptotic critical value tabulation, we use the bootstrap procedure. In
order to establish validity of the block bootstrap under SGMM with the presence of
PEE, the simulated sample size should be chosen to grow at a faster rate than the
historical sample, i.e., T/N ! 0.
Thus, we can follow the steps in appropriate bootstrap procedure in Sect. 56.3.4.
For instance, if the SGMM estimator is used, the bootstrap statistic is
ð
V T , N, h ¼ V 2
2
T , N , h ðuÞpðuÞdu,
U

where
T 

1 X  
V 2
T, N, h ¼ p ffiffiffi 1 Xt  u
1fXt  ug
T t¼1
    

^ ^

F u; y T , N, h
F u; y T , N, h :
 
In the above expression, y^T , N, h is the bootstrap analog of y^T , N, h and is estimated
by the bootstrap sample X1, . . ., XT (see Sect. 56.3.4). With appropriate conditions,
* *
2*
Corradi and Swanson (2005) show that under the null, VT,N,h has a well-defined
2
limiting distribution which coincides with that of VT,N,h. We then can straightfor-
wardly derive the bootstrap critical value by following Steps 1–5 in Sect. 56.3.4. In
particular, in Step 5, the idea is to perform B bootstrap replications (B large) and
compute the percentiles of the empirical distribution of the B bootstrap statistics.
Reject H0 if V2T,N,h is greater than the (1
a)th percentile of this empirical distribu-
tion. Otherwise, do not reject H0.

56.3.1.2 Conditional Distribution Tests

Hypothesis 2 tests correct specification of the conditional distribution, implied by
a proposed diffusion model. In practice, the difficulty arises from the fact that the
functional form of neither t-step ahead conditional distributions Ft(u|Xt, y{) nor
F0,t(u|Xt, y0) is unknown in most cases. Therefore, Bhardwaj et al. (2008) develop
1528 D. Duong and N.R. Swanson

bootstrap specification test on the basis of simulated distribution using the

SGMM estimator.15 With the important inputs leading to the test such as simulated
estimator, distribution simulation, and bootstrap procedures to be presented in the
next section,16 the test statistic is defined as
ZT ¼ sup jZT ðu; vÞj,
uv2UV
where
0 X n ^ o1
1
1 X T
t S y T, N, h
@S 1 X ,  u A,
ZT ðu; vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi s¼1 s tþt
T
t t¼1
ð1fX  ugÞ tþt

with U and V compact sets on the real line. y^T , N, h is the simulated estimator using
entire sample X1, . . . , XT, and S is the number of simulated replications used in the
estimation of conditional distributions as described in Sect. 56.3.3. If SGMM estimator
is used (similar to unconditional distribution case and the same as in Bhardwaj et al.
(2008)), then N ¼ L, where L is the simulation length used in parameter estimation.
The above statistic is a simulation-based version of the conditional Kolmogorov
test of Andrews (1997), which compare the joint empirical distribution:

1 X T
t
pffiffiffiffiffiffiffiffiffiffiffi 1fXtþt  ug1fXt  vg,
T
t t¼1

with its semi-empirical/semi-parametric analog given by the product of

1 X T
t
pffiffiffiffiffiffiffiffiffiffiffi F0, t ðujXt ; y0 Þ1fXt  vg:
T
t t¼1

Intuitively, if the null is not rejected, the metric distance between the two should
asymptotically disappear. In the simulation context with parameter estimation
error, the asymptotic limit of ZT however is a nontrivial one. Bhardwaj et al.
(2008) show that with the proper choice of T, N, S, h, i.e., T, N, S, T2/S ! 1
and h, T/N, T/S, Nh, h2T ! 0, then
d
ZT ! sup jZ ðu; vÞj,
uv2UV

where Z(u, v) is a Gaussian process with a covariance kernel that characterizes

(1) long-run variance we would have if we knew F0,t(u|Xt, y0)), (2) the contribution

15
In this chapter, we assume that X(·) satisfies the regularity conditions stated in Corradi and
Swanson (2011), i.e., assumptions A1–A8. Those conditions also reflect requirements A1–A2 in
Bhardwaj et al. (2008). Note that the SGMM estimator used in Bhardwaj et al. (2008) satisfies the
root-N consistency condition that Corradi and Swanson (2011) impose on their parameter estima-
tor (see Assumption 4).
16
See Sects. 56.3.3 and 56.3.4 for further details.
56 Density and Conditional Distribution-Based Specification Analysis 1529

of parameter estimation error, and (3) The correlation between the first two.
Furthermore, under HA, there exists some e > 0 such that
 
1
lim Pr pffiffiffi ZT > e ¼ 1:
P!1 T
As T/S ! 0, the contribution of simulation error is asymptotically negligible. The
limiting distribution is not nuisance parameter-free and hence critical values cannot
be tabulated directly from it. The appropriate bootstrap statistic in this context is

ZT ¼ sup ZT ðu; vÞ ,
uv2UV

where

1 X T
t

Z T ¼ ðu; vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1 Xt  v
T
t t¼1
 !
1X S 
y^T , N, h
n o

 1 Xs, tþt  u
1 Xtþt  u
S s¼1
1 X T
t

pffiffiffiffiffiffiffiffiffiffiffi 1fXt  vg
T
t t¼1
!
S  y^
1X T, N, h
 1 Xs, tþt  u
1fXtþt  ug :
S s¼1

In the above expression, y^ T , N, h is the bootstrap parameter estimated using the

y^
resampled data X*t , N, h , s ¼ 1, . . . , S and t ¼ 1, . . . , T
t is
for t ¼ 1, . . . , T
t. Xs,Ttþt

the simulated date under y^T, N, h and X*t ; t ¼ 1, . . ., T
t is a resampled series
constructed using standard block bootstrap methods as described in Sect. 56.3.4.
Note that in the original paper, Bhardwaj et al. (2008) propose bootstrap SGMM
estimator for conditional distribution of diffusion processes. Corradi and Swanson
(2011) extend the test to the case of simulated recursive NPSQML estimator.
Regarding the generation of the empirical distribution of Z*T (asthmatically the
same as ZT), follow Steps 1–5 in the bootstrap procedure in Sect. 56.3.4. This
yields B bootstrap replications (B large) of Z*T. One can then compare ZT with the
percentiles of the empirical distribution of Z*T, and reject H0 if ZT is greater than the
(1
a)th percentile. Otherwise, do not reject H0. Tests carried out in this manner
are correctly asymptotically sized and have unit asymptotic power.

56.3.1.3 Predictive Density Tests for Multiple Competing Models

In many circumstances, one might want to compare one (benchmark) model
(model 1) against multiple competing models (models k, 2  k  m). In this case,
recall in the null in Hypothesis 3 that no model can outperform the benchmark
model. In testing the null, we first choose a particular interval, i.e., (u1, u2) 2 U  U,
1530 D. Duong and N.R. Swanson

where U is a compact set so that the objective is evaluation of predictive densities for
a given range of values. In addition, in the recursive setting (not full sample is used to
estimate parameters), if we use the recursive NPSQML estimator, say y^1:t, N , h and
y^k:t, N, h , for models 1 and k, respectively, then the test statistic is defined as

k, P, S ðu1 ; u2 Þ ¼ max Dk, P, S ðu1 ; u2 Þ,

DMax
k¼2, ..., m

where
"
S n o
1 X T
t
1X y^ N, h
Dk, P, S ðu1 ; u2 Þ ¼ pffiffiffi 1 u1  X1,1,i,t,tþt ð X t Þ  u2
P t¼R S i¼1
#2

1fu1  Xtþt  u2 g
"
S n o
1X y^ N, h

1 u1  Xk,k,i,t,tþt ðXt Þ  u2
S i¼1
#2 1

fu1  Xtþt  u2 g A:

All notations are consistent with previous sections where S is the

number of simulated replications used in the estimation of conditional distributions.
y^ N, h y^ N, h
X1,1,i,t,tþt ðXt Þ and Xk,k,i,t,tþt , i ¼ 1, . . ., S, t ¼ 1, . . ., T
t, are the ith simulated path
^ ^
under y 1, t, N, h and y k, t, N, h . If models 1 and k are nonnested for at least one, k ¼
2, . . ., m. Under regular conditions and if P, R, S, h are chosen such as P, R, N ! 1
and h, P/N, h2P ! 0, P/R ! p where p is finite then
 
max Dk, P, N ðu1 ; u2 Þ
mk ðu1 ; u2 Þ ! max Zk ðu1 ; u2 Þ:
k¼2, ..., m k¼2, ..., m

where, with an abuse of notation, mk(u1, u2) ¼ m1(u1, u2)
mk(u1, u2), and
00 ! 12 1
B B F y{ ð u 2 Þ
F y{ ð u1 Þ C C
mj ðu1 ; u2 Þ ¼ E@@ Xj,j tþt ðXt Þ Xj,j tþt ðXt Þ A A,

ðF0 ðu2 jXt Þ
F0 ðu1 jXt ÞÞ

for j ¼ 1,. . ., m, and where (Z1(u1, u2), . . ., Zm(u1, u2)) is an m-dimensional Gaussian
random variable the covariance kernels that involves error in parameter estimation.
Bootstrap statistics are thereforerequired to reflect this parameterestimation errorissue.17
In the implementation, we can obtain the asymptotic critical value using a
recursive version of the block bootstrap. The idea is that when forming block

17
See Corradi and Swanson (2011) for further discussion.
56 Density and Conditional Distribution-Based Specification Analysis 1531

bootstrap samples in the recursive setting, observations at the beginning of the

sample are used more frequently than observations at the end of the sample. We can
replicate Steps 1–5 in bootstrap procedure in Sect. 56.3.4. It should be stressed
that the resampling in the Step 1 is the recursive one. Specifically, begin
by resampling b blocks of length l from the full sample, with lb ¼ T. For any
given t, it is necessary to jointly resample Xt, Xt+1, . . ., Xt+t. More precisely, let
Zt,t ¼ (Xt, Xt+1, . . ., Xt + t), t ¼ 1, . . .,T
t. Now, resample
 b overlapping blocks of
t,    
length l from Z . This yields Z ¼ X ; X ; . . . ; X , t ¼ 1,. . ., T
t. Use these
t,t
t tþ1 tþt

data to construct bootstrap estimator y^k, t, N, h. Recall that N is chosen in Corradi and
Swanson (2011) as the number of simulated series used to estimate the parameters
(N ¼ M ¼ S) and such as N/R, N/P ! 1. Under this condition, simulation error
vanishes and there is no need to resample the simulated series.
Corradi and Swanson (2011) show that
T 
1 X 
pffiffiffi y^k, t, N, h
y^k, t, N, h Þ
P t¼R
has the same limiting distribution as
T 
1 X
pffiffiffi y^k, t, N, h
y{k Þ,
P t¼R
conditional on all samples except a set with probability measure approaching zero.
Given this, the appropriate bootstrap statistic is
( "
S n o
1 X T
t
1X y^ N, h   
Dk, P, S ðu1 ; u2 Þ ¼ pffiffiffi 1 u1  X1,1,i,t,tþt X t  u2
P t¼R S i¼1
#2




1 u1  Xtþt  u2
"
S n   o
1XT
1X y^ N , h

1 u1  X1,1,i,t,tþt Xj  u2
T j¼1 S i¼1
#2



1 u1  Xjþt  u2
"
S n o
1X y^ N , h   

1 u1  Xk,k,i,t,tþt Xt  u2
S i¼1
#2



1 u1  Xtþt  u2
"
S n o
1X S
1X y^ N, h  

1 u1  Xk,k,i,t,tþt X j  u2
S i¼1 S i¼1
# 2 119

 =

1 u1  Xjþt  u2 AA :
;
1532 D. Duong and N.R. Swanson

As the bootstrap statistic is calculated from the last P resampled observations, it

is necessary to have each bootstrap term recentered around the (full) sample mean.
This is true even in the case there is no need to mimic PEE, i.e., the choice of P, R is
such that P/R ! 0. In such a case, above statistic can be formed using y^k, t, N, h rather

than y^k, t, N, h .
For any bootstrap replication, repeat B times (B large) bootstrap replications
*
which yield B bootstrap statistics Dk,P,S . Reject H0 if Dk,P,S is greater than the
(1
a)th percentile of the bootstrap empirical distribution. For numerical imple-
mentation, it is of importance to note that

  in thecase where P/R !0, P, T, R  ! 1,
^ ^ ^ ^
there is no need to re-estimate y 1, t, N, h y k, t, N , h . Namely, y 1, t, N, h y k, t, N, h can be
used in all bootstrap experiments.
Of course, the above framework can also be applied using entire simulated
distributions rather than predictive densities, by simply estimating parameters
once, using the entire sample, as opposed to using recursive estimation
techniques, say, as when forming predictions and associated predictive
densities.

56.3.2 Multifactor Models

Now, let us turn our attention to multifactor diffusion models of the form
(X(t), V(t))0 ¼ (X(t), V1(t), . . ., Vd(t))0 , where only the first element, the diffusion
process Xt, is observed while V(t) ¼ (V1(t), . . ., Vd(t))0 is latent. The most
popular class of the multifactor models is stochastic volatility model expressed as
below:

   !  !  !
dXðtÞ b1 XðtÞ,y{ s11 V ðtÞ,y{ s12 V ðtÞ,y{
¼   dt þ dW 1 ðtÞ þ   dW 2 ðtÞ,
dV ðtÞ b2 V ðtÞ,y{ 0 s22 V ðtÞ,y{
(56:10)

where W1(t)11 and W2(t)11 are independent Brownian motions.18 For instance,
many term structure models require the multifactor specification of the above form
(see Dai and Singleton (2000)). In a more complicated case, the drift function can
also be specified to be a stochastic process which poses even more challenges to
testing. As mentioned earlier, the hypotheses (Hypothesis 1, 2, 3) and the test

18
Note that the dimension of X(·) can be higher and we can add jumps to the above specification
such that it satisfies the regularity conditions outlined in the one-factor case. In addition, Corradi
and Swanson (2005) provide a detailed discussion of approximation schemes in the context of
stochastic volatility models.
56 Density and Conditional Distribution-Based Specification Analysis 1533

construction strategy for multifactor models are the same as for one-factor model.
All theory essentially applies immediately to multifactor cases. In implementation,
the key difference is in the simulated approximation scheme facilitating parameter
and CDF estimation. X(t) cannot simply be expressed as a function
Ðt of d + 1 driving
Brownian motions but instead involves a function of (Wjt, 0 WjsdWis), i, j ¼ 1, .. . ,
d + 1 (see, e.g., Pardoux and Talay (1985, pp. 30–32) and Corradi and Swanson
(2005)).
For illustration, we hereafter focus on the analysis of a stochastic volatility
model (56.10) where drift and diffusion coefficients can be written as
  !
b1 XðtÞ, y{
b¼  
b2 V ðtÞ, y{
   !
s11 V ðtÞ, y{ s12 V ðtÞ, y{
s¼  
0 s22 V ðtÞ, y{

We also examine a three-factor model (i.e., the Chen model as in Eq. 56.5) and
a three-factor model with jumps (i.e., CHENJ as in Eq. 56.6). By presenting two-
and three-factor models as an extension of our above discussion, we make it clear
that specification tests of multiple factor diffusions with d  3 can be easily
constructed in similar manner.
In distribution estimation, the important challenge for multifactor models lies in
the missing variable issue. In particular, for simulation of Xt, one needs initial
values of the latent processes V1, . . . , Vd, which are unobserved. To overcome this
problem, it suffices to simulate the process using different random initial values for
the volatility process; then construct the simulated distribution using those initial
values and average them out. This allows one to integrate out the effect of
a particular choice of volatility initial value.
For clarity of exposition, we sketch out a simulation strategy for a general model
of d latent variables in Sect. 56.3.3. This generalizes the simulation scheme of
three-factor models in Cai and Swanson (2011). As a final remark before moving to
the statistic presentation, note that the class of multifactor diffusion processes
considered in this chapter is required to match the regular conditions as in previous
section (assumption from A1 to A8 in Corradi and Swanson (2011) with A4 being
replaced by A40 ).

56.3.2.1 Unconditional Distribution Tests

Following the above discussion on test construction, we specialize to the case of
two-factor stochastic volatility models. Extension to general multidimensional and
multifactor models follows similarly. As the CDF is rarely known in closed form for
stochastic volatility models, we rely on simulation technique. With the simulation
scheme, estimators, simulated distribution, and bootstrap procedures to be presented
in the next sections (see Sects. 56.3.3 and 56.3.4), the test statistics for Hypothesis 1
turns out to be
1534 D. Duong and N.R. Swanson

T  S  
1 X 1X y^
SV T , S, h ¼ pffiffiffi 1fX t  ug
1 Xt,Th, N, L, h  u :
T t¼1 S t¼1

In the above expression, recall that S is the number of simulation paths used in
distribution simulation; y^T , N, L, h is a simulated estimator (see Sect. 56.3.3). N is
a generic notation throughout this chapter, i.e., N ¼ L, the length of each simulation
path for SGMM, and N ¼ M, the number of random draws (simulated paths) for
NPQML estimator. h is the discretization interval used in simulation. Note that
y^T , N, L, h is chosen in Corradi and Swanson (2005) to be SGMM estimator using
full sample and therefore N ¼ L ¼ S.19 To put it simply, one can write
y^T , S, h ¼ y^T , N, L, h .
Under the null, choose T, S to satisfy T, S ! 1, Sh ! 0, T/S ! 0, then
ð
SV 2T , S, h ! SV 2 ðuÞpðuÞ,
U

where Z is a Gaussian process with covariance kernel that reflects both PEE and the
time-dependent nature of the data. The relevant bootstrap statistic is

1 X T 
 
SV 2
T , S, h ¼ p ffiffiffi 1 Xt  u
1fXt  ug
T t¼1
S     ^ 
1 X T
1X y^ y

pffiffiffi 1 Xt,Th, N, L, h  u
1 Xt,Th, N, L, h  u ,
T t¼1 S t¼1

where y^T, S, h is the bootstrap analog of y^T , S, h . Repeat Steps 1–5 in the bootstrap
procedure in Sect. 56.3.4 to obtain critical values which are the percentiles of
the empirical distribution of Z*T. Compare SVT,S,h with the percentiles of the
empirical distribution of the bootstrap statistic and reject H0 if SVT,S,h is greater
than the (1
a)th percentile thereof. Otherwise, do not reject H0.

56.3.2.2 Conditional Distribution Tests

To test Hypothesis 2 for the multifactor models, first we present the test statistic
for the case of the stochastic volatility model (Xt, Vt) in Eq. 56.10 (i.e., for
two-factor diffusion), and then we discuss testing with the three-factor model
(Xt, V1t , V2t ) as in Eq. 56.5. Other multiple factor models can be tested analogously.

19
As seen in assumption A40 in Corradi and Swanson (2011) and Sect. 56.3.3 of this chapter,
^
y T , N, L, h can be other estimators such as the NPSQML estimator. Importantly, y^T , N, L, h satisfies
condition A40 in Corradi and Swanson (2011).
56 Density and Conditional Distribution-Based Specification Analysis 1535

Note that for illustration, we again assume use of the SGMM estimator y^T , N, L, h ,
as in the original work of Bhardwaj et al. (2008) (namely, y^T , N, L, h is the
simulated estimator described in Sect. 56.3.3). Specifically, N is chosen as
the length of sample path L used in parameter estimation. The associated test
statistic is

SZT ¼ sup jSZ T ðu, vÞj

uv2UV
1 X T
t
SZT ðu, vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S n o
1 X N X
y^T , N, L, h
 1 Xj, i, tþt  u
1fXtþt  ug ,
NS j¼1 i¼1

y^
where Xj,Ti,,Ntþt
, L, h is t-step ahead simulated skeleton obtained by simulation procedure
for multifactor model in Sect. 56.3.3.1.
In a similar manner, the bootstrap statistic analogous to SZT is


SZ T ¼ sup SZ ðu; vÞ ,
T
uv2UV
1 X
T
t

SZ T ðu; vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1 Xt  v
T
t t¼1
!
S  ^
1 X N X
y T , N, L, h
 
 1 Xj, i, tþt  u
1 Xtþt  u
NS j¼1 i¼1
1 X T
t

pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S n o
1 X N X
y^T , N, L, h
1 Xj, i, tþt  u
1fXtþt  ug
NS j¼1 i¼1


where y^T , N, L, h is the bootstrap estimator described in Sect. 56.3.4. For the three-
factor model, the test statistic is defined as

MZT ¼ sup jMZ T ðu; vÞj,

uv2UV
1 X T
t
MZT ðu; vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S n o
1 X L X L X
y^T , N , L, h
 1 Xs, tþt  u
1fXtþt  ug ,
L2 S j¼1 k¼1 i¼1

and bootstrap statistics is

1536 D. Duong and N.R. Swanson

1 X T
t
MZ T ðu; vÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S  ^
1 X L X L X
y t, N , L , h
 
 2 1 Xs, tþt  u
1 Xtþt  u
L S j¼1 k¼1 i¼1
1 X T
t

pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S n o
1 X L X L X
y^t, N, L, h
 2 1 Xs, tþt  u
1fXtþt  ug
L S j¼1 k¼1 i¼1

  
y^ N, L, h
y^ , N , L, h y^ , N, L, h 1, y^ 2, y^
where Xs,Ttþt ¼ Xs,Ttþt Xt ; V j T, N, L, h ; V k T, N, L, h and Xs,t,tþt ¼

  

y^ N , L, h 1, y^ 2, y^
Xs,t,tþt Xt ; V j t, N, L, h ; V k t, N, L, h .
The first-order asymptotic validity of inference carried out using bootstrap
statistics formed as outlined above follows immediately from Bhardwaj
et al. (2008). For testing decision, one compares the test statistics SZT,S,h and
MZT,S,h with the percentiles or the empirical distributions of SZ*T and MZT,S,h *
,
respectively. Then, reject H0 if the actual statistic is greater than the (1
a)th
percentile of the empirical distribution of the bootstrap statistic, as in Sect. 56.3.4.
Otherwise, do not reject H0.

56.3.2.3 Predictive Density Tests for Multiple Competing Models

For illustration, we present the test for the stochastic volatility model (two-factor
model). Again, note that extension to other multifactor models follows immediately.
In particular, all steps in the construction of the test in the one-factor model case carry
through immediately to the stochastic volatility case with the statistic defined as

DV P, L, S ¼ max DV k, P, L, S ðu1 ; u2 Þ,
k¼2, ..., m

where

1 X T
t
DV k, P, L, S ðu1 ; u2 Þ ¼ pffiffiffi
P t¼R
!2
S n   o
1 X L X
y^1, t, N , L, h y^1, t, N , L, h
1 u1  X1, tþt, i, j Xt ; V 1, j  u2
1fu1  Xtþt  u2 g
SL j¼1 i¼1
0 !2 1
XL X S n   o
1 y^k, t, N , L, h y^k, t, N, L, h

@ 1 u1  Xk, tþt, i, j Xt ; V k, j  u2
1fu1  Xtþt  u2 g A:
SL j¼1 i¼1
56 Density and Conditional Distribution-Based Specification Analysis 1537

Critical values for these tests can be obtained using a recursive version of the
block bootstrap. The corresponding bootstrap test statistic is

DV P, L, S ¼ max DV k, P, L, S ðu1 ; u2 Þ

k¼2, ..., m

where

1 X T
t
DV k, P, L, S ðu1 ;u2 Þ ¼ pffiffiffi
P t¼R
80" #2
< X L X S  
 



1 y^1, t, N , L, h y^1, t, N , L, h
@ 1 u1  X1, tþt, i, j Xt ;V 1, j  
 u2
1 u1  Xtþt  u2
: SL j¼1 i¼1
" #2 1
S n   o
1X T
1X L X
^
y , t , N , L, h ^
y 1, t, N , L, h

1 u1 < X1,1tþt , i, j Xl ;V 1, j  u2
1fu1  Xlþt  u2 g A
T l¼1 SL j¼1 i¼1
0" #2
L X
X S  

^



1 y^k, t, N , L, h y

@ 1 u1  Xk, tþt, i, j Xt ;V k, j k , t , N , L , h
 u2
1 u1  Xtþt  u2
SL j¼1 i¼1
" # 2 19
1 XT
1 XL X S n
^  ^  o =
y , t, N , L, h y k, t, N , L, h A

1 u1  Xk,ktþt X ;V
, i, j l k , j  u2
1fu 1  X lþt  u2 g
T l¼1 SL j¼1 i¼1 ;

Of note is that we follow Cai and Swanson (2011) by adopting the recursive
NPSQML estimator y^1, t, N, L, h and y^k, t, N, L, h for model 1 and k, respectively, as
 
introduced in Sect. 56.3.3.4 with the choice N ¼ M ¼ S. y^1, t, N , L, h and y^k, t, N, L, h are
bootstrap analogs of y^1, t, N, L, h and y^k, t, N, L, h , respectively (see Sect. 56.3.4). In
addition, we do not need to resample the volatility process, although volatility is

simulated under both y^k, t, N, L, h and y^k, t, N, L, h , k ¼ 1, . . ., m.
Repeat Steps 1–5 in the bootstrap procedure in Sect. 56.3.4 to obtain critical
values. Compare DVP, L, S with the percentiles of the empirical distribution of
L, S, and reject H0 if DVP, L, S is greater than the (1
a)th percentile.
*
DVP,
Otherwise, do not reject H0. Again, in implementation, there is no need to

re-estimate y^k, t, N, L, h for each bootstrap replication if P/R ! 0, P, T, R ! 1, as
parameter estimation error vanishes asymptotically in this case.

56.3.3 Model Simulation and Estimation

56.3.3.1 Simulating Data

Approximation schemes are used to obtain simulated distributions and simulated
parameter estimators, which are needed in order to construct the test statistics
outlined in previous sections. We therefore devote the first part of this section to
1538 D. Duong and N.R. Swanson

a discussion of the Milstein approximation schemes that have been used in

Corradi and Swanson (2005), Bhardwaj et al. (2008), and Corradi and Swanson
(2011). Let L be the length of each simulation path and h be the discretization
interval, L ¼ Qh, and y be a generic parameter in simulation expression. We
consider three cases:
The pure diffusion process as in Eq. 56.9:
    1    
Xyqh
Xyðq
1Þh ¼ b Xyðq
1Þh ; y h þ s Xðyq
1Þh ; y ϵ qh
s Xyðq
1Þh ; y 0 s Xyðq
1Þh ; y h
2
1  y  
0 y

þ s Xðq
1Þh ; y s Xðq
1Þh ; y ϵ qh ,2
2

where
  iid
W qh
W ðq
1Þh ¼ ϵ qh N ð0; hÞ,

q ¼ 1, . . . , Q, with ϵ qh iid N ð0; hÞ, and where s0 is the derivative of s(

) with respect
to its first argument. Hereafter, Xyqh denotes the values of the diffusion at time qh,
simulated under generic y, and with a discrete interval equal to h, and so is a
y
fine-grain analog of Xt,h .
The pure jump diffusion process without stochastic volatility as in Eq. 56.1:

    1  0  
Xðyqþ1Þh
Xyqh ¼ b Xyqh ; y h þ s Xyqh ; y ϵ ðqþ1Þh
s Xyqh ; y s Xyqh ; y h
2
1  0  
þ s Xyqh ; y s Xyqh ; y ϵ ð2qþ1Þh
lmy h
2
XJ

þ yj 1 qh  U j  ðq þ 1Þh :
j¼1

(56:11)

The only difference between this approximation and that used for the pure diffusion
is the jump part. Note that the last term on the right-hand side (RHS) of Eq. 56.11
is nonzero whenever we have one (or more) jump realization(s) in the interval
[(q
1)h, qh]. Moreover, as neither the intensity nor the jump size is state
dependent, the jump component can be simulated without any discretization
error, as follows. Begin by making a draw from a Poisson distribution with intensity
parameter l^t , say J . This gives a realization for the number of jumps over the
simulation time span. Then draw J uniform random variables over [0, L], and sort
them in ascending order so that U 1  U 2  . . .  U J . These provide realizations
for the J independent draws from f, say y1 , . . . , yJ .
SV models without jumps as in Eq. 56.4 (using a generalized Milstein scheme):
56 Density and Conditional Distribution-Based Specification Analysis 1539

     
Xðyqþ1Þh ¼ Xyqh þ eb 1 Xyqh ; y h þ s11 V yqh ; y ϵ 1, ðqþ1Þh þ s12 V yqh ; y ϵ 2, ðqþ1Þh
   
y  @s11 V yqh ; y
1  y  @s12, k V qh ; y 2 
þ s22 V qh ; y ϵ 2, ðqþ1Þh þ s22 V yqh ; y
2ð @V @V
ðqþ1Þh ð s 
 dW 1, t dW 2:s ,
qh qh
(56:12)
   
V ðyqþ1Þh ¼ V yqh þ e
b 2 V yqh ; y h þ s22 V yqh ; y ϵ 2, ðqþ1Þh
 
  y (56:13)
1 @s 22 V qh ; y
y
þ s22 V qh ; y ϵ 2, ðqþ1Þh ,
2
2 @V
 
where h
1/2ϵi,qh N(0,1), i ¼ 1, 2, E ϵ 1, qh ϵ 2, q0 h ¼ 0 for all q 6¼ q0 , and
0 1
  1 @s12 ðV; yÞ
e
b 1 ðV; yÞ b
B 1 ð V; y Þ
s22 ð V; y Þ C
e
b ðV; yÞ ¼ ¼@ 2 @V :
e
b 2 ðV; yÞ 1 @s22 ðV; yÞ A
b2 ðV; yÞ
s22 ðV; yÞ
2 @V

The last terms on the RHS of Eq. 56.12 involve stochastic integrals and
cannot be explicitly computed. However, they can be approximated up to an
error of order o(h) by (see, e.g., Eq. 3.7, pp. 347 in Kloeden and Platen (1999))
ð ðqþ1Þh ð s   
1 pffiffiffiffiffi 
dW 1, t dW 2, s  h x1 x2 þ rp m1, p x2
m2, p x1
qh qh 2
h X1 pffiffiffi  pffiffiffi 
p
þ B1, r 2x2 þ 2, r
B2, r 2x1 þ  1, r ,
2p r
r¼1

where for j ¼ 1, 2, xj, mj, p, Bj, r, j, r are i.i.d. N(0, 1) random variables, rp ¼ 12
1
Xp 1

2p1 2 r¼1 r 2
, and p is such that as h ! 0, p ! 1.

Stochastic Volatility with Jumps

Simulation of sample paths of diffusion processes with stochastic volatility and
jumps follows straightforwardly from the previous two cases. Whenever both
intensity and jump size are not state dependent, a jump component can be simulated
and added to either X(t) or the V(t) in the same manner as above. Extension to
general multidimensional and multifactor models both with and without jumps also
follows directly.
1540 D. Duong and N.R. Swanson

56.3.3.2 Simulating Distributions

In this section, we sketch out methods used to construct t-step ahead simulated
conditional distributions using simulated data. In applications, simulation tech-
niques are needed when the functional form conditional distribution is unknown.
We first illustrate the technique for one-factor models and then discuss multifactor
models.

One-factor Models
Consider the one-factor model as in Eq. 56.7. To estimate the simulated CDFs:
Step 1: Obtain y^T , N, h (using the entire sample) or y^t, N, h (recursive estimator) where
y^T , N, h and y^t, N, h are estimators as discussed in Sects. 56.3.3.3 and 56.3.3.4.
Step 2: Under y^T , N , h or y^t, N, h ,20 simulate S paths of length t, all having the same
starting value, Xt. In particular, for each path i ¼ 1, . . ., S of length t, generate
y^ , N, h
Xi,Ttþt ðXt Þ according to a Milstein schemes detailed in previous section, with
iid
y ¼ y^T , N, h or y^t, N, h. The errors used in simulation are eqh e N ð0; hÞ, and Qh ¼ t.
eqh is assumed to be independent across simulations, so that E(ei,qhej,qh) ¼ 0, for
all i 6¼ j and E(ei,qhei,qh) ¼ h, for any i, j. In addition, as the simulated diffusion is
ergodic, the effect of the starting value approaches zero at an exponential rate,
as t ! 1.  
Step 3: If y^T , N, h y^t, N, h is used, an estimate for the distribution, at time t + t,
^ ^
conditional on Xt, with estimator y T , N, h y t, N, h , is defined as

  n ^ o
^ ^ 1X S
y , N, h
F t u X t ; y T , N , h ¼ 1 Xi,Ttþt ðX t Þ  u :
S i¼1

Bhardwaj et al. (2008) show that if the model is correctly specified, then
XS n y^T, N, h o
1
S i¼1
1 X iþtþt ðX t Þ  u provides a consistent of the conditional distribution
Ft(u|Xt, y{) ¼ Pr(Xt+t
y{
 u|Xy{
t ¼ Xt).
Specifically, assume that T, N, S ! 1. Then, for the case of SGMM estimator,
if h ! 0, T/N ! 0, and h2T ! 0, T2/S ! 1, the following result holds for any
Xt, t  1, uniformly in u
    pr

F^t u Xt ; y^T , N, h
Ft ujXt ; y{ ! 0,

20
Note that N ¼ L for the SGMM estimator while N ¼ M ¼ S for NSQML estimator.
56 Density and Conditional Distribution-Based Specification Analysis 1541

In addition, if the model is correctly specified (i.e., if m(,) ¼ m0(,) and

s(,) ¼ s0(,)), then
 

F^t u Xt ; y^T , N, h
F0, t ðujXt , y0 Þ! 0:
pr

Step 4: Repeat Steps 1–3 for t ¼ 1, . . ., T
t. This yields T
t conditional

distributions that are t-steps ahead which will be used in the construction of the
specification tests.
The CDF simulation in the case selection test of multiple models with recursive
estimator is similar. For model k, let y^k, t, N, h be the recursive estimator of “pseudo
y^ N, h
true” y{k computed using all observations up to varying time t. Then, Xk,k,i,t,tþt ðXt Þ is
^
generated according to a Milstein schemes as in Sect. 56.3.3.1, with y ¼ y k, t, N, h
and the initial value Xt, Qh ¼ t. The corresponding empirical distribution of the
y^ N , h
simulated series Xk,k,i,t,tþt XðtÞ can then be constructed. Under some regularity
conditions,

S n o pr
1X y^ N , h
1 u1  Xk,k,i,t,tþt ð X t Þ  u 2 ! F y{ ð u 2 Þ
F y{ ðu1 Þ, t ¼ R, . . . , T
t,
S i¼1 Xk,k tþt ðXt Þ Xk,k tþt ðXt Þ

y{
where F y
{ ðuÞ is the marginal distribution of Xtþt
k
ðXt Þ implied by k model (i.e.,
Xk, tþt ðXt Þ
k

by the model used to simulate the series), conditional on the (simulation) starting
{
value Xt. Furthermore, the marginal distribution of Xytþt ðXt Þ is the distribution of Xt+t
 

conditional on the values observed at time t. Thus, FXy{ ðX Þ ðuÞ ¼ FTk u Xt ; y{k .
kþtþt t

y^ N , h
Of important note is that in the simulation of Xk,k,i,t,tþt ðXt Þ,
i ¼ 1, . . ., S, for each t,
t ¼ R, . . ., T
t, we must use the same set of randomly drawn errors and similarly
the same draws for numbers of jumps, jump times, and jump sizes. Thus, we
only allow for the starting value to change. In particular, for each i ¼ 1, . . ., S,
y^ , N, h y^
we generate Xk,k,i,RRþt ðXR Þ, . . . , Xk,k,i,T
t
t
, N, h ðX Þ . This yields an S  P matrix
T
t
of simulated values, where P ¼ T
R
t + 1 refers to the length of the out-of-
y^
sample period. Xk,k,i,Rþj , N, h  
(at time R + j + t) can be seen as t periods
Rþjþt XRþj
ahead value “predicted” by model k using all available information up to
time R + jR+j, j ¼ 1, . . ., P (the initial value XR+j and y^k, Rþj, N, h estimated using
X1, . . ., XR+j). The key feature of this setup is that it enables us to compare
“predicted” t periods ahead values (i.e., Xk,k,i,Rþj
y^ , N, h  
Rþjþt X Rþj ) with actual values
that are t periods ahead (i.e., XR+j+t), for j ¼ 1, . . ., P. In this manner,
simulation-based tests under ex-ante predictive density comparison framework
can be constructed.
1542 D. Duong and N.R. Swanson

Multifactor Model
Consider the multifactor model with a skeleton (Xt, V1t , . . .,Vdt )0 (e.g., stochastic
mean, stochastic volatility models, stochastic volatility of volatility) where only the
first element Xt is observed. For simulation of the CDF, the difficulty arises as we do
not know the initial values of latent variables (V1t , . . .,Vdt )0 at each point in time. We
generalize the simulation plan of Bhardwaj et al. (2008) and Cai and Swanson
(2011) to the case of d dimensions. Specifically, to overcome the initial value
difficulty, a natural strategy is to simulate a long path of length L for each latent
variable V1t , . . ., Vdt and use them to construct Xt+t and the corresponding simulated
CDF of Xt+t; and finally, we average out the volatility values. Note that there are Ld
combinations of the initial values V1t , . . ., Vdt . For illustration, consider the case of
stochastic volatility (d ¼ 1) and the Chen three-factor model as in Eq. 56.5 (d ¼ 2),
using recursive estimators.
For the case of stochastic volatility (d ¼ 1), i.e., (Xt, Vt)0 , the steps are as
follows:
Step 1: Estimate y^t, N, L, h using recursive SGMM or NSQML estimation methods.
y^
Step 2: Using the scheme in Eq. 56.13 with y ¼ y^t, N, L, h, generate the path V qht, N, L, h
y^
for q ¼ 1/h, . . ., Qh with Qh ¼ L and hence obtain V t, N, L, h j ¼ 1, . . . L.
j
Step 3: Using schemes in Eqs. 56.12, 56.13, simulate L  S paths of length t, setting
the initial value for the observable state variable to be Xt. For the initial values of
y^
unobserved volatility, use V j,t,qhN, L, h j ¼ 1,. . ., L as retrieved in Step 2. Also,
Ð Ð
keep the simulated random innovations (i.e., e1,qh,. e1,qh, (q+1)h qh ( sqhdW1,t)dW2,s)
to be constant across ^each j and t. Hence, for each replication i, using
y , N, h y^ , L, h
initial values Xt and V j,Tqh , we obtain Xj,t,i,Ntþt ðXt Þ which is a t-step ahead
simulated value.
Step 4: Now the estimator of Ft(u|Xt,y{) is defined as

  1 XL X S n ^ o
y ,h
F^t u Xt ; y^t, N, L, h ¼ 1 Xj,t,i,Ntþt ðX t Þ  u :
LS j¼1 i¼1

Note that by averaging over the initial value of the volatility process, we have
1 X S n y t, N , h o
integrated out its effect. In other words, 1 X j, i, tþt ð X t Þ  u is an
  S i¼1
y^
estimate of Ft u Xt ; V j,t,hN, h , y{ .
Step 5: Repeat Steps 1–4 for t ¼ 1, . . ., T
t. This yields T
t conditional
distributions that are t-steps ahead which will be used in the construction of the
specification tests.
For three-factor model (d ¼ 2), i.e., (Xt, V1t , V2t ), consider model (56.5), where
Wt ¼ (W1t , W2t , W3t ) are mutually independent standard Brownian motions.
Step 1: Estimate y^t, N, L, h using SGMM or NSQML estimation methods.
56 Density and Conditional Distribution-Based Specification Analysis 1543

1, y ^
Step 2: Given the estimated parameter y^t, N, L, h , generate the paths V qh t, N, L, h and
2, y^ 1, y^
V ph t, N, L, h for q, p ¼ 1/h, . . ., Qh with Qh ¼ L, and hence, obtain V j T, N, L, h ,
2, y^
V k T, N, L, h , k ¼ 1, . . ., L.
j
1, y^
Step 3: Given the observable X and the L  L simulated latent paths (V t, N, L, h and
t j
2, y^t, N, L, h y^t, N, L, h
Vk j, k ¼ 1, . . ., L) as the start values, we simulate t-step ahead Xtþt
 
1, y^ 2, y^
Xt ; V j t, N, L, h ; V k t, N, L, h . Since the start values for the two latent variables are
L  L length, so for each Xt we have N2 path. Now to integrate out
the initial effect of latent variables, form the estimate of conditional
distribution as

  1XL X L n ^   o
y N , L, h 1, y^ 2, y^
F^t, s u Xt y^ ¼ 2 1 Xs,t,tþt Xt ; V j t, N, L, h ; V k t, N, L, h  u ,
L j¼1 k¼1

where s denotes the sth simulation.

y^ N, L, h
Step 4: Simulate Xs,t,tþt S times, that is, repeat Step 3 S times, i.e., s ¼ 1, . . ., S.
The estimate of Ft(u|Xt,y{) is

  1X S  

F^t u Xt y^ ¼ F^t, s u Xt ; y^T , N, h :
S i¼1

Step 5: Repeat Steps 1–4 for t ¼ 1, . . ., T
t. This yields T
t conditional

distributions that are t-steps ahead which will be used in the construction of the
specification tests.
As a final remark, for the case of multiple competing models, we can
proceed similarly. In addition, in the next two subsections, we present the
exactly identified simulated (recursive) general method of moments and recursive
nonparametric simulated quasi-maximum likelihood estimators that can be used
in simulating distributions as well as constructing test statistics described in
Sect. 56.3.2. The bootstrap analogs of those estimators will be discussed
in Sect. 56.3.4.

56.3.3.3 Estimation: (Recursive) Simulated Generalized Method

of Moments (SGMM) Estimators
Suppose that we observe a discrete sample (skeleton) of T observations, say
(X1, X2, . . ., XT)0 , from the underlying diffusion in Eq. 56.7. The (recursive)
SGMM estimator y^t, L, h with 1  t  T is specified as
1544 D. Duong and N.R. Swanson

!
1 X t   1X L   0
y^t, L, h ¼ arg min g Xj
g Xyj, h
y2Y t j¼1 L j¼1
!
Xt   1X L  

1 1 y (56:14)
 Wt g Xj
g X j, h
t j¼1 L j¼1
¼ arg min Gt, L, h ðyÞ0 W t Gt, L, h ðyÞ,
y2Y

where g is a vector of p moment conditions, Y  ℜp (so that we have as many

y
moment conditions as parameters), and Xj,h ¼ Xy[Qjh/L], with L ¼ Qh is the simulated
y
path under generic parameter y and with discrete interval h. Xj,h is simulated using
the Milstein schemes.
Note that in the above expression, in the context of the specification test, y^t, L, h is
estimated using the whole sample, i.e., t ¼ T. In the out-of-sample context, the
recursive SGMM estimator y^t, L, h is estimated recursively using the using sample
from 1 up to t.
Typically, the p moment conditions are based on the difference between sample
moments of historical and simulated data or between sample moments and model
implied moments, whenever the latter are known in closed form. Finally, Wt is the
heteroskedasticity and autocorrelation (HAC) robust covariance matrix estimator,
defined as

! !0
1X lt X
t
lt
  1X t     1X t  
W
1
t ¼ wn g Xj
g Xj g Xj
n
g Xj ,
t n¼
l j¼nþ1þlt t j¼1 t j¼1
t

(56:15)

where wv ¼ 1
v/(lT + 1). Further, the pseudo true value, y{, is defined to be

y{ ¼ arg min G1 ðyÞ0 W 0 G1 ðyÞ,

y2Y

where

G1 ðyÞ0 W 0 G1 ðyÞ ¼ p lim GT , L, h ðyÞ0 W T GT , L, h ðyÞ,

L, T!1, h!0

and where y{ ¼ y0 if the model is correctly specified.

In the above setup, the exactly identified case is considered rather than the
overidentified SGMM. This choice guarantees that G1(y{) ¼ 0 even under
misspecification, in the sense that the model differs from the underlying DGP. As
pointed out by Hall and Inoue (2003), the root-N consistency does not hold for
overidentified SGMM estimators of misspecified models. In addition,
56 Density and Conditional Distribution-Based Specification Analysis 1545

 0  
∇y G1 y{ W { G1 y{ ¼ 0:

However, in the case for which the number of parameters and the number of
moment conditions are the same, ∇yG1(y{)0 W{ is invertible, and so the first-order
conditions also imply that G1(y{) ¼ 0.
Also note that other available estimation methods using moments include the
efficient method of moments (EMM) estimator as proposed by Gallant and Tauchen
(1996, 1997), which calculates moment functions by simulating the expected value
of the score implied by an auxiliary model. In their setup, parameters are then
computed by minimizing a chi-square criterion function.

56.3.3.4 Estimation: Recursive Nonparametric Simulated

Quasi-maximum Likelihood Estimators
In this section, we outline a recursive version of the NPSQML estimator of
Fermanian and Salanié (2004), proposed by Corradi and Swanson (2011). The
bootstrap counterpart of the recursive NPSQML estimator will be presented in
the next section.

One-Factor Models
Hereafter, let f(Xt|Xt
1, y{) be the conditional density associated with the above
jump diffusion. If f is known in closed form, we can just estimate y{ recursively,
using standard QML as21

1X t  
y^t ¼ arg max ln f Xj Xj
1 ; y , t ¼ R, . . . R þ P
1: (56:16)
y2Y t j¼2

Note that, similarly to the case of SGMM, the pseudo true value y{ is optimal in
the sense:

y{ ¼ arg max Eðlnf ðXt jXt
1 ; yÞÞ: (56:17)

y2Y

For the case f is not known in closed form, we can follow Kristensen and Shin
(2008) and Cai and Swanson (2011) to construct the simulated analog f^of f and then
use it to estimate y{. f^ is estimated as function of the simulated sample paths
y
Xt,i(Xt
1), for t ¼ 2, . . ., T
1, i ¼ 1, . . ., M. First, generate T
1 paths of length
one for each simulation replication, using Xt
1 with t ¼ 1, . . ., T as starting values.
y
Hence, at time t and simulation replication i, we obtain skeletons Xt,i (Xt
1), for
t ¼ 2, . . ., T
1; i ¼ 1,. . .M where M is the number of simulation paths (number of
y y
random draws or Xt,j (Xt
1) and Xt,l (Xt
1) are i.i.d.) for each simulation replication.

21
Note that as model k is, in general, misspecified, ∑ T
1 t¼1 fk(Xt|Xt
1,yk) is a quasi-likelihood and
fk(Xt|Xt
1,y{k ) is not necessarily a martingale difference sequence.
1546 D. Duong and N.R. Swanson

M is fixed across all initial values. Then the recursive NPSQML estimator is defined
as follows:

1X t  
y^t, M, h ¼ arg max lnf^M, h ðXi jXi
1 ; yÞtM  f^M, h ðXi jXi
1 ; yÞ , t  R,
y2Y t i¼2

where
!
1 XM Xyt, i, h ðXt
1 Þ
Xt
f^M, h ðXt jXt
1 ; yÞ ¼ K :
MxM i¼1 xM

Note that with abuse of notation, we define y^t, L, h for SGMM and y^t, M, h for
NPSQML estimators where L and M have different interpretations (L is the length
of each simulation path  and M is number  of random draws).
The function tM f^M, h ðXt jXt
1 ; yÞ is a trimming function.  It has some
characteristics such as positive and increasing, tM f^M, h ðXt ; Xt
1 ; yÞ ¼ 0, if f^M, h
 
ðXt ; Xt
1 ; yÞ < xdM and tM f^M, h ðXt ; Xt
1 ; yÞ ¼ 1 if f^M, h ðXt ; Xt
1 ; yÞ > 2xdM , for
some d > 0.22 Note that when the log density is close to zero, the derivative tends to
infinity and thus even very tiny simulation errors can have a large impact on the
likelihood. The introduction of the trimming parameter into the optimization
function ensures the impact of this case to be minimal asymptotically.

Multifactor Models
Since volatility is not observable, we cannot proceed as in the single-factor case
when estimating the SV model using NPSQML estimator. Instead, let Vyj be
generated according to Eq. 56.13, setting qh ¼ j, and j ¼ 1, . . ., L. The idea is to
simulate L different starting values for unobservable volatility, construct the sim-
ulated likelihood functions accordingly, and then average them out. For each
simulation replication at time t, we simulate L different values of Xt(Xt
1, Vyj ) by
generating L paths of length one, using fixed observable Xt
1 and unobservable Vyj ,
j ¼ 1, . . . , L as starting values. Repeat this procedure for any t ¼ 1, . . . , T
1 and
for any set j, j ¼ 1, . . . , L of random errors e1,t + (q + 1)h,j and e2,t+(q+1)h,j, q ¼ 1, . . . ,
1/h. Note that it is important to use the same set of random errors e1,t+(q+1)h,j and
e2, t+(q+1)h,j across different initial values for volatility. Denote the simulated value

22
Fermanian and Salanie (2004) suggest using the following trimming function:

4ðx
aN Þ3 3ðx
aN Þ4
tN ðxÞ ¼
,
a3N a4N

for aN  x  2aN.
56 Density and Conditional Distribution-Based Specification Analysis 1547

at time t and simulation replication i, under generic parameter y, using Xt
1, Vyj as
starting values as Xyt,i,h(Xt
1, Vyj ). Then
0   1
y y
1 XL
1 X M X t, i , h X t
1 ; V j
X t
f^M, L, h ðXt jXt
1 ; yÞ ¼ K@ A,
L j¼1 MxM i¼1 xM

and note that by averaging over the initial values for the unobservable volatility, its
effect is integrated out. Finally, define23

1X t  
y^t, M, L, h ¼ arg min ln f^M, L, h ðXs jXs
1 ; yÞtM  f^M, L, h ðXs jXs
1 ; yÞ , t  R:
y2Y t s¼2

Note that in this case, Xt is no longer Markov (i.e., Xt and Vt are jointly
Markovian, but Xt is not). Therefore, even in the case of true data generating
process, the joint likelihood cannot be expressed as the product of the conditional
and marginal distributions. Thus, y^t, M, L, h is necessarily a QML estimator. Further-
more, note that ∇yf(Xt|Xt
1,y{) is no longer a martingale difference sequence;
therefore, we need to use HAC robust covariance matrix estimators, regardless of
whether the model is the “correct” model or not.

56.3.4 Bootstrap Critical Value Procedures

The test statistics presented in Sects. 56.3.1 and 56.3.2 are implemented
using critical values constructed via the bootstrap. As mentioned earlier, motivation
for using the bootstrap is clear. The covariance kernel of the statistic limiting
distributions contains both parameter estimation error and the data-related
time-dependence components. Asymptotic critical value cannot thus be tabulated
in a usual way. Several methods have been proposed to tackle this issue.
One is the block bootstrap procedures which we discuss. Others have been
mentioned above.
With regard to the validity of the bootstrap, note that, in the case of dependent
observations without PEE, we can tabulate valid critical value using a simple
empirical version of the K€unsch (1989) block bootstrap. Now, the difficulty in
our context lies in accounting for parameter estimation error. Goncalves and White
(2002) establish the first-order validity of the block bootstrap for QMLE
(or m-estimator) for dependent and heterogeneous data. This is an important result

For discussion of asymptotic properties of y^k, t, M, L, h , as well as of regularity conditions, see

23

Corradi and Swanson (2011).

1548 D. Duong and N.R. Swanson

for the class of SGMM and NSQML estimators surveyed in this chapter and allows
Corradi and Swanson in CS (2011) and elsewhere to develop asymptotically valid
version of the bootstrap that can be applied under generic model misspecification,
as assumed throughout this chapter.
For the SGMM estimator, as shown in Corradi and Swanson (2005), the first-
order validity of the block bootstrap is valid in the exact identification case, and
when T/S ! 0. In this case, SGMM is asymptotically equivalent to GMM, and
consequently, there is no need to bootstrap the simulated series. In addition, in the
exact identification case, GMM estimators can be treated the same way that QMLE
estimators are treated. For the NSQML estimator, Corradi and Swanson (2011)
point out that the NPSQML estimator is asymptotically equivalent to the QML
estimator. Thus, we do not need to resample the simulated observations as the
negligible contribution of simulation errors.
Also note that critical values for these tests can be obtained using a recursive
version of the block bootstrap. When forming block bootstrap samples in the
recursive case, observations at the beginning of the sample are used more
frequently than observations at the end of the sample. This introduces
a location bias to the usual block bootstrap, as under standard resampling with
replacement, all blocks from the original sample have the same probability of
being selected. Also, the bias term varies across samples and can be either
positive or negative, depending on the specific sample. A first-order valid
bootstrap procedure for nonsimulation-based m-estimators constructed using
a recursive estimation scheme is outlined in Corradi and Swanson (2007a).
Here we extend the results of Corradi and Swanson (2007a) by establishing
asymptotic results for cases in which simulation-based estimators are
bootstrapped in a recursive setting.
Now the details of bootstrap procedure for critical value tabulation can be
outlined in 5 steps as follows:
Step 1: Let T ¼ bl, where b denotes the number of blocks and l denotes the length of
each block. We first draw a discrete uniform random variable, I1, that can take
values 0, 1, . . ., T
l with probability 1/(T
l + 1). The first block is given by
XI1 þ1 , . . . , XI1 þl. We then draw another discrete uniform random variable, say I2,
and a second block of length l is formed, say XI2 þ1 , . . . , XI2 þl . Continue in the
same manner, until you draw the last discrete uniform, say Ib, and so the
last block is XIb þ1 , . . . , XIb þl . Let’s call the X*t the resampled series, and note
that X*1, X*2, . . ., X*T corresponds to XI1 þ1 , XI1 þ2 , . . . , XIb þl . Thus, conditional
on the sample, the only random element is the beginning of each block.
In particular,

X1 , . . . , Xl , Xlþ1 , . . . , X2l , XT
lþ1 , . . . , XT ,

conditional on the sample, can be treated as b i.i.d. blocks of discrete uniform

random variables. Note that it can be shown that except a set of probability
measure approaching zero,
56 Density and Conditional Distribution-Based Specification Analysis 1549

!
1X T
1X T
E 
X ¼ Xt þ OP ðl=T Þ (56:18)
T t¼1 t T t¼1
! ! !
1 X
T
1XT
l X l
1X T
1X T  
Var  Xt ¼ Xt
Xt Xtþi
Xt þ OP l2 =T ,
T 1=2 t¼1
T t¼l i¼
l T t¼1 T t¼1
(56:19)

where E* and Var* denote the expectation and the variance operators with
respect to P* (the probability law governing the resampled series or the proba-
bility law governing the i.i.d. uniform random variables, conditional on the
sample) and where OP*(l/T)(OP*(l2/T)) denotes a term converging in probability
P* to zero, as l/T ! 0(l2/T ! 0).
In the case of recursive estimators, we proceed the bootstrap similarly as
follows. Begin by resampling b blocks of length l from the full sample, with
lb ¼ T. For any given t, it is necessary to jointly resample Xt, Xt+1, . . ., Xt+t. More
precisely, let Zt,t ¼ (Xt, Xt+1, . . ., Xt+t), t ¼ 1,. . ., T
t. Now, resample
b overlapping blocks of length l from Zt,t. This yields Zt,* ¼ (X*t , X*t+1, . . .,
*
Xt+t ), t ¼ 1, . . ., T
t.  
 
Step 2: Re-estimate y^t, N, h y^T , N, L, h using the bootstrap sample Zt,* ¼ (X*t , X*t+1,
. . .,Xt+t
*
), t ¼ 1, . . ., T
t (or full sample X*1, X*2, . . ., X*T). Recall that if we use the
entire sample for the estimation, as the specification test in Corradi and Swanson
 
(2005) and Bhardwaj et al. (2008), then y^t, N, h is denoted as y^T , N, h. The bootstrap
estimators for SGMM and NPSQML are presented below:
Bootstrap (Recursive) SGMM Estimators
If the full sample is used in the specification test as in Corradi and Swanson
(2005) and Bhardwaj et al. (2008), the bootstrap estimator is constructed
straightforward as

!
 1X T  
1X L   0
y^T , N, h ¼ arg min 
g Xj
y
g X j, h
y2Y T j¼1 L i¼1
!
XT   XL  
1 1
 W 
1
T g Xj
g Xyj, h ,
T j¼1 L i¼1

where W
1T and g(.) are defined in Eq. 56.15 and L is the length of each
simulation path.
Note that it is convenient not to resample the simulated series as the
simulation error vanishes asymptotically. In implementation, we do not
mimic its contribution to the covariate kernel.
In the case of predictive density-type model selection where recursive
estimators are needed, define the bootstrap analog as
1550 D. Duong and N.R. Swanson

0 00 1
 1 X t   1X T  
y^t, L, h ¼ arg min @ @@g X 

j g X j0 A
y2Y t j¼1 T j0 ¼1
!!!0
L   1X L 
X ^t, L, h 
y
y

L1 g X j, h
g X j, h
i¼1
L i¼1
0 00 1
1 X t   1X T  
 O
1 @ @@g X 
g Xj0 A
t j
t j¼1 T j0 ¼1
!!!
L   1X L 
1X ^t, L, h 
y

g Xyj, h
g Xj, h
L i¼1 L i¼1
¼ arg min Gt, L, h ðyÞ0 O
1
t Gt, L, h ðyÞ,
y2Y

where

2 32 3
1X lt X
t
lt   1X T     1X T  
O
1 ¼ w n, t 4g X


g Xj0 54g Xj
n


g X j0 5 :
t j
t n¼
l j¼nþ1þl
T j0 ¼1 T j0 ¼1
t t

Note that each bootstrap term is recentered around the (full) sample mean.
The intuition behind the particular recentering in bootstrap recursive
SGMM estimator is that it ensures that the mean of the bootstrap moment
conditions, evaluated at y^t, L, h , is zero, up to a negligible term. Specifically,
we have

0 0 1
1 Xt   1X T  
E@ @g X 

j g Xj0 A
t j¼1 T j0 ¼1
!!
L  ^  L  
1X y t, L , h 1X y^t, L, h

g X j, h
g Xj, h
L i¼1 L i¼1
   1 X T  
¼ E  g Xj
g X j0
T j0 ¼1
 
¼ Oðl=T Þ, with l ¼ o T 1=2 ,

where the O(l/T) term is due to the end block effect (see Corradi and Swanson
(2007b) for further discussion).
56 Density and Conditional Distribution-Based Specification Analysis 1551

Bootstrap Recursive NPSQML Estimators

Let Zt,* ¼ (X*t , X*t+1, . . ., Xt+t
*
), t ¼ 1, . . ., T
t. For each simulation
replication, generate T
1 paths of length one, using X*1, . . ., X*T
1 as starting
y
values, and so obtaining Xt,j (X*t
1) for t ¼ 2, . . ., T
1, i ¼ 1,.., M. Further, let:

  !
   1 X M Xyt, i, h Xt
1
Xt
f^M, h Xt Xt
1 , y ¼ K ,
MxM i¼1 xM

Now, for t ¼ R, . . ., R + P
1, define

 1X t  
y^t, M, h ¼ arg max ln f^M, h Xl Xl
1 ; y tM
y2Y t l¼2
    
 f M, h Xl Xl
1 , y
^

1X T
∇y f^M, h ðXl0 jXl0
1 ; yÞ

y0   y¼^yt, M, h
T l0 ¼2 f^M, h Xl0 jXt
l0 y
  

tM f^M, h Xl0 Xl0
1 ; y^t, M, h
  

þt0M f^M, h Xl0 Xl0
1 ; y^t, M, h


∇y f^M, h ðXl0 jXl0
1 ; yÞ ^yt, M, h
!
lnf^M, h ðXl0 jXl0
1 y^t, M, h Þ ,

where tM 0 (
) denotes the derivative of tM(
) with respect to its argument. Note
that each term in the simulated likelihood is recentered around the (full)
sample mean of the score, evaluated at y^t, M, h. This ensures that the bootstrap
score has mean zero, conditional  on the sample.
 The recentering term requires

computation of ∇y f^M, h Xl0 Xl0
1 ; y^t, M, h , which is not known in closed
form. Nevertheless, it can be computed numerically, by simply taking the
numerical derivative of the simulated likelihood.
Bootstrap Estimators for Multifactor Model
The SGMM and the bootstrap SGMM estimators in the case of multifactor
model are similar as in one-factor model. The difference is that the simulation
schemes (56.12) and (56.13) are used instead of Eq. 56.11.
For recursive NPSQML estimators, to construct the bootstrap counterpart

y^t, M, L, h of y^t, M, L, h , since M/T ! 1 and L/T ! 1, the contribution of
simulation error is asymptotically negligible. Hence, there is no need to resample
the simulated observations or the simulated initial values for volatility. Define
1552 D. Duong and N.R. Swanson

0   1
y  y 
  1 XL
1 XM X t, i, h X t
1 ; V j
Xt
f^M, L, h Xt Xt
1 , y ¼
 
K@ A:
L j¼1 MxM i¼1 xM

Now, for t ¼ R, . . ., R + P
1, define

 1 X
t
y^t, M, L, h ¼ arg max
y2Y t l¼2
    
logf^t, M, L, h Xl Xl
1 , y tM f^t, M, L, h Xl Xl
1 , y

1X T
∇y f^t, M, L, h ðXl0 jXl0
1 ; yÞ

y0   yt, M, L, h
T l0 ¼2 f^t, M, L, h Xl0 Xt
l0 ; y
  

tM ¼ f^t, M, L, h Xl0 Xl0
1 ; y^t, M, L, h
  

þt0M f^t, M, L, h Xl0 Xl0
1 ; y^t, M, L, h


∇y f^t, M, L, h ðXl0 jXl0
1 ; yÞ ^yt, M, L, h
!!
lnf^t, M, L, h ðXl0 jXl0
1 , y^t, M, L, h Þ ,

where tM 0 (
) denotes the derivative with respect to its argument.
Of note is that each bootstrap term is recentered around the (full) sample
mean. This is necessary because the bootstrap statistic is constructed using
the last P resampled observations, which in turn have been resampled from
the full sample. In particular, this is necessary regardless of the ratio, P/R. In
addition, in the case P/R ! 0, so that there is no need to mimic parameter
estimation error, the bootstrap statistics can be constructed using y^t, M, L, h

instead of y^t, M, L, h .
Step 3: Using the same set of random variables used in the construction of the
 
y^ N, h y^t, N , h
actual statistics, construct Xi,t,tþt , or X k, i, tþt, , i ¼ 1, . . ., S and t ¼ 1,. . ., T
t.
Note that we do not need resample the simulated series (as L/T ! 1,
simulation error is asymptotically negligible). Instead, simulate the
series using bootstrap estimators and using bootstrapped values as starting
values.
2* * * 2 * *
Step 4: Corresponding bootstrap statistics VT,N,h (or ZT,N,h , Dk,P,S , SV
 T,N,h, SZT,N,h,
 
*
SDk,P,S depending on the types of tests) which are built on y^t, N, h y^t, N, L, h Þ then
are followed correspondingly. For the numerical implementation, again, of
important note is that in the case where we pick the choice P/R ! 0, P, T,
56 Density and Conditional Distribution-Based Specification Analysis 1553


      
R ! 1, there is no need to re-estimate y^t, N, h y^t, N, L, h . y^t, N, h y^t, N, L, h can be
used in all the bootstrap replications.
Step 5: Repeat the bootstrap Steps 1–4 B times, and generate the empirical
distribution of the B bootstrap statistics.

56.4 Summary of Empirical Applications of the Tests

In this section, we briefly review some empirical applications of the methods

discussed above. We start with unconditional distribution test, as in Corradi
and Swanson (2005) and then give a specific empirical example using the
conditional distribution test from Bhardwaj et al. (2008). Finally, we briefly
discuss on conditional distribution specification test applied to multiple
competing models. The list of the diffusion models considered is provided in
Table 56.1.
Note that specification testing of the first model – a simplified version of
the CIR model (we refer to this model as Wong) – is carried out using the
unconditional distribution test. With the cumulative distribution function
known in closed form as in Eq. 56.8, the test statistic can be straightforwardly
calculated. It is also convenient to use GMM estimation in this case as the first two
moments are known in closed form, i.e., a
l and a/2(a
b), respectively. Corradi
and Swanson (2005) examine Hypothesis 1 using simulated data. Their Monte
Carlo experiments suggest that the test is useful, even for samples as small as
400 observations.
Hypothesis 2 is tested in Bhardwaj et al. (2008) and Cai and Swanson (2011).
For illustration, we focus on the results in Bhardwaj et al. (2008) where CIR, SV,
and SVJ models are empirically tested using the 1-month Eurodollar deposit rate
(as a proxy for short rate) for the sample period January 6, 1971–September
30, 2005, which yields 1,813 weekly observations. Note that one might apply
these tests to other datasets including the monthly federal funds rate, the weekly
3-month T-bill rate, the weekly US dollar swap rate, the monthly yield on zero-
coupon bonds with different maturities, and the 6-month LIBOR. Some of these
variables have been examined elsewhere, for example, in Ait-Sahalia (1999),
Andersen et al. (2004), Dai and Singleton (2000), Diebold and Li (2006), and
Piazzesi (2001).
The statistic needed to apply the test discussed in Sect. 56.3.1.2 is

ZT ¼ sup jZ T ðvÞj,
v2V
1554 D. Duong and N.R. Swanson

Table 56.1 Specification test hypotheses of continuous time spot rate processa
Reference and
Model Specification data Hypothesis
pffiffiffiffiffiffiffiffiffiffi
Wong drðtÞ ¼ ða
l
rðtÞÞdt þ ar ðtÞdW r ðtÞ Corradi and H1
Swanson (2005)
Simulated data
Bhardwaj
et al. (2008)
Eurodollar rate H2
(1971–2005)
pffiffiffiffiffiffiffiffiffi
CIR drðtÞ ¼ kr ðr
rðtÞÞdt þ V ðtÞdW r ðtÞ, Cai and Swanson
(2011)
Eurodollar Rate H2, H3
(1971–2008)
Cai and Swanson
(2011)
CEV drðtÞ ¼ kr ðr
rðtÞÞdt þ sr rðtÞr dW r ðtÞ Eurodollar rate H2, H3
(1971–2008)
pffiffiffiffiffiffiffiffiffi
SV drðtÞ ¼ kr ðr
rðtÞÞdt þ V ðtÞdW r ðtÞ, Bhardwaj H2
et al. (2008)
pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kr ðv
V ðtÞÞdt þ sv V ðtÞdW v ðtÞ, Cai and Swanson H2, H3
(2011)
pffiffiffiffiffiffiffiffiffi
SVJ drðtÞ ¼ kr ðr
rðtÞÞdt þ V ðtÞdW r ðtÞ þ J u dqu
J d dqd , Bhardwaj H2
et al. (2008)
pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kv ðv
V ðtÞÞdt þ sv V ðtÞdW v ðtÞ, Cai and Swanson H2, H3
(2011)
pffiffiffiffiffiffiffiffiffi
CHEN drðtÞ ¼ kr ðyðtÞ
rðtÞÞdt þ V ðtÞdW r , Cai and Swanson
(2011)
pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kv ðv
V ðtÞÞdt þ sv V ðtÞdW v ðtÞ, Eurodollar rate H2, H3
  pffiffiffiffiffiffiffiffi
dyðtÞ ¼ ky y
yðtÞ dt þ sy yðtÞdW y ðtÞ, (1971–2008)
pffiffiffiffiffiffiffiffiffi
CHENJ drðtÞ ¼ kr ðyðtÞ
rðtÞÞdt þ V ðtÞdW r ðtÞ þ J u dqu
J d dqd , Cai and Swanson
(2011)
pffiffiffiffiffiffiffiffiffi
dV ðtÞ ¼ kv ðv
V ðtÞÞdt þ sv V ðtÞdW v ðtÞ, Eurodollar rate H2, H3
  pffiffiffiffiffiffiffiffi
dyðtÞ ¼ ky y
yðtÞ dt þ sy yðtÞdW y ðtÞ, (1971–2008)
a
Note that the third column, “Reference and data,” provides the referenced papers and data used in
empirical applications. In the fourth column, H1, H2, and H3 denote Hypothesis 1, Hypothesis
2, and Hypothesis 3, respectively. The hypotheses are presented corresponding to the references in
the third column. For example, for CIR model, H2 corresponds to Bhardwaj et al. (2008) and H2
and H3 correspond to Cai and Swanson (2011)

where

1 X T
t
ZT ðvÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1fXt  vg
T
t t¼1
!
S n o
1X y^T , N, h

 1 u  Xs, tþt  u
1 u  Xtþt  u ,
S s¼1
56 Density and Conditional Distribution-Based Specification Analysis 1555

and

Z T ¼ sup Z T ðvÞ ,
v2V

Where

1 X T
t

Z T ðvÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1 Xt  v
T
t t¼1
!
S 
1X 
y^T , N, h


 1 u  Xs, tþt,   u
1 u  Xtþt  u
S s¼1
1 X T
t

pffiffiffiffiffiffiffiffiffiffiffi 1fXt  vg:
T
t t¼1
!
S n o
1X y^T , N , h

 1 u  Xs, tþt  u
1 u  Xtþt  u :
S s¼1

For the case of stochastic volatility models, similarly we have

SZT ¼ sup jSZ T ðvÞj,

v2V

where

1 X T
t
SZT ðvÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1fXt  vg
T
t t¼1
!
S n o
1 X L X
y^T , N, h

 1 u  Xj, s, tþt  u
1 u  Xtþt  u ,
LS j¼1 s¼1

and its bootstrap analog

SZ T ¼ sup SZ T ðvÞ ,
v2V

where

S 
1 X T
t
 1 XL X
y^


SZT ðvÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi 1 Xt  v  1 u  Xj,i,s,T,tþt

N, h
,  u
T
t t¼1 LS j¼1 s¼1
!

 1 X T
t

1 u  Xtþt  u
pffiffiffiffiffiffiffiffiffiffiffi 1fX t  v g
T
t t¼1
!
S n o
1 X L X
y^i, T , N, h

 1 u  Xj, s, tþt  u
1 u  Xtþt  u :
LS j¼1 s¼1
1556 D. Duong and N.R. Swanson

Bhardwaj et al. (2008) carry out these tests using t-step ahead confidence inter-
vals. They set t ¼ {1, 2, 4, 12} which  corresponds
  to 1-week, 2-week,
 1-month, and
one-quarter ahead intervals and set uu ¼ X  0:5sX , X  sX , covering 46.3 %
and 72.4 % coverage, respectively. X and sX are the mean and variance of an initial
sample of data. In addition, S ¼ {10T, 20T} and l ¼ {5, 10, 20, 50}.
For illustrative purposes, we report one case from Bhardwaj et al. (2008). The test is
implemented by setting S ¼ 10T and l ¼ 25 for the calculation of both ZT and SZT. In
Table 56.2, single, double, and triple starred entries represent rejection using 20 %,
10 %, and 5 % size tests, respectively. Not surprisingly, the findings are consistent with
some other papers in the specification test literature such as Aı̈t-Sahalia (1996) and
Bandi (2002). Namely, the CIR model is rejected using 5 % size tests in almost all cases.
When considering SV and SVJ models, smaller confidence intervals appear to lead to
more model rejections. Moreover, results are somewhat mixed when evaluating the
SVJ model, with a slightly higher frequency of rejection than in the case of SV models.
Finally, turning to Hypothesis 3, Cai and Swanson (2011) use an extended
version of the above dataset, i.e., the 1-month Eurodollar deposit rate from January
1971 to April 2008 (1,996 weekly observations). Specifically, they examine
whether the Chen model is the “best” model amongst multiple alternative models
including those outlined in Table 56.1. The answer is “yes.” In this example, the test
was implemented using Dk,p,N(u1, u2), as described in Sects. 56.3.1 and 56.3.2,
where P ¼ T/2 and predictions are constructed using recursively estimated models
and the simulation sample length used to address latent variable initial values
  is set
at L ¼ 10T. The choice of other inputs to the test such as t and interval u, u is the
same as in Bhardwaj et al. (2008). The number of replications S, the block length l,
and number of bootstrap replications are S ¼ 10T, l ¼ 20, and B ¼ 100.
Cai and Swanson (2011) also compare the Chen model with the so-called smooth
transition autoregression (STAR) model defined as follows:

r t ¼ ðy1 þ b1 r t
1 ÞGðg; zt ; cÞ þ ðy1 þ b2 r t
1 Þð1
Gðg; zt ; cÞÞ þ ut ,

whereut is a disturbance term; y1, b1,g, b2, and c are constants; G(·) is the logistic
CDF i:e:; G; ðg; zt ; cÞ; ¼; 1þexpð1gðzt
cÞÞ and the number of lags; and p is selected via
the use of Schwarz information criterion. Test statistics and predictive density-type
“mean square forecast error” (MSFEs) values are again calculated as in Sects.
56.3.1 and 56.3.2.24 Their results indicate that at a 90 % level of confidence, one
cannot reject the null hypothesis that the Chen model generates predictive densities
at least as accurate as the STAR model, regardless of forecast horizon and confi-
dence interval width. Moreover, in almost all cases, the Chen model has lower
MSFE, and the magnitude of the MSFE differential between the Chen model and
STAR model rises as the forecast horizon increases. This confirms their in-sample
findings that the Chen model also wins when carrying out in-sample tests.

24
See Table 6 in Cai and Swanson (2011) for complete details.
 56

Table 56.2 Empirical illustration of specification testing – CIR, SV, SVJ modelsa
 
u, u CIR SV SVJ
ZT 5 % CV 10 % CV SZT 5 % CV 10 % CV SZT 5 % CV 10 % CV
l ¼ 25
1 X  0:5sX 0.5274*** 0.2906 0.3545 0.9841*** 0.8729 0.9031 1.1319 1.8468 2.1957
X  sX 0.4289*** 0.2658 0.3178 0.6870 0.6954 0.7254 1.2272* 1.1203 1.3031
***
2 X  0:5sX 0.6824 0.4291 0.4911 0.4113 1.3751 1.4900 0.9615* 0.8146 1.1334
*
X  sX 0.4897 0.4264 0.5182 0.3682 1.1933 1.2243 1.2571 1.3316 1.4096
4 X  0:5sX 0.8662** 0.7111 0.8491 1.2840 2.3297 2.6109 1.5012* 1.1188 1.6856
X  sX 0.8539* 0.7512 0.9389 1.0472 2.2549 2.2745 0.9901* 0.9793 1.0507
12 X  0:5sX 1.1631* 1.0087 1.3009 1.7687 4.9298 5.2832 2.4237* 2.0818 3.0640
X  sX 1.0429 1.4767 2.0222 1.7017 5.2601 5.6522 1.4522 1.7400 2.1684
a
Tabulated entries are test statistics and 5 %, 10 %, and 20 % level critical values. Test intervals are given in the second column of the table, for t ¼ 1, 2, 4, 12.
All tests are carried out using historical 1-month Eurodollar deposit rate data for the period January 1971–September 2005, measured at a weekly frequency.
* **
Density and Conditional Distribution-Based Specification Analysis

, , and *** denote rejection at the 20 %, 10 %, and 5 % levels, respectively. Additionally, X and sX are the mean and standard deviation of the historical data.
See above for complete details
1557
1558 D. Duong and N.R. Swanson

56.5 Conclusion

This chapter reviews a class of specification and model selection-type tests devel-
oped by Corradi and Swanson (2005), Bhardwaj et al. (2008), and Corradi and
Swanson (2011) for continuous time models. We begin with outlining the setup
used to specify the types of diffusion models considered in this chapter. Thereafter,
diffusion models in finance are discussed, and testing procedures are outlined.
Related testing procedures are also discussed, both in contexts where models are
assumed to be either correctly specified under the null hypothesis or generically
misspecified under both the null and alternative test hypotheses. In addition to
discussing tests of correct specification and test for selecting amongst alternative
competing models, using both in-sample methods and via comparison of predictive
accuracy, methodology is outlined allowing for parameter estimation, model and
data simulation, and bootstrap critical value construction.
Several extensions that are left to future research are as follows. First, it remains
to construct specification tests that do not integrate out the effects of latent factors.
Additionally, it remains to examine the finite sample properties of the estimators
and bootstrap methods discussed in this chapter.

Acknowledgments The authors thank the editor, Cheng-Few Lee, for many useful suggestions
given during the writing of this chapter. Duong and Swanson would like to thank the Research
Council at Rutgers University for research support.

References
Ait-Sahalia, Y. (1999). Transition densities for interest rate and others nonlinear diffusions.
Journal of Finance, LIV, 1361–1395.
Aїt-Sahalia, Y. (1996). Testing continuous time models of the spot interest rate. Review of
Financial Studies, 9, 385–426.
Aїt-Sahalia, Y. (2002). Maximum likelihood estimation of discretely sampled diffusions: A closed
form approximation approach. Econometrica, 70, 223–262.
Aїt-Sahalia, Y. (2007). Estimating continuous-time models using discretely sampled data. In
R. Blundell, P. Torsten, & W. K. Newey (Eds.), Econometric Society World Congress invited
lecture, in Advances in economics and econometrics, theory and applications, Ninth World
Congress, Econometric Society Monographs. Cambridge University Press.
Aїt-Sahalia, Y., Fan, J., & Peng, H. (2009). Nonparametric transition-based tests for diffusions.
Journal of the American Statistical Association, 104, 1102–1116.
Altissimo, F., & Mele, A. (2009). Simulated nonparametric estimation of dynamic models with
application in finance. Review of Economic Studies, 76, 413–450.
Andersen, T. G., & Lund, J. (1997). Estimating continuous time stochastic volatility models of
short term interest rates. Journal of Econometrics, 72, 343–377.
Andersen, T. G., Benzoni, L., & Lund, J. (2004). Stochastic volatility, mean drift, and jumps in the
short-term interest rate (Working paper). Northwestern University.
Andrews, D. W. K. (1993). An introduction to econometric applications of empirical process
theory for dependent random variables. Econometric Reviews, 12, 183–216.
Andrews, D. W. K. (1997). A conditional Kolmogorov test. Econometrica, 65, 1097–1128.
Bai, J. (2003). Testing parametric conditional distributions of dynamic models. Review of
Economics and Statistics, 85, 531–549.
56 Density and Conditional Distribution-Based Specification Analysis 1559

Bandi, F. (2002). Short-term interest rate dynamics: A spatial approach. Journal of Financial
Economics, 65, 73–110.
Beckers, S. (1980). The constant elasticity of variance model and US implications for option
pricing. Journal of Finance, 35, 661–673.
Bhardwaj, G., Corradi, V., & Swanson, N. R. (2008). A simulation based specification test for
diffusion processes. Journal of Business and Economic Statistics, 26, 176–193.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81, 637–654.
Bontemps, C., & Meddahi, N. (2005). Testing normality: A GMM approach. Journal of Econo-
metrics, 124, 149–186.
Brennan, M. J., & Schwartz, E. S. (1979). A continuous-time approach to the pricing bonds.
Journal of Banking and Finance, 3, 133–155.
Cai, L., & Swanson, N. R. (2011). An empirical assessment of spot rate model stability. Journal of
Empirical Finance, 18, 743–764.
Chan, K. C., Karolyi, G. A., Longstaff, F. A., & Sanders, A. B. (1992). An empirical comparison
of alternative models of the short-term interest rate. Journal of Finance, 47, 1209–1227.
Chen, L. (1996). Stochastic mean and stochastic volatility – A three-factor model of term structure
of interest rates and its application to the pricing of interest rate derivatives. Oxford, UK:
Blackwell.
Clements, M. P., & Smith, J. (2000). Evaluating the forecast densities of linear and nonlinear
models: Applications to output growth and unemployment. Journal of Forecasting, 19,
255–276.
Clements, M. P., & Smith, J. (2002). Evaluating multivariate forecast densities: A comparison of
two approaches. International Journal of Forecasting, 18, 397–407.
Corradi, V., & Swanson, N. R. (2005). A bootstrap specification test for diffusion processes.
Journal of Econometrics, 124, 117–148.
Corradi, V., & Swanson, N. R. (2006). Predictive density and conditional confidence interval
accuracy tests. Journal of Econometrics, 135, 187–228.
Corradi, V., & Swanson, N. R. (2007a). Nonparametric bootstrap procedures for predictive
inference based on recursive estimation schemes. International Economic Review February,
2007(48), 67–109.
Corradi, V., & Swanson, N. R. (2007b). Nonparametric bootstrap procedures for
predictive inference based on recursive estimation schemes. International Economic Review,
48, 67–109.
Corradi, V., & Swanson, N. R. (2011). Predictive density construction and accuracy testing with
multiple possibly misspecified diffusion models. Journal of Econometrics, 161, 304–324.
Corradi, V. & Swanson, N. R. (2002). A consistent test for nonlinear out of sample predictive
accuracy. Journal of Econometrics, 110, 353–381.
Courtadon, G. (1982). The pricing of options on default-free bonds. Journal of Financial and
Quantitative Analysis, 17, 75–100.
Cox, J. C., Ingersoll, J. E., & Ross, S. A. (1985). A theory of the term structure of interest rates.
Economerica, 53, 385–407.
Dai, Q., & Singleton, K. J. (2000). Specification analysis of affine term structure models. Journal
of Finance, 55, 1943–1978.
Diebold, F. X., & Li, C. (2006). Forecasting the term structure of government bond yields. Journal
of Econometrics, 130, 337–364.
Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business and
Economic Statistics, 13, 253–263.
Diebold, F. X., Gunther, T., & Tay, A. S. (1998a). Evaluating density forecasts with applications to
finance and management. International Economic Review, 39, 863–883.
Diebold, F. X., Tay, A. S., & Wallis, K. D. (1998b). Evaluating density forecasts of inflation: The
survey of professional forecasters. In R. F. Engle & H. White (Eds.), Festschrift in honor of
C.W.J. Granger. Oxford: Oxford University Press.
1560 D. Duong and N.R. Swanson

Diebold, F. X., Hahn, J., & Tay, A. S. (1999). Multivariate density forecast evaluation and
calibration in financial risk management: High frequency returns on foreign exchange. Review
of Economics and Statistics, 81, 661–673.
Duan, J. C. (2003). A specification test for time series models by a normality transformation
(Working paper). University of Toronto.
Duffie, D., & Kan, R. (1996). A yield-factor model of interest rates. Mathematical Finance, 6,
379–406.
Duffie, D., & Singleton, K. (1993). Simulated moment estimation of Markov models of asset
prices. Econometrica, 61, 929–952.
Duffie, D., Pan, J., & Singleton, K. (2000). Transform analysis and asset pricing for affine
jump-diffusions. Econometrica, 68, 1343–1376.
Duong, D., & Swanson, N. R. (2010). Empirical evidence on jumps and large fluctuations in
individual stocks (Working paper). Rutgers University.
Duong, D., & Swanson, N. R. (2011). Volatility in discrete and continuous time models: A survey
with new evidence on large and small jumps. Missing Data Methods: Advances in
Econometrics, 27B, 179–233.
Emanuel, D., & Mcbeth, J. (1982). Further results on the constant elasticity of variance call option
pricing model. Journal of Financial Quantitative Analysis, 17, 533–554.
Fermanian, J. D., & Salanié, B. (2004). A non-parametric simulated maximum likelihood
estimation method. Econometric Theory, 20, 701–734.
Gallant, A. R., & Tauchen, G. (1996). Which moments to match. Econometric Theory, 12,
657–681.
Gallant, A. R., & Tauchen, G. (1997). Estimation of continuous time models for stock returns and
interest rates. Macroeconomic Dynamics, 1, 135–168.
Goncalves, S., & White, H. (2002). The bootstrap of the mean for dependent heterogeneous arrays.
Econometric Theory, 18, 1367–1384.
Hall, A. R., & Inoue, A. (2003). The large sample behavior of the generalized method of moments
estimator in misspecified models. Journal of Econometrics, 114, 361–394.
Hansen, B. E. (1996). Inference when a nuisance parameter is not identified under the null
hypothesis. Econometrica, 64, 413–430.
Heston, S. L. (1993). A closed form solution for option with stochastic volatility with applications
to bond and currency options. Review of Financial Studies, 6, 327–344.
Hong, Y. (2001). Evaluation of out-of-sample probability density forecasts with applications to
S&P 500 stock prices (Working paper). Cornell University.
Hong, Y. M., & Li, H. T. (2005). Nonparametric specification testing for continuous time models
with applications to term structure interest rates. Review of Financial Studies, 18, 37–84.
Hong, Y. M., Li, H. T., & Zhao, F. (2007). Can the random walk model be beaten in out-of-sample
density forecasts? Evidence from intraday foreign exchange rates. Journal of Econometrics,
141, 736–776.
Hull, J., & White, A. (1987). The pricing of options on assets with stochastic volatility. Journal of
Finance, 42, 281–300.
Inoue. (2001). Testing for distributional change in time series. Econometric Theory, 17, 156–187.
Jiang, G. J. (1998). Nonparametric modelling of US term structure of interest rates and implica-
tions on the prices of derivative securities. Journal of Financial and Quantitative Analysis, 33,
465–497.
Karlin, S., & Taylor, H. M. (1981). A second course in stochastic processes. San Diego: Academic.
Kloeden, P. E., & Platen, E. (1999). Numerical solution of stochastic differential equations.
New York: Springer.
Kolmogorov, A. N. (1933). Sulla Determinazione Empirica di una Legge di Distribuzione.
Giornale dell’Istitutodegli Attuari, 4, 83–91.
Kristensen, D., & Shin, Y. (2008). Estimation of dynamic models with nonparametric simulated
maximum likelihood (Creates Research paper 2008-58). University of Aarhus and Columbia
University.
56 Density and Conditional Distribution-Based Specification Analysis 1561

K€unsch, H. R. (1989). The Jackknife and the bootstrap for general stationary observations. Annals
of Statistics, 17, 1217–1241.
Marsh, T., & Rosenfeld, E. (1983). Stochastic processes for interest rates and equilibrium bond
prices. Journal of Finance, 38, 635–646.
Meddahi, N. (2001). An Eigenfunction approach for volatility modeling (Working paper).
University of Montreal.
Merton, C. R. (1973). Theory of rational option pricing. The Bell Journal of Economics and
Management Science, 4, 141–183.
Nelson, D. B. (1990). ARCH as diffusion approximations. Journal of Econometrics, 45, 7–38.
Pardoux, E., & Talay, D. (1985). Discretization and simulation of stochastic differential equations.
Acta Applicandae Mathematicae, 3, 23–47.
Piazzesi, M. (2001). Macroeconomic jump effects and the yield curve (Working paper). UCLA.
Piazzesi, M. (2004). Affine term structure models. Manuscript prepared for the Handbook of
Financial Econometrics. University of California at Los Angeles.
Piazzesi, M. (2005). Bond yields and the federal reserve. Journal of Political Economy, 113,
311–344.
Politis, D. N., Romano, J. P., & Wolf, M. (1999). Subsampling. New York: Springer.
Pritsker, M. (1998). Nonparametric density estimators and tests of continuous time interest rate
models. Review of Financial Studies, 11, 449–487.
Rosenblatt, M. (1952). Remarks on a multivariate transformation. Annals of Mathematical
Statistics, 23, 470–472.
Singleton, K. J. (2006). Empirical dynamic asset pricing – Model specification and econometric
assessment. Princeton: Princeton University Press.
Smirnov, N. (1939). On the estimation of the discrepancy between empirical curves of distribution
for two independent samples. Bulletin Mathematique de l’Universite’ de Moscou, 2, fasc. 2.
Sullivan, R., Timmermann, A., & White, H. (1999). Data-snooping, technical trading rule perfor-
mance, and the bootstrap. The Journal of Finance, 54(5), 1647–1691.
Sullivan, R., Timmermann, A., & White, H. (2001). Dangers of data-driven inference: The case of
calendar effects in stock returns. Journal of Econometrics, 105, 249–286.
Thompson, S. B. (2008). Identifying term structure volatility from the LIBOR-swap curve. Review
of Financial Studies, 21, 819–854.
van der Vaart, A. (1998). Asymptotic statistics. Cambridge, MA: Cambridge University Press.
Vasicek, O. A. (1977). An equilibrium characterization of the term structure. Journal of Financial
Economics, 5, 177–188.
Vuong, Q. (1989). Likelihood ratio tests for model selection and non-nested hypotheses.
Econometrica, 57, 307–333.
White, H. (1982). Maximum likelihood estimation of misspecified models. Econometrica, 50,
1–25.
White, H. (2000). A reality check for data snooping. Econometrica, 68, 1097–1126.
Wong, E. (1964). The construction of a class of stationary Markov processes. In R. Bellman (Ed.),
Sixtenn symposia in applied mathematics: Stochastic processes in mathematical physics and
engineering. Providence: American Mathematical Society.
Assessing the Performance of Estimators
Dealing with Measurement Errors 57
Heitor Almeida, Murillo Campello, and Antonio F. Galvao

Contents
57.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1564
57.2 Dealing with Mismeasurement: Alternative Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1570
57.2.1 The Erickson–Whited Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1570
57.2.2 An OLS-IV Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1574
57.2.3 GMM Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1578
57.3 Monte Carlo Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1581
57.3.1 Monte Carlo Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1582
57.3.2 The EW Identification Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1585
57.3.3 Bias and Efficiency of the EW, OLS-IV, and AB-GMM Estimators . . . . . . . 1587
57.3.4 Heteroscedasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1593
57.3.5 Identification of the EW Estimator in Panel Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1594
57.3.6 Revisiting the OLS-IV Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1598
57.3.7 Distributional Properties of the EW and OLS-IV Estimators . . . . . . . . . . . . . . . 1600
57.4 Empirical Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1602
57.4.1 Theoretical Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1603
57.4.2 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1604
57.4.3 Testing for the Presence of Fixed Effects and Heteroscedasticity . . . . . . . . . . . 1604
57.4.4 Implementing the EW Identification Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1605

This article is based on Almeida, H., M. Campello, and A. Galvao, 2010, Measurement Errors in
Investment Equations, Review of Financial Studies, 23, pages 3279–3382.
H. Almeida (*)
University of Illinois at Urbana-Champaign, Champaign, IL, USA
e-mail: halmeida@illinois.edu
M. Campello
Cornell University, Ithaca, NY, USA
e-mail: campello@cornell.edu
A.F. Galvao
University of Iowa, Iowa City, IA, USA
e-mail: antonio-galvao@uiowa.edu

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1563
DOI 10.1007/978-1-4614-7750-1_57,
# Springer Science+Business Media New York 2015
1564 H. Almeida et al.

57.4.5 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1606

57.4.6 Robustness of the Empirical OLS-IV Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1611
57.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1613
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1615

Abstract
We describe different procedures to deal with measurement error in linear
models and assess their performance in finite samples using Monte Carlo
simulations and data on corporate investment. We consider the standard instru-
mental variable approach proposed by Griliches and Hausman (Journal of
Econometrics 31:93–118, 1986) as extended by Biorn (Econometric Reviews
19:391–424, 2000) [OLS-IV], the Arellano and Bond (Review of Economic
Studies 58:277–297, 1991) instrumental variable estimator, and the higher-order
moment estimator proposed by Erickson and Whited (Journal of Political Econ-
omy 108:1027–1057, 2000, Econometric Theory 18:776–799, 2002). Our anal-
ysis focuses on characterizing the conditions under which each of these
estimators produce unbiased and efficient estimates in a standard “errors-in-
variables” setting. In the presence of fixed effects, under heteroscedasticity, or in
the absence of a very high degree of skewness in the data, the EW estimator is
inefficient and returns biased estimates for mismeasured and perfectly measured
regressors. In contrast to the EW estimator, IV-type estimators (OLS-IV and
AB-GMM) easily handle individual effects, heteroscedastic errors, and different
degrees of data skewness. The IV approach, however, requires assumptions
about the autocorrelation structure of the mismeasured regressor and the mea-
surement error. We illustrate the application of the different estimators using
empirical investment models. Our results show that the EW estimator produces
inconsistent results when applied to real-world investment data, while the IV
estimators tend to return results that are consistent with theoretical priors.

Keywords
Investment equations • Measurement error • Monte Carlo simulations • Instru-
mental variables • GMM • Bias • Fixed effects • Heteroscedasticity • Skewness •
High-order moments

57.1 Introduction

OLS estimators are the workhorse of empirical research in many fields in applied
economics. Researchers see a number of advantages in these estimators. Most
notably, they are easy to implement and the results they generate are easy
to replicate. Another appealing feature of OLS estimators is that they easily
accommodate the inclusion of individual (e.g., firm and time) idiosyncratic effects.
Despite their popularity, however, OLS estimators are weak in dealing with the
problem of errors in variables. When the independent (right-hand side) variables of
an empirical model are mismeasured, coefficients estimated via standard OLS are
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1565

inconsistent (attenuation bias). This poses a problem since, in practice, it is hard to

think of any empirical proxies in applied research whose measurement is not
a concern.
In this chapter, we describe three estimators that deal with the problem of
mismeasurement, namely, the standard instrumental variable approach extended
by Biorn (2000) [OLS-IV], the Arellano and Bond (1991) instrumental variable
estimator [AB-GMM], and the higher-order moment estimator proposed by
Erickson and Whited (2000, 2002) [henceforth, EW]. We also assess the perfor-
mance of these estimators in finite samples using Monte Carlo simulations and
illustrate their application using data on corporate investment.
While we provide a formal presentation in the next section, it is useful to discuss
the intuition behind the estimation approaches we analyze. All approaches share the
attractive property that they do not require the researcher to look for instruments
outside the model being considered.1 They differ, however, on how identification is
achieved.
Both instrumental variable approaches rely on assumptions about the serial
correlation of the latent variable and the innovations of the model (the model’s
disturbances and the measurement error). There are two conditions that must hold
to ensure identification. First, the true value of the mismeasured regressor must
have some degree of autocorrelation. In this case, lags of the mismeasured regressor
are natural candidates for the instrumental set since they contain information about
the current value of the mismeasured regressor.2 This condition is akin to the
standard requirement that the instrument be correlated with the variable of interest.
The other necessary condition is associated with the exclusion restriction that is
standard in IV methods and relates to the degree of serial correlation of the
innovations. A standard assumption guaranteeing identification is that the
measurement-error process is independently and identically distributed. This con-
dition ensures that past values of the observed variables are uncorrelated with the
current value of the measurement error, validating the use of lags of observed
variables as instruments. Under certain conditions, one can also allow for autocor-
relation in the measurement error. Examples in which identification works are when
autocorrelation is constant over time or when it evolves according to a moving
average process.3 The first assumption ensures identification because it means that
while lagged values of the measurement error are correlated with its current value,
any past shocks to the measurement-error process do not persist over time. The
moving average assumption allows for shocks to persist over time, but it imposes
restrictions on the instrumental set. In particular, as we show below, it precludes the

1
Naturally, if extraneous instruments are available, they can help solve the identification problem.
See Rauh (2006) for the use of discontinuities in pension contributions as a source of variation in
cash flows in an investment model. Bond and Cummins (2000) use information contained in
financial analysts’ forecasts to instrument for investment demand.
2
Lags of the well-measured variable may also be included in the instrument set if they are believed
to also contain information about the mismeasured one.
3
See, among others, Biorn (2000), Wansbeek (2001), and Xiao et al. (2008).
1566 H. Almeida et al.

use of shorter lags of observed variables in the instrument set, as the information
contained in short lags may be correlated with the current value of the measurement
error.
The EW estimator is based on high-order moments of residuals obtained by
“partialling out” perfectly measured regressors from the dependent, observed
mismeasured, and latent variables, as well as high-order moments of the innova-
tions of the model. The key idea is to create a set of auxiliary equations as a function
of these moments and cross-moments. Implementation then requires a high degree
of skewness in the distribution of the partialled out latent variable. Our analysis also
shows that the presence of individual fixed effects and heteroscedasticity also
impacts the performance of the EW estimator, particularly so if they are both
present in the data process.
We perform a series of Monte Carlo simulations to compare the performance of
the EW and IV estimators in finite samples. Emulating the types of environments
commonly found by empirical researchers, we set up a panel data model with
individual fixed effects and potential heteroscedasticity in the errors. Monte Carlo
experiments enable us to study those estimators in a “controlled environment,”
where we can investigate the role played by each element (or assumption) of an
estimator in evaluating its performance. Our simulations compare the EW and IV
(OLS-IV and AB-GMM) estimators in terms of bias and root mean squared error
(RMSE), a standard measure of efficiency.
We consider several distributional assumptions to generate observations and
errors. Experimenting with multiple distributions is important because researchers
often find a variety of distributions in real-world applications and because one
ultimately does not observe the distribution of the mismeasurement term. Since the
EW estimator is built around the notion of skewness of the relevant distributions,
we experiment with three skewed distributions (lognormal, chi-square, and
F-distribution), using the standard normal (non-skewed) as a benchmark. The
simulations also allow for significant correlation between mismeasured and well-
measured regressors (as in Erickson and Whited 2000, 2002), so that the attenuation
bias of the mismeasured regressor affects the coefficient of the well-measured
regressor.
Our simulation results can be summarized as follows. First, we examine the
identification test proposed by Erickson and Whited (2002). This is a test that
the data contain a sufficiently high degree of skewness to allow for the
identification of their model. We study the power of the EW identification
test by generating data that do not satisfy its null hypothesis of non-skewness.
In this case, even for the most skewed distribution (lognormal), the test rejects
the null hypothesis only 47 % of the time – this is far less than desirable, given
that the null is false. The power of the test becomes even weaker after we treat
the data for the presence of fixed effects in the true model (“within transfor-
mation”). In this case, the rejection rate under the lognormal distribution drops
to 43 %. The test’s power declines even further when we consider alternative
skewed distributions (chi-square and F-distributions). The upshot of this first
set of experiments is that the EW model too often rejects data that are
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1567

generated to fit its identifying assumptions. These findings may help explain
some of the difficulties previous researchers have reported when attempting to
implement the EW estimator.
We then study the bias and efficiency of the EW and IV estimators. Given that
the true model contains fixed effects, it is appropriate to apply the within transfor-
mation to the data. However, because most empirical implementations of the EW
estimator have used data in “level form” (i.e., not treated for the presence of fixed
effects),4 we also experiment with cases in which we do not apply the within
transformation. EW propose three estimators that differ according to the number
of moments used: GMM3, GMM4, and GMM5. We consider all of them in our
experiments.
In a first round of simulations, we impose error homoscedasticity. When we
implement the EW estimator with the data in level form, we find that the
coefficients returned are significantly biased even when the data have a high
degree of skewness (i.e., under the lognormal case, which is EW’s preferred
case). Indeed, for the mismeasured regressors the EW biases are in excess of
100 % of their true value. As should be expected, the performance of the EW
estimator improves once the within transformation is used. In the case of the
lognormal distribution, the EW estimator bias is relatively small. In addition,
deviations from the lognormal assumption tend to generate significant biases for
the EW estimator.
In a second round of simulations, we allow for heteroscedasticity in the data.
We focus our attention on simulations that use data that are generated using
a lognormal distribution after applying the within transformation (the best case
scenario for the EW estimator). Heteroscedasticity introduces heterogeneity to the
model and consequently to the distribution of the partialled out dependent
variable, compromising identification in the EW framework. The simulations
show that the EW estimator is biased and inefficient for both the mismeasured
and well-measured regressors. In fact, biases emerge even for very small amounts
of heteroscedasticity, where we find biases of approximately 40 % for the
mismeasured regressor. Paradoxically, biases “switch signs” depending on the
degree of heteroscedasticity that is allowed for in the model. For instance, for
small amounts of heteroscedasticity, the bias of the mismeasured regressor is
negative (i.e., the coefficient is biased downwards). However, the bias turns
positive for a higher degree of heteroscedasticity. Since heteroscedasticity is
a naturally occurring phenomenon in corporate data, our simulations imply that
empirical researchers might face serious drawbacks when using the EW
estimator.
Our simulations also show that, in contrast to the EW estimator, the bias in the
IV estimates is small and insensitive to the degree of skewness and heterosce-
dasticity in the data. Focusing on the OLS-IV estimator, we consider the case of
time-invariant correlation in the error structure and use the second lag of the

4
Examples are Whited (2001, 2006), Hennessy (2004), and Colak and Whited (2007).
1568 H. Almeida et al.

observed mismeasured variable as an instrument for its current (differenced) value.5

We also allow the true value of the mismeasured regressor to have a moderate
degree of autocorrelation. Our results suggest that the OLS-IV estimator renders
fairly unbiased estimates. In general, that estimator is also distinctly more efficient
than the EW estimator.
We also examine the OLS-IV estimator’s sensitivity to the autocorrelation
structures of the mismeasured regressor and the measurement error. First, we
consider variations in the degree of autocorrelation in the process for the true
value of the mismeasured regressor. Our simulations show that the IV bias is
largely insensitive to variations in the autoregressive (AR) coefficient (except for
extreme values of the AR coefficient). Second, we replace the assumption of time-
invariant autocorrelation in the measurement error with a moving average
(MA) structure. Our simulations show that the OLS-IV bias remains small if one
uses long enough lags of the observable variables as instruments. In addition,
provided that the instrument set contains suitably long lags, the results are robust
to variations in the degree of correlation in the MA process. As we discuss below,
understanding these features (and limitations) of the IV approach is important given
that the researcher will be unable to pin down the process followed by the
measurement-error process.
To illustrate the performance of these alternative estimators on real data, in the
final part of our analysis, we estimate empirical investment models under the EW
and IV frameworks. Concerns about measurement errors have been emphasized in
the context of the empirical investment model introduced by Fazzari et al. (1988),
where a firm’s investment is regressed on a proxy for investment demand
(Tobin’s q) and cash flows. Theory suggests that the correct proxy for a firm’s
investment demand is captured by marginal q, but this quantity is unobservable and
researchers use instead its measurable proxy, average q. Since the two variables are
not the same, a measurement problem naturally arises (Hayashi 1982; Poterba
1988). Following Fazzari et al. (1988), investment-cash flow sensitivities became
a standard metric in the literature that examines the impact of financing imperfec-
tions on corporate investment (Stein 2003). These empirical sensitivities are also
used for drawing inferences about efficiency in internal capital markets (Lamont
1997; Shin and Stulz 1998), the effect of agency on corporate spending (Hadlock
1998; Bertrand and Mullainathan 2005), the role of business groups in capital
allocation (Hoshi et al. 1991), and the effect of managerial characteristics on
corporate policies (Bertrand and Schoar 2003; Malmendier and Tate 2005).
Theory does not pin down exact values for the expected coefficients on q and
cash flow in an investment model. However, two conditions would seem reasonable
in practice. First, given that the estimator is addressing measurement error in q that
may be “picked up” by cash flow (joint effects of attenuation bias and regressor

5
The results for the Arellano–Bond GMM estimator are similar to those of the OLS-IV estimator.
To save space and because the OLS-IV estimator is easier to implement, we focus on this
estimator.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1569

covariance), we should expect the q coefficient to go up and the cash flow coeffi-
cient to go down, when compared to standard (likely biased) OLS estimates.
Second, we would expect the q and cash flow coefficients to be nonnegative after
addressing the problem of mismeasurement. If the original q-theory of investment
holds and the estimator does a good job of addressing mismeasurement, then the
cash flow coefficient would be zero. Alternatively, the cash flow coefficient could
be positive because of financing frictions.6
Using data from Compustat from 1970 to 2005, we estimate investment equa-
tions in which investment is regressed on proxies for q and cash flow. Before doing
so, we perform standard tests to check for the presence of individual fixed effects
and heteroscedasticity in the data. In addition, we perform the EW identification
test to check whether the data contain a sufficiently high degree of skewness.
Our results are as follows. First, our tests reject the hypotheses that the data do
not contain firm-fixed effects and that errors are homoscedastic. Second, the EW
identification tests indicate that the data fail to display sufficiently high skewness.
These initial tests suggest that the EW estimator is not suitable for standard
investment equation applications. In fact, we find that, when applied to the data,
the EW estimator returns coefficients for q and cash flow that are highly unstable
across different years. Moreover, following the EW procedure for panel models
(which comprises combining yearly cross-sectional coefficients into single esti-
mates), we obtain estimates for q and cash flow that do not satisfy the conditions
discussed above. In particular, EW estimators do not reduce the cash flow coeffi-
cient relative to that obtained by standard OLS, while the q coefficient is never
statistically significant. In addition, those estimates are not robust with respect to
the number of moments used: EW’s GMM3, GMM4, and GMM5 models procedure
results that are inconsistent with one another. These results suggest that the pres-
ence of heteroscedasticity and fixed effects in real-world investment data hampers
identification when using the EW estimator.
In contrast to EW, the OLS-IV procedure yields estimates that are fairly sensi-
ble. The q coefficient goes up by a factor of 3–5, depending on the set of instruments
used. At the same time, the cash flow coefficient goes down by about two-thirds of
the standard OLS value. Similar conclusions apply to the AB-GMM estimator. We
also examine the robustness of the OLS-IV to variations in the set of instruments
used in the estimation, including sets that feature only longer lags of the variables in
the model. The OLS-IV coefficients remain fairly stable after such changes. These
results suggest that real-world investment data likely satisfies the assumptions that
are required for identification of IV estimators.
The remainder of the chapter is structured as follows. We start the next section
discussing in detail the EW estimator, clarifying the assumptions that are needed for
its implementation. Subsequently, we show how alternative IV models deal with

6
See Hubbard (1998) and Stein (2003) for comprehensive reviews. We note that the presence of
financing frictions does not necessarily imply that the cash flow coefficient should be positive. See
Chirinko (1993) and Gomes (2001) for arguments suggesting that financing frictions are not
sufficient to generate positive cash flow coefficients.
1570 H. Almeida et al.

the errors-in-variables problem. In Sect. 57.3, we use Monte Carlo simulations to

examine the performance of alternative estimators in small samples and when we
relax the assumptions that are required for identification. In Sect. 57.4, we take our
investigation to actual data, estimating investment regressions under the EW and IV
frameworks. Section 57.5 concludes the chapter.

57.2 Dealing with Mismeasurement: Alternative Estimators

57.2.1 The Erickson–Whited Estimator

In this section, we discuss the estimator proposed in companion papers by Erickson

and Whited (2000, 2002). Those authors present a two-step generalized method of
moments (GMM) estimator that exploits information contained in the high-order
moments of residuals obtained from perfectly measured regressors (similar to
Cragg 1997). We follow EW and present the estimator using notation of cross-
section estimation. Let (yi, zi, xi), i ¼ 1,. . .,n, be a sequence of observable vectors,
where xi  (xi1,. . .,xiJ) and zi  (1, zi1,. . ., ziL). Let (ui, wi, ei) be a sequence of
unobservable vectors, where wi  (wi1,. . .,wiJ) and ei  (ei1,. . .,eiJ). Consider the
following model:
yi ¼ zi a þ wi b þ ui (57.1)

where yi is the dependent variable, zi is a perfectly measured regressor, wi is

a mismeasured regressor, ui is the innovation of the model, and a  (a0, ai,
. . .,aL)0 and b  (b1, . . .,bJ)0 . The measurement error is assumed to be additive
such that
x i ¼ wi þ e i (57.2)

where xi is the observed variable and ei is the measurement error. The observed
variables are yi, zi, and xi; and by substituting Eq. 57.2 in Eq. 57.1, we have

yi ¼ zi a þ xi b þ v i ,

where vi ¼ ui  eib. In the new regression, the observable variable xi is correlated

with the innovation term vi, causing the coefficient of interest, b, to be biased.
To compute the EW estimator, it is necessary to first partial out the effect of the
well-measured variable, zi, in Eqs. 57.1 and 57.2 and rewrite the resulting expres-
sions in terms of residual populations:

yi  zi my ¼
i b þ ui (57.3)

xi  zi mx ¼
i þ ei , (57.4)
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1571

where (my, mx, mw) ¼ [E(z0izi)]1E[z0i(yi, xi, wi)] and
i  wi  zimw. For the details of
this derivation, see Erickson and Whited (2002, p. 779). One can then consider
a two-step estimation approach, where the first step is to substitute least squares

 Xn 0 1 Xn 0
estimates m^y ; m^x  zz
i¼1 i i
z ðy ; x Þ into Eqs. 57.3 and 57.4 to
i¼1 i i i
obtain a lower dimensional errors-in-variables model. The second step consists
of estimating b using GMM using high-order sample moments of yi  zi m^y and
xi  zi m^x . Estimates of a are then recovered via my ¼ a + mxb. Thus, the estimators
are based on equations giving the moments of yi  zimy and xi  zimx as functions of
b and the moments of (ui, ei,
i).
To give a concrete example of how the EW estimator works, we explore the case
of J ¼ 1. The more general case is discussed below. By substituting

!1
X
n X
n
m^y ; m^x  z0i zi zi 0 ðyi ; xi Þ
i¼1 i¼1

into Eqs. 57.1 and 57.2, one can estimate b, E(u2i ), E(e2i ), and E(
2i ) via
GMM. Estimates of the lth element of a are obtained by substituting the estimate
of b and the lth elements of m^y and m^x into

al ¼ myl  mxl b, l 6¼ 0:

There are three second-order moment equations:

h
2 i


E yi  zi m y ¼ b2 E
2i þ E u2i (57.5)


 

E yi  zi my ðxi  zi mx Þ ¼ bE
2i (57.6)
h i


E ðxi  zi mx Þ2 ¼ E
2i þ E e21 : (57.7)

The left-hand side quantities are consistently estimable, but there are only three
equations with which to estimate four unknown parameters on the right-hand side.
The third-order product moment equations, however, consist of two equations in
two unknowns:
h
2 i

E yi  zi my ðxi  zi mx Þ ¼ b2 E
3i , (57.8)

h
 i

E yi  zi my ðxi  zi mx Þ2 ¼ bE
3i : (57.9)

It is possible to solve these two equations for b. Crucially, a solution exists if the
identifying assumptions b 6¼ 0 and E(
3i ) 6¼ 0 are true, and one can test the contrary
1572 H. Almeida et al.

hypothesis (i.e., b 6¼ 0 and/or E(
3i ) ¼ 0) by testing whether their sample counter-

parts are significatively different from zero.
Given b, Eqs. 57.5, 57.6, 57.7, and 57.9 can be solved for the remaining right-
hand side quantities. One obtains an overidentified equation system by combining
Eqs. 57.5, 57.6, 57.7, 57.8, and 57.9 with the fourth-order product moment equa-
tions, which introduce only one new quantity, E(
4i ):
h
3 i



E yi  zi my ðxi  zi mx Þ ¼ b3 E
4i þ 3E
2i E u2i , (57.10)

h
2 i 



E yi  zi my ðxi  zi mx Þ2 ¼ b2 E
4i þ E
2i E u2i




þ E u2i E
2i þE e2i , (57.11)
h
 i 



E yi  zi my ðxi  zi mx Þ3 ¼ b E
4i þ 3E
2i E e2i : (57.12)

The resulting eight-equation system Eqs. 57.5, 57.6, 57.7, 57.8, 57.9, 57.10,
57.11, and 57.12 contains the six unknowns (b, E(u2i ), E(e2i ), E(
2i ), E(
3i ), E(
4i )).
It is possible to estimate this vector by numerically minimizing a quadratic form
that minimizes asymptotic variance.
The conditions imposed by EW imply restrictions on the residual moments of
the observable variables. Such restrictions can be tested using the corresponding
sample moments. EW also propose a test for residual moments that is based on
several assumptions.7 These assumptions imply testable restrictions on the residuals
from the population regression of the dependent and proxy variables on the
perfectly measured regressors. Accordingly, one can develop Wald-type partially
adjusted statistics and asymptotic null distributions for the test. Empirically, one
can use the Wald test statistic and critical values from a chi-square distribution to
test whether the last moments are equal to zero. This is an identification test, and if
in a particular application one cannot reject the null hypothesis, then the model is
unidentified and the EW estimator may not be used. We study the finite sample
performance of this test and its sensitivity to different data-generating processes in
the next section.
It is possible to derive more general forms of the EW estimator. In particular, the
EW estimators are based on the equations for the moments of yi  zimy and xi  zimx
as functions of b and the moments ui, ei, and
i. To derive these equations, write
Eq. 57.3 as yi  zimy ¼ ∑ jJ ¼ 1
ijbj + ui, where J is the number of well-measured
regressors and the jth equation in Eq. 57.4 as xij  zimxj ¼
ij + eij, where mxj is the jth
column of and mx and (
ij, eij) is the jth row of (
0ij, e0ij). Next write

7
First, the measurement errors, the equation error, and all regressors have finite moments of
sufficiently high order. Second, the regression error and the measurement error must be indepen-
dent of each other and of all regressors. Third, the residuals from the population regression of the
unobservable regressors on the perfectly measured regressors must have a nonnormal distribution.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1573

" !r0 #
h
r0 YJ i X
J YJ
rj rj
E yi  zi my j¼1
ðxi  zi mx Þ ¼E
i b þ ui j¼1
ð
i þ ei Þ ,
j¼1

(57.13)

where (r0, r1,. . ., rJ) are nonnegative integers. After expanding the right-hand side of
Eq. 57.13, using the multinomial theorem, it is possible to write the above moment
condition as

E½gi ðmÞ ¼ cðyÞ,

r
where m ¼ vec(my, mx), gi(m) is a vector of distinct elements of the form yi  zi my 0
YJ
ðx  zi mx Þrj , cðyÞ contains the corresponding expanded version of the right-
j¼1 i
hand side of Eq. 57.13, and y is a vector containing the elements of b and the
moments of (ui, ei,
i). The GMM estimator y is defined as

y^ ¼ arg min ðgi ðm^Þ  cðtÞÞ W

0
^ ðgi ðm^Þ  cðtÞÞ,
t2Y

Xn
where g- i ðsÞ  ^ is a positive definite matrix. Assuming a
g ðsÞ for all s, and W
t¼1 i
8
number of regularity conditions, the estimator is consistent and asymptotically normal.
It is important to notice that the estimator proposed by Erickson and Whited
(2002) was originally designed for cross-sectional data. To accommodate a panel-
like structure, Erickson and Whited (2000) propose transforming the data before the
estimation using the within transformation or differencing. To mimic a panel
structure, the authors propose the idea of combining different cross-sectional
GMM estimates using a minimum distance estimator (MDE).
The MDE estimator is derived by minimizing the distance between the auxiliary
parameter vectors under the following restrictions:
 
f b; y^ ¼ Hb  y^ ¼ 0,

where the R · K  K matrix H imposes (R  1) · K restrictions on y. The R K  1

vector y^ contains the R stacked auxiliary parameter vectors, and b is the parameter
of interest. Moreover, H is defined by an R · K  K – dimensional stacked identity
matrix.

8
More specifically, these conditions are as follows: (zi, wi, ui, ei) is an independent and identically
distributed sequence; ui and the elements of zi, wi, and ei, have finite moments of every order; (ui, ei)
is independent of (zi, wi), and the individual elements in (ui, ei) are independent of each other;
E(ui) ¼ 0 and E(ei) ¼ 0; E[(zi, wi)0 (zi, wi)] is positive definite; every element of b is nonzero; and the
distribution of
satisfies E[(
ic)3] 6¼ 0 for every vector of constants c ¼ (c1,  ,cJ) having at least
one nonzero element.
1574 H. Almeida et al.

The MDE is given by the minimization of

 0 h i1  
DðbÞ ¼ f b; y^ V^ y^ f b; y^ , (57.14)

^ is the common estimated variance–covariance matrix of the auxiliary

where V^ ½y
parameter vectors.
In order to implement the MDE, it is necessary to determine the covariances
between the cross-sections being pooled. EW propose to estimate the covariance by
using the covariance between the estimators’ respective influence functions.9 The
procedure requires that each cross-section have the same sample size, that is, the
panel needs to be balanced.
Thus, minimization of D in Eq. 57.14 leads to
 h i1 1 0 h i1
b^ ¼ H V^ y^ H ^
H V^ y^ y,
0

with variance–covariance matrix:

h i  0 h i1 1
V^ b^ ¼ H V^ y^ H :

H is a vector in which R is the number of GMM estimates available (for each

time period) and K ¼ 1, y^ is a vector containing allh the
i EW estimates for each
period, and b is the MDE of interest. In addition, V y^ is a matrix carrying the
estimated variance–covariance matrices of the GMM parameter vectors.

57.2.2 An OLS-IV Framework

In this section, we revisit the work of Griliches and Hausman (1986) and Biorn
(2000) to discuss a class of OLS-IV estimators that can help address the errors-in-
variables problem.
Consider the following single-equation model:

yit ¼ gi þ wit b þ uit, i ¼ 1, . . . N, t ¼ 1, . . . , T, (57.15)

where uit is independently and identically distributed, with mean zero and variance
s2u, and Cov(wit, uis) ¼ Cov(gi, uis) ¼ 0 for any t and s, but Cov(gi, wit) 6¼ 0, y is an
observable scalar, w is a 1  K vector, and b is K  1 vector. Suppose we do not
observe wit itself, but rather the error-ridden measure:

9
See Erickson and Whited (2002) Lemma 1 for the definition of their proposed influence function.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1575

xit ¼ wit þ eit , (57.16)

where Cov(wit, eit) ¼ Cov(gi, eit) ¼ Cov(uis, eit) ¼ 0, Var(eit) ¼ s2e ,

Cov(eit, eit  1) ¼ ges2e , and e is a 1  K vector. If we have a panel data with
T > 3, by substituting Eq. 57.16 in Eq. 57.15, we can take first differences of the
data to eliminate the individual effects gi and obtain

yit  yit1 ¼ ðxit  xit1 Þb þ ½ðuit  eit bÞ  ðuit1  eit1 bÞ (57.17)

Because of the correlation between the mismeasured variable and the

innovations, the coefficient of interest is known to be biased.
Griliches and Hausman (1986) propose an instrumental variable approach to
reduce the bias. If the measurement error eit is i.i.d. across i and t, and x is serially
correlated, then, for example, xit2, xit3, or (xit2  xit3) are valid as instruments
for (xit  xit1). The resulting instrumental variable estimator is consistent even
though T is finite and N might tend to infinity.
As emphasized by Erickson and Whited (2000), for some applications, the
assumption of i.i.d. measurement error can be seen as too strong. Nonetheless, it
is possible to relax this assumption to allow for autocorrelation in the measurement
errors. While other alternatives are available, here we follow the approach
suggested by Biorn (2000).10
Biorn (2000) relaxes the i.i.d. condition for innovations in the mismeasured
equation and proposes alternative assumptions under which consistent IV estima-
tors of the coefficient of the mismeasured regressor exists. Under those assump-
tions, as we will show, one can use the lags of the variables already included in the
model as instruments. A notable point is that the consistency of these estimators is
robust to potential correlation between individual heterogeneity and the latent
regressor.
Formally, consider the model described in Eqs. 57.15 and 57.16 and assume that
(wit, uit, eit, gi) are independent across individuals. For the necessary orthogonality
assumptions, we refer the reader to Biorn (2000), since these are quite standard.
More interesting are the assumptions about the measurement errors and distur-
bances. The standard Griliche–Hausman’s assumptions are
(A1) E(e0iteiy) ¼ 0KK, t 6¼ y,
(A2) E(uituiy) ¼ 0, t 6¼ y
which impose non-autocorrelation on innovations. It is possible to relax these
assumptions in different ways. For example, we can replace (A1) and (A2) with
(B1) E(e0iteiy) ¼ 0KK, jt  yj > t,
(B2) E(uituiy) ¼ 0, jt  yj > t,

10
A more recent paper by Xiao et al. (2008) also shows how to relax the classical Griliches–
Hausman assumptions for measurement error models.
1576 H. Almeida et al.

This set of assumptions is weaker since (B1) and (B2) allow for a vector moving
average (MA) structure up to order t (1) for the innovations. Alternatively, one
can use the following assumptions:
(C1) E(e0itety) is invariant to t, y, t 6¼ y.
(C2) E(uituiy) is invariant to t, y, t 6¼ y.
Assumptions (C1) and (C2) allow for a different type of autocorrelation,
more specifically they allow for any amount of autocorrelation that is time
invariant. Assumptions (C1) and (C2) will be satisfied if the measurement
errors and the disturbances have individual components, say eit ¼ e1i + e2it,
uit ¼ u1i + u2it, where e1i, e2it, u1i, u2it i.i.d. Homoscedasticity of eit and/or uit across
i and t need not be assumed; the model accommodates various forms of
heteroscedasticity.
Biorn also considers assumptions related to the distribution of the latent regressor
vector wit:
(D1) E(wit) is invariant to t.
(D2) E(giwit) is invariant to t.
Assumptions (D1) and (D2) hold when wit is stationary. Note that wit and i need
not be uncorrelated.
To ensure identification of the slope coefficient vector when panel data are
available, it is necessary to impose restrictions on the second-order moments of
the variables (wit, uit, eit, gi). For simplicity, Biorn assumes that this distribution is
the same across individuals and that the moments are finite. More specifically,
C(wit, wit) ¼ ∑ ww wn 0 ee uu 2


ty , E(wit
i) ¼ ∑ t , E(eiteit) ¼ ∑ty, E(uituit) ¼ sty , E(
i ) ¼ s , where
C denotes the covariance matrix operator. Then, it is possible to derive the
second-order moments of the observable variables and show that they only depend
on these matrices and the coefficient b.11 In this framework, there is no need to use
assumptions based on higher-order moments.
Biorn proposes several strategies to estimate the slope parameter of interest.
Under the OLS-IV framework, he proposes estimation procedures of two kinds:
• OLS-IV A: The equation is transformed to differences to remove individual
heterogeneity and is estimated by OLS-IV. Admissible instruments for
this case are the level values of the regressors and/or regressands for other
periods.
• OLS-IV B: The equation is kept in level form and is estimated by OLS-IV.
Admissible instruments for this case are differenced values of the regressors
and/or regressands for other periods.
Using moment conditions from the OLS-IV framework, one can define
the estimators just described. In particular, using the mean counterpart and the
moment conditions, one can formally define the OLS-IV A and OLS-IV B
estimators.

Formally, one can show that C(xit, xiy) ¼ ∑ ww

11 ee ww w

ty + ∑ ty, E(xit, yiy) ¼ ∑ ty b + ∑ t , and E(yit, yiy)
¼ b0 ∑ ww
ty b + ∑ xn
t b + b H0
(∑ xn 0
y ) + s uu
ty + s


.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1577

In particular, the estimator for OLS-IV A can be defined as

" #1 " #
XN X
N
b^ ¼
xpðtyÞ
0
x ðDxity Þ
ip
0
x ðDy Þ ,
ip ity
i¼1 i¼1

where (t, y, p) are indices. Let the dimension of b be defined by K. If K ¼ 1, it is

possible to define the following estimator for a given (t, y, p):
" #1 " #
X
N X
N
b^ ¼
ypðtyÞ y ðDxity Þ
ip y ðDy Þ :ip ity
i¼1 i¼1

If K > 1, the latter estimator is infeasible, but it is possible to modify the former
estimator by replacing one element in x0ip by yip.
The estimator for OLS-IV B (equation in level and instruments in difference) can
be defined as
" #1 " #
XN
0 XN
0
b^ ¼
xðpqÞt Dxipq xit Dxipq y : it
i¼1 i¼1

As in the previous case, if the dimension of b, K is equal to 1, it is possible to

define the following estimator for (t, p, q):
" #1 " #
XN
 X
N

^
b ¼ Dy xit Dy y :
yðpqÞt ipq ipq it
i¼1 i¼1

If K > 1, the latter estimator is infeasible, but it is possible to modify the former
estimator by replacing one element in Dxip by Dyip.
For some applications, it might be useful to impose weaker conditions on the
autocorrelation of measurement errors and disturbances. In this case, it is necessary
to restrict slightly further the conditions on the instrumental variables. More
formally, if one replaces assumptions (A1) and (A2), or (C1) and (C2), by the
weaker assumptions (B1) and (B2), then it is necessary to ensure that the IV set has
a lag of at least t  2 and/or lead of at least t + 1 periods of the regressor in order to
“clear” the t period memory of the MA process. Consistency of these estimators is
discussed in Biorn (2000).12
To sum up, there are two simple ways to relax the standard assumption of i.i.d.
measurement errors. Under the assumption of time-invariant autocorrelation, the set of
instruments can contain the same variables used under Griliches–Hausman.

12
In particular, if jt  pj, jy  pj > t, then (B1) and rank (E[w0ip(Dwity)]) ¼ K for some
p 6¼ t 6¼ 0 ensure consistency of OLS–IV B, b^xpðtyÞ , and (B2) and the same rank condition ensure
consistency of b^xyðtyÞ . In the same way, if jp  tj, q  t > t, (B1), (D1), (D2), and rank
(E[(Dwipq)0 wit)]) ¼ K for some p 6¼ q 6¼ t ensure consistency of OLS-IV B, b^xðpqÞt , and (B2),
(D1), (D2), and the same rank condition ensure consistency of b^yðpqÞt .
1578 H. Almeida et al.

For example, if one uses the OLS-IV A estimator (equation in differences and instru-
ments in levels), then twice-lagged levels of the observable variables can be used as
instruments. Under a moving average structure for the innovations in the measurement
error [assumptions (B1) and (B2)], identification requires the researcher to use longer
lags of the observable variables as instruments. For example, if the innovations follow
an MA(1) structure, then consistency of the OLS-IV A estimator requires the use of
instruments that are lagged three periods and longer. Finally, identification requires the
latent regressor to have some degree of autocorrelation (since lagged values are used as
instruments). Our Monte Carlo simulations will illustrate the importance of these
assumptions and will evaluate the performance of the OLS-IV estimator under
different sets of assumptions about the structure of the errors.

57.2.3 GMM Estimator

Within the broader instrumental variable approach, we also consider an entire class
of GMM estimators that deal with mismeasurement. These GMM estimators are
close to the OLS-IV estimator discussed above but may attain appreciable gains in
efficiency by combining numerous orthogonality conditions [see Biorn (2000) for
a detailed discussion]. GMM estimators that use all the available lags at each period
as instruments for equations in first differences were proposed by Holtz-Eakin
et al. (1988) and Arellano and Bond (1991). We provide a brief discussion in turn.
In the context of a standard investment model, Blundell et al. (1992) use GMM
allowing for correlated firm-specific effects, as well as endogeneity
(mismeasurement) of q. The authors use an instrumental variable approach on
a first-differenced model in which the instruments are weighted optimally so as to
form the GMM estimator. In particular, they use qit2 and twice-lagged investments
as instruments for the first-differenced equation for firm i in period t. The Blundell,
Bond, Devereux, and Schiantarelli estimators can be seen as an application of the
GMM instrumental approach proposed by Arellano and Bond (1991), which was
originally applied to a dynamic panel.
A GMM estimator for the errors-in-variables model of Eq. 57.17 based on IV
moment conditions takes the form
h   0 i1  0   0 
b^ ¼
0
Dx Z V 1
N Z Dx Dx Z V 1
N Z Dy ,

where Dx is the stacked vector of observations on the first difference of the

mismeasured variable and Dy is the stacked vector of observations on the first
difference of the dependent variable. As in Blundell et al. (1992), the instrument
matrix Z has the following form13:

13
In models with exogenous explanatory variables, Zi may consist of sub-matrices with the block
diagonal (exploiting all or part of the moment restrictions), concatenated to straightforward
one-column instruments.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1579

0 1
x1 0 0  0  0
B 0 x1 x2  0  0 C
Zi ¼ B
@⋮ ⋮
C:
⋮ ⋮ ⋮ ⋮ ⋮ A
0 0 0    x1    xT2

According to standard GMM theory, an optimal choice of the inverse weight

matrix VN is a consistent estimate of the covariance matrix of the orthogonality
conditions E(Z0iDviDv0iZi), where Dvi are the first-differenced residuals of each
XN
individual. Accordingly, a one-step GMM estimator uses V^ ¼
0
Z0 DD Zi ,
i1 i
where D is the first-difference matrix operator. A two-step GMM estimator uses
XN
e¼
a robust choice V ZD^v i D^
v 0Z i where D^
v i are one-step GMM residuals.
i¼1 i
Biorn (2000) proposes estimation of linear, static regression equations from
panel data models with measurement errors in the regressors, showing that if the
latent regressor is autocorrelated or nonstationary, several consistent OLS-IV and
GMM estimators exist, provided some structure is imposed on the disturbances and
measurement errors. He considers alternative GMM estimations that combine all
essential orthogonality conditions. The procedures are very similar to the one
described just above under non-autocorrelation in the disturbances. In particular,
the required assumptions when allowing autocorrelation in the errors are very
similar to those discussed in the previous section. For instance, when one allows
for an MA(t) structure in the measurement error, for instance, one must ensure
that the variables in the IV matrix have a lead or lag of at least t + 1 periods to the
regressor.
We briefly discuss the GMM estimators proposed by Biorn (2000). First,
consider estimation using the equation in differences and instrumental variables
in levels. After taking differences of the model, there are (T  1) + (T + 1) equations
that can be stacked for individual i as

2 3 2 3 2 3
Dyi21 Dxi21 Dei21
6 Dyi32 7 6 Dxi32 7 6 Dei32 7
6 7 6 7 6 7
6 ⋮ 7 6 ⋮ 7 6 ⋮ 7
6 7 6 7 6 7
6 Dyi , T, T  1 7 6 Dxi T, T  1 7 6 Dei , T, T  1 7
6 7¼6 7b þ 6 7,
6 Dyi31 7 6 Dxi31 7 6 Dei31 7
6 7 6 7 6 7
6 Dyi42 7 6 Dxi42 7 6 Dei42 7
6 7 6 7 6 7
4 ⋮ 5 4 ⋮ 5 4 ⋮ 5
Dyi , T, T  2 Dxi , T, T  2 Dei , T, T  2

or compactly

Dyi ¼ DXi b þ D2i :

The IV matrix is the ((2T  3)  KT(T  2)) diagonal matrix with the
instruments in the diagonal defined by Z. Let
1580 H. Almeida et al.

h 0 0
i0 h 0 0
i0
Dy ¼ ðDy1 Þ ; . . . ; ðDyN Þ , D2 ¼ ðD21 Þ ; . . . ; ðD2N Þ
h 0 0
i0  0
DX ¼ ðDX1 Þ ; . . . ; ðDXN Þ , Z ¼ Z01 ; . . . ; Z0N :

The GMM estimator that minimizes [N1(D2)0 Ζ](N2V)1[N1Z0 (D2)] for V ¼ Z0 Z

can be written as
2" #" #1 " #31
X 0 X X
b^Dx ¼ 4 ðDXi Þ Zi Z 0i Z i Z 0i ðDXi Þ 5
i i i
2" #" #1 " #3
X 0 X X
4 ðDXi Þ Zi Z0 Zi i Z0 ðDyi Þ 5: i
i i i

If D2 has a non-scalar covariance matrix, a more efficient GMM estimator, e b Dx,

can be obtained setting V ¼ VZ(D2) ¼ E[Z0 (D2)(D2)0 Z] and estimating V^ zðD2Þ by

V^ ZðD2Þ 1 X 0  c  c 0
¼ Z D2 D2 Z,
N N i
where D2 c i ¼ Dy  ðDXi Þb^ . This procedure assumes that (A1) and (A2) are
i Dx
satisfied. However, as Biorn (2000) argues, one can replace them by (B1) or
(B2) and then ensure that the variables in the IV matrix have a lead or lag of at
least t + 1 periods to the regressor, to “get clear of” the t period memory of the
MA(t) process. The procedure described below is also based on the same set of
assumptions and can be extended similarly.14
The procedure for estimation using equation in levels and IVs in difference is
similar. Consider the T stacked level equations for individual i:
2 3 2 3 2 3 2 3
yi1 c xi1 2i1
4 ⋮ 5 ¼ 4 ⋮ 5 þ 4 ⋮ 5b þ 4 ⋮ 5,
yiT c xiT 2iT

or more compactly,

yi ¼ eT c þ Xi b þ 2,

where eΤ denotes a (T  1) vector of ones. Let the (T  T(T  2)K) diagonal matrix
of instrument be denoted by DZi. This matrix has the instruments in difference in
the main diagonal. In addition,

14
See Propositions 1* and 2* in Biorn (2000) for a formal treatment of the conditions.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1581

define:
 0  0
y ¼ y01 ; . . . ; y0N , 2 ¼ 201 ; . . . ; 20N
 0 h 0 0
i0
X ¼ X01 ; . . . ; X0N , DZ ¼ ðDZ 1 Þ ; . . . ; ðDZN Þ :

The GMM estimator that minimizes [N120 (DZi)0 ]|(N2VD)1[N1(DZ)0 2]

for VD ¼ (DZ)0 (DZ) is
2" #" #1 " #31
X X X
b^Lx ¼ 4 X0i ðDZi Þ ðDZ i Þ0 DZi ðDZi Þ0 Xi 5
i i i
2" #" #1 " #3
X 0 X X
4 Xi ðDZi Þ ðDZ i Þ0 DZi ðDZi Þ0 yi 5:
i i i

If 2 has a non-scalar covariance matrix, a more efficient GMM estimator, eb Lx ,

can be obtained setting VD ¼ V(DZ)2 ¼ E[(DZ)220 (DZ)] and estimating V^ ðDZÞ2 by

V^ ðDZÞ2 1 X 0 0
¼ ðDZÞ 2^ 2^ ðDZÞ,
N N i
where 2^ ¼ yi  Xi b^Lx .
Finally, let us briefly contrast the OLS-IV and AB-GMM estimators. The advantages
of GMM over IV are clear: if heteroscedasticity is present, the GMM estimator is more
efficient than the IV estimator, while if heteroscedasticity is not present, the GMM
estimator is no worse asymptotically than the IV. Implementing the GMM estimator,
however, usually comes with a high price. The main problem, as Hayashi (2000, p. 215)
points out, concerns the estimation of the optimal weighting matrix that is at the core of the
GMM approach. This matrix is a function of fourth moments, and obtaining reasonable
estimates of fourth moments requires very large sample sizes. Problems also arise when
the number of moment conditions is high, that is, when there are “too many instruments.”
This latter problem affects squarely the implementation of the AB-GMM, since it relies
on large numbers of lags (especially in long panels). The upshot is that the efficient GMM
estimator can have poor small sample properties [see Baum et al. (2003) for a discussion].
These problems are well documented and remedies have been proposed by, among
others, Altonji and Segal (1996) and Doran and Schmidt (2006).

57.3 Monte Carlo Analysis

We use Monte Carlo simulations to assess the finite sample performance of the EW
and IV estimators discussed in Sect. 57.2. Monte Carlo simulations are an ideal
experimental tool because they enable us to study those two estimators in
a controlled setting, where we can assess and compare the importance of elements
1582 H. Almeida et al.

that are key to estimation performance. Our simulations use several distributions to
generate observations. This is important because researchers will often find
a variety of distributions in real-world applications and because one ultimately
does not see the distribution of the mismeasurement term. Our Monte Carlos
compare the EW, OLS-IV, and AB-GMM estimators presented in Sect. 57.2 in
terms of bias and RMSE.15 We also investigate the properties of the EW identifi-
cation test, focusing on the empirical size and power of this test.

57.3.1 Monte Carlo Design

A critical feature of panel data models is the observation of multiple data points
from the same individuals over time. It is natural to consider that repeat samples are
particularly useful in that individual idiosyncrasies are likely to contain information
that might influence the error structure of the data-generating process.
We consider a simple data-generating process to study the finite sample perfor-
mance of the EW and OLS-IV estimators. The response variable yit is generated
according to the following model:

yit ¼ gi þ bwit þ z0it að1 þ rwit Þuit , (57.18)

where gi captures the individual-specific intercepts, b is a scalar coefficient associ-

ated with the mismeasured variable wit, a ¼ (a1,a2,a3)0 is 3  1 vector of coefficients
associated with the 3  1 vector of perfectly measured variables zit ¼ (zit1, zit2, zit3),
uit is the error in the model, and r modulates the amount of heteroscedasticity in the
model. When r ¼ 0, the innovations are homoscedastic. When r > 0, there is
heteroscedasticity associated with the variable wit, and this correlation is stronger as
the coefficient gets larger. The model in Eq. 57.18 is flexible enough to allow us to
consider two different variables as wit: (1) the individual-specific intercept gi and
(2) the well-measured regressor zit.
We consider a standard additive measurement error

xit ¼ wit þ vit , (57.19)

where wit follows an AR(1) process:

15
The mean squared error (MSE) of an estimator y^ incorporates a component measuring the
variability of the estimator (precision) and another measuring its bias (accuracy). An estimator
with good MSE properties has small combined variance and bias. The MSE of y^ can be defined as
  h   i2
Var y^ þ Bias y^ . The root mean squared error (RMSE) is simply the square root of the
^ For an
MSE. This is an easily interpretable statistic, since it has the same unit as the estimator y.
approximately unbiased estimator, the RMSE is just the square root of the variance, that is, the
standard error.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1583

ð1  fLÞwit ¼ 2it : (57.20)

In all simulations, we set w*i, 50 ¼ 0 and generate wit for t ¼ 49, 48, . . .,T,
such that we drop the first 50 observations. This ensures that the results are not
unduly influenced by the initial values of the wit process.
Following Biorn (2000), we relax the assumption of i.i.d. measurement error.
Our benchmark simulations will use the assumption of time-invariant autocorrela-
tion [(C1) and (C2)]. In particular, we assume that uit ¼ u1i + u2it and vit ¼ v1i + v2it.
We draw all the innovations (u1i,u2it,v1i,v2it,) from a lognormal distribution; that is,
we exponentiate two normal distributions and standardize the resulting variables to
have unit variances and zero means (this follows the approach used by EW). In
Sect. 57.3.6, we analyze the alternative case in which the innovations follow an MA
structure.
The perfectly measured regressor is generated according to

zit ¼ mi þ 2it : (57.21)

And the fixed effects, mi and gi, are generated as

mi ¼ e1i
1 X T
(57.22)
gi ¼ e2i þ pffiffiffi W it ,
T t¼1

where Wit is the sum of the explanatory variables. Our method of generating mi and
gi ensures that the usual random effects estimators are inconsistent because of the
correlation that exists between the individual effects and the error term or the explan-
atory variables. The variables (e1i, e2i) are fixed as standard normal distributions.16
We employ four different schemes to generate the disturbances (2it, eit). Under
Scheme 1, we generate them under a normal distribution, N(0,s2u). Under Scheme 2,
we generate them from a lognormal distribution, LN(0,s2u). Under Scheme 3, we use
a chi-square with 5 degrees of freedom, w25. Under Scheme 4, we generate the
innovations from a Fm,n-distribution with m ¼ 10 and n ¼ 40. The latter three
distributions are right-skewed so as to capture the key distributional assumptions
behind the EW estimator. We use the normal (non-skewed) distribution as
a benchmark.
Naturally, in practice, one cannot determine how skewed – if at all – is the
distribution of the partially out latent variable. One of our goals is to check how this
assumption affects the properties of the estimators we consider. Figure 57.1 pro-
vides a visual illustration of the distributions we employ. By inspection, at least, the
three skewed distributions we study appear to be plausible candidates for the
distribution governing mismeasurement, assuming EW’s prior that measurement
error must be markedly rightly skewed.

16
Robustness checks show that the choice of a standard normal does not influence our results.
1584 H. Almeida et al.

Normal Density Chi−Square Density

0.4 0.15

0.3
0.10
0.2
y

y
0.05
0.1

0.0 0.00
−3 −2 −1 0 1 2 3 0 5 10 15 20
x x
F Density Lognormal Density

0.8 0.6

0.6
0.4
y

0.4
0.2
0.2

0.0 0.0
0 5 10 15 20 0 5 10 15 20
x x

Fig. 57.1

As in Erickson and Whited (2002), our simulations allow for cross-sectional

correlation among the variables in the model. We do so because this correlation
may aggravate the consequences of mismeasurement of one regressor on the
estimated slope coefficients of the well-measured regressors. Notably, this source
of correlation is emphasized by EW in their argument that the inferences of Fazzari
et al. (1988) are flawed in part due to the correlation between q and cash flows. To
introduce this correlation in our application, for each period in the panel, we
generate (wi, zi1, zi2, zi3) using the correspondent error distribution and then multiply
the resulting vector by [var(wi, zi1, zi2, zi3)]1/2 with diagonal elements equal to 1 and
off-diagonal elements equal to 0.5.
In the simulations, we experiment with T ¼ 10 and N ¼ 1,000. We set the
number of replications to 5,000 and consider the following values for the remaining
parameters:

ðb; a1 ; a2 ; a3 Þ ¼ ð1,  1, 1,  1Þ
f ¼ 0:6, s2u ¼ s2e1 ¼ s2e2 ¼ 1,

where the set of slope coefficients b, ai is set similarly to EW.

57 Assessing the Performance of Estimators Dealing with Measurement Errors 1585

Notice that the parameter f controls the amount of autocorrelation of the latent
regressor. As explained above, this autocorrelation is an important requirement for
the identification of the IV estimator. While we set f ¼ 0.6 in the following
experiments, we also conduct simulations in which we check the robustness of
the results with respect to variations in f between 0 and 1 (see Sect. 57.3.6).

57.3.2 The EW Identification Test

We study the EW identification test in a simple panel data setup. In the panel
context, it is important to consider individual fixed effects. If the data contain fixed
effects, according to Erickson and Whited (2000), a possible strategy is to transform
the data first and then apply their high-order GMM estimator. Accordingly,
throughout this section, our estimations consider data presented in two forms:
“level” and “within.” The first refers to data in their original format, without the
use of any transformation; estimations in level form ignore the presence of fixed
effects.17 The second applies the within transformation to the data – eliminating
fixed effects – before the model estimation.
We first compute the empirical size and power of the test. Note that the null
hypothesis is that the model is incorrectly specified, such that b ¼ 0 and/or
E[
3i ] ¼ 0. The empirical size is defined as the number of rejections of the null
hypothesis when the null is true – ideally, this should hover around 5 %. In our case,
the empirical size is given when we draw the innovations (2it, eit) from a
non-skewed distribution, which is the normal distribution since it generates
E[
3i ] ¼ 0. The empirical power is the number of rejections when the null hypoth-
esis is false – ideally, this should happen with very high probability. In the present
case, the empirical power is given when we use skewed distributions: lognormal,
chi-square, and F-distribution.
Our purpose is to investigate the validity of the skewness assumption once we
are setting b 6¼ 0. Erickson and Whited (2002) also restrict every element of b to be
nonzero. We conduct a Monte Carlo experiment to quantify the second part of this
assumption. It is important to note that we can compute the E[(
i)3] since, in our
controlled experiment, we generate wi and therefore observe it.
Since the EW test is originally designed for cross-sectional data, the first
difficulty the researcher faces when implementing a panel test is aggregation.
Following EW, our test is computed for each year separately. We report the
average of empirical rejections over the years.18 To illustrate the size and power
of the test for the panel data case, we set the time series dimension of the

17
To our knowledge, all but one of the empirical applications of the EW model use the data in level
form. In other words, firm-fixed effects are ignored outright in panel setting estimations of
parameters influencing firm behavior.
18
The results using the median are similar.
1586 H. Almeida et al.

Table 57.1 The performance of the EW identification test

Distribution Null is Data form Frequency of rejection
Normal True Level 0.05
Within 0.05
Lognormal False Level 0.47
Within 0.43
w23 False Level 0.14
Within 0.28
F10, 40 False Level 0.17
Within 0.28
This table shows the performance of the EW identification test for different distributional
assumptions displayed in column 1. The tests are computed for the data in levels and after applying
a within transformation. Column 4 shows the frequencies at which the null hypothesis that the
model is not identified is rejected at the 5 % level of significance

panel to T ¼ 10. Our tests are performed over 5,000 samples of cross-sectional size
equal to 1,000. We use a simple homoscedastic model with r ¼ 0, with the other
model parameters given as above.
Table 57.1 reports the empirical size and power of the statistic proposed by EW
for testing the null hypothesis H0 : E(y˙2i x˙i) ¼ E(y˙ix˙2i ) ¼ 0. This hypothesis is
equivalent to testing H0 : b ¼ 0 and/or E(
3i ) ¼ 0. Table 57.1 reports the frequencies
at which the statistic of test is rejected at the 5 % level of significance for,
respectively, the normal, lognormal, chi-square, and F-distributions of the data-
generating process. Recall that when the null hypothesis is true, we have the size of
the test, and when the null is false, we have the power of the test.
The results reported in Table 57.1 imply an average size of approximately 5 %
for the test. In particular, the first two rows in the table show the results in the case
of a normal distribution for the residuals (implying that we are operating under the
null hypothesis). For both the level and within cases, the empirical sizes match the
target significance level of 5 %.
When we move to the case of skewed distributions (lognormal, chi-square, and F),
the null hypothesis is not satisfied by design, and the number of rejections delivers
the empirical power of the test. In the case when the data is presented in levels and
innovations are drawn from a lognormal distribution (see row 2), the test rejects
about 47 % of the time the null hypothesis of no skewness. Using within data, the
test rejects the null hypothesis 43 % of the time. Not only are these frequencies low,
but comparing these results, one can see that the within transformation slightly
reduces the power of the test.
The results associated with the identification test are more disappointing when
we consider other skewed distributions. For example, for the F-distribution, we
obtain only 17 % of rejections of the null hypothesis in the level case and only 28 %
for the within case. Similarly, poor statistical properties for the model identification
test are observed in the chi-square case.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1587

57.3.3 Bias and Efficiency of the EW, OLS-IV, and AB-GMM

Estimators

In this section we present simulation results that assess the finite sample performance of
the estimators discussed in Sect. 57.2. The simulations compare the estimators in terms
of bias and efficiency under several distributional assumptions. In the next subsection,
we consider the cross-sectional setting, focusing on the properties of the EW estimator.
Subsequently, we examine the panel case in detail, comparing the performance of the
EW, OLS-IV, and AB-GMM estimators in terms of bias and efficiency.

57.3.3.1 The Cross-Sectional case

We generate data using a simple model as in Eqs. 57.18, and 57.19 with T ¼ 1, such
that there are no fixed effects, no autocorrelation(f ¼ 0), and no heteroscedasticity
(r ¼ 0). The other parameters are (b, a1, a2, a3) ¼ (1, 1, 1, 1). Table 57.2 shows
the results for bias and RMSE for four different distributions: lognormal,
chi-square, F-distribution, and standard normal. For each distribution we estimate
the model using three different EW estimators: EW-GMM3, EW-GMM4, and
EW-GMM5. These estimators are based on the respective third, fourth, and fifth
moment conditions. By combining the estimation of 4 parameters, under 4 different
distributions, for all 3 EW estimators – a total of 48 estimates – we aim at
establishing robust conclusions about the bias and efficiency of the EW approach.
Panel A of Table 57.2 presents the results for bias and RMSE when we use the
lognormal distribution to generate innovations (2i ei) that produce wi and zi. Under
this particular scenario, point estimates are approximately unbiased, and the small
RMSEs indicate that coefficients are relatively efficiently estimated.
Panels B and C of Table 57.2 present the results for the chi-square and
F-distribution, respectively. The experiments show that coefficient estimates pro-
duced by the EW approach are generally very biased. For example, Panel B shows
that the b coefficient returned for the EW-GMM4 and EW-GMM5 estimators is
biased downwards by approximately 35 %. Panel C shows that for EW-GMM3, the
b coefficient is biased upwards about 35 %. Paradoxically, for EW-GMM4 and
EW-GMM5, the coefficients are biased downwards by approximately 25 %. The
coefficients returned for the perfectly measured regressors are also noticeably
biased. And they, too, switch bias signs in several cases. Panels B and C show
that the EW RMSEs are very high. Notably, the RMSE for EW-GMM4 under the
chi-square distribution is 12.23, and under F-distribution, it is 90.91. These RMSE
results highlight the lack of efficiency of the EW estimator. Finally, Panel D
presents the results for the normal distribution case, which has zero skewness. In
this case, the EW estimates are severely biased and the RMSEs are extremely high.
The estimated coefficient for the mismeasured variable using EW-GMM3 has
a bias of 1.91 (about three times larger than its true value) and an RMSE of 2305.
These results reveal that the EW estimators only have acceptable performance in
the case of very strong skewness (lognormal distribution). They relate to the last
section in highlighting the poor identification of the EW framework, even in the
1588 H. Almeida et al.

Table 57.2 The EW estimator: cross-sectional data

b a1 a2 a3
Panel A. Lognormal distribution
EW-GMM3 Bias 0.0203 0.0054 0.0050 0.0056
RMSE 0.1746 0.0705 0.0704 0.0706
EW-GMM4 Bias 0.0130 0.0034 0.0033 0.0040
RMSE 0.2975 0.1056 0.1047 0.1083
EW-GMM5 Bias 0.0048 0.0013 0.0013 0.0019
RMSE 0.0968 0.0572 0.0571 0.0571
Panel B. Chi-square distribution
EW-GMM3 Bias 0.0101 0.0092 0.0101 0.0060
RMSE 61.9083 16.9275 16.1725 14.7948
EW-GMM4 Bias 0.3498 0.0938 0.0884 0.0831
RMSE 12.2386 3.2536 2.9732 3.1077
EW-GMM5 Bias 0.3469 0.0854 0.0929 0.0767
RMSE 7.2121 1.8329 1.8720 1.6577
Panel C. F-distribution
EW-GMM3 Bias 0.3663 0.1058 0.0938 0.0868
RMSE 190.9102 53.5677 52.4094 43.3217
EW-GMM4 Bias 0.2426 0.0580 0.0649 0.0616
RMSE 90.9125 24.9612 24.6827 21.1106
EW-GMM5 Bias 0.2476 0.0709 0.0643 0.0632
RMSE 210.4784 53.5152 55.8090 52.4596
Panel D. Normal distribution
EW-GMM3 Bias 1.9179 0.6397 0.5073 0.3512
RMSE 2305.0309 596.1859 608.2098 542.2125
EW-GMM4 Bias 1.0743 0.3012 0.2543 0.2640
RMSE 425.5931 111.8306 116.2705 101.4492
EW-GMM5 Bias 3.1066 1.0649 0.9050 0.5483
RMSE 239.0734 60.3093 65.5883 58.3686
This table shows the bias and the RMSE associated with the estimation of the model in Eqs. 57.17,
57.18, 57.19, 57.20, and 57.21 using the EW estimator in simulated cross-sectional data. b is the
coefficient on the mismeasured regressor, and a1 to a3 are the coefficients on the perfectly
measured regressors. The table shows the results associated with GMM3, GMM4, and GMM5
for all the alternative distributions. These estimators are based on the respective third, fourth, and
fifth moment conditions

most basic cross-sectional setup. Crucially, for the other skewed distributions we
study, the EW estimator is significantly biased for both the mismeasured and the
well-measured variables. In addition, the RMSEs are quite high, indicating low
efficiency.

57.3.3.2 The Panel Case

We argue that a major drawback of the EW estimator is its limited ability to handle
individual heterogeneity – fixed effects and error heteroscedasticity – in panel data.
This section compares the impact of individual heterogeneity on the EW, OLS-IV,
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1589

and AB-GMM estimators in a panel setting. In the first round of experiments, we

assume error homoscedasticity by setting the parameter r in Eq. 57.18 equal to
zero. We shall later allow for changes in this parameter.
Although the EW estimations are performed on a period-by-period basis, one
generally wants a single coefficient for each of the variables in an empirical model.
To combine the various (time-specific) estimates, EW suggest the minimum dis-
tance estimator (MDE) described below. Accordingly, the results presented in this
section are for the MDE that combines the estimates obtained from each of the ten
time periods considered. For example, EW-GMM3 is the MDE that combines the
ten different cross-sectional EW-GMM3 estimates in our panel.
The OLS-IV is computed after differencing the model and using the second lag
of the observed mismeasured variable x, xt2, as an instrument for Dxt. The
AB-GMM estimates (Arellano and Bond 1991) use all the orthogonality conditions,
with all available lags of x’s as instrumental variables. We also concatenate the
well-measured variables z’s in the instruments’ matrix. The AB-GMM estimator is
also computed after differencing Eq. 57.18. To highlight the gains of these various
estimators vis-à-vis the standard (biased) OLS estimator, we also report the results
of simulations for OLS models using equation in first difference without
instruments.
We first estimate the model using data in level form. While the true model
contains fixed effects (and thus it is appropriate to use the within transformation), it
is interesting to see what happens in this case since most applications of the EW
estimator use data in level form, and as shown previously, the EW identification test
performs slightly better using data in this form.
The results are presented in Table 57.3. The table makes it clear that the EW
method delivers remarkably biased results when ignoring the presence of fixed
effects. Panel A of Table 57.3 reports the results for the model estimated with the
data under strong skewness (lognormal). In this case, the coefficients for the
mismeasured regressor are very biased, with biases well in excess of 100 % of
the true coefficient for the EW-GMM3, EW-GMM4, and EW-GMM5 estimators.
The biases for the well-measured regressors are also very strong, all exceeding
200 % of the estimates’ true value. Panels B and C report results for models under
chi-square and F-distributions, respectively. The EW method continues to deliver
very biased results for all of the estimates considered. For example, the EW-GMM3
estimates that are returned for the mismeasured regressors are biased downwardly
by about 100 % of their true values – those regressors are deemed irrelevant when
they are not. Estimates for the well-measured regressors are positively biased by
approximately 200 % – they are inflated by a factor of 3. The RMSEs reported in
Panels A, B, and C show that the EW methodology produces very inefficient
estimates even when one assumes pronounced skewness in the data. Finally,
Panel D reports the results for the normal distribution. For the non-skewed data
case, the EW framework can produce estimates for the mismeasured regressor that
are downwardly biased by about 90 % of their true parameter values for all models.
At the same time, that estimator induces an upward bias of larger than 200 % for the
well-measured regressors.
1590 H. Almeida et al.

Table 57.3 The EW estimator: panel data in levels

b a1 a2 a3
Panel A. Lognormal distribution
EW-GMM3 Bias 1.6450 2.5148 2.5247 2.5172
RMSE 1.9144 2.5606 2.5711 2.5640
EW-GMM4 Bias 1.5329 2.5845 2.5920 2.5826
RMSE 1.9726 2.6353 2.6443 2.6354
EW-GMM5 Bias 1.3274 2.5468 2.5568 2.5490
RMSE 1.6139 2.5944 2.6062 2.5994
Panel B. Chi-square distribution
EW-GMM3 Bias 1.0051 2.2796 2.2753 2.2778
RMSE 1.1609 2.2887 2.2841 2.2866
EW-GMM4 Bias 0.9836 2.2754 2.2714 2.2736
RMSE 1.0540 2.2817 2.2776 2.2797
EW-GMM5 Bias 0.9560 2.2661 2.2613 2.2653
RMSE 1.0536 2.2728 2.2679 2.2719
Panel C. F-distribution
EW-GMM3 Bias 0.9926 2.2794 2.2808 2.2777
RMSE 1.1610 2.2890 2.2904 2.2870
EW-GMM4 Bias 0.9633 2.2735 2.2768 2.2720
RMSE 1.0365 2.2801 2.2836 2.2785
EW-GMM5 Bias 0.9184 2.2670 2.2687 2.2654
RMSE 2.0598 2.2742 2.2761 2.2725
Panel D. Normal distribution
EW-GMM3 Bias 0.8144 2.2292 2.228 2.2262
RMSE 0.9779 2.2363 2.2354 2.2332
EW-GMM4 Bias 0.9078 2.2392 2.2363 2.2351
RMSE 0.9863 2.2442 2.2413 2.2400
EW-GMM5 Bias 0.8773 2.2262 2.2225 2.2217
RMSE 0.9846 2.2316 2.2279 2.2269
This table shows the bias and the RMSE associated with the estimation of the model in Eqs. 57.17,
57.18, 57.19, 57.20, and 57.21 using the EW estimator in simulated panel data. The table reports
results from data in levels (i.e., without applying the within transformation). b is the coefficient on
the mismeasured regressor, and a1, a2, a3 are the coefficients on the perfectly measured regressors.
The table shows the results for the EW estimator associated with EW-GMM3, EW-GMM4, and
EW-GMM5 for all the alternative distributions. These estimators are based on the respective third,
fourth, and fifth moment conditions

Table 57.4 reports results for the case in which we apply the within transforma-
tion to the data. Here, we introduce the OLS, OLS-IV, and AB-GMM estimators.
We first present the results associated with the set up that is most favorable for the
EW estimations, which is the lognormal case in Panel A. The EW estimates for the
lognormal case are relatively unbiased for the well-measured regressors (between 4 %
and 7 % deviation from true parameter values). The same applies for the mismeasured
regressors. Regarding the OLS-IV, Panel A shows that coefficient estimates are
unbiased in all models considered. AB-GMM estimates are also approximately
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1591

Table 57.4 OLS, OLS-IV, AB-GMM, and EW estimators: panel data after within transformation
b a1 a2 a3
Panel A. Lognormal distribution
OLS Bias 0.7126 0.1553 0.1558 0.1556
RMSE 0.7131 0.1565 0.1570 0.1568
OLS-IV Bias 0.0065 0.0019 0.0014 0.0015
RMSE 0.1179 0.0358 0.0357 0.0355
AB-GMM Bias 0.0248 0.0080 0.0085 0.0081
RMSE 0.0983 0.0344 0.0344 0.0340
EW-GMM3 Bias 0.0459 0.0185 0.0184 0.0183
RMSE 0.0901 0.0336 0.0335 0.0335
EW-GMM4 Bias 0.0553 0.0182 0.0182 0.0183
RMSE 0.1405 0.0320 0.0321 0.0319
EW-GMM5 Bias 0.0749 0.0161 0.0161 0.0161
RMSE 0.1823 0.0303 0.0297 0.0297
Panel B. Chi-square distribution
OLS Bias 0.7126 0.1555 0.1553 0.1556
RMSE 0.7132 0.1565 0.1563 0.1567
OLS-IV Bias 0.0064 0.0011 0.0017 0.001
RMSE 0.1149 0.0348 0.0348 0.0348
AB-GMM Bias 0.0231 0.0083 0.0077 0.0081
RMSE 0.0976 0.0339 0.0338 0.0342
EW-GMM3 Bias 0.3811 0.0982 0.0987 0.0982
RMSE 0.4421 0.1133 0.1136 0.1133
EW-GMM4 Bias 0.3887 0.0788 0.0786 0.0783
RMSE 0.4834 0.0927 0.0923 0.0919
EW-GMM5 Bias 0.4126 0.0799 0.0795 0.0798
RMSE 0.5093 0.0926 0.0921 0.0923
Panel C. F-distribution
OLS Bias 0.7123 0.1554 0.1549 0.1555
RMSE 0.7127 0.1565 0.1559 0.1566
OLS-IV Bias 0.0066 0.0013 0.0023 0.001
RMSE 0.1212 0.0359 0.0362 0.0361
AB-GMM Bias 0.0232 0.0079 0.0072 0.0085
RMSE 0.0984 0.0343 0.0342 0.0344
EW-GMM3 Bias 0.3537 0.0928 0.0916 0.0917
RMSE 0.4239 0.1094 0.1086 0.1095
EW-GMM3 Bias 0.3906 0.0802 0.0790 0.0791
RMSE 0.4891 0.0939 0.0930 0.0932
EW-GMM3 Bias 0.4188 0.0818 0.0808 0.0813
RMSE 0.5098 0.0939 0.0932 0.0935
(continued)
1592 H. Almeida et al.

Table 57.4 (continued)

b a1 a2 a3
Panel D. Normal distribution
OLS Bias 0.7119 0.1553 0.1554 0.1551
RMSE 0.7122 0.1563 0.1564 0.1562
OLS-IV Bias 0.0060 0.0011 0.0012 0.0014
RMSE 0.1181 0.0353 0.0355 0.0358
AB-GMM Bias 0.0252 0.0086 0.0085 0.0084
RMSE 0.0983 0.0344 0.0339 0.0343
EW-GMM3 Bias 0.7370 0.1903 0.1904 0.1895
RMSE 0.7798 0.2020 0.2024 0.2017
EW-GMM4 Bias 0.8638 0.2141 0.2137 0.2137
RMSE 0.8847 0.2184 0.218 0.2182
EW-GMM5 Bias 0.8161 0.1959 0.1955 0.1955
RMSE 0.8506 0.2021 0.2018 0.2017
This table shows the bias and the RMSE associated with the estimation of the model in Eqs. 57.17,
57.18, 57.19, 57.20, 57.21 using the OLS, OLS-IV, AB-GMM, and EW estimators in simulated
panel data. The table reports results from the estimators on the data after applying the within
transformation. b is the coefficient on the mismeasured regressor, and a1, a2, a3 are the coefficients
on the perfectly measured regressors. The table shows the results for the EW estimator associated
with EW-GMM3, EW-GMM4, and EW-GMM5 for all the alternative distributions. These esti-
mators are based on the respective third, fourth, and fifth moment conditions

unbiased, while standard OLS estimates are very biased. In terms of efficiency, the
RMSEs of the EW-GMM3 are somewhat smaller than those of the OLS-IV and
AB-GMM for the well-measured and mismeasured regressors. However, for the
mismeasured regressor, both OLS-IV and AB-GMM have smaller RMSEs than
EW-GMM4 and EW-GMM5.
Panel B of Table 57.4 presents the results for the chi-square distribution. One can
see that the EW yields markedly biased estimates in this case. The bias in the
mismeasured regressor is approximately 38 % (downwards), and the coefficients
for the well-measured variable are also biased (upwards). In contrast, the OLS-IV
and AB-GMM estimates for both well-measured and mismeasured regressors are
approximately unbiased. In terms of efficiency, as expected, the AB-GMM presents
slightly smaller RMSEs than the OLS-IV estimator. These IV estimators’ RMSEs
are much smaller than those associated with the EW estimators.
Panels C and D of Table 57.4 show the results for the F and standard normal
distributions, respectively. The results for the F-distribution in Panel C are essentially
similar to those in Panel B: the instrumental variable estimators are approximately
unbiased while the EW estimators are very biased. Finally, Panel D shows that deviations
from a strongly skewed distribution are very costly in terms of bias for the EW estimator,
since the bias for the mismeasured regressor is larger than 70 %, while for the well
measured, it is around 20 %. A comparison of RMSEs shows that the IV estimators are
more efficient in both the F and normal cases. In all, our simulations show that standard IV
methods almost universally dominate the EW estimator in terms of bias and efficiency.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1593

We reiterate that the bias and RMSE of the IV estimators in Table 57.4 are all
relatively invariant to the distributional assumptions, while the EW estimators are
all very sensitive to those assumptions. In short, this happens because the EW relies
on the high-order moment conditions as opposed to the OLS and IV estimators.

57.3.4 Heteroscedasticity

One way in which individual heterogeneity may manifest itself in the data is via error
heteroscedasticity. Up to this point, we have disregarded the case in which the data has
a heteroscedastic error structure. However, most empirical applications in corporate
finance entail the use of data for which heteroscedasticity might be relevant. It is important
that we examine how the EW and the IV estimators are affected by heteroscedasticity.19
The presence of heteroscedasticity introduces heterogeneity in the model and
consequently in the distribution of the partialled out dependent variable. This
compromises identification in the EW framework. Since the EW estimator is
based on equations giving the moments of (yi  zimy) and (yi  zimw) as functions
of b and moments of (ui, ei,
i), the heteroscedasticity associated with the fixed
effects (ai) or with the perfectly measured regressor (zit) distorts the required
moment conditions associated with (yi  zimy), yielding biased estimates. These
inaccurate estimates enter the minimum distance estimator equation and conse-
quently produce incorrect weights for each estimate along the time dimension. As
our simulations of this section demonstrate, this leads to biased MDE estimates,
where the bias is a function of the amount of heteroscedasticity.
We examine the biases imputed by heteroscedasticity by way of graphical analysis.
The graphs we present below are useful in that they synthesize the outputs of numerous
tables and provide a fuller visualization of the contrasts we draw between the EW and
OLS-IV estimators. The graphs depict the sensitivity of those two estimators with
respect to heteroscedasticity as we perturb the coefficient p in Eq. 57.18.
In our simulations, we alternatively set wit ¼ gi or wit ¼ zit. In the first case,
heteroscedasticity is associated with the individual effects. In the second, heterosce-
dasticity is associated with the well-measured regressor. Each of our figures
describes the biases associated with the mismeasured and the well-measured
regressors for each of the OLS-IV, EW-GMM3, EW-GMM4, and EW-GMM5
estimators.20 In order to narrow our discussion, we only present results for the
highly skewed distribution case (lognormal distribution) and for data that is treated
for fixed effects using the within transformation. As Sect. 57.3.3 shows, this is the
only case in which the EW estimator returns relatively unbiased estimators for the
parameters of interest. In all the other cases (data in levels and for data generated by

19
We focus on the OLS-IV estimator hereinafter for the purpose of comparison with the EW
estimator.
20
Since estimation biases have the same features across all well-measured regressors of a model,
we restrict attention to the first well-measured regressor of each of the estimated models.
1594 H. Almeida et al.

chi-square, F, and normal distributions), the estimates are strongly biased even
under the assumption of homoscedasticity.21
Figure 57.2 presents the simulation results under the assumption that wit ¼ gi as we
vary the amount of heteroscedasticity by changing the parameter r,22 the results for
the mismeasured coefficients show that biases in the EW estimators are generally
small for r equal to zero (this is the result reported in Sect. 57.3.3). However, as this
coefficient increases, the bias quickly becomes large. For example, for r ¼ 0.45, the
biases in the coefficient of the mismeasured variable are, respectively, 11 %,
20 %, and 43 %, for the EW-GMM3, EW-GMM4, and EW-GMM5 estimators.
Notably, those biases, which are initially negative, turn positive for moderate values
of r. As heteroscedasticity increases, some of the biases diverge to positive infinite.
The variance of the biases of the EW estimators is also large. The results regarding
the well-measured variables using EW estimators are analogous to those for the
mismeasured one. Biases are substantial even for small amounts of heterosce-
dasticity, they switch signs for some level of heteroscedasticity, and their variances
are large. In sharp contrast, the same simulation exercises show that the OLS-IV
estimates are approximately unbiased even under heteroscedasticity. While the EW
estimator may potentially allow for some forms of heteroscedasticity, it is clear that it
is not well equipped to deal with this problem in more general settings.

57.3.5 Identification of the EW Estimator in Panel Data

Our Monte Carlo experiments show that the EW estimator has a poor handle of
individual fixed effects and that biases arise for deviations from the assumption of
strict lognormality. Biases in the EW framework are further magnified if one allows
for heteroscedasticity in the data (even under lognormality). The biases arising from
the EW framework are hard to measure and sign, ultimately implying that it can be
very difficult to replicate the results one obtains under that framework.
To better understand these results, we now discuss in more mathematical details
the identification of the EW estimator for the panel data case for both the model in
level and after the within transformation. Extending the EW estimator to panel data
seems to be a nontrivial task. EW have proposed to break the problem for each time
series, estimate a cross-section model for each t, and after that combine the
estimates using a minimum distance estimator. In what follows we show that this
procedure might affect the identification condition.
Consider the following model:

yit ¼ ai þ bwit þ uit , i ¼ 1, . . . , N; t ¼ 1, . . . , T, (57.23)

21
Our simulation results (available upon request) suggest that introducing heteroscedasticity
makes the performance of the EW estimator even worse in these cases.
22
The results for wit ¼ zit are quite similar to those we get from setting wit ¼ gi. We report only one
set of graphs to save space.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1595

Fig. 57.2 a Mismeasured Variable

Bias 2

0 1 2 3 4 5
ρ

b Perfectly−Measured Variable
0.5

0.0
Bias

−0.5

−1.0

0 1 2 3 4 5
ρ

where uit is independent and identically distributed, with mean zero and variance
s2u. Assume that the Var(xit) ¼ s2w. The independent variable and unobserved effects
are exogenous, that is, Cov(wit, uis) ¼ Cov(ai, uit) ¼ 0 for any t and s. However,
Cov(ai, wit) 6¼ 0. Now, assume that we do not observe the true variable wit, but rather
a mismeasured variable, that is, you observe the following variable with an error:

xit ¼ wit þ eit , i ¼ 1, . . . , N . . . , N; t ¼ 1, . . . T, (57.24)

1596 H. Almeida et al.

where Cov(xit, eis) ¼ Cov(ai, eis) ¼ Cov(uit, eis) ¼ 0, and Var(eit) ¼ s2e ,
Cov(eit, eit  1) ¼ gs2e .
In addition, assume here that there is no variable zit (a ¼ 0 in Eq. 57.18) to
simplify the argument.

57.3.5.1 Model in Level

As mentioned before, EW propose to fix a particular time series and estimate the
model using the cross-section data. Without loss for generality, fix T ¼ 1. Thus,
Eqs. 57.23 and 57.24 become

yi1 ¼ ai þ bwi1 þ ui1 , i ¼ 1, . . . , N, (57.25)

and

xi1 ¼ wi1 þ ei1 , i ¼ 1, . . . , N: (57.26)

However, the unobserved individual-specific intercepts, ai, are still present

in Eq. 57.25 and in addition Cov(ai, wi1) 6¼ 0. Therefore, one can see that it is
impossible to estimate b consistently since ai’s are unobserved. This argument
is easily extended for every t ¼ 1, . . ., T. Thus, the estimator for each fixed t is
inconsistent, and consequently the minimum distance estimator is inconsistent by
construction. Therefore, we conclude that the EW minimum distance estimator
produces inconsistent estimates for panel data model with fixed effects.

57.3.5.2 Model After Within Transformation

Given the inconsistency of the model in levels presented in the last section, one
strategy is to previously transform the data to eliminate the fixed effects. One
suggestion is to use the within transformation in the data before estimation.
In order to analyze the model after the transformation, let’s assume that T ¼ 2 for
simplification. Using the within transformation in Eqs. 57.23 and 57.24, we obtain

yit  yit ¼ bðwit  wi Þ þ ðuit  ui Þ,

and

xit  xi ¼ ðwit  wi Þ þ ðeit  ei Þ,

where y- i ¼ 12 ðyil þ yi2 Þ, w- i ¼ 12 ðwi1 þ wi2 Þ, and so on.

Now, again EW propose to use a particular time series and estimate the model
using the cross-section data. Let’s use t ¼ 1 for ease of exposition. The model can
be written as
yil  yi ¼ bðwi1  wi Þ þ ðui1  ui Þ
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1597

and

xi1  xi ¼ ðwi1  wi Þ þ ðeit  ei Þ:

Now substituting the definition of the deviations and rearranging, we have

 
1 1 1
yi1  ðyi1 þ yi2 Þ ¼ b wi1  ðwi1 þ wi2 Þ þ ui1  ðui1 þ ui2 Þ ,
2 2 2
yil þ yi2 ¼ bðwi1 þ wi2 Þ þ ðui1 þ ui2 Þ,

and
 
1 1 1
xi1  ðxi1 þ xi2 Þ ¼ wi1  ðwi1 þ wi2 Þ þ ei1  ðei1 þ ei2 Þ ,
2 2 2
xi1 þ xi2 ¼ ðwi1 þ wi2 Þ þ ðei1 þ ei2 Þ:

Finally, our model can be described as

yi1 þ yi2 ¼ bðwi1 þ wi2 Þ þ ðui1 þ ui2 Þ,

and

xi1 þ xi2 ¼ ðwi1 þ wi2 Þ þ ðei1 þ ei2 Þ:

Let’s now define Yi ¼ yi1 + yi2, Xi ¼ xi1 + xi2, Ui ¼ ui1 + ui2, vi ¼ wi1 + wi2, and
Ei ¼ ei1 + ei2. So, the model could be rewritten as

Y i ¼ bvi þ Ui

and

X i ¼ v i þ Ei :

Notice that the requirements for identification now are on the high-order
moments of (V, U, E). However, note that vi ¼ wi1 + wi2, which is a sum of two
random variables. As it is well known from the econometrics literature, convolution
of random variables is in general a nontrivial object.
One example of why the identification condition may worsen considerably
is the following. Consider a model where wi1 and wi2 are independent
chi-square distributions with 2 degrees ofpffiffiffiffiffiffiffi
freedom.
ffi The skewness of the
chi-square with k degrees of freedom is 8=k . Note that the sum of two
independent chi-squares with k degrees of freedom is a chi-square with 2k degrees
1598 H. Almeida et al.

of freedom. Therefore, the skewness of the vi ¼ wi1 + wi2 drops from two
for the model using one distribution to 1.41 for the model using the summation
of both wi1 and wi2.
From this simple analysis one could conclude that the identification conditions
required for EW estimator can deteriorate considerably when using the within
transformation to eliminate the fixed effects in panel data. Thus, the required
conditions to achieve unbiased estimates with EW are very strong.

57.3.6 Revisiting the OLS-IV Assumptions

Our Monte Carlo simulations show that the OLS-IV estimator is consistent even
when one allows for autocorrelation in the measurement-error structure. We have
assumed, however, some structure on the processes governing innovations. In this
section, we examine the sensitivity of the OLS-IV results with respect to our
assumptions about measurement-error correlation and the amount of autocorrela-
tion in the latent regressor. These assumptions can affect the quality of the instru-
ments and therefore should be examined in some detail.
We first examine conditions regarding the correlation of the measurement
errors and disturbances. The assumption of time-invariant autocorrelation for
measurement errors and disturbances implies that past shocks to measurement
errors do not affect the current level of the measurement error. One way to relax
this assumption is to allow for the measurement-error process to have a moving
average structure. This structure satisfies Biorn’s assumptions (B1) and (B2). In
this case, Proposition 1 in Biorn (2000) shows that for an MA(t), the instruments
should be of order of at most t  t  2. Intuitively, the set of instruments must be
“older” than the memory of the measurement-error process. For example, if the
measurement error is MA(1), then one must use third- and longer-lagged
instruments to identify the model.
To analyze this case, we conduct Monte Carlo simulations in which we
replace the time-invariant assumption for innovation uit and vit with an MA(1)
structure for the measurement-error process. The degree of correlation in the
MA process is set to y ¼ 0.4. Thus, the innovation in Eqs. 57.18 and 57.19 has
the following structure:

uit ¼ u1it  yu1it1 and vit ¼ v1it ¼ v1it  yv1it1 ,

with jyj  1, and u1it and v1it are i.i.d. lognormal distributions. The other parameters
in the simulation remain the same.
The results are presented in Table 57.5. Using MA(1) in the innovations and
the third lag of the latent regressor as an instrument (either on its own or in
combination with the fourth lag), the bias of the OLS estimator is very small
(approximately 2–3 %). The bias increases somewhat when we use only the fourth
lag. While the fourth is an admissible instrument in this case, using longer lags
decreases the implied autocorrelation in the latent regressor [which follows an
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1599

AR(1) process by Eq. 57.20]. This effect decreases somewhat the quality of the
instruments. Notice also that when we do not eliminate short lags from the instru-
ment set, the identification fails. For example, the bias is 60 % when we use the
second lag as an instrument. These results thus underscore the importance of using
long enough lags in this MA case. Table 57.5 also reports results based on an MA
(2) structure. The results are qualitatively identical to those shown in the MA
(1) case. Once again, the important condition for identification is to use long enough
lags (no less than four lags in this case).23
The second condition underlying the use of the OLS-IV is that the latent
regressor is not time invariant. Accordingly, the degree of autocorrelation in
the process for the latent regressor is an important element of the identification
strategy. We assess the sensitivity of the OLS-IV results to this condition by
varying the degree of autocorrelation through the autoregressive coefficient in the
AR(1) process for the latent regressor. In these simulations, we use a time-invariant
autocorrelation condition for the measurement error, but the results are very similar
for the MA case.
Figure 57.3 shows the results for the bias in the coefficients of interest for the well-
measured and mismeasured variables, using the second lag of the mismeasured
variable as an instrument. The results show that the OLS-IV estimator performs well
for a large range of the autoregressive coefficient. However, as expected, when the f
coefficient is very close to zero or one, we have evidence of a weak instrument
problem. For example, when f ¼ 1, then Dwit is uncorrelated with any variable
dated at time t  2 or earlier. These simulations show that, provided that one uses
adequately lagged instruments, the exact amount of autocorrelation in the latent
variable is not a critical aspect of the estimation.
The simulations of this section show how the performance of the OLS-IV
estimator is affected by changes in assumptions concerning measurement errors
and latent regressors. In practical applications, it is important to verify whether
the results obtained with OLS-IV estimators are robust to the elimination of
short lags from the instrumental set. This robustness check is particularly
important given that the researcher will be unable to pin down the process
followed by the measurement error. Our empirical application below incorporates
this suggestion. In addition, identification relies on some degree of autocorrelation
in the process for the latent regressor. While this condition cannot be directly
verified, we can perform standard tests of instrument adequacy that rely on
“first-stage” test statistics calculated from the processes for the observable
variables in the model.
Another important assumption in the OLS is non-autocorrelation in both uit
and vit. For example, these innovations cannot follow an autoregressive process.

23
We note that if the instrument set uses suitably long lags, then the OLS-IV results are robust to
variations in the degree of correlation in the MA process. In unreported simulations under MA(1),
we show that the OLS bias is nearly invariant to the parameter y.
1600 H. Almeida et al.

Table 57.5 Moving average structures for the measurement-error process

Instrument MA(1) MA(2)
Xit2 0.593 0.368
(0.60) (0.38)
Xit3 0.028 0.707
(0.30) (0.71)
Xit3, Xit4 0.025 0.077
(0.63) (1.01)
Xit3, Xit4, Xit5 0.011 0.759
(0.30) (0.76)
Xit4 0.107 0.144
(1.62) (2.01)
Xit4, Xit5 0.113 0.140
(0.58) (0.59)
Xit5 0.076 0.758
(0.31) (0.76)
This table shows the bias in the well-measured coefficient for OLS-IV using moving average
structure for the measurement-error process. Numbers in parentheses are the RMSE

When this is the case, the IV strategy of using lags of mismeasured variable as valid
instruments is invalid (see Biorn 2000).

57.3.7 Distributional Properties of the EW and OLS-IV Estimators

A natural question is whether our simulation results are rooted in the lack of
accuracy of the asymptotic approximation of the EW method. Inference in models
with mismeasured regressors is based on asymptotic approximations; hence infer-
ence based on estimators with poor approximations might lead to wrong inference
procedures. For instance, we might select wrong critical values for a test under poor
asymptotic approximations and make inaccurate statements under such circum-
stances. In this section, we use the panel data simulation procedure of Sect. 3.3.2 to
study and compare the accuracy of the asymptotic approximation of the EW and IV
methods. To save space, we restrict our attention to the mismeasured regressor
coefficient for the EW-GMM5 and OLS-IV cases. We present results where we
draw the data from the lognormal, chi-square, and F-distributions. The EW-GMM5
estimator is computed after the within transformation and the OLS-IV uses second
lags as instruments.
One should expect both the IV and EW estimators to have asymptotically normal
representations, such that when we normalize the estimator by subtracting the true
parameter and divide by the standard deviation, this quantity behaves asymptoti-
cally as a normal distribution. Accordingly, we compute the empirical density and
the distribution functions of the normalized sample estimators and their normal
approximations. These functions are plotted in Fig. 57.4. The true normal density
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1601

Mismeasured Variable
2

1
Bias

−1

−2
0.0 0.2 0.4 0.6 0.8 1.0
φ
Perfectly−Measured Variable

0.4
Bias

0.0

−0.4

0.0 0.2 0.4 0.6 0.8 1.0

φ

Fig. 57.3

and distribution functions (drawn in red) serve as benchmarks. The graphs in

Fig. 57.4 depict the accuracy of the approximation. We calculate the density of
the estimators using a simple Gaussian Kernel estimator and also estimate the
empirical cumulative distribution function.24
Consider the lognormal distribution (first panel). In that case, the OLS-IV (black
line) displays a very precise approximation to the normal curve in terms of both
density and distribution. The result for the OLS-IV is robust across all of the
distributions considered (lognormal, chi-square, and F). These results are in sharp
contrast to those associated with the EW-GMM5 estimator. This estimator presents
a poor asymptotic approximation for all distributions examined. For the lognormal
case, the density is not quite centered at zero, and its shape does not fit the normal
distribution. For the chi-square and F-distributions, Fig. 57.4 shows that the shapes
of the density and distribution functions are very unlike the normal case, with the
center of the distribution located far away from zero. These results imply that
inference procedures using the EW estimator might be asymptotically invalid in
simple panel data with fixed effects, even when the relevant distributions present
high skewness.

24
The empirical cumulative distribution function Fn is a step function with jumps i/n at observation
values, where i is the number of tied observations at that value.
1602 H. Almeida et al.

Fig. 57.4

57.4 Empirical Application

We apply the EW and OLS-IV estimators to Fazzari et al. (1988) investment

equation. This is the most well-known model in the corporate investment literature,
and we use this application as a way to illustrate our Monte Carlo-based results. In
the Fazzari, Hubbard, and Petersen model, a firm’s investment spending is
regressed on a proxy for investment demand (Tobin’s q) and the firm’s cash flow.
Theory suggests that the correct proxy for the firm’s investment demand is
marginal q, but this quantity is unobservable and researchers use instead its
measurable proxy, average q. Because average q measures marginal
q imperfectly, a measurement problem naturally arises. Erickson and Whited
(2002) uses the Fazzari, Hubbard, and Petersen model to motivate the adoption of
their estimator in applied work in panel data.
A review of the corporate investment literature shows that virtually all empirical
work in the area considers panel data models with firm-fixed effects (Kaplan and
Zingales 1997; Rauh 2006; Almeida and Campello 2007). From an estimation point
of view, there are distinct advantages in exploiting repeated observations from
individuals to identify the model (Blundell et al. 1992). In an investment model
setting, exploiting firm effects contributes to estimation precision and allows for
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1603

model consistency in the presence of unobserved idiosyncrasies that may be

simultaneously correlated with investment and q. The baseline model in this
literature has the form

I it =K it ¼
i þ bq it þ aCFit =K it þ uit , (57.27)

where I denotes investment, K capital stock, q* is marginal q, CF cash flow,
is the

firm-specific effect, and u is the innovation term.
As mentioned earlier, if q* is measured with error, OLS estimates of b will be
biased downwards. In addition, given that q and cash flow are likely to be positively
correlated, the coefficient a is likely to be biased upwards in OLS estimations. In
expectation, these biases should be reduced by the use of estimators like the ones
discussed in the previous section.

57.4.1 Theoretical Expectations

In order to better evaluate the performance of the two alternative estimators, we

develop some hypotheses about the effects of measurement-error correction on the
estimated coefficients b and a from Eq. 57.27. Theory does not pin down the exact
values that these coefficients should take. Nevertheless, one could argue that the
two following conditions should be reasonable.
First, an estimator that addresses measurement error in q in a standard
investment equation should return a higher estimate for b and a lower estimate
for a when compared with standard OLS estimates. Recall that measurement error
causes an attenuation bias on the estimate for the coefficient b. In addition,
since q and cash flow are likely to be positively correlated, measurement error
should cause an upward bias on the empirical estimate returned under the standard
OLS estimation. Accordingly, if one denotes the OLS and the measurement-error
consistent estimates, respectively, by (bOLS, aOLS) and (bMEC, aMEC), one
should expect:
Condition 1. bOLS < bMEC and aOLS > aMEC. Second, one would expect the coefficients
for q and the cash flow to be nonnegative after treating the data for measurement
error. The q-theory of investment predicts a positive correlation between invest-
ment and q (e.g., Hayashi 1982). If the theory holds and the estimator does a good
job of adjusting for measurement error, then the cash flow coefficient should be zero
(“neoclassical view”). However, the cash flow coefficient could be positive either
because of the presence of financing frictions (as posited by Fazzari et al. 1988)25 or

25
However, financial constraints are not sufficient to generate a strictly positive cash flow
coefficient because the effect of financial constraints is capitalized in stock prices and may thus
be captured by variations in q (Chirinko 1993; Gomes 2001).
1604 H. Almeida et al.

due to fact that cash flow picks up variation in investment opportunities even after
we apply a correction for mismeasurement in q. Accordingly, one should observe:
Condition 2. bMEC  0 and aMEC  0. Notice that these conditions are fairly weak.
If a particular measurement-error consistent estimator does not deliver these basic
results, one should have reasons to question the usefulness of that estimator in
applied work.

57.4.2 Data Description

Our data collection process follows that of Almeida and Campello (2007). We consider
a sample of manufacturing firms over the 1970–2005 period with data available from
Compustat. Following those authors, we eliminate firm years displaying asset or sales
growth exceeding 100 %, or for which the stock of fixed capital (the denominator of
the investment and cash flow variables) is less than $5 million (in 1976 dollars).
Our raw sample consists of 31,278 observations from 3,084 individual firms. Summary
statistics for investment, q, and cash flow are presented in Table 57.6. These statistics
are similar to those reported by Almeida and Campello, among other papers. To save
space we omit the discussion of these descriptive statistics.

57.4.3 Testing for the Presence of Fixed Effects and

Heteroscedasticity

Before estimating our investment models, we conduct a series of tests for the
presence of firm-fixed effects and heteroscedasticity in our data. As a general
rule, these phenomena might arise naturally in panel data applications and should
not be ignored. Importantly, whether they appear in the data can have concrete
implications for the results generated by different estimators.
We first perform a couple of tests for the presence of firm-fixed effects. We allow
for individual firm intercepts in Eq. 57.27 and test the null hypothesis that the
coefficients associated with those firm effects are jointly equal to zero (Baltagi
2005). Table 57.7 shows that the F-statistic for this test is 4.4 (the associated
p-value is 0.000). Next, we contrast the random effects OLS and the fixed effects
OLS estimators to test again for the presence of fixed effects. The Hausman test
statistic reported in Table 57.7 rejects the null hypothesis that the random effects
model is appropriate with a test statistic of 8.2 (p-value of 0.017). In sum, standard
tests strongly reject the hypothesis that fixed effects can be ignored.
We test for homoscedasticity using two different panel data-based methods.
First, we compute the residuals from the least squares dummy variables estimator
and regress the squared residuals on a function of the independent variables [see
Frees (2004) for additional details]. We use two different combinations of inde-
pendent regressors – (qit, CFit) and (qit, q2it, CFit, CFit, CF2it) – and both of them
robustly reject the null hypothesis of homoscedasticity. We report the results for the
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1605

Table 57.6 Descriptive statistics

Variable Obs. Mean Std. dev. Median Skewness
Investment 22,556 0.2004 0.1311 0.17423 2.6871
q 22,556 1.4081 0.9331 1.1453 4.5378
Cash flow 22,556 0.3179 0.3252 0.27845 2.2411
This table shows the basic descriptive statistics for q, cash flow, and investment. The data are taken
from the annual Compustat industrial files over the 1970–2005 period. See text for details

Table 57.7 Diagnosis tests

Test Test statistic p-value
Pooling test 4.397 0.0000
Random effects vs. fixed effects 8.17 0.0169
Homoscedasticity 1 55.19 0.0000
Homoscedasticity 2 7,396.21 0.0000
This table reports results for specification tests. Hausman test for fixed effects models considers
fixed effects models against the simple pooled OLS and the random effects model.
A homoscedasticity test for the innovations is also reported. The data are taken from the annual
Compustat industrial files over the 1970–2005 period. See text for details

first combination in Table 57.7, which yields a test statistic of 55.2 (p-value of
0.000). Our second approach for testing the null of homoscedasticity is the standard
random effects Breusch-Pagan test. Table 57.7 shows that the Breusch-Pagan test
yields a statistic of 7,396.2 (p-value of 0.000). Our tests hence show that the data
strongly reject the hypothesis of error homoscedasticity.

57.4.4 Implementing the EW Identification Test

Our preliminary tests show that one should control for fixed effects when estimating
investment models using real data. In the context of the EW estimator, it is thus
appropriate to apply the within transformation before the estimation. However, in this
section, we also present results for the data in level form to illustrate the point made in
Sect. 57.3.2 that applying the within transformation compromises identification in the
EW context. Prior papers adopting the EW estimator have ignored (or simply
dismissed) the importance of fixed effects (e.g., Whited 2001, 2006).
We present the results for EW’s identification test in Table 57.8. Using the data in
level form, we reject the hypothesis of no identification in 12 out of 30 years (or 36 %
rejection). For data that is transformed to accommodate fixed effects (within trans-
formation), we find that in only 7 out of 33 (or 21 %) of the years between 1973 and
2005, one can reject the null hypothesis that the model is not identified at the usual
5 % level of significance. These results suggest that the power of the test is low and
decreases further after applying the within transformation to the data. These results
are consistent with Almeida and Campello’s (2007) use of the EW estimator.
Working with a 15-year Compustat panel, those authors report that they could only
find a maximum of 3 years of data passing the EW identification test.
1606 H. Almeida et al.

The results in Table 57.8 reinforce the notion that it is quite difficult to
operationalize the EW estimator in real-world applications, particularly in situa-
tions in which the within transformation is appropriate due to the presence of fixed
effects. We recognize that the EW identification test rejects the model for most of
the data at hand. However, recall from Sect. 57.3.2 that the test itself is likely to be
misleading (“over-rejecting” the data). In the next section, we take the EW estima-
tor to the data (a standard Compustat sample extract) to illustrate the issues applied
researchers face when using that estimator, contrasting it to an easy-to-implement
alternative.

57.4.5 Estimation Results

We estimate Eq. 57.27 using the EW, OLS-IV, and AB-GMM estimators. For
comparison purposes, we also estimate the investment equation using standard OLS
and OLS with fixed effects (OLS-FE). The estimates for the standard OLS are likely to
be biased, providing a benchmark to evaluate the performance of the other estimators.
As discussed in Sect. 57.4.1, we expect estimators that improve upon the problem of
mismeasurement to deliver results that satisfy Conditions 1 and 2 above.
As is standard in the empirical literature, we use an unbalanced panel in our
estimations. Erickson and Whited (2000) propose a minimum distance estimator
(MDE) to aggregate the cross-sectional estimates obtained for each sample year,
but their proposed MDE is designed for balanced panel data. Following Riddick and
Whited (2009), we use a Fama–MacBeth procedure to aggregate the yearly EW
estimations.26
To implement our OLS-IV estimators, we first take differences of the model in
Eq. 57.27. We then employ the estimator denoted by OLS-IV A from Sect. 57.2.2,
using lagged levels of q and cash flow as instruments for (differenced) qit. Our
Monte Carlos suggest that identification in this context may require the use of
longer lags of the model variables. Accordingly, we experiment with specifications
that use progressively longer lags of q and cash flow to verify the robustness of our
results.
Table 57.9 reports our findings. The OLS and OLS-FE estimates, reported in
columns (1) and (2), respectively, disregard the presence of measurement error
in q. The EW-GMM3, EW-GMM4, and EW-GMM5 estimates are reported in
columns (3), (4), and (5). For the OLS-IV estimates reported in column (6), we
use qt—2 as an instrument.27 The AB-GMM estimator, reported in column (7), uses
lags of q as instruments. Given our data structure, this implies using a total of
465 instruments. We account for firm-fixed effects by transforming the data.

26
Fama–MacBeth estimates are computed as a simple standard errors for yearly estimates. An
alternative approach could use the Hall–Horowitz bootstrap. For completeness, we present in the
appendix the actual yearly EW estimates.
27
In the next section, we examine the robustness of the results with respect to variation in the
instrument set.
 57

Table 57.8 The EW identification test using real data

Level Within transformation
#Rejections null #Rejections null
1973 t-statistic 1.961 0 1973 t-statistic 1.349 0
p-value 0.375 p-value 0.509
1974 t-statistic 5.052 0 1974 t-statistic 7.334 1
p-value 0.08 p-value 0.026
1975 t-statistic 1.335 0 1975 t-statistic 1.316 0
p-value 0.513 p-value 0.518
1976 t-statistic 7.161 1 1976 t-statistic 5.146 0
p-value 0.028 p-value 0.076
1977 t-statistic 1.968 0 1977 t-statistic 1.566 0
p-value 0.374 p-value 0.457
1978 t-statistic 9.884 1 1978 t-statistic 2.946 0
p-value 0.007 p-value 0.229
1979 t-statistic 9.065 1 1979 t-statistic 1.042 0
p-value 0.011 p-value 0.594
1980 t-statistic 9.769 1 1980 t-statistic 7.031 1
p-value 0.008 p-value 0.03
1981 t-statistic 10.174 1 1981 t-statistic 7.164 1
p-value 0.006 p-value 0.028
1982 t-statistic 3.304 0 1982 t-statistic 2.991 0
p-value 0.192 p-value 0.224
1983 t-statistic 5.724 0 1983 t-statistic 9.924 1
Assessing the Performance of Estimators Dealing with Measurement Errors

p-value 0.057 p-value 0.007

1984 t-statistic 15.645 1 1984 t-statistic 6.907 1
p-value 0 p-value 0.032
(continued)
1607
 1608

Table 57.8 (continued)

Level Within transformation
#Rejections null #Rejections null
1985 t-statistic 16.084 1 1985 t-statistic 1.089 0
p-value 0 p-value 0.58
1986 t-statistic 4.827 0 1986 t-statistic 5.256 0
p-value 0.089 p-value 0.072
1987 t-statistic 19.432 1 1987 t-statistic 13.604 1
p-value 0 p-value 0.001
1988 t-statistic 5.152 0 1988 t-statistic 1.846 0
p-value 0.076 p-value 0.397
1989 t-statistic 0.295 0 1989 t-statistic 0.687 0
p-value 0.863 p-value 0.709
1990 t-statistic 0.923 0 1990 t-statistic 1.3 0
p-value 0.63 p-value 0.522
1991 t-statistic 3.281 0 1991 t-statistic 3.17 0
p-value 0.194 p-value 0.205
1992 t-statistic 2.31 0 1992 t-statistic 2.573 0
p-value 0.315 p-value 0.276
1993 t-statistic 1.517 0 1993 t-statistic 1.514 0
p-value 0.468 p-value 0.469
1994 t-statistic 2.873 0 1994 t-statistic 4.197 0
p-value 0.238 p-value 0.123
1995 t-statistic 0.969 0 1995 t-statistic 1.682 0
p-value 0.616 p-value 0.431
H. Almeida et al.
 57

1996 t-statistic 17.845 1 1996 t-statistic 4.711 0

p-value 0 p-value 0.095
1997 t-statistic 0.14 0 1997 t-statistic 1.535 0
p-value 0.933 p-value 0.464
1998 t-statistic 0.623 0 1998 t-statistic 5.426 0
p-value 0.732 p-value 0.066
1999 t-statistic 0.354 0 1999 t-statistic 2.148 0
p-value 0.838 p-value 0.342
2000 t-statistic 13.44 1 2000 t-statistic 13.502 1
p-value 0.001 p-value 0.001
2001 t-statistic 3.159 0 2001 t-statistic 3.309 0
p-value 0.206 p-value 0.191
2002 t-statistic 13.616 1 2002 t-statistic 0.693 0
p-value 0.001 p-value 0.707
2003 t-statistic 12.904 1 2003 t-statistic 4.006 0
p-value 0.002 p-value 0.135
2004 t-statistic 5.212 0 2004 t-statistic 2.801 0
p-value 0.074 p-value 0.246
2005 t-statistic 2.365 0 2005 t-statistic 4.127 0
p-value 0.306 p-value 0.127
Sum 12 Sum 7
% of years 0.3636 % of years 0.2121
This table shows the test statistic and its p-value for the EW identification test, which tests the null hypothesis that the model is not identified. The tests are
Assessing the Performance of Estimators Dealing with Measurement Errors

performed on a yearly basis. In the last columns, we collect the number of years in which the null hypothesis is rejected (sum) and compute the percentage of
years in which the null is rejected. The data are taken from the annual Compustat industrial files over the 1970–2005 period. See text for details
1609
 1610

Table 57.9 EW, GMM, and OLS-IV coefficients, real-world data

Variables OLS OLS-FE EW-GMM3 EW-GMM4 EW-GMM5 OLS-IV AB-GMM
q 0.0174*** 0.0253*** 0.0679 0.3031 0.0230 0.0627*** 0.0453***
(0.002) (0.003) (0.045) (0.302) (0.079) (0.007) (0.006)
Cash flow 0.1310*** 0.1210*** 0.1299*** 0.3841* 0.1554*** 0.0434*** 0.0460***
(0.011) (0.017) (0.031) (0.201) (0.052) (0.007) (0.016)
Observations 22,556 22,556 22,556 22,556 22,556 17,348 19,748
F-stat p-value (first step) – – – – – 0.000 –
This table shows the coefficients and standard deviations that we obtain when we use the OLS, EW, and the GMM estimators in Eq. 57.22. The table also
displays the standard OLS-FE coefficients (after applying the differencing transformation to treat the fixed effects) in column (2) and OLS-IV in the last
column. Robust standard errors in parentheses for OLS and GMM and clustered in firms for OLS-FE and OLS-IV. Each EW coefficient is an average of the
yearly coefficients reported in Table 57.11 and the standard error for these coefficients is a Fama–MacBeth standard error. The table shows the EW coefficients
for the data after applying the within transformation. The data are taken from the annual Compustat industrial files over the 1970–2005 period. See text for
details. *, **, and *** represent statistical significance at the 10 %, 5 %, and 1 % levels, respectively
H. Almeida et al.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1611

When using OLS and OLS-FE, we obtain the standard result in the literature that
both q and cash flow attract positive coefficients [see columns (1) and (2)]. In the
OLS-FE specification, for example, we obtain a q coefficient of 0.025 and a cash
flow coefficient of 0.121. Columns (3), (4), and (5) show that the EW estimator does
not deliver robust inferences about the correlations between investment, cash flow,
and q. The q coefficient estimate varies significantly with the set of moment
conditions used, even flipping signs. In addition, none of the q coefficients is
statistically significant. The cash flow coefficient is highly inflated under EW, and
in the case of the EW-GMM4 estimator, it is more than three times larger than the
(supposedly biased) OLS coefficient. These results are inconsistent with Conditions
1 and 2 above. These findings agree with the Monte Carlo simulations of Sect.
57.3.3, which also point to a very poor performance of the EW estimator in cases in
which fixed effects and heteroscedasticity are present.
By comparison, the OLS-IV delivers results that are consistent with Conditions
1 and 2. In particular, the q coefficient increases from 0.025 to 0.063, while the cash
flow coefficient drops from 0.131 to 0.043. These results suggest that the proposed
OLS-IV estimator does a fairly reasonable job at addressing the measurement-error
problem. This conclusion is consistent with the Monte Carlo simulations reported
above, which show that the OLS-IV procedure is robust to the presence of fixed
effects and heteroscedasticity in simulated data. The AB-GMM results also gener-
ally satisfy Conditions 1 and 2. Notice, however, that the observed changes in the
q and cash flow coefficients (“corrections” relative to the simple, biased OLS
estimator) are less significant than those obtained under the OLS-IV estimation.

57.4.6 Robustness of the Empirical OLS-IV Estimator

It is worth demonstrating that the OLS-IV we consider is robust to variations in the

set of instruments that is used for identification. While the OLS-IV delivered results
that are consistent with our priors, note that we examined a just-identified model,
for which tests of instrument quality are not available. As we have discussed
previously, OLS-IV estimators should be used with care in this setting, since the
underlying structure of the error in the latent variable is unknown. In particular, the
Monte Carlo simulations suggest that it is important to show that the results remain
when we use longer lags to identify the model.
We present the results from our robustness checks in Table 57.10. We start by
adding one more lag of q (i.e., qt3) to the instrumental set. The associated
estimates are in the first column of Table 57.10. One can observe that the
slope coefficient associated with q increases even more with the new instrument
(up to 0.090), while that of the cash flow variable declines further (down to 0.038).
One problem with this estimation, however, is the associated J-statistic.
If we consider a 5 % hurdle rule, the J-statistic of 4.92 implies that, with this
particular instrumental set, we reject the null hypothesis that the identification
restrictions are met (p-value of 3 %). As we have discussed, this could be
expected if, for example, the measurement-error process has an MA structure.
 1612

Table 57.10 OLS-IV coefficients, robustness tests

Variables (1) (2) (3) (4) (5) (6) (7)
q 0.0901*** 0.0652*** 0.0394*** 0.0559*** 0.0906*** 0.0660*** 0.0718***
(0.014) (0.014) (0.015) (0.012) (0.033) (0.024) (0.026)
Cash flow 0.0383*** 0.0455*** 0.0434*** 0.0449*** 0.0421*** 0.0450*** 0.0444***
(0.007) (0.011) (0.011) (0.010) (0.008) (0.012) (0.012)
Observations 15,264 11,890 12,000 13,448 12,000 10,524 10,524
F-stat p-value (first step) 0.000 0.000 0.000 0.000 0.001 0.000 0.000
J-stat 4.918 3.119 0.122 0.00698 0.497 8.955 5.271
J-stat p-value 0.0266 0.210 0.727 0.933 0.481 0.111 0.261
This table shows the results of varying the set of instruments that are used when applying the OLS-IV estimator to Eq. 57.22. In the first column we use the
second and third lags of q as instruments for current (differenced) q, as in Table 57.6. In column (2), we use third, fourth, and fifth lags of q as instruments. In
column (3), we use the fourth and fifth lags of q and the first lag of cash flow as instruments. In column (4), we use the third lag of q and fourth lag of cash flow
as instruments. In column (5), we use the fourth and fifth lags of cash flow as instruments. In column (6), we use {qt4, qt5, qt6, CFt3, CFt4, CFt5} as
instruments. Finally, in column (7) {qt5,qt6,CFt3,CFt4,CFt5} as instruments. The estimations correct the errors for heteroscedasticity and firm-
clustering. The data are taken from the annual Compustat industrial files over the 1970–2005 period. See text for details. *, **, and *** represent statistical
significance at the 10 %, 5 %, and 1 % levels, respectively
H. Almeida et al.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1613

This suggests that the researcher should look for longer lagging schemes, lags that
“erase” the MA memory of the error structure.
Our next set of estimations use longer lagging structures for our proposed
instruments and even an instrumental set with only lags of cash flow, the exogenous
regressors in the model. We use combinations of longer lags of q (such as the fourth
and fifth lags) and longer lags of cash flow (fourth and fifth lags). This set of tests
yields estimates that more clearly meet standard tests for instrument validity.28
Specifically, the J-statistics now indicate we do not reject the hypothesis that the
exclusion restrictions are met. The results reported in columns (2) through (7) of
Table 57.10 also remain consistent with Conditions 1 and 2. In particular, the
q coefficient varies from approximately 0.040 to 0.091, while the cash flow
coefficient varies roughly from 0.044 to 0.046. These results are consistent with
our simulations, which suggest that these longer lag structures should deliver
relatively consistent, stable estimates of the coefficients for q and cash flow in
standard investment regressions.

57.5 Concluding Remarks

OLS estimators have been used as a reference in empirical work in financial

economics. Despite their popularity, those estimators perform poorly when dealing
with the problem of errors in variables. This is a serious problem since in most
empirical applications, one might raise concerns about issues such as data quality
and measurement errors.
This chapter uses Monte Carlo simulations and real data to assess the perfor-
mance of different estimators that deal with measurement error, including EW’s
higher-order moment estimator and alternative instrumental variable-type
approaches. We show that in the presence of individual fixed effects, under
heteroscedasticity, or in the absence of high degree of skewness in the data, the
EW estimator returns biased coefficients for both mismeasured and perfectly mea-
sured regressors. The IV estimator requires assumptions about the autocorrelation
structure of the measurement error, which we characterize and discuss in the chapter.
We also estimate empirical investment models using the two methods. Because
real-world investment data contain firm-fixed effects and heteroscedasticity, the
EW estimator delivers coefficients that are unstable across different specifications
and not economically meaningful. In contrast, a simple OLS-IV estimator yields
results that conform to theoretical expectations. We conclude that real-world
investment data is likely to satisfy the assumptions that are required for identifica-
tion of OLS-IV but that the presence of heteroscedasticity and fixed effects causes
the EW estimator to return biased coefficients.

28
All of the F-statistics associated with the first-stage regressions have p-values that are close to
zero. These statistics (reported in Table 57.10) suggest that we do not incur a weak instrument
problem when we use longer lags in our instrumental set.
1614 H. Almeida et al.

Table 57.11 EW coefficients for real data (within transformation)

q coefficient Cash flow coefficient
Year GMM3 GMM4 GMM5 GMM3 GMM4 GMM5
1973 0.029 0.000 0.000 0.347 0.265 0.264
(0.075) (0.073) (4.254) (0.207) (0.207) (11.968)
1974 0.050 0.029 0.019 0.168 0.199 0.214
(0.037) (0.012) (0.016) (0.073) (0.043) (0.043)
1975 0.225 0.001 0.000 0.161 0.292 0.292
(0.475) (0.149) (0.125) (0.281) (0.095) (0.094)
1976 0.137 0.001 0.000 0.156 0.276 0.276
(0.094) (0.273) (0.042) (0.090) (0.251) (0.048)
1977 0.082 0.243 0.000 0.203 0.091 0.261
(0.263) (0.109) (0.108) (0.179) (0.090) (0.083)
1978 0.263 0.514 0.281 0.122 (0.067) 0.108
(0.282) (0.927) (0.146) (0.224) (0.689) (0.125)
1979 0.020 0.001 0.001 0.249 0.266 0.266
(0.161) (0.048) (0.031) (0.155) (0.056) (0.044)
1980 0.349 0.116 0.183 0.021 0.219 0.163
(0.294) (0.071) (0.055) (0.273) (0.074) (0.067)
1981 0.334 0.185 0.324 0.145 0.061 0.131
(0.165) (0.045) (0.128) (0.248) (0.093) (0.191)
1982 0.109 0.383 0.238 0.125 0.206 0.031
(0.155) (0.316) (0.126) (0.195) (0.398) (0.174)
1983 0.081 0.001 0.001 0.132 0.184 0.184
(0.037) (0.041) (0.059) (0.033) (0.034) (0.040)
1984 0.230 0.210 0.185 0.125 0.138 0.154
(0.083) (0.050) (0.043) (0.067) (0.052) (0.048)
1985 0.198 0.349 0.230 0.050 (0.018) 0.035
(0.483) (0.137) (0.024) (0.212) (0.086) (0.032)
1986 0.672 0.244 0.593 0.179 0.070 0.133
(0.447) (0.089) (0.162) (0.303) (0.079) (0.128)
1987 0.102 0.104 0.115 0.078 0.078 0.077
(0.039) (0.020) (0.003) (0.021) (0.021) (0.020)
1988 0.129 0.179 0.148 0.030 0.027 0.029
(0.051) (0.029) (0.014) (0.011) (0.007) (0.007)
1989 0.365 0.015 0.111 0.285 0.162 0.196
(1.797) (0.082) (0.196) (0.642) (0.063) (0.078)
1990 0.437 0.419 0.529 0.395 0.386 0.440
(0.404) (0.137) (0.024) (0.214) (0.094) (0.093)
1991 0.384 0.260 0.240 0.098 0.007 0.023
(0.225) (0.105) (0.038) (0.199) (0.099) (0.055)
1992 0.105 0.102 0.040 0.086 0.088 0.148
(0.016) (0.008) (0.016) (0.034) (0.033) (0.037)
1993 0.274 0.322 0.452 0.076 0.118 0.232
(0.394) (0.352) (0.273) (0.360) (0.297) (0.276)
(continued)
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1615

Table 57.11 (continued)

q coefficient Cash flow coefficient
Year GMM3 GMM4 GMM5 GMM3 GMM4 GMM5
1994 0.110 4.436 0.047 0.255 3.550 0.207
(0.136) (86.246) (0.011) (0.108) (65.488) (0.045)
1995 0.574 8.847 2.266 0.537 5.898 1.633
(1.862) (145.827) (5.565) (1.275) (94.154) (3.749)
1996 0.220 0.167 0.196 0.101 0.106 0.103
(0.068) (0.022) (0.013) (0.036) (0.033) (0.030)
1997 0.089 0.177 0.158 0.059 0.020 0.028
(0.082) (0.042) (0.021) (0.042) (0.041) (0.034)
1998 0.620 0.245 0.119 0.688 0.355 0.242
(1.634) (0.187) (0.027) (1.446) (0.169) (0.037)
1999 0.031 0.003 0.000 0.160 0.126 0.123
(0.059) (0.028) (0.055) (0.074) (0.038) (0.068)
2000 0.071 0.126 0.118 0.032 0.029 0.021
(0.024) (0.030) (0.020) (0.043) (0.057) (0.051)
2001 0.050 0.077 0.055 0.034 0.020 0.031
(0.021) (0.020) (0.013) (0.016) (0.016) (0.012)
2002 0.047 0.048 0.048 0.030 0.030 0.030
(0.128) (0.016) (0.014) (0.033) (0.013) (0.012)
2003 0.131 0.066 0.157 0.013 0.025 0.027
(0.043) (0.025) (0.010) (0.031) (0.026) (0.014)
2004 0.005 0.030 0.030 0.092 0.079 0.079
(0.066) (0.018) (0.009) (0.045) (0.034) (0.034)
2005 0.049 0.029 0.026 0.078 0.095 0.098
(0.025) (0.009) (0.011) (0.040) (0.032) (0.034)
Fama–MacBeth standard error 0.0679 0.3031 0.0232 0.1299 0.3841 0.1554
0.0455 0.3018 0.0787 0.0310 0.2027 0.05222
This table shows the coefficients and standard deviations that we obtain when we use the EW
estimator in Eq. 57.22, estimated year by year. The table also shows the results for the EW
estimator associated with GMM3, GMM4, and GMM5. The table also shows the EW coefficients
for the data that is treated for fixed effects via the within transformation. The data are taken from
the annual Compustat industrial files over the 1970–2005 period. See text for details

References
Agca, S., & Mozumdar, A. (2007). Investment-cash flow sensitivity: Myth or reality? Working
paper, George Washington University.
Almeida, H., & Campello, M. (2007). Financial constraints, asset tangibility and corporate
investment. Review of Financial Studies, 20, 1429–1460.
Altonji, J., & Segal, L. (1996). Small-sample bias in GMM estimation of covariance structures.
Journal of Business & Economic Statistics, 14, 353–366.
Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence
and an application to employment equations. Review of Economic Studies, 58, 277–297.
1616 H. Almeida et al.

Bakke, T., & Whited, T. (2009). What gives? A study of firms’ reactions to cash shortfalls.
Working paper, University of Rochester.
Baltagi, B. (2005). Econometric analysis of panel data (3rd ed.). Chichester: Wiley.
Baum, C., Schaffer, M., & Stillman, S. (2003). Instrumental variables and GMM: Estimation and
testing. Stata Journal, 3, 1–31.
Bertrand, M., & Mullainathan, S. (2005). Bidding for oil and gas leases in the Gulf of Mexico:
A test of the free cash flow model?. mimeo, University of Chicago and MIT.
Bertrand, M., & Schoar, A. (2003). Managing with style: The effect of managers on firm policies.
Quarterly Journal of Economics, 118, 1169–1208.
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-in-
differences estimates? Quarterly Journal of Economics, 119, 249–275.
Biorn, E. (2000). Panel data with measurement errors: Instrumental variables and GMM pro-
cedures combining levels and differences. Econometric Reviews, 19, 391–424.
Blanchard, O., Lopez-de-Silanes, F., & Shleifer, A. (1994). What do firms do with cash windfalls?
Journal of Financial Economics, 36, 337–360.
Blundell, R., Bond, S., Devereux, M., & Schiantarelli, F. (1992). Investment and Tobin’s Q:
Evidence from company panel data. Journal of Econometrics, 51, 233–257.
Bond, S., & Cummins, J. (2000). The stock market and investment in the new economy. Brookings
Papers on Economic Activity, 13, 61–124.
Chirinko, R. (1993). Business fixed investment spending: Modeling strategies, empirical results,
and policy implications. Journal of Economic Literature, 31, 1875–1911.
Colak, G., & Whited, T. (2007). Spin-offs, divestitures, and conglomerate investment. Review of
Financial Studies, 20, 557–595.
Cragg, J. (1997). Using higher moments to estimate the simple errors-in-variables model. RAND
Journal of Economics, 28, 71–91.
Doran, H., & Schmidt, P. (2006). GMM estimators with improved finite sample properties using
principal components of the weighting matrix, with an application to the dynamic panel data
model. Journal of Econometrics, 133, 387–409.
Erickson, T., & Whited, T. (2000). Measurement error and the relationship between investment
and Q. Journal of Political Economy, 108, 1027–1057.
Erickson, T., & Whited, T. (2002). Two-step GMM estimation of the errors-in-variables model
using high-order moments. Econometric Theory, 18, 776–799.
Fazzari, S., Hubbard, R. G., & Petersen, B. (1988). Financing constraints and corporate invest-
ment. Brooking Papers on Economic Activity, 1, 141–195.
Frees, E. (2004). Longitudinal and panel data analysis and applications in the social sciences.
New York: Cambridge University Press.
Gan, J. (2007). Financial constraints and corporate investment: Evidence from an exogenous shock
to collateral. Journal of Financial Economics, 85, 709–734.
Gomes, J. (2001). Financing investment. American Economic Review, 91, 1263–1285.
Griliches, Z., & Hausman, J. A. (1986). Errors in variables in panel data. Journal of Econometrics,
31, 93–118.
Hadlock, C. (1998). Ownership, liquidity, and investment. RAND Journal of Economics, 29, 487–508.
Hayashi, F. (1982). Tobin’s marginal q and average q: A neoclassical interpretation.
Econometrica, 50, 213–224.
Hayashi, F. (2000). Econometrics. Princeton: Princeton University Press.
Hennessy, C. (2004). Tobin’s Q, debt overhang, and investment. Journal of Finance, 59,
1717–1742.
Holtz-Eakin, D., Newey, W., & Rosen, H. S. (1988). Estimating vector autoregressions with panel
data. Econometrica, 56, 1371–1395.
Hoshi, T., Kashyap, A., & Scharfstein, D. (1991). Corporate structure, liquidity, and investment:
Evidence from Japanese industrial groups. Quarterly Journal of Economics, 106, 33–60.
Hubbard, R. G. (1998). Capital market imperfections and investment. Journal of Economic
Literature, 36, 193–227.
57 Assessing the Performance of Estimators Dealing with Measurement Errors 1617

Kaplan, S., & Zingales, L. (1997). Do financing constraints explain why investment is correlated
with cash flow? Quarterly Journal of Economics, 112, 169–215.
Lamont, O. (1997). Cash flow and investment: Evidence from internal capital markets. Journal of
Finance, 52, 83–110.
Lyandres, E. (2007). External financing costs, investment timing, and investment-cash flow
sensitivity. Journal of Corporate Finance, 13, 959–980.
Malmendier, U., & Tate, G. (2005). CEO overconfidence and corporate investment. Journal of
Finance, 60, 2661–2700.
Petersen, M. (2009). Estimating standard errors in finance panel data sets: Comparing approaches.
Review of Financial Studies, 22, 435–480.
Poterba, J. (1988). Financing constraints and corporate investment: Comment. Brookings Papers
on Economic Activity, 1, 200–204.
Rauh, J. (2006). Investment and financing constraints: Evidence from the funding of corporate
pension plans. Journal of Finance, 61, 33–71.
Riddick, L., & Whited, T. (2009). The corporate propensity to save. Journal of Finance, 64,
1729–1766.
Shin, H., & Stulz, R. (1998). Are internal capital markets efficient? Quarterly Journal of Eco-
nomics, 113, 531–552.
Stein, J. (2003). Agency information and corporate investment. In G. Constantinides, M. Harris, &
R. Stulz (Eds.), Handbook of the economics of finance. Amsterdam: Elsevier/North-Holland.
Wansbeek, T. J. (2001). GMM estimation in panel data models with measurement error. Journal of
Econometrics, 104, 259–268.
Whited, T. (2001). Is it inefficient investment that causes the diversification discount? Journal of
Finance, 56, 1667–1691.
Whited, T. (2006). External finance constraints and the intertemporal pattern of intermittent
investment. Journal of Financial Economics, 81, 467–502.
Xiao, Z., Shao, J., & Palta, M. (2008). A unified theory for GMM estimation in panel data models
with measurement error. Working paper, University of Wisconsin Madison.
Realized Distributions of Dynamic
Conditional Correlation and Volatility 58
Thresholds in the Crude Oil, Gold, and
Dollar/Pound Currency Markets

Tung-Li Shih, Hai-Chin Yu, Der-Tzon Hsieh, and Chia-Ju Lee

Contents
58.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1620
58.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1622
58.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1623
58.4 Dynamic Conditional Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1625
58.4.1 Dynamic Conditional Correlation Between Gold, Oil, and
the Dollar/Pound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1625
58.4.2 Empirical Results of Dynamic Conditional Correlation . . . . . . . . . . . . . . . . . . . . . 1628
58.4.3 Volatility Threshold Dynamic Conditional Correlation . . . . . . . . . . . . . . . . . . . . . 1636
58.4.4 Does Investors’ Behavior Change over Subperiods? . . . . . . . . . . . . . . . . . . . . . . . . 1639
58.5 Conclusions and Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1643
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1644

Abstract
This chapter proposes a modeling framework for the study of co-movements in
price changes among crude oil, gold, and dollar/pound currencies that are
conditional on volatility regimes. Methodologically, we extend the dynamic

T.-L. Shih
Department of Hospitality Management, Ming Dao University, Changhua Peetow, Taiwan
e-mail: tungli@mdu.edu.tw
H.-C. Yu (*)
Department of International Business, Chung Yuan University, Chungli, Taiwan
e-mail: haichin@cycu.edu.tw; haichinyu@hotmail.com
D.-T. Hsieh
Department of Economics, National Taiwan University, Taipei, Taiwan
e-mail: dthsieh@ccms.ntu.edu.tw
C.-J. Lee
College of Business, Chung Yuan University, Chungli, Taiwan
e-mail: g9604601@cycu.edu.tw

C.-F. Lee, J. Lee (eds.), Handbook of Financial Econometrics and Statistics, 1619
DOI 10.1007/978-1-4614-7750-1_58,
# Springer Science+Business Media New York 2015
1620 T.-L. Shih et al.

conditional correlation (DCC) multivariate GARCH model to examine the

volatility and correlation dynamics depending on the variances of price returns
involving a threshold structure. The results indicate that the periods of market
turbulence are associated with an increase in co-movements in commodity (gold
and oil) prices. By contrast, high market volatility is associated with a decrease
in co-movements between gold and the dollar/pound or oil and the dollar/pound.
The results imply that gold may act as a safe haven against major currencies
when investors face market turmoil. By looking at different subperiods based on
the estimated thresholds, we find that the investors’ behavior changes in different
subperiods. Our model presents a useful tool for market participants to engage in
better portfolio allocation and risk management.

Keywords
Dynamic conditional correlation • Volatility threshold • Realized distribution •
Currency market • Gold • Oil

58.1 Introduction

Commodity markets in recent years have experienced dramatic growth in trading

volume as well as widespread price volatility. With few exceptions, most of the
commodities have experienced an impressive bull run and have generally
outperformed traditional investments. For example, the prices of commodities
such as crude oil have risen dramatically, and the crude oil price almost reached
a new high of US$200 per barrel in 2011. In the meantime, the price of gold hit
a new high of US$1,700 in 2011. These price surprises have influenced not only the
commodity markets but also the currency markets and the international parity of
foreign exchange. By the fall of 2007, the increasing speculation in commodity
markets was associated with the devaluation of the US dollar.
Among these commodities, gold appears to have exhibited a more stable price
trend than crude oil. From the beginning of the financial crisis in 1997 up until
2011, the price of gold has risen by almost 42 %. For many years, gold has been
viewed as a safe haven from market turbulence. However, very few empirical
studies have examined the role of gold as a safe-haven asset and even fewer have
examined gold’s safe-haven role with respect to major currency exchange rates,
especially those of the two major currencies – the US dollar and the British pound.
The reason why we choose exchange rates as a comparative baseline is that, for
commodities that are traded continuously in organized markets, a change in a major
currency exchange rate will result in an instant adjustment in the prices of com-
modities in at least one currency and perhaps in both currencies if both countries are
“large.” For instance, when the dollar depreciates against the pound, the dollar
prices of commodities tend to rise (and pound prices fall) even though the funda-
mentals of the markets remain unchanged.
58 Realized Distributions of Dynamic Conditional Correlation 1621

This widely expanded and complex volatility in commodity prices increases the
importance of modeling real volatility and correlation, because a good estimate
helps facilitate portfolio optimization, risk management, and hedging activities.
Although some of the literature assumes volatility and correlation to be constant
in the past years, it is widely recognized that they indeed vary over time. This
recognition has spurred a vibrant body of work regarding the dynamic properties of
market volatility. To date, very little is known about the volatility dynamics
between the commodity and currency markets, for instance, in the case of gold
and its possible correlations with oil and major currencies. This chapter intends to
address this gap.
The main purpose of this study is to examine the dynamic relationships among
gold, oil, and the dollar/pound to further understand the hedging ability of gold
relative to another commodity or currency. That is to say, if gold acts as a financial
safe haven against the dollar (or oil), it allows for systematic feedback between
changes in the price of gold, oil, and the dollar/pound exchange rate. Specifically,
this chapter asks, does gold act as a safe haven against the dollar/pound, as a hedge,
or as neither? Are gold and oil highly correlated with each other? Movements in the
price of gold, oil, and the dollar/pound are analyzed using a model of dynamic
conditional correlation covering 20 years of daily data.
Studies related to this issue are few. Capie et al. (2005) point out that gold acts as
an effective hedge against the US dollar by estimating elasticity relative to changes
in the exchange rate. However, their approach involves the use of a single-equation
model in which the exchange rate is assumed to be unaffected by the time of
the dependent variable, the price of gold. Our chapter improves their work
by employing a dynamic model of conditional correlations in which all variables
are treated symmetrically. Besides, although Baur and Lucey (2010) find evidence
in support of gold providing a haven from losses incurred in the bond and
stock markets, they neglect the interactions with the currency market and,
like Capie et al. (2005), do not consider feedback in their model of returns. Nikos
(2006) uses correlation analysis to estimate the correlation of returns between gold
and the dollar and shows that the correlation between the dollar and gold is
0.19
and
0.51 for two different periods. These findings imply that gold is
a contemporaneous safe haven in extreme currency market conditions. Steinitz
(2006) utilizes the same method to estimate the correlations of weekly
returns between gold and Brent oil for two periods of 1 year and 5 years,
respectively, and shows that the correlations between gold and Brent oil are 0.310
and 0.117, respectively. Boyer and Fillion (2007) report on the financial
determination of Canadian oil and gas stock returns and conclude that
a weakening of the Canadian dollar against the US dollar has a negative impact
on stock returns.
If correlations and volatilities vary over time, the hedge ratio should be adjusted
to account for the new information. Other work, such as Baur and McDermott
(2010), similarly neglects feedback in its regression model. Further studies
1622 T.-L. Shih et al.

investigate the concept of a safe-haven asset without reference to gold. For example,
Ranaldo and Soderlind (2009) and Kaul and Sapp (2006) examine safe-haven
currencies, while Upper (2000) examines German government bonds as safe-haven
instruments. Andersen et al. (2007) show that exchange rate volatility outstrips bond
volatility in the US, British, and German markets. Thus, currency risk is worth
exploring and being hedged.
As a general rule, commodities are priced in US dollars. Since the US currency
has weakened that a bull run of commodity prices appeared, the question arises as to
which the increases in commodity prices have been a product of the depreciation in
the US dollar. Furthermore, it would be interesting to examine how to provide
a hedge against the dollar that varies across different commodities. It also needs to
be asked which investment instruments are more suitable for diversification pur-
poses to protect against changes in the US currency.
This chapter investigates the following issues. First, how do the time-varying
correlations and associated distributions appear in the crude oil, gold, and dollar/
pound markets? Second, what is the shape of each separate distribution of various
volatility levels among the crude oil, gold, and dollar/pound markets? Third, by
employing the volatility threshold DCC model put forward by Kasch and Caporin
(2012), is the high volatility (exceeding a specified threshold) of the assets associ-
ated with an increasing degree of correlation?
We find that the volatility thresholds of oil and gold correspond to two major
events – the First Gulf War in 1990 and the 911 event in 2001. We also find that
the increase in commodity (crude oil and gold) prices was a reflection of the falling
US dollar, especially after the 911 event. The evidence shows that the DCC
between crude oil and gold was 0.1168, while those for the gold/dollar/pound and
oil/dollar/pound markets were
0.2826 and
0.0369, respectively, with the latter
being significantly higher than in the other subperiods.
This remainder of this chapter is organized as follows. Section 58.2 provides
a review of the literature. Section 58.3 describes the data and summary statistics for
crude oil, gold, and the dollar/pound exchange rate. Section 58.4 presents the
dynamic conditional correlation model and reports the results of its volatility
threshold. And also provides the results for subperiods separated by the thresholds
found. Finally, Sect. 58.5 discusses the results and concludes.

58.2 Literature Review

Engle et al. (1994) investigate how the returns and volatilities of stock indices
between Tokyo and New York are correlated and find that, except for a lagged
return spillover from New York to Tokyo after the crash, there was no significant
lagged spillover in returns or in volatilities. Ng (2000) examines the size and the
impact of volatility spillover from Japan and the USA to six Pacific Basin equity
markets. Using four different specifications of correlation by constructing volatility
spillover models, he distinguishes the volatility between local idiosyncratic shock,
regional shock from Japan, and global shock from the USA and finds significant
58 Realized Distributions of Dynamic Conditional Correlation 1623

spillover effects from regional to Pacific Basin economies. Andersen et al. (2001a)
find strong evidence that volatilities and correlations move together in a manner
broadly consistent with the latent factor structure. Andersen et al. (2001b) found
that volatility movements are highly correlated across the deutsche mark and yen
against the US dollar. Furthermore, the correlation between the two exchange rates
increases with volatility. Engle (2002) finds that the breakdown of the correlations
between the deutsche mark and the pound and lira in August 1992 is very apparent.
In addition, after the euro is launched, the estimated currency correlation essentially
moves to 1.
Recently, Doong et al. (2005) examined the dynamic relationship and pricing
between stocks and exchange rates for six Asian emerging markets. They found that
the currency depreciation is accompanied by a fall in stock prices. The conditional
variance-covariance process of changes in stock prices and exchange rates is time
varying. Lanza et al. (2006) estimate the dynamic conditional correlations in the
daily returns for West Texas Intermediate (WTI) oil forward and future prices from
January 3, 1985 to January 16, 2004, and find that the dynamic conditional
correlations vary dramatically. Chiang et al. (2009) investigate the probability
distribution properties, autocorrelations, dynamic conditional correlations, and
scaling analysis of Dow-Jones and NASDAQ Intraday returns from August 1,
1997 to December 31, 2003. They find the correlations to be positive and to mostly
fluctuate in the range of 0.6–0.8. Furthermore, the variance of the correlation
coefficients has been declining and appears to be stable during the post-2001
period. Pérez-Rodrı́guez (2006) applies a multivariate DCC-GARCH technique to
examine the structure of the short-run dynamics of volatility returns on the euro,
yen, and British pound against the US dollar over the period from 1999 to 2004 and
finds strong dynamic relationships between currencies. Tastan (2006) applies
multivariate GARCH to capture the time-varying variance-covariance matrix for
stock market returns (Dow-Jones Industrial Average Index and S&P500 Index) and
changes in exchange rates (euro/dollar exchange rates). He also plots news impact
surfaces for variances, covariances, and correlation coefficients to sort out the
effects of shocks. Chiang et al. (2007a) apply a dynamic conditional correlation
model to nine Asian daily stock-return series from 1990 to 2003 and find evidence
of a contagion effect and herding behavior. Chiang et al. (2007b) examine A-share
and B-share market segmentation conditions by employing a dynamic multivariate
GARCH model and show that stock returns in both A- and B-shares are positively
correlated with the daily change in trading volume or abnormal volume.

58.3 Data

Our data consist of the daily prices of crude oil and gold, and the US dollar/British
pound exchange rate, and are obtained from the AREMOS database over the period
from January 1, 1986 to December 31, 2007 for a total of 5,165 observations.
The West Texas Intermediate crude oil price is chosen to represent the oil
spot market, and the price of 99.5 % fine gold, the London afternoon fixing,
1624 T.-L. Shih et al.

The Price Movement of Crude Oil and Gold Market

100 900
Oil Gold 800
80
700
60 600

40 500
400
20
300
0 200
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006
The Price Movement of Foreign Exchange Market
0.75

0.70

0.65
USD/Pound

0.60

0.55

0.50

0.45
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006
Time

Fig. 58.1 The price movement for the sampled markets from January 1, 1986 to December
31, 2007, for a total of 5,405 observations

is chosen to represent the gold spot market. The daily dollar/pound exchange rate,
which represents the major currencies, is selected to estimate the volatility of
the FX market.
In our sample period, the crude oil price was testing the $100 per barrel threshold
by November 2007. Meanwhile, the price of gold was relatively stable varying
between $415 and $440 per ounce from January to September 2005. However, in
the fourth quarter of 2005, the gold price jumped dramatically and hit $500 per
ounce. In April 2006, the gold price broke through the $640 level. 2007 was a strong
year, with the price steadily rising from $640 on January 2 with a closing London
fixed price of over $836 on December 31, 2007. Since then, prices have continued
to increase to reach new record highs of over $1,700 in 2011.
Figure 58.1 displays the price movements for oil, gold, and the dollar/pound over
the sample period. As shown in Fig. 58.1, gold traded between a low of $252
(August 1999) and a high of $836 (December 31, 2007) per ounce at the fixing,
while oil traded between a low of $10 (in late 1998, in the wake of the Asian
58 Realized Distributions of Dynamic Conditional Correlation 1625

Table 58.1 Summary statistics of the daily returns among crude oil, gold, and dollar/pounda
(January 1, 1986 to December 31, 2007)
Crude oil Gold Dollar/pound
Mean 0.024 0.017
0.006
(0.536) (0.159) (0.4769)
Max 0.437 0.070 0.0379
Min
0.404
0.063
0.0329
Standard dev. 0.029 0.009 0.006
Skewnessb
0.012
0.031 0.164**
(0.729) (0.354) (0.0000)
Kurtosisb 37.915** 5.968** 2.439**
(0.0000) (0.0000) (0.0000)
Jarque-Berac 323,692.99** 8,019.20** 1,363.59**
(0.0000) (0.0000) (0.0000)
a
The table summarizes the daily returns of estimates for the West Texas Intermediate crude oil,
gold, and dollar/pound markets. The sample covers the period from January 1, 1986 through
December 31, 2007 for a total of 5,405 observations
b
The three markets are far away from the skewness and kurtosis of 0 and 3, respectively, implying
that the three markets are not normally distributed
c
Jarque-Bera is the Jarque-Bera test statistic, distributed w22
**Denotes significance at the 0.05 level

Financial Crisis and the United Nations’ oil-for-food program) and a high of $99.3
(November 2007) per barrel. These large variations in the price of both gold and oil
indicate that the DCC and realized distribution are better approaches for detecting
the trading pattern of investors.
Table 58.1 reports the statistics of daily returns for crude oil, gold, and the dollar/
pound exchange rate. The daily returns are calculated as the first differences of the
natural log of the prices times 100. The results show that the crude oil has the
highest return, followed by gold and the dollar/pound.

58.4 Dynamic Conditional Correlation

58.4.1 Dynamic Conditional Correlation Between Gold, Oil, and the

Dollar/Pound

It is generally recognized that financial markets are highly integrated in terms of

price movements, since prices soaring in one market can spill over to another
market instantly. One simple method to explore the relationship between the two
markets is to calculate the correlation coefficient. We then specify a multivariate
model, which is capable of computing the dynamic conditional correlation (DCC)
that is capable of capturing ongoing market elements and shocks. The DCC model
is specified as Eq. 58.1.
1626 T.-L. Shih et al.

Table 58.2 The correlation among crude oil, gold, and FX of dollar/pound (January 1, 1986 to
December 31, 2007)
Oil Gold FX
Oil 1
Gold 0.7488 1
FX
0.5592
0.6260 1


Et
1 r i,t r j,t
rij,t ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
   ffi (58.1)
Et
1 r 2i,t Et
1 r 2j,t

where the conditional correlation rij,t is based on information known in the

previous period Et
1 and i, j represent the three markets 1, 2, and 3. Based on the
laws of probability, all correlations defined in this way must lie within the interval
[
1, 1]. This is different from the constant correlation we have usually used and
assumed throughout a given period. To clarify the relationship between the
conditional correlations and conditional variances, it is convenient to express the
returns as the conditional standard deviation times the standardized disturbance as
suggested by Engle (2002) in Eq. 58.2 below:
  pffiffiffiffiffiffiffiffiffiffiffiffi
hi,t ¼ Et
1 r 2i,t , r i,t ¼ hi , t e i , t , i ¼ 1, 2, 3 (58.2)

Since the correlation coefficients among crude oil, gold, and dollar/pound FX
markets provide useful measures of the long-term relationship between each pair of
markets, Table 58.2 presents a simple correlation matrix in which the calculation is
based on the constant coefficient given by Eq. 58.1. Some preliminary information
is obtained below. First, the crude oil and gold are highly correlated with
a coefficient of 0.7488, a result that is in line with Steinitz (2006). Secondly, both
gold and crude oil are highly negatively related to the dollar/pound with coefficients
of
0.6260 and
0.5592, respectively, which is consistent with the report of
Nikos (2006).
As the autoregressive conditional heteroskedasticity (ARCH) model has become
the most useful model in investigating the conditional volatility since Engle (1982),
we then follow this model in our analysis. The ARCH model adopts the effect of
past residuals that helps explain the phenomenon of volatility clustering. Bollerslev
(1986) proposed the generalized autoregressive conditional heteroskedasticity
(GARCH) model, which has created a new field in the research on volatility and
is widely used in financial and economic time series. Some of his research attempts
to discuss the effects of more than one variable simultaneously. For instance,
Bollerslev (1990) proposed the constant conditional correlation (CCC) model
which makes a strong assumption, namely, that the correlation among the variables
remains constant in order to simplify the estimation. Engle (2002) later proposed
58 Realized Distributions of Dynamic Conditional Correlation 1627

a dynamic conditional correlation (DCC) model, which allows the correlation to be

time varying and, by involving fewer complicated calculations, is capable of
dealing with numerous variables.
In this chapter, we follow the Engle (2002) approach, which has clear compu-
tational advantages over multivariate GARCH models in that the number of
parameters to be estimated remains constant and loosens the assumptions of the
multivariate conditional correlations, in order to develop the dynamic conditional
correlation (DCC) model. The DCC model can be viewed as a generalization of the
Bollerslev (1990) constant conditional correlation (CCC) estimator. It differs only
in that it allows the correlation to be time varying, which parameterizes the
conditional correlations directly. The estimation takes place in two steps, in that
a series of univariate GARCH estimates are first obtained followed by the
correlation coefficients. The characteristics of the DCC model are that the
multivariate conditional correlations are dynamic and not constant and confirm
that the real conditional correlations of financial assets in general and the time-
varying covariance matrices can be estimated. This model involves a less compli-
cated calculation without losing too much generality and is able to deal with
numerous variables.
Following Engle (2002) and Chiang et al. (2009), the mean equation is
assumed to be represented by Eq. 58.1, where the multivariate conditional
variance is given by
H t,t ¼ Dt,t V t,t Dt,t , (58.3)

where t is a time interval, which can be a day, an hour, or 1 min. Here t is a daily
interval. Vt,t is a symmetric conditional correlation matrix of et and Dt,t is a (2  2)
matrix with the conditional variances ht,ii,t for two stock returns (where i ¼ gold,
oil, or the dollar/pound exchange rate) on the diagonal. That is,
hqffiffiffiffiffiffiffiffiffiffi i
Dt,t ¼ diag s2t,ii,t . Equation 58.3 suggests that the dynamic properties of
ð2;2Þ
the covariance matrix Ht,t are determined by Dt,t and Vt,t for a given t, a time
interval that can be 1 min, 1 day, or 1 week and so on. The DCC model proposed by
Engle (2002) involves a two-stage estimation of the conditional covariance matrix
Ht in Eq. 58.3. In the first stage, univariate volatility models are fitted for each of the
qffiffiffiffiffiffiffiffiffiffiffi
returns, and estimates of s2t, ii, t ði ¼ 1, 2, and 3Þ are obtained by using Eq. 58.4.
In the second stage, return residuals are transformed by their estimated standard
qffiffiffiffiffiffiffiffiffiffi
deviations from the first stage. That is t,i,t ¼ et,i,t = s2t,ii,t , where t,i,t is used to
estimate the parameters of the conditional correlation. The evolution of the corre-
lation in the DCC model is given by Eq. 58.5:

s2t,ii,t ¼ ct,i þ at, i e2t,i,t
1 þ bt, i s2t,ii,t
1 , i ¼ 1, 2 (58.4)


Qt,t ¼ 1
at,i
bt,i Qt þ at,i t, i,t
1 0t, i,t
1 þ bt,i Qt,t
1 , (58.5)
1628 T.-L. Shih et al.

where
h Qt,t ¼i (qt,ij,t) is the 2  2 time-varying covariance matrix of t, i, t , Qt ¼
E t, i, t 0t, i, t is the 2  2 unconditional variance matrix of t,i,t, and at,i and bt,i are
non-negative scalar parameters satisfying (at,i + bt,i) < 1. Since Qt does
not generally have ones on its diagonal, we scale it to obtain a proper correlation
matrix Vt,t Thus,



1=2


1=2
V t, t ¼ diag Qt, t Qt, t diag Qt, t , (58.6)



1=2  
pffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffi
where diag Qt, t ¼ diag 1= qt,11,t , 1= qt,22,t :
Here Vt,t in Eq. 58.6 is a correlation matrix with ones on the diagonal and
off-diagonal elements of less than one in absolute value terms, as long as Qt,t is
positive definite. A typical element of Vt,t takes the form:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rt, 12, t ¼ qt, 12, t = qt, 11, t qt, 22, t (58.7)

The dynamic correlation coefficient, rt,12,t, can be obtained by using the element
of Qt,t in Eq. 58.5, which is given by Eq. 58.8 below:

 0
qt, ij, t ¼ 1
at, i
bt, i rt, ij þ at, i t, i, t
1 t, j, t
1 þ bt, i qt, ij, t
1 , (58.8)

The mean reversion requires that (at,i + bt,i) < 1. In general terms, the essence of
this concept is the assumption that both an asset’s high and low prices are temporary
and that the asset’s price will tend to move toward the average price over time.
Besides, the estimates of the dynamic correlation coefficients, rij,t, between each
pair of the three markets have been specified as in Eq. 58.1.

58.4.2 Empirical Results of Dynamic Conditional Correlation

In this section, we present the estimation results of the models outlined above. The
estimation results are presented in Table 58.1, which provides the dynamic corre-
lations of returns across crude oil, gold, and the dollar/pound foreign exchange rate
with each other. The estimated a and b for three markets are listed in Tables 58.3
and 58.4. The likelihood ratio does not support the rejection of the null hypothesis
of the scalar dynamic conditional correlation. It can be seen that the sum of the
estimated coefficients in the variance equations (a + b) is close to 1 for all of the
cases, implying that the volatility appears to be highly persistent. As for the Ljung-
Box Q-statistic of the serial correlation of the residuals, the results show that that
the serial correlations in the error series are regarded as adequate.
Calvert et al. (2006) observed that through the dynamic conditional correlation
distribution, we can more fully understand the real impacts in international markets.
58 Realized Distributions of Dynamic Conditional Correlation 1629

Table 58.3 DCC estimates: three marketsa (January 1, 1986 to December 31, 2007)
DCC
a 0.0202** b 0.9651**
(52.3020) (1,050.395)
a
The t-statistic is given in parentheses
**Denotes significance at the 0.05 level

Table 58.4 Estimation results from the DCC-GARCH modela

Mean equation Variance equation

Constant a b Persistence Ljung-Box Q-statistic
Oil 0.0004** 0.1399** 0.8306** 0.9705 185.7203**
(1.6549) (21.6431) (143.5303)
Gold
1.54E-06** 0.0664** 0.9272** 0.9936 61.8887**
(
0.1658) (15.0320) (202.223)
FX
0.0001** 0.0417** 0.9479** 0.9896 32.6628
(
1.6045) (9.2628) (168.722)
a
The persistence level of the variance is calculated as the summation of the coefficients in the
variance equations (a + b). The z-statistic is given in parentheses. The Ljung-Box Q-statistic tests
the serial correlation of the residuals
**
Denotes significance at the 0.05 level

It can also help with portfolio selection and risk management. In Fig. 58.2, which
reports the results of the dynamic conditional correlation, the estimated correlation
coefficients are time varying, reflecting some sort of portfolio shift between each
two items.
The correlation between crude oil and gold was estimated using the DCC
integrated method, and the results, shown in Fig. 58.2a, are quite interesting.
The correlations are found to be generally positive around 0.2 except for
mid-1990 which turns out to be highly correlated with a correlation of around
0.6605. The possible interpretation for the high correlation is due to the Iraqi
invasion of Kuwait and the Gulf War. The crude oil price jumped from about $15
to over $33 per barrel during that time, so that investors channeled their money into
the gold market because of their fear of inflation. This fact accords with the “flight
to quality” concept, which represents the action of investors moving their capital
away from riskier or more volatile assets to the ones considered to be safer and less
volatile. The correlation between gold and the dollar/pound exchange rate is shown
in Fig. 58.2b for the integrated DCC in the last 20 years. Whereas for most of the
period the correlations were between
0.1 and
0.3, there were two notable drops,
where the stock market crashed in October 1987 and in late 2002, and we also find
two peaks, one in the middle of 1990 and the other in late 1998 where the gold price
dropped to $252 per ounce. Fig. 58.2c shows the correlation between crude oil and
1630 T.-L. Shih et al.

a 0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
−0.1
−0.2
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006
b 0.3
0.2
0.1
−0.0
−0.1
−0.2
−0.3
−0.4
−0.5
−0.6
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006
c 0.40
0.32
0.24
0.16
0.08
0.00
− 0.08
− 0.16
− 0.24
− 0.32
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006

Fig. 58.2 The time series of dynamic conditional correlations (DCC) among each pair of three
markets: (a) daily DCC between crude oil and gold, (b) daily DCC between gold and dollar/pound,
and (c) daily DCC between crude oil and dollar/pound
58 Realized Distributions of Dynamic Conditional Correlation 1631

the dollar/pound that was estimated using the DCC integrated method. Except in the
mid-1990s when they are highly correlated with a coefficient of 0.32, the correla-
tion between crude oil and the dollar/pound is generally negative (with a coefficient
of
0.08 at the beginning of 1986 and a coefficient of
0.16 at the beginning of
2003, respectively).
Key issues relevant in financial economic applications include, for example,
whether and how volatility and correlation move together. It is widely recognized
among both finance academics and practitioners that they vary importantly over
time (Andersen et al. 2001a, b; Engle 2002; Kasch and Caporin 2012). Such
questions are difficult to answer using conventional volatility models, and so we
wish to use the dynamic conditional correlation model to explain the
phenomenon. From Fig. 58.3, the bivariate scatter plots of volatilities and
correlations, it is hard to tell if there is a strong positive association between
each of the two markets sampled. The correlation between two financial assets
will affect the diversification of the portfolio. If two financial assets are highly
negatively correlated, the effect of diversification will be significant, meaning that
the portfolio can balance the returns. This is the so-called idea of not putting all
one’s eggs in the same basket.
According to the empirical data, the dynamic conditional correlation for the
overall gold and dollar/pound (at
0.1986) or the overall oil and dollar/pound
(at
0.0116) can moderately diversify investment risks and can thereby increase
the rate of return. Investors can add commodities and their related derivatives to
portfolios, in an effort to diversify away from traditional investments and assets.
These results are in line with Capie et al. (2005) who found a negative relationship
between the gold price and the sterling/dollar and yen/dollar foreign exchange rates
and Nikos (2006) who found that the correlation between the dollar and gold is
significantly negative. The conclusion we can draw from the results is that gold is
by far the most relevant commodity in hedging against the US dollar. Capie
et al. (2005) observed that gold has served as a hedge against fluctuations in the
foreign exchange value of the dollar. Secondly, gold has become particularly
relevant during times of US dollar weakness. In addition to that, the dynamic
conditional correlation for the overall crude oil and gold markets is 0.0889, and
a similar correlation was documented for the Brent crude oil and gold markets by
Steinitz (2006).
To characterize the distributions of dynamic conditional correlation among the
sampled markets, the summary statistics of the probability distributions for DCC
are shown in Table 58.5, and the associated distributions of DCC for the sampled
markets are shown in Fig. 58.4. We can find that the average DCC between the
crude oil and gold markets is 0.0889 with a standard deviation of 0.0916. The
distribution of the daily DCC between the crude oil and gold markets reflects
a slightly right-skewed (at 1.2021) and leptokurtic distribution (at 4.6799), imply-
ing that a positive DCC occurs more often than a negative DCC between the
crude oil and gold markets. Furthermore, the average DCC between the gold and
 1632

a DCC between oil and gold markets b DCC between oil and gold markets
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3

DCC
DCC
0.2 0.2
0.1 0.1
0.0 0.0
−0.1 −0.1
−0.2 −0.2
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.0000 0.0001 0.0002 0.0003 0.0004 0.0005 0.0006 0.0007
oil volatility gold volatility
DCC between gold and FX markets DCC between FX and gold markets
c 0.6 d 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2

DCC
DCC
0.1 0.1
−0.0 −0.0
−0.1 −0.1
−0.2 −0.2
−0.3 −0.3
0.0000 0.0001 0.0002 0.0003 0.0004 0.0005 0.0006 0.0007 0.000000 0.000050 0.000100 0.000150
gold volatility FX volatility
DCC between oil and FX markets DCC between oil and FX markets
e 0.36 f 0.36
0.24 0.24
0.12 0.12
0.00 0.00

DCC
DCC
−0.12 −0.12
−0.24 −0.24
−0.36 −0.36
0.000000 0.000050 0.000100 0.000150 0.00 0.01 0.02 0.03 0.04 0.05 0.06
FX volatility oil volatility

Fig. 58.3 Bivariate scatter plots of volatilities and correlations. The scatter plots of the daily DCC between (a) crude oil and gold and (b) crude oil and dollar/
pound against the crude oil volatility; the daily DCC between (c) gold and crude oil and (d) gold and dollar/pound against the gold volatility; and the daily
DCC between (e) dollar/pound and crude oil and (f) dollar/pound and gold against the dollar/pound volatility have been shown
T.-L. Shih et al.
58 Realized Distributions of Dynamic Conditional Correlation 1633

Table 58.5 Summary statistics of probability distributions of DCC for each pair of oil, gold,
and FX
Panel A The DCC distributions for the sampled markets from January 1, 1986 to December 31, 2007
Mean Standard dev. Max Min Skewness Kurtosis
DCC between oil and gold 0.0889 0.0916 0.6605
0.1744 1.2021 4.6799
DCC between gold and FX
0.1986 0.1197 0.2019
0.5596
0.1529 0.3304
DCC between FX and oil
0.0116 0.0799 0.3349
0.3076
0.0201 0.8678
Panel B The DCC distributions for the sampled markets from January 1, 1986 through July 31, 1990
Mean Standard dev. Max Min Skewness Kurtosis
DCC between oil and gold 0.0910 0.0669 0.3003
0.0783 0.3752 0.1478
DCC between gold and FX
0.2567 0.1052 0.0246
0.5506
0.2751 0.0666
DCC between FX and oil
0.0030 0.0734 0.2308
0.3076
0.5582 1.9547
Panel C The DCC distributions for the sampled markets from August 1, 1990 through
August 31, 2001
Mean Standard dev. Max Min Skewness Kurtosis
DCC between oil and gold 0.0720 0.1040 0.6605
0.1744 1.8367 6.5149
DCC between gold and FX
0.1258 0.0857 0.2019
0.3222 0.7307 0.8036
DCC between FX and oil
0.0004 0.0772 0.3349
0.2469 0.1648 1.3984
Panel D The DCC distributions for the sampled markets from September 1, 2001 through
December 31, 2007
Mean Standard dev. Max Min Skewness Kurtosis
DCC between oil and gold 0.1168 0.0758 0.3270
0.0772
0.0521
0.4706
DCC between gold and FX
0.2826 0.1003
0.0688
0.5596
0.2534
0.4195
DCC between FX and oil
0.0369 0.0832 0.2413
0.2459 0.1549
0.0225

dollar/pound markets is
0.1986 with a standard deviation of 0.1197. The distri-

bution of the daily DCC between the gold and dollar/British pound markets reflects
a slightly left-skewed (at
0.1529) and platykurtic distribution (at 0.3304), imply-
ing that a negative DCC occurs more often than a positive DCC between gold and
the dollar/pound. Moreover, the average DCC between the dollar/pound and crude
oil is
0.0116 with a standard deviation of 0.0799. The distribution of daily DCCs
between the dollar/pound and crude oil markets reflects a slightly left-skewed
(at
0.0201) and platykurtic distribution (at 0.8678), implying that negative
DCCs occur more often than positive DCCs between the dollar/pound and crude
oil markets.
According to the empirical results, we rank the sequence of the volatility in order
to analyze the volatility effect in the correlation. We show the panel of volatility
and dynamic conditional correlation in Tables 58.6, 58.7, and 58.8. We can clearly
realize from the sample mean, from the daily DCC between the crude oil and gold
markets, that higher volatility can accompany the larger DCC. However, the higher
volatility can also accompany the smaller DCC between the gold and dollar/pound
markets and the crude oil and dollar/pound markets. This is because the correlations
between these markets are negative.
1634 T.-L. Shih et al.

7 7

6 6

Probability Density
Probability Density

5 5

4 4

3 3

2 2

1 1

0 0
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 −0.6 −0.4 −0.2 0 0.2 0.4 0.6
The DCC between oil and gold market The DCC between gold and FX market
7 7
WTI and GOLD
6 6
GOLD and FX
WTI and FX
Probability Density

Probability Density
5 5

4 4

3 3

2 2

1 1

0 0
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8
The DCC between oil and FX market DCC

Fig. 58.4 The distributions of dynamic conditional correlations among gold, oil, and FX from
1986 to 2007

To further quantify this volatility effect in correlation, we classify the volatility

into two categories, low volatility days and high volatility days,1 and according to
the results, we rank the sequence of the volatility. The group of low volatility days
means that the volatility is less than the 10th percentile value and the group of high
volatility days means that the volatility is greater than the 90th percentile value. The
results are shown in Fig. 58.5a–c that reports the DCC distributions for low
volatility days and high volatility days.
It is found that some special characteristics of DCC exist among the oil, gold,
and FX markets. First, distributions for low volatility days are obviously differ-
ent from those for high volatility days. Those for low volatility days approximate
leptokurtic distributions, whereas those for high volatility days approximate

1
Following Andersen et al. (2001b), the authors classify the days into two groups: low volatility
days and high volatility days. The empirical results show that the distribution of correlations shifts
rightward when volatility increases.
58 Realized Distributions of Dynamic Conditional Correlation 1635

Table 58.6 DCC distributions of crude oil and gold markets

Panel A DCC between crude oil and gold against the crude oil market volatility
Crude oil and gold against the crude oil market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 % 0.0779 0.0379 0.6205 1.8582
10–20 % 0.0785 0.0589 0.7672 0.6458
20–30 % 0.0887 0.0680 0.3950
0.1434
30–40 % 0.0924 0.0754 0.6638 0.9140
40–50 % 0.0904 0.0769 0.4299 0.8993
50–60 % 0.0900 0.0825 0.1764 0.5715
60–70 % 0.0839 0.0820 0.3908 0.7808
70–80 % 0.0735 0.0872 0.4688 1.4680
80–90 % 0.0908 0.1273 1.2185 2.3300
90–100 % 0.1212 0.1530 0.8691 1.2595
Panel B DCC between crude oil and gold against the gold market volatility
Crude oil and gold against the gold market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 % 0.0537 0.0516
0.1667
0.6070
10–20 % 0.0817 0.0567 0.0538 1.4446
20–30 % 0.0790 0.0704 0.4538 1.2475
30–40 % 0.0908 0.0734 0.7867 1.8984
40–50 % 0.0842 0.0879 0.7319 2.4974
50–60 % 0.0723 0.0918 0.2724 1.5955
60–70 % 0.0838 0.0899 0.6821 1.9649
70–80 % 0.0868 0.0880 0.9889 3.9028
80–90 % 0.1002 0.0996 0.9919 3.0296
90–100 % 0.1579 0.1386 0.9773 1.9451

platykurtic distributions. Secondly, the average DCCs of the high volatility

days are greater than the average DCCs of the low volatility days for the
DCCs between crude oil and gold, implying that the correlation between
gold and oil increases with volatility. Furthermore, the distribution of DCCs
shifts rightward when volatility increases. Similar results are found for
equity returns as reported by Sonlinik et al. (1996) and in realized exchange
rate returns by Andersen et al. (2001b). Thirdly, the average DCCs for the high
volatility days are smaller than the average DCCs for lower volatility days across
gold and the dollar/pound, and oil and the dollar/pound. This implies that
the correlation between gold (oil) and foreign exchange rates decreases with
volatility, and the distribution of DCCs shifts leftward when volatility increases.
Finally, the standard deviations of the distributions for high volatility days
are obviously greater than the standard deviations of the distributions for
low volatility days.
1636 T.-L. Shih et al.

Table 58.7 DCC distributions of gold and dollar/pound markets

Panel A DCC between gold and FX against the gold market volatility
Gold and FX against the gold market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 %
0.1507 0.0452
0.2642
0.2864
10–20 %
0.1468 0.0811
0.1898 0.1166
20–30 %
0.2048 0.1069 0.0397
0.2823
30–40 %
0.2209 0.1039 0.3412 0.0578
40–50 %
0.2105 0.1184 0.2681
0.3950
50–60 %
0.2070 0.1131 0.0924
0.2283
60–70 %
0.1984 0.1239
0.0848
0.2003
70–80 %
0.2188 0.1411
0.2455
0.1963
80–90 %
0.2231 0.1425
0.2051 0.0326
90–100 %
0.2064 0.1498 0.3059 0.0513
Panel B DCC between gold and FX against the FX market volatility
Gold and FX against the FX market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 %
0.1419 0.0621 0.1926 0.8518
10–20 %
0.1618 0.0861 0.1325
0.1298
20–30 %
0.1839 0.0943 0.0866
0.0156
30–40 %
0.1968 0.1036 0.5927 0.7147
40–50 %
0.2012 0.1163 0.4529 1.0610
50–60 %
0.2314 0.1150 0.3684 0.9286
60–70 %
0.2189 0.1239 0.2181 0.1255
70–80 %
0.2201 0.1217 0.1057
0.1704
80–90 %
0.2170 0.1510
0.2916
0.4897
90–100 %
0.2170 0.1558
0.1548
0.6496

58.4.3 Volatility Threshold Dynamic Conditional Correlation

To further check if different subperiods have various patterns, we utilize

the volatility threshold model addressed by Kasch and Caporin (2012)2 to
examine whether increasing volatility (exceeding a specified threshold)
is associated with an increasing correlation. The volatility threshold DCC model
is specified as Eq. 58.9:


qij, t ¼ 1
a2
b2 qij
gi gj vij þ a2 ei, t
1 ej, t
1 þ b2 qij, t
1 þ gi gj vij, t (58.9)

2
Kasch and Caporin (2012) extended the multivariate GARCH dynamic conditional correlation of
Engle to analyze the relationship between the volatilities and correlations. The empirical results
indicated that high volatility levels significantly affect the correlations of the developed markets,
while high volatility does not seem to have a direct impact on the correlations of the transition blue
chip indices with the rest of the markets. It is easy to see that the volatility and correlation move
together.
58 Realized Distributions of Dynamic Conditional Correlation 1637

Table 58.8 DCC distributions of crude oil and dollar/pound markets

Panel A DCC between crude oil and FX against the FX market volatility
Crude oil and FX against the FX market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 %
0.0147 0.0530
0.2458 0.3703
10–20 %
0.0142 0.0651
0.4224
0.1734
20–30 %
0.0140 0.0695 0.0757 0.2648
30–40 %
0.0028 0.0804 0.0465
0.2710
40–50 %
0.0108 0.0832 0.0264 0.1784
50–60 %
0.0206 0.0720 0.1618 0.0328
60–70 %
0.0081 0.0789 0.5080 1.0521
70–80 %
0.0081 0.0773 0.4689 1.6993
80–90 %
0.0161 0.1042 0.1100 0.5576
90–100 %
0.0068 0.1019
0.7085 0.2311
Panel B DCC between crude oil and FX against the crude oil market volatility
Crude oil and gold against the gold market volatility
Volatility Mean Standard dev. Skewness Kurtosis
0–10 %
0.0001 0.0453 0.5049 0.7991
10–20 % 0.0045 0.0609
0.1145 0.1341
20–30 %
0.0016 0.0675
0.2397
0.2455
30–40 %
0.0087 0.0715
0.2668 0.1219
40–50 %
0.0164 0.0736
0.2498 0.1421
50–60 %
0.0228 0.0743
0.2262 0.0489
60–70 %
0.0174 0.0732 0.0187
0.2162
70–80 %
0.0167 0.0811
0.0200 0.0117
80–90 %
0.0028 0.0934 0.0397 0.0698
90–100 %
0.0310 0.1234 0.4996 0.1905

where vt is a dummy variables matrix defined as


1 if hi, t > f hi ðkÞ or hj, t > f hj ðkÞ
vij, t ¼ (58.10)
0 otherwise

where fhi(k) is the kth fractional of the volatility series hi,t.

When thresholds are found, the whole period will be divided into various sub-
periods based on these thresholds. This separation helps detect any changes in
investor behavior after crucial events. It is then known whether a time horizon is
a key factor influencing the patterns of return and volatility.
Table 58.9 presents the estimation results of the volatility threshold DCC models.
The estimation was based on various volatility threshold levels at 50 %, 75 %, 90 %,
and 95 %. The results in Table 58.9 show that the correlation between the oil and gold
prices is significantly affected by the volatility of oil at the 50 %, 75 %, and 90 %
1638 T.-L. Shih et al.

a 12 12
High Volatility Days High Volatility Days
Low Volatility Days Low Volatility Days
10 10

Probability Density
Probability Density

8 8

6 6

4 4

2 2

0 0
−0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
The DCC between oil and gold markets The DCC between oil and gold markets

b 8 10
High Volatility Days High Volatility Days
7 9
Low Volatility Days Low Volatility Days
8
Probability Density
6
Probability Density

7
5 6
4 5

3 4
3
2
2
1 1
0 0
−0.5 −0.4 −0.3 −0.2 −0.1 0 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0
The DCC between gold and FX markets The DCC between gold and FX markets

c 14 10
Low Volatility Days Low Volatility Days
9
12 High Volatility Days High Volatility Days
8
Probability Density
Probability Density

10 7
6
8
5
6 4
4 3
2
2
1
0 0
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 −0.2 −0.1 0 0.1 0.2 0.3
The DCC between oil and FX markets The DCC between oil and FX markets

Fig. 58.5 (a) The distributions of DCC between crude oil and gold on low and high volatility
days. (b) The distributions of DCC between gold and dollar/pound on low and high volatility days.
(c) The distributions of DCC between oil and dollar/pound on low and high volatility days

(with the exception of 95 %) thresholds. Interestingly, these estimated thresholds

were quite consistent with the real events of the First Gulf War in 1990 and the
911 attack in 2001. We then separate the period into three subperiods to further
examine whether the investors’ behaviors change after the events.
58 Realized Distributions of Dynamic Conditional Correlation 1639

Table 58.9 The volatility threshold dynamic conditional correlationa

Panel A Crude oil and gold
50 % 75 % 90 % 95 %
a2 0.0171 0.0146 0.0187 0.0229
(1.956) (1.870) (2.072) (2.424)
b2 0.9546 0.9548 0.9362 0.9233
(36.377) (41.162) (27.21) (23.705)
gWTIgGOLD 0.0075 0.0116 0.0206 0.0253
(2.045) (2.300) (2.376) (1.636)
Panel B Gold and dollar/pound
50 % 75 % 90 % 95 %
a2 0.0161 0.0150 0.0150 0.0149
(3.016) (3.052) (2.938) (2.894)
b2 0.9784 0.9808 0.9804 0.9803
(108.063) (124.94) (120.37) (119.94)
gGOLDgFX
0.00032 0.0012 0.0023 0.0046
(
0.190) (0.805) (0.932) (1.079)
Panel C Crude oil and dollar/pound
50 % 75 % 90 % 95 %
a2 0.0131 0.0121 0.0109 0.0126
(2.294) (1.825) (1.633) (2.013)
b2 0.9519 0.9580 0.9639 0.9518
(33.260) (28.286) (33.633) (28.095)
gWTIgFX 0.00006 0.0015 0.0034 0.0099
(0.016) (0.418) (0.785) (0.991)
a
This table presents the quasi-maximum likelihood estimates of volatility threshold of dynamic
conditional correlation. The t-statistics are given in parentheses

58.4.4 Does Investors’ Behavior Change over Subperiods?

The three subperiods based on our estimated thresholds are before the first Gulf War
(January 1, 1986 to July 31, 1990), after the first Gulf War up to the 911 attack
(August 1, 1990 to August 31, 2001), and after the 911 attack (September 1, 2001 to
December 1, 2007). We then examine the dynamic co-movement in each pair of
markets in various subperiods (Rigobon and Sack 2005; Guidi et al. 20073).
Our sampled period covers economic hardship and soaring energy prices.
Soaring energy prices make gold an attractive hedging asset against inflation in
that a positive correlation with oil is expected over time, especially after the
911 event. The evidence in Table 58.10 shows that oil and gold are highly

3
Guidi et al. (2007) examined the impact of relevant US decisions on oil spot price movements
from January 1986 to December 2005. They identified the following conflict periods: the Iran-Iraq
conflict, January 1985 until July 1988; Iraq’s invasion of Kuwait, August 1990 until February
1991; and the US-led forces’ invasion of Iraq, March 2003 until December 2005.
1640 T.-L. Shih et al.

Table 58.10 Simple correlation matrix of oil, gold, and dollar/pound markets among the three
subperiods
Panel A The correlation coefficients between crude oil, gold, and dollar/pound markets from January 1,
1986 through July 31, 1990
Oil Gold FX
Oil 1
Gold 0.1648 1
FX
0.1517
0.5228 1
Panel B The correlation coefficients between crude oil, gold, and dollar/pound markets from August 1,
1990 through August 31, 2001
Oil Gold FX
Oil 1
Gold
0.2465 1
FX 0.0628
0.2096 1
Panel C The correlation coefficients between crude oil, gold, and dollar/pound markets from September 1,
2001 through December 31, 2007
Oil Gold FX
Oil 1
Gold 0.9264 1
FX
0.8124
0.8142 1

correlated with a high coefficient of 0.9264 between September 1, 2001 and

December 31, 2007 (after the 911 event). In the meantime, the correlation between
gold and the dollar/pound is
0.8142, and between oil and the dollar/pound
is
0.8124, showing the fears of a depreciation in the US dollar push commodity
prices up significantly.
The results of Panels B, C, and D in Table 58.5 along with Figs. 58.6 and 58.7
show that the DCCs between oil and gold are all positive in the three subperiods
(with coefficients of 0.0910, 0.0720, and 0.1168, respectively). Obviously, higher
oil prices spark inflationary concerns and make gold a value reserve for wealth.
By contrast, the correlation between oil and the dollar/pound has been negative,
and a DCC of
0.0369 is shown after the 911 attack. In this period, the oil price
increased from a low of US$19 per barrel in late January 2002 to US$96 per barrel
in late December 2007. Meanwhile, the dollar conversely tumbled 29 % from
US$0.7 to a US$0.5 per pound.
Historical data also show that the prices of oil and gold are rising over time. The
increase in the prices of gold was a reflection of the falling US dollar. The evidence
confirms that the DCCs between gold and the dollar/pound in the three subperiods
are all negative (with average DCC of
0.2567,
0.1258, and
0.2826 in the first,
second, and third periods, respectively). Generally speaking, during the subperiods
of market crises, increasingly high correlations in commodity prices were observed,
with oil and gold moving in the same direction and the dollar/pound moving in
opposite directions.
58 Realized Distributions of Dynamic Conditional Correlation 1641

Fig. 58.6 The distributions a 30

of dynamic conditional WTI and GOLD
GOLD and FX
correlations among each pair WTI and FX
25
of the three markets over three
subperiods: (a) January 1,

Probability Density
1986 to July 31, 1990, (b) 20
August 1, 1990 to August 31,
2001, and (c) September 1, 15
2001 to December 31, 2007
10

0
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3
DCC during 19860101-19900731
b 12
WTI and GOLD
GOLD and FX
WTI and FX
10
Probability Density

0
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1
DCC during 19900801-20010831

c 8
WTI and GOLD
GOLD and FX
7 WTI and FX

6
Probability Density

0
−0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
DCC during 20010901-20071231
1642 T.-L. Shih et al.

Fig. 58.7 The distributions a 30

of dynamic conditional 19860101–19900731
19900801–20010831
correlations for (a) crude oil 20010901–20071231
25
and gold, (b) gold and
dollar/pound, (c) crude oil and

Probability Density
dollar/pound over three 20
subperiods
15

10

0
−0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
DCC between crude oil and gold markets

b 25
19860101–19900731
19900801–20010831
20010901–20071231
20
Probability Density

15

10

0
−0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3
DCC between gold and FX markets
c 30
19860101–19900731
19900801–20010831
20010901–20071231
25
Probability Density

20

15

10

0
−0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
DCC between crude oil and FX markets
58 Realized Distributions of Dynamic Conditional Correlation 1643

58.5 Conclusions and Implications

Using the dynamic conditional correlation model, we have estimated the cross
correlation and volatility among crude oil, gold, and dollar/pound currencies from
1986 to 2007. After exploring the time-varying correlations and realized distribu-
tions, several regularities have been found that help illustrate the characteristics of
the crude oil, gold, and currency markets.
First, the correlation coefficients between each pair of the three assets are found
to be time varying instead of constant. As such, besides considering the mean and
standard deviation of the underlying assets, investors need to follow the
co-movement in the relevant assets in order to make better portfolio hedging and
risk management decisions across these assets. The results of the dynamic correla-
tion coefficients between the gold and dollar/pound show that gold is by far the
most relevant commodity in terms of serving as a hedge against the US dollar.
These results are in line with the reports suggested by Nikos (2006) and Capie
et al. (2005). Our findings are helpful in terms of arriving at a more optimal
allocation of assets based on their multivariate returns and associated risks.
Besides, the distributions of low volatility days are found to approximate
leptokurtic distributions in the gold, oil, and dollar/pound markets, whereas the
high volatility days approximate platykurtic distributions. Furthermore, the DCC
between oil and gold is increasing with volatility, indicating that the distribution of
DCC shifts rightward when volatility increases. By contrast, the DCCs between
gold and the dollar/pound and crude oil and the dollar/pound are decreasing with
the volatility. Our findings in terms of oil and gold are consistent with the reports of
Sonlinik et al. (1996) and Andersen et al. (2001b) who use different approaches.
Moreover, by estimating the volatility threshold dynamic conditional correlation
model addressed by Kasch and Caporin (2012), we find that high volatility values
(exceeding some specified thresholds) are associated with an increase in correlation
values in various subperiods. Remarkably, investors’ behaviors are seen to have
changed in different subperiods. During periods of market turmoil, such as the First
Gulf War in 1990 and the 911 terror attack in 2001, an increase in correlation
between the prices of oil and gold, as well as a decrease in correlation between the
oil (gold) and dollar/pound currencies, is observed. These behaviors make gold
an attractive asset against major currencies for value-preserving purposes.
For market participants from long-term hedging perspective, our results provide
useful information on asset allocation across commodity and currency markets
during market turmoil.

Acknowledgments This chapter was presented at the Seventh International Business Research
Conference in Sydney, Australia, and the Taiwan Finance Association Annual Meeting in
Taichung, Taiwan. The authors are grateful for the helpful and suggestive comments from Ken
Johnson, C. L. Chiu, Ming-Chi Lee, and other participants. The authors also appreciate the
financial grants for attending the conference from the National Science Council.
1644 T.-L. Shih et al.

References
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Ebens, H. (2001a). The distributions of realized
stock return volatility. Journal of Financial Economics, 61, 43–76.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2001b). The distribution of realized
exchange rate volatility.