Lecture Notes in Economics

and Mathematical Systems 506

Founding Editors:
M. Beckmann
H. P. Kiinzi

Managing Editors:
Prof. Dr. G. Fandel
Fachbereich Wirtschaftswissenschaften
Femuniversitat Hagen
Feithstr. 140/AVZ II, 58084 Hagen, Germany
Prof. Dr. W. Trockel
Institut fUr Mathematische Wirtschaftsforschung (IMW)
Universitat Bielefeld
Universitatsstr. 25, 33615 Bielefeld, Germany

Co-Editors:
C. D. Aliprantis, Dan Kovenock

Editorial Board:
P. Bardsley, A. Basile, M.R. Baye, T. Cason, R. Deneckere, A. Drexl,
G. Feichtinger, M. Florenzano, W Giith, K. Inderfurth, M. Kaneko, P. Korhonen,
W. Kiirsten, M. Li Calzi, P. K. Monteiro, Ch. Noussair, G. Philips, U. Schittko,
P. Schonfeld, R. Selten, G. Sorger, R. Steuer, F. Vega-Redondo, A. P. Villamil,
M. Wooders
Springer-Verlag Berlin Heidelberg GmbH
B. Philipp Kellerhals

Financial Pricing Models

in Continuous Time
and Kalman Filtering

Springer
Author
Dr. B. Philipp Kellerhals
Deutscher Investment-Trust
Gesellschaft fUr Wertpapieranlagen mbH
Mainzer LandstraBe 11-13
60329 Frankfurt am Main, Germany

Cataloging-in-Publication data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme

Kellerhals, Philipp B.:

Financial pricing models in continuous time and Kalman filtering 1 B.
Philipp Kellerhals. - Berlin; Heidelberg; New York; Barcelona; Hong Kong
; London; Milan; Paris; Singapore; Tokyo: Springer, 2001
(Lecture notes in economics and mathematical systems; 506)
ISBN 3-540-42364-8

ISSN 0075-8450
ISBN 978-3-540-42364-5 ISBN 978-3-662-21901-0 (eBook)
DOI 10.1007/978-3-662-21901-0

This work is subject to copyright. All rights are reserved, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, re-use
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other
way, and storage in data banks. Duplication of this publication or parts thereof is
permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from
Springer-Verlag. Violations are liable for prosecution under the German Copyright
Law.

http://www.springer.de
© Springer-Verlag Berlin Heidelberg 2001
Originally published by Springer-Verlag Berlin Heidelberg New York in 2001.

The use of general descriptive names, registered names, trademarks, 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.
Typesetting: Camera ready by author/editors
Cover design: design & production, Heidelberg
Printed on acid-free paper SPIN: 10844260 55/3142/du 5432 I 0
To my parents
Foreword

Straight after its invention in the early sixties, the Kalman filter approach
became part of the astronautical guidance system of the Apollo project and
therefore received immediate acceptance in the field of electrical engineer-
ing. This sounds similar to the well known success story of the Black-Scholes
model in finance, which has been implemented by the Chicago Board of Op-
tions Exchange (CBOE) within a few month after its publication in 1973.
Recently, the Kalman filter approach has been discovered as a comfortable
estimation tool in continuous time finance, bringing together seemingly un-
related methods from different fields.
Dr. B. Philipp Kellerhals contributes to this topic in several respects.
Specialized versions of the Kalman filter are developed and implemented
for three different continuous time pricing models: A pricing model for
closed-end funds, taking advantage from the fact, that the net asset value is
observable, a term structure model, where the market price of risk itself is a
stochastic variable, and a model for electricity forwards, where the volatility
of the price process is stochastic. Beside the fact that these three models can
be treated independently, the book as a whole gives the interested reader a
comprehensive account of the requirements and capabilities of the Kalman
filter applied to finance models. While the first model uses a linear version
of the filter, the second model using LIBOR and swap market data requires
an extended Kalman filter. Finally, the third model leads to a non-linear
transition equation of the filter algorithm.
Having some share in the design of the first two models, I am rather
impressed by the potential and results of the filtering approach presented
by Dr. Kellerhals and I hope, that this monograph will play its part in
making Kalman filtering even more popular in finance.

Tiibingen, May 2001 Rainer SchObel

Acknowledgements

The book on hand is based on my Ph.D. thesis submitted to and accepted

by the College of Economics and Business Administration at the Eberhard-
Karls-University TUbingen, Germany. There, I had the opportunity to work
as a researcher and lecturer in the Department of Finance.
First and foremost, I would like to thank my academic supervisor and
teacher Prof. Dr.-Ing. Rainer Schobel who introduced me to the challenging
field of continuous time modeling in financial economics. He has created an
ideal environment for productive research in which I have received valuable
advice and support. Moreover, I am grateful to the further members of
my thesis committee Prof. Dr. Werner Neus, Prof. Dr. Gerd Ronning, and
Prof. Dr. Manfred Stadler.
The first remarkable contact with financial theory I gratefully received
during my academic year I could spend at Arizona State University, Tempe.
During my final graduate studies at the University of Mannheim, Ger-
many, Prof. Dr. Wolfgang Buhler and Dr. Marliese Uhrig-Homburg started
to create my affinity to interest rate modeling which I could deepen in a
Ph.D. seminar with Prof. Nick Webber on current theoretical and empir-
ical fixed-income issues. I also had the opportunity to participate in a
Ph.D. seminar with Prof. Yacine Ait-Sahaliah on statistical inference prob-
lems with discretely observed diffusion models. Finally, I am thankful to
Prof. Dr. Herbert Heyer for his introduction to the theory of stochastic par-
tial differential equations.
Last but not least, lowe much more than they imagine to my colleagues
at the faculty Dr. Stephan Heilig, Dr. Roman Liesenfeld, Dr. Hartmut Nagel,
Dr. Ariane Reiss, and Dr. Jianwei Zhu for the inspiring, fruitful, and pleasant
working atmosphere they provided.

TUbingen, May 2001 B. Philipp Kellerhals

Contents

1 Overview of the Study 1

I Modeling and Estimation Principles 5

2 Stochastic Environment 7

3 State Space Notation 11

4 Filtering Algorithms 15
4.1 Linear Filtering 16
4.2 MMSE . . . . . . 17
4.3 MMSLE . . . . . 21
4.4 Filter Recursions 22
4.5 Extended Kalman Filtering 24

5 Parameter Estimation 29

II Pricing Equities 35
6 Introduction 37
6.1 Opening Remarks . . . . . . . . 37
6.2 The Case of Closed-End Funds 38

7 Valuation Model 43
7.1 Characteristics of Closed-End Funds 43
7.2 Economic Foundation . . . . . . . 47
7.3 Pricing Closed-End Fund Shares .. . 50
XII Contents

8 First Empirical Results 55

8.1 Sample Data. . . . . 55
8.2 Implemented Model . 60
8.3 State Space Form . . 61
8.4 Closed-End Fund Analysis 64

9 Implications for Investment Strategies 71

9.1 Testing the Forecasting Power . . . . 71
9.1.1 Setup of Forecasting Study .. 71
9.1.2 Evidence on Forecasting Quality. 73
9.2 Implementing Trading Rules . . . . . . . 75
9.2.1 Experimental Design . . . . . . . 75
9.2.2 Test Results on Trading Strategies 77

10 Summary and Conclusions 83

III Term Structure Modeling 85

11 Introduction 87
11.1 Overview. . . . . . . . . . . . . 87
11.2 Bond Prices and Interest Rates 88
11.3 Modeling an Incomplete Market 92

12 Term Structure Model 97

12.1 Motivation for a Stochastic Risk Premium 97
12.2 Economic Model . . . . . . . . . . . . . . 100

13 Initial Characteristic Results 105

13.1 Valuing Discount Bonds . . . . . . . . . . . . . . 105
13.2 Term Structures of Interest Rates and Volatilities 112
13.2.1 Spot and Forward Rate Curves 112
13.2.2 Term Structure of Volatilities . . . . . . . 113
13.3 Analysis of Limiting Cases . . . . . . . . . . . . . 116
13.3.1 Reducing to an Ornstein-Uhlenbeck Process 116
13.3.2 Examining the Asymptotic Behavior 118
13.4 Possible Shapes of the Term Structures . 120
13.4.1 Influences of the State Variables. . . 121
Contents XIII

13.4.2 Choosing the Model Parameters . . . . 123

14 Risk Management and Derivatives Pricing 129

14.1 Management of Interest Rate Risk. 129
14.2 Martingale Approach . 131
14.3 Bond Options. . . . . . . . . . 133
14.4 Swap Contracts . . . . . . . . . 141
14.5 Interest Rate Caps and Floors. 143

15 Calibration to Standard Instruments 147

15.1 Estimation Techniques for Term Structure Models. 147
15.2 Discrete Time Distribution of the State Variables 150
15.3 US Treasury Securities . . . . 153
15.3.1 Data Analysis. . . . . . . . . . 153
15.3.2 Parameter Estimation . . . . . 160
15.3.3 Analysis of the State Variables. 165
15.4 Other Liquid Markets. . . . . . . . . . 169
15.4.1 Appropriate Filtering Algorithm. 169
15.4.2 Sample Data and Estimation Results 170

16 Summary and Conclusions 175

IV Pricing Electricity Forwards 177

17 Introduction 179
17.1 Overview . . . . . . . . . . . . 179
17.2 Commodity Futures Markets. 179
17.3 Pricing Commodity Futures 183

18 Electricity Pricing Model 189

18.1 Pricing Electricity Derivatives . . . . . . . . . 189
18.2 Model Assumptions and Risk Neutral Pricing 191
18.3 Valuation of Electricity Forwards . . . . . . . 194

19 Empirical Inference 201

19.1 Estimation Model. 201
19.1.1 Distribution ofthe State Variables . 201
XIV Contents

19.1.2 State Space Formulation and Kalman Filter Setup . 205

19.2 Data Analysis and Estimation Results . . . . . . . . . . . 208

20 Summary and Conclusions 221

List of Symbols and Notation 223

List of Tables 227

List of Figures 229

Bibliography 231
1 Overview of the Study

The research objective of this study can be grasped by the question: Are
Kalman filter algorithms applicable to efficiently estimate financial models
of contingent claim pricing? This formulation asks for two more clarifica-
tions: First, which financial pricing models for which markets are selected
and second, which type of Kalman filters are applied? In answering these
questions, we hope to generate both theoretical and econometric contribu-
tions to the scientific field of financial economics.
The Kalman filter framework was originally developed by Kalman (1960)
and Kalman and Bucy (1961) for applications in aerospace engineering and
has successfully been used in electrical engineering. Only in the last decade,
marked by Harvey (1989), Kalman filters have found ground as economet-
ric tool for economic and financial estimation problems. The technique of
Kalman filtering generally leads to more efficient parameter estimates since
it imposes greater theoretical modeling restrictions on the data than other
commonly used methods. Given a hypothesized financial pricing model, the
estimation procedure of Kalman filtering imposes cross-sectional as well as
time-series restrictions, making it a true maximum likelihood method. We
use different types of Kalman filter algorithms to estimate current pricing
problems in the field of financial economics. Three major financial markets
are to be examined for which we select the equity market, the bond market,
and the electricity market. In each market we derive new contingent claim
valuation models to price selected financial instruments in continuous time.
The decision criterium for choosing a continuous time modeling framework is
the richness of the stochastic theory available for continuous time processes
with Merton's pioneering contributions to financial economics, collected in
Merton (1992). The continuous time framework, reviewed and assessed
by Sundaresan (2000), allows us to obtain analytical pricing formulae that
2 1. Overview of the Study

would be unavailable in a discrete time setting. However, at the time of

implementing the derived theoretical pricing models on market data, that
is necessarily sampled at discrete time intervals, we work with so-called ex-
act discrete time equivalents a la Bergstrom (1984). Thereupon we perform
empirical maximum likelihood inferences implementing linear and extended
specifications of Kalman filter algorithms.
The selected pricing issues within the chosen financial markets are pre-
sented in three distinct essays building parts II, III, and IV of this study:

1. In part II we investigate the pricing of equities. Within this asset class

we specialize in the valuation of closed-end funds which share distinct
financial characteristics of common joint-stock companies and mutual
fund investment companies. In building our stochastic pricing model
we especially draw our attention on the market anomaly that the
market value of closed-end funds determined on organized exchanges
differs dynamically over time from the reported value on their under-
lying investment portfolios by a discount or a premium. In light of this
anomaly we develop a two-factor valuation model of closed-end funds
that takes into account both the price risk of the funds as well as the
risk associated with altering discounts. Based on a state space formu-
lation of the pricing model we estimate the relevant model parameters
using a Kalman filter framework. Having validated our pricing model
on a sample of emerging market closed-end funds traded on the New
York Stock Exchange, we further assess the quality of the developed
model for investor applications. There, we test the forecasting power
of the valuation model to predict closed-end fund market prices and
implement portfolio strategies using trading rules. The results of these
applications indicate that the pricing model generates valuable invest-
ment information.

2. For the bond market we present a new term structure model in part III
which belongs to the affine term structure models. Motivated by the
incomplete market situation of fixed income instruments we propose a
pricing model that shifts from modeling static tastes to beliefs of the
dynamic behavior of tastes. Besides a stochastic short interest rate
the proposed two-factor term structure model allows for a stochastic
behavior of the market price of interest rate risk. Given the theoretical
1. Overview of the Study 3

term structure model we derive a closed-form expression for valuing

discount bonds and examine the implied term structures of interest
rates and volatilities as initial characteristic results. We further show
how to apply the term structure model in a consistent framework to
manage interest rate risk and price interest rate derivative securities.
There we start with the concept of immunization using factor dura-
tion and price bond options, swap contracts, and interest rate caps
and floors as the most relevant derivative contracts in risk manage-
ment. Using US market data on Treasury Securities, LIB OR rates,
and swap rates we implement the theoretical pricing model based on
a corresponding state space form. Linear and extended Kalman fil-
ter routines give us the possibility of estimating the model as well as
extracting the time-series of the underlying state variables.

3. In part IV we present a continuous time pricing model for the val-

uation of electricity forwards. From investigating more mature com-
modity markets than the electricity market, we gather information on
the price behavior of commodities and pricing models for commodity
futures. From properly chosen model assumptions, especially captur-
ing the unique characteristic of non-storability of electricity and the
marked volatility in electricity prices being both high and variable
over time, we build an appropriate valuation model to price electric-
ity forwards. The proposed stochastic volatility pricing model yields a
closed-form solution to electricity forwards using risk neutral pricing
techniques. Next, we empirically adapt the theoretical pricing model
in state space form dealing with state variables that follow a non-
Gaussian distribution; thereupon we implement an extended Kalman
filter algorithm to estimate the model by means of maximum likeli-
hood. Finally, we report empirical results of the pricing model based
on electricity data from the largely deregulated Californian power mar-
ket.

Each of these essays can be read independently and contains an intro-

ductory chapter which motivates the presented research and clears its con-
text in the respective literature. Furthermore, each essay concludes with a
summary chapter where we finally present our findings. The modeling and
estimation principles common to all essays are described in part I.
 Part I

Modeling and Estimation

Principles
2 Stochastic Environment

Typical financial pricing models in continuous time are built upon hypoth-
esized stochastic processes to assess the models' dynamics over time. A
common and convenient method to describe the evolution of chosen state
variables over time is to specify appropriate stochastic differential equations.
By these equations we try to capture possibly all behavior of the state vari-
able during time evolution or equivalently set up a hypothesis on how the
state variables could evolve over time with the final objective of finding the
real data generating process underlying the analyzed sample of data. To
set up such stochastic differential equations we first take a look at stochas-
tic processes in general defined on a probability space which represents the
uncertainty of the intertemporal stochastic economy.
Definition 2.0.1 (Probability Space) The underlying set-up for stochas-
tic processes consists of a complete probability space (n, F, JPl), equipped with
a filtmtion, i.e. a non-decreasing family IF = {Ft}tET of sub-a-fields of F :
Fs ~ F t ~ F for 0 ~ s < t < 00 where T = [0,00). Here, F t represents
the information available at time t, and the filtmtion {Ft }tET represents
the information flow evolving over time and accruing to all agents in the
economy.
Definition 2.0.2 (Stochastic Process) A scalar (n-vector) stochastic
process, denoted by {XthET' is a family of mndom variables (n-vectors)
indexed by the pammeter set T, where the pammeter t will refer to time
in our applications. The process is defined on a filtered probability space
(n, F, JPl, IF) with values in lR,n. We say the process is adapted if X t E F t
for each time t, i.e. X t is Ft-measumble; thus, X t is known when :ft
is known. Further, if a filtmtion is genemted by a stochastic process,
i.e. F t = a (Xs; 0 ~ s ~ t), we call F t the natuml filtmtion of the process
{XthET' Thus, a process is always adapted to its natuml filtmtion.
8 2. Stochastic Environment

Definition 2.0.3 (Classification of Stochastic Processes) If the ran-

dom variables (vectors) X t are discrete, we say that the stochastic process
has a discrete state space. If they are continuous, the process is said to have
a continuous state space. The parameter set may also be discrete (for exam-
ple, T = {1, 2, , ... ,n}, or T = {1, 2, 3, ... }) or continuous (for example,
T = [0, 1], or T = [0,(0)). If the parameter set T is discrete, the stochastic
process is a discrete parameter process; if it is continuous, we say that the
stochastic process is a continuous parameter process.

An overview of this classification is shown in table (2.1) following Jazwin-

ski (1970).

Table 2.1: Classification of Stochastic Processes

Classifying Parameter Set T

Characteristics Discrete Continuous
Discrete Parameter Continuous Parameter
State Discrete
Chain Chain
Space X Continuous Random Sequence Stochastic Process

Remark 2.0.4 Note that the process is measurable with respect to two dis-
tinct spaces: For every t E T the process X t is a random variable with
w ~ Xdw);w E n; if we fix the value for wEn we obtain the mapping
t ~ X t (w); t E T, a so-called path of X t . Thereby, we need to look at a
stochastic process as a function of two distinct variables (t, w) ~ X (t, w)
from T x n into the space JR 1 •

Among the most popular building blocks of stochastic models in financial

economics is the standard Brownian motion. 1 This elementary stochastic
process satisfies some distinct properties which are defined as follows.

lThe earliest modeling is by Bachelier (1900), who used this type of stochastic pro-
cesses to describe stock price movements. To cover jumps in the dynamics of a security,
one can add a Poisson process component; see, for example, Merton (1976).
2. Stochastic Environment 9

Definition 2.0.5 (Standard Brownian Motion) A continuous param-

eter process {Xt h~o is a Brownian motion under the probability measure ]p>
if
(i) every increment Xt+T - X t is normally distributed with N(/-LT, (J2T)
under ]P>, and
(ii) for every pair of disjoint time intervals [tb t 2], [t3, t4J (with tl <
t2 ~ t3 < t 4 ) the increments X 2 - Xl and X 4 - X3 are independent random
variables, and
(iii) ]P> (Xo = 0) = 1 and X t is continuous at t = O.
For {Xdt~o to be a standard Brownian motion, we need fixed parameters
of /-L = 0 and (J = 1.

This means that the increment Xt+T - X t is independent of the past, or

alternatively, if we know X t = X o, then further knowledge of the values of
Xs for s < t has no effect on our knowledge of the probability law governing
X t+T - X t . Written formally, this says with to < tl < ... < tn < t:

which states that the standard Brownian motion is a Markovian process.

Also, the standard Brownian motion is a Gaussian process, since it has a
normal distribution with specific first two moments. 2

2See, for example, Karlin and Taylor (1975, p. 376).

3 State Space Notation

In this chapter we examine the state space formulation of a financial model

for a general case. The state space form is a very popular and useful way
to write a dynamic model for a further analysis with the Kalman filter. 3
It is based on two important sets of system equations: The measurement
equation which relates the state variables to the variables which can only
be observed with measurement noise; the second set of equations, called
the transition or process equation, describes the dynamic evolution of the
unobservable variables. The vector of unknown parameters, on which the
system matrices and error term specifications of the state space form depend
on, will be denoted by 'l/J. 4

Definition 3.0.6 (Measurement Equation) For the functional relation-

ship of the measurable observations Yt with the possibly unobservable state
vector et,we define the linear form

Yt = ad'l/J) + Bd'l/J) et + ed'l/J) , (3.1)

gxl gxl gxk kxl gxl

with the parameters 'l/J, an additive component at ('l/J), a multiplicative ma-

trix B t ('l/J), and a noise term et ('l/J). Further, we assume a normal distri-
bution with

E [etl = 0, and
for s = t
E [ete~l = { H t~'l/J)
otherwise

for the error term et ('l/J).

3See, for example, Harvey (1989), Aoki (1990), and Hamilton (1994a).
41.e. we treat both the original parameters and the variances of the measurement errors
as part of the vector 1/J, as we will see further down in chapter 5 on parameter estimation.
12 3. State Space Notation

Definition 3.0.7 (Transition Equation) For the system, that describes

the evolution of the state variables et over time, we assume the linear equa-
tion

et = cd"p) + <Pt ("p )et-l + l1t ("p) (3.2)

kxl kxl kxk kxl kxl

with the transition matrix <Pt ("p), an additive component Ct ("p) and a Gaus-
sian noise term 'fit ("p) with

E [l1tl = 0, and
for s = t
E [l1tl1~l = { Qt ~"p)
otherwise.

The model order is given by the dimension of eu

i. e. it is of k-th order.
Further, we assume independence between the error terms Et ("p) and 'fit ("p),
and a normally distributed initial state vector eo,
with E [eol = eOl o and
Cov [eol = }:olo, that is assumed independent of the error terms in the sense
that E [eoe~l = 0 and E [eol1~l = 0 for all t. This system is covariance-
stationary provided that all eigenvalues of <P ("p) lie inside the unit circle. 5

These two equations constitute a linear first-order Gauss-Markov state

space representation for the dynamic behavior of the observable variables
Yt. This framework admits a particularly tractable solution for the further
analysis of filtering. 6 However, it can be further generalized to allow for
time-varying system matrices, non-normal disturbances, correlated distur-
bances and non-linear dynamics, as will be discussed in particular later. 7
Now, for the derivation of the Kalman filter, we just focus on the general
system characterized by equations (3.1) and (3.2).
Before we come to specific applications, we need to address the problem
of identification which is due to the generality of the state space form defined
in equations (3.2) and (3.1). Identification is of considerable conceptual
interest, since it is a necessary condition for the properties of the maximum
5See, for example, Hamilton (1994b, ch. 10); in the Gaussian state-space model we
have strict stationarity (see, for example, Hamilton (1994b, p. 45 f.)), since the normal
density is completely specified by its first two moments.
6See, for example, Jazwinski (1970, ch. 5).
7S ee, for example, Jazwinski (1970), Tanizaki (1996) and Gourieroux and Monfort
(1997).
3. State Space Notation 13

likelihood estimates, described in chapter 5, as well as of practical interest

when we deal with empirical estimation.
Generally, techniques to infer a model from measured data typically con-
tain two steps. First a family of candidate models is decided upon which
we call the modeling step. In financial applications this step heavily draws
on economic theory and mathematics, especially the field of stochastic the-
ory. In a second step, we look for the particular member of this family
that optimally describes the information content revealed by the data. In
our applications this step is in fact a parameter estimation problem in that
we maximize a likelihood function based on the prediction error decom-
position. 8 Given this general description, identification is a link between
the mathematical model world and the real world of data. As such we
need to ensure identification on the level of model building before we pro-
ceed with the statistical inference. A model is said to be not identifiable,
if there are observationally equivalent structures of model parameters 1/J
that imply the same distribution for the observable random outcomes Yt,
i.e. there exist different parameter vectors that lead to the same likelihood
L (YT, YT-l, . .. ,Yl; 1/J). To be more precise we define the concept of iden-
tification following Rothenberg (1971).
Definition 3.0.8 (Concept of Identification) With denoting F a para-
metrized cumulative distribution junction, 1/J the parameter vector contained
in the admissible parameter space W, we call the set {F(YT,YT-l, ... ,Yll
1/J)} tjlEW a model and an element F (YT, YT-l, ... , Yll1/J) of this set a struc-
ture. A model is called globally identifiable at particular parameter values
1/Jo if
1/J =1= 1/Jo =* F (YT, YT-l, ... , Yll1/J) =1= F (YT,yT-l, ... , Yll1/JO) ,
i. e. there exist no other parameter values besides the vector 1/Jo that give rise
to the same structure. A model is said to be locally identified at 1/Jo if there
exists a 8 > 0 such that for any value of 1/J satisfying (1/J -1/J O)' (1/J - 1/JO) <
o there exist realizations YT,yT-l, ... , Yl for which we obtain two different
structures.
Dealing with an unidentifiable parameter vector 1/J is a severe problem,
since alternative parameter structures usually lead to different causal inter-
pretations in the context of the investigated model. Thus the uniqueness
8See chapter 5.
14 3. State Space Notation

of the likelihood must be guaranteed for all possible parameter vectors, be-
cause the true parameter structure is not known. For an unidentified model
example in the case of maximum likelihood estimation using the Kalman
filter see Hamilton (1994b, p. 387 f.).
4 Filtering Algorithms

According to the assumed state space form of equations (3.1) and (3.2),
we specify stochastic processes of explaining factors eo
which describe the
state of the system. Instead of being able to observe the factors directly, we
can only observe some noisy function Yt of et.
The problem of determining
the state of the system from noisy measurements Yt is called estimation.
The special estimation problem of filtering has the object of obtaining an
et
expression for the optimal estimate of given the observations up to time t.
The most successful result of this kind is that obtained for linear systems by
Kalman (1960), Kalman and Bucy (1961), and Kalman (1963) which we are
further dealing with. In general, the estimation problem can be classified
upon the available and processed information into three different problems
according to the following definition.

Definition 4.0.9 (Estimation) Considering the problem of estimating et

using information up to time s, denoted by the information set Fs = {ys, . . . ,
Y2, Yl}, we differentiate between the three cases of a

filtering problem for t = s,

smoothing problem for t < s, and
prediction problem for t > s,

dependent on which information set we use in estimation. 9

Furthermore, we need a clear concept of what constitutes a statistically

optimal estimate in our context. The concept we use is the statistical cri-
terium of mean square error which we define next. We also present the
solution to the problem of the optimal estimator. Finally, we introduce
convenient notations for the exposition of the Kalman filter derivation.
9See, for example, Jazwinski (1970, p. 142 f.) and Harvey (1993, sec. 4.2).
16 4. Filtering Algorithms

Definition 4.0.10 (Optimality Criterium) Let e;lt-1 denote an estimate

of et based on the information set Ft- I . In order to choose the optimal of
various possible forecasts, we need to specify a criterion of what optimal
means. Very convenient results are obtained from assuming a quadratic loss
function L( et) on the estimation error et = et - e;lt_I' lO Choosing the
estimator as to minimize the mean square error

results in the minimum mean square estimator (MMSE) as the best or opti-
mal estimator with respect to any quadratic function of the estimation error.

Theorem 4.0.11 (Optimal Estimator) Given the availability of an in-

formation set :Ft.-I, the optimal estimator etlt-I of et among all estimators
e;lt-1 is the expected value of et-I conditional on the information at time
t - 1; thus the MMSE is given by ~tlt-I = E [etIFt-d.ll The corresponding
MSE is given by

i. e. in this case the mean square error matrix is equal to the variance-
covariance matrix.

Notation 4.0.12 (Conditional Expectation and Variance) In consid-

ering the problem of estimating et using information up to time s, i. e. the
information set Fs = {ys, . .. ,y2, Y I}, we denote the conditional expecta-
tion of et given Fs in the further analysis for convenience by E [etlFsl == etls'
For the second conditional moment we will further use Cov [etlFsl == ~tIs'
the conditional variance-covariance matrix of et given Fs.

4.1 Linear Filtering

Having gone through the necessary preliminaries, we will see in this section
that the Kalman filter is a set of equations which allows an estimator to
be updated once new observational information becomes available. This
process is carried out in two distinct parts:
lOSee, for example, Jazwinski (1970, p. 146 f.).
llSee, for example, Hamilton (1994b, p. 72 f.).
4.2. MMSE 17

1. Prediction Step: The first step consists of forming an optimal pre-

dictor of the next observation, given all the information available up
to time t - 1. We extrapolate the state vector by means of conditional
expectation utilizing the information set Ft.-I and calculate a so-called
a priori estimate for time t.

2. Updating Step: The a priori state estimate is updated with the

new information arriving at time t that is combined with the already
available information from time t - 1. The result of this step is called
the filtered estimate or the a posteriori estimate. Therein the Kalman
gain in estimation is realized.

We will find that the Kalman filter provides an optimal solution to the
presented problems of prediction and updating. In deriving the Kalman
filter we can choose among several approaches available in the literature
that are linked to different interpretations of the filter. Besides the original
derivation in the work by Kalman (1960), where he used the idea of orthog-
onal projection (see, for example, Aoki (1990) and Brockwell and Davies
(1987)), we can derive the Kalman filter from the properties of a multi-
variate normal distribution as shown, for example, in Harvey (1989), while
Burridge and Wallis (1988) and Hamilton (1994b), for example, deduce the
Kalman filter to be the minimum mean square linear estimator. Moreover,
alternative derivations are found in Jazwinski (1970). In the following we
present the elementary derivation of the Kalman filter under the normal-
ity assumption which yields the interpretation of the Kalman filter as the
optimal filter in the sense of a minimum mean square estimator (MMSE).
Also, we obtain the Kalman filter for the general case of non-normality by
exploiting the given linear relationships of the observations and dynamics
stated respectively in the state space equations (3.1) and (3.2); in this case
we speak of the Kalman filter as the optimal filter in the sense of a minimum
mean square linear estimator (MMSLE).

4.2 MMSE
In order for the Kalman filter to yield a MMSE, we assume that the ad-
ditive error terms et and "1t are independently and normally distributed.
FUrthermore, the error terms are treated as independent of the initial state
18 4. Filtering Algorithms

vector, which is assumed to be normally distributed with IE [eol = eo and

Cov [eol = r:o. Therefore, since the transition equation (3.2) is linear in et-l
and the error term "1u the state vector et - as the sum of et-l and "1t - is also
normally distributed. Furthermore, for the measurement equation (3.1) we
also have a normally distributed Yt, since et and et are both normal. Thus,
we are dealing with a model with Gaussian signals perturbed by additive
Gaussian noise.
In the prediction step, we forecast the state vector by calculating the
conditional expectation of the state variables on both sides of equation (3.2)
given the information up to time t - 1:
(4.1)
The corresponding variance-covariance matrix for the state variables is given
by

r:tlt-I - IE [(et -IE [etIFt-l]) {et -IE [etIFt-I])'IFt-l]

IE [ (et - etlt-I) (et - etlt-I)' 1Ft-I]
IE [( Ct + <I>tet-I + "1t - Ct - <I>tet-Ilt-I)
(Ct + e~-l <I>~ + '1Jt - Ct - e~-llt-l <I>~) IFt-d
<I>t IE [et-Ilt-I e~-llt-IIFt-l] <I>~ + IE ["1t"1~IFt-ll
<I>tr:t-Ilt-I <I>~ + Qt· (4.2)
Equations (4.1) and (4.2) are known as the prediction equations of the
Kalman filter.
As an intermediate result we define the prediction error denoted by Vt,
using the given data for the measurable observations Yt, as

Vt = Yt - Ytlt-I = Yt -IE [YtIFt-ll = Yt - at - Btetlt-I'

and are able to calculate the variance-covariance matrix of Vt as

Ftlt- I Cov [vtIFt-d

IE [(Yt -IE [YtIFt-l]) {Yt -IE [YtIFt-I])'IFt- l]
= IE [(at + Btet + et - at - Btetlt-I)
(~+ e~B~ + e~ - ~ - e~lt-IBD 1Ft-I]
BtIE [etlt-Ie~lt-IIFt-l] B~ + IE [ete~IFt_ll
(4.3)
4.2MMSE 19

In the last step, the updating step, we update the inference on etlt-l
by including the newly available information at time t, which results in
the filtered estimate etit.
Hereby, we can realize the so-called Kalman gain
K t . In order to obtain the updating equations, we first consider the joint
distribution of et and Yt. Since both variables are normally distributed, the
joint distribution given the information F t - 1 is

[ et ] '" N ([ et1t-l] [ :Etl t- 1 (4.4)

Yt Ytlt-l ' Bt:Etlt- 1
with the conditional means

etlt-l - E [etIFt-d = Ct + 4>tet-llt-l

Ytlt-l E [YtIFt-1l = at + Btetlt-l'
and the conditional variance-covariance matrices stated in equations (4.2)
and (4.3), and

E [(et - etlt-l)
(et - etlt-l)' 1Ft-I] B~ = :Etlt-lB~
E [(Yt - Ytlt-I) (et - etlt-l)' 1F t-I] = Bt:Etlt-l.

For the further derivation, we take advantage of the following result for
normal variables.

Lemma 4.2.1 (Conditional Normal Distribution) 12 Letzl andz 2 de-

note (nl xl) and (n2 xl) vectors of random variables respectively that fol-
low a joint normal distribution:

Then the distribution of Zl conditional on Z2 is N (m,:E) where

m 1-'1 + 012 0 z-l (Z2 - 1-'2), and

:E = Ou - 0120z-lo~2.
12See, for example, Gourieroux and Monfort (1995, p. 481 f.) and for a proof Anderson
and Moore (1979).
20 4. Filtering Algorithms

Thus the optimal forecast of Zl conditional on having observed Z2 is given

by

with ~ characterizing the mean square error of this forecast:

Applying this result to equation (4.4) yields the following distribution

for et given the information F t of the observable data Yt:

et '" N (etlt-l + ~tlt-1B~F~Ll (Yt - Ytlt-l) ,

~tlt-l - ~tlt-1B~F~LlBt~tlt-l) .

Comparing this distribution with the result for the conditional moments of
et given the information F t ,

we obtain the following relationships, where Vt denotes the prediction error:

etjt etlt-l + ~tlt-1B~F~Ll Vt

~tlt ~tlt-l - ~tlt-1B~F~LlBt~tlt-l.

This finally results in the update for the state vector

(4.5)

and its variance-covariance matrix

~tlt = E [(et - etlt) (et - etlt)' 1Ft]

= ~tlt-l - KtBt~tlt-l
(I - KtBt ) ~tlt-b (4.6)

where we substituted K t = ~tlt-1B~F~Ll for the Kalman gain matrix.

4.3. MMSLE 21

4.3 MMSLE
Dealing with a non-normally distributed state vector but confining ourselves
at the same time to the class of linear estimators, we derive the Kalman filter
as the MMSLE without requiring the normality assumption for the error
terms. The prediction equations (4.1) and (4.2) are obtained similarly to
the previous derivation by taking the conditional expectation and variance
given the information up to time t - l.
For obtaining the updating equations, we begin with stating a general
linear relationship of the updated estimator etit
and the information of the
present sample (yt, at, Ct) and the past information F t - 1 . We will assume
the general linear form

with arbitrary matrices K t , L, M, and N. These matrices are now to be

chosen such that etit
is the MMSLEj therefore we define the estimation error

et et - etit
= Ct + <pet-l + '11t - K t (at + B t (Ct + <ptet-l + '11t) + et)
-Lat - MCt - N (et-l - et-l)
(<pt - N - KtBt<p) et-l - (L + K t ) at + (I - M - KtB t ) Ct
+Net-l + (I - KtB t ) '11t - Ktet. (4.7)

For etitbeing the MMSLE the estimation error et needs to be (i) uncondi-
tionally unbiased, i.e. the estimation error has zero expectation, which gives
us the conditions

0,
0, and

for equation (4.7). This yields, with the substitution N = (I - KtB t ) <Pt ,
the estimation error
22 4. Filtering Algorithms

Furthermore, the estimation error et is required to have (ii) minimum vari-

ance, i.e. we need to minimize the estimation errors variance-covariance ma-
trix

L: t1t lEJ [( (I - KtB t ) (4)tet-l - 'TJt) - Ktet)

((I - KtB t ) (4)tet-l - 'TJt) - Ktet)']
= (I - KtBt) (4)tL:t-llt-l4>~ + Qt) (I - KtBS + KtHtK~
= (I - KtB t ) L: t1t - 1 (I - KtBt)' + KtHtK~

with respect to Kt; for the partial derivative we get: 13

8~t (L:t1t- 1 - KtBtL:tlt-l - L:tlt-lB~K~

+KtBtL:tlt-lB~K~ + KtHtKD
- (BtL:tlt-l)' - L:tlt-lB~ + 2BtL:tlt-lB~Kt + 2HtK t
-2L:tlt-lB~ + 2BtL:tlt-lB~Kt + 2HtK t.
Setting the result equal to zero according to the necessary minimization
condition, we can solve for the Kalman gain matrix

which is equivalent to the expression derived in equation (4.6) for the case
of the Kalman filter resulting in the MMSE.

4.4 Filter Recursions

Having derived the Kalman filter algorithm in two different ways yielding
the MMSE and the MMSLE, we are now able to apply the sequence of
filtering equations recursively as each new observation becomes available.
Please compare the flowchart of figure (4.1) for a graphical illustration of
the Kalman filter algorithm. The recursion algorithm starts with a feasible
choice of the parameter vector 'ljJ and corresponding state vector with eo
variance-covariance matrix L:o. Thereupon, we use the prediction equations
to calculate the a priori estimates
13For derivatives of functions with arguments in matrix form see, for example, Llitke-
pohl (1996).
4.4 Filter Recursions 23

Kalman Filter and MLE

Initialization of
~and;
at Time t=O
The system is initialized
with specific (prior) values
for the state variables and
their covariance matrix.

Essentially, these three

New distinctive calculations
Observations constitute the Kalman
filter algorithm.
Yt
at Time
t=t+1

t=T?
The recursion algorithm is
performed until we reach the
yes last observations.

Choose
other
Parameter The conditional likelihood
Values function is evaluated at the
current parameter values using
'V the results of the Kalman
recursions.

L----no'----<
Maximization
Criterium met?
\. The Kalman algorithm is run
. until we reach the predefined
abortion Criterium.
yes

8
Figure 4.1: Flowchart of the Kalman Filter and MLE
24 4. Filtering Algorithms

etlt-i Ct+ cI>tet-ilt-i, and

~tlt-i = cI>t~t-ilt-i cI>~ + Qt.
Using the current market information on observable variables Yt on date
t = 1, we next derive the prediction error and the corresponding MSE as

Vt = Yt - at - Btetlt-i, and
Ftlt- i = Bt~tlt-iB~ + H t .

Given these intermediate results we then update the a priori estimates yield-
ing the filtered estimates

etit etlt-i + KtVt

~tit (I - KtBt) ~tlt-i

as optimal in the sense of MMSE or MMSLE using the Kalman gain K t .

Now for each discrete time step, we recursively feed the results for the
state vector etlt and its variance-covariance matrix ~tlt from the updating
equations into the prediction equations until we reach the last observations
YT. 14 Finally, with the obtained time-series of the state variables we can
evaluate the likelihood function as we describe in chapter 5 to choose other
more appropriate values for the parameters ..p.

4.5 Extended Kalman Filtering

In the previous section we have been dealing with the case of linear measure-
ment and transition equations and normally distributed error terms as spec-
ified in equations (3.1) and (3.2). However, some applications derived from
financial theory exhibit non-linear functional relationships of non-normally
distributed state variables in the measurement and transition equations. In
such cases explicit expressions for the filtering algorithms cannot be derived,
i.e. some approximations are necessary for the estimation procedure. There
are two main approaches to obtain non-linear filtering algorithms. 15 The
14Note that these are then formally indexed with t - 1, since we denote the actual
information we are working with by the time index t.
15See, for example, Anderson and Moore (1979) and Tanizaki (1996).
4.5 Extended Kalman Filtering 25

first approach is to approximate the non-linear measurement and transi-

tion equations. The linearized non-linear functions are then applied to a
modification of the linear Kalman filter algorithm as derived in the previ-
ous section. The second types of algorithms can be summarized under the
approach of approximating the underlying density functions of the state
vector. Then a recursive algorithm on the densities is derived using Bayes's
formula. The advantage of this approach is that it results in asymptotically
unbiased filtering estimates. However, the estimators based on the density
approach require a great amount of computational burden compared with
those based on the Taylor series approximations. 16
FUrther, we constrain our analysis of Kalman filters on algorithms that
are derived from the Taylor series expansions, i.e. we only deal with extended
Kalman filters. For the exposition of the extended Kalman filter algorithms
we work with the following general state space model.

Definition 4.5.1 (Non-Linear State Space Model) We treat the non-

linear filtering problem based on a state-space model with the measurement
and transition equations being specified as

Yt = gt (~t' E:t ('ljJ) ,'ljJ), and (4.8)

gxl

~t
kxl
= h t (~t-l' "'t ('ljJ) , 'ljJ) , (4.9)

with g (~t' E:t ('ljJ) , 'ljJ) and h (~t-l' "'t ('ljJ) , 'ljJ) denoting the possible non-linear
functional relationships. The error terms E:t ('ljJ) and"'t ('ljJ) are assumed to
follow the properties

IE [ E:t ('ljJ) ] 0, and

"'t ('ljJ)
[E:t ('ljJ) ]
\U

'" ar "'t('ljJ)

as normally distributed random vectors.

In the case of the extended Kalman filters we approximate the non-linear

equations gd~t,E:t,'ljJ) and h t (~t-l''''t''ljJ) around the conditional means
16For a detailed comparison of different non-linear filtering algorithms see Tanizaki
{1996}.
26 4. Filtering Algorithms

and error terms (et,et) = (etlt-l,O) and (et-l,1]t) = (et-llt-l,O). The

first-order Taylor series expansions around the vectors (et,et) = (etlt-l,O)
result in the approximate expression

Yt gt (etlt-l, 0, 1/J) + B tlt- l (et - etlt-l) + Rtlt-let,

:::::J (4.10)
wl· th B tlt-l -- 8g.(et,Et"p)
8e' ,
Iand
t (et,Et,t/J )=(et t - 1,O,t/J )
l
D _ 8gt (et ,Et,t/J) I
~"tlt-l - 8E' )
t (et,Et,t/J)=(etlt_1,O,t/J
for the measurement equation. The transition equation is approximated
around (et-l,1]t) = (et-llt-l,O) which results in:
et :::::J h t (et-llt-l'O,-rP) + <I>tlt-l (et-l -et-llt-l) +Stlt-l1]t, (4.11)
with <I> - 8ht(et_1,71t,t/J)
tlt-l - 8e'
I ,
and
(et-1,71t,t/J )=(et- 1It- 1,o,t/J)
I
t-1
Stlt-l _- 8h t (et-1 ,71t,t/J )
877i .
(et-1 ,71t,t/J )=( et-1It-1 ,O,t/J )
The state space model of equations (4.10) and (4.11) can further be stated
as
Yt :::::J at + Btlt-let + Rtlt-let
et :::::J Ct + <I>tlt-let-l + Stlt-l1]t,
with at = gt (etlt-l, 0, -rP) - Btlt-letlt-l, and
Ct = ht (et-llt-l, 0, 1/J) - <I>tlt-let-llt-l
following the treatment of the linear state space model of equations (3.1)
and (3.2). For the approximated state space model we are able to derive
the following modification of the linear Kalman filter algorithm:17

etlt-l = h t (et-llt-l, 0, -rP) ,

~tlt-l = <I>tlt-l~t-llt-l <I>~lt-l + Stlt-l QtS~lt-ll
Vt = Yt - gt (etlt-l, 0, -rP) ,
F tlt-l B tlt- l ~tlt-l B~lt-l + R tlt- l Ht~lt-l ,
Kt = (Btlt-l~tlt-l)/F~Ll'
etit = etlt-l + KtVt, and
~tit = ~tlt-l - KtFtlt-lK~.
~~------------------
17See Tanizaki (1996, ch. 3)
4.5 Extended Kalman Filtering 27

This system of equations can recursively be implemented in analogy to the

algorithm in the linear case presented in the previous section.
5 Parameter Estimation

The overview of the Kalman filter estimation procedure in figure 4.1 shows
the likelihood function and the maximization procedure as the last rele-
vant steps. We start to examine the appropriate log-likelihood function
within our framework. Let 'I/J Ewe ]Rn be the vector of the n unknown
parameters, which various system matrices of the state space model for-
mulated in equations (3.1) and (3.2), or (4.10) and (4.11), as well as the
variances of the measurement errors depend on. The likelihood function of
the state space model is given by the joint density of the observational data
Y = (YT, YT-l, ... ,Yl)

which reflects how likely it would have been to have observed the data if 'I/J
were the true values for the parameters. Using the definition of conditional
probability, we can split the likelihood up into conditional densities using
Bayes's theorem recursively and write the joint density as the product of
conditional densities

l(y;'I/J) =p(YTIYT-l, ... ,yl;1/J)· ····p(YtIYt-l, ... ,Yl;'I/J)· ... ·P(Yl;'I/J) ,

(5.1)

where we approximate the initial density function P (Yl; 'I/J) by P (YIlyo; 'I/J).
Moreover, since we deal with a model with a Markovian structure according
to the state space model, i.e. future values of Yu, with u > t, only depend
on (Yt, Yt-l,··· ,Yl) through the current value Yt, the expression in equa-
tion (5.1) is equivalent to conditioning on only the last observed vector of
observations, i.e. it reduces to
30 5. Parameter Estimation

For our purpose of estimating the parameter vector 1/; given the data
Y and the structural form of the specified state space model, we use the
approach of Schweppe (1965) known as the prediction error decomposition
of the likelihood function to be maximized with respect to 1/;. That is, we
express the likelihood function in terms of the prediction errors; since the
variance of the prediction error Vt = Yt - lE [YtIFt-1l is the same as the
conditional variance of Yt, i.e.

we are able to state the density function. The general density function
p (YtIYt-l; 1/;) is given by the Gaussian distribution with conditional mean
lE [YtIFt-ll = a + Betlt-l and the conditional variance-covariance matrix
Cov [YtIFt-d = Ftlt- 1, which takes the form:

(5.3)

using the substitution Vt = Yt -lE [YtIFt-1l. Taking the logarithm of equa-

tion (5.3) yields the expression

which gives us the whole log-likelihood function in the innovation form as

the sum

1 T
L (y; 1/;) = -"2 L kIn (271") + In IFtlt-11 + v~Ftlt-lVt (5.4)
t=l
instead of the product in equation (5.1).
In order to estimate the unknown parameters from the log-likelihood
function in equation (5.4) we need to choose an appropriate optimization
algorithm to minimize L (y; 1/;) with respect to 1/;. Given the simplicity of
the innovation form of the likelihood function, we can find the values for
the conditional expectation lE [YtIFt-1l and conditional variance-covariance
matrix Cov [YtIFt-1l for every given parameter vector 1/; using the Kalman
filter algorithm. Thus, we are able to compute the value of the conditional
5. Parameter Estimation 31

log-likelihood function numerically. This can then be used in an optimiza-

tion algorithm of the log-likelihoodjunction L (y; 1/J) in order to calculate
the maximum likelihood estimates 1/JML of the parameter vector 1/J:

:¢ML = argmaxL(y;1/J), (5.5)

t/JE\II

where \II c ]Rn denotes the parameter space. In the case of implementing
extended Kalman filter algorithms where we use approximations to the con-
ditional normal distribution of equation (5.3) the log-likelihood of equation
(5.5) can be used to yield quasi maximum likelihood parameter estimates. 18
However, the general statistical properties of such estimates must be studied
separately and are assessed in detail by White (1982), Gallant and White
(1988), and White (1994).
In the numerical implementation to obtain the optimal parameter es-
timates we actually minimize the log-likelihood function of equation (5.5)
including its opposite sign. In infinite precision the necessary conditions
for a local minimum :¢ML of L(y; 1/J) are defined by the conditions for the
gradient:

(5.6)

assuming the relevant partial derivatives exist. A sufficient condition is that

the matrix of second order de~vatives exists and is positive definite at the
optimal parameter estimates 1/JML' i.e.

(5.7)

is needed.
In order to estimate the model parameters 1/J M L we decided to choose a
quasi-Newton method, which is similar to the methods used by Lund (1997)
and Nunes and Clewlow (1999) based on Dennis and Schnabel (1996). The
iteration rule for Newton's method for unconstrained minimization is given
by the parameter estimates at the ith iteration step of

18For an exploration of quasi (or pseudo) maximum likelihood estimation methods see,
for example, Gourieroux, Monfort, Renault, and Thognon (1984).
32 5. Parameter Estimation

where the variable Si is found by solving

For the construction of \7 L (y; 1/Ji) and \7 2 L (y; 1/Ji) we use numerical ap-
proximations. We found that the available closed-form solutions19 do not
lead to superior results in that they accumulate numerical errors in their re-
quired extensive calculations while asking for multiples of computer time. In
cases where \7 2 L (y; 1/Ji) is not positive definite Dennis and Schnabel (1996,
sec. 5.5) suggest a modification of the modified Cholesky factorization.
This algorithm changes the model when necessary so that it has a unique
minimizer and uses this minimizer to define the Newton step. Specifically,
at each iteration we take steps of the form

by adding small parts Pi of an identity matrix I. Therein, Pi is ideally not

much larger than the smallest P that will make \7 2 L (y; 1/Ji) + pI positive
definite and reasonably well conditioned.
By now the optimization algorithm implies a full quasi-Newton step
which will be taken whenever possible. In addition, we implement a back-
tracking line search framework for the step length Ai. For properly chosen
steps Dennis and Schnabel (1996, sec. 6.3) show that the algorithm ensures
global convergence. Thereby, the optimization algorithm we use is a glob-
ally convergent modification of Newton's method. Its iteration rule of the
optimization process can be represented as

where 1/Ji are the parameter estimates at the ith iteration step.
Moreover, looking at the bottom part of the flowchart in figure 4.1, we
need to discuss how to terminate our optimization algorithm in finite preci-
sion, i.e. we further have to examine the conditions stated in equations (5.6)
and (5.7).20 Although \7 L(y; 1/J) = 0 can also occur at a maximum or sad-
dle point, our globalizing strategy and our method of perturbing the model
Hessian to be positive definite make convergence impossible to maxima and
19 As given, for example, in Harvey (1989, ch. 3.4).
20See Dennis and Schnabel (1996, ch. 7.2).
5. Parameter Estimation 33

saddle points. Therefore, we consider V'L(y; 'ljJ) = 0 to be a necessary and

sufficient condition for 'ljJ to be a minimizer of L(y; 'ljJ).
To test whether V'L(y; 'ljJ) = 0, a test such as

IV'L(y;'ljJ)1 ~ c, (5.8)

with a tolerance level of, for example, c = 10-4 , is inadequate, because

it is strongly dependent on the scaling of both L(y; 'ljJ) and 'ljJ. A direct
modification of equation (5.8) is to define the relative gradient of L(y; 'ljJ)
at 'ljJ by
L(Y;'1/J+OLi)-L(Y;1/1)
reI grad ('ljJ). = lim L(y;1/1)
V'L(y; 'ljJ )i'ljJi
t 6--+0 ..§....
1/1;
L(y; 'ljJ)

With the given relative gradient we perform the test

(5.9)

which is independent of any change in the units of L(y; 'ljJ) and'ljJ. However
the idea of relative change in 'l/Ji and L(y; 'ljJ) breaks down if'ljJi or L(y; 'ljJ)
happen to be close to zero. This problem can be fixed by altering the test
of equation (5.9) to

(5.10)

where typ (x) denotes an estimate of a typical magnitude of x. Thus, we

use the test stated in equation (5.10) in the Kalman filter applications.
 Part II

Pricing Equities
6 Introduction

6.1 Opening Remarks

In part II of the study we investigate the pricing of equities. Within this

asset class we draw our attention to the case of closed-end funds. We first
describe the specific characteristics of closed-end funds and their motivating
features for financial research, especially for issues of pricing. Thereupon,
we develop a valuation model in chapter 7 by means of contingent claim pric-
ing techniques attempting to capture the certain financial characteristics of
closed-end funds. The stochastic pricing model especially takes into account
both the price risk of the funds as well as their risk associated with altering
discounts. Chapter 8 describes a possible implementation of the pricing
model and introduces the estimation techniques used in the further analy-
sis. For the empirical adaptation of the valuation model we use a sample
of closed-end equity funds that invest in emerging markets and are traded
on the New York Stock Exchange. The statistical inferences are based on
a historical sample for the five-year-period of 1993 to 1997 using maximum
likelihood estimation based on a Kalman filter algorithm. We estimate the
relevant parameters and validate our model. In chapter 9 we are able to
infer insights into two potential applications for investors. First, we test
the forecasting power of the pricing model to predict closed-end fund mar-
ket prices. Second, based on information that is revealed in the valuation
model, we implement portfolio strategies using trading rules. The results
on these suggested applications indicate that our pricing model generates
valuable investment information. Finally, we summarize and interpret our
theoretical and empirical findings in chapter 10.
38 6. Introduction

6.2 The Case of Closed-End Funds

In the categories of asset classes closed-end funds lie in between the class
of common equity and open-end funds, which are generally known as mu-
tual funds. In thus, they exhibit features of both asset classes which are
worthwhile examining: (i) Closed-end funds are companies whose opera-
tions resemble those of any business corporation. Their shares are regularly
traded on organized exchanges as any other publicly traded corporation.
Closed-end funds only differ because their corporate business largely con-
sists of investing funds in the securities of other entities and managing these
investment holdings for income and profit. The important characteristic
which makes closed-end funds unique among other joint-stock companies is
that they provide simultaneous price quotations for both their stocks and
their underlying investment portfolio. Especially, the share price behavior
often results in prices being different from the value of their underlying in-
vestments as we will see further on. (ii) Closed-end funds are investment
companies that provide investment management and bookkeeping services
to investors. Especially, investors who do not have the time or expertise
to manage their own funds rely on the services of investment companies.
To qualify as such in the United States their activities are regulated by
the investment Company Act of 1940 and by the 1950 amendments to the
act. 21 The two major types of investment companies in the United States
are open-end and closed-end investment companies. Open-end funds con-
tinuously issue and redeem ownership shares which are not traded on a
secondary market or organized exchange. Instead investors purchase shares
from the company and redeem shares by selling them back to the company.
In thus, the equity capital and assets of a mutual fund are increased when
shares are sold and reduced when shares are repurchased. The second class
of investment companies are closed-end funds whose capitalization or out-
standing number of shares is fixed or "closed". This implies that the supply
of closed-end fund shares is inelastic, i.e. the price of the shares is a function
of the supply and demand for the shares on the market. Therefore, there is
only an indirect link with the value of the underlying investment portfolio,
the so-called net asset value, corresponding to each share.
Combining these two aspects, the interesting feature of closed-end funds

21S ee, for example, Anderson and Born (1992).

6.2 The Case of Closed-End Funds 39

in comparison to common joint-stock companies and mutual funds is, that

the market value determined on an organized exchange differs dynamically
over time from the value of the underlying investment portfolio by a dis-
count or a premium. This pricing phenomenon contradicts the well-known
efficient market hypothesis put forward by Fama (1970) and Fama (1991)
which states that securities markets are expected to be informationally effi-
cient. As noted by Malkiel (1977), the market for closed-end funds provides
a "startling counter-example to the general rule" .22 He investigates several
real world obstacles that could prevent the parity of the closed-end funds
share prices and the net asset values. As a result he finds that the consid-
ered hypotheses can only account for a fraction in the observed discounts.
Especially, the tax liability hypothesis of unrealized capital gains gives sup-
port to explain discount values of up to six percent. However, US domestic
equity closed-end funds have traded at an average discount of around ten
percent over the last thirty years. 23
The various proposed hypotheses aiming at an explanation of the puzzle
with the existing and observable closed-end fund discounts have grown to a
wide body.24 Broadly, the research put forward can be categorized into the
standard economic theories and the behavioral explanations. The standard
economic theories attempt to explain the observable premia within the ef-
ficient markets framework. At first, there are explanations that aim at the
possibility that the closed-end funds' underlying portfolio values may be
overestimated. This could be the case in that the net asset values do not
reflect the capital gains tax that must be paid by the closed-end funds if
appreciated assets in the fund are sold. 25 A further theory is given by the
restricted stock hypothesis which states that substantial holdings of letter
stock could be overvalued in the calculation of the net asset value. 26 Sec-
ond, agency costs could create closed-end fund discounts if management
expenses and fees are too high or if future portfolio management and per-
formance is expected to be inadequate. 27 A last group of hypotheses focuses

22Malkiel (1977, p. 847).

23See Swaminathan (1996).
24For a rich review of the evolving literature on closed-end funds see, for example,
Anderson and Born (1992) and Dimson and Minio-Kozerski (1998).
25See, for example, Malkiel (1977), Lee, Shleifer, and Thaler (1990), and Pontiff (1995).
26See the studies by Malkiel (1977) and Lee, Shleifer, and Thaler (1990).
27See, for example, Roenfeldt and Thttle (1973), Ingersoll (1976), Malkiel (1977), and
40 6. Introduction

on various forms of market segmentation. Especially, market segmentation

arises internationally when US investors diversify into inaccessible, less well
regulated and under-researched foreign markets. 28 However, these standard
arguments, even when considered together, do not explain all of the closed-
end fund puzzle. 29
The observed price differences with closed-end funds are even more in-
teresting since closed-end funds are assets of high transparency in the sense
that the companies underlying fundamental values are quite easy to deter-
mine compared to common equity. Instead, traded on organized exchanges
the market prices of the closed-end funds seem to reflect more the supply
and demand for these shares than the value of their underlying assets. The
latter would be expected in a rational and efficient market. This brings us to
the second line of closed-end fund literature which concentrates on behav-
ioral explanations of the observable discounts. Here, the most prominent
theory is the investor sentiment hypothesis proposed by Lee, Shleifer, and
Thaler (1991).30 The proposition is that fluctuations in premia are driven
by changes in individual investor sentiment. By finding that closed-end fund
premia are a measure of the sentiment of individual investors they conclude
that a changing sentiment makes the funds riskier than their underlying
portfolios and such causes the average underpricing of closed-end funds rel-
ative to their fundamentals. Thus, if the dynamic behavior of closed-end
fund discounts reflects investor sentiment and is not conformable to rational
asset pricing models, then there may be trading possibilities for investors on
exploiting closed-end fund discounts. However, Swaminathan (1996) shows
that the negative correlation between changes in discounts and small firm
returns reported by Lee, Shleifer, and Thaler (1991) can be explained by the
ability of the discounts to predict future small firm returns which are also
affected by investor sentiment. He further provides evidence that informa-
tion contained in the discounts may be related to fundamental explanations
which suggests the possibility of a rational explanation for the described

Thompson (1978).
28 Especially, see the studies by Bonser-Neal, Brauer, Neal, and Wheatley (1990) and
Diwan, Errunza, and Senbet (1995).
29See Lee, Shleifer, and Thaler (1991).
3 0 The hypothesis is based on the noise trader model of De Long, Shleifer, Summers,
and Waldmann (1990) who argue that this concept can explain both the persistent and
variable premia on closed-end funds as well as the creation of such funds.
6.2 The Case of Closed-End Funds 41

negative correlation. This leads Chordia and Swaminathan (1996) to build

a noisy rational expectations model which includes market imperfections31
instead of the existence of rational and irrational noisy traders as in the irra-
tional asset pricing model of Lee, Shleifer, and Thaler (1991). Their model
is capable of creating closed-end fund discounts endogenously within a ra-
tional setting which brings us back to the standard economic explanations.
They identify further market imperfections that may prevent investors from
arbitrage gains using information on closed-end fund discounts. 32
An empirical study that examines the relevance of both, the standard
theories as well as the hypothesis of investor sentiment, is provided by Gem-
mill and Thomas (2000). For a sample of UK closed-end funds they find
support for a rational basis for the existence of a persistent discount which
is considered to be driven by management expenses and asymmetric arbi-
trage. Further, they confirm that short and medium-term fluctuations of
the premia are related to investor sentiment while they reject the hypothesis
that noise is the cause of the observable discounts.
We currently see various theories of the determination and the variation
of closed-end fund discounts available with no distinct prevailing hypothesis.
The consensus of the different aspects is basically revealed in the investors'
supply and demand behavior and is finally disclosed in the realized market
prices of the closed-end fund shares. Upon the fact that dynamically chang-
ing premia and discounts predominantly exist in the market of closed-end
funds we develop a valuation model to price closed-end fund shares. The
objective is to provide a theoretical and empirical analysis of the dynamic
price behavior of closed-end funds by incorporating the major types of risk
associated with an investment in closed-end funds. These types of risk are
considered to be the common price risk due to changing underlying portfo-
lio values as well as the risk associated with altering values of the premia.
In the following chapter we derive a two-factor pricing model that utilizes
the net asset values of the closed-end funds and a dynamically changing
premium as state variables.

31 They assume market segmentation and imperfect information.

3 2 The main issues are found to be related to securities regulations, fiduciary responsi-
bilities, and free-rider problems.
7 Valuation Model

7.1 Characteristics of Closed-End Funds

The capital markets for closed-end funds provide two important prices for
financial analysis: The market value of the closed-end funds' shares as de-
termined on organized exchanges, and the net asset value of their foreign
assets which is reported by the investment companies.

Definition 7.1.1 The net asset value (NAV) or the 'book value of share-
holder's equity' for time t is given by

N Avt = (Total Assets - Total Liabilities) / (Shares Outstanding)

which generally differs from the market price Pt of the closed-end funds. 33
This price difference per share is generally quantified by

PREMt = In (Pt!NAvt) (7.1)

which defines the empirical premium.

The empirical premium serves as a characteristic number for closed-end

funds and is regularly published in the financial press and contained in
financial data bases. From its definition the empirical premia can realize
positive and negative values. 34 This is confirmed empirically whereas we
rarely find values for the premia which are zero or even close to zero. This
partly stems from the fact compared to open-end mutual funds, that closed-
end fund investors agree to trade their shares at prices differing from the
33The net asset value and the market price are generally stated per share.
34Negative empirical premia correspond to the case of closed-end funds trading at
discounts.
44 7. Valuation Model

market value of the closed-end funds assets on organized exchanges. In order

to liquidate a holding in a fund, the investors need to sell their shares to
other investors for the market price instead of selling and redeeming them
at or near their net asset value as with open-end funds. Thus, closed-end
funds are endowed with an inelastic supply of shares.

20 0.8
. . . . . .. net asset value
16 - - market price 0.6

12 0.4

.~
Co 8 0.2

0.0 E
~
o+_--.--_,_-----,.-----.----r---.------r-____,.------.----+ -0.2
E 30% 0 52 104 156 208 260
.~ 20%
!!! 10%
~ 0%~~4r~~-_4H4~--_r~~------~~.-------------
.g -10%
'~-20%
m_~%+_--.--_,_--____,.------.-----r----.---____r--____,.------~~~
o 52 104 156 208 260

Figure 7.1: ROC Taiwan FUnd (ROC)

Considering the closed-end funds in our sample35 we can identify three

different patterns of empirical premia behavior over time: (i) Premia which
change from positive to negative values over the examined period of time,
(ii) mainly positive premia over time, and (iii) negative premia or discounts
for the major part of the sample. Out of our examined sample we choose
three representative funds for each case: The ROC Taiwan FUnd (ROC),
the Indonesia Fund (IF), and the GT Global Eastern Europe (GTF). The
time-series properties of these funds are illustrated, respectively, in figures

35For the detailed description of the data set see section 8.1 on page 55.
7.1 Characteristics of Closed-End Funds 45

20 1.00
....... net asset value
16 - - market price 0.75
Ul 12
Q)
0
.~ 0.50
8

4 0.25
0
0.00 Ul
-4 E
~
-8 -0.25

o 52 104 156 208 260

Figure 7.2: Indonesia FUnd (IF)

7.1, 7.2 and 7.3. 36

The figures show the closed-end fund prices for the market shares and
the reported net asset values in the upper graph combined with a graph of
the total returns to these prices. The lower graph displays the time-series
of the empirical premia of the funds. In the case of the ROC Taiwan FUnd
(figure 7.1) the premium oscillates between positive and negative values.
Looking at figures 7.2 and 7.3, we see that the Indonesia Fund and the GT
Global Eastern Europe, respectively, exhibit primarily positive and negative
values for the premium. For all three funds, we further notice that the price
returns show a higher volatility for the market prices than for the net asset
values.
The visual examination of the three selected closed-end funds from the
sample seems to be in line with the existing literature on closed-end fund
valuation. The literature sees the empirical premia of the closed-end funds
to be subject to wide fluctuations over time and by no means a constant
fraction of the net asset value (see Lee, Shleifer, and Thaler (1991)). As we

36Further information on the empirical premia of our sample is provided in table 8.2
on page 58.
46 7. Valuation Model

28 0.6

24
net asset value
o 0 0 0 0 0 °

0.4
20 - - market price
..........
'" 16
.g 0.2
Q. 12

0.0
4 '"
E
~
o -0.2
5°1< 0 52 104 156 208 260

§Oo~
°Ee -5%10~V' ,tAA~
~ -10%
.g -15%
.~ -20%
m -25%+--~-~-~~-~-~--r-~-~--~--'
o 52 104 156 208 260

Figure 7.3: GT Global Eastern Europe (GTF)

take a look at the values of the premia across all closed-end funds in our
sample (see figure 7.4 on page 47) over the examined period of time, we can
confirm this view based on the whole sample. For the average premia we
observe mainly positive values for the first half of the sample period with a
peak of 13 percent at the end of the first year, which tends to continually fall
to discount values of about 14 percent in the fourth and fifth year. Looking
at the maximum positive and negative values, we notice a higher volatility
for the positive premia whereas the discounts oscillate in a narrow band of
minus 10 to 30 percent.
Moreover, for Sharpe and Sosin (1975) the fluctuations of the empir-
ical premia appear to be mean-reverting. For their sample, they find a
slight tendency for the premia to revert to a long-term mean discount value
of approximately seven percent. Further, in an examination of the rela-
tion between closed-end fund premia and returns, Pontiff (1995) finds no
economically motivated explanation for discounted closed-end funds having
higher expected returns than non-discounted closed-end funds, but rather
accounts a mean-reverting nature of the empirical premia for this effect. For
the market of closed-end funds in the UK, the study of Minio-Paluello (1998)
7.2. Economic Foundation 47

60%

50%

40%
Maximum Values
30%

20%

10%

0%

-10%

-20%

-30%

-40%
0 52 104 156 208 260

Figure 7.4: Time Series of the Empirical Premia

shows the empirical tendency of premia to revert to their mean. These re-
sults from the literature give rise to capturing this pricing phenomenon in
a stochastic framework of closed-end fund share valuation which we layout
in the next section.

7.2 Economic Foundation

The proposed model structure links the share prices of the closed-end funds
to the value of their underlying portfolio invested in the originating mar-
kets by incorporating a dynamically changing premium under a stochastic
framework. Before performing this linkage, we first specify the frictionless
economy underlying our continuous time valuation model.

Assumption 7.2.1 (Frictionless Economy) We assume a risk-free as-

set at a constant interest rate r that is available to all market participants
in the economy. The risky assets in the economy are traded on either of the
following markets: .
(i) the primary market, i. e. the local home market of the countries where
the closed-end fund portfolios are originally invested, and
48 7. Valuation Model

(ii) the secondary market for the shares of the closed-end funds, i. e. the
US host market.

In the case of international closed-end funds traded in the US the pri-

mary market could, for example, be Brazil. Thereby, the closed-end fund
shares and their underlying portfolio of foreign securities are priced in dif-
ferent market segments. In modeling the closed-end fund prices on the
secondary markets we choose the net asset value to be the first model fac-
tor. This choice is motivated by the fact that the value of the component
assets in the originating countries are fundamental to value closed-end funds
on the advanced secondary markets. From the literature we know that the
net asset value explains most of the return on closed-end fund share prices
but not all of it. 37

Assumption 7.2.2 (Net Asset Value) For the primary market we model
the price movements for the investment portfolio of the closed-end funds
measured by the reported net asset value as geometric Brownian motion38

dNAVt/NAVt = I-Ldt + uxdWx,t,

as in Ingersoll (1976).

Corollary 7.2.3 For the assumed price behavior of the net asset value we
conveniently work with the stochastic dynamics of the log-transformed vari-
able X t = In N A Vt given by

dXt = (I-L - ~u3c ) dt + uxdWx,t, (7.2)

where dXt is the continuous return on the net asset value portfolio.

In addition to the price risk modeled by the first factor we introduce

a time continuous premium 'lrt on the expected return of closed-end funds.
3 7 Forexample, Chen, Kan, and Miller (1993) find that 72.8 percent is explained by
the net asset value variance on a closed-end fund sample covering the period from 1965
to 1985.
38 A stochastic process corresponding to an exponential growth of the state variable
which is a widespread asset pricing assumption for modeling equities in financial eco-
nomics.
7.2 Economic Foundation 49

How this dynamically modeled premium is related to the empirical pre-

mium will be clarified further on. 39 Such a premium is intended to capture
the stochastic price behavior of the closed-end fund market prices that is
superimposed on the dynamics of the underlying asset values. The stochas-
tic specification of this second factor is motivated by the observed mean-
reverting dynamic fluctuations of the empirical premia.

Assumption 7.2.4 (Dynamic Premium) The dynamic premium is as-

sumed to follow a stochastic mean-reverting specification of the form

(7.3)
which is a type of Ornstein- Uhlenbeck process. In this special case the con-
tinuous premium 7rt is modeled with a volatility of (J1r and reverts with a
speed of /'i, to its long-run mean (). Further, the derived valuation model will
allow for a correlation of the two model factors X t and 7rt as the standardized
Brownian motions are set up as dWx,tdW1r,t = pdt.

Now, we link the price of the closed-end fund shares that are traded on
the secondary market both to the values of their underlying net asset value
and the dynamic premium. We model the market value of the closed-end
fund share prices, denoted by P (X, 7r, t; 'l/J M) with parameter vector 't/J M, as
a derivative security dependent on the values of the N A Vi and the premium
7rt. In analogy to the standard technique known from contingent claim
pricing40, we derive the following stochastic dynamics for P (X, 7r, t; 't/J M)
by applying Ito's formula
dP
(7.4)
P

with J-Lp

'f/x =
39 Specifically,
see the description on page 61 for this relationship.
40See, for example, Merton (1992).
50 7. Valuation Model

where the time indices are temporarily dropped for legibility.

Taking a closer view on the model economy with its three specified mar-
kets, we further need to ensure the consistent pricing of the different assets
across all markets. Therefore, we take the view of an US investor that com-
pares returns of alternative investments denominated in home currency. In
an equilibrium state of the modeled economy, investors are indifferent in
holding either of the available assets according to their risk preferences.
Using the standard finance condition for a market equilibrium41 , we spec-
ify the necessary expected yield above the risk-free rate as fL - r = ax)..x
including the market price of risk )..x. This ensures that there are no arbi-
trage possibilities on the primary market to gain risk-less profits. For the
secondary market, where the closed-end fund shares are traded, investors
require an expected return of fLp = r + TJx)..x + 7r + TJ7r)..7r including the
dynamic premium 7r and its market price of risk )..7r. 42 Interpreting the
latter equilibrium condition from a discounting perspective, we expect the
implicit discount rate fLp of the true terminal value to be high when the
premium is positive. Contrarily, closed-end funds with negative values for
the premium reveal investors' attitudes towards a low expected return. In
addition, by incorporating both risk premia for the modeled risks "Ix and
TJ7r our valuation formula will contain both the price risk and the premium
risk inherent, respectively, in changes of the NAV and varying values for the
continuous premium.

7.3 Pricing Closed-End Fund Shares

Based on the outlined economic foundations of the valuation model we now

derive the fair market price of closed-end fund shares. In a first step, we
use the underlying equilibrium conditions in conjunction with the stochastic
dynamics of the closed-end funds' share prices from equation (7.4). This
yields the fundamental partial differential equation underlying our valuation

41See, for example, Ingersoll (1987).

42The two market prices of risk, AX and A1T , are treated as constants in our model.
7.3 Pricing Closed-End Fund Shares 51

model

+ (r _ 12(JX2) ap(X,1r,t;t/J
ax
M) + ( (() _ ) _
K 11"
\) ap(X,1r,t;t/J M) + ap(X,1r,t;t/J
(J7r/\7r a7r at
M)

(7.5)

which describes how the closed-end funds' share price P varies for differ-
ent values of the state variables X and 11" as well as for changing trading
time t. Additionally, to fully reflect the dynamic behavior of the closed-end
funds' share price P (X, 11", t; '¢M) we further need to specify the appropri-
ate boundary condition for this security. To defer the relevant condition, we
exploit the economic behavior of the closed-end funds at the fourth stage
of their life cycle at time T.43 According to Brauer (1984) and Brickley
and Schallheim (1985) the market values of the closed-end funds equal their
net asset values with empirical regularity at the time of the funds liqui-
dation or open-ending. From this fact we obtain the boundary condition
P (X, 11", T; 'l/JM) = eXT for our valuation model.
The next step in deriving the price of closed-end fund shares in our
model is to solve the partial differential equation of equation (7.5). For the
solution to this equation we consider the exponential affine form 44

~/. ) = eA (t)X+B(t)7r+C(t) ,
P (X , 11" , t·,'I'M (7.6)

with model parameters '¢M and factor loadings A (t), B (t), and C (t). Using
the partial derivatives of the proposed solution

P .= ap(x,7r,t;t/JM) = AP
x· ax '

P7r .-
.-
ap(x,7r,t;t/J M) -
a7r -
BP,

PX7r .=
•
a 2p(x,7r,t;t/JM)
axa>.
= ABP' n._ ap(x,7r,t;t/JM) -
.rt·- at -
(A t X t + Bt1l"t + C)t P
43 According to the nomenclature of Lee, Shleifer, and Thaler (1991).
44This type of solution is known from interest rate modeling; see, for example, Duffie
and Kan (1996).
52 7. Valuation Model

for the differential equation (7.5) results in

(7.7)

where we already separated for the variables.

In order for equation (7.7) to hold for every values of X t and 7rt, we need
the expressions in square brackets to be zero. This leads us to a system
of ordinary differential equations in A (t), B (t), and C (t). For successively
solving this system we make use of the original economic boundary condition
that the market values of the closed-end funds equal their net asset values
at the fourth stage of their life cycle which translates into

A(T) = 1, and B(T) = C(T) = o.

Making use of the first two conditions we obtain the following solutions for
the functions A (t) and B (t):

A(t) = 1, and
B (t) -.!. (1 - e-K(T-t») .
'"
The last function of interest is C (t). Solving the integral

J
T

C(d( = C(T) - C(t)

t

in conjunction with the boundary condition C (T) = 0 yields:

C (t) = - [e - ~(77r.x7r + ~P(7X(77r ;;2] (T - t)

-

+-1 [e - -(77r/\7r
1 \ +1 (7~] (1 - e- K(T - t»)
-P(7x(77r - -

(72
'" '" '" ",2

+~ (1 - e- 2K (T-t») .
4",3

To further derive the market price of the closed-end fund shares, we now
use the calculated functions A (t), B (t), and C (t) along with the suggested
solution in equation (7.6).
7.3 Pricing Closed-End Fund Shares 53

Solution 7.3.1 (Closed-End Funds' Market Price) With the log-trans-

formation of the closed-end funds , market price Y (X, 7r, t; 'l/JM) = In P(X, 7r,
t; 'l/J M) we finally obtain the general valuation formula

for the log-market price of closed-end funds with a maturity ofT.

The derived valuation formula states that the log-market price yt should
basically be equal to the log-value X t of the underlying assets in that the
market prices carry the price risk with a factor loading of A (t) = 1. Second,
the closed-end funds' market prices are affected by changing values of the
dynamic premium 7rt. The realizations of the premia influence the market
prices by a negative loading B (t), i.e. closed-end funds with positive values
for the dynamic premium trade at a discount. Finally, the valuation formula
contains a time varying term C (t) which depends on the maturity of the
closed-end funds and the model parameters 'l/JM. In the following chapter
we motivate a possible implementation of the valuation model upon market
data of closed-end funds.
8 First Empirical Results

8.1 Sample Data

For the empirical implementation of our pricing model we choose to select
a broad market sample of emerging market closed-end funds. 45 The sample
examined in this study exists of closed-end equity funds for a complete
five year period from January 1993 to December 1997. All NYSE traded
closed-end funds are included, if their date of issue lies before January 1993
in order to avoid post-offering pricing effects. 46 The data is collected on a
weekly basis to yield a multiple time-series of 33 closed-end funds with 260
observations each. The market prices are obtained from the last trade on
Fridays, and the net asset values are calculated by the funds on Thursdays
balances. However, the net asset values are reported on Fridays after the
NYSE closes which makes trading on the basis of these prices impossible, but
are treated as being observed contemporaneously with the market prices. 47
Finally, the share prices and net asset values are adjusted for distributions
including income dividends and capital gains. In table 8.1 we report the
figures for the total returns and their correlation along with the net proceeds
and the date of the initial public offering (IPO) of the funds.
The returns on the different closed-end funds show a wide range ac-
cording to their core countries of investment. Funds invested in Brazil, for
example, realize about 15 percent on their market prices, whereas invest-
ment portfolios in Thailand show a 18 percent loss on average over the five
year period. The returns on the net asset values also show a wide range, but
45 1 greatly thank Lipper Analytical Services, and especially Donald Cassidy for kindly
providing the closed-end fund data.
46See, for example, Peavy (1990) and Weiss (1989).
47Thus, the empirical premia will include some undetermined amount of measurement
error.
56 8. First Empirical Results

Table 8.1: Characteristics of the Closed-End Fund Sample

Fund Sym- IPO Pro- P_Returnsb}C} NAV-Returns C}

Name bol Date ceedsa ) MEAN STD MEAN STD Corrd)
Argentina Fund AF 10/91 63.3 4.13 33.09 11.35 22.15 0.699
Asia Pacific Fund APB 04/87 79.0 -0.09 33.60 3.42 20.84 0.554
Brazil Fund BZF 03/88 138.4 14.44 38.55 18.33 37.32 0.700
Brazilian Equity Fund BZL 04/92 63.4 15.35 45.91 17.65 41.04 0.635
Chile Fund CH 09/89 64.0 8.06 27.11 12.59 17.52 0.674
China Fund CHN 07/92 1l0.7 1.09 32.33 3.87 21.47 0.508
Clemente Glob Growth CLM 06/87 54.9 10.05 16.62 10.24 11.17 0.500
Emerging Mexico Fund MEF 10/90 55.0 -3.30 39.69 1.98 32.96 0.677
Emerg MKTS Telecom ETF 06/92 115.3 10.03 27.77 13.09 16.05 0.605
First Israel Fund ISL 10/92 68.7 2.33 27.32 3.93 18.33 0.503
First Philippine Fund FPF 1l/89 98.9 -2.09 28.88 -5.27 22.40 0.578
Greater China Fund GCH 07/92 93.1 5.14 33.47 7.45 24.53 0.595
GT Global East Europe GTF 03/90 224.6 14.42 23.97 14.12 17.59 0.560
India Growth Fund IGF 08/88 55.0 -5.82 30.92 -2.09 24.48 0.477
Indonesia Fund IF 03/90 63.6 -13.36 37.74 -15.54 30.11 0.486
Jakarta Growth Fund JGF 04/90 54.9 -13.37 35.10 -14.74 29.07 0.517
Jardine Fleming China JFC 07/92 94.1 -3.78 32.44 -0.75 22.48 0.513
Korea Fund KF 08/84 54.9 -12.34 40.26 -13.14 32.71 0.581
Korean Investm Fund KIF 02/92 45.8 -17.49 37.12 -23.43 34.66 0.613
Latin Amer Equity FD LAQ 10/91 83.1 6.18 32.97 10.28 21.32 0.579
Latin Amer Investment LAM 07/90 54.6 7.19 31.94 9.84 20.88 0.612
Latin Amer Discovery LDF 06/92 80.2 14.83 35.38 12.04 28.47 0.602
M S Emerg MKTS MSF 10/91 148.1 3.37 29.07 7.33 16.59 0.495
Malaysia Fund MF 05/87 80.2 -9.37 31.69 -14.28 28.87 0.534
Mexico Equity&Income MXE 08/90 70.4 5.17 36.83 11.05 27.83 0.672
Mexico Fund MXF 06/81 112.1 -0.80 39.04 3.50 36.25 0.759
Roc Taiwan Fund ROC 05/89 54.5 6.06 34.98 8.97 22.64 0.605
Scudder New Asia Fund SAF 06/87 77.5 -1.46 29.98 1.92 19.13 0.572
Singapore Fund SGF 07/90 54.9 0.33 30.85 -0.33 18.51 0.471
Taiwan Fund TWN 12/86 23.9 7.39 35.87 10.12 25.83 0.493
Tempelt Emerg MKTS EMF 02/87 106.4 12.67 29.09 14.56 16.46 0.334
Taih Capital Fund TC 05/90 67.9 -16.77 35.26 -23.50 28.32 0.586
Taih Fund TTF 02/88 105.7 -18.71 35.28 -27.18 31.46 0.570
MEAN 82.3 0.89 33.03 2.04 24.83 0.571
STD 37.5 9.91 5.40 12.52 7.11 0.084
MAX 224.6 15.35 45.91 18.33 41.04 0.759
MIN 23.9 -18.71 16.62 -27.18 11.17 0.334

Notes:
a) Net proceeds are in million $.
b) P-returns denote returns on the market prices of the closed-end fund shares.
c) Returns are annualized log-returns, calculated on a weekly basis, and given in percent.
d) Denotes the correlation coefficient between the P- and NAV-returns.
8.1 Sample Data 57

are not as volatile. The standard deviation of the net asset value returns
averages on 25 percent and is about one-half lower than the 33 percent on
the market returns. 48 Looking at the correlation between the two returns,
we see a strong positive comovement with a correlation coefficient of 0.57
on average.
In table 8.2 we show the descriptive statistics for the empirical premia
of each closed-end fund. Across all closed-end funds in the sample, we find
an average discount of 3.25 percent which is in line with previous research
on emerging market closed-end funds. 49 The range of premia is given by
the spanning from a 20 percent discount to positive premia of 15 percent
on average. However, only eight out of the 33 closed-end funds trade at
positive premia on average. Especially, the four closed-end funds including
the symbols IGF, IF, KF and EMF show premia that seem to be persistently
on a level above zero over time. For these funds we observe positive values
for at least 75 percent of the sample.
To finally characterize the sample data, we further analyze the time-
series properties of the empirical premia. According to the described liter-
ature in section 7.1 the empirical premia show some evidence to behave in
a mean-reverting manner. Therefore, we model the premia P REMt ad hoc
by an Ornstein-Uhlenbeck process with a specification of the form

dPREMt = i.p (X - PREMt ) dt + WdWPREM,t, (8.1)

where the speed of mean-reversion is denoted by i.p, the long-run mean

given by X, and w stands for the volatility term. With this specification
the empirical premia follow a Gaussian distribution which can be estimated
using an exact maximum likelihood method for the discretized process.
The results of the empirical premia in our sample for the discretized
stochastic specification of equation (8.1) are shown in table 8.3. The param-
eter estimates result in average values for the coefficient of mean-reversion
of 3.9, for the long-run mean value of the empirical premia of -0.04 and
an annual standard deviation of 28 percent. The mean-reversion coefficient
shows a high fluctuation of values from 1.4 up to 9.8. The estimations
48This excess price volatility compared to the fundamental volatility is consistent with
the implications of the De Long, Shleifer, Summers, and Waldmann (1990) noise trader
model.
49See, for example, Dimson and Minio-Kozerski (1998).
58 8. First Empirical Results

Table 8.2: Descriptive Statistics on the Closed-End Fund Premia

Fund Empirical Premiaa )

Name MEAN STD MAX UQ MED LQ MIN SKO) KUO)
AF -0.36 9.11 20.23 6.95 1.38 -7.60 -21.38 -0.26 2.18
APB -0.28 13.57 28.51 11.26 2.14 -13.63 -23.30 0.02 1.71
BZF -5.73 10.55 16.24 2.37 -4.05 -16.46 -28.65 -0.21 1.95
BZL -5.77 12.00 26.68 3.69 -5.56 -16.99 -33.41 0.14 2.30
CH -8.63 7.79 12.35 -4.58 -9.59 -14.20 -24.29 0.49 2.55
CHN -1.30 13.57 30.51 9.45 0.97 -11.84 -29.51 -0.25 2.02
CLM -20.46 5.96 -6.01 -15.60 -21.41 -25.64 -30.66 0.32 1.98
MEF -5.20 14.81 34.15 2.46 -4.40 -19.11 -28.91 0.59 2.72
ETF -7.13 12.07 23.31 2.63 -8.36 -18.43 -24.50 0.50 2.10
ISL -6.79 12.67 23.39 1.41 -10.97 -16.23 -28.65 0.74 2.44
FPF -18.14 6.38 8.12 -15.76 -19.26 -22.14 -30.56 1.46 5.94
GCH -7.53 11.38 17.35 2.13 -7.13 -18.35 -30.68 0.01 1.86
GTF -12.22 6.06 5.37 -8.63 -13.50 -16.65 -22.48 0.74 2.83
IGF 8.10 10.49 43.75 14.49 7.21 0.31 -18.23 0.37 3.44
IF 14.57 12.41 39.97 23.31 16.82 9.95 -13.80 -0.74 2.79
JGF 2.77 9.44 29.96 9.84 3.69 -3.23 -16.27 -0.16 2.43
JFC -6.36 12.77 27.51 4.76 -7.49 -18.34 -35.00 0.12 1.94
KF 10.44 9.79 34.56 16.97 8.81 3.87 -20.32 0.29 2.82
KIF 1.81 10.75 37.02 9.43 -1.29 -6.20 -18.02 0.73 2.89
LAQ -8.09 10.06 20.11 -1.44 -8.53 -16.72 -26.49 0.38 2.38
LAM -8.32 10.25 20.39 -1.43 -9.37 -17.06 -27.35 0.56 2.68
LDF -11.58 7.03 9.59 -7.58 -12.14 -16.26 -27.69 0.35 2.95
MSF -3.68 7.15 17.47 1.25 -3.08 -8.81 -20.76 -0.07 2.63
MF -3.13 9.90 40.98 0.81 -5.81 -10.05 -17.71 1.78 6.80
MXE -4.66 14.86 36.61 5.76 -3.44 -19.76 -28.81 0.36 2.33
MXF -8.98 10.82 20.99 -3.07 -8.18 -18.34 -31.92 0.40 2.82
ROC -2.36 11.74 25.83 6.65 -1.51 -11.29 -30.15 -0.22 2.29
SAF -3.60 10.56 20.51 4.19 -2.45 -13.06 -22.50 0.03 1.98
SGF 2.42 7.60 27.83 7.76 2.63 -3.54 -13.73 0.26 3.06
TWN 1.33 16.24 48.65 13.04 4.81 -11.79 -34.50 -0.15 2.39
EMF 13.52 6.78 35.73 17.96 13.50 7.92 -1.32 0.35 2.60
TC -0.49 15.54 51.08 1.74 -5.64 -9.65 -21.17 1.56 4.88
TTF -1.41 16.53 53.44 2.96 -7.25 -11.96 -26.99 1.42 4.39
MEAN -3.25 10.81 26.73 3.79 -3.59 -11.86 -24.54 0.36 2.82
STD 7.74 2.97 13.42 8.66 8.32 8.34 7.16 0.56 1.13
MAX 14.57 16.53 53.44 23.31 16.82 9.95 -1.32 1.78 6.80
MIN -20.46 5.96 -6.01 -15.76 -21.41 -25.64 -35.00 -0.74 1.71

Notes:
a) The premia are given in percent, measured by In(PtfNAVd. UQ and LQ denote
the upper and lower quartile, respectively. The median is abbreviated by MED.
b) We also report the skewness (SK) and kurtosis (KU) of the premia distribution.
8.1 Sample Data 59

Table 8.3: Parameter Estimates of the Empirical Premia Process

Fund <p X W neg.
Name EST STD EST STD EST STD LogL
AF 3.398*** 1.233 -0.025 0.033 0.240*** 0.011 2.007
APB 2.020** 0.956 -0.020 0.063 0.283*** 0.013 1.831
BZF 3.785*** 1.326 -0.067** 0.036 0.299*** 0.014 1.792
BZL 4.996*** 1.503 -0.062** 0.034 0.383*** 0.018 1.559
CH 3.556*** 1.214 -0.099*** 0.026 0.203*** 0.009 2.175
CHN 2.072** 0.964 -0.026 0.062 0.285*** 0.013 1.823
CLM 3.128*** 1.153 -0.205*** 0.021 0.149*** 0.007 2.483
MEF 1.935** 0.930 -0.079 0.071 0.301*** 0.013 1.770
ETF 1.632** 0.841 -0.090* 0.062 0.223*** 0.010 2.067
ISL 1.742** 0.873 -0.077 0.062 0.242*** 0.011 1.984
FPF 7.578*** 1.913 -0.177*** 0.015 0.251*** 0.012 2.002
GCH 2.900*** 1.121 -0.083** 0.043 0.276*** 0.012 1.864
GTF 5.927*** 1.633 -0.122*** 0.016 0.208*** 0.010 2.173
IGF 3.416*** 1.311 0.071** 0.039 0.294*** 0.013 1.808
IF 4.093*** 1.342 0.151*** 0.039 0.357*** 0.016 1.620
JGF 5.930*** 1.660 0.030 0.025 0.329*** 0.015 1.719
JFC 2.447*** 1.035 -0.076* 0.053 0.288*** 0.013 1.819
KF 6.359*** 1.726 0.105*** 0.025 0.351*** 0.016 1.658
KIF 3.532*** 1.421 0.034 0.042 0.323*** 0.015 1.715
LAQ 3.531*** 1.270 -0.092*** 0.035 0.274*** 0.012 1.877
LAM 2.976*** 1.158 -0.092*** 0.039 0.256*** 0.012 1.938
LDF 9.756*** 2.135 -0.113*** 0.014 0.304*** 0.015 1.831
MSF 6.309*** 1.759 -0.043*** 0.019 0.261*** 0.012 1.951
MF 3.588*** 1.434 -0.018 0.038 0.299*** 0.014 1.791
MXE 1.573** 0.850 -0.084 0.081 0.276*** 0.012 1.851
MXF 2.840*** 1.137 -0.105*** 0.042 0.266*** 0.012 1.900
ROC 2.770*** 1.110 -0.034 0.046 0.282*** 0.013 1.841
SAF 2.630*** 1.090 -0.049 0.043 0.249*** 0.011 1.962
SGF 7.092*** 1.797 0.025* 0.018 0.284*** 0.013 1.876
TWN 1.857** 0.916 -0.001 0.079 0.325*** 0.015 1.692
EMF 9.436*** 2.143 0.133*** 0.014 0.295*** 0.014 1.859
TC 1.570** 0.877 0.038 0.088 0.298*** 0.013 1.776
TTF 1.408** 0.868 0.046 0.106 0.313*** 0.014 1.724
MEAN 3.872 1.294 -0.037 0.043 0.281 0.013 1.871
STD 2.245 0.374 0.082 0.023 0.046 0.002 0.178
MAX 9.756 2.143 0.151 0.106 0.383 0.018 2.483
MIN 1.408 0.841 -0.205 0.014 0.149 0.007 1.559

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.
60 8. First Empirical Results

for the long-run mean exhibit positive as well as negative values related to
whether the closed-end fund trades mainly with a premium or a discount
over the sample period. The range is from plus 15 percent to discounts of 21
percent which is close to the descriptive results for the mean values of table
8.2. When comparing the estimation results with the three closed-end funds
shown in figures 7.1, 7.3 and 7.2, we can confirm the high mean-reversion
tendency of 5.9 for the GT Global Eastern Europe (GTF). We can also
infer the high volatility for the Indonesia Fund (IF) from the figure. The
long-run means for the closed-end funds IF and GTF are estimated in line
with values of plus 15 percent and minus 12 percent.

8.2 Implemented Model

In order to implement the theoretical valuation model for closed-end fund
shares on capital market data we need to choose an appropriate empirical
model which comes close to the properties of the general model of equation
(7.8). Examining the available data on closed-end funds we notice that there
exist virtually no closed-end funds that feature automatic windup dates. 50
Therefore, we deal with infinite time horizons compared to an explicit expi-
ration date T as generally given in applications to price contingent claims.
Thus, for the purpose of the empirical implementation of our valuation
model we need to ensure the boundedness of the price stated in equation
(7.8) for the limit of T reaching infinity. Examining the equation for the
limits T --4 00 yields:

Thus, we actually reach a steady state except for the last expression
which remains from the first term of the function C (t). Therefore, we
constrain our model to the case where the long-run mean of the continuous

50There are very rare exceptions which show this feature and no closed-end funds
included in our sample.
8.3. State Space Form 61

premium is given by

(8.2)

which reduces our parameter space by one parameter. Using equation (8.2)
we obtain a functional mapping of the two state variables X t and 7rt to the
value of the observable asset Y (X, 7r, t; 1/JM ). The pricing relationship for
the implementation of the closed-end fund market prices according to the
valuation model is given by

(8.3)

where P (X, 7r, t; 1/JM) depends on the model parameters 1/JM = {O'x, /'i" 0'1r, p,
>'1r}. In further examining the valuation formula of equation (8.3) we are
finally able to show how the second factor we choose in the valuation model
is related to the empirical premium as defined in equation (7.1). By rewrit-
ing the valuation formula in terms of logarithms we obtain the connecting
formula
1
P REMt = --7rt
/'i,
+ const (8.4)

for the two premia, with a constant term denoted by const = f (1/J M ). From
this linear relationship between the empirical and the modeled continuous
premia we expect to observe positive dynamic premia on closed-end funds
with corresponding empirical discounts, and vice versa, in the empirical
analysis.

8.3 State Space Form

In this section we present the empirical adaptation of the theoretical pricing
model as stated in equation (8.3). In order to empirically estimate the
model parameters, we present the dynamic model in a state space form. 51
Therefore, we first specify the measurement and the transition equation,
to then be able to use a Kalman filter algorithm to perform a maximum
likelihood estimation of the parameters and to obtain a time-series of the
51 For more details on state space modeling compare chapter 3 and see, for example,
Aoki (1990).
62 8. First Empirical Results

continuous premia. For the state variables we choose et

which are assumed to follow the transition equation

(8.5)
for a time interval of 6..t. In order to specify the corresponding state space
form to equations (7.2) and (7.3) with the exact discrete time equivalents,
i.e. the discrete approximation yields exactly the same distributions as in
the continuous case, we need to match the following moments. 52

Remark 8.3.1 (Moments of the State Variables) For the first factor
el,t = X t we use equation {7.2} which can be integrated over the time period
6..t = [t - 6..t, t] to yield the almost explicit solution

J
t

X t = Xt-t:J.t + (JL - ~O"~ ) 6..t + O"x dWx,t.

t-t:J.t
Based on equation {7.3} for the second state variable, ,t = 7rt - (J, we e2
consider the differential of the function F (7rtl t) = e"'t (7rt - (J) by applying
Ito's formula

which we also integrate to find

J
t

(7rt - ()) = e-",t:J.t (7r t-t:J.t - ()) + e-"'tO"7r e"'(dW7r ,(

t-t:J.t
besides solving for the Ito integral. Since the remaining integrals are mar-
tingales we are able to calculate the conditional moments to be5 3

(8.7)

For the variance term we use an Euler approximation.

5 2 See,
for example, Bergstrom (1984).
53Using Ito's isometry and the fact that the expectation of an Ito integral is zero. See,
for example, Oksendal (1995, ch. III).
8.3 State Space Form 63

Thereby, the appropriate matrix and moment specifications of the tran-

sition equation (8.5) are given by

E ['1Jtl = [0,0]' ,

using a discretisation length of f:lt = 1/52 for weekly data. 54

With the two observable prices on closed-end funds, the net asset values
and their market values defined by Zt = [Xt, Yt]', we obtain the measurement
equation

(8.8)

Equation (8.8) serves as a link between the latent state variables et

and the observable net asset values N A lit and the market prices Pt for
the shares. These prices are assumed to be only measurable with a noise
et = [c:x,t, c:y,tl'.55 For the error terms we further assume E [etl = 0, and
E [ete~l = H·O t ,s.56 In order to obtain our empirical model of equation (8.3)
developed in the previous section, we specify the distinct coefficients

Combining the transition and the measurement equations, the full state
space model depends on the time homogeneous parameter vector tPSSM =
{K;, A7r , 0"7r' hy , /-l, o"x, hx, p}. These parameters are estimated in a statistical
inference using a recursion algorithm based on linear Kalman filtering as
described in chapter 4. With the specified state space form of equations
(8.5) and (8.8), we are able toperform a maximum likelihood estimation of

54The function Oi,j stands for the Kronecker symbol which takes value 1 if i = j, and
o else.
55The small errors et are introduced to capture imperfections in observed price quota-
tions.
56 Further, we assume independent error terms et and 'TJt.
64 8. First Empirical Results

the parameters

ibSSM = argmaxL (ZT' ZT-l, ... ,Zl; '¢SSM)· (8.9)

..p SSM E'iI!

Additional to the parameter estimation, the Kalman filter enables us to

extract a time-series of the unobservable state variables et
from the data.
In the following section we present a closed-end fund analysis based on the
examined sample of closed-end fund data.

8.4 Closed-End Fund Analysis

In this empirical analysis we examine the characteristics of our pricing model
by estimating its model parameters and the resulting time-series for the dy-
namic premia. The inference is based on the maximum likelihood estimator
of equation (8.9) using the defined Kalman filter setup of section 8.3.
The results on the estimated values of the parameters contained in '¢ SSM
are shown in table 8.4 and are continued in table 8.5. We first analyze
the parameters corresponding to the second state variable, i.e. the dynamic
premium'Trt. The coefficient of mean-reversion on the continuous premia is
estimated by 1.6 on average across all funds. Interpreting these parameter
estimates of /'i, as half-life of the premia57 , we infer a range of 0.1 to 1.6
years for the different closed-end funds with a mean value of 0.4 years.
These numbers indicate high fluctuations of the dynamic premia around
their long-run means B. This fact is emphasized in that closed-end funds
with high /'i, values also show a high standard deviation of the premium. For
the volatility on the dynamic premia (In we average around 0.3 with outliers
that go up to 1.0. Further, the estimated values for the market price of risk
An show no distinct pattern in that it varies from -3 to +5 with a mean
close to zero of 0.3. Finally, the volatility of the error terms hy is estimated
significantly around the value of 1.8% on average.
The outcome on the remaining parameter estimates are reported in table
8.5. First, we show the results on the drift coefficient J-l for which we obtain
an average value of 5 percent with a standard deviation of 12 %; the esti-
mates, however, are not statistically significant in general. The estimates of
57The half-life of the dynamic premium is given by In(2)/K. This can be deferred from
equation (8.6) solving for T = t - s.
8.4 Closed-End Fund Analysis 65

Table 8.4: Parameter Estimates of'l/JssM

Fund K, A11" (711" hy
Name EST STD EST STD EST STD EST STD
AF 0.976*** 0.011 0.197 0.568 0.143*** 0.001 0.018*** 0.000
APB 0.773*** 0.000 0.024 0.144 0.125*** 0.004 0.022*** 0.001
BZF 0.810*** 0.003 0.519 0.340 0.113*** 0.001 0.019*** 0.000
BZL 1.288*** 0.005 0.499 0.348 0.248*** 0.000 0.027*** 0.000
CH 3.091 *** 0.000 1.919*** 0.519 0.453*** 0.001 0.014*** 0.000
CHN 1.118*** 0.119 0.375 0.393 0.219*** 0.001 0.017*** 0.000
CLM 1.017*** 0.000 2.687*** 0.703 0.078*** 0.017 0.011*** 0.000
MEF 0.891 *** 0.001 0.247 0.330 0.182*** 0.000 0.021*** 0.000
ETF 0.423*** 0.002 0.230 0.237 0.049*** 0.000 0.019*** 0.000
ISL 1.337*** 0.001 0.739** 0.356 0.247*** 0.000 0.015*** 0.000
FPF 4.630*** 0.013 5.045*** 0.437 0.778*** 0.003 0.016*** 0.000
GCH 1.194*** 0.004 0.720** 0.366 0.194*** 0.000 0.020*** 0.000
GTF 2.119*** 0.003 2.439*** 0.384 0.242*** 0.001 0.014*** 0.000
IGF 1.548*** 0.000 -0.185 0.397 0.391*** 0.001 0.010*** 0.000
IF 2.494*** 0.006 -1.369*** 0.399 0.637*** 0.004 0.020*** 0.000
JGF 2.034*** 0.002 -0.401 0.416 0.379*** 0.002 0.021*** 0.000
JFC 1.263*** 0.005 0.508 0.407 0.241*** 0.001 0.019*** 0.000
KF 1.220*** 0.000 -0.886*** 0.356 0.383*** 0.007 0.010*** 0.000
KIF 1.840*** 0.006 -0.264 0.346 0.456*** 0.000 0.013*** 0.000
LAQ 1.047*** 0.002 -0.082 0.358 0.165*** 0.000 0.021*** 0.000
LAM 1.056*** 0.058 -0.066 0.292 0.224*** 0.017 0.010*** 0.002
LDF 0.813*** 0.002 0.089 0.358 0.162*** 0.001 0.020*** 0.000
MSF 1.130*** 0.003 -0.076 0.344 0.193*** 0.000 0.019*** 0.000
MF 2.070*** 0.001 -0.320 0.370 0.459*** 0.004 0.018*** 0.000
MXE 0.768*** 0.002 0.060 0.251 0.145*** 0.002 0.019*** 0.000
MXF 1.037*** 0.009 0.498** 0.224 0.175*** 0.001 0.042*** 0.000
ROC 1.785*** 0.002 -0.033 0.170 0.362*** 0.001 0.019*** 0.000
SAF 1.265*** 0.002 0.214 0.478 0.200*** 0.007 0.018*** 0.000
SGF 4.513*** 0.000 -0.575 0.451 1.010*** 0.004 0.015*** 0.000
TWN 0.822*** 0.000 -0.021 0.036 0.189*** 0.001 0.019*** 0.000
EMF 4.298*** 0.021 -3.177*** 0.436 0.802*** 0.002 0.020*** 0.000
TC 0.845*** 0.002 -0.230 0.345 0.157*** 0.000 0.021*** 0.000
TTF 0.911 *** 0.000 -0.048 0.281 0.221*** 0.002 0.018*** 0.000
MEAN 1.589 0.009 0.281 0.359 0.304 0.003 0.D18 0.000
STD 1.085 0.022 1.319 0.122 0.223 0.004 0.006 0.000
MAX 4.630 0.119 5.045 0.703 1.010 0.017 0.042 0.002
MIN 0.423 0.000 -3.177 0.036 0.049 0.000 0.010 0.000

Note:
Statistically significant parameter estimates at the 1%-, 5%- and 10%-levels
are denoted by ***, **, and *, respectively.
66 8. First Empirical Results

Table 8.5: Parameter Estimates of 'l/JSSM (continued)

Fund J.L Ux hx p neg.

Name EST STD EST STD EST STD EST STD LogL
AF 0.148 0.099 0.220*** 0.004 0.002*** 0.000 -0.119*** 0.000 1061.88
APB 0.056 0.091 0.207*** 0.000 0.000*** 0.000 0.024*** 0.000 1039.75
BZF 0.229 0.145 0.328*** 0.001 0.018*** 0.000 0.153*** 0.000 898.02
BZL 0.237 0.160 0.370*** 0.002 0.017*** 0.000 0.278*** 0.000 808.01
CH 0.143* 0.078 0.175*** 0.000 0.000*** 0.000 -0.049*** 0.000 1161.79
CHN 0.058 0.084 0.194*** 0.000 0.009*** 0.000 0.008*** 0.001 1026.79
CLM 0.108** 0.046 0.100*** 0.041 0.005*** 0.000 -0.011*** 0.002 1378.02
MEF 0.065 0.141 0.327*** 0.000 0.002*** 0.000 0.304*** 0.000 901.30
ETF 0.149** 0.072 0.160*** 0.000 0.000*** 0.000 -0.138*** 0.001 1167.38
ISL 0.062 0.083 0.180*** 0.000 0.003*** 0.000 0.153*** 0.000 1106.93
FPF 0.000 0.052 0.223*** 0.001 0.000*** 0.000 0.255*** 0.001 1062.24
GCH 0.102 0.098 0.231*** 0.001 0.008*** 0.000 0.079*** 0.000 1004.23
GTF 0.152** 0.069 0.153*** 0.000 0.009*** 0.000 -0.113*** 0.000 1176.75
IGF 0.015 0.048 0.228*** 0.001 0.009*** 0.000 0.237*** 0.000 989.07
IF -0.113 0.121 0.282*** 0.001 0.010*** 0.000 0.304*** 0.002 892.83
JGF -0.118 0.116 0.267*** 0.003 0.012*** 0.000 0.250*** 0.001 935.04
JFC 0.013 0.594 0.212*** 0.000 0.008*** 0.000 0.098*** 0.000 1012.46
KF -0.133 0.145 0.325*** 0.002 0.003*** 0.000 0.301*** 0.002 865.20
KIF -0.187 0.134 0.307*** 0.001 0.015*** 0.000 0.236*** 0.000 884.25
LAQ 0.149* 0.089 0.208*** 0.000 0.007*** 0.000 -0.252*** 0.000 1037.28
LAM 0.109 0.087 0.199*** 0.010 0.007*** 0.002 -0.160*** 0.000 1048.51
LDF 0.145 0.120 0.282*** 0.001 0.004*** 0.000 0.284*** 0.001 946.94
MSF 0.071 0.074 0.163*** 0.000 0.000*** 0.000 0.048*** 0.000 1115.73
MF -0.155 0.120 0.288*** 0.000 0.000*** 0.000 0.501*** 0.004 959.44
MXE 0.214* 0.126 0.275*** 0.001 0.001*** 0.000 0.104*** 0.000 963.84
MXF 0.030 0.159 0.355*** 0.003 0.004*** 0.000 0.246*** 0.000 869.54
ROC 0.067 0.079 0.226*** 0.000 0.004*** 0.000 -0.115*** 0.000 1007.59
SAF 0.035 0.104 0.191*** 0.000 0.001*** 0.000 0.087*** 0.003 1090.93
SGF 0.015 0.078 0.177*** 0.000 0.003*** 0.000 0.116*** 0.000 1069.41
TWN 0.133 0.105 0.233*** 0.001 0.011*** 0.000 0.158*** 0.000 946.06
EMF 0.154** 0.072 0.164*** 0.002 0.000*** 0.000 0.289*** 0.000 1106.40
TC -0.176 0.125 0.275*** 0.001 0.006*** 0.000 0.271*** 0.000 951.12
TTF -0.137 0.131 0.311 *** 0.001 0.001*** 0.000 0.416*** 0.000 911.17
MEAN 0.050 0.116 0.237 0.002 0.005 0.000 0.129 0.001 1012.00
STD 0.120 0.091 0.066 0.007 0.005 0.000 0.180 0.001 115.54
MAX 0.237 0.594 0.370 0.041 0.Ql8 0.002 0.501 0.004 1378.02
MIN -0.187 0.046 0.100 0.000 0.000 0.000 -0.252 0.000 808.01

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels are denoted
by ***, **, and *, respectively.
8.4 Closed-End Fund Analysis 67

Table 8.6: Values of e and Statistic of Dynamic Premia 1ft

Fund 1f~)
Name ea ) MEAN STD MAX UQ MED LQ MIN SK c) KU c)
AF 4.335 -0.20 8.60 19.30 6.83 -2.33 -7.17 -15.97 0.34 2.15
APB 1.624 -0.45 10.24 16.68 9.89 -1.77 -9.19 -20.96 0.Dl 1.67
BZF 7.489 4.15 8.13 20.71 12.91 2.50 -2.65 -10.40 0.33 1.86
BZL 9.468 6.51 14.58 37.56 21.50 5.53 -5.48 -26.77 -0.05 2.05
CH 29.341 26.13 23.28 70.94 43.26 29.07 12.91 -31.60 -0.48 2.42
CHN 9.233 0.54 14.87 30.41 12.41 -1.49 -11.51 -31.02 0.27 1.95
CLM 20.841 20.67 5.85 30.11 25.83 21.61 15.81 7.51 -0.33 1.91
MEF 5.114 3.56 12.98 23.45 15.96 2.71 -2.94 -29.03 -0.55 2.64
ETF 3.601 2.67 5.02 9.56 7.49 2.95 -1.12 -8.84 -0.47 2.03
ISL 14.874 8.26 16.71 36.87 20.94 14.03 -1.66 -30.75 -0.74 2.42
FPF 85.265 83.45 27.75 135.50 101.18 87.94 75.33 -26.02 -1.58 6.15
GCH 12.752 8.35 13.15 34.58 21.72 8.44 -3.31 -17.61 0.02 1.78
GTF 28.684 25.65 12.06 45.39 34.32 28.93 18.88 -7.04 -0.82 2.83
IGF -2.854 -14.10 16.09 25.88 -2.50 -12.60 -23.52 -68.83 -0.37 3.40
IF -33.910 -37.88 29.96 30.00 -28.14 -43.34 -60.28 -94.02 0.81 2.82
JGF -6.969 -6.49 18.08 29.98 5.82 -9.40 -20.19 -46.33 0.32 2.30
JFC 11.124 7.13 15.75 39.80 22.58 8.32 -6.70 -31.14 -0.10 1.85
KF -25.960 -15.21 11.84 21.17 -7.27 -13.09 -23.52 -44.24 -0.30 2.77
KIF -5.262 -4.84 19.24 29.31 10.20 1.02 -18.54 -65.34 -0.71 2.69
LAQ 0.779 7.72 10.18 27.64 16.38 7.68 1.25 -16.25 -0.32 2.26
LAM 1.522 7.64 10.73 27.50 16.69 8.53 0.47 -22.35 -0.55 2.68
LDF 2.164 8.40 5.45 19.55 11.82 8.91 4.76 -6.66 -0.31 2.82
MSF 0.029 3.37 7.69 21.03 8.86 2.72 -1.70 -16.41 0.12 2.51
MF -7.817 5.18 19.95 32.59 18.63 10.60 -2.25 -76.84 -1.81 6.79
MXE 2.364 2.65 11.25 20.36 14.41 1.73 -5.20 -26.62 -0.34 2.26
MXF 8.367 8.49 10.60 29.24 18.03 7.31 2.88 -19.34 -0.36 2.72
ROC 1.917 3.12 20.42 51.26 18.29 1.42 -12.36 -41.07 0.26 2.24
SAF 4.356 3.90 12.96 25.66 15.91 2.45 -5.56 -23.49 0.01 1.94
SGF -10.817 -12.15 33.02 56.99 13.36 -14.12 -34.29 -113.76 -0.22 3.01
TWN 1.319 -2.42 13.14 25.98 8.09 -5.24 -12.22 -35.95 0.18 2.31
EMF -58.447 -59.01 26.67 -8.21 -35.08 -58.70 -77.13 -139.25 -0.30 2.36
TC -3.932 -0.42 12.90 16.08 6.96 3.68 -1.98 -42.19 -1.58 4.81
TTF -1.375 -0.17 14.96 22.96 9.15 5.37 -4.14 -49.40 -1.43 4.38
MEAN 3.310 2.85 14.97 32.00 14.44 3.37 -6.74 -37.21 -0.34 2.75
STD 21.975 21.38 6.97 23.24 21.33 22.58 23.95 31.53 0.60 1.17
MAX 85.265 83.45 33.02 135.50 101.18 87.94 75.33 7.51 0.81 6.79
MIN -58.447 -59.01 5.02 -8.21 -35.08 -58.70 -77.13 -139.25 -1.81 1.67

Notes:
a) The long-run means of the dynamic premia (in percent) are given by equation (8.2).
b) In the statistics for the dynamic premia (in percent), UQ and LQ denote the upper
and lower quartile respectively; the median is abbreviated by MED.
c) For each closed-end fund we also report the skewness (SK) and kurtosis (KU) of the
dynamic premium distribution.
68 8. First Empirical Results

the net asset value volatility ax closely resemble the values obtained empir-
ically from the data (compare table 8.1). The differences take a maximum
of 4.6 percentage points with an average of 1.1%. Further, the estimated
model error h y of 0.005 across all funds translates to average pricing devia-
tions of 0.54% with the observable prices on the net asset values. 58 Lastly,
the values for the correlation coefficient p primarily show positive values up
to +0.5 with a mean of +0.13, i.e. changes in the net asset value and the
dynamic premium are related.

160%

120% Maximum Values

/
80%

40%

-40%

-80%

-120%

o 52 104 156 208 260

Figure 8.1: Time Series of (Inverted) Dynamic Premia

Further, the results on the parameter estimates of the volatility terms

a 11" and a x of the used state variables show that the risk of altering premia
with an average of 30.4% is about 1.3 times higher than the risk involved
with the net asset value showing a value of 23.7%. This, however, is a result
directly related to our valuation model; the empirical standard deviations
on the closed-end fund sample show a mean of the empirical premia of only
10.8% which is lower than both, the empirical as well as the estimated net
asset value volatility (compare tables 8.1 and 8.2).

58S ee, for example, Aitchison and Brown (1957, ch. 2.2) for the log-normal transfor-
mation.
8.4 Closed-End Fund Analysis 69

40%
20%
.~
E 0%
e!
a. .~,~yv~~~
-20%
-40% ROC Taiwan Fund (ROC) I ',' ••••••' / .....,'.,

80%
60%
ttl
'E 40%
e! 20%
a.
0%
-20%

10%

'E -10%
ttl
0%
;.;;;t'''f.
e! -20%
a. -30% .. / ... /./' :~
:····:······/":.·· ....t···· .. ·. . . . . :..... ··..... r·:· ~" . :'...:. . :\. . ::..":.;. .: . ./
-40% GT Global Eastern Europe (GTF) '.' .......... :.. . .
-50%
o 52 104 156 208 260
- - empirical premium ....... (inverted) dynamic premium

Figure 8.2: Comparison of Empirical and Dynamic Premia

Taking a final look at the significance levels of the estimated param-

eters, we find that they show highly significant values for the parameters
throughout the various funds, except for the estimates of the drift param-
eters '\71" and J1 which is a typical phenomenon in financial econometrics.
However, the estimates for the market price of risk >'71" show significant val-
ues for selected funds, such as CLM, FPF, KF, and EMF which trade at
high discounts and premia (compare table 8.2).
The results for the time-series of the dynamic premia for the differ-
ent closed-end funds are reported in table 8.6 along with their theoretical
long-run mean values. The sample means (MEAN) take values close to the
long-run means e. The differences are below eight percent, except for the
funds CHN, IGF, KF, and MF, where the MF fund shows the maximum
difference of 13%. The means themselves exhibit a wide range of -53 to +83
percent. But only 13 out of the 33 closed-end funds show a higher empirical
than theoretical mean value. Examining the sign of the dynamic premia
indicated by our theoretical result of equation (8.4), we find a negative re-
lation between the signs of the dynamic and the empirical premia. Our
observations coincide with this theoretical implication except for the minor
70 8. First Empirical Results

deviations of the AF, APB, TC, and TTF funds. However, the sign needs
to be put into perspective when incorporating the reported standard devia-
tions. Looking at the realized dynamic premia, we find high fluctuations as
indicated by the estimated values for the speed of mean-reversion and the
volatility from above. With 135.5 percent FPF marks the maximum and
EMF shows the minimum of minus 139.3 percent.
In figure 8.1 we graph the evolution of the dynamic premia with inverted
sign over the sample period, i.e. we visualize the numerical results of table
8.6. The figure contains the maximum and minimum values across all funds,
along with the mean values and the 90 percent interval. In general we see a
higher fluctuation in the positive premia from the maximum values as well
as the confidence interval. The negative premia seem to be estimated in a
much closer range. Overall, the visual impression resembles the results for
the empirical premia as shown in figure 7.4 on page 47 except for a scaling
factor. This fact also holds when we look at the three selected funds from
section 7.1 which we graph in figure 8.2. In each of the graphs we plot the
time-series of the empirical as well as the (inverted) dynamic premia for the
three closed-end funds ROC, IF, and GTF. We see that for the funds IF
and GTF the inverted dynamic premia emphasizes the positive and negative
values.
9 Implications for Investment
Strategies

In this chapter we further examine potential implications of the pricing

model for investment strategies. Based on the empirical findings presented
in the previous chapter we infer insights into two suggested applications
relevant for closed-end fund investors. First, we address the investors' fun-
damental question of the forecasting power of the pricing model, i.e. we
investigate the ability of the pricing model to predict the market prices of
the closed-end funds based on current information. The second question we
raise is to test on how much information content lies in the estimated values
of the closed-end fund premia. For this purpose we study trading strategies
that exploit the observable differences between the dynamic premia and
their long-run equilibria.

9.1 Testing the Forecasting Power

9.1.1 Setup of Forecasting Study

The study to test the forecasting quality of the pricing model is setup as
follows and is illustrated in figure 9.1. From the available five years of the
closed-end fund sample we choose the first four years, which contain 208
weekly observations, to estimate the parameter vector 'l/J SSM for the vari-
ous funds. The estimation results are close to the reported parameters for
the whole sample period (see tables 8.4 and 8.5) and thus are not reported
separately. The last 52 weeks of the sample are used to test whether fore-
casts up to four weeks beginning at the X-th week can generate meaningful
price predictions of the closed-end fund market shares in an out of sample
72 9. Implications for Investment Strategies

setting. In order to get current values for the state variables at initiation of
the forecast we filter the data based on the estimated four year parameters
for the period of the first week in the sample to the X-th week.

Period of Filtering Forecasting

...
I-----~ - - - -+--t---I---~--t---Avi
4 Years X 1W 2W 3W 4W
01/1993 01/1997 1211997
...
Parameter Estimation 1 Year

Figure 9.1: Setup of Forecasting Study

The finite sample forecasts are obtained from the same state space form
as used with the Kalman filter. The expected values for the s-period-ahead
predictions of the state variables and the observable prices are the condi-
tional expectations

et+s.~tlt = <I>se*, and Yt+s.~* = a + Bet+s.~*,

which can be deferred from the law of iterated projections. 59
Further, we increment the initial week successively by one week starting
from the 208th week and ending with week 256. This forecasting procedure
creates 49 sets of four week forecasts on the various closed-end fund market
prices.
In order to evaluate the quality of the forecast results we additionally
implement a reference model. For this purpose we choose a most general
Gaussian ARMA(I, I)-model for log-share prices yt given by

yt - m = a (yt-~t - m) + Et + bEt_~t
with b..t = 1/52. This process specification of the closed-end funds market
prices translates with the matrix specifications

c=o, <I>=[~ ~], "It = . 8t ,s,

a=m, B = [1,b], and
59S ee, for example, Harvey (1989).
9.1 Testing the Forecasting Power 73

into a two-factor state space representation comparable to that of equa-

tions (8.5) and (8.8).60 Thereby, the setup for parameter estimation and
forecasting is analogous to the one for the original pricing model. 61

9.1.2 Evidence on Forecasting Quality

The results on the forecasting study are reported in table 9.1. The table
contains the descriptive statistics stated in prices for the valuation model
as well as their differences to the reference model. Generally, to calculate
the prediction error we subtract the forecast results of the implemented
models from the realized closed-end fund prices on the market. The five
selected statistics are calculated separately for the 33 funds in the sample
and described by their average values (MEAN), their standard deviation
(STD), the maximum (MAX), and the minimum (MIN) for all closed-end
funds in the sample.
The first five rows show the mean results over the various funds. The
average forecasting error of our pricing model in absolute terms (mean ab-
solute deviations, MAD) is 59 cents for one week ahead predictions and
continuously increases to $1.29 for the one month period; the correspond-
ing root mean square errors (RMSE) are slightly higher with values from
82 cents for one week up to $1. 76 for the longest period. Compared to
the reference model the predictions are better throughout the different fore-
casting periods by 3 to 8 cents; for the RMSE the results are slightly better
for the first week but weaken for longer forecasting periods. Looking at
the mean maximum and minimum price deviations we find our model to
perform better for all four weeks ranging from 14 up to 69 cents.
The next rows contain information on the variability (STD) of the five
statistics over the different funds. The goodness of fit measured by the
MAD (RMSE) ranges within a one-sigma-interval from 36 to 83 cents (49
cents to $1.14) for the one week ahead predictions. For one month forecasts
the interval widens to 85 cents and $1.74 ($1.10 and $2.43). Compared to
the reference model the results on the variability are better for the MAD
measure, but higher in the case of RMSE.
In the last part of table 9.1 we document the maximum and minimum
60See Hamilton (1994b, p. 387).
61The maximum likelihood results are reliably estimated and not reported.
74 9. Implications for Investment Strategies

Table 9.1: Results on the Forecasting Study in Terms of Market Prices

[Results and Model Results Difference to ARMA(l,l)d)

Differences in $1 lW 2W 3W 4W lW 2W 3W 4W
SSE a ) 37.553 77.659 126.912 173.589 -0.410 0.401 4.850 9.074
MADb) 0.595 0.862 1.100 1.293 -0.029 -0.045 -0.059 -0.078
MEAN RMSE b) 0.815 1.174 1.504 1.764 -0.007 -0.001 0.021 0.037
c)
RES MAX 1.774 2.082 2.317 2.256 -0.137 -0.298 -0.361 -0.574
c)
RES MIN -2.450 -3.657 -4.498 -5.248 -0.179 -0.275 -0.538 -0.692
SSE a ) 29.968 62.577 101.710 135.607 -0.035 1.621 7.732 12.954
MADb) 0.231 0.324 0.397 0.446 -0.019 -0.026 -0.028 -0.039
STD RMSE b) 0.324 0.461 0.581 0.667 0.004 0.01l 0.032 0.045
c)
RES MAX 0.939 1.208 1.222 1.214 0.006 0.013 0.023 -0.047
c)
RES MIN 1.148 1.882 2.208 2.460 0.163 0.286 0.409 0.503
SSE a ) 118.122 290.061 488.989 644.715 -1.973 9.839 36.883 65.832
MADb) 1.128 1.670 2.078 2.375 -0.122 -0.130 -0.163 -0.195
MAX RMSE b) 1.553 2.433 3.159 3.627 -0.013 0.042 0.121 0.190
c)
RES MAX 4.256 6.240 5.494 5.315 -0.041 0.695 -0.028 -0.739
c)
RES MIN -0.919 -1.308 -1.654 -1.720 0.039 0.130 -0.329 -0.483
SSE a ) 5.193 10.038 15.472 20.068 -1.302 -2.788 -4.684 -6.935
MADb) 0.233 0.351 0.460 0.535 -0.038 -0.069 -0.096 -0.121
MIN RMSE b) 0.326 0.453 0.562 0.640 -0.039 -0.059 -0.079 -0.102
c)
RES MAX 0.672 0.379 0.462 0.479 0.099 -0.203 -0.341 -0.452
c)
RES MIN -5.076 -8.831 -10.683 -12.038 -0.241 -0.902 -1.283 -1.661

Notes:
a) We first report the sum of squared errors (SSE).
b) The two measures to evaluate the goodness of fit are the mean absolute deviation
(MAD) and the root mean square error (RMSE).
c) We further show the results on the maximum and the minimum of the prediction
errors (RESMAX and RES MIN ).
d) The differences are calculated as the result from our model minus the result from
the ARMA(l,l) model, i.e. negative values signal better performance than the reference
model.
9.2. Implementing Trading Rules 75

values of the statistics. We find the realized maximum MAD to lie between
$1.13 and $2.37 which is throughout lower than the results for the reference
model. The highest individual prediction errors take values from $4.26 up
to $5.32 for one month. The corresponding minimum realizations of the
forecasting errors are found to range from -$5.08 to -$12.04. This indi-
cates that the pricing model tends to overestimate the forthcoming price
movements and provides evidence that the forecasting power is better in
times of rising market prices. Negative price jumps are only captured with
higher forecast errors which, however, are lower than those predicted by the
reference model.
The different aspects covered by the implemented forecasting study re-
veal a good prediction quality of the pricing model in terms of the closed-end
fund market prices in our sample. This motivates us to examine the second
application on how the information behind the pricing model could be used
to create abnormal return generating trading portfolios.

9.2 Implementing Trading Rules

9.2.1 Experimental Design

Several studies show that abnormal returns can be earned by following trad-
ing strategies based on the information content revealed in the empirical
premia. 62 In the second application of the derived pricing model we follow
the idea of exploiting the information content using the modeled dynamic
premia. For the implemented trading rule study we split the available five
year sample into a four year parameter estimation period and a one year
out of sample implementation period similar to the setup of the forecasting
study.
The filter strategies are based on the information content of the residual
premium difference defined by RESt,i = 7rt,i - (}i for a closed-end fund ion
day t. For this residual value the theoretical mean-reversion property of the

62For the US market of closed-end funds see, for example, Thompson (1978), Anderson
(1986), and Pontiff (1995).
76 9. Implications for Investment Strategies

continuous premia indicate the following information:

> E : buying indication,

{
RESt,i < -E: shortselling indication, and (9.1)
else : no indication.

The selection criterion for going long or short in a closed-end fund is

whether its value of the dynamic premium 1ft lies above or below its long-
run mean O. Upon the realizations of the premium residual we then build
equally weighted trading portfolios containing closed-end funds that are
hold long, short or not at all. The implemented symmetric filter strategies
are based on a minimum amount E, and -E respectively, of the residual
RESt,i in order for closed-end fund i to be included in the trading portfolio.
The trading portfolios are revised on a monthly basis where new closed-end
funds are selected according to the selection criterion and already contained
closed-end funds are excluded when their premium residual becomes zero.
In order to be able to examine how long the information implicit in
the premium residuals lasts in the market we run the filter rules for two
scenarios. We use the information content collected (i) on week t and (ii)
on week t -1 to actually rebalance the trading portfolios on week t. Thereby,
we allow for a time lag of up to one week between the decision to trade and
the actual implementation of the trade.
To evaluate the outcomes of the trading portfolios we utilize the concept
of abnormal returns. 63 Let H Pt,i denote the log-return of closed-end fund
i for a one week holding period. Define ARt,i as the abnormal return for
security i on week t, i.e. the excess return over a benchmark return:

ARt,i = HPt,i - BMt·

As risk adjusting benchmark BMt for week t we use the cross-sectional

average BMt = -k L:!1 H Pt,i' with N = 33 closed-end funds. This risk
adjustment for general market movements within the same market segment
ensures that a naive buy-and-hold strategy of holding all closed-end funds
in the sample over the fifth year earns a risk adjusted return of zero. 64 The
63See, for example, Bjerring, Lakonishok, and Vermaelen (1983), Peavy (1990), and
Campbell, Lo, and MacKinlay (1997).
64The naive buy-and-hold strategy is similar to what serves Thompson (1978) as control
strategy ('All Funds' trading strategy).
9.2 Implementing Trading Rules 77

amounts invested in each selected closed-end fund are assumed to be equal

such that the abnormal return from the trading portfolio is given by the
equally weighted average abnormal return defined by
1
L ARt,i1REst,i'
nt

ARt = ;;:
t i=l
+1 if filter strategy indicates buying,
with lREst,i = { -1 if filter strategy indicates shortselling, and
o otherwise
across the number nt of selected funds contained in the trading portfolios
on week t. The average abnormal returns are then cumulated over the fifth
year and the reported t-tests on a given trading strategy are computed as
in Campbell, Lo, and MacKinlay (1997).
Analogously to the described trading strategies based on the information
content of the dynamic premia we implement the same trading rules using
the empirical values for the premia as defined in equation (7.1). For this
benchmark test we use the residual of the empirical premium RESt,i =
PREMt,i - PREM i , where PREM i serves us as proxy for the long-run
mean value of the empirical premia and is calculated as the sample mean
over the first four years. We then apply the selection criterion of equation
(9.1) with opposite signs, since empirical discounts correspond to positive
dynamic premia. The final procedure to calculate cumulated abnormal
returns is done as described above.

9.2.2 Test Results on Thading Strategies

The results on the implemented filter strategies are reported separately for
the two implementation lags as well as for the different trading strategies,
i.e. the long, short and combined strategies that use the full information
on long and short positions. The results based on the information content
of the dynamic premia and the results for the benchmark test are shown,
respectively, in tables 9.2 and 9.3 and can be summarized as follows.
From table 9.2 we first see that nearly all trading portfolios significantly
earn positive cumulated abnormal returns (CARs) over the fifth year. Us-
ing no filter restriction, i.e. including all 33 funds, the immediate implemen-
tation of a combined strategy yields a risk adjusted cumulated abnormal
78 9. Implications for Investment Strategies

Table 9.2: Results on Portfolio Trading Strategies

Long Strategies Short Strategies Combined Strategies
Filter Obs. nT CAR t-stat Obs. nT CAR t-stat Obs. nT CAR t-stat
(i) No Implementation Lag
0 27.3 90 0.136 3.26*** 5.7 42 0.583 8.92*** 33.0 132 0.218 4.73***
0.025 25.8 92 0.138 3.32*** 4.9 34 0.615 9.17*** 30.7 126 0.222 4.82***
0.05 24.1 94 0.133 3.17*** 4.3 30 0.552 8.09*** 28.4 124 0.201 4.35***
0.075 21.5 80 0.126 2.96*** 3.9 28 0.612 8.99*** 25.4 108 0.199 4.24***
0.1 19.5 82 0.121 2.78*** 3.6 30 0.566 8.44*** 23.1 112 0.193 4.06***
0.125 16.0 86 0.102 2.31** 3.2 22 0.652 9.79*** 19.1 108 0.183 3.76***
0.15 11.6 74 0.064 1.48* 2.7 18 0.586 8.19*** 14.3 92 0.150 3.08***
0.175 9.4 60 0.055 1.24 2.6 22 0.449 6.50*** 11.9 82 0.145 2.88***
0.2 7.3 48 0.009 0.19 2.4 18 0.380 5.53*** 9.7 66 0.122 2.25**
0.225 6.2 34 -0.052 -1.12 2.0 14 0.340 6.25*** 8.3 48 0.073 1.51*
0.25 5.1 30 0.046 0.98 2.0 14 0.328 5.95*** 7.1 44 0.145 2.96***
0.275 4.3 24 -0.011 -0.22 1.9 12 0.349 6.31*** 6.2 36 0.160 3.24***
0.3 3.9 22 -0.023 -0.48 1.6 10 0.330 6.02*** 5.5 32 0.120 2.42***
(ii) Implementation Lag of One Week
0 27.3 90 0.124 2.91*** 5.7 42 0.557 8.35*** 33.0 132 0.200 4.24***
0.025 25.8 92 0.116 2.76*** 4.9 34 0.563 8.19*** 30.7 126 0.194 4.12***
0.05 24.1 94 0.110 2.58*** 4.3 30 0.574 8.04*** 28.4 124 0.183 3.83***
0.075 21.5 80 0.093 2.16** 3.9 28 0.620 8.71*** 25.4 108 0.175 3.62***
0.1 19.5 82 0.093 2.10** 3.6 30 0.561 7.90*** 23.1 112 0.170 3.45***
0.125 16.0 86 0.062 1.39* 3.2 22 0.608 8.60*** 19.1 108 0.147 2.92***
0.15 11.6 74 0.017 0.37 2.7 18 0.589 7.76*** 14.3 92 0.112 2.20**
0.175 9.4 60 -0.003 -0.06 2.6 22 0.557 7.58*** 11.9 82 0.111 2.09**
0.2 7.3 48 -0.061 -1.23 2.4 18 0.476 6.51*** 9.7 66 0.069 1.19
0.225 6.2 34 -0.070 -1.45* 2.0 14 0.318 5.63*** 8.3 48 0.053 1.06
0.25 5.1 30 0.007 0.14 2.0 14 0.298 5.29*** 7.1 44 0.116 2.29**
0.275 4.3 24 -0.014 -0.28 1.9 12 0.305 5.39*** 6.2 36 0.129 2.56***
0.3 3.9 22 -0.020 -0.41 1.6 10 0.312 5.57*** 5.5 32 0.123 2.45***

Notes:
For each type of strategy we first report the average number of CEFs (Obs.) contained
in the trading portfolios and the number of trades (nT) needed to realize the trading
strategy. Statistically significant cumulated abnormal returns (CARs) at the 1%-, 5%-
and lO%-levels are denoted by ***, **, and *, respectively.
9.2 Implementing Trading Rules 79

Table 9.3: Benchmark Results of Trading Rules Based on Empirical Premia

Long Strategies Short Strategies Combined Strategies
Filter Obs. nT CAR t-stat Obs. nT CAR t-stat Obs. nT CAR t-stat
(i) No Implementation Lag
0 27.5 94 0.132 3.07*** 5.5 44 0.607 9.70*** 33.0 138 0.213 4.59***
0.025 26.5 82 0.140 3.35*** 4.5 32 0.649 9.85*** 31.0 114 0.214 4.69***
0.05 25.3 72 0.133 3.18*** 3.5 20 0.731 10.24*** 28.8 92 0.208 4.54***
0.075 23.0 58 0.129 3.07*** 3.3 18 0.686 9.44*** 26.3 76 0.205 4.43***
0.1 22.0 54 0.144 3.43*** 3.0 14 0.623 8.16*** 25.0 68 0.211 4.56***
0.125 20.5 48 0.136 3.48*** 2.8 14 0.557 7.62*** 23.3 62 0.193 4.41***
0.15 17.8 44 0.146 3.72*** 2.7 12 0.536 7.70*** 20.5 56 0.199 4.62***
0.175 14.7 44 0.128 3.17*** 2.6 12 0.504 7.64*** 17.3 56 0.183 4.14***
0.2 9.1 36 0.148 3.57*** 2.2 8 0.490 7.61*** 11.3 44 0.213 4.60***
0.225 5.5 28 0.078 1.73** 1.8 8 0.481 7.36*** 7.3 36 0.167 3.30***
0.25 4.3 18 0.084 1.86** 1.8 8 0.481 7.36*** 6.1 26 0.199 3.83***
0.275 2.5 12 -0.064 -1.29 1.8 8 0.481 7.36*** 4.3 20 0.152 2.60***
0.3 1.3 4 -0.011 -0.22 1.1 4 0.421 6.71*** 2.4 8 0.210 3.58***
(ii) Implementation Lag of One Week
0 27.5 94 0.116 2.67*** 5.5 44 0.580 8.75*** 33.0 138 0.188 3.95***
0.025 26.5 82 0.130 3.06*** 4.5 32 0.640 9.24*** 31.0 114 0.199 4.23***
0.05 25.3 72 0.120 2.83*** 3.5 20 0.616 8.39*** 28.8 92 0.182 3.90***
0.075 23.0 58 0.122 2.88*** 3.3 18 0.566 7.57*** 26.3 76 0.184 3.90***
0.1 22.0 54 0.137 3.26*** 3.0 14 0.533 6.69*** 25.0 68 0.197 4.15***
0.125 20.5 48 0.135 3.43*** 2.8 14 0.520 6.80*** 23.3 62 0.189 4.20***
0.15 17.8 44 0.141 3.57*** 2.7 12 0.539 7.48*** 20.5 56 0.200 4.56***
0.175 14.7 44 0.124 3.07*** 2.6 12 0.533 7.75*** 17.3 56 0.189 4.17***
0.2 9.1 36 0.144 3.45*** 2.2 8 0.575 8.61*** 11.3 44 0.231 4.89***
0.225 5.5 28 0.044 0.96 1.8 8 0.560 8.15*** 7.3 36 0.163 3.13***
0.25 4.3 18 0.022 0.48 1.8 8 0.560 8.15*** 6.1 26 0.177 3.35***
0.275 2.5 12 -0.075 -1.53* 1.8 8 0.560 8.15*** 4.3 20 0.176 2.94***
0.3 1.3 4 -0.082 -1.63* 1.1 4 0.505 7.67*** 2.4 8 0.233 3.84***

Notes:
For each type of strategy we first report the average number of CEFs (Obs.) contained
in the trading portfolios and the number of trades (nT) needed to realize the trading
strategy. Statistically significant cumulated abnormal returns (CARs) at the 1%-, 5%-
and lO%-levels are denoted by ***, **, and *, respectively.
80 9. Implications for Investment Strategies

return of 21.8 percent which marginally falls to 20.0 percent for an imple-
mentation lag of one week. Second, from the average number of observations
(Obs.) for a given portfolio we see that the long strategies contain two to
six times more closed-end funds in their trading portfolios on average than
do the short strategies. This indicates that the dynamic premia lie above
their long-run equilibrium () in most cases of the fifth sample year which cor-
responds to observable closed-end fund discounts of the empirical premia.
Third, we notice in general that the short strategies earn higher cumu-
lated abnormal returns than do the buying strategies. With an immediate
implementation of a long strategy the cumulated abnormal return results
start from 13.6% and then gradually fall to negative values for higher filter
values. However, the cumulated abnormal returns of around 13% are re-
mained for filter values up to 0.1 which signal that the optimal filter height
for a trading strategy lies between zero and ten percent. The comparable
short strategy selects a daily average of 5.7 funds and yields a cumulated
abnormal return of 58.3% with no filter implemented. Here, the highest
cumulated abnormal return with 65.2% is realized for an intermediate fil-
ter height of 0.125. Comparing the overall results for the immediate and
the lagged implementation of the trading strategies show that the main re-
sults are preserved and the information content persists over one week. The
realizable values for the cumulated abnormal returns, however, are not as
high any more. Fourth, in implementing the portfolio strategies with the
aim of earning high cumulated abnormal returns one needs to consider the
transactions costs associated with these strategies. The reference number
of trades (nT) is 66 which stems from building a buy-and-hold portfolio
containing all closed-end funds and its liquidation at the end of the year. 65
In order to administer the long and the combined strategies the number of
trades lies above 66 for attractive cumulated abnormal returns. The highly
profitable short strategies can be followed with twenty to forty transactions.
Compared to the short strategies we further find the long portfolio holdings
more stable over time in that the relation of the number of trades to the
included closed-end funds is generally lower. With the objective to lower
transactions costs it seems optimal to pursue an intermediate filter strat-
egy of around 0.1 which only marginally sacrifices the returns. Finally, in

65The return of this naive buy-and-hold strategy is already incorporated as benchmark

return in our reported cumulated abnormal returns.
9.2 Implementing Trading Rules 81

order to evaluate the outcomes of the trading test we compare the model
based results to those obtained from the information content of the empir-
ical premia which we report in table 9.3. The facts already described for
the model based results principally hold for the benchmark results as well.
When comparing the realizable cumulated abnormal returns our model per-
forms better for filter values close to zero in terms of long and combined
strategies. Overall, we see the similar pattern that the cumulated abnor-
mal returns remain stable up to certain filter values while the number of
closed-end funds and the number of transactions decrease.
10 Summary and Conclusions

Part II of the study contains a theoretical and empirical analysis on the

special equity asset class of closed-end funds. The analysis contributes to
the pricing literature on closed-end funds in that we develop a valuation
model that captures the distinct pricing characteristics of closed-end fund
market shares. Given the pricing model we are able to perform a general
closed-end fund analysis and derive insights into two potential applications
of the model valuable for investment decisions.
In chapter 7 we focus on the theoretical modeling perspective where we
derive a closed-form, two-factor valuation model for closed-end fund market
shares which captures the distinct pricing characteristics of closed-end funds
trading at discounts which vary over time. As explaining factors we utilize
the underlying price of the NAV and a stochastic premium which is modeled
with dynamic mean-reversion. Thereby, the pricing model incorporates both
the price risk stemming from the originating portfolio investments and the
risk associated with time varying values for the premia.
For the empirical implementation of the theoretical pricing model in
chapter 8 we use a state space representation to estimate the model pa-
rameters and the time-series of the dynamic premium by maximum likeli-
hood inferences. In the empirical estimations we use a historical sample of
emerging market closed-end funds for the period of 1993 to 1997. For the
examined closed-end fund sample we observe an average of the empirical
discount of 3.25 percent which is consistent with the literature. The sample
further confirms a higher volatility of the share prices than the prices of the
underlying net asset value which is in line with the noise trader argument.
In the empirical analysis of our pricing model we find a high mean-reversion
of the dynamic premia with an approximate half-life of five months. This
indicates high fluctuations of the premia over time which is emphasized by
84 10. Summary and Conclusions

a premium volatility of 30 percent on average and the outcome we report

for the time-series of the dynamic premia. For the results on the parame-
ter estimates of the factor volatilities we obtain a higher volatility for the
dynamic premia than for the net asset value on average. This presents evi-
dence that the risk with altering premia is higher than the underlying price
risk.
The possible use of the pricing model to investors is evaluated by infer-
ring insights into two suggested applications in chapter 9. First, we imple-
ment an out of sample forecasting study in order to evaluate the prediction
quality of the model in terms of closed-end fund market prices. The esti-
mation results based on our sample document that the models performance
is better in most aspects when compared to a well specified econometric
reference model. In a second application we test whether the information
contained in the dynamic premia can be used to build portfolio strategies
with the objective to generate risk adjusted abnormal returns. In an out
of sample setting we find that long trading strategies can earn cumulated
abnormal returns of around fourteen percent which can be increased to
over twenty percent following both long and short strategies. FUrther, we
see that an immediate implementation of the strategies is more valuable
than implementing them with a time lag which indicates that the informa-
tion content of the premia vanishes over time. Overall, we find that our
proposed valuation model to price closed-end funds can generate valuable
information for investment purposes.
 Part III

Term Structure Modeling

11 Introduction

11.1 Overview
In part III of the study we explore a new approach in modeling the term
structure of interest rates and the pricing of fixed income instruments. The
pricing model we present is a type within the affine class and is estimated
using interest rate panel data. This first chapter considers the peculiarities
of modeling and pricing assets in an incomplete market such as the market
for fixed income instruments. We clear the basic definitional terminology
when dealing with bond prices and interest rates and motivate our choice of
model type. In chapter 12 we present our dynamic term structure model and
derive a closed-form solution for the price of discount bonds as the security
of primary interest. In the next chapter we perform a comparative statistic
in order to evaluate the characteristic theoretical properties of our interest
rate model. Here, we especially examine the distinct behavior of the term
structure of interest rates as well as the term structure of volatilities which
are relevant in valuing derivatives. The following chapter 14 is dedicated
to the management of interest rate risk. Starting with a clarification of the
types of risk involved, we show how to implement the duration technique for
our model, and finally show how to price term structure derivatives. Among
the most relevant fixed income instruments in risk management we special-
ize in valuing bond options, swap contracts, and interest rate floor and
cap agreements. Chapter 15 considers the calibration of the proposed term
structure model to standard fixed income instruments. First, we describe
the different econometric methods used in the literature to then specialize
our implemented version that resembles the theoretical model properties at
best. For the different interest rate markets we choose data from US govern-
ment issues, LIBOR rates and swap rates. Upon these data we report our
88 11. Introduction

inferences on the model parameters and analyze the filtered state variables.
Finally, we conclude in the last chapter with a summary of our findings.
Before we describe our proposed term structure model in the next chap-
ter, we first take a look at the incomplete market for fixed income instru-
ments. There, we unfortunately realize that interest rates are not directly
tradeable. Instead different financial instruments are traded on the markets
which depend on interest rates themselves. Guided by this fact, we initially
consider the basic financial instruments in the bond market and how they
are related to each other. Second, we closer examine the peculiarity of the
risk premium concept that generally need to be taken into account in inter-
est rate modeling. Last but not least, we theoretically motivate our choice
of a special interest rate model with a risk premium following stochastic
dynamics.

11.2 Bond Prices and Interest Rates

We start with the framework of calculating interest rates in continuous
time. The framework of continuous compounding can be traced back to
two different origins: (i) The field of econometrics66 , where in time-series
analysis the differences in logarithmic security prices Pt, which can be equity
prices, bond prices or foreign exchange rates,

are called logarithmic yields r, or just "first differences" in terms of levels.

The use in the field of finance lays (ii) in the theory of finance as described,
for example in Ingersoll (1987), where the continuous compounding evolves
from within period compounding. Instead of one compounding period, we
incorporate n subperiods, and obtain for the limiting case of getting interest
payments continuously

-P- = lim
t (r)n
1+- ;
Pt - 1 n-+oo n

this result can be stated in the form of the Taylor series expansion
. (
1~m
noon
r)
1+ -
n
= 1 + -1'1. 1 2 1 3
r +.
2' r +
.
3' r + 0
( 3)
r = er .
66S ee, for example, Campbell, Lo, and MacKinlay (1997).
11.2 Bond Prices and Interest Rates 89

The equivalence between the use in econometrics and finance is given by

the law of logarithms that In (.Pt) -In (Pt - l ) = In (Pt./ Pt.-l). More formally,
we now define the basic building blocks of the bond market, consisting of the
prices of pure discount bonds, coupon bonds, the money market account,
and the interest rates, as follows.

Definition 11.2.1 (Bond Market) A discount bond is defined as a non-

defaultable security, which guamntees the holder to receive $1 at a prespec-
ified maturity date T; its price at time t will be denoted by P (t, T). The
theoretical price of a coupon bond pc (t, T) will then be

pc (t, T) = L c(u)· P(t, u)

uE(tjT]

where c (u) denotes the cash flows (coupon and notional) received from the
coupon bond holder at time U. 67

The relationship between discount and coupon bond prices is rather

crucial, because in real world capital markets discount bonds are rarely
available. 68 Instead, we are unluckily stuck with extracting our necessary
information from market prices of traded coupon bonds, which even might
be additionally equipped with special features that make the analysis even
worse. 69 A common way to extract the term structure of interest rates is
to use market prices of coupon bonds, interest rate futures, or interest rate
swaps. 70 In the latter case we reinterpret a coupon bond in light of the
portfolio theory as a portfolio of distinct cash flows discounted with their
corresponding interest rates which we define next.

67 Otherwise there would exist arbitrage opportunities of Type 2, according to the

nomenclature of Ingersoll {1987}. Farkas's Lemma (see Farkas (1902)) can serve as cri-
terion for this arbitrage-free relationship between discount and coupon bond prices.
68 However, stripping coupon bonds into separate cash flows is becoming more popular
and widespread in the US and Germany. For an empirical analysis of the US Treasury
STRIPS program see, for example, Love and McShea {1997} and Grinblatt and Schwartz
{2000}.
69For example, a bond with embedded options such as callable US Treasury Bonds,
will probably deteriorate from the price of an equivalent pure coupon bond {Bliss and
Ronn {1998}}.
70 See, for example, Bliss {1997} and Uhrig and Walter {1996}.
90 11. Introduction

Definition 11.2.2 (Interest Rates) Given the price P (t, T) of a discount

bond we define the continuously compounded spot rates for the period [t, TJ
by

( T) = _In P (t, T) (11.1)

y t, T
-t
'

The functional mapping T 1----+ Y (t, T), for a given time t, is called the yield
curve. The spot rate y (t, T) collapses at the short end of the term structure
to

=l' _lnp(t,T)=_OP(t,tl)1 =f( )

r ()
t 1m T I ):l t, t ,
TIt - t vt t'=t

which equivalently is the instantaneous short and forward rate. The general
forward rate for [T1' T2J contracted at t is defined as

In P (t, T2 ) -In P (t, T1 )

(11.2)
T2 -T1

By taking the limit of T2 1 T1 , we converge at the instantaneous forward

rate for the maturity T contracted at t

f (t, T) = _ 8ln:it, T),

which constitutes the rate for instantaneous borrowing and lending in T.

By these definitions on interest rates, we can characterize the relation-

ship between the term structure of spot and forward rates. We differentiate
the formula for the spot rates (11.1) with respect to T:

oR(t,T) = _ (81np(t,T)_I_ -lnP(t T) 1 ).

8T 8T T- t ' ( T - t)2

Hence, we are able to deduce the rearrangement

oR (t T)
f(t,T)=y(t,T)+(T-t) cJ,
which shows that the forward curve superficially resembles the spot curve
dependent on the slope of the spot curve. The forward curve is higher
11.2 Bond Prices and Interest Rates 91

than the spot curve if the yield curve is increasing, and vice versa. More-
over, using the definitions on interest rates we gain the ordinary but crucial
relationships
P(t, u) = e-(u-t).r(t,u), and
pc (t,T) = L CF(u)· e-(u-t).r(t,u)
uE(tjT]

for discount and coupon bond prices, respectively. By restating the price of
a discount bond as

Jf
T

P(t, T) = exp (t, () d(

t
we see that we need to take information on future interest rates or the
forward rate into account for a fair valuation of bonds. Equivalently, market
prices on bonds reveal information about these interest rates.
A different security which will be of interest as so-called numeraire asset
is the money market account. This security looks into the past and present
information on interest rates.
Definition 11.2.3 (Money Market) For the money market, we define
the money account1 1 by
t

B(O,t)=exp jr(()d(,
o
i.e. the account starts with an initial value of B (0,0) = 1 and accumulates
interest with the dynamics dB (0, t) = r (t) B (0, t) dt over time.
Finally, to value financial instruments such as interest rate swaps, cap
or floor agreements we introduce a special interest rate, the LIBOR rate.
This interest rate uses simple instead of continuous compounding as we see
from its definition.
Definition 11.2.4 (LmOR Rate) The simple spot rate or LIBOR rate
is defined as
P(t,T) -1
L (t, T) = - (T _ t) P (t, T) (11.3)

for the time interval [t, TJ.

71 Also known as accumulator or savings account.
92 11. Introduction

The relationship between interest rates and forward rates shows that
interest rates of different maturities are heavily correlated. 72 Therefore and
to keep an interest rate model arbitrage-free and analytically tractable, the
stochastic processes governing r (t) or f (t, T) are assumed to be functions
of a fixed number of Brownian motions.
The considered relationships between bond prices and different types
of interest rates along with the fact that there are several interest rate
dependent securities traded on the financial markets and interest rates on
the other hand are non-tradable call for a special treatment as we will discuss
in the next section.

11.3 Modeling an Incomplete Market

First, we investigate some general aspects of using a factor model within
the affine class of Duffie and Kan (1996) to price interest rate derivatives.
Among the general aspects found in modeling bond markets the following
interrelated peculiarities are examined: the incompleteness of the market,
the maturity independent risk premium, and the pricing of interest rate
derivatives. To study these aspects we use a single-factor model that chooses
the instantaneous short rate as the state variable:

"More recently it has been recognized that, if assumptions

are made about the stochastic evolution of the instantaneous
rate of interest in a continuous time model, much richer theories
of bond pricing can be derived, which constrain the relationship
between the risk premia on bonds of different maturities."73

Intuitively, the choice of the short rate is a natural starting point in

that the price of a discount bond P (t, T) depends on the behavior of the
short rate during the interval [t, T] as we demonstrated in the previous sec-
tion. From an empirical point of view, statistical approaches to interest rate
modeling adopted and followed by, for example, Litterman, Scheinkman,
and Weiss (1991) and Rebonato (1998) show that the first factor can be
721£ one would implement a multi factor regression, this would possibly give rise to a
multicolinnearity problem. See the analysis of chapter 15.
73Brennan and Schwartz (1979, p. 134).
11.3 Modeling an Incomplete Market 93

attributed with the level of interest rates. In fact, this is the classical ap-
proach to interest rate theory adopted in the field of financial economics
with the most prominent representatives being the arbitrage-free Vasicek
(1977) model and the utility based methodology of Cox, Ingersoll, and Ross
(1985b). The models make assumptions about the stochastic evolution of
the short rate in a continuous time model. The specification of such dynam-
ics are chosen with the aim of incorporating the fluctuations of the short
rate during time at best. The peculiarity about these types of models is
that the short rate as chosen factor is a non-traded underlying object in the
interest rate market. Therefore, we deal with the bond market as an incom-
plete market in comparison to, for example, the well known stock market
model dating back to Black and Scholes (1973). Since the bond market
is incomplete by assumption, we will further see that it is not possible to
hedge a generic derivative contract as well as we will not be able to derive
a unique price on such a contract. We begin with the assumption of the
relevant model for the bond market.

Assumption 11.3.1 (Bond Market) For the state variable in our model
we assume the a priori specified belief of the short-rote dynamics

dr (t) = J-l (r, t) dt + (7 (r, t) dWt

under the objective probability measure l? By assumption the short rote is

the only object given a priori, i. e. there exists only one exogenously given
asset which is the risk-free asset. This money market account follows the
l?-dynamics given in definition 11.2.3. We further assume the existence
of a market for discount bonds P (t, T) for every choice of maturity T.
The bond price is modeled by a function of the short rote r (t) and time t,
i.e. P (t, T) = f (t, r; T), where the maturity date T is treated as a constant
parometer.

In order to find the arbitrage-free price of the discount bonds, we use the
idea of an investor seeking for arbitrage opportunities. He has the possibility
to hold three different assets: The money market account and two discount
bonds P (t,~) of different maturities i E [1,2]. The discount bonds follow
94 11. Introduction

the price dynamics

dP (t, Ii) (11.4)

1 [aP(t,Ii) aP(t,Ii) 1 2a2P (t,Ii)] d

with /-Li P (t, Ii) at + /-L or + 20- ar2 ' an

'T/i =
1 [0- aP (t, Ii)]
P (t, Ii) or

found from the above assumption using Ito's formula. The portfolio struc-
ture of the arbitraguer is given by the sum of the investments in the three
tradeable assets in the market:

V (t) = no (t) B (t) + nl (t) P (t, TI ) + n2 (t) P (t, T2) ,

with the absolute portions denoted by nj with j E [1,2, 3J. His portfolio
increases or decreases in wealth over the interval [t, t + dtJ by

dV (t) = no (t) dB (t) + nl (t) dP (t, T I ) + n2 (t) dP (t, T2)

+dno (t) B (t) + dnl (t) P (t, T I ) + dn2 (t) P (t, T2)
+dno (t) dB (t) + dnl (t) dP (t, T I ) + dn2 (t) dP (t, T2) .(11.5)

To ensure that the formed portfolio is self-financing over time, it needs

to qualify the condition

+ dB (t)) + dnl (t) (P (t, T I) + dP (t, TI ))

dno (t) (B (t)
+dn2 (t) (P (t, T2) + dP (t, T2)) = o.

This condition states that the portfolio need to be restructured without

requiring any cash inflow and does not permit any outflows which reduces
equation (11.5) to:

dV (t) = [no (t) r (t) B (0, t) + nl (t) /-LIP (t, T I ) + n2 (t) /-L2P (t, T2)] dt
+ [nl (t) 'T/IP (t, T I ) + n2 (t) 'T/2P (t, T2)] dW.

To constrain this wealth dynamics to an arbitrage-free portfolio we use the

following conditions.
11.3 Modeling an Incomplete Market 95

Condition 11.3.2 (No Arbitrage) 74 Conditions for there being no pos-

sibilities of arbitrage: (i) Zero investment, i.e. the portfolio is self-financing
at the beginning

(ii) No market risk which requires

This in return means that there can be (iii) no return:

i. e. the arbitraguer should not be able to realize any profits on his portfolio.

Using the abbreviations Wo (t) = no (t) B (0, t) and Wi (t) = ni (t) P (t, Ii)
for the relative portfolio weights the conditions of no-arbitrage reduce to

1 1 1] [wo (t) ]
[ o "11 "'2 W1 (t) = o. (11.6)
r (t) ILl IL2 W2 (t)
Therein, solutions for the portfolio weights only exist if the market satisfies
the internal consistency relation, i.e. the matrix in equation (11.6) needs to
be singular. With one degree of freedom we choose, for example, the last
row being a linear combination of the remaining rows. Using>.. for the free
coefficient not determined in the model for the bond market yields

These equations reduce to the internal consistency result for the bond mar-
ket:
\ ILl - r IL2 - r
A=--=--. (11.7)
"'1 "'2
More specifically, the quantity>.. is called the market price of interest rate
risk, since it gives the extra increase in expected instantaneous rate of return
74 For a valuation of derivative contracts based on no-arbitrage see, for example, Schobel

(1995b, ch. 2).

96 11. Introduction

J.Li - r on a bond per an additional unit of risk 'fJi' The result of equation
(11.7) now states that all bonds in an arbitrage-free market have the same
market price of risk regardless of the time to maturity. This result entails
even more information on the bond dynamics. We substitute the drift and
volatility coefficient of equation (11.4) into one of equations (11.7) to obtain
the so-called term structure equation

~
2a
2 1j2 P
ar2
(t, Ii)
+
(-.A)
J.L a
ap ar
(t, Ii) ap (t, Ii)
+ at
- P ( 1',.)
- r t,..
This equation governs the price of discount bonds in an arbitrage-free mar-
ket when we additionally incorporate the appropriate boundary condition
P (Ii, Ii) = 1 for discount bonds. We also note that the price of a particular
bond P (t, Ii) is not uniquely given by the bond market model in that the
market price of risk .A remains undetermined in the model. It constitutes
an exogenous parameter. The reason for this fact is that arbitrage pricing
is always a case of pricing a derivative (in our case P (t, Ii)) in terms of
some underlying assets (in our case r). But, with the bond market choosing
a value for .A to satisfy the internal consistency relation of equation (11.7)
arbitrage possibilities are avoided. At this point we see the possibility to
establish a new dimension in term structure modeling in that we build a
term structure model that captures not a time homogeneous but a time
varying market price of risk .A (t).
12 Term Structure Model

12.1 Motivation for a Stochastic Risk Pre-

mium
In financial economics we model investors' attitudes towards risk in order
to come up with a reasonable relationship of risk and return in the relevant
market. In standard microeconomic theory investors' tastes are modeled
by specifying their utility function, being, for example, an investor with a
logarithmic utility function as in the production-exchange economy of Cox,
Ingersoll, and Ross (1985a). To get a utility function for a representative
investor, we then aggregate the utility functions across investors with po-
tential aggregation problems. 75 In an equilibrium term structure model this
utility function indirectly enters into the form of the market price of risk.
However, looking at real world security markets we see different investors
present on different trading days. Thus, we could model these changing
representative tastes over time by allowing for a dynamic behavior of the
market price of risk A (t). This approach is inspired by a shift from mod-
eling static tastes as with previous term structure models to capture the
beliefs of the dynamic behavior of tastes. Instead of assuming the same
representative investor being present over time, i.e. assuming A (t) = canst,
we set up a plausible dynamic model of how the market price of risk possibly
fluctuates over time.
Besides the purpose of modeling the dynamic behavior of investors'
tastes we additionally want to build a term structure model that has promis-
ing features in comparison to extant models. A range of desirable theoreti-
cal, technical, and empirical features can be summarized by requiring some

75S ee Ingersoll (1987).

98 12. Term Structure Model

criteria that need to be met by the model. For Rogers (1995) a term struc-
ture model should be:

" (a) flexible enough to cover most situations in practice; (b)

simple enough that one can compute answers in reasonable time;
(c) well-specified, in that required inputs can be observed or es-
timated; (d) realistic, in that the model will not do silly things.
Additionally, the practitioner shares the view of an econometri-
cian who wants (e) a good fit of the model to data; and a theoret-
ical economist would also require (f) an equilibrium derivation
of the model." 76

A class of term structure models that broadly meet these criteria and
has recently been attended considerable focus is the affine class. 77 This
class is defined by factor models in which the drift and volatility coefficients
of the state variable processes are affine functions of the underlying state
vector. 78 Our approach is to work within this class of factor models and
extend the single-factor term structure models, such as the model used in
the previous section, to include the market price of risk A (t) as a second
state variable. Thereby, we establish a two-factor model that enables us to
analyze the investors' beliefs with respect to the interest rate risk inherent
in the observable term structures on the markets. We aim at building a two-
factor model that accommodates the above mentioned criteria. Especially,
as we will see in the following section, we try to build a maximal model that
does not suffer from implicitly imposing potentially strong over-identifying
restrictions. 79
Looking at the extant financial literature on this subject we already find
several propositions of two-factor model specifications. The main idea is
that single-factor models based on specifying a dynamic structure of the
short interest rate exhibit some shortcomings that are rather unrealistic. 8o
The most problematic feature of such models is that they endogenously
result in a perfect correlation among bond prices of different maturities.
76Rogers (1995, p. 93).
77For an overview of different classes in term structure modeling see, for example,
Uhrig (1996, ch. 1).
78See, for example, Duffie and Kan (1996).
79See, for example, the specification analysis of Dai and Singleton (2000).
80See, for example, Chen (1996b, ch. 2).
12.1 Motivation for a Stochastic Risk Premium 99

Several studies have shown that such single-factor models are not suitable
in modeling the term structure of interest rates. 8l

Table 12.1: Developments in Two-Factor Models

Authors Model Specification Constants

Brennan/ dr = (al + bl (1 - r)) dt + raldz l aI, bl , aI,
Schwartz d1 = 1(a2 + b2r + c21) dt + la2dz2 a2, b2, C2, a2,
(1979,80,82) dz l dz 2 = pdt P
Schaefer/ ds = m (/-l - s) dt + "'(dz l m,/-l,"'(,
Schwartz dl = (a 2 -ls) dt + aVZdz2 a,
(1984) dz l dz 2 = pdt p
Pennacchi dr = (al + bll (r + ak) + b121T) dt + ardzr al, bll , b12 , ak, a r
(1991) d1T = (a2 + b2l (r + ak) + b22 1T) dt + a 7r dz7r a2, b21 , b22 , a 7r ,
dzrdz7r = pdt p
Beaglehole/ dr = Kl (0 - r) dt + aldwl,t Kl, aI,
Tenney dO = K2 (0 2 - 0) dt + a2dw2,t K2, O2, a2,
(1991) dWl,tdw2,t = pdt p
Longstaff/ dX = (a - bX) dt + cVXdZl a,b,c,
Schwartz dY = (d - eY) dt + fVYdZ 2 d,e,f
(1992) dZl dZ2 = 0

An overview of proposed two-factor specifications of term structure mod-

els is given in table 12.1. The first extensions were to incorporate the shape
of the yield curve in some way. The ideas of Brennan and Schwartz (1979),
Brennan and Schwartz (1980), Brennan and Schwartz (1982), and Schaefer
and Schwartz (1984) were to model the processes for the long-term interest
rate 1 in addition to the short rate of, respectively, r and s. In the context
of an equilibrium asset pricing model Pennacchi (1991) uses the instanta-
neous real interest rate r along with the expected inflation 1T as factors to
establish a nominal term structure model. The term structure model of

81The empirical studies on this subject include Chen and Scott (1992) and Pearson
and Sun (1994) who use a maximum likelihood estimation, the works by Heston (1992)
and Gibbons and Ramaswamy (1993) based on the generalized methods of moments,
and the factor analysis of Litterman and Scheinkman (1991).
100 12. Term Structure Model

Beaglehole and Tenney (1991) corresponds to an economy where the short

e
rate r reverts to a drift rate which itself reverts to a fixed mean rate over
time. From a specification point of view we will see in the next section
that our model has similarities with those derived by Pennacchi (1991) and
Beaglehole and Tenney (1991). Another approach is the idea to model the
stochastic volatility of the short rate as a factor. In this line Longstaff and
Schwartz (1992) extend the general equilibrium term structure model of
Cox, Ingersoll, and Ross (1985b). Based on two orthogonal economic fac-
tors X and Y they are able to derive a term structure model which utilizes
the short rate and the volatility of the short rate as factors.
The idea to further add an additional third factor to the term structure
models has also been followed in the literature. 82 However, the specifica-
tion analysis performed by Dai and Singleton (2000) raises doubts on the
enhanced richness of such models.

12.2 Economic Model

The proposed model belongs to the class of factor models. This model
class has two distinguished implications for valuation purposes that can be
inferred from the description of the theoretical valuation of these models:
the time-series implications from the specification of the dynamic evolution
of the chosen factors during time, and the cross-section dimension with a
model for the functional relationship of the value of discount bonds and time
to maturity at a single point of time. Thus, the parameters of the models
fulfill two tasks. First, they determine the form of the random evolution
of the factors through time by the specific form of the stochastic processes.
Second, they fix the various types of term structures out of the realizable
spectrum of different time to maturities. In the following we focus on the
first aspect whereas the consecutive two sections focus on the second aspect
and will clarify the distinction between the two dimensions.
To derive the term structure model we start with the stochastic dynamics
of the chosen underlying factors from a time-series perspective. The idea is
to extend the specification of the short rate r (t) as an Ornstein-Uhlenbeck

82Chen (1996a), for example, develops a model using the short rate, the short-term
mean, and the short-term volatility as explaining factors.
12.2 Economic Model 101

process83 by a stochastic market price of risk>. (t).

Assumption 12.2.1 (State Variables) Under the objective probability

measure JP> (where increments of Brownian motions are denoted as dWIP)
we specify the stochastic dynamics for the state variables x = [r (t), >. (t)]',
with r (t) being the instantaneous interest rate and>' (t) the market price of
the interest rate risk, as:

dx [J.l + TlPx] dt + udWf, (12.1)

with J-LIP = [/-lr' /-loX]' , TIP = [ :: ~oX ] ,

and dWf = [dW~t, dWf,t]'·

Therein, we allow for a correlation of the two sources of uncertainty dW~t,

dwf,t = pdt between the short rate process r (t) and the price process for
>. (t). This process specification assumes a mean-reversion for the short rate
as well as for the market price of risk.

Corollary 12.2.2 (Equivalent Martingale Measure) Going through

t
the Girsanov transformationsB 4 W~t = W~+ J >. (() d(
with the Girsanov
o
Kernel>. (t), which also holds for our stochastic specification of the drift
t
>. (t), and Wf,t = W~t + J >'2d(, we obtain under the equivalent martingale
o
measure Q (with Brownian motions wQ) for equations (12.1) the risk
adjusted processes:

dx [J-LQ + TQx] dt + udW?, (12.2)

··"th J-LQ -- [/-lr' /-loX

UfO
\J'
- 11 oX "2 ,

83Proposed by Vasicek (1977). For the reasonability of a mean-reversion specification

for the short rate see, for example, Backus, Foresi, and Zin (1998).
84See the original work by Girsanov (1960) and, for example, Oksendal (1995).
102 12. Term Structure Model

Under the drift transformations the mean-reversion property and the corre-
lation d~dW~t = pdt of the two processes are retained.

Remark 12.2.3 (Pull System Matrix) Note that the system of stochas-
tic differential equations (12.2) incorporates a full system matrix TQ, i. e. we
model a two-fold interdependence: The two modeled state variables r (t) and
..\ (t) may both depend on their own value as well as on the other factor's
value. Furthermore, the state variables are related through the correlation
between the Brownian increments.

First, we apply 1t6's formula on a general, not yet specified interest

rate derivative security with price P(t, T, x) on date t and a maturity date
T based on the risk adjusted stochastic differential equations of equation
(12.2).85 This results in the following stochastic differential:
dP(t, T,x)
(12.3)
P(t, T,x)

with /-LP P(t, T, x)

1 [1 2f]2P(t,T,x)
"2O"r ar2 + PO"rO")"
a 2P(t,T,x)
ara..\
1 2 0 2 P (t, T, x) aP (t, T, x)
+"20")" 0..\2 + f-Ll or
aP (t, T, x) aP (t, T, X)]
+f-L2 0..\ + at '
f-Lr + Krr - O"rA,

f-L)" - 0")"..\2 + K)"r - (h..\, and

1 [aP(t,T,x) aP(t,T,X)]'
P(t, T, x) O"r or 0")" 0..\

To get to the fundamental partial differential equation, we use the stan-

dard finance condition86 as a necessary condition87 to ensure the pricing
of the derivative in absence of arbitrage opportunities: In a risk-adjusted
environment all financial claims have to earn the same rate of return as
a risk-less asset, i.e. the drift of the price process /-Lp needs to equal the
85The derivative security price P (t, T, x) is a function of the variables time t, the short
rate r, and the market price of risk ,x, whereas the maturity date T is regarded as a
parameter.
86See, for example, Heath, Jarrow, and Morton (1992, pp. 86-87).
87See Morton (1988).
12.2 Economic Model 103

risk-less interest rate r (t). This approach is equal to Merton's argument

that a risk-less security or portfolio must earn the spot rate r (t) in equilib-
rium. 88 Thus, we have a partial equilibrium condition which implies that
there are no possibilities for arbitrage gains. 89 This leads us in our case to
the following partial differential equation:

(12.4)

ordered in terms of derivatives. Equation (12.4) is called the fundamental

partial differential equation in option pricing since it holds for any fixed
income instrument P. Based on this equation we derive arbitrage-free prices
of discount bonds in the following chapter in order to evaluate the theoretical
properties of the implied term structures. With the obtained cross-section
we then fully specify the dimensions of the term structure model besides
the already given time-series properties. In the chapter thereafter we use
equation (12.4) to price several fixed income instruments and derivative
securities that are crucial in risk management.

881.e. the stochastic part in the models corresponding stochastic differential equation
drops out. See, for example, Merton (1973).
89Note that the vice versa argument does not hold; freedom of arbitrage does not imply
an equilibrium.
13 Initial Characteristic
Results

In this chapter we present initial characteristic results of our term structure

model. At first, we show how to price discount bonds considered to be the
most basic interest rate contracts serving as pure securities in the sense of
Arrow /Debreu. Based on the discount bond pricing formula we are able
to derive the term structures of interest rates and volatilities. Further, we
analyze limiting cases of our model where we first demonstrate that our
model contains the Ornstein-Uhlenbeck process model proposed by Vasicek
(1977) as special case. Second, we examine the asymptotic behavior of the
term structure of spot interest rates at the short and long end. In the last
section we discuss the core influences of the state variables and the model
parameters on the shape of the term structures in a comparative statistic
analysis. Such an analysis is considered relevant in order to know which
type of term structures are realizable within the model and to get an idea
on how changes in the values of the state variables and the model parameters
influence the shape of the term structures.

13.1 Valuing Discount Bonds

In order to value discount bonds we specialize the partial differential equa-
tion (12.4) on the case of P being the value of a discount bond. Thereby,
we impose the boundary of P (T, T, x) = 1 for the discount bond price at
maturity.

Proposition 13.1.1 (Discount Bond Price) For our Gaussian term

structure model we propose in light of Duffie and Kan (1996) an exponential
affine structure for the discount bond prices P (t, T, x) as:
106 13. Initial Characteristic Results

P (t, T, x) = e-[A(t),B(t)]x-C(t). (13.1)

For the boundary conditions we need to fulfill A (T) = B (T) = C (T) = 0
in order to ensure a final discount bond price of one at maturity. With this
functional mapping of the short rate and its market price of risk we only
need to calculate the functions A (t), B (t), and C (t) in order to fully specify
the cross-section dimension of our model.
To prove the proposed solution we need to show that the discount bond
price from equation (13.1) indeed solves the partial differential equation
(12.4). Utilizing the standard separation of variables technique, as used
for example by Chen (1995), we first calculate the partial derivatives and
denote them by subindexes as
Pr := 8P(i;'x) = -AP, Prr := 82 PJ:..r,x) = A2 P,

P>. ..= 8P(t,T,x)

8>'
= -BP, > .p . - 82 P(t,T,x)
>. .- 8>.2
a d
-- B 2p , n

P, .-
r>. . -
82 P(t,T,x) p,.=
8r8>. -- ABP, t .
8P(t,T,x)
8t
= (-A t r - B t ).. - C)
t,
P

which we insert into equation (12.4) to yield

~0";A2 + PO"rO">.AB + ~O"iB2 - (/-Lr + Krr - O"r)..) A

- (/-L>. - 0">.)..2 + K>.rt + B>.)..) B - Atr - B t ).. - Ct - r = o.

We then separate for the state variables rand ).., and the constant terms:

[At+ KrA + K>.B + 1] . r

+ [Bt + B>.B - O"rA] . ).. (13.2)
+ [~0";A2 - /-LrA + ~O"iB2 - (/-L>. - 0">.)..2) B + PO"rO">.AB - Ct] = o.
In order to hold for every values of r (t) and)" (t), we need to have a value of
zero for the expressions in brackets. This leads us to a system of decoupled
partial differential equations in A (t ), B (t), and C (t):

At + KrA + K>.B -1 (13.3)

Bt + B>.B - O"rA o (13.4)
122
20"r A - /-LrA 1 2 2- (/-L>. - 0">.)..2 ) B + PO"rO">.AB - C
+ 20">.B O. (13.5)
t
13.1 Valuing Discount Bonds 107

Looking closer at the system, we fortunately see that function C only ap-
pears in the last equation (13.5) and only in its first derivative with respect
to t. Thus, we can start solving the system by examining the first two equa-
tions (13.3) and (13.4) in order to derive a solution for the functions A (t)
and B (t). In matrix notation we have

with the system matrix [TQ]' and the driving term d = [-1 0]'. The gen-
eral solution to this non-homogeneous subsystem is the sum of a particular
solution to the non-homogeneous equation (denoted by AP, and BP) and the
general solution to the homogeneous equation (complementary functions AC,
and BC):

A (t) AP + A C , and
B (t) BP + B C •

For the particular solution we can gain the constant solution by solving the
following system:

Thus, we obtain the functions

I ~1 ~A
(h
I I ~T
-(h = -aT
AP= = , and BP -aT

~A I ~T(h + ~AaT
;~ I
~T ~T

I -aT BA I -aT
for the particular solution. Now, we look at the general solutions AC and
BC to the homogeneous equations

(13.6)

Here, we adopt the trial solution

108 13. Initial Characteristic Results

Putting these with the corresponding time derivatives into equation (13.6),
we get

AaeAt + Krae At + K)J3e At 0

A/3e At - arae At + O)"/3e At 0

and

+ A) . a + K),. . /3
(Kr o (13.7)
-ar . a + (0),. + A) . /3 O. (13.8)

We note that equations (13.7) and (13.8) only have non-trivial solutions, if
the characteristic polynomial is equal to zero, i.e.

I Kr-a+ A K),.
0),. +A
I= 0'
r

or

(13.9)

This results in Ai = -!
(Kr + (h) ± h/(Kr + 0),.)2 - 4 (Krfh + K),.ar ) for the
eigenvalues, with i E {1, 2}.

Remark 13.1.2 (Special Eigenvalues) We need to note the special cases

of the eigenvalues where the following calculations are not defined: (i) Ai =
0, and (ii) Al = A2 .90 However, if the parameters are estimated empirically,
there seems to be a neglectable possibility of ending up with this special cases.

Now, we can resolve the values for a and /3 from, for example, equation
(13.8) to aj/3 = (0),. + A) jar. Choosing arbitrary constants ki' with i E
{1,2}, we get ai = (0),. + Ai) ki' and /3i = arki . Thus we get for the com-
plementary functions

AC + AI) kIe A1t + (8),. + A2 ) k2 eA2t ,

(8),. and
B C arkieAlt + a r k2e A2t •

90 As we discuss in section 13.3.1, to get a solution for Ai = 0 we need to go back to

the original system of partial differential equations of (13.3) to (13.5).
13.1 Valuing Discount Bonds 109

For an intermediate result, we now combine the particular and the comple-
mentary solution for the functions A (t) and B (t). Making use of Vieta's
Theorem, by which the product of the eigenvalues Ai, with i E {I, 2}, equals
the constant term "'rfh. + "'>/l'r in equation (13.9)91, we obtain:

A (t) (13.10)

B (t) (13.11)

In order to derive the constants k1 and k2' we now make use of the first two
boundary conditions A (T) = 0 and B (T) = O. For equations (13.10) and
(13.11) we get in matrix form:

with the solutions

...!!..l......
A1A2
(h +A2
1 1
A1A2 -1
k1e A1T
(h. + Al 0>. +A2
= Al (Al - A2 )
1 1
-1
k1 = A1 (Al - A2) e-A1T , and

1
(h +A1 (h +A2
1 1
1 -A2T
k2 = A2 (A1 _ A2) e .

9 1 This equivalently is the determinant of the system matrix.

110 13. Initial Characteristic Results

This finally results in the values for the functions A (t, T) and B (t, T):92

A (t, T) ao + a 1e-A1(T-t) + a 2e-A2 (T-t) , (13.12)

0),. 0),. +Ai 0),. +A2
with ao
- AiA2'
ai - -
- Ai(Ai-A2)'
a2 = and
A2 (Ai - A2)'

B (t, T) bo + b1 e-A1(T-t) + b2e-A2(T-t) , (13.13)

O'r O'r b _ O'r
with bo bi =
- AiA2' Ai (Ai - A2)' 2 - A2 (Ai - A2)"
With these results for A (t, T) and B (t, T), we can now solve equation (13.5)
for the last function ofrelevance C (t).

Remark 13.1.3 (Alternative Derivation) Instead of solving the system

of differential equations consisting of {13.3} and (13.4), we alternatively
can derive the solutions of A (t) and B (t) by reducing the two first-order
differential equations to a single second-order equation. By differentiating
equation {13.3} with respect to t, and inserting it into equation (13.4), we
can defer the following single second-order equation for A (t):

Having derived a solution for A (t), we then resolve equation (13.4) for the
function B (t). With the boundary conditions A(T) = 0 and B(T) = 0, we
finally get the same results for the functions A (t), and B (t).93

Finally, we start calculating C (t) by isolating the partial derivative Ct

in the third equation of our system (13.5):

122 122 )
Ct = 20'rA - JLrA + 20'),.B - (JL),. - 0'),.A2 B + PO'rO'),.AB,
where we now enter our results of equations (13.12) and (13.13) for A (t)
and B (t). A straightforward calculation gives us in a nice rearrangement
92 For a better understanding of the independent variables of the functions A (t) and
B (t) we add the parameter T to the final expressions, i.e. denote the functions as A (t, T)
and B(t,T).
93This alternative derivation is of interest in terms of studying the equilibrium behav-
ior of the system, because it nicely relates to the nonhomogeneous partial differential
equation which is known from physics to model oscillation. However, our approach is
more symmetric in its formulation and derivation of the solutions to A (t) and B (t).
13.1 Valuing Discount Bonds 111

with constants Ci, i E {O, 1, ... , 5}:

Ct = eo + cle-A1(T-t) + ~e-A2(T-t)
+C3 e - 2A1 (T-t) + C4 e - 2A2 (T-t) + cse-(Al+A2)(T-t) (13.14)

with eo '122
2 urao - J.trao + '2122
u ),bo - (J.t), - U),A2 ) bo + paru),ao b0
CI = (u~ao - J.tr + paru),bo) al + (u~bo - (J.t), - U),A2) + parU),ao) bl
C2 = (u~ao - J.tr + puru),bo) a2 + (u~bo - (J.t), - U),A2) + puru),ao) ~
1 22 1 22
C3 = '2 Ural + paru),a1bl + '2 U),bl
1 22 1 22
C4 = '2 Ura2 + purU),a2 b2 + '2 U),b2
Cs U~ala2 + parU)' (al~ + a2bl) + U~bl~'
Finally, in view of the boundary condition C (T) = 0, we integrate
equation (13.14) to derive the result for the function C (t)

-C (t) = Jeo
T
+ cle-A1(T-() + C2e-A2(T-() + C3 e- 2A1 (T-()
t
+C4 e - 2A2 (T-() + cse-(A1+A2)(T-()d(,
which yields: 94
C (t, T) = -eo(T - t) + ~ (e- A1 (T-t) - 1) + ~ (e- A2 (T-t) - 1)
Al A2
+~ (e- 2A1 (T-t) _ 1) + ~ (e- 2A2 (T-t) - 1)
2AI 2A2
+ Cs (e-(Al +A2)(T-t) _ 1) , (13.15)
Al +A2
where the constants Ci, with i E {O, 1, ... , 5}, are given in equation (13.14).
Solution 13.1.4 (Discount Bond Price) Putting our results together, we
are able to state the price of a discount bond in the closed-form expression
P (t, T, x) = e-[A(t,T),B(t,T)]x-C(t,T) (13.16)

where we use the formulas for the functions A (t, T), B (t, T) and G (t, T)
from equations (13.12), (13.13), and (13.15), respectively.
94We also add the parameter T to the final expression on the function C (t) for a better
understanding, i.e. we denote the function as C (t, T).
112 13. Initial Characteristic Results

Moreover, with the explicit formulas of the functions A (t, T), B (t, T)
and C (t, T) we can readily calculate the risk neutral dynamics of the dis-
count bond price. We begin with the P-dynamics under the martingale
measure Q of equation (12.3) which generates the stochastic form:

dP
[( -At - ~rA - ~).B) r + (-Bt - (hB + O'rA) A
P
122 - j.£rA + 1
+20'rA 20'),.B2- 2 ()
j.£),. - 0'),.A2 B + PO'rO'),.AB - Ct] dt

-O'rAdWZ - O'),.BdW~t·

In conjunction with the results for A (t, T), B (t, T), and C (t, T) the stochas-
tic dynamics for the discount bond simplifies to

dP = j.£pPdt + 1]rPdWZ + 1]),.PdW~t, (13.17)

with j.£p = r, 1]r = -O'rA, and 1]). = -O').B.

Equation (13.17) shows explicitly that the modeled discount bond P (t, T, x)
actually has a drift rate of r in the risk neutralized world.

13.2 Term Structures of Interest Rates and

Volatilities
On the basis of the main result in the previous section - the formula for the
value of a discount bond - we are now able to explore the cross-sectional
dimensions of our interest rate model in detail. We examine both the term
structure of interest rates as well as the term structure of volatilities in the
following.

13.2.1 Spot and Forward Rate Curves

From the general functional relationship between the value of discount bonds
and spot rates in section 11.2 we accomplish to determine the yield curve
implied by our term structure model. We conveniently compute the yield to
maturities y (t, T) of different discount bond prices by the affine relationship
between the zero bond yield y (t, T) and the state variables r (t) and A (t)
13.2 Term Structures of Interest Rates and Volatilities 113

with the relationship from equation (11.1)

1
Y (t, T) = - T [[A (t, T), B (t, T)] x + C (t, T)]; (13.18)
-t

the factor loadings A (t, T) and B (t, T) are given in equations (13.12) and
(13.13), and the intercept term C (t, T) is from equation (13.15). Especially
this affine structure of the yields in the driving factors - the short rate r
and the market price of risk A(t) - will be of special importance for the
empirical implementation and analysis which we perform in chapter 15.
Further, it is of interest how the forward rate curve looks in this model.
Starting from equation (11.2) we can see that forward rates are related to
spot rates by

rp) = Y (t, T2) (T2 - t) - Y (t, T1) (Tl

f (t, T b.L2 - t)
rp T . (13.19)
.L2 - 1

Thus, the term structure of forward rates can immediately be derived from
the zero bond yield curve given in equation (13.18).

13.2.2 Term Structure of Volatilities

Besides the well known meaning and usage of the phrase term structure
as the term structure of interest rates - as in the previous section - it is
also used in the context of volatilities. The term describes the functional
relationship of either the volatilities of the spot or forward rates over the
spectrum of different time to maturities. This second meaning probably
stems from the importance of volatilities in option pricing as we will see
further down in chapter 14.
In order to derive the volatility structure of the spot mtes we first need
to determine the dynamics of the T-maturity spot rate y (t, T). In our
specification of the term structure with the risk neutral factor processes
given in equation (12.2), we attain the following stochastic differential on
y (t, T) by applying Ito's formula:

dy (t, T)

with Jl. y
114 13. Initial Characteristic Results

Therein, the partial derivatives Yr, Y).., Yt, Yrr, Yr).., and Y)..).. are easily found
from the result of equation (13.18). Using these derivatives leads us to

dy (t, T) ( ILl A(t,T)

T_t + IL2 T _ t
B(t,T)) d
t
A(t,T)dWQ B(t,T)dTrrO
+0"r T _ t r,t + 0").. T _ t VV )..:t· (13.20)

To further calculate the volatility structure of spot rates, we use the

integrated form of equation (13.20) over the period [t, t + .6.t] to yield

T)
Y (t + .6.t, T + .6.t) Y (t, + (A(t,T) B(t,T)).6.
ILl T _ t + IL2 T _ t t

+ J
Hll.t
A (t,T)dW Q
O"r T _ t r,(+
J
Hll.t

where the two stochastic integrals follow a normal distribution. Thus, the
spot rate Y (t, T) follows a bivariate Gaussian distribution. From this dy-
namics we obtain the variance per unit of time, i.e . .6.t = 1, as:

VAR [y (t, T)] ]EQ [Y;IFs] -]EQ [YtIFs]2

( O"r A(t,T))2 ( B(t,T))2

T -t + 0").. T - t
A (t, T) B (t, T)
+ 2pO"rO").. 2' (13.21)
(T - t)

which describes the term structure of spot rate volatilities.

One important capability of a two-factor model is its implication of an
imperfect correlation of spot rates of different maturities. Therefore we
calculate the covariance between the interest rates of discount bonds with
13.2 Term Structures of Interest Rates and Volatilities 115

different maturities:

= EQ [O"rA (t, T1) d~ + O")..B (t, T1) d~t

T1 - t

x O"rA (t, T2) d~ + O")..B (t, T2) dW~tl

T2 - t

[ O"~A (t, T1) A (t, T2) + O"~B (t, T1) B (t, T2)] dt
(T1 - t) (T2 - t)
+ [A (t, T1) B (t, T2) + A (t, T2) B (t, T1)]
(T1 - t) (T2 - t)

We further use the general relationship for the covariance

to reveal our final result for the correlation coefficient of the spot rates with
maturities T1 - t and T2 - t:

In analogy to the previous findings we are able to derive the volatility

structure on the forward rotes. From the result of equation (13.19) we obtain
the stochastic dynamics

dlf(t,T) = ( A(t,T) + B(t,T)) dt

1-'1 T - t 1-'2 T - t
BA (t, T) dwQ BB (t, T) dWQ
+O"r 8T r,t + 0").. BT )..,t
for the instantaneous forward rate. By using the integrated form of this
dynamics as an intermediate step, we attain the term structure of forward
rote volatilities as:

VAR [f (t, T)] = (~BA(t,T))2 + (~BB(t,T))2

T-t 8T T-t 8T
2pC7rO").. BA (t, T) BB (t, T)
+ (T - t)2 8T 8T
116 13. Initial Characteristic Results

We finally calculate the covariance between forward rates of different

maturities Tl - t and T2 - t as

to complete the analysis on the term structure of volatilities.

13.3 Analysis of Limiting Cases

13.3.1 Reducing to an Ornstein-Uhlenbeck Process

Since our model is an extended version of modeling the short rate dynamics
with an Ornstein-Uhlenbeck process, we would naturally think of comparing
our pricing formula in equation (13.16) to that of Vasicek (1977). However,
the derived pricing formula is not valid in the case where we set the time
homogeneous parameters J-t>.., 0>.., (7).., and >'2 equal to zero. Instead, as
noted in remark 13.1.2, we already need to adjust the parameters to their
appropriate values in the system of partial differential equations in (13.3),
(13.4), and (13.5). This reduces the system of equations to

(13.23)

In solving this reduced system, we use the same boundary conditions

as in proposition 13.1.1 for the affine solution P (t, T, x) = e-[A(t),B(t)]x-C(t)
where, however, the market price of interest rate risk is now a constant,
i.e. >. = const. First, we calculate the solution to the function A (t) out
of the first equation in (13.23). A particular solution is easily seen to be
AP = -1/"'r; we further use the homogeneous equation to calculate the
13.3 Analysis of Limiting Cases 117

complementary solution with integration constants k and k':

J
T T

J ~dAc
Ac = -Kr d(
t t
-lnAC(t) -Kr (t - T) +k
A (t)
C k' eK.r(T-t) .

°
Now, we add the particular and complementary solution, and incorporate
the boundary condition A (T) = to find the general solution to be

1
A (t) = -- (1 - eK.r(T-t)) • (13.24)
Kr

°
For the function B (t), we get from the second homogeneous equation in
(13.23) by using B (T) = the following result:

- CT r
Kr
J(1 -
T

eK.r(T-()) d(
t

-B (t) = _ CTr
Kr
((T _t) + ~ (1 _ Kr
eK.r(T-t)))

B (t) = - CT r [A (t) - (T - t)]. (13.25)

Kr

In order to derive the function C (t) with its boundary condition C (T) = 0,
we use the last equation of equations (13.23):

-C(t)

C (t)
118 13. Initial Characteristic Results

To see the equivalence between the bond price formula from Vasicek
(1977) and our solution

..L (l_e"r(T-t») ~[A(t)-(T-t)J]x+(b:+4) ((T-t)-A(t))+~A2(t)

P (t , T , x) = e [ "r ' "r "r 2"r 4"r ,
we conveniently change the symbols for the used parameters. The original
notation of the Ornstein-Uhlenbeck process used by Vasicek (1977, p. 185)
is

dr = a b - r) dt + dz.
With the substitutions -Kr = a and -J-lr/Kr = 'Y we are easily able to
transfer our results into that found by Vasicek (1977).

13.3.2 Examining the Asymptotic Behavior

When looking at possible shapes of the term structure of interest rates two
points of the term structure are of special interest: The short end and the
long end. These are given by two extreme values for the maturity parameter
T. We examine the behavior of the yield function from equation (13.18) for
the case when T approximates t from above, and for the case where T
approaches infinity.
For theoretical reasons the short end of the term structure should co-
incide with the instantaneous rate of return r (t). This implies that the
functions A (t), B (t), and C (t) should have limit values of

lim A (t) = 1, and lim B (t) = lim C (t) = O.

T!t T - t T!t T - t T!t T - t
For calculating the limit values we use the method of Taylor series ex-
pansion. In our case a first order approximation is sufficient. For the limit
values of the functions A (t), B (t), and C (t) we derive:

limA(t) 1 [ao+ad1- A d T -t)+o(T-t))

-T
T!tT - t - t
+a2 (1 - A2 (T - t) + 0 (T - t))]
_1_ [-rho
(Al - A2) - (0).. + A1) A2 + (0).. + A2) Al
T- t A1A2 (Al - A2)
+ 8).. + A~l-_(~: + A2) (T - t)] = 1, and
13.3 Analysis of Limiting Cases 119

lim B (t) =
TltT - t

~w~ ~~ = T ~ t [-Co (T - t) + ~ll (1 - AdT - t) + 0 (T - t) - 1)

+~: (I-A2(T-t)+o(T-t)-I)
+ 2~l (1- AdT - t) + o(T - t) -1)

+ 2~2 (1 - 2A2 (T - t) + 0 (T - t) - 1)

+ Al : A2 (1 - (Al + A2) (T - t) + 0 (T - t) - 1)]

= - [Co + Cl + C2 + C3 + C4 + cs] = o.

By having calculated these limiting values for the functions A (t), B (t), and
C (t) we indirectly verified our pricing formula of equation (13.16) in that
we obtain

Ys (t)
T!t - tT 1 - [[A (t), B (t)] x + C (t)] = r (t)
= limy),t, T) = l i m
T!t - t

for the short end of the term structure of interest rates. Thus the short end
coincides with the first state variable of our term structure model.
Now, we analyze the asymptotic behavior of the term structure of in-
terest rates at its long end. For the long end we need to investigate the
limiting value

lim y (t, T) = lim - T

T-+oo
1 [[A (t) B (t)] x
T-+oo - t
+ C (t)] .
120 13. Initial Characteristic Results

Again, we examine the three functions A (t), B (t) , and C (t) separately:

Thus, we obtain the constant for the long end of the term structure

Yl (t)

which is independent of the state variables x = [r (t) ,A (t)]'.

13.4 Possible Shapes of the Term Structures

From the various term structures we derived in section 13.2 we consider the
effects on the two special cases of the term structure of spot interest rates
as well as the term structure of volatilities for spot interest rates. Our two-
factor model endogenously produces these term structures determined by
equations (13.18) and (13.21). Both term structures depend on the market
expectations of the two state variables, as inherently modeled by the spec-
ifications of the stochastic processes under the measures JP> and Q. These
process specifications, respectively, model the dynamics of the state vari-
ables through time and determine the term structure relationship between
the interest rates of different maturities for a given date. Furthermore, the
term structure of spot interest rates is also a function of the current levels
of the two state variables - the short rate r and the market price of risk A -
which we therefore also need to include in our comparative statistics.
13.4 Possible Shapes of the Term Structures 121

13.4.1 Influences of the State Variables

We first examine the influences of the state variables on the term structure of
spot interest rates. As determined by equation (13.18) the term structure
is influenced both by the current level of the short rate rt and the value
of the market price of risk At whereas the term structure of volatility is
independent of the state variables. 95
For the exposition of the comparative static analysis we choose to use
the empirical estimates for the parameter values 1/J reported in table 15.5
as basis. 96 Given these parameter values we work with the specification

dr = [ltr + K,rr - <TrA] dt + <TrdWZ

= [0.0071 - 0.0744r - 0.0225A] dt + 0.0225dWZ, and (13.27)
dA = [itA - <TAA2 + K,Ar + OAA] dt + <TAdW~t
[-0.0441 - 0.6198r - 0.4434A] dt + 0.3863d~t, (13.28)

for the stochastic processes for the state variables including a correlation
coefficient of p = 0.7102 between the two Brownian motions. For the cur-
rent values of the state variables we use rt = 0.0699 and At = -0.2624,
respectively, which reflect the empirical mean values of the realized state
variables. 97
How the term structure of spot interest rates reacts to different levels
of the state variables is shown in figure 13.1. The left graph is based on
a market price of risk of At = -0.2624 and models the impact of short
rates in the range rt = {0.02, 0.04, ... ,0.18}. The different values for the
short rate can be traced back to the crossing points of the yield curves
with the ordinate as demonstrated in section 13.3.2. All yield curves are
a little upward sloping since the market price of risk (At = -0.2624) lies
below its risk neutral value of At = O. The changes in the short rate have
the impact of an approximately parallel shift of the yield curve. The effect is
slightly higher on the short end of the yield curve than for longer maturities.
Therefore, the first state variable can be attributed as a level factor.
95See equation (13.21).
96The empirical parameter estimates will be considered in depth in chapter 15 on the
calibration of the model to US interest rate data.
97The empirical mean values are those obtained for the US Treasury data set shown
in table 15.7 on page 168.
122 13. Initial Characteristic Results

0.11
0.18 r- values: A. - values: ·1.0

0.16 0.10

0.14 -0.5
0.14 0.09

0.12
0.10
0.10
0.07
0.08 0.06

0.06 0.06
0.02
0.04
0.05
0.02
0.04
0 2 4 6 8 10 0 2 4 6 8 10

Figure 13.1: Influence of the State Variables on the Term Structure

The graph on the right side of figure 13.1 underlies a value of Tt =

0.0699 for the current short rate. Depicted are yield curves for varying
risk attitude values At. To represent the spectrum of risk averse, risk neu-
tral, up to risk taking investors we choose values in the range of At =
{-1.0, -0.5,0,0.5, 1.0}. These values are obtained on the basis of empiri-
cal observable quantities with estimates on the US term structure. 98 For a
constant short rate Tt the different levels of the market price of risk have
a significant effect on the slope of the yield curve. For the underlying sce-
nario we obtain normal shapes of the yield curve with negative values for
the market price of risk, i.e. risk averse market attitudes towards interest
rate risk. This is in accordance with the intuition that more risk averse
investors require a higher return on their fixed income investments. Inverse
yield curves on the other hand are attainable for positive market prices of
risk. We recognize that the level of the market price of risk is the major
factor determining the slope of the yield curve. In the conventional sense
the second factor can therefore be considered as the steepness factor.
Combining variations in both state variables, while holding the model
parameters 1/J fixed, results in a rich complexity of different shapes of the
yield curve. This flexibility is rather relevant in recalibrating the model
98S ee the empirical results shown in table 15.7 on page 168.
13.4 Possible Shapes of the Term Structures 123

from time to time with a close fit to the then prevailing term structures
without reestimating the whole range of model parameters 1/J.
The next subject of analysis is the term structure of volatilities. Since
the volatility curve is independent of the levels of the state variables, we
now move to the influences of choosing particular model parameters on the
term structures.

13.4.2 Choosing the Model Parameters

In this section we establish comparative statistical facts on the behav-
ior of the two term structures for different values of the model param-
eters. The parameters of the theoretical term structure model include
1/J = {ltr, /'i,r, O"r, It)", /'i,)", 0")", (1).,, >'2, pl· The range of possible values for the
parameters is found on the ground of the empirical results on the US term
structure data. 99 The analyzed range includes all empirical values as well
as some higher and lower values. The standard parameter values are the
same as in the previous section.
We first examine the behavior of the term structure of spot interest
rates. The possible term structures for the nine parameters are plotted in
figure 13.2. The graphs are further based on the values of rt = 0.0699
and >'t = -0.2624 for the state variables. Since the rt-value lies below
its long-run mean of rl = 0.0950 and the risk attitude is risk averse the
resulting yield curve shapes are all upward sloping. lOo These values for the
state variables correspond to the scenario 'risk averse' and 'r < rl' reported
in the eighth column of table 13.1. In the table we present the effects of
positive changes in the parameter values for all possible scenarios of the
state variables. When a rise in parameter values has an increasing effect on
the yield curve it is denoted by '11"; decreasing and neglect able effects are
symbolized by '.u..' and ':::::', respectively.
In explaining the contents of table 13.1 we begin with the influences
of the parameters Itr' It)", and >'2. We consider these parameters as level
parameters since they show up as additive constants in the drift terms of
equations (13.27) and (13.28). They all have a constant influence on the
99S ee the results given in table 15.6 on page 163.
lOaThe long-run means of the state variables in the data set are Xl = [0.0950, -0.1311]'
which are based on the US interest data; for further analysis of the state variables see
section 15.3.3.
124 13. Initial Characteristic Results

0.12 1', - values: 0.10 0.084

0.11 -0.015 0.09
K, - values: cr, - values:
0.012/0.01/0.008 0.080
0.10 -0.005 0.08
0.07 0.076
0.09
0.06
0.072 0.021 0.015 I 0.01
0.05
-.(l.25 -0.005
0.04 F====!'~~"""""'---'8'---~10 0·068±-0-'2-~4-~6'-----'8'---~10
8 10 0 2 4 6
0.100
0.095 'S. - values:

0.0008 I 0.001 I 0.0012

-0.0015 -0.7
0.068
0 2 4 6 8 10 0 2 4 6 8 2 4 6 8 10
0.092 0.11 0.084
0.088 a, - values: A2 - values:
0.10
-0.25
0.09

--0.1 0.08
.(l.21 .(l.31 -0.4
0.07 -.(l.5
0.068
2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10

Figure 13.2: Analysis of the Term Structure of Interest Rates

yield curve throughout the different scenarios. The parameter /-Lr gives the
short rate higher presumed values in the future in thus it has an increasing
effect on the yield curve. The additive terms in the process of the market
price of risk are /-LA and A2. Numerically positive changes in /-LA result in
lower yields, but within the empirically estimated values for the parameter
we only observe neglect able effects. The opposite influence is attained by
A2 since it enters the stochastic process with an opposite sign. The values
for these parameters both make the expected value of the market price of
risk either rise or fall. This further gives a feedback on the short rate which
is lower under the measure Q. Thus, higher values for the market price of
risk lower the term structure of interest rates as we demonstrated in the
previous section.

The next parameters we analyze are the direct factor loadings "-r, "-A,
and 0>.. in the stochastic processes of equations (13.27) and (13.28). With
respect to the short rate higher values of "-r assume higher expectations of
the investors in the future in that it has an increasing effect on the yield
curve. Only in the case of high values of A we observe a contrary reaction.
13.4 Possible Shapes of the Term Structures 125

Table 13.1: Effects ofIncreasing Parameter Values on the Yield Curve

Model Market Price of Risk

Para- Risk Taking (A > 0) Risk Neutral (A = 0) Risk Averse (A < 0)
meters Short Interest Rate
1/J r < rl1 r = rll r > rl r < rl I r = rl Ir > rl r < rl Ir = rl Ir > rl
f-lr li
lir il 1 li I li li li
ar il il 1 ::::::l
I li li
f-l>.. ::::::l

Ii>.. li I il I il il il
a>.. il
fh. il il 1liil I li li
A2 li
p li
Effects on Yield Curve: li: increasing il: decreasing ::::::l: neglect able

The market price of risk is directly influenced by the parameters Ii>.. and
0>...With positive changes in Ii>.. we generally lower the expectations in the
market price of risk which makes the investors require higher returns in the
future. This results in an increase of the yield curve. The impact of the
level of the market price on itself is modeled by 0>... Its influence on the yield
curve varies in dependence of the sign of At, i.e. the current risk attitude of
investors. In risk taking scenarios we attain a negative influence and with
risk averse investors we obtain rising yield curves. For risk neutral investors
with the current short rate being at its long-run level the short rates are
pushed up whereas the long rates are lowered.
Moreover, the state variables are modeled with a possible correlation
factor p. With high values for this parameter it is more likely to observe
high values of the short rate whenever the risk attitude is high. It only
influences the term structure of interest rates by the term C (t, T) according
to equation (13.18). Through all different scenarios of the state variables we
observe increasing yields with positive changes in the correlation coefficient.
Finally, we examine the influences of the volatility parameters a r and a>...
126 13. Initial Characteristic Results

0.030 0.030 0.030

or - values:
Kr - values:
0.025 0.025 --0.025 0.025
---0.05
0.02 / 0.015 / 0.Q1
--0.005 0.020

0.Q15 0.015

0.010 K, - values:
---2
0.005 0.005 -1.5/-1/-0.5
--0
0.0000
2 4 6 8 100 .0000 2 4 6 8 100 .0000 2 4 6 8 10
0.030 0.030 0.030

0.025 0.025

0.020

0.015

0.010 01. - values: 0.010 S, - values: 0.010

p - values:
--0.3 ---0.1 ---0.5
0.005 0.4 / 0.5 / 0.6 0.005 -0.2 / -0.3 / -0.4 0.005 -0.25 / 0 / 0.25
--0.7 ---0.5 --0.75
0.0000 0.000
2 4 6 8 10 0 2 4 6 8 100 .0000 2 4 6 8 10

Figure 13.3: Typical Patterns of the Term Structure of Volatilities

The volatility of the short rate does not directly impact the expected value
of the short rate. However, with higher values for a r risk averse investors
need to be compensated with higher returns. Thus, the yield curve increases
with higher ar-values. Risk taking investors on the other side love to have
more volatility in that they are satisfied with lower returns. The volatility
a>. on the risk premium models the frequency of changing risk attitudes of
the investors through time. Thereby, high fluctuations result in lower levels
of the yield curve.
Now, we move on to consider the comparative analysis of the term struc-
ture of spot rate volatilities. The possible cases given by the standard pa-
rameter values for the possible cases of volatility structures are shown in
figure 13.3. From equation (13.21) we see that the exogenous parameters
that influence the volatility structure are given by the reduced number of
't/J = {KT) aT) K>., a>., (h, p}. Thus, aside from the state variables the level
parameters J.Lr' J.L>., and '\2 do not impact the volatility structure as it is
expected from finance theory.
The volatility structures show us how the short rate volatility a r is
13.4 Possible Shapes of the Term Structures 127

transformed and extrapolated towards yields of higher maturities. At the

ordinate we coincide with the value of the short rate volatility O"r = 0.0225.
This fact can conveniently be seen from the O"r-graph in figure 13.3 as the
levels of the volatility curves change with different values for O"r. Herein,
the volatilities for the ten year rate range between 51 to 55 percent of the
short rate value, i.e. the relative volatility is about 50 percent.
All other parameters examined show a positive effect on the yield curve.
With positive changes in the parameter values the volatilities are shifted
upwards for longer maturities. In the fourth graph we observe that we
are able to let the relative volatility decrease faster over the spectrum of
maturities by altering the values for the volatility 0">. of the market price
of risk. The last two graphs show volatility structures that increase with
higher time to maturities which is a rather unrealistic phenomenon when
compared to empirical findings. However, these patterns are obtained with
altering the values for the parameters (h, and p and at the same time holding
the other parameters constant at their standard values. This is only done
for explaining comparative results and is not observable in the empirical
results. 101

101 Compare the results presented on the empirical estimations in chapter 15.
14 Risk Management and
Derivatives Pricing

14.1 Management of Interest Rate Risk

After examining the implications of our theoretical term structure model
on the different term structures that are relevant in pricing term structure
derivatives we now answer the question on how the model can be used in risk
management. The management of interest rate risk is especially concerned
with rebalancing a fixed income portfolio exposed to interest rate risk due
to the desired risk return characteristics. The types of interest rate risk
considered to be of relevance in interest rate risk management are generally
considered to be: (i) Market risk is the risk of changing prices due to general
changes of the overall level of interest rates on default free securities, (ii)
yield curve risk is considered the risk associated with non-parallel shifts
in the yield curve, i.e. a reshaping of the yield curve due to, for example,
steepening, flattening, or twisting, and (iii) credit risk which is related to
altering security prices caused by changes in the creditworthiness of the
issuer.
The oldest risk, risk managers have aimed at, is the market risk for
which the conventional concepts of duration and convexity enabled them
to successfully immunize their portfolios against market risk. 102 The yield
curve risk asked for better measures of interest rate risk beyond capturing
infinitesimal parallel shifts in the yield curve. Here, especially, the key rate
durations ofHo (1992) gained popularity.103 However, to cope with this risk

l02See, for example, Macaulay (1938) and Bierwag, Kaufman, and Toevs (1983).
l030ther recent concepts include the functional duration from Klaflky, Ma, and Nozari
(1992) and the partial duration by Waldman (1992).
130 14. Risk Management and Derivatives Pricing

type managers already use interest rate derivative securities such as swap
contracts, or cap and floor agreements. Lastly, the credit risk is the newest
risk exposure fixed income managers start to deal with.104
The risks captured by our proposed term structure model include the
market and the yield curve risk whereas it is not capable to incorporate
any risk associated with default risky securities. As we analyzed in section
13.4 the first model factor can be seen as level factor and thus attributed to
describe general market movements of interest rates. With the market price
of risk factor we additionally consider the steepness of the term structure of
interest rates and thereby capture the yield curve risk. The two factors can
successfully be applied to the concept of duration using factor durations to
manage the market and yield curve risk.

Definition 14.1.1 (Factor Duration) 105 Given the price of a coupon

bond pc (t, T) the factor dumtion is defined by
1 apc (t, T)
Dp =
pc (t, T) of
1 ~ ()aP(t,u)
= pc (t T) L..J c u of
, UE(t;T]

for a factor F.

With respect to the two factors in our term structure model we get the
factor durations

Dr = - 1 ~ c(u)aPC(t,u,x) and (14.1)

pc (t, T, x) L..J(T]
uE t;
or'
D).. - 1 2: c(u)aPC(t,u,x) (14.2)
PC (t,T,x) uE(t;
T] a)..
for the short rate r and the market price of risk)". The factor durations
measure the sensitivity of bond prices to the factors. In the case of a
discount bond with c(T) = 1 we get the measures of risk exposure as Dr =
104For a comparative analysis of current credit risk models used as internal models by
banks see, for example, Crouhy, Galai, and Mark (2000).
105See, for example, Wu (1996) and Chen (1996a).
14.2. Martingale Approach 131

A (t, T), and D>. = B (t, T), where the functions A (t, T) and B (t, T) are
from equations (13.12) and (13.13).
They duration measures of equations (14.1) and (14.2) can directly be
derived from the stochastic dynamics of coupon bonds. Starting with the
infinitesimal time changes of a discount bond P (t, T, x) = exp[-y (t, T)
(T - t)] based on equation (12.1) using 1M's formula yields:

dP(t, T,x)
(14.3)
P(t,T, x)

with J.£p =
1 [1
P(t,T, x) 20'r
2{P P (t, T, x)
or2 +PO'rO'>.
{P P (t, T, x)
ora)'
1 202p(t,T,x) ( ) oP(t,T,x)
+20'>. 0),2 + J.£r + Krr Or
(
+ J.£>. + K>.r
) oP (t, T, x) oP (t, T,
A), + at
X)]
' an
d

1 [OP(t,T,X) OP(t,T,X)]'
"1 = P(t, T, x) O'r Or 0'>. A),

To obtain the dynamics of coupon bonds we use equation (14.3):

(14.4)

where we already substituted for the duration measures of equations (14.1)

and (14.2). From equation (14.4) we see that investors who want to im-
munize their portfolios against market and yield curve risk will construct
a portfolio that has the same factor durations as the relevant fixed income
benchmark.
In the next sections we show how to price selected derivative contracts
relevant in risk management: Options on discount bonds, interest rate swap
contracts, and interest rate cap and floor agreements. Before that we intro-
duce the martingale approach on which our pricing derivations are based.

14.2 Martingale Approach

With the value of a pure discount bond we have priced the first interest
rate security in section 13.1. Based on this price we have further derived
132 14. Risk Management and Derivatives Pricing

the various term structures and correlation interdependencies of the yields

that were relevant in order to characterize the term structure model. Now,
we want to examine the prices of further specifications of derivatives on the
term structure. Therefore, we use the method of martingale pricing which
will be proved to be a powerful tool in pricing derivative securities. 106
In the case of discount bonds we obtain an explicit closed-form solu-
tion to the fundamental partial differential equation of equation (12.4); the
specific boundary value was set to P (T, T, x) = 1. To price other contin-
gent claims we appropriately need to change the boundary condition. The
different solution to the differential equation then reflects the specified con-
tingent claim.lo7 However, instead of deriving the value for the contingent
claim from the partial differential equation we can equivalently obtain the
value by calculating an expectation. This stochastic representation of the
solution to the partial differential equation is known as the Feynman-Kac
formula named after the contributions of Feynman (1948) and Kac (1949).

Theorem 14.2.1 (Feynman-Kac) 108 Under the relevant conditions, we

can derive the arbitmge-free value of an contingent claim, denoted by II(t, T,
x), by calculating the expectation under the martingale measure Q

_
II (t,T,x) -E
Q
[
e
- {r()d(
T II(T,T,x) Ft 1
, (14.5)

where the chamcteristic boundary condition is given by the time T value of

the contingent claim, i.e. by II (T, T, X).109

The peculiarity about the calculation of this expectation is that it is

done not with respect to the original measure JP> but with respect to the
equivalent martingale measure Q. The relevant martingale measure is found
106For a broad variety of applications of this method see, for example, Musiela and
Rutkowski (1997).
107 A method studied extensively for different contingent claims by, for example,
Wilmott, Dewynne, and Howison (1993).
108For the complete definitions and the relevant conditions see, for example, Duffie
(1996), Karatzas and Shreve (1991), and Oksendal (1995).
109For example, the previously derived discount bond formula is the solution to

I&Q [ exp - I
r (() d(1 FJ
14.3. Bond Options 133

by applying the right numeraire asset which makes the contingent claim a
martingale. In evaluating the prices of equity options one conveniently
works with the money market account B (t, T) as numeraire. 110 However,
in pricing term structure derivatives we further use the forward martingale
measure which we describe and use in depth further on. In the coming
sections we demonstrate how to value interest rate derivatives based on the
martingale approach.

14.3 Bond Options

In this section we examine the most basic of all derivatives on discount
bonds, namely call and put options. After defining the specific contract of
a European call option, we derive its todays fair value. Finally, we transfer
the pricing result of call options to value put options which can be easily
obtained using a theoretical parity relation between call and put prices.

Definition 14.3.1 (European Call Option) A European style call op-

tion with expiration date T on a zero-coupon bond which matures at time
U 2:: T is defined by the contingent payoff
II (T, T,x) = (P (T, U) - K)+,
which is the value of the call option CT at time T. It can only be exercised
at the expiration date T. 111

From this definition we know the terminal condition of the derivative

contract which we use in conjunction with the general value of a derivative
contract as given in equation (14.5). In order to get a value of the call at
a date prior to the expiration T of the option, we need to calculate the
discounted expected value of the option under the equivalent martingale
measure Q, i.e.

(14.6)

110This results in discounted asset prices being martingales with respect to Q; see
Harrison and Pliska (1981) and Geman, Karoui, and Rochet (1995).
111 From now onwards we drop the notation for the x-dependence of discount bonds,
i.e. we denote their price by P (t, T).
134 14. Risk Management and Derivatives Pricing

Now, we use the fact that a call option is a piecewise linear instrument,
which gives us the possibility of using indicator functions from probability
theory within our specification in equation (14.6). This yields

lffiQ [ e
- J r(()d(
tT P (T, U) l{p(T,u)2:K} F t 1
- KlffiQ [ e
- J r(()d(
tT l{p(T,u)2:K} 1
Ft . (14.7)

However, we cannot split the terms in the first expectation since they are
not independent under Q. To further analyze the expectations, we use
the powerful idea of choosing the optimal numeraire in the following when
pricing and hedging a given contingent claim as described in Geman, Karoui,
and Rochet (1995).112 They change measure from the accumulator measure
in use to a new equivalent forward measure by introducing a discount bond
as a practical numeraire.

Definition 14.3.2 (Forward Martingale Measure) For a forward mar-

tingale measure QT, equivalent to the measure Q, there exists a random
variable (T, for any random variable x, 'l11ith the property

J
n
x(w)dQT(w) = J
n
x(w)(TdQ(w).

°
The function (T, often denoted by ~, is the Radon-Nikodym derivative of
Q 'l11ith respect to QT defined on F t for ~ t ~ T. The Radon-Nikodym
derivative ~ is given by the ratio of the numeraires N in use, i. e. (T =
NQT / NQ. For the old numeraire we use the money market account

NQ = B(O,t)
B (0, 0)
'l11ith B (0,0) = 1, whereas the numeraire for the measure QT is given by the
discount bond maturing at date T
NQT = P(t,T)
P(O, T)'
112 The technique of changing numeraire was developed by Geman (1989) for the general
case of stochastic interest rates. Jamshidian (1989) implicitly uses this technique in the
Gaussian interest rate framework of Vasicek (1977).
14.3 Bond Options 135

which is normalized to force a unit value at time t = 0.

Conveniently, we now change measure from the risk adjusted measure
Q to the equivalent forward martingale measures QU and QT, which yield
the change of measure or likelihood ratio processes

lffi
Q [d QU I ] _ B (0,Pt)(t,PU)(0, U) , and
dQ F t -

rTt = Q [dQTI ] _ pet, T) .

':. lffi dQ F t - B (0, t) P (0, T) ,
these stochastic processes are martingales with respect to the measure Q.
For the further calculations we make use of the following abstract version
of the Bayes's formula, the 'Bayes's formula for change of numeraire'.
Theorem 14.3.3 (Bayes's Formula) 113 For a process Y adapted to F t
and the Radon-Nikodym derivative dQd dQ2 = ( we have
lffiQl [YI F,j = lffiQ2 [Y(I Ftj = ~lffiQ2 [Y(I F,j
t lffiQ2 [(I Ftl (t t

or equivalently
lffiQl [Ycll F,]
lffiQ2[YIF,j= t =rlffiQ1 [yr- l lF,]
t lffiQl [cll F t ] ':.t ':. t

under the usual conditions.

Coming back to equation (14.7) we calculate the two expectations sepa-
rately. For the first expectation we use the Radon-NikodYm derivatives t:;Y
and (¥ to change measure from Q to the U - forward measure QU:

Q
lffil =
U QU [ e - J
(t lffi t
r()d( P (T, U) (TU- 1 l{P(T,U)~K} F
t
1
= P(t,U) lffiQU [ -Ir()d(p(T U)
B (0, t) P (0, U) e ,

B(O,T)P(O,U)
P (T, U) l{P(T,U)~K} F t
I]
= P(t,U)P(O,U)lffiQU [ -lr()d(B(O,T)l F,]
P (0, U) e B (0, t) {P(T,U)~K} t

P(t,U)lffiQU [l{P(T,U)~K}IFt].
113See, for example, Dothan (1990, p. 288).
136 14. llisk Management and Derivatives Pricing

Similarly, to make a change from the accumulator measure to the T-forward

measure we use (f and (~:

IQ
lE2 = (tT lEIQT [ e - f r()d( (TT - 1 l{p(T,U)2:K}
t
]
Ft

P(t,T) lElQT [-fr()d(B(O,T)P(O,T) ;:,]

= B (0, t) P (0, T) e t 1 l{p(T,u)2: K} t

P(t,T)P(O,T)lEIQT [ -fr()d(B(O,T) ;:,]

= P (0, T) e t B (0, t) l{p(T,u)2:K} t

= P (t, T) lElQT [l{p(T,u)2:K} 1Ft] .

When we insert the results for the expectations lE? and lE~ back into equa-
tion (14.7), we get

Ct = I I
P (t, U) lElQu [l{p(T,U)2: K} F t ] - P(t, T) KlElQT [l{p(T,U)2: K} F t]
= P (t, U) QU (P (T, U) ;:::: K) - P (t, T) KQT (P (T, U) ;:::: K) .(14.8)

For the purpose of further calculating the probabilities in equation (14.8)

we use relative discount bond prices. For the calculation of the first proba-
bility, which is defined in the U-forward measure, we work in tenns of the
P (t, U) bond as numeraire, defining us the relative price ~u = P (t, T) /
P (t, U):

Using the explicit formula for discount bonds in equation (13.1) we calculate
the forward bond price ~T as

~T = e[A(t,U)-A(t,T)]r+[B(t,U)-B(t,T)].HC(t,U)-C(t,T).
14.3 Bond Options 137

Applying It6's formula based on the r (t) - and A(t) -dynamics of equation
(12.2) we get the dynamics of Y;T under the measure Ql:

d vT OY;T d oy[ d OY;T d \

It = 7ft t+ or r+ OA 1\

! 02Y;T d 2 02Y;T drdA ! 02Y;T dA 2

+ 2 or2 r + oroA + 2 OA 2
dy;T Q
y;T = ( ... ) dt + O"r (A (t, U) - A (t, T)) dWr

+0",\ (B (t, U) - B (t, T)) d~.

Then, under the equivalent U-forward measure Y;T is a martingale which
can be described by
(14.10)
with O"Y,r = O"r (A (t, U) - A (t, T)), and
O"y,,\ = 0",\ (B (t, U) - B (t, T)).
From equation (14.10) we see that Y;T follows a log-normal distribution.
Applying It6's formula to the function In Y;T - with partial derivatives
olny;TjoY;T = 1jY[ and o2Iny;Tjo2y;T = _1jy;T 2_ we obtain the Qlu_
dynamics as
dIn y'tT = 0 In y;T dy'T ! 0 2 In Y;T (dy'T) 2
oY{ t + 2 02Y;T t

O"y,rdW~U + O"y,,\dW~u
-"21 (O"Y,r
2
+ 2{XTy,rO"y,,\ + O"y,,\
2 )
dt. (14.11)
We now integrate equation (14.1,1) over the time period [t, Tl:

JO"y,rdW~U + JO"y,,\d~u
T T
In yj = In y;T +
t t

-"21 J T
2 + 2{XTY,rO"Y,'\ + O"y,,\d(.
O"Y,r 2

Therein, the stochastic integrals are Gaussian random variables with zero
mean. Since It6 integrals have an expectation of zero, the sum of the two
random variables possesses a zero mean. The variance of the two integrals
can be calculated with 1M's isometry.
138 14. Risk Management and Derivatives Pricing

Theorem 14.3.4 (The Ito Isometry) 114 For a function f (t, w) we have

Using this theorem we can calculate the variance as:

J
T

(O"h + 2PO"Y,rO"Y,>. + O"~>.) 2 d(

t

J0";
T

(A ((, U) - A ((, T»2

t
+2PO"rO">. (A ((, U) - A ((, T» (B ((, U) - B ((, T»
+O"~ (B ((, U) - B ((, T» d(. (14.12)
The expression for v~ gained in equation (14.12) actually is the variance of
the relative bond price y"u. Thereby, we gained all relevant information to
calculate the probability from equation (14.9):

QU (Y,fo' ::; ~) = QU (In Y,fo' ::; In ~)

QU J
(In Y,u - uy"d~ - J Uy"tiWf

1 ~ In ~)
t t

-~ u}d(

QU (J uy"dw,<'U + J uy"dJ¥1lU

2)
t t

1 1
::; In KY,.u + '2Vy .
114See, for example, Oksendal (1995).
14.3 Bond Options 139

Using the fact that the Ito integrals in this equation is Gaussian with
N (0, v~), we finally get

Using similar arguments we now calculate the second probability in equa-

tion (14.8). Here, we naturally work in terms of relative prices with respect
to the discount bond P (t, T) because of the T-forward measure QT. We
use the forward price defined by zr
= P (t, U) / P (t, T):

For the stochastic dynamics of ZY

we get the analogous expression to
equation (14.1O) by using Ito's formula which yields

dZr = {7Z,T ZTdWlQu

t T + (7z,).. ZTdWlQu
t ).., (14.14)
with {7Z,T = {7T (A (t, T) - A (t, U)), and
(7Z,).. = (7)..(B(t,T)-B(t,U)).

Comparing the result of equation (14.14) with equation (14.10), we note the
equivalence of the volatility terms just by a reversion of the sign, i.e. {7Z,T =
-(7y,).. and {7Z,T = -{7y,)... Further, we obtain

J{7z'TdW~u + J{7z,)..dW~u
T T
In Zj: In Zr +
t t

- 2"1 J T

2
(7 Z,T + 2p{7 Z,T{7 Z,).. + (7z,)..
2 d(.

Thereby, we result in the same variance for the forward bond price P (T, U)
as in the previous case of the relative price ~u, i.e. we get the identity
relationship

2
V = V 2z = v 2y .
140 14. Risk Management and Derivatives Pricing

With this result, we obtain for the probability in equation (14.13)

Restating our results, we have shown that the value of a European call
option can be calculated within our term structure framework using the
closed-form expression

Ct = P (t, U) N (dd - P (t, T) KN (d2 ) , where (14.15)

1~±12
n KP(t T)"2 v
,
v
and v
2
=
J T
2
(CTP(<;,T) - CTP(<;,u») d(.
t

In order to derive the arbitrage price of the European Put option belonging
to the same option series, we conveniently use the put-call parity relation-
ship.

Corollary 14.3.5 (Put-Call Parity) 115 The put-call parity relationship

for European options Ct and Pt on discount bonds is immediately obtained
as

P, EO [e -[.I()d( (K - P (T, U))+ .1; 1

EO [e -['I()<I( (P(T, U) - K)+ + K - P(T, U) .1; 1
Ct + K P (t, T) - P (t, U),
115The put-call parity was first described by Stoll (1969).
14.4. Swap Contracts 141

using the equality (K - P (T, U)t = (P (T, U) - Kt + K - P (T, U).

Thus, we have the price of a European put option given by

Pt = P (t, U)(N (d1) - 1) - P (t, T) K (1- N (d 2 ))

= P (t, T) KN (-d2 ) - P (t, U) N (-d1) , where (14.16)
I P(t,U) ± 1 2
n KP(t T)
,
v
2V and
T
v
2= J( G'p«,T) - G'p«,u)
)2 de.
t

The pricing formulas of equations (14.15) and (14.16) based on our two-
factor term structure model are actually in line with the Gaussian option
pricing formula of Geman, Karoui, and Rochet (1995, p. 456). Having gone
through the calculations for the two fair option prices we need to remark
the advantage of having the price of discount bonds available in closed-form
expressions. This especially eases the work with changing numeraires.

14.4 Swap Contracts

Besides the prices of put and call options one of the most important interest
rate derivative securities are interest rate swaps. In financial history, no
other market than the swap market has grown as rapidly since its initial
inception in 1981. 116 Until today the interest rate swap markets have also
grown in depth in that they offer a wide variety of swap contracts designed
to serve specific needs. The variants of swap contracts include swaps with
contingency features, such as forward swaps, swap rate locks, and swaptions,
swaps that use indices in their contracts, such as basis swaps, and yield curve
swaps, and finally swaps with varying notional principals, like amortizing
and accreting swaps.117
However, for our purposes we specialize in so called fixed-for-floating
interest rate swaps which exchange a stream of varying payments for a
stream of fixed amount payments. This type of swap contract is formalized
in the following definition in order to be exactly analyzable.
116For the origin and the development of the swap markets see, for example, Das (1994,
ch. 1).
117For variant forms of interest rate swaps see, for example, Marshall and Kapner (1993,
ch. 3).
142 14. Risk Management and Derivatives Pricing

Definition 14.4.1 (Interest Rate Swap) A fixed-for-floating interest

rate swap is an agreement whereby two parties undertake to exchange a
fixed set of payments with a spot swap rate k for a floating set of LIBOR
payments with a rate L (0, Tj- 1 , Tj ) at known dates Tj , j = 1, ... ,n. The
value of such a swap settled in arrears with a tenor of 0 = [Tj- b 1j] and a
notional principal of N is given by

n
SWPt = 8N'L)L(0,Tj- b Tj ) - k)P(t,Tj) ,
j=l

i. e. the difference of the present values of the floating and the fixed payments.

Since financial markets give quotes on swap rates, i.e. the fixed rates at
which financial institutions offer their clients interest rate swap contracts of
different maturities, we further derive a valuation formula for the fair spot
swap rate. In arbitrage-free markets the fair swap value is given by:
n
oN L (L (0, Tj- b Tj ) - k) P (t, Tj ) ~ 0,
j=1

i.e. a fair swap contract is worth zero today for both counterparties; Sub-
stituting the implicit forward rate L (0, Tj - 1 , Tj ) from equation (11.3) we
obtain

8N~(P(t,1j-d-P(t,Tj)
~ 8P(t,Tj)
-k)P( T)
t, J
°
°
n
L P (t, Tj-d - P (t, Tj ) - 8kP (t, Tj )
j=l
n n
LP(t,Tj-d -P(t,Tj ) k8LP(t,1j).
j=l j=l

The final equation can be solved for the swap rate

k = 1 - P (t, Tn)
n (14.17)
8~P(t,Tj)
j=l
14.5. Interest Rate Caps and Floors 143

of a fair fixed-for-floating interest rate swap settled in axreaxs with tenor

8. With our interest rate model we axe able to calculate the equilibrium
swap rate of equation (14.17) in closed-form. This advantage over numerical
solutions is offered since we can use the bond price formula derived in section
13.1 to obtain the required bond prices P (t, Tj).
In chapter 15 we actually implement this approach in that we use the
observable swap rates from the very liquid US swap maxkets to obtain in-
formation about the underlying term structure of interest rates. Based on
the swap rates we implement an empirical model and calibrate the model
to the information inherent in the swap rates.

14.5 Interest Rate Caps and Floors

A more complex but nonetheless very populax interest rate derivative of-
fered by financial institutions axe interest rate caps. us Such a contractual
agreement is designed to provide an insurance for a long floating rate po-
sition against rising above a certain interest rate level known as the cap
rate. The seller of the agreement is obliged to make a cash payment to the
buyer if a particulax interest rate exceeds a mutually prespecified rate at
some future date or dates. Similaxly, an interest rate floor is an equivalent
contract where the specific interest rate must be below a preassigned level
to pay cash.

Definition 14.5.1 (Interest Rate Cap and Floor) An interest rate cap
(floor) with tenor 8 = [Tj- 1 , Tj] is a series of caplets (floorlets) which re-
semble European call (put) options that are settled in arrears at maturity
dates Tj , j = 1, ... ,n, on the LIBOR interest rate L (Tj- b Tj). The payoff
characteristic of a so-called caplet (floorlet) is defined, respectively, as

Cpl (Tj, Tj ) = 8N (L (Tj- b Tj) - Kct, and

Flt (Tj, Tj ) = 8N (KF - L (Tj- b Tj ))+ ,
in the case of an in arrears settlement on a notional principal of N with a
cap (floor) rate of Kc(KF).
11 8 For the pricing of default-free interest rate caps, for example, under the direct ap-
proach see Briys, Crouhy, and Schobel (1991) and under the log-normal approach see
Miltersen, Sandmann, and Sondermann (1997).
144 14. Risk Management and Derivatives Pricing

To calculate the price of an interest rate cap, we start with valuing a

single caplet. Using the expression for the LIBOR rate in terms of dis-
count bonds from equation (11.3), we can restate the payoff of the caplet
at maturity Tj as:

Now, we discount the payoff at time Tj to calculate the value of the caplet
at time t. We conveniently work with the substitution K'C = 1 + 8Kc:

NJFl~ [e-7 r()d( (

P(Tj -
1
b Tj )
- Ke) + Ft]

NE
Q [ - Tjtr()d(
e t
(1
(7j-l, 7j) -
P
*)+ Kc P (Tj - 1, Tj ) F t
]
NEQ [e - Y,«)d( (1- KcP(T;-" T ))+ .r;] j

* Q [ - Tj.t r()d( ( 1 ) + ]
Kc NE e t K'C - P (7j-l, T j ) Ft. (14.18)

Reinterpreting the result from equation (14.18) in light of the charac-

teristic payoff for a put option we can state the price of a single caplet in
terms of the fair value of European put options as given in equation (14.16)

Cpl (t, Tj ) = KeNPt (K = liKe, T = Tj - 1 , U = Tj );

i.e. its value is equivalent to that of Ke put options on a discount bond
e,
with a strike of II K a maturity of Tj , and a delivery date Tj - 1 . Thus,
we can conveniently use our results from the previous section on options on
discount bonds to obtain the closed-form solution for an interest rate caplet

In P(t,Tj)Kc ±
P(t'Tj~)
.!v2
2 and v2 = J
t
T
(O"P(,Tj_J) - O"P(,Tj)) 2 d(.
14.5 Interest Rate Caps and Floors 145

Finally, the value of a cap which is a bundle of caplets is obtained by

the sum of all caplet values
n
Cap (t, T) = L Cpl (t, T
j ), (14.19)
j=l

i.e. it is equivalent to that of a portfolio of European put options on discount

bonds. The fair value of an interest rate floor agreement can be derived
analogously in closed-form.
15 Calibration to Standard
Instruments

15.1 Estimation Techniques for Term Struc-

ture Models
The empirical finance literature on the econometric analysis of term struc-
ture models provides a variety of estimation methods. With the objective
of building an empirical model that comes close to the theoretical model
properties we basically distinguish two approaches in the literature: The
time-series and the cross-section approaches.n 9 Both methods are easy to
implement, but suffer from the fact that they only use part of the available
information of the market for fixed income instruments in the estimation
procedure. A comparison of the two approaches in an estimation of the
two well studied models by Vasicek (1977) and Cox, Ingersoll, and Ross
(1985b) on data of the Dutch interest rates market is found in, for exam-
ple, DeMunnik and Schotman (1994). However, there is growing literature
following an advanced approach of a hybrid estimation in both, the time-
series and the cross-section domains simultaneously. These new attempts of
integrating the dynamic properties of the term structure models with their
cross-sectional implications using interest rate data of different maturities
stem from some general problems and shortcomings when estimating term
structure models.
The key difficulty in estimating term structure models is the unobserv-
ability of the underlying factors. To deal with this issue, the model factors
are usually either approximated, transformed or treated as parameters for

119For an overview see, for example, Campbell, Lo, and MacKinlay (1997, ch. 11).
148 15. Calibration to Standard Instruments

the estimation procedure: (i) The approximation method is often applied in

time-series estimation, where, for example, Chan, Karolyi, Longstaff, and
Sanders (1992) and Longstaff and Schwartz (1992) use the one month 'frea-
sury bill rate as an approximation for the instantaneous spot rate. This
approach can induce significant measurement errors, since we face a possi-
ble inconsistency that the proxy variable is not in line with the short end of
the estimated term structure. 120 Also, the instantaneous interest rate does
not depend on the risk premium while, for example, the one month proxy
rate does. Thus, with the approximation approach we need to deal with
theoretical inconsistencies in the estimation. Other works that can be sub-
sumed under this approach include Broze, Scaillet, and Zakoian (1995), the
non-parametric approach by Ait-Sahalia (1996), and the stochastic volatility
model by Andersen and Lund (1997). (ii) With the transformation method
Chen and Scott (1993), for example, use the thirteen week 'freasury bill
rate as a perfect substitute for the short rate. With at least one bond price
observed without error they are able to derive a system of equations where
they utilize the conditional density of the state variables to estimate the
parameters of a multi-factor Cox, Ingersoll, and Ross (1985b) model. Other
examples following this approach include Pearson and Sun (1994), Duan
(1994), and Duffie and Singleton (1997). (iii) Finally, we can treat the spot
rate as an unobserved parameter among other parameters. This approach
is close to the procedure of inverting the yield curve to extract implied
estimates of the underlying factors as known from the implied volatility es-
timates of the Black and Scholes (1973) formula. 121 Examples of estimating
the short rate along with the model parameters in a cross-section analysis
include the works by Brown and Dybvig (1986), Titman and Torous (1989),
and Brown and Schaefer (1994). However, thereby we ignore the dynamic
properties of the interest rates, but we may exploit some information about
mispricing of bonds which can lead to trading profits. 122
A second problem arises from the risk premia inherent in the observable
prices of fixed income instruments but not in the instantaneous short rate
as discussed, for example, in Chen (1995). For the purpose of bond valua-
120For a study on proxies of the short rate see, for example, Chapman, Long, and
Pearson (1999).
121See, for example, Hull and White (1990).
122See, for example, Kellerhals and Uhrig-Homburg (1998) for empirical results on the
market of German government bonds.
15.1 Estimation Techniques for Term Structure Models 149

tion, we need to obtain parameter estimates for the risk adjusted processes.
However, the parameters denoting the risk premia cannot be estimated sep-
arately from neither a pure time-series approach nor a pure cross-section
analysis only. ScMbel (1995a), for example, uses a two step procedure
where he first estimates parameters associated with the dynamic behavior
of the factors from time-series data and identifies the risk premium in a
second step from a cross-section sample of the term structure.
The described difficulties with estimating term structure models have
created a branch of econometric literature devoted to the analysis of panel
data. This combination of time-series and cross-section data is considered
very suitable for the interest rate market, since on every trading day a
whole spectrum of interest rates, i.e. a term structure, is being provided
from the markets of fixed income instruments. An estimation based on
interest rate panel data takes into account the whole information embedded
in the term structure and should therefore result in more efficient parameter
estimates. The approach we implement is based on a convenient state space
representation of the term structure model and treats the underlying state
variables correctly as unobservables. In conjunction with the customized
state space model we then exploit the econometric methods of Kalman
filtering and maximum likelihood estimation.
This estimation procedure is considered an appropriate method for both
extracting time-series of the unobservable state variables and estimating the
relevant model parameters. 123 The literature along this line can be traced
back to the work by Pennacchi (1991) who estimates an equilibrium inflation
term structure model with a number of four Treasury bills with different
maturities. Duan and Simonato (1995) and Ball and Torous (1996) provide
encouraging simulation results on Kalman filters working with single-factor
models of the Vasicek (1977) and Cox, Ingersoll, and Ross (1985b) type.
Especially, the mean-reversion parameters are estimated accurately when
compared to other estimation procedures. 124 The work by Lund (1997)
uses extended Kalman filters in order to use other marketable securities than
pure yield curve data like, for example, coupon bond prices and swap rates.
123For further advantages see, for example, Geyer and Pichler (1997) and De Jong
(1997).
124For an exploration of different estimation methods in the case of mean-reverting
interest rate dynamics that are sufficiently close to a non-stationary process with a unit
root see, for example, Ball and Torous (1996).
150 15. Calibration to Standard Instruments

FUrther, De Jong (1997) and Babbs and Nowman (1999) demonstrate how
to implement affine multi-factor term structure models on panel interest rate
data. Nunes and Clewlow (1999) present estimation results on simulations
and real market data for extended Kalman filter algorithms using interest
rate caps and swaptions.
In the following sections we calibrate our term structure model to stan-
dard fixed income instruments. For this purpose we customize the general
state space model as presented in chapter 3 to reflect the time-series and
cross-sectional properties of our theoretical term structure model at best
and setup the Kalman filters along the lines of chapter 4. We implement
two special types of Kalman filter algorithms, one for linear and one for
non-linear functional relationships in the measurement equation, in order
to calibrate our term structure model to the liquid US markets of Treasury
Securities, LIB OR rates, and swaps.

15.2 Discrete Time Distribution of the State

Variables
In setting up the specific state space model usable for Kalman filter algo-
rithms, we begin with deriving the transition equation from the continuous
time dynamics of the state variables. The aim for the transition equation is
to represent the stochastic evolution of the state variables x = [r (t), A (t)]'
as good as possible. In section 12.2 we made the assumption of

(15.1)

for the underlying factors of our term structure model. This stochastic
representation of the evolution of the state variables can be integrated to
yield the solution125

Xt eTP(t-s)xs + J
t

eTP(t-()p,JI>dt+ Jt

eTI'(t-()udWJI>

s s

eTP(t-s)xs - (TJI>r 1 (I - eTP(t-S)) p,JI> + J

t

eTP(t-()udWJI>

125See, for example, Hirsch and Smale (1974, ch. 5).

15.2 Discrete Time Distribution of the State Variables 151

for 8 < t. From this almost explicit integral solution we obtain for the dis-
crete time interval of!:1t = [8, t] the following exact discrete time equivalent:

Xt = eTPLltXt_Llt - (TPrl (I - e TPLlt ) ,."P + 'lJt, (15.2)

with 'lJt = Jt

eTP(t-()udWIP .

t-Llt

Further, with the constant vector of the unconditional long-run means of

the state variables m defined by126

i.e. the constant is given by m =- (TIP) -1 ,."IP, we obtain for equation (15.2):

Xt = eTPLltXt_Llt - ((TIP) -1 TIP _ (T") -1 eTPLltTIP) m + 'lJt

= eTPLltXt_Llt - (I - eT'Llt) m + 'lJt
for which we finally get

Thus, we derived our setting for the transition equation of the state space
model as

(15.3)

using the substitution et = Xt - m. For the remaining matrix exponential

e TPLlt of equation (15.3) we use diagonal Pade approximations with scaling
and squaring defined as follows. 127

Definition 15.2.1 (Pade Approximation) 128 The (p, q) Pade approxi-

126In section 15.3.3 we will see that m is the fix point of the deterministic part of the
system described in equation (15.1).
12 7 For a comparison in terms of computational stability of various algorithms to calcu-
late matrix exponentials see Moler and van Loan (1978).
12 8 See, for example, Golub and van Loan (1996).
152 15. Calibration to Standard Instruments

mation to eA , where A is a n x n matrix, is defined by

Rp,q (A) [Dp,q (A)r 1 Np,q (A),

with Np,q (A) = ~ (p + q - k) !p! A k and

~ (p+q)!k!{p-k)! '
(p+q-k)!q! -A k
2:
q
Dp,q (A)
= k=O
(p + q)!k! (q - k)! ( ).

Diagonal Pade Approximations are given if p = q. The method of scaling

and squaring exploits the fundamental property

of the exponential function. With a j - th order scaling, i. e. choosing m = 2j ,

we use

as approximating algorithm to the matrix exponential.

FUrther, we need to examine the statistical properties of the noise term

'1Jt in equation (15.3). For the conditional expectation of the error term we
obtain]ffi1P ['1JtIFt-6.tl = OJ its covariance matrix Q (t/J) will be approximated
by

Bringing the results together, we can state the transition equation - as

examined in general in chapter 3 - corresponding to the time-series impli-
15.3. US TI-easury Securities 153

cations of our term structure model by

et = Ct ('ljJ) + <I>t ('ljJ) et-l + '11t ('ljJ) , (15.4)

with Ct ('ljJ)

<I>t ('ljJ)
[0,0]' ,

[ (T..a,)
E ['11tIFt-lltl = 0, and
Rq,q r
Var [11tIFt-lltl = [ a
2 para>. ] flt
par~>. a >.2

for the matrix and error term specifications. The obtained transition equa-
tion can be used in our empirical models independently of the interest data
we calibrate our term structure model to. However, the functional rela-
tionship of the measurement equation is conditional to the fixed income
instruments we use as sample data. Based on the sample data we use for
calibration we need to differentiate between implementations of the linear
and the extended Kalman filter algorithm as we show in the following two
sections.

15.3 US Treasury Securities

15.3.1 Data Analysis
For a first calibration of our term structure model we employ the very
liquid markets of US Treasury Securities. The database we use is set up
from the Bliss (1999) data which is an updated version of the widely applied
McCulloch-Kwon data. 129 The data set consists of month end price quotes
for Treasury issues for the period January 1970 through December 1998
which yields 348 sample dates.
The price quotes are taken from the Center of Research in Security
Prices (CRSP) Government Bonds files. Included are all eligible issues with
maturities running from 1 month to 30 years. 130 From these issues the con-
129For a detailed description of the data set see McCulloch and Kwon (1993).
130 Bonds with option features and special liquidity problems were eliminated; see Bliss
(1999).
154 15. Calibration to Standard Instruments

10y

Figure 15.1: Estimated Zero Bond Yield Curve from 1970 to 1998

tinuous time yields are estimated by approximating the discount function

by cubic splines. 131 This results in a full term structure of continuous time
zero bond interest rates. In order to get an idea on how the term structure
evolves over time we visualize the historical evolution of the interest rates
in figure 15.1 and provide accompanying descriptive statistics in table 15.1.
Looking at the shape of the term structure the data show an upward sloping
yield curve on average. Moreover, we can identify the unstable period of
the years 1979 to 1982 when the Federal Reserve monetary policy shifted to
the 'New Operating Procedure'. The Federal Reserve Bank's anti inflation
strategy from October 1979 to October 1982 was to deemphasize the federal
funds rate as operating target and to turn to non-borrowed reserves target-
ing. 132 During this period high levels of interest rates coincide with large
volatilities. The development of interest rates during this period is also the
reason for the positive skewness of around one for the different maturities.

In terms of maturities used for our state space model, i.e. for the ob-

l31See McCulloch (1975) for the details of this method.

132For the Federal Reserve Bank's monetary operating procedures see, for example,
Miller and VanHoose (1993, ch. 25).
15.3 US TI-easury Securities 155

Table 15.1: Descriptive Statistics of the Term Structure Sample

Maturity 3M 6M lY 2Y 3Y 5Y 7Y 10 Y
MEAN 0.0684 0.0709 0.0731 0.0761 0.0777 0.0802 0.0816 0.0828
STD 0.0271 0.0272 0.0263 0.0248 0.0237 0.0226 0.0220 0.0209
MAX 0.1619 0.1640 0.1621 0.1586 0.1573 0.1535 0.1507 0.1477
MED 0.0613 0.0644 0.0679 0.0702 0.0724 0.0749 0.0763 0.0776
MIN 0.0278 0.0288 0.0309 0.0380 0.0420 0.0435 0.0430 0.0451
SK 1.1801 1.1135 1.0447 1.0292 1.0448 1.0425 1.0247 1.0133
KU 4.3848 4.1363 3.9071 3.7443 3.6703 3.4985 3.3949 3.3956

Note:
For each maturity we report the sample mean (MEAN), standard deviation
(STD), maximum values (MAX), median (MED), minimum values (MIN),
skewness (SK), and kurtosis (KU).

servable number of yields Yt in the measurement equation (15.5), we need

to choose a number of distinct maturities. For the choice of maturities we
consider two relevant aspects: First, we need to constrain our measurement
equation to a fixed number of yields. Recent choices from the empirical
literature on Kalman filtering cover the range from three yields as used in
the simulation analysis of Ball and Torous (1996) to eight yields as with the
empirical study by Babbs and Nowman (1999). In line with the works by
Duan and Simonato (1995) and De Jong (1997) we consider a number of
four yields as appropriate. Second, to cover the dynamics of the whole spec-
trum of yields we need to select the yields appropriately over the available
maturities. At the long end of the term structure we only have a sparse
number of available issues, especially in the 1970's. At the short end we
only use Treasury bills that have maturities above one month to avoid in-
cluding unexplainable price movements of issues close to their redemption
time. For finally choosing the distance between the maturities we also take
the volatility term structure into account.

From table 15.1 we infer that the longer interest rates are less volatile
than the shorter rates. This calls for a closer representation of the shorter
part of the term structure. The 10-years rate, for example, is only 77
percent as volatile as the 3-months rate. As we have seen in section 13.2.2
our interest rate model is capable to cover such volatility structures. Thus,
156 15. Calibration to Standard Instruments

we select the four maturities containing three months, one year, five years,
and ten years to govern the measurement equation (15.5).

Table 15.2: Cross-Correlations among Interest Rates

Maturity 3M 6M lY 2Y 3Y 5Y 7Y lOY
3M 1
6M 0.9963 1
lY 0.9849 0.9944 1
2Y 0.9584 0.9713 0.9890 1
3Y 0.9355 0.9500 0.9732 0.9959 1
5Y 0.8973 0.9128 0.9419 0.9789 0.9928 1
7Y 0.8713 0.8874 0.9196 0.9639 0.9826 0.9973 1
10 Y 0.8484 0.8644 0.8985 0.9480 0.9702 0.9905 0.9972 1

1.00

~
0.98
3Y
0.96
2Y
0.94
g
~ 0.92
~ 3M 6M
8 0.90

0.88

0.86

0.84 +--...--,.--.----,r-----.----,--.........----r---.----,
o 2 4 6 8 10
Maturity
Figure 15.2: Cross-Correlations among Interest Rates

Furthermore, we show the cross-correlation among the estimated yields

in table 15.2 and figure 15.2. Surprising is the high degree of comovements
among the different interest rates along the term structure. Even the short
and the long end with the 3-months and 10-years rate exhibit a correlation of
0.8484. Especially the interest rates with adjacent time to maturities reveal
15.3 US Treasury Securities 157

an almost perfect correlation whereas the degree of correlation decreases

with increasing difference in maturity. However, the amount of decorrela-
tion inherent in the term structure intuitively displays the need to model
the term structure not only by a one but a multi-factor interest rate model
of the yield curve. This stems from the fact that single-factor models imply
perfect instantaneous correlation between the movements in yields of differ-
ent maturities. However, this does not mean that with single-factor models
the yield curve is forced to move in parallel shifts; the yields are affected by
changes in the driving variable in as a complex way as the richness of the
model can allow. 133
The high correlations among the different interest rates give rise to the
idea of analyzing the data set with the descriptive device of principal com-
ponent analysis. 134 This analysis asks, whether we can describe each of a
set of original variables by a linear function of a small number of so called
common factors - the principal components - with a high degree of accuracy.
This would be trivially true if all variables moved proportionally. Then a
single variable would suffice to describe the behavior of all original variables.
In general, the principal component technique transforms the original vari-
ables into orthogonal new variables called factors. By this transformation
the total variance of the variables is preserved. Finally, the first principal
component accounts for the maximum fraction of the total variance, the
second factor explains the largest part of the residual variance until the last
factor accounts for all of the original variance.
In the case of a set of interest rates combined to build up a full spectrum
of maturities, i.e. a yield curve, we face a high degree of comovement across
the yields as described above. 135 The idea to determine common factors that
affect interest rates of different maturities was first developed by Litterman
and Scheinkman (1991). For US Treasury Securities their analysis suggests
that most of the variation in returns on all fixed income securities can be
explained by three factors. A first factor is attributed to the level of interest

133For the shortcomings of a single-factor model see, for example, Chen (1996b).
13 4 This multivariate analysis technique is described in detail, for example, in Theil
(1971, sec. 1.9) and Jolliffe (1986). For testing cointegration with principal component
methods see, for example, Phillips and Ouliaris (1988).
135For a discussion of cointegration in the case of the term structure of interest rates see,
for example, Campbell and Shiller (1987). The formal development of the key concepts
of cointegration is given in Engle and Granger (1987).
158 15. Calibration to Standard Instruments

rates in that it is basically constant across maturities. The second factor

they call steepness factor even though it does not correspond exactly to any
of the steepness measures commonly used. This factor lowers the yields
with shorter maturities and raises the interest rates with higher dates of
maturity, and vice versa. Finally, the last factor, which they call curvature,
increases the curvature of the yield curve. The purpose of obtaining these
independent factors is to be able to describe the dynamics of the entire yield
curve without a large loss of information.

Table 15.3: Principal Component Analysis

Principal Component First Second Third Fourth Fifth

Accounted separate 0.871168 0.098036 0.017786 0.007466 0.002927

Variance cumulated 0.871168 0.969204 0.986990 0.994456 0.997383

Table 15.4: Factor Loadings of the Common Factors

Principal Factor Loadings
Component 3-M 6-M l-Y 2-Y 3-Y 5-Y 7-Y lO-Y

First 0.1076 0.1126 0.1081 0.0973 0.0873 0.0731 0.0647 0.0570

Second -0.0521 -0.0308 -0.0046 0.0144 0.0239 0.0302 0.0317 0.0320
Third 0.0159 -0.0015 -0.0145 -0.0139 -0.0093 0.0026 0.0123 0.0209
Fourth 0.0102 -0.0101 -0.0097 0.0039 0.0076 0.0083 0.0011 -0.0109
Fifth -0.0038 0.0082 -0.0026 -0.0065 -0.0001 0.0067 0.0037 -0.0057

Based on our data set we further examine whether we are able to dras-
tically reduce the number of independent factors with little loss of informa-
tion. This will further support our proposed term structure model in that it
only uses two factors, the instantaneous interest rate and the market price
of risk. The results of the principal components analysis are presented in
tables 15.3 and 15.4. The original variables are the returns of the different
interest rates from the yield curve. 136
Table 15.3 shows the resulting values for the explained fractions of vari-
136For the principal component analysis we use all eight interest rates with maturities
given in table 15.1.
15.3 US Treasury Securities 159

0.12

0.10

0.08

0.06
(/)
C> 2. Factor
/
c:
'6 0.04
<II 4. Factor
0
...J
0.02
----- -.-.-.-.-.-.-.-.-
~
u.
0.00
.- -

-0.02 ~3.Factor
-0.04

·0.06
0 2 4 6 8 10
Time to Maturity

Figure 15.3: Factor Loadings of Principal Components

ance. The first line consists of the separate numbers of the accounted vari-
ance. Therein, the first principal component already contains 87.1 percent
of the total variance incorporated in the original variables. The second fac-
tor still has an explaining power of 9.8 percent. Thus, as can be seen from
the second line, the first two factors add up to 96.9 percent of cumulated
accounted variance. An additional third factor can raise the possible vari-
ance by another 1.8 percent, whereas the fourth and fifth factor only reveal
another 0.7 and 0.3 percent, respectively. Thus, by implementing our two
factor interest rate model we at best only loose 3.1 percent of the original
variance inherent in the data sample.
Further, we present the factor loadings of each of the first five princi-
pal components in table 15.4. These weights represent the sensitivity of
the original variables to the common factors. Thus, we can interpret the
loadings as the impact of the factors on the different yields of the term struc-
ture. The results on the weights of the first three principal components are
graphed in figure 15.3. For the first factor weights we obtain positive load-
ings in a fairly constant range of 0.1126 to 0.0570. This means that the yield
changes caused by the first factor essentially represent a parallel change of
the yield curve with the short maturities being somewhat more sensitive to
the factor movement. This result gives evidence to the interpretation of the
first principal component as being the average level of the yield curve. The
second principal component is made up by continually increasing weights
160 15. Calibration to Standard Instruments

across the term structure. The factor loadings are of opposite signs at the
short and long end of the maturity spectrum with a minimum at the 3-
months yield of -0.0521 and a maximum of 0.0320 for the lO-years rate.
Thereby, the second factor lowers the interest rates with shorter maturities
and increases the longer maturity rates which leads to refer to this factor as
steepness factor. Besides the first two principal components the remaining
higher order factors exhibit significant lower weights. The highest weight
of 0.0209 we obtain with the third factor. For this factor the loadings are
positive at the very short end and for maturities above five years. In the
intermediate maturity range we observe negative factor loadings. Thus, the
third principal component makes the yield curve twist in the way that it
raises it at the ends and at the same time lowers interest rates with interme-
diate maturities. The subsequent factors show even lower weights, at most
half of the weights of the third factor, and do not exhibit distinct patterns
in that we see theoretical interpretable facts. We conclude that for our data
set we obtain generally comparable results to those presented in the study
of Litterman and Scheinkman (1991).
In the light of these findings, the reduction of modeling the yield curve
with only specifying the dynamics of two factors, as with our theoretical
term structure model, is further given ground from a statistical analysis of
the historical interest rate data. Next, we present the estimation results on
the parameter values of our term structure model before we turn to analyze
the behavior of the underlying state variables.

15.3.2 Parameter Estimation

In the case of extracting the information on interest rates from zero coupon
spot rates, we are able to set up a linear Kalman filter algorithm as laid
out in chapter 4. In order to obtain a measurement equation we exploit the
affine functional relationship

1
y(t,T) = - T [[A(t,T),B(t,T)]x+C(t,T)]
-t

from section 13.2.1. Restating this affine structure for a vector of 9 zero
bond yields y (t, T) and the driving factors, the short rate r (t) and the
15.3 US Treasury Securities 161

market price of risk A (t), yields the measurement equation

(15.5)

= [A *, B*] m + C*, and

[A*, B*],

where the modified factor loadings are given by 0* = [0 (t, Tt) / (Tl - t),
o (t, T2) / (T2 - t), ... ,0 (t, Tg) / (Tg - t)l' with 0* E {A*, B*, C*}. In the
measurement equation we further add a normally distributed measurement
error term €t (1/J) for that the observable yields Yt are only measurable with
noise. For the moment specification of the error term €t (1/J) we assume
that the errors have zero mean, are serially uncorrelated, and have a time
invariant covariance matrix. Notationally, we have the assumptions

IE [€tIFt_~t] 0, and
Var [€tIFt_~tl 17; • Igxg

which adds with 17e one more parameter to our parameter space 1/J.137
Thus, with the functional mapping of the state variables et
on the ob-
servable yield vector Yt from the bond market, we establish our specific
measurement equation. Based on the state space model defined by the tran-
sition equation (15.4) and the measurement equation (15.5) we are able to
run a linear Kalman filter algorithm. Applying the filter algorithm we esti-
mate the parameter values 1/J as laid out in chapter 5 by means of maximum
likelihood. Based on the term structure data set of US Treasury Security
issues we estimate the parameter values for the full available period ranging
from 1970 to 1998.
In table 15.5 we present the results for the entire sample period of our
two-factor model as well as of the reduced model which we derived in section
13.3.1. For the parameters common to both models we come to comparable
estimates besides for parameters /-LA and 17f:. The parameter /-LA actually
corresponds to the constant market price of risk A in the reduced model.
13 7 Notethat this homoskedastic error term specification leaves space to incorporate the
full cross-correlation structure among interest rates of different maturities as given in
equation (13.22) in order to obtain a close mapping with the empirical values reported
in table 15.2.
162 15. Calibration to Standard Instruments

Table 15.5: Estimated Parameter Values

Parameters OUf Model Reduced Model

t/J EST STn EST STn

J.Lr 0.007068** 0.003397 0.004255* 0.002253
/'i,r -0.074414*** 0.008277 -0.067469·** 0.004440
(7r 0.022532*** 0.000926 0.017075·** 0.000992

J.L),. 0.000768 0.122034 -0.255242·* 0.131353

/'i,),. -0.619839·** 0.120246
(7),. 0.386329*** 0.022684
fJ;,. -0.443361 *. * 0.019980

A2 -0.116206 0.172479

P 0.710261*** 0.033481
(7c; 0.001400**· 0.000036 0.006011*** 0.000125

L(Y;t/J) -6844.58 -4916.87

Notes:
Statistically significant parameter estimates at the 1%-, 5%- and
lO%-levels are denoted by ***, **, and *, respectively. The para-
meter J.L),. corresponds to the constant A in the reduced model.

With a value of -0.2552 the constant value actually lies close to the average
value of the stochastic market price of risk in our model which is given in
table 15.7 with a mean value of -0.2624. Looking at the values of the
measurement error term (7c; we see a significant improvement by using the
two-factor model. Adding one more factor lowers the error term from 60
b.p. to 14 b.p. Overall, the parameter estimates are statistically significant
besides the two level parameters of the market price of risk process, J.L)"
and A2. This, however, is not surprising since drift parameters generally
show wide confidence intervals. The mean-reversion parameters /'i,r and /'i,),.
imply mean half-lives, i.e. the expected time for the state variables to return
halfway to their long-run means, of 9.3 years and 1.1 years, respectively.
These results correspond to the estimates on the volatility parameters in
that (7r = 2.25% is much lower than the volatility of the market price of
risk with (7),. = 38.6%. The point estimates on the parameters J.Lr' /'i,r, (7r,
and (7c;, which are comparable to the one and two-factor specifications of
15.3 US 'freasury Securities 163

term structure models tested by Duan and Simonato (1995) and Babbs and
Nowman (1999), are found to be in line with their reported findings. For the
correlation factor p in our model we obtain a point estimate of 0.71. This
demonstrates that in times of high short rates the market price of short
rate risk A is also high which means that market participants then parallel
become more risk taking. 138

0.8 0.025
0.6
0.4 j_---:.;;---_ _ _ _ _ *""---- 0.020
0.2 CJ}../

0.0 Kr ' "

-0.2 F===/7=::::::::::::====~=~~C:==:::::::==>~-_t 0.015 ::l
::l -0.4 t.. :§
:J / "0

~ ~~:: s}.. ........................":-.::. . . . . . . / . . . . . . . . .::... 0.01 o~

-1.0 Il r / -':::"-":::~-
-1.2 _____ CJ r 0.005
-1.4
K}.. ~:E
-1.6
:::::::::::::::::::::.:.::::::: .. :::: ... ::::~.....~ .........................
,""', .. .
-1 . 8 -+-~,....-,,.........---r--'T=:::.,.........,.-~.-.......,.........-,.--.--....,.-~~......,r--"........-+ 0 . 0 0 0
1970 1975 1980 1985 1990

Figure 15.4: Estimated Values for the Parameters t/J over Time

Further, we examine the stability of the parameter estimates for different

choices of the sample period. In detail we run inferences for the whole
sample period from 1970 to 1998 and continuously decrease the sample
length by one year each time ending in a minimum sample of five years from
1993 to 1998. The results on these fourteen inferences are shown graphically
in figure 15.4 and the accompanying descriptive statistics are reported in
table 15.6. At a first glance of the graph we see that the point estimates
of the parameters are fairly stable until 1979 and also for the period from
1982 onwards. In the previous section we saw that interest rates follow
a different pattern during the late seventies and early eighties due to the
Federal Reserve Bank's changing policy. Our results support the finding in

138For the impact of the state variables on the term structure compare section 13.4.1.
164 15. Calibration to Standard Instruments

Table 15.6: Statistics on Historical Parameter Estimates

Parameter MEAN STD MAX UQa) MED a ) LQa) MIN SK b) KU b)

f.-lr 0.0100 0.0017 0.0121 0.0115 0.0104 0.0088 0.0069 -0.48 1.90
/'i,r -0.1577 0.0396 -0.0744 0.0097 0.0090 0.0083 -0.2222 0.37 2.45
ar 0.0159 0.0070 0.0238 0.0085 0.0100 0.0118 0.0070 -0.02 1.14

f.-l),. 0.0008 0.0001 0.0013 0.0120 0.0087 0.0117 0.0007 2.10 7.67
/'i,),. -0.6801 0.7048 -0.0109 0.0119 0.0089 0.0096 -1.7787 -0.37 1.50
a),. 0.4049 0.0766 0.6373 0.0097 0.0076 0.0096 0.3238 1.27 4.63
0),. -0.2608 0.1275 -0.0918 0.0113 0.0100 0.0109 -0.4434 -0.01 1.29

>'2 -0.0442 0.1473 0.2139 0.0105 0.0089 0.0096 -0.2379 0.21 1.43

P 0.1145 0.4652 0.7103 0.0109 0.0100 0.0117 -0.4187 0.03 1.16

aE 0.0012 0.0002 0.0015 0.0117 0.0100 0.0117 0.0008 -0.14 1.41

Notes:
a) In the statistics for the parameter estimates, UQ and LQ denote the upper and lower
quartile respectively; the median is abbreviated by MED.
b) For each parameter we also report the skewness (SK) and kurtosis (KU).

the literature that the shift in the monetary policy during the years of 1979
to 1982 caused a structural break in the interest rate process. 139 This can
especially be seen from the volatility of the short rate a r which starts at
225 b.p. fairly constant until 1980 where it falls monotonically to a value
of 7 b. p. for the minimum sample length. Therein, the impact of the high
volatilities during these years becomes evident. Also, the correlation factor
between the two state variables changes from positive values of 0.7 to -0.3
which marks a shift in the market participants changes of risk aversion
attitudes. Further, during time the mean-reversion tendency of the market
price of risk with a maximum of -1.78 does not become significant any
more after 1982 whereas the central tendency parameter of the short rate
/'i,r rather stays constant with maximum and minimum values of -0.07 and

-0.22, respectively. Finally, we see the measurement error a E to become

smaller with a decreasing sample size starting from a value of 14 b.p. to end
with values 8 b.p. This is a reasonable fact for the parameter a e since it is
a parameter to measure the goodness of fit which needs to get better for a
smaller sample size.

139See, for example, Duan and Simonato (1995) and Hansen (1998).
15.3 US Treasury Securities 165

15.3.3 Analysis of the State Variables

In the analysis of the state variables we first examine the dynamic behavior
of the two-factor term structure model in the absence of randomness before
we examine the obtained time-series of the state variables from the sample
data.
The description of the dynamic behavior of the state variables can either
be based on equation (12.1) or equation (12.2) dependent on whether we
look at the dynamics under the lP-measure or the Q-measure. Without
stochastic randomness we obtain the system

(15.6)

for the state variables x under measure M E {lP, Q}. Choosing the market
price of risk as the dependent state variable we divide the given equations
by each other which results in the differential equations

/.LA + K,>.rlP + (h)...1P and d)...Q = /.L>. - a>.)...2 + K,>.rQ + (h)...Q

~+K,~ ~ ~+~Q_~~

These equations define the phase space of the dependent state variable )....
The solutions to these first order differential equations can be plotted in a
vector field as graphed in figure 15.5. 140 The direction, the angle, and the
length of the vector line segments represent, respectively, the movement,
the way, and the speed of the state variables from the starting values at
that point on an orbital trajectory.
In order to algebraically analyze the stability 141 of our state space system
we first calculate the eigenvalues of the system as

ewlP = {-0.4434; -0.0744}, and eWQ = {-0.4797; -0.0398}

with corresponding eigenvectors of

eVp =
{ [ 0] [-0.5115]}
1 ; 0.8593 ' and eVQ =
{ [ 0.0557] [-0.5456]}
0.9984; 0.8380 .

140 Thevector field is based on the estimated parameter values reported in table 15.5.
141 Fora systematic classification of the possible equilibria of a two dimensional linear
system see, for example, Brock and Malliaris (1989, ch. 3.6).
166 15. Calibration to St8Jldard Instruments

Eigenline 1 ___ Eigenline 1 _ _

0.2
0.2 I I I I I Ii I I III l 111111:1 I j
1111:1 I I I I I ~ • I I I II 1 j
I I Ii I I I j I l N
~ ~ ~ I I il I I
I Ii I I I I j I CD
:§ 0.1
+.~~. Iii I I
I: I I I I I 1 5i - .~. I+ .! I I I
m
• • j

I I I I 1 ""'~. • • j j! j I
I I I I 8l
f
,
~
f
.~.
,
I ,
• '! j

I I I '"
~ 0.0
...... t
, , , , .,..... ,~ ; • j j j
I I
, ... ~ I • .: • ~ • t " , .. ···t: • ; 1 1 j
~~"··"··l+ttt t "".(-"';11
-0.1 ' , , , ...•. I~ , , ! • j j -0.1 1 I I 'i I ". , • ;
, , f , j • , , ,
, ....• j j tIl ~II"""
" " , , [ ' - - ·. • • '!I
ttl t , , • ~"'~
I I , t t l : • ' .. ~ , • • ; t tt il 'I~
-0.2 I I t t , I: I • ....• , , -0.2 h-t..,.....,~t,...",t~..,.....,,..;..;,...,,,1~"Tt~'.....'
3% 6% 9% 12% 15% 18% 3% 6% 9% 12% 15% 18%
r" - values rQ - values

Figure 15.5: Phase Plane of State Variables x as Vector Field

Under both measures we obtain eigenvalues with the properties142

i.e. they are distinct and lie both below zero. In this case we obtain a
stable tangent node at xl!' = [0.0950, -0.1311]' and xlQ = [0.1106, -0.0517]',
respectively, as indicated by the point of intersection of the two dotted
lines. 143 The dotted lines represent the eigenlines of the system which are
lines through the asymptotic point of the system. They can be derived from
the general solutions to the partial differential equation of (15.6)

0.0950]
[ -0.1311 [ 0] -0.4434 [ -0.5115] -0.0744 d
+ Cl 1 e + C2 0.8593 e , an

[ 0.1106] [ 0.0557] -0.4797 C [-0.5456] e-0.0398

-0.0517 + C3 0.9984 e + 4 0.8380 '

142See, for example, Brock and Malliaris (1989, p. 80).

143The coordinates of the nodes are found by setting the first order derivatives of equa-
tion (15.6) equal to zero and then solve for the state variables.
15.3 US 'freasury Securities 167

0.8

0.6

0.4
Eigenline2

I
0.2

en 0.0 .......
(I)
~
(ij
> -0.2
I
a..
<-< -0.4 V/-'_'--01/82

-0.6

-0.8
01/83
-1.0
0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18

rP - values

Figure 15.6: Historical Evolution of State Variables from 1970 to 1998

with arbitrary constants Gi, with i E {I, 2, ... ,4}. The eigenlines are found
by consecutively setting the constants Ci equal to zero and solving the para-
metric representation of the functional relationships between r and A for the
state variable A. Generated by the eigenvectors [0,1]' and [-0.5115,0.8593],
for the dynamic system under the measure IP' we obtain

rf = 0.0950, and A~ = -1.6800r~ + 0.0285

for the eigenlines. For the dynamic system under Q the eigenlines become

A? = 17.9100r? - 2.0332, and A~ = -1.5359r~ + 0.1182

for the eigenvectors [0.0557,0.9984], and [-0.5456,0.8380]', respectively.
Both eigenlines are plotted in dashed lines for the two probability measures
in figure 15.5. From the vector line segments we see that the dynamic sys-
tem is stable. Starting from any point in the phase space the path is forced
on different ways at a decreasing speed towards the asymptotic eigenlines
and finally towards the long-term equilibrium state of the system.
In the analysis up to now we have eliminated the randomness of the
dynamic system of the two modeled state variables. Now, we consider the
168 15. Calibration to Standard Instruments

dynamic behavior of the term structure model including random shocks. We

especially examine the time-series for the unobservable state variables Xt re-
trieved from the data sample using the Kalman filter algorithm. In figure
15.6 we plot the historical evolution of the state variables Xt obtained from
the filter algorithm for the whole sample period; included are the asymp-
totic lines from the deterministic analysis. The accompanying descriptive
statistics are reported in table 15.7.

Table 15.7: Statistics on Realized State Variable Values

X MEAN STD MAX UQa) MED a ) LQa) MIN SK b) KU b)

r 0.0699 0.0280 0.1654 0.0826 0.0631 0.0519 0.0233 1.06 4.10

A -0.2624 0.3777 0.8084 0.0099 -0.2782 -0.5301 -0.9772 0.23 2.41

Notes:
a) In the statistics for the state variables x, UQ and LQ denote the upper and
lower quartile respectively; the median is abbreviated by MED.
b) For each state variable we also report the skewness (SK) and kurtosis (KU).

From the numerical results we first see the short rate in a range of
realized values from a minimum of 2.3% to a maximum of 16.5% with a
mean value of 7.0%. These values closely resemble the statistics of the
short rate (3-months rate) included in the data sample as can be seen from
the reported values in table 15.1. Also, the positive skewness stemming
from the 1979-1982 period (see the short rate values in graph 15.6) and
the excess kurtosis are retained. Second, with our filtering technique we
can extract the market prices of risk over time. The values we obtain for
At take a minimum of -0.98, a mean value of -0.26, and a maximum of
0.81. With an upper quartile value of close to zero we can infer that the
market participants were risk averse over 75 percent of the sample period
from 01/1970 - 12/1998. Finally, we see from graph 15.6 how the two state
variables evolve over time. Therein, the years of 1979 to 1982 mark a period
of high short rates along with risk taking values for the market price of risk.
We further note that the path of the state variables actually crosses the
asymptotic lines when we include the historical randomness. The central
tendency is especially effective when the realized values are far away from
their long-run means.
15.4. Other Liquid Markets 169

15.4 Other Liquid Markets

The analysis of the previous section is based on interest rates obtained
from the US Treasury Security market. Other possible observables from
the fixed income markets include, for example, LIB OR or swap rates and
prices for bond options, cap and floor agreements. Using such fixed income
instruments give rise to non-linear functional relationships between the state
variables of our term structure model and the observable data. In thus, we
further implement extended Kalman filter algorithms instead of the linear
algorithms we could use for the US Treasury Security market. From the
possible market prices of fixed income instruments we choose to work with
data from the LIB OR and swap markets because of their availability and
high liquidity.144

15.4.1 Appropriate Filtering Algorithm

In order to calibrate our term structure model to the LIB OR and swap mar-
kets we implement an extended Kalman filter algorithm. The evolution of
the state variables over time stays unchanged in that we can keep the speci-
fication of the transition equation (15.4) in the state space setting. However,
the measurement equation needs to be linearized when we use LIB OR and
swap rates for the observable vector Yt as we show in the following. Using
our closed-form solution to the discount bond price from equation (13.16)
the LIB OR and swap rates are given by the formulae

1 - P (t, T)
L (t, T) = (T - t) P (t, T)' and (15.7)

1 - P (t, T)
k (t, T) = n (15.8)
OLP(t,Tj )
j=l

as defined in equations (11.3) and (14.17). From these formulae we see that
the LIBOR curve represents the natural extension of the swap rate curve

144Liquidity in the US interest rate swap market is generally agreed to be due to the
high demand on swaps and the ease of hedging swaps with the US Treasury and futures
markets. Also standard terms introduced by the ISDA and BBA in 1985 have assisted
the growth in the swap markets. See, for example, Das (1994).
170 15. Calibration to Standard Instruments

for shorter maturities for which no swap contracts are traded. 145 Thus, in
order to cover the whole spectrum of the term structure we choose LIBOR
and swap rates as observables 0 (t, Ii) E {L (t, Ii), k (t, Ii)} with which we
specify the non-linear measurement equation as

gt(et,ed'l/J),'l/J) = s(et,'l/J) +et ('l/J) ,

with s (et, 'l/J) = [0 (t, T 1 ), 0 (t, T 2 ), ... ,0 (t, Tg)]' .

This non-linear equation can be linearized with a Taylor series extension

as described in section 4.5. The first order approximations lead to the
linearized form of the measurement equation

in the case of the extended Kalman filter. With the discount bonds P (t, T)
and P (t, 1j) being a function of the state variables the derivatives for the
LIBOR and the swap rate can be calculated from equations (15.7) and (15.8)
in closed-form. The moment assumptions regarding the error term et ("p)
are identical with those in the linear case. With equation (15.9) we obtain
a linearized version of the non-linear functional relationship between the
unobservable state variables et and the LIBOR and swap rates Yt which are
observable in the financial markets.
In analogy to the linear case, we are now able to run the extended
Kalman filter algorithm of section 4.5 based on the state space model defined
by the transition equation (15.4) and the measurement equation (15.9). The
empirical implementation and the estimation results are presented in the
next section.

15.4.2 Sample Data and Estimation Results

In this section we examine other liquid US fixed income markets besides
the US Treasury Security market underlying the analysis of section 15.3.
With the specific form of the extended Kalman filter algorithm described
145The similarity of the two rates can be traced back to definition 14.4.1 where the
floating part of the interest rate swap is defined on LIBOR.
15.4 Other Liquid Markets 171

in the previous section we are able to implement our term structure model
on LIBOR and swap rates. The calibration procedure is based on interest
rate data sampled monthly over the longest available period from 04/1987
to 01/2000. We obtain 154 observation dates and at each date we use four
distinct maturities as in the case of the Treasury Security sample. In both,
the LIB OR and the swap market, we select the maturities with the aim of
capturing a maximum spectrum of the available term to maturities. For the
LIB OR rates we choose to include the 1-, 3-, 6-, and 12-months rates and for
the swap rates the maturities of 2, 3, 5, and 10 years. The closer maturity
choices at the short end are intended to better capture the information of
this more volatile part of the term structure.

Table 15.8: Descriptive Statistics of LIBOR and Swap Data

LIBOR Data Swap Data

Maturity I-M 3-M 6-M 12-M 2-Y 3-Y 5-Y lO-Y

MEAN 0.0598 0.0607 0.0617 0.0639 0.0672 0.0698 0.0731 0.0771

STD 0.0175 0.0174 0.0172 0.0171 0.0163 0.0156 0.0147 0.0141
MAX 0.1006 0.1025 0.1050 0.1094 0.1064 0.1042 0.1022 0.1070
MED 0.0569 0.0576 0.0586 0.0602 0.0632 0.0653 0.0697 0.0733
MIN 0.0313 0.0319 0.0325 0.0338 0.0391 0.0434 0.0493 0.0518
SK 0.3406 0.3224 0.3212 0.3288 0.3148 0.3240 0.3147 0.2733
KU 2.5095 2.4863 2.4952 2.4996 2.1952 2.0009 1.8238 1.7921

Note:
For each data series we report the sample mean (MEAN), standard
deviation (STD), maximum values (MAX), median (MED), minimum
values (MIN), skewness (SK), and kurtosis (KU).

In table 15.8 we report the summary statistics separately for each time-
series. For the short-term market of LIB OR rates the values of the statistics
for the different maturities show very similar results. The term structure
rises on average with mean sample values of 5.98% for the one-month rate
up to 6.39% for the 12-months rate. The volatility of the rates slightly
decreases with higher maturities. For the criteria of skewness and kurtosis
we see modest positively skewed rates that do not show excess kurtosis.
172 15. Calibration to Standard Instruments

Table 15.9: Parameter Estimates for LIBOR and Swap Data

Parameter LIB OR Data Swap Data Combined Data

1/J EST STD EST STD EST STD

J.Lr 0.0174*** 0.0045 0.0029*** 0.0001 0.0041*** 0.0009

/'i,r -0.2772*** 0.0751 -0.0453*** 0.0004 -0.0424*** 0.0117

Ur 0.0100*** 0.0006 0.0156*** 0.0003 0.0100*** 0.0008

J.L>. 0.0113 0.0199 0.0037*** 0.0003 0.0060*** 0.0010

/'i,>. -0.5956 0.6618 -0.6168*** 0.0044 -0.6925*** 0.2204

U>. 3.2225*** 0.3644 0.3811*** 0.0132 1.3362*** 0.2231

(J>, -1.8756*** 0.2707 -0.4581*** 0.0137 -0.8193*** 0.1507

>'2 0.1847 0.1532 0.0224*** 0.0006 0.0196* 0.0109

P 0.1365 0.1255 0.6363*** 0.0218 0.5566*** 0.0745

Ue; 0.0007*** 0.0000 0.0003*** 0.0000 0.0020*** 0.0001

L(y;1/J) -3393.86 -3674.51 -2969.80

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.

These sample facts mostly resemble those for the swap rates. Due to the
broader maturity spectrum we only see the term structure of swap rates
increasing up to a 10-years rate of 7.71%.
The estimation results on the parameter values 1/J of our term structure
model are given in table 15.9. We calibrate the term structure model to
three different panel data: The four LIBOR rates, the four swap rates,
and a combination of the 1- and 12-months LIBOR rates along with the 5-
and 10-years swap rates. The combined data set is motivated by the close
relation of the two types of interest rate markets as we notice from equations
(15.7) and (15.8) with the aim of obtaining a smooth term structure pattern
over all available maturities. First, we compare the model fit to the three
different samples in terms of the measurement error. The LIBOR rates
exhibit a standard deviation of 7 b.p. which falls to 3 b.p. for the swap
sample but increases to 20 b. p. for the combined data. This indicates that
the swap market is the most homogeneous sample despite its high spanning
of maturities from 2 to 10 years. Second, the results on the point estimates
15.4 Other Liquid Markets 173

of J.Lr' ""r, and a r are comparable among the three samples in use. Third,
the parameters corresponding to the stochastic specification of the market
price of short rate risk are estimated with plausible values for the swap and
the combined data. However, based on the LIB OR data we obtain diverging
estimates for the parameters J.L)..., "")..., a)..., (h and ).2. This provides evidence
that a sample including only a small one year window of the term structure
is not a rich enough sample in order to get reliable estimates of the market
price of risk parameters. Fourth, the mean-reversion parameters ""r and "")...
estimated for the swap and combined data at around -0.045 and -0.65
imply mean half-lives of 15.4 years and 1.1 years, respectively. Finally, the
correlation coefficient is estimated at significantly positive values of around
+0.6. Overall, all parameters are estimated at statistically significant levels
besides the estimates for J.L)..., "")..., ).2, and p based on the LIBOR data and
).2 for the combined data.
16 Summary and Conclusions

In part III of this study we present a new continuous time term struc-
ture model and provide its implications on risk management and the pric-
ing of different fixed income derivatives based on a model calibration to
standard fixed income instruments. The theoretical and empirical analysis
contributes to the term structure literature in that it shifts the theoret-
ical attention of modeling static investor tastes towards capturing beliefs
of investor risk attitudes towards interest rate risk using an affine model
structure.
In the introductory chapter we consider the specific financial instru-
ments most relevant in modeling fixed income markets and describe the
peculiarities of an incomplete market such as the market for fixed income
instruments. In the following chapter 12 we motivate our desire to build
a new term structure model that takes into account time varying market
prices of interest rate risk. The model we suggest chooses the short interest
rate and its market price of risk as the two explaining state variables where-
upon we establish an affine term structure model. In chapter 13 we present
initial characteristic results and implications of the proposed term structure
model. First, we derive a closed-form solution for arbitrage-free prices of dis-
count bonds. Thereupon, we are able to state and analyze the implied term
structures of interest rates and volatilities. In a comparative analysis we
examine limiting cases of the term structures and the influence of the state
variables and the model parameters on the shape of the term structures.
We identify the first model factor as a level factor and the market price of
risk as steepness factor that captures at the same time the investors' risk
attitudes. Chapter 14 shows how to implement risk management techniques
based on the proposed model. With the duration concept we illustrate the
implementation of a simple model for interest rate risk management. More
176 16. Summary and Conclusions

complex risk management strategies can be followed by using the derived

closed-form expressions for the arbitrage-free prices of bond options, swap
contracts, and cap and floor agreements.
In chapter 15 we show how to calibrate our term structure model to the
standard fixed income instruments of US Treasury Securities, LIB OR rates,
and swap rates. , The maximum likelihood estimation technique we imple-
ment is based on a linear as well as an extended Kalman filter algorithm.
The parameter estimates and the data fit we obtain are similar to the re-
ported values of comparable models in the literature. With the filtering
technique we are able to extract the dynamic behavior of the risk attitudes
of fixed income investors. We infer that the market participants were risk
averse over 75 percent of the sample period from 01/1970 - 12/1998 with a
market price of risk ranging from a minimum of -0.98 to a maximum of 0.81
with a mean value of -0.26. From the historical evolution of the state vari-
ables over time we see that the years of 1979 to 1982 mark a period of high
short rates along with risk taking values for the market price of risk. Finally,
we report the estimation results on the liquid markets of US LIBOR and
swap rates which are comparable to those obtained on the Treasury sample.
 Part IV

Pricing Electricity Forwards

17 Introduction

17.1 Overview
In part IV of the study we develop a continuous time pricing model for one
of the latest cash markets to be transformed by derivative securities - the
electricity market. Based on available data on electricity derivatives we fo-
cus our attention on the valuation of short-term electricity forwards. First,
we investigate the more mature futures markets contracted on established
commodity underlyings and describe proposed pricing models for traditional
commodity futures contracts from the literature. Thereby, we gather rele-
vant information about the commodity electricity in order to build an ap-
propriate valuation model to price electricity forwards in chapter 18. From
given model assumptions we derive a closed-form valuation model for elec-
tricity forwards using risk neutral pricing techniques. The suggested model
especially captures the high and time varying volatility seen in electricity
prices. In chapter 19 we present an empirical adaptation of the theoretical
pricing model in state space form. Using maximum likelihood estimation
based on extended Kalman filtering we report empirical results on electric-
ity data from the largely deregulated Californian electricity market. In the
last chapter we conclude.

17.2 Commodity Futures Markets

Basically, a futures contract is an agreement between two parties to make
an exchange of financial products or commodities at a particular future
date. Futures contracts are traded on futures exchanges which evolved from
entrepreneurs pursuing an economic advantage by organizing the forward
exchange in a standardized and regulated manner. With the evident risk
180 17. Introduction

of fluctuating commodity prices in the grain market of Chicago during the

18th century a group of businessmen formed the Chicago Board of Trade
(CBOT) in 1848. The aims were to maintain a central market that treats
market participants equally and fairly, to collect and disseminate commodity
and economic information, and to establish quality control standards for all
deliverable grades of grain. With the futures market an efficient mechanism
has been established to allocate price risk in commodities among market
participants. 146

, ,
[.Com
·Coff..
'Cotton
t·
• Uvecata.
Pork Bellies
• Lean Hogs
r-~r
• AlUminum
'Lead
r-G~
' Silver
-Platinum
[. Crude Oil
'HeetingOiI
'Natural Gas
[ • Electricity

Figure 17.1: Classification of Commodity Futures Underlyings

Today, futures exchanges exist at different places worldwide specialized

in various types of futures contracts, According to their different underly-
ings the futures contracts can be grouped into the two major categories of
financial futures contracts and commodity futures contracts. 147 In our par-
ticular study we are interested in the commodity futures markets where we
further focus our attention on energy markets. In order to get an overview
of commodities that are relevant in futures markets see figure 17.1 which
follows the systematization of the Standard International Trade Classifica-
tion. 148
The agricultural futures contracts show the longest history with the
major exchanges being the Chicago Board of Trade (CBOT), the Chicago

146For historical, institutional and regulatory issues of futures exchanges see, for exam-
ple, Markham (1987).
147See, for example, Duffie (1989).
148See UN (1996).
17.2 Commodity Futures Markets 181

Mercantile Exchange (CME), and the Coffee, Sugar, and Cocoa Exchange
(CSCE). Futures contracts are traded on vegetable goods such as grains
(corn, oats, wheat, and rice), foodstuffs (coffee, cocoa, orange juice, and
sugar), textiles (cotton), oil and meal (soybeans, -meal, -oil, sunflower seed
and oil), timber products (lumber and plywood) and on animal goods such
as live hogs, live cattle, lean hogs, and pork bellies. For many of these
commodities various futures contracts are available due to the necessary
standardization of futures contracts. The specification of contracts for dif-
ferent grades and types of the commodity ensures a homogeneous deliverable
commodity. In the metallurgical futures market the underlying commodities
are classified by either being industrial or precious metals. The spectrum
ranges from copper, aluminum, palladium, lead, zinc, nickel, and tin up
to the other end with silver, platinum, and gold. Major metal futures ex-
changes are the Commodity Exchange (COMEX), the New York Mercantile
Exchange (NYMEX), and the London Metal Exchange (LME).
Among the various commodities the energy market is the most recent
market to be transformed by derivative securities and risk management. A
major reason for this late development is considered to be the postponed
deregulation of energy markets giving the spot markets ultimately the nec-
essary liquidity and price volatility. During the 1960s and 1970s prices of oil
as one of the most crucial commodities in the world were tightly controlled
by the Organization of Petroleum Exporting Countries (OPEC) together
with large globally integrated oil companies. The trading of oil futures at
the NYMEX began with heating oil futures in 1978, followed by crude oil fu-
tures introduced in 1983, and unleaded gasoline futures introduced in 1985
establishing the NYMEX as the world's leading energy futures exchange. A
further major energy commodity besides oil is natural gas which accounts
for almost a quarter of total US energy consumption. The gas industry has
been deregulated since the enactment of the Natural Gas Policy Act of 1978
changing it to an industry that largely operates as a free market today. An
accompanying futures contract to the natural gas spot market centers was
launched in 1990 traded on the NYMEX.
The transformation of the electricity market into a competitive mar-
ket has been essential for the development of electricity futures markets.
Since electricity is considered a regional commodity not only deregulation
in electricity markets but also in the transmission system has been crucial in
182 17. Introduction

creating a electricity derivatives market. Vigorous restructuring and dereg-

ulation efforts have been undertaken by both the Federal Energy Regulatory
Commission (FERC) and key state commissions. 149 The passage of FERC
Order 888 in 1997, along with the passage of the Public Utility Regulatory
Policies Act (PURPA) of 1978 and the Energy Policy Act (EPAct) of 1992,
opened the bulk power transmission system to competition and has allowed
new market participants to enter the generation and wholesale power busi-
ness. The state-by-state process of deregulation began in California where
the Government signed restructuring legislation in 1996 with the aim of
developing the first competitive marketplace for electricity in the US. The
revised Californian market structure established two new entities by 1998,
the California Power Exchange (CaIPX) to conduct an electricity auction
market and an Independent System Operator (CaISO) to manage transmis-
sion from generation sources to power marketers and finally consumers.150
Besides the centralized dispatch and real time market pricing of the CaISO,
the CalPX created the first one-hour and one-day forwards on electricity.151
For the rapidly developing cash electricity markets in the Western US ex-
change traded futures contracts were launched by the NYMEX in 1996,
one contract based on delivery at the California-Oregon border (COB) and
another for the Palo Verde switchyard in Arizona. This early introduc-
tion of electricity futures contracts is also seen critical in that a maturely
evolving cash market provides the benefit of being able to identify the most
desirable trading locations and the most demanded trade terms and quanti-
ties. 152 However, with the high volatility observed in electricity prices high
potential of commercial applications for electricity futures are seen: Power
generating companies can sell futures contracts to lock in a specific sales
prices for the electricity they intend or expect to produce in the future; by
buying futures power purchasers, such as large utilities or major industrial
concerns, can protect their purchase prices; power marketers who have ex-
posure on both the generating and delivery sides of the market can hedge
their risks using futures contracts.

149For a detailed record of deregulation history in the US power market at the federal
and the state levels see, for example, Bell and Lilyestrom (1997).
150 See CalPX (2000a).
151See CalPX (2000b).
152See Johnson and Sogomonian (1997).
17.3. Pricing Commodity Futures 183

17.3 Pricing Commodity Futures

In the previous section we presented the variety of different commodities
serving as underlying assets for futures contracts. Across all these com-
modities ranging from agricultural products to pure financial assets certain
common principles of futures valuation and futures price behavior apply.153
As we demonstrate further on, the essence of the price relation between a
futures contract and its underlying spot asset is captured by an arbitrage
argument; this argument, however, need to be relaxed when it comes to
pricing electricity derivatives. 154 For the valuation of futures prices in an
arbitrage-free world we assume deterministic risk free interest rates which
makes forward prices equal to futures prices. 155

Definition 17.3.1 (Futures Price) The futures price F (t, T) of an asset

St for a futures contract with delivery at time T is given by the formula

F (t, T) = ]EQ [STIFtl, (17.1)

i. e. todays arbitrage-free futures price F (t, T) is the expected spot price St

at delivery under the martingale measure Q.

The futures pricing formula of equation (17.1) is arbitrage-free in that

there are possibilities of risk-less profits if the relationship is violated. This
can be directly recovered from a cash-and-carry and a reverse cash-and-carry
arbitrage argument. Taking into account carry costs CO and carry returns
C R when holding the deliverable asset the pricing relationship of equation
(17.1) yields the standard cost-of-carry formula for futures pricing:

(17.2)

The two arbitrage arguments are as follows: First, if equation (17.2) is

violated in that the futures price is strictly larger than the expected spot
153 See,for example, Duffie (1989) and Stoll and Whaley (1993).
154See the exposition in chapter 18 for the peculiarities in the pricing of derivatives on
a non-storable underlying such as electricity.
155Por a rigorous examination of the relationship between futures and forward prices see
Cox, Ingersoll, and Ross (1981), Jarrow and Oldfield (1981), and Richard and Sundaresan
(1981).
184 17. Introduction

price including carry costs and returns under the martingale measure Q,
investors follow a cash-and-carry arbitrage. The arbitraguer borrows funds
to buy the spot asset, takes a short position in a futures contract, and
carries the deliverable asset until the futures delivery date. Second, if the
futures price is below the discounted expected spot price a reverse cash-and-
carry arbitrage can be followed. The arbitraguer sells the spot asset short,
lends the proceeds, and takes a long position in a futures contract. Together
both arbitrage possibilities ensure that the equality of equation (17.2) holds.
However, in view of implementing the two arbitrage arguments there is a
marked difference. The cash-and-carry arbitrage can be set up and pursued
easily with financial underlyings as well as storable commodities in that it
only asks for carrying the deliverable asset until the futures delivery date.
Thus, the cost-of-carry model can be used to determine an upper bound on
futures prices in that the futures price F (t, T) can not exceed the cost of
stored underlying assets for delivery. If in this case the future-spot price
relationship is mainly characterized by the carry costs the futures market is
said to be in contango. This is the standard market situation where carry
costs are higher than returns from carrying the asset and the futures prices
lie above the spot prices. However, in the case of reverse cash-and-carry ar-
bitrage we need to sell the deliverable asset short in order to realize risk-less
profits. The implementation of this arbitrage can either be complicated or
impossible when there are no or little amounts of deliverable assets avail-
able in storage. 156 Thus, the cost-of-carry model can not provide a lower
bound for futures pricing. The modified cost-oj-carry Jormula for non-carry
commodities resolves this pricing problem by including a convenience yield
CY:157

The convenience yield is the return that users of the commodity realizes
for carrying inventory of the spot commodity. This scenario of positive
156The reverse cash-and-carry arbitrage also breaks down in cases of storable commodi-
ties for which there exist no leasing market as, for example, the natural gas market. See,
for example, Leong (1997).
157The concept of a convenience yield was first described by Kaldor (1939) and Working
(1949).
17.3 Pricing Commodity Futures 185

returns or negative storage costs arises when the users will not be willing
to sell their inventory in the spot market and buy futures contracts that
will replenish their supplies at a later date. A commodity's convenience
yield can be different for various users and can vary over time. The return
will be at highest when there are spot shortages of the spot commodity. In
instances where the convenience yield plays a major role for users the spot
prices will lie above futures prices, Le. a market situation which is known as
backwardation.

Table 17.1: Commodity Pricing Models

Authors Model Specification Constants

Brennan/ dS = I-'Sdt + uSdz 1-', u,
Schwartz (1985) C(S,t)=cS c
Gibson/ dS = (I-' - 8) Sdt + ulSdz l 1-', Ul,
Schwartz d8 = K. (0: - 8) dt + U2dz2 K., 0:, U2,
(1990) dzl dz2 = pdt P
Brennan dS = I-'Sdt + uSdzs I-',u
(1991) dC = 0: (m - C) dt + 'T]dzc o:,m,'T]
dzsdzc = pdt P
SchObel (1992) dI = K. (In I - In I) I dt + u I dw K.,I,u,
P = P (1/1) -<p <p
Ross (1997) dS = k (0 - S) dt + u (.) dz k,O,u
Schwartz (1997, M. 1) dS = K. (I-' -InS) Sdt + uSdz K., 1-', u
Schwartzi dXt = -K.Xt dt + u'Xdz'X K., u'X'
Smith (1997) df.t = I-'f.dt + uf.dzf. I-'f., uf.,
InSt = Xt + et dz'Xdzf. = p,)(£dt P')(£
Schwartz dS = (r - 8) Sdt + ulSdzi ul.
(1997, M. 3) d8 = K. (a - 8) dt + U2dZ2 K., a, U2,
dr = (m* - r) dt + uadza m*,ua,
dzidz2 = Pl dt,dz2dza = P2dt,dzidza = Padt Pl' P2, Pa
Clewlow/ dS = 0: (I-' -In S) Sdt + uSdz o:,u
Strickland (1999) I-' = 81n~o,l~ + In F (0, 1) + ~ (1 - e- )
2ctt
186 17. Introduction

The evolving literature on valuation models for pricing commodity fu-

tures in a continuous time framework is summarized in table 17.1. The lit-
erature can broadly be categorized into the two approaches of no-arbitrage
models and equilibrium models.
In the class of no-arbitrage models the underlying commodity is assumed
to be tradable. The first stochastic futures pricing model of this class is
presented by Brennan and Schwartz (1985) who use a convenience yield
C (S, t) that is proportional to the spot asset price. They illustrate the
nature of their solution on a copper mine using stylized facts. However,
the assumption of a constant convenience yield only holds under restrictive
assumptions, since the theory of storage is rooted in an inverse relationship
between the convenience yield and the level of inventories. The proposed
models by Gibson and Schwartz (1990) and Brennan (1991) therefore ex-
tend the single-factor models to two-factor models with a mean-reverting
stochastic specification of the convenience yield denoted, respectively, by 8
and C. In extending these approaches Schwartz (1997, Model 3) presents
a three-factor model which includes a stochastic risk free rate of interest r.
He implements the pricing model on empirical data of oil futures contracts.
The paper by Hilliard and Reis (1998) investigates the impact of stochastic
convenience yields, stochastic interest rates, and Poisson-distributed jumps
in the spot price process in the valuation of commodity futures, forwards,
and futures options.
The valuation approach common to these pricing models is their reliance
on the information content revealed by the spot prices and their foundation
in the theory of storage. Instead of using price information available at
the spot commodity markets there are pricing models that use the com-
modity futures or forward price curve, which is observable at the market,
as given. Similar to the technique of modeling the evolution of the entire
forward curve known from interest rate derivatives pricing, Clewlow and
Strickland (1999) derive analytical pricing formulae for standard commod-
ity derivatives. 158 Based on a mean-reverting specification of the commodity
spot price process they develop a trinomial tree approach and are thereby
able to price general path-dependent derivative contracts. The approach of
modeling the movement of the term structure of commodity futures prices is

15 8 This
modeling approach can be traced back to Ho and Lee (1986) and Heath, Jarrow,
and Morton (1992).
17.3 Pricing Commodity Futures 187

also followed by Cortazar and Schwartz (1994) and Amin, Ng, and Pirrong
(1995).
The category of equilibrium models assumes the underlying commodity
to be a non-traded asset. Thereby, the no-arbitrage hedging argument does
not apply and the value of a futures contract is determined by the expected
spot price at maturity of the contract. One-factor models based on a dy-
namic specification of the spot price behavior are presented by Ross (1997)
and Schwartz (1997, Modell). Ross (1997) considers an underlying asset
which cannot be stored and held over time except at high cost making the
pricing of derivative contracts impossible with the usual risk neutral hedg-
ing argument. The quite general volatility specification of Ross (1997) is
put in concrete terms by Schwartz (1997, Modell) who assumes a volatility
that is linear in the spot price. In the two-factor model of Schwartz and
Smith (1997) the log-spot price behavior is decomposed into a process to
capture the equilibrium level ~t and a dynamic specification of transitory
disturbances Xt from the long-run leveP59
FUrthermore, two regime models are proposed in the literature which
are hybrid approaches grasping both arbitrage and equilibrium pricing sit-
uations of futures contracts. The regime switching approach of Schobel
(1992) models the spot price P via the dynamics of the level of inventories
I. Dependent on the availability of the spot asset the model captures a com-
plete and an incomplete market situation which are given for a spot asset
that is, respectively, available or too scarce for arbitrage purposes. Based
on Deaton and Laroque (1992) and Deaton and Laroque (1996), Routledge,
Seppi, and Spatt (2000) determine spot and forward prices endogenously
from an immediate net-demand process and the resulting dynamic structure
of equilibrium inventory.160 They additionally discuss a tractable two-factor
augmentation of their model which leads to better calibration results of the
volatility term structures. In the paper by Buhler, Korn, and Schobel (2000)
a unifying approach a la Schobel (1992) is presented. Conditional on the
levels of current spot prices and inventories the derived futures prices span a
wide price range which is equivalently obtainable from using a cost-of-carry
model up to using a pure equilibrium model.

159Schwartz and Smith (1997) show that their model is equivalent to the convenience
yield model of Gibson and Schwartz (1990).
160 However, it is not a continuous time pricing model but formulated in discrete time.
18 Electricity Pricing Model

18.1 Pricing Electricity Derivatives

Comparing the pricing of electricity derivatives to the existing models for
traditional commodities as presented in section 17.3 we need to encompass
the unique characteristic of non-stombility of power. This peculiarity of the
electricity market has several crucial implications for derivatives pricing:

(i) The non-storability feature means that due to a lack of any power
inventories electricity must be produced at exactly the same time as
it is consumed. This creates a load-matching problem in that the
utility industry needs to discover the value of lost load and decide on
the optimum amount of reserve margin to provide. 161

(ii) The non-storability problem could be diminished if there would be a

physical possibility of economically transferring power from an over-
production region to a consumer region. However, the US power mar-
kets are geographically distinct with several regions serving as delivery
points for electricity futures contracts. Currently, there are no satis-
factory transmission policies but inter-regional price differences will
decrease as deregulation proceeds and the utility industry is required
to open up its transmission systems. 162

(iii) If electricity is practically non-storable, no inventories can be hold and

thereby it is impossible to specify the positive returns from owning the
commodity for delivery if power is not storable. Thus, the convenience
161See, for example, Woodley and Hunt (1997, p. 44 f.).
1620n the isssue of rationalising the US power transmission business see, for example,
Barber (1997).
190 18. Electricity Pricing Model

yield models can hardly be extended to price electricity derivatives.

This incomplete market situation asks for different pricing models to
value electricity derivatives. 163

(iv) Due to the load-matching problem electricity prices show highly volatile
seasonal, weekday and even intra-day patterns. These patterns per-
sist since there are no arbitrage possibilities due to the non-storability
feature of electricity. The hourly changing price behavior needs also
to be considered in specifying appropriate electricity derivative con-
tracts which are useful for pricing, hedging, and risk management
applications. 164

(v) Finally, the fact of non-storability has an enormous impact on the

investment and operating decisions of utilities and other major pro-
ducers and consumers of electricity. Thus, the non-storability plays a
crucial role in determining the relevance and the range of applications
of electricity derivatives. 165

All these implications need to be considered with their impact on the

pricing of electricity derivatives. The literature put forward in this rela-
tively new pricing segment can be categorized into two approaches: The
production cost models, also known as physical models, and the financial
models. The stochastic specifications of the proposed models are given in
table 18.I.
Models of the production cost type, also known as physical models, in-
clude models which simulate the physical operation of the generation and
transmission system of a single utility or a whole regional electricity mar-
ket. Using certain variables such as fuel prices, hydro condition and load
the models are capable of developing different market price scenarios of
power possibly with accompanying statements about probability distribu-
tions. 166 A publicly available pricing model of the production cost type is
163For suggestions of different pricing models see, for example, Eydeland and Geman
(1998).
164For the marked volatility in energy prices that is both high and variable over time
see, for example, Duffie and Gray (1995) and Pokalsky and Robinson (1997).
165See, for example, Pokalsky and Robinson (1997) and Leong (1997).
166In applying these models for pricing purposes it needs to be ensured that power
derivatives are valued risk neutrally. Leong (1997), for example, sees the need of properly
distinguishing between power forward prices and forecasts of power prices.
18.2. Model Assumptions and Risk Neutral Pricing 191

Table 18.1: Electricity Pricing Models

Authors Model Specification Constants

Pilipovic dS = K, (L - S) dt + aSdz K"a,
(1998) dL = J.LLdt + ~ Ldw J.L,~
dzdw = 0
Pirrong dq = (J.L + k (lnq - 8)) qdt + aqqdu - dL k,q
(1999) dJ = af (j, t) Jdt + af (j, t) Jdz
dudz = Pafdt Paf

proposed and implemented by Pirrong and Jermakyan (1999). In light of

the time-dependent variations in power prices and the non-linear relation
between load and spot power prices they develop an equilibrium pricing
model where the spot price is a function of two state variables. For the
independent variables they choose the load q and the fuel price J resulting
in an equilibrium contingent claim pricing model where they especially care
about the non-storability of power.
A second category is build by the financial models which derive fair
values for electricity derivative prices based on standard contingent claim
pricing models in continuous and discrete time. A continuous time val-
uation model for electricity forwards is presented by Pilipovic (1998) who
derives a factor pricing model based on a mean-reverting specification of the
electricity spot price S and its long-run equilibrium value L. A proposed
implementation of the model uses the observable electricity forward curve
to calibrate the model parameters to the market volatility matrix.

18.2 Model Assumptions and Risk Neutral

Pricing
A proper specification of the model assumptions goes along with the deci-
sion on the pursued pricing objective. Our further attention concentrates
on pricing short-term electricity derivatives, a choice which is partly moti-
vated by the scarcely available and often illiquid data history on electricity
derivatives.
Our choice is the Californian power market which provides historical
spot as well as forward price data. In figure 18.1 we graph the power prices
192 18. Electricity Pricing Model

of electricity in California for the period of 04/01/1998 to 12/31/1999. 167

Depicted are daily spot and day-ahead prices for 2 p.m. From the graph
we note that the spot and forward prices exhibit non-linearities. The prices
show extreme spikes as well as high volatility which changes rapidly over
short time periods.

250 - - Spot Prices

........_........ Forward Prices

_200
..s:::
~
:iE

-
......
~ 150
Q)
u
.;::
a.
~ 100
'0
.;::

~
iIi 50

04-----r------r~~---,--------~----__,

01/01/99 03114/99 OS/26/99 08/07/99 10/19/99 12131/99

Figure 18.1: Electricity Spot and Day-Ahead Forward Prices in California

For the selected pricing segment of day-ahead forwards we develop a

pricing model that incorporates a stochastic volatility pattern following the
suggestions in the current literature:

"Prices in the energy markets are marked by a volatility that

is both high and variable over time. These characteristics mean
that [... ] the ability to track and forecast volatility is of paramount
importance when trading and hedging energy related portfolios
of derivatives." 168
167See section 19.2 for a detailed description of the selected data sample and the esti-
mation results of the empirical inferences.
168Duffie and Gray (1995, p. 39).
18.2 Model Assumptions and Risk Neutral Pricing 193

" Stochastic volatility is certainly necessary if we want a diffusion

representation to be compatible with the extreme spikes as well
as the fat tails displayed by the distribution of the realized power
prices." 169

In developing a pricing model for short-term electricity derivatives, we

see the mean-reversion pattern modeled on the spot price, as with the mod-
els by Pilipovic (1998) and Pirrong and Jermakyan (1999), dominated by a
stochastic specification of the spot volatility. However, for long-term elec-
tricity contracts we find it necessary to model the spot price by a mean-
reverting diffusion. 17o The stochastic volatility model we present is based
on the model of Heston (1993).171

Assumption 18.2.1 (State Variables) We specify the stochastic dynam-

ics for the state variables St and Vt, with St being the spot price of electricity
and Vt the variance rate of the spot price, as in Heston {1993} by the pro-
cesses under the objective probability measure JP>.

J-tStdt + StJVtdWE,t, and (18.1)

J-tv dt + aJVtdW~t, (18.2)

with the drift of the variance rate process specified as J-t v = K (8 - Vt). We
further assume that dWE,tdW:,t = pdt to cover possible correlation between
the two state variables.

Examining the stochastic behavior of St assumed in equation (18.1)

shows that the spot price follows a geometric Brownian motion with a
stochastic specification of the volatility term. In our case of developing a
valuation model for pricing electricity forwards we need to modify the stan-
dard Heston (1993) valuation model which is originally intended to price
equity, bond, and currency claims.

Assumption 18.2.2 (Electricity Extension) The extension of the Hes-

ton {1993} model is based on the idea of Ross {1997} on equilibrium pricing
169Eydeland and Geman (1998, p. 73).
170For an analysis on the impact of mean-reversion processes on, for example, interest
rate contingent claims see Uhrig-Homburg (1999).
171 Other proposed stochastic volatility models from the literature include Hull and
White (1987), Stein and Stein (1991), and Sch5bel and Zhu (1999).
194 18. Electricity Pricing Model

of assets that cannot be stored and held over time except at high cost. Given
electricity as an underlying commodity that is non-storable, we face a market
situation that is incomplete. In this case both state variables - the spot price
of electricity St and its variance rate Vt - cannot be hedged. 172 In thus, the
state variables must be risk adjusted by incorporating market prices of risk in
order to develop arbitrage-free pricing formulas for electricity derivatives. 173

Using the market prices of risk, we are able to derive the risk neutral
dynamics of the state variables in order to price electricity derivatives under
the corresponding martingale measure in an incomplete market. In order to
obtain the corresponding martingale process for the electricity spot price,
we use the Girsanov transformation dW~t = dW%,t + A* yVidt with a time
invariant market price of risk A*. Based on equation (18.1) the stochas-
tic development of the transformed spot price X t = In St leads under the
Girsanov transformation to the stochastic behavior

(18.3)

for the first model factor using the substitution A = ~ + A*. In the case of
the spot price variance rate Vt as the second state variable, we use the same
adjustment as suggested in Heston (1993) . This is given by

(18.4)

with the market price of risk being Av (St, Vt, t) = AvVt where Av is a con-
stant. Based on the risk neutralized processes (18.3) and (18.4) of the two
state variables we are able to price forward contracts arbitrage-free as we
demonstrate in the following section, using martingale methods developed
by Geman, Karoui, and Rochet (1995) and Scott (1997).

18.3 Valuation of Electricity Forwards

Let the current market price of a forward contract with time T - t until
maturity be denoted by F (t, T, St, Vt; 1/J) in our model, where 1/J denotes the
1721n Heston (1993) the spot price diffusion models the dynamics of a stock price which
is a tradeable and therefore hedgeable asset.
1730n the relevance of incorporating risk premia in the valuation of power derivatives
see, for example, Pirrong and Jermakyan (1999).
18.3 Valuation of Electricity Forwards 195

vector of model parameters. From equation (17.1) we know that forward

prices are equal to the expected future spot price under the risk neutral
measure Q

(18.5)

assuming deterministic interest rates.

To calculate the expectation of equation (18.5) we first derive the in-
tegral solutions of the log-transformed spot state variable X t . Based on
the stochastic behavior of the Xt-process in equation (18.3) we obtain the
solution

J + JVV;d~t
T T

XT = X t + /-l (T - t) - A Vt dt (18.6)
t t

for the relevant time horizon [t, TJ. Therein, we can further get an expression
for the time integral from the variance rate specification of equation (18.4).
Working analogously with an integral solution we get

1~ v,dt " : A. [I<i! (T - t) - (Vr - v,) + 0" 1yV;~,] . (18.7)

Using the results of equations (18.6) and (18.7) we intermediately resolve

the following expression for the expectation of equation (18.5):

F (t, T) =]EQ e
Xt+J1.{T-t)- ;>. J J
[t<:IJ{T-t)-{VT-Vt)+". yVidw:!t] + yVidw2 t
[ 0< v t ' t •

Therein, we need to take a further look at the two remaining corre-

lated 174 stochastic integrals in the exponent. To compare these two martin-
gales we use the fact that the transformation

174The correlation coefficient is assumed to be dW~tdW~t = pdt.

196 18. Electricity Pricing Model

results in two uncorrelated processes, i.e. dW'?tdWIQi

, = O. This brings us to

F (t, T) = ]E1Qi [ e
Xt+JL(T-t)-
I<
;>. v
[1t1J(T-t)-(VT-vt)+<TJ
t '
ytitd~t]

(18.8)

in which we can work with Ito's isometry on the last integral since the
Brownian motion WtlQi is independent of the variance rate variable. Thereby,
we obtain for the last term

J J
T T

";1- p2 .jV;dWtlQi = ~ (l-l) Vt dt

t t

under the Q-expectation in the exponent.

However, the first type integral in equation (18.8) further remains un-

1Jiit~, ~ ~ [-.e(T - 1Wit] .

solved. But from equation (18.4) we are able to obtain the expression

t) + (t>r - v,) + (K + \,)

Putting the results together and bringing the constant terms with re-
spect to the information available in F t outside the conditional expectation
we get the expression
F (t, T) = eXt-;Vt+(JL-t;!p)(T-t)

lllQ [e'''T+-"''''P-l(l-P'»{'''''' F.]. (18.9)

With equation (18.9) we have almost derived the price of the forward

-1 V,dS)
contract F (t, T) besides solving for the remaining expectation

y(t, Vt) = lll'l [exp ( k, exp (!:'t>r) F,] , (18.10)

with kl ~ +.Av
.A - -(J-P 1( 2)
- 2 1 - P ,and
p
k2 = (J
18.3 Valuation of Electricity Forwards 197

A common tool in solving such types of conditional expectations is applying

the Feynman-Kac formula which states the following.

Theorem 18.3.1 (Feynman-Kac) 175 Let {Xtlt~O be an Ito diffusion,

f(X t ) a non-negative function and assume that q is lower bounded. Then
consider the function

y(t,x)=Ex [ e
- Jq(X.)ds
0t 1
f(Xt).

For this function the backward differential equation

ay
at = Ay - q (Xt)y

holds with the necessary initial condition being y (0, x) = f (x); A denotes
the differential operator. 176

In our case of equation (18.10) the corresponding partial differential

equation under the equivalent martingale measure Q takes the form

~ a 2y(t,Vt)
2
2
a Vt a 2
vt
+ (()
/'i,
_ (
/'i, + \ )Vt) ay(t,Vt)
Av a
Vt
+ ay(t,Vt)
a
t
_ k ( )
- I VtY t, Vt ,
(18.11)

with the boundary condition for the value at maturity y (T, Vt) = exp (k 1VT).
For the solution for the differential equation we propose the exponential
affine structure

Y (t, Vt) = exp (A (t, T) Vt + B (t, T)), (18.12)

. h ay(t,Vt) (aA (t, T) ( )

WIt at at Vt + aB at
(t, T))
y t, Vt ,
ay (t, Vt)
aVt
= A (t, T) Y (t, Vt), and
a 2y (t, Vt)
A (t, T)2 Y (t, Vt),
aVt2
175The formula is named after the contributions of Feynman (1948) and Kac (1949).
176For a description of the differential operator see, for example, Oksendal (1995, p.
128).
198 18. Electricity Pricing Model

including the transformed boundary conditions A (T, T) = k2 and B (T, T) =

O.
The result of bringing equations (18.11) and (18.12) together is given by
the system of two partial differential equations:

(18.13)

0, (18.14)

where we applied the separation of variables technique. 177 The first equation
(18.13) is known as a lliccati type differential equation 178 for which we get
the solution

A (t, T) 2. [( A)_ '1'1 sinh ~'1 (T - t) + 12 cosh ~'1 (T - t)]

(J 2 '" + v 1 (T
11 II cosh 211 - )
t . h 2'1
+ 12 sm 1 (T - t) '

with 11 V('" + Av)2 + 2(J2k b and (18.15)

12 '" + Av - (J2k2

incorporating the boundary condition A (T, T) = k1 J7 9 Given the solution

to the function A (t, T) of equation (18.12) we can integrate the second
differential equation (18.14) as

-"'(} J
T

B (T, T) - B (t, T) A (t, T) dt,

t
with B (T,T) o
to get a result for the function B (t, T). Working out the integral we obtain

"'(}
(J [('" + Av) (T - t)
B(t,T') = 2"

-2ln {COSh ~'1

2
(T - t) + 12 sinh ~'1 (T -
11 2
t)}](18.16)

177 See, for example, Ingersoll (1987, p. 397) for an application with interest rate con-
tingent claims.
178 This type of differential equation commonly arises with the squared Gaussian model
of Cox, Ingersoll, and Ross (1985b); see, for example, Rogers (1995, p. 99 f.).
179For a similar solution to equation (18.13) see, for example, Rogers (1995, p. 100)
and Stein and Stein (1991, p. 730).
18.3 Valuation of Electricity Forwards 199

for the second function in equation (18.12). Thus, we have finally calculated
an expression for the remaining expectation y(t, Vt) of equation (18.9).

Solution 18.3.2 (Forward Price) Having gone through the calculations

we are now able to state our final result for the time t value of a forward
contract on electricity:

F (t , T , S t, v·t, ~/.)
0/
= S t e(A(t,T)-;)vt+B(t,T)+(JL-~p)(T-t) , (18.17)

where the functions A (t, T) and B (t, T) are given by the expressions of
equations (18.15) and (18.16), respectively, with 1/J = {/-L, K" (), a, p, A, Av}.

The full curve of forward prices at a given date t is given by applying

different times to maturity T - t to equation (18.17) for the same values of
the state variables X t and Vt.
19 Empirical Inference

19.1 Estimation Model

The pricing model we propose for the valuation of short-term electricity
forwards is based on a factor specification of the spot electricity price and
its variance yielding a closed-form solution for the forward curve as given in
equation (18.17). This theoretical model can conveniently be transformed
into an equivalent empirical state space model whereupon we are able to
build a Kalman filtering algorithm to calibrate our forward pricing model
to available empirical data. From chapter 3 we know that a state space
model consists of two major pillars: the transition and the measurement
equation. In the case of our forward pricing model the transition equation
captures the dynamics of the state variables. The measurement equation
provides us with a link of the spot electricity price to the forward prices
of electricity. In the following we first focus on specifying the transition
equation by examining the distribution of the state variables.

19.1.1 Distribution of the State Variables

The distribution of the state variables is approximated by the first two ex-
act conditional moments for the state variables. These moments are implied
by the stochastic process specifications of the factors X t and Vt. For conve-
nience we restate the dynamic specifications of the state variables in integral
form as follows:

J J J
s s s

dX( = [p - ~V(] d( + JV(dWk,(, and (19.1)

t t t
202 19. Empirical Inference

s s s

j dV(;=j K(O-vdd(+ j O"y'v(dW?,(, (19.2)

t t t

for a period oftime T = [t, s]. These equations fully contain the information
to derive the conditional moments of the state variables. In order to calcu-
late the moments we employ a technique of putting up differential equations
that contain the conditional moments of interest as functions. With the rel-
evant boundary conditions we then solve these differential equations for the
moments. 180
For the conditional first moments we take the expectations on equations
(19.1) and (19.2). To calculate the conditional mean of Vs we take the
derivative with respect to the upper bound s of the integrals which yields

olElP[vsl.Ttl_ 0 IT]
oS - K + KJrJlU'1!'[ vs.rt· (19.3)

Exploiting the initial condition of IEI!' [Vtl .Tt ] = Vt we solve the differential
equation (19.3) as

(19.4)

which resembles the solution to the conditional expectation in the interest

rate model of Cox, Ingersoll, and Ross {1985b).181 The conditional mean of
the first state variable X t can then be derived from equation (19.1) where
we take expectations:
s

IBt [Xsi .Tt ] = X t + f-£ (s - t) - ~ j IEIP [v(l.Ttl d(.

t

Using the expression for lEI!> [vsl.Ttl from equation (19.4) we can calculate
the solution

(19.5)

180 This method is, for example, used by De Jong and Santa-Clara (1999) in the context
of a factor based interest rate model a la Heath, Jarrow, and Morton (1992).
181Compare Cox, Ingersoll, and Ross (1985b, p. 392).
19.1 Estimation Model 203

for the conditional expectation of the log-spot price X t •

In order to derive the second moments of the state variables we work on
the basis of the statistical relationship:

Given the conditional means EIP' [Xa IFtl and EIP' [Va IFtl we further need
to calculate the first terms in equation (19.6) to get solutions for the three
conditional second moments. Therefore, we first obtain the stochastic dy-
namics for the functions it = Xl, h = XtVt, and fa = v; using Ito's
formula. This yields the following dynamics:

dX; (2/l'xt + Vt - XtVt) dt + 2Xtyfv;dWI,t,

dXtvt (""()Xt + (p, + P(1) Vt - ~V; - ""XtVt) dt

+v:12dWr,t + (1Xt yfv;dW?'t, and

dV; = ( (2,.,,() + (12) Vt - 2,."v;) dt + (1 yfv;dW?'t.

Based on these stochastic differential equations we can again - as with the

conditional first moments - take expectations and differentiate with respect
to s. This directly transfers these equations to a system of three differ-
ential equations with the functions being the second moments EIP' [X;I Fd,
EIP' [Xsvsl Ftl, and EIP' [v;1 Fd:

(19.7)

Additionally, the corresponding initial conditions are given by EIP' [xli Fd =

xl, EIP' [XtVtl Ftl = XtVt, and EIP' [v;1 Fd = v;. Further, we solve the system
of equations given in (19.7) beginning with the last equation and ending
204 19. Empirical Inference

with solving the first equation. Finally, we use these results along with
the statistical relationship stated in equation (19.6) to derive the following
expressions for the second moments. First, we obtain the solution to the
conditional variance of the second state variable from the third equation of
(19.7) as:

which matches the formula stated in Cox, Ingersoll, and Ross (1985b, eq.
19). Given the result of equation (19.8) we are further able to calculate the
covariance term of the state variables as

(19.9)

Finally, using the covariance term we obtain

for the conditional variance of the log-spot price X t .

To summarize the derived results, we obtained formulas for the exact
theoretical first two conditional moments of the state variables. The so-
lutions for the conditional expectations are given in equations (19.4) and
(19.5) and for the variance terms with equations (19.8), (19.9), and (19.1O).
19.1 Estimation Model 205

19.1.2 State Space Formulation and Kalman Filter

Setup
In section 4.5 we introduced the concept of extended Kalman filtering in
addition to the linear Kalman filter, where the transition and measurement
equations are linear in the state variables and the error terms are assumed
to be normally distributed. In the case of implementing our forward pric-
ing model we choose to implement an extended Kalman filter algorithm,
since the state variables et = [Xt, vtl' do not follow a Gaussian distribution.
With the hypothesized forward pricing model we can impose high theoretical
modeling restrictions both in the time-series as well as in the cross-section
properties. First, assuming the stochastic dynamics of the state variables
following the process specifications of equations (18.1) and (18.2) we can
infer the corresponding transition equation. The time-series properties of
the model can be deduced from the results of the first and second moments
as derived in the previous section. Second, given the explicit forward pricing
formula of equation (18.17) we can implement cross-sectional restrictions on
the data. There, we assume that the theoretical forward prices match the
observable prices on the power exchange only with small discrepancies.
Let us begin with the general specification of the tmnsition equation by

as formulated in section 4.5. The conditional expectations of equations

(19.4) and (19.5) directly lead to the expression

~e) t:lt + :1<

(1- e-I<~t) ]
[ (JL - e (1- e-I<~t)
1 _-.L (1- e-I<~t) ]
+[ 0 21< e-I<~t et-~t

+7Jt(et-~t,¢) (19.11)

for the transition equation. However, within this equation the error term
7Jt (et-~t, 'I/J) still contains the state variable et-~t which is not observ-
able. Thus, we approximate the transition equation around (et-~t, 7Jt)
(et-~tlt-~t' 0) which results in:
206 19. Empirical Inference

Therein, the transition matrix is given by

8ht (et-At, "'t, 1/J)

<Ptlt-At =
8e~-At

= [~ (19.12)

and the error term matrix Stlt-At can be derived as

Stlt-At =
(et-at,TJt,1/J )=(et-atlt-at,o,1/J )

[
VarF!' [ Xt 1Ft - Ad 1F
CovF!' [XtVt t - At ] ]
(19.13)
CovF!' [ XtVt 1Ft - Ad 1F
VarF!' [ Vt t - At ]

where we use the expressions of equations (19.10), (19.9), and (19.8) and
substitute values Vt-Atlt-At for Vt-At. Thus, we can finally state our transi-
tion equation of the state space model as

et ~ Ct + <Ptlt-Atet-At + Stlt-At"'t, (19.14)

with Ct = ht (et-Atlt-At, 0, 1/J) - <Ptlt-Atet-Atlt-At

where the formulae of h t (et-Atlt-At, 0, 1/J), <Ptl t - At , and Stlt-At are given in
equations (19.11), (19.12), and (19.13).
Next, we work out the specification of the measurement equation where
we use the forward pricing formula derived in section 18.3. The idea is to
comprehensively use market information contained in both the spot electric-
ity prices as well as the information from the forward prices of electricity.
Bringing both market informations together, we formulate the measurement
equation as an identity between the observable market prices and the the-
oretical prices. However, the two pairs of prices do not need to fully match
in that we indude a measurement error to allow for some discrepancies be-
tween the market and model prices. We obtain the measurement equation
by exploiting the functional relationship of the state variables et with the
forward price from the original pricing formula given in equation (18.17)

In F (t, T) = (It - ~ p) (T - t) + B (t, T) + [1, A (t, T) - ;] et,

19.1 Estimation Model 207

restated in logarithmic form. Denoting the observable prices on the elec-

tricity spot and forward markets by Yt = [Xt, In F (t, T)]' we are able to use
the general specification of the measurement equation

(19.15)

as known from chapter 3. The measurement equation (19.15) is further

particularized by the system matrices

ad1/J) [ (It - ~ p) (T ~ t) + B (t, T) ] , and

B t (1/J) = [~ A (t, ~) _ ~ ] .
For the moment specification of the error term et (1/J) in the measurement
equation (19.15) we assume that the errors have zero mean, Le.lE [etIFt-atl =
0, and are serially uncorrelated. The errors' time invariant covariance ma-
trix is further modeled by

o ].
2
ac:,F

This covariance specification models the noise observable in the spot

market separately from the noise seen in the forward markets. Thereby, we
add two more parameters, ac:,s and ac:,F, to our parameter space 1/J of the
theoretical futures pricing model.
Thereby, we set up the specific characteristics of the state space model
for the electricity spot and forward market using the transition and mea-
surement equations from (19.14) and (19.15). Based on this state space
formulation we are able to run an extended Kalman filter algorithm in or-
der to estimate our parameter values 1/J = {It, K, 0, a, p, A, Av, ac:,S, ac:,F} by
means of maximum likelihood according to

(19.16)

In our case, where the conditional normal distribution is approximated

by the derived second moments of equations (19.8), (19.9), and (19.10), the
log-likelihood can be used to yield quasi maximum likelihood parameter
208 19. Empirical Inference

estimates. 182 The asymptotic statistical properties of quasi maximum like-

lihood estimates for heteroskedastic models, such as the Cox, Ingersoll, and
Ross (1985b) and our stochastic volatility model, are discussed in Boller-
slev and Wooldridge (1992). They show that the quasi maximum likelihood
estimates are consistent and asymptotically normal under fairly weak reg-
ularity conditions. Even though the assumption of normality is violated
when maximizing the normal log-likelihood, these results are obtained since
the score of the normal log-likelihood has the martingale difference property
when the first two conditional moments are correctly specified. Thereby, one
can correctly estimate not only the mean parameters but also the second
moments.
For our model the conditional means and variances should probably be
different from the true moments of the system since we use a linear projec-
tion of the state variables. This linear approximation is needed to incorpo-
rate the non-normal nature of the model into the typical extended Kalman
filter framework. The approximation, however, can be expected to work
well because it is linearly optimal. The Monte Carlo studies in Bollerslev
and Wooldridge (1992) and Duan and Simonato (1995) indicate that the
asymptotic results carryover to finite samples, i.e. it is reasonable to expect
the parameter vector ib to be approximately consistent and asymptotically
normal. Further, the small sample properties of the quasi maximum like-
lihood estimates in the case of single and multi-factor models a la Cox,
Ingersoll, and Ross (1985b) are studied in detail by Duan and Simonato
(1995), Chen and Scott (1995), and Ball and Torous (1996). The various
simulation evidence therein shows that the sampling properties of the quasi
maximum likelihood estimator based on the extended Kalman filter provide
satisfactory results.
In the following section we perform a general analysis of the electric-
ity data set and present the results on empirical inferences of a one-factor
pricing model and our suggested stochastic volatility valuation model.

19.2 Data Analysis and Estimation Results

In this section we apply the electricity forward pricing models to study the
Californian electricity market. This first and largest totally open electric-
182See, for example, Gourieroux, Monfort, Renault, and Trognon (1984).
19.2 Data Analysis and Estimation Results 209

ity wholesale and retail market has two core institutions, the California
Power Exchange (CalPX) and the California Independent System Operator
(CaISO). We briefly describe their operations on the Californian electricity
market, before we come to the data analysis and present our estimation
results. 183
The CalPX is a non-profit, public benefit corporation which is regulated
by the FERC. As a power commodity exchange it provides an efficient,
competitive marketplace by conducting an open, non-discriminatory trad-
ing process for qualified electricity suppliers and purchasers. The CalPX
determines the price of electricity in the hour-ahead and day-ahead mar-
kets of forward contracts to deliver a given quantity of electricity over a
one hour period. For each delivery hour the exchange collects demand and
generation bids in prices and quantities to settle for unconstrained market
clearing prices. Since its opening on March 31, 1998 the CalPX's day-ahead
market has traded between 80 and 90 percent of California's electricity mar-
ket. This day-ahead market is open from 6 a.m. to 1 p.m. when sellers and
buyers of electricity can submit their portfolio supply and demand bids for
24 hourly auction periods for electricity delivery on the next day. There-
upon, the market clearing prices are determined separately for each auction
hour. Besides providing marketplaces for electricity forwards, the CalPX
serves as scheduling coordinator in that it submits balanced supply and
demand schedules along with information on the generating units and the
locations for delivery to the CaISO.
The CaISO serves as the control area operator for the Californian elec-
tricity market where it provides open access to the transmission and man-
ages the real time operations of the transmission grid. The objective is to
provide reliable system operations, primarily keeping the right frequency
and providing sufficient generation, which are maintained by buying and
providing ancillary services. The CaISO matches the power output of the
generating units with the power demand within the controlled electric power
system including the scheduled interchange with adjacent control areas.
This real time balancing of load and generation is based on the coordina-
tion of the hour-ahead and day-ahead schedules dispatched by the CalPX.
On the basis of the real time operations, the CaISO determines and pub-
lishes real time spot electricity prices by hour.

183The Californian power market is explained in detail in, for example, CalPX (1999).
210 19. Empirical Inference

In tables 19.1 and 19.2 we present descriptive statistics on the Califor-

nian electricity prices from the spot and day-ahead forward markets. The
data are sampled separately for each hour of the day from the opening of the
CalPX on 04/01/1998 until 12/31/1999. The spot prices for delivery of one
MWh of electricity average around $29.62 with a mean standard deviation
of $27.83 which is very high. The prices take a wide spectrum from prices
close to zero up to values of $492.20. Comparing the spot price averages of
the mean with the median value of $24.24 indicates the spiky price behav-
ior with large positive outliers. This price pattern is also described by the
right skewness with values ranging from 0.6 to 7.5 with a mean of 3.5. The
distribution of the spot prices further exhibit large excess kurtosis for all
24 time-series. The statistics on the forward prices in table 19.2 are simi-
lar to the spot price results. The average forward prices are slightly lower
with an average of $28.67 but exhibit a higher average median value with
$25.75. The standard deviation of $18.53 is remarkably lower than the spot
price volatility showing that forward contracts are more stable over time
than spot prices. However, the data sample also contains rather extreme
time-series with those of the 6th, 10th, 19th, 20th, and 21th hour which
are characterized by a high positive skewness and an excess kurtosis. From
these characteristics we see that the Californian electricity prices exhibit a
spiky pattern and are highly volatile which was already expected ahead of
starting to trade electricity.184
A further examination of the relationship between spot and forward
prices is reported in table 19.3. There we analyze the difference between
the forward prices of day t - 1 for electricity delivery on day t with the
realized spot prices on day t. The left part of the table contains the sample
statistics which show that the average spot price lies above the forward price
by $1.10. However, the range of the price differences is rather large with a
maximum value of $714.49 and a low of minus $426.40 where the medians
take values around $0.94. Further, the right side of table 19.3 reports the
results of a regression of the realized day t spot prices against the forward
prices of day t - 1. This enables us to evaluate the question whether the
forward prices are biased or unbiased predictors of next day's realized spot

184The major factors for the price patterns are seen in the fuel price, load uncertainty,
variations in hydroelectricity production, generation uncertainty, and transmission con-
gestion; see, for example, CalPX (1998).
19.2 Data Analysis and Estimation Results 211

Table 19.1: Sample Statistics of Electricity Spot Prices

Hour MEAN STD MAX MED MIN SK KU
1 20.66 14.85 107.14 19.03 0.00 1.6 8.9
2 20.10 14.58 134.35 19.36 0.00 1.5 10.0
3 19.35 13.99 99.99 18.51 0.00 1.2 6.8
4 17.82 12.16 72.11 17.00 0.01 0.6 3.3
5 17.36 12.21 75.00 17.50 0.00 0.6 3.4
6 18.99 14.08 148.78 18.52 0.00 2.6 22.7
7 22.88 19.19 172.14 20.65 0.00 2.9 18.6
8 29.11 33.63 434.37 25.00 0.00 6.5 61.1
9 29.21 25.72 331.36 25.50 0.01 5.1 45.7
10 30.30 28.70 360.00 25.76 0.01 6.4 60.4
11 32.56 32.03 492.20 26.79 1.00 7.5 86.4
12 33.49 32.74 414.48 27.00 0.01 6.3 57.5
13 32.66 28.98 250.00 26.04 0.01 4.3 27.1
14 37.26 40.09 250.00 26.70 0.07 3.7 17.7
15 39.19 45.78 250.00 26.90 0.01 3.5 15.2
16 41.28 52.77 274.44 26.00 0.10 3.1 11.8
17 38.48 47.00 250.00 26.05 0.43 3.4 14.9
18 39.85 43.64 308.55 28.37 0.01 3.4 16.2
19 38.46 41.56 252.06 27.85 0.00 3.2 14.7
20 36.25 34.83 256.00 29.00 0.01 3.7 19.9
21 34.48 28.03 250.00 28.82 0.00 4.1 26.0
22 30.82 18.98 152.00 27.84 0.01 2.8 15.6
23 27.96 18.47 242.76 25.89 0.01 4.8 50.2
24 22.33 14.00 117.59 21.65 0.01 1.3 8.4
MEAN 29.62 27.83 237.31 24.24 0.07 3.5 25.9
STD 7.89 12.64 114.34 3.98 0.22 1.9 22.1
MAX 41.28 52.77 492.20 29.00 1.00 7.5 86.4
MIN 17.36 12.16 72.11 17.00 0.00 0.6 3.3

Notes:
Prices are stated in $/MWh. For each distribution of hourly prices
we report the mean, the standard deviation, the maximum, the
median, the minimum, the skewness, and the kurtosis. Summary
statistics across all time-series are given at the bottom of the table.
212 19. Empirical Inference

Table 19.2: Sample Statistics of Electricity Forward Prices

Hour MEAN STD MAX MED MIN SK KU
1 20.90 9.75 82.00 21.00 0.01 0.7 5.9
2 19.34 9.79 82.00 19.50 0.00 0.8 6.0
3 18.29 9.96 82.00 18.00 0.00 0.8 6.0
4 18.00 9.95 82.00 17.53 0.01 0.8 6.0
5 18.73 10.16 101.35 18.09 0.00 1.1 9.3
6 21.06 13.66 249.00 20.23 0.00 7.4 122.9
7 23.14 13.08 182.51 23.10 0.00 3.3 37.5
8 26.42 13.22 146.59 25.49 1.84 2.5 18.0
9 28.41 15.56 250.00 27.00 0.99 5.9 73.1
10 30.74 30.17 725.00 28.15 4.00 19.3 440.4
11 31.43 15.87 250.00 28.96 4.99 5.4 61.8
12 31.80 16.36 250.00 28.81 4.99 5.0 54.8
13 32.96 18.93 250.00 28.69 4.93 4.3 36.5
14 35.01 21.98 178.65 28.97 4.93 3.0 15.5
15 36.91 26.72 213.13 28.92 4.93 3.1 14.4
16 37.64 29.67 250.00 28.99 3.01 3.3 16.1
17 37.54 27.92 221.27 29.34 4.00 3.1 14.2
18 37.38 24.90 200.20 30.56 4.94 3.0 14.4
19 36.18 33.71 725.00 30.17 4.50 13.9 275.6
20 34.12 32.53 725.00 29.41 4.99 15.5 321.3
21 33.88 30.87 725.00 29.99 6.79 17.8 395.2
22 29.07 10.24 104.02 28.01 6.79 1.7 10.5
23 26.56 10.28 104.60 26.03 4.99 1.4 10.4
24 22.56 9.35 82.10 22.98 3.01 0.7 5.9
MEAN 28.67 18.53 260.89 25.75 3.11 5.2 82.1
STD 6.92 8.70 221.68 4.43 2.39 5.6 131.8
MAX 37.64 33.71 725.00 30.56 6.79 19.3 440.4
MIN 18.00 9.35 82.00 17.53 0.00 0.7 5.9

Notes:
Prices are stated in $/MWh. For each distribution of hourly prices
we report the mean, the standard deviation, the maximum, the
median, the minimum, the skewness, and the kurtosis. Summary
statistics across all time-series are given at the bottom of the table.
19.2 Data Analysis and Estimation Results 213

Table 19.3: Comparison of Electricty Forward and Realized Spot Prices

Sample Statistics on Difference of Regression Results
Forward Prices - Realized Spot Prices Intercept Slope
Hour MEAN STD MAX MED MIN SK KU EST STD EST STD
1 -0.01 12.75 57.00 -0.03 -77.80 -1.2 11.3 4.68 1.27 0.78 0.05
2 -0.97 11.91 66.12 0.13 -98.76 -1.3 15.2 3.99 1.13 0.85 0.05
3 -1.30 11.09 49.43 -0.59 -77.05 -1.3 11.6 4.28 1.00 0.84 0.05
4 0.06 10.08 55.54 0.01 -52.81 0.2 6.1 5.45 0.87 0.71 0.04
5 1.16 9.98 71.55 1.03 -39.73 0.7 8.9 4.48 0.86 0.71 0.04
6 1.95 14.29 212.95 1.97 -105.79 4.2 87.4 9.67 0.95 0.47 0.04
7 -0.01 16.35 143.76 1.60 -124.86 -1.2 25.9 5.15 1.40 0.79 0.05
8 -3.00 30.36 73.98 0.21 -381.73 -6.6 66.2 -0.63 2.77 1.14 0.09
9 -0.94 22.94 101.24 0.64 -278.35 -4.6 48.8 7.34 1.90 0.78 0.06
10 0.41 35.89 649.00 1.63 -286.31 7.1 184.0 22.83 1.57 0.25 0.04
11 -1.07 29.35 218.00 1.28 -426.40 -6.4 90.5 6.55 2.57 0.83 0.07
12 -1.83 30.02 218.11 1.07 -339.61 -5.1 57.8 7.49 2.60 0.82 0.07
13 0.15 26.62 215.76 1.40 -210.10 -1.9 25.4 10.33 2.07 0.68 0.05
14 -2.27 34.60 137.92 1.48 -217.23 -2.6 15.9 4.85 2.58 0.93 0.06
15 -2.33 40.85 164.40 1.81 -219.80 -2.4 14.7 9.66 2.75 0.80 0.06
16 -3.71 46.43 192.86 2.50 -220.92 -2.0 11.4 8.93 2.97 0.86 0.06
17 -0.97 40.69 177.85 3.00 -218.41 -2.2 13.7 6.35 2.69 0.86 0.06
18 -2.67 37.03 158.87 1.91 -263.25 -2.2 14.8 5.16 2.67 0.93 0.06
19 -2.54 46.79 714.49 2.34 -220.00 4.3 93.9 28.07 2.37 0.30 0.05
20 -2.28 40.82 671.09 0.60 -224.31 5.7 124.2 26.78 1.94 0.29 0.04
21 -0.73 37.34 705.30 0.20 -211.62 9.7 207.1 28.60 1.62 0.18 0.04
22 -1.88 17.47 70.43 -0.81 -114.65 -1.9 13.0 9.24 2.09 0.75 0.07
23 -1.48 17.12 62.00 -1.10 -200.77 -4.3 51.5 8.97 1.88 0.72 0.07
24 -0.11 11.42 47.17 0.17 -72.59 -0.8 8.0 3.89 1.23 0.84 0.05
MEAN -1.10 26.34 218.12 0.94 -195.12 -0.7 50.3 9.67 1.91 0.71 0.05
STD 1.38 12.65 221.70 1.08 104.09 4.1 56.0 8.16 0.70 0.24 0.01
MAX 1.95 46.79 714.49 3.00 -39.73 9.7 207.1 28.60 2.97 1.14 0.09
MIN -3.71 9.98 47.17 -1.10 -426.40 -6.6 6.1 -0.63 0.86 0.18 0.04

Notes:
The price differences are stated in $/MWh and their distribution is characterized by
values for the mean, the standard deviation, the maximum, the median, the minimum,
the skewness, and the kurtosis. Further, we report results on a ordinary least squares
regression of forward prices on spot prices. At the bottom of the table we show summary
statistics across all time-series.
214 19. Empirical Inference

prices. The null hypothesis for unbiasedness requires the intercept term
and the slope coefficient to be statistically not different from zero and one,
respectively. Looking at the results indicates that the null hypothesis can be
rejected at conventional significance levels for almost all 24 time-series. The
intercept term lies above zero at an average of 9.67 with a corresponding
standard error of 1.91. The slope coefficients show values that are almost
all significantly below one with an average of 0.71 at a standard deviation of
0.05. Thus, even for a time period of only one-day ahead the forward prices
are not an unbiased predictor of realized spot prices which demonstrates
the need for sound theoretical valuation models.
In order to capture the spot price behavior and to price the day-ahead
forward contracts we select to calibrate two theoretical pricing models to
the given electricity data. First, we choose to implement the one-factor
model as parametrized in Schwartz (1997, Model 1) which originally is in-
tended to capture the price processes on oil spot and futures markets. We,
however, use this equilibrium pricing model to describe the electricity spot
price behavior by the continuous time process

dS = '" (J..l -In S) Sdt + erSdzP (19.17)

yielding a corresponding arbitrage-free log-forward price of

er 2
lnF (t, T) e-K(T-t) lnSt + a* (1- e-K(T-t)) + 4", (1- e- 2K (T-t)) ,

er 2
with a* J..l - 2", - A (19.18)

under the martingale measure Q including the time invariant market price of
risk parameter A. Based on equations (19.17) and (19.18) as transition and
measurement equation, respectively, we estimate the model using the stan-
dard Kalman filter setup. The results on the parameter estimates including
the two measurement errors erg,S and erg,F are shown in tables 19.4 and 19.5.
The coefficient of mean-reversion is estimated by 0.6 on average across all
24 time-series. The values translate to mean half-lives of the electricity spot
price with a minimum of 0.6 years for hour 16 and a maximum period of
3.8 years for the 22nd hour where the average lies at 1.1 years. The rather
not statistically significant estimates for the long-run mean J..l reach values
around 2.8. Related to the coefficient of mean-reversion is the er-parameter
which exhibits how volatile the spot prices fluctuate around their long-run
19.2 Data Analysis and Estimation Results 215

Table 19.4: Parameter Estimates for Schwartz (1997, Modell)

r;, 11 (f

Hour EST STD EST STD EST STD

1 0.287*** 0.005 2.516 2.351 2.405*** 0.381
2 0.518*** 0.042 2.569** 1.187 3.019*** 0.202
3 0.724*** 0.073 2.498 3.854 3.060*** 0.415
4 0.370*** 0.014 2.536** 1.131 2.883*** 0.352
5 0.770*** 0.077 2.561*** 0.388 3.060*** 0.532
6 0.611*** 0.237 2.444 3.373 2.668*** 0.557
7 0.645 0.690 2.871 17.650 2.746*** 0.536
8 0.294 0.390 2.501 60.805 2.161*** 0.830
9 1.022 1.291 3.451 27.236 4.126** 2.062
10 0.341 *** 0.062 2.743 61.953 2.383*** 0.881
11 0.467*** 0.102 2.829 24.289 2.810** 1.185
12 0.445*** 0.138 2.873 10.067 3.042*** 0.656
13 0.483*** 0.088 3.299 6.596 3.136*** 0.440
14 0.901*** 0.080 3.250 9.597 4.026*** 0.412
15 1.101 *** 0.311 3.505 12.857 4.526*** 0.364
16 1.170** 0.569 3.371 12.982 4.629*** 0.407
17 1.076*** 0.075 3.238 11.428 4.391*** 0.370
18 0.972*** 0.199 3.120 6.447 4.337*** 0.392
19 0.752*** 0.256 3.134 2.480 3.815*** 0.485
20 0.458*** 0.072 2.600 5.658 2.992*** 0.451
21 0.249*** 0.011 2.651 7.278 2.459*** 0.404
22 0.182*** 0.030 2.292 28.804 2.319*** 0.208
23 0.289*** 0.073 2.400 26.339 2.615*** 0.666
24 0.351*** 0.035 2.624 2.495 2.561*** 0.460
MEAN 0.603 0.205 2.828 14.469 3.174 0.569
STD 0.304 0.291 0.372 16.879 0.773 0.385
MAX 1.170 1.291 3.505 61.953 4.629 2.062
MIN 0.182 0.005 2.292 0.388 2.161 0.202

Note:
Statistically significant parameter estimates at the 10/0-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.
216 19. Empirical Inference

Table 19.5: Parameter Estimates for Schwartz (1997, Modell) (continued)

.x C7 g ,S C7 g ,F neg.
Hour EST STD EST STD EST STD LogL
1 -0.921* 0.513 0.650*** 0.033 0.181*** 0.018 653.2
2 -1.087*** 0.268 0.641*** 0.033 0.169*** 0.001 669.4
3 -0.955 0.774 0.574*** 0.029 0.232*** 0.029 727.3
4 -1.131*** 0.064 0.634*** 0.034 0.276*** 0.032 850.9
5 -0.943*** 0.333 0.675*** 0.036 0.258*** 0.031 865.2
6 -0.829 0.740 0.613*** 0.028 0.303*** 0.034 862.1
7 -0.739 5.129 0.619*** 0.033 0.367*** 0.030 965.7
8 -0.753 8.213 0.634*** 0.033 0.331*** 0.031 922.9
9 -1.237 6.387 0.638*** 0.037 0.257*** 0.081 933.7
10 -0.820 9.350 0.544*** 0.033 0.237*** 0.038 690.3
11 -0.953 4.871 0.525*** 0.034 0.193*** 0.058 612.8
12 -1.110 1.264 0.510*** 0.040 0.173*** 0.035 568.2
13 -1.075 0.752 0.466*** 0.028 0.200*** 0.027 578.4
14 -1.310 2.103 0.572*** 0.036 0.187*** 0.026 751.8
15 -1.435 2.816 0.557*** 0.031 0.181*** 0.027 763.7
16 -1.494 3.568 0.608*** 0.031 0.194*** 0.028 845.3
1" -1.444 2.674 0.591*** 0.031 0.173*** 0.029 777.0
18 -1.498 1.505 0.635*** 0.032 0.142*** 0.026 753.3
HI -1.318*** 0.215 0.687*** 0.034 0.182*** 0.035 826.4
20 -1.112 0.783 0.600*** 0.030 0.192*** 0.031 700.0
21 -0.969* 0.519 0.532*** 0.024 0.177*** 0.029 547.6
22 -0.983 2.302 0.539*** 0.031 0.111*** 0.013 361.7
23 -1.050 3.240 0.618*** 0.038 0.089** 0.04 422.7
24 -0.940 0.598 0.622*** 0.027 0.107*** 0.026 458.9
MEAN -1.088 2.458 0.595 0.032 0.205 0.032 712.9
STD 0.230 2.605 0.055 0.004 0.068 0.015 163.7
MAX -0.739 9.350 0.687 0.040 0.367 0.081 361.7
MIN -1.498 0.064 0.466 0.024 0.089 0.001 965.7

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.
19.2 Data Analysis and Estimation Results 217

Table 19.6: Inference Results for the Stochastic Volatility Model

JL A /'i, (J

Hour EST STD EST STD EST STD EST STD

1 1.865 1.375 -10.739*** 2.183 1.156*** 0.253 1.815*** 0.635
2 0.153 0.813 -4.753*** 1.289 7.359*** 1.542 9.476*** 2.053
3 -0.311 1.183 -3.298** 1.247 1.607*** 0.165 3.773*** 0.761
4 -0.092 0.446 -3.486 2.205 2.148 2.351 2.614 3.175
5 3.466 2.367 -8.236*** 1.243 0.434 1.994 0.319 0.221
6 -0.093 0.605 -11.086*** 2.862 0.982 1.819 0.686 1.276
7 -0.534 1.621 -13.006*** 2.691 5.664*** 1.698 7.107*** 1.454
8 0.010 0.037 -7.490*** 2.464 2.279* 1.211 3.301 *** 0.761
9 -0.982 1.512 -2.131*** 0.074 0.620*** 0.101 3.423*** 0.679
10 0.535 1.365 -1.788 2.411 1.760 1.862 0.678 0.513
11 -0.248 2.837 -7.294*** 1.495 2.069** 1.039 2.772*** 0.645
12 2.067 1.226 -6.055*** 1.181 1.092** 0.477 3.871*** 0.888
13 0.108 0.371 -5.549*** 1.047 1.424*** 0.488 4.222*** 0.706
14 -0.072 0.309 -3.445*** 0.602 1.535*** 0.461 6.240*** 0.985
15 0.051 0.229 -1.348*** 0.403 0.944 2.382 0.281 0.407
16 4.103** 2.129 -3.175*** 0.527 0.894* 0.469 4.488*** 0.677
17 4.562* 2.546 -4.153*** 0.724 0.399 0.587 2.487*** 0.639
18 7.266*** 2.185 -4.991*** 0.958 4.297*** 1.006 6.757*** 1.034
19 3.440* 2.056 -6.187*** 1.178 3.763*** 0.919 5.896*** 0.949
20 0.347 3.188 -4.282*** 1.101 2.638' 1.462 3.112*** 1.110
21 0.720 1.519 -2.517*** 0.697 0.803 0.496 1.185*** 0.264
22 0.006 0.068 -1.824 2.259 0.229 0.297 1.244*** 0.403
23 0.115 0.473 -1.107 1.797 0.937 2.019 0.285 0.298
24 0.019 0.070 -1.860 2.167 0.776 1.490 0.970 0.615
MEAN 1.104 1.272 -4.992 1.450 1.909 1.108 3.208 0.881
STD 2.034 0.947 3.259 0.789 1.752 0.731 2.478 0.637
MAX 7.266 3.188 -1.107 2.862 7.359 2.382 9.476 3.175
MIN -0.982 0.037 -13.006 0.074 0.229 0.101 0.281 0.221

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.
218 19. Empirical Inference

Table 19.7: Inference Results for the Stochastic Volatility Model (continued)

p ae,s ae,F neg.

Hour EST STD EST STD EST STD LogL
1 -0.646*** 0.132 0.640*** 0.019 0.183*** 0.008 640.3
2 -0.432*** 0.106 0.619*** 0.020 0.173*** 0.010 640.4
3 -0.697*** 0.082 0.554*** 0.018 0.245*** 0.012 701.3
4 -0.757 0.580 0.628*** 0.020 0.285*** 0.015 846.1
5 -0.467 0.795 0.668*** 0.021 0.261*** 0.011 859.0
6 -0.952*** 0.034 0.608*** 0.023 0.310*** 0.014 852.4
7 -0.333** 0.157 0.561*** 0.018 0.360*** 0.014 901.2
8 -0.596** 0.258 0.625*** 0.020 0.336*** 0.012 908.9
!) -0.657*** 0.118 0.639*** 0.018 0.275*** 0.011 903.6
10 -0.606 0.963 0.553*** 0.019 0.238*** 0.010 697.7
11 -0.032 0.149 0.507*** 0.015 0.191*** 0.009 580.5
12 0.868*** 0.065 0.506*** 0.015 0.178*** 0.009 532.0
13 0.654*** 0.091 0.452*** 0.014 0.201 *** 0.010 539.3
14 0.654*** 0.114 0.554*** 0.017 0.164*** 0.011 674.0
15 0.111 0.235 0.558*** 0.017 0.179*** 0.011 761.3
16 0.802*** 0.070 0.594*** 0.018 0.205*** 0.011 802.6
1. r 0.146 0.447 0.580*** 0.018 0.188*** 0.011 746.2
18 0.854*** 0.059 0.616*** 0.018 0.148*** 0.010 703.3
19 0.721 *** 0.099 0.664*** 0.020 0.179*** 0.011 785.0
201 0.023 0.131 0.582*** 0.019 0.195*** 0.009 677.3
21 -0.033 0.056 0.531*** 0.016 0.176*** 0.008 538.9
22 -0.855** 0.427 0.535*** 0.016 0.122*** 0.007 345.7
23 -0.429 0.405 0.619*** 0.018 0.090*** 0.006 419.0
24 -0.999** 0.420 0.627*** 0.019 0.118*** 0.008 470.1
MEAN -0.152 0.250 0.584 0.018 0.208 0.010 688.6
STD 0.624 0.247 0.054 0.002 0.068 0.002 158.2
MAX 0.868 0.963 0.668 0.023 0.360 0.015 908.9
MIN -0.999 0.034 0.452 0.014 0.090 0.006 345.7

Note:
Statistically significant parameter estimates at the 1%-, 5%- and lO%-levels
are denoted by ***, **, and *, respectively.
19.2 Data Analysis and Estimation Results 219

mean. At statistically significant levels the volatility parameter takes values

of 216 to 463 percent p.a. which underlines the characteristic of extreme
volatility patterns in electricity prices. All volatility point estimates are
statistically significant at the one percent level, except for hour 11. Fur-
ther, the market price of risk parameter .A is estimated at a mean value
of approximately -1.0 across all 24 time-series. These negative inferred pa-
rameter values for the market price of risk can be seen as a compensation
of the market participants for the high positive skewness of the spot prices
when forward contracts are to be traded on such an underlying. Finally, the
measurement errors give evidence that the spot electricity prices on average
are only captured by a three times higher standard deviation of 0.60 than
the 0.20 estimated for the forward prices. Overall, the information we get
from calibrating the forward pricing model of Schwartz (1997, Modell) to
the electricity data is that the spot prices follow a mean-reverting process
which is highly volatile, and the cross-section combining the spot and the
forward prices reveals negative values for the market price for unhedgeable
spot price movements.

Second, we estimate our theoretical stochastic volatility forward pricing

model derived in section 18.3 using the setup for the empirical inferences
as described in section 19.1. The results on the model parameter estimates
are shown in tables 19.6 and 19.7. In the estimation procedure we set
the market price of risk parameter .Au for the variance rate equal to zero
and use the squared point estimates of the volatility parameter a from the
empirical inferences on Schwartz (1997, Modell) as input for the long-run
mean () of the variance rate. Thus, with the parameter space of interest
given by 1/J = {IL,.A, K, a, p, a e,S, a e,F} we have one state variable and one
parameter more to calibrate our model to the spot and forward electricity
prices as compared to the single-factor model of Schwartz (1997, Model
1). The higher flexibility of the stochastic volatility model is underlined
by a lower measurement error a e,S in the spot prices averaging at 0.58
instead of 0.60 as with the single-factor model. This gives evidence that
the time-series behavior of the electricity data is captured closer when we
include the variance rate as second factor. The cross-sectional fit of the
term structure calibrated by the second measurement equation of (19.15),
however, is not improved by the two-factor model; the measurement error
a e,F stays approximately unchanged at a value of 0.21. Next, we look at
220 19. Empirical Inference

the other parameter estimates. The expected rate of return J-t of the spot
prices is estimated at levels ranging from -0.98 up to 7.2. AB expected
for a level parameter, the inferences on the mean rate are not statistically
different from zero. The values for the market price of spot price risk >.
are deduced from the data at a negative spectrum starting from -1.1 and
ending at a value of -13.0. With the stochastic volatility pricing model the
market prices of risk are mostly estimated at significant negative levels. The
inferred negative values are in line with the results from the estimation of
Schwartz (1997, Modell). The highly negative parameter values for the
market price of risk further demonstrate that it is crucial to calibrate an
equilibrium forward pricing model in that the non-storable commodity of
electricity creates unhedgeable risk. This risk need to be compensated for on
the short sellers' side of the forward contracts. The remaining parameters
K, (J, and p describe the time-series behavior of the second state variable,
i.e. the variance rate. The variance rate is estimated with high tendencies
to revert to its long-run equilibrium value; the inferences on the coefficients
of mean-reversion K translate to mean half-lives with a minimum of 1.1
months and a maximum of 36.3 months. The volatility of the variance
rate is inferred at generally highly significant values with an average of 320
percent across all 24 time series. Finally, the correlation coefficient p, which
shows whether the two state variables generally move in the same or in the
opposite direction, is estimated at a negative average value of -0.15. Only
37.5 percent or 9 out of the 24 time-series exhibit correlation coefficients
above zero.
20 Summary and Conclusions

In part IV of this study we focus on pricing short-term electricity forward

contracts. Based on the peculiarities of electricity as underlying commodity
of forward contracts we develop a theoretical valuation model. In analyz-
ing electricity spot and forward market data we further present results on
empirical inferences for the Californian power market.
In the introductory chapter we investigate the more mature markets of
commodity futures to clear the facts on the pricing of electricity forwards.
Especially the fact of non-storability of electricity is found to be important
in that it causes extreme spikes and high volatility in the electricity spot
prices and asks for pricing in an incomplete market situation. In achieving
to capture such price behavior of electricity we choose to build a stochastic
factor model to price electricity forwards as described in chapter 18. We use
the non-tradeable spot price of electricity and the variance rate of the spot
prices as underlying state variables including their market prices of risk.
Thereupon, we present a closed-form solution on valuing electricity forward
contracts using risk neutral pricing techniques.
In chapter 19 we then show how to implement the theoretical pricing
model on empirical data. There, we first clear distributional issues of the
stochastic state variables and discretize the continuous time pricing model
into a convenient state space model. Grounded on the state space represen-
tation we show how to run an extended Kalman filter algorithm in order to
calibrate our pricing model to empirical spot and forward data on electric-
ity by means of maximum likelihood inferences. For the historical sample
we use hourly spot and day-ahead forward electricity data from the largely
deregulated Californian power market available for the period of April 1998
to December 1999. In describing the data sample we find the spot and
forward prices to average around $29 per MWh and exhibit high standard
222 20. Summary and Conclusions

deviations. The distribution of the spot prices is further characterized by a

high positive skewness and an excess kurtosis. The findings for the forward
prices are similar, but some time series exhibit even higher values for the
skewness and kurtosis than they do in the spot prices. Both the spot and the
forward time series show extreme spikes in prices and a highly volatile price
behavior. Further regression results provide evidence that the day-ahead
forward prices are not unbiased predictors for realized spot prices. This
creates the need for theoretical valuation models to price electricity forward
contracts. We choose to calibrate a single-factor model and our stochastic
volatility pricing model to the data. In comparing the two models, we find
the higher flexibility of the stochastic volatility model resulting especially
in a better fit of the time-series behavior of the spot prices. Furthermore,
the inferences on the model parameters primarily provide evidence for the
need of choosing an equilibrium valuation model to price electricity for-
wards. This result stems from largely negative estimates for the market
prices of unhedgeable spot electricity price risk in both models. Finally, it
would have been interesting to calibrate a two-factor model that captures
the mean-reversion property in both state variables.
List of Symbols and Notation

The symbols and notation used throughout this study are introduced at
their first appearance. The choices of symbols and notation are adopted
from common uses in the literature. In the following we enumerate and
explain the most frequently used symbols and notation separately for each
part of the study.

Part I
Cov[···I···] conditional covariance operator
lBJ [ •. ·1 ... ] conditional expectation operator
et vector of measurement errors
"It vector of transition equation noise terms
Ft variance-covariance matrix of the prediction error
F( ... ) distribution function
Ft information set up to time t
gt general non-linear measurement function
ht general non-linear transition function
Kt Kalman gain matrix
L ( ... ) log-likelihood function
MMSE minimum mean square error
MMSLE minimum mean square linear error
MSE (... ) mean square error
N( .. . ) normal distribution function
p ( ... ) density function
'ljJ vector of parameters
Et variance-covariance matrix of the state variables
t variable for calendar time
Xt general stochastic process
224 List of Symbols and Notation

vector of state variables

vector of measurable observations

Part II
A,B,C factor loadings
ARt,i abnormal return of security i on date t
BMt benchmark return
CAR cumulated abnormal return
dt infinitesimal time increment
dWt infinitesimal Brownian increment
E[ .. ·I···] conditional expectation operator
HPt,i holding period return of security i on date t
/'i, speed of mean-reversion
KU kurtosis
L( ... ) log-likelihood function
MAD mean absolute deviation
JL expected mean return of the net asset value
NAVt net asset value of closed-end funds
Pt market price of closed-end funds
1ft instantaneous dynamic premium
PREMt empirical premium
1/1 vector of parameters
r riskless interest rate
RESt residual premium
p coefficient of correlation
RMSE root mean square error
(Jx volatility of the net asset value
volatility of the dynamic premium
SK skewness
SSE sum of squared errors
STD standard deviation
f) long-run mean value
VAR[···I···] conditional variance operator
Xt logarithm of the net asset value
et vector of state variables
Yt logarithm of the market price
List of Symbols and Notation 225

Part III
A,B,C factor loadings
B (0, t) money market account
Cov [... 1... ] conditional covariance operator
dt infinitesimal time increment
dWt infinitesimal Brownian increment
1& [... 1... ] conditional expectation operator
f (t, T) forward interest rate
Ft information set up to time t
k (t, T) swap rate
KU kurtosis
L (t, T) LIBOR rate
A market price of risk
P (t, T) price of a discount bond
lP empirical probability measure
r instantaneous interest rate
p coefficient of correlation
Q risk neutral probability measure
QT forward martingale measure
SK skewness
STD standard deviation
Var [... 1... ] conditional variance operator
x vector of theoretical state variables
~t vector of empirical state variables
Yt vector of measurable observations
y (t, T) spot interest rate
r7t forward price of a discount bond

Part IV
Cov[···I···] conditional covariance operator
dt infinitesimal time increment
dWt infinitesimal Brownian increment
1&[···1···] conditional expectation operator
F (t, T) forward or futures price
Ft information set up to time t
K, speed of mean-reversion
KU kurtosis
226 List of Symbols and Notation

L ( ... ) log-likelihood function

A market price of risk
J-t expected mean return
P(t, T) price of a discount bond
IP' empirical probability measure
Q risk neutral probability measure
p coefficient of correlation
St spot price
instantaneous volatility
SK skewness
STD standard deviation
() long-run mean volatility
Vt stochastic variance rate
Var[ .. . I ... J conditional variance operator
Xt logarithm of spot price
et vector of empirical state variables
Yt vector of measurable observations
List of Tables

2.1 Classification of Stochastic Processes . . . . . . . . . . 8

8.1 Characteristics of the Closed-End FUnd Sample . . . . 56
8.2 Descriptive Statistics on the Closed-End Fund Premia. 58
8.3 Parameter Estimates of the Empirical Premia Process . 59
8.4 Parameter Estimates of'l/JssM . . . . . . . . . . 65
8.5 Parameter Estimates of'l/JssM (continued) . . . . . . . 66
8.6 Values of () and Statistic of Dynamic Premia 'lrt . . • • 67
9.1 Results on the Forecasting Study in Terms of Market Prices 74
9.2 Results on Portfolio Trading Strategies . . . . . . . . . . .. 78
9.3 Benchmark Results of Trading Rules Based on Empirical Pre-
mia. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 79
12.1 Developments in Two-Factor Models . . . . . . . . . . .. 99
13.1 Effects of Increasing Parameter Values on the Yield Curve 125
15.1 Descriptive Statistics of the Term Structure Sample 155
15.2 Cross-Correlations among Interest Rates 156
15.3 Principal Component Analysis. . . . . . 158
15.4 Factor Loadings of the Common Factors 158
15.5 Estimated Parameter Values . . . . . . . 162
15.6 Statistics on Historical Parameter Estimates 164
15.7 Statistics on Realized State Variable Values 168
15.8 Descriptive Statistics of LIBOR and Swap Data 171
15.9 Parameter Estimates for LIBOR and Swap Data. 172
17.1 Commodity Pricing Models . . . . . . . . . 185
18.1 Electricity Pricing Models . . . . . . . . . . . 191
19.1 Sample Statistics of Electricity Spot Prices. . 211
19.2 Sample Statistics of Electricity Forward Prices 212
19.3 Comparison of Electricty Forward and Realized Spot Prices. 213
228 List of Tables

19.4 Parameter Estimates for Schwartz (1997, Modell) . . . . . 215

19.5 Parameter Estimates for Schwartz (1997, Modell) (continued)216
19.6 Inference Results for the Stochastic Volatility Model . . . . . 217
19.7 Inference Results for the Stochastic Volatility Model (contin-
ued) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
List of Figures

4.1 Flowchart of the Kalman Filter and MLE . 23

7.1 ROC Taiwan FUnd (ROC) . . . . . 44
7.2 Indonesia Fund (IF) . . . . . . . . . 45
7.3 GT Global Eastern Europe (GTF) . 46
7.4 Time Series of the Empirical Premia 47
8.1 Time Series of (Inverted) Dynamic Premia 68
8.2 Comparison of Empirical and Dynamic Premia. 69
9.1 Setup of Forecasting Study . . . . . . . . . . . . 72
13.1 Influence of the State Variables on the Term Structure 122
13.2 Analysis of the Term Structure of Interest Rates . . . 124
13.3 Typical Patterns of the Term Structure of Volatilities 126
15.1 Estimated Zero Bond Yield Curve from 1970 to 1998 154
15.2 Cross-Correlations among Interest Rates . . . . . . 156
15.3 Factor Loadings of Principal Components. . . . . . 159
15.4 Estimated Values for the Parameters '¢ over Time. 163
15.5 Phase Plane of State Variables x as Vector Field. . 166
15.6 Historical Evolution of State Variables from 1970 to 1998. 167
17.1 Classification of Commodity Futures Underlyings . . . .. 180
18.1 Electricity Spot and Day-Ahead Forward Prices in California 192
Bibliography

Ait-Sahalia, Y. (1996): "Testing Continuous-Time Models of the Spot In-

terest Rate," Review of Financial Studies, 9(2), 385-426.

Aitchison, J., and J. A. C. Brown (1957): The Lognormal Distribution with

Special References to Its Uses in Economics. Cambridge: Cambridge Uni-
versity Press.

Amin, K., V. Ng, and S. C. Pirrong (1995): "Valuing Energy Derivatives,"

in Managing Energy Price Risk, ed. by R. Jamson, pp. 57-70. London:
Risk Publications.

Andersen, T. G., and J. Lund (1997): "Estimating Continuous-Time

Stochastic Volatility Models of the Short-Term Interest Rate," Journal
of Econometrics, 77(77), 343-377.

Anderson, B. D.O., and J. B. Moore (1979): Optimal Filtering. Englewood

Cliffs: Prentice Hall.

Anderson, S. C. (1986): "Closed-End Funds versus Market Efficiency,"

Journal of Portfolio Management, 13(1), 63-65.

Anderson, S. C., and J. A. Born (1992): Closed-End Investment Companies:

Issues and Answers. Boston: Kluwer.

Aoki, M. (1990): State Space Modeling of Time Series. Berlin: Springer, 2

edn.

Babbs, S. H., and K. B. Nowman (1999): "Kalman Filtering of Generalized

Vasicek Term Structure Models," Journal of Financial and Quantitative
Analysis, 34(1), 115-130.
232 Bibliography

Bachelier, L. (1900): "Theorie de la Speculation," in The Random Character

of Stock Market Prices (English translation, 1964), ed. by P. H. Cootner,
pp. 17-78. Cambridge: MIT Press.

Backus, D., S. Foresi, and S. Zin (1998): "Arbitrage Opportunities in

Arbitrage-Free Models on Bond Pricing," Journal of Business and Eco-
nomic Statistics, 16(1), 13-26.

Ball, C. A., and W. N. Torous (1996): "Unit Roots and the Estimation of
Interest Rate Dynamics," Journal of Empirical Finance, 3(2), 215-238.

Barber, P. (1997): "Rationalising the Transmission Business," in The US

Power Market, ed. by R Jameson. London: Risk Publications.

Beaglehole, D. R, and M. S. Tenney (1991): "General Solutions of Some

Interest Rate-Contingent Claim Pricing Equations," Journal of Fixed In-
come, pp. 69-83.

Bell, J. C., and J. R Lilyestrom (1997): "The March of the Regulator:

Regulatory Dynamics at the State and Federal Level," in The US Power
Market, ed. by R Jameson, pp. 17-32. London: Risk Publications.

Bergstrom, A. (1984): "Continuous Time Stochastic Models and Is-

sues of Aggregation over Time," in Handbook of Econometrics, ed. by
Z. Griliches, and M. Intriligator, vol. 2, pp. 1145-1212. Amsterdam: El-
sevier.

Bierwag, G., G. C. Kaufman, and A. Toevs (1983): "Innovations in Bond

Portfolio Management: Duration Analysis and Immunization," Green-
wich: JAI Press.

Bjerring, J. H., J. Lakonishok, and T. Vermaelen (1983): "Stock Prices

and Financial Analysts' Recommendations," Journal of Finance, 38(1),
187-204.

Black, F., and M. Scholes (1973): "The Pricing of Options and Corporate
Liabilities," Journal of Political Economy, 81, 637-654.

Bliss, R R (1997): "Testing Term Structure Estimation Methods," Ad-

vances in Futures and Options Research, 9, 197-231.
Bibliography 233

Bliss, R R (1999): "Bliss Term Structures," Technical report, Federal

Reserve Bank of Chicago.

Bliss, R R, and E.!. Ronn (1998): "Callable U.S. Treasury Bonds: Optimal
Calls, Anomalies, and Implied Volatilities," Journal of Business, 71(2),
211-251.

Bollerslev, T., and J. M. Wooldridge (1992): "Quasi-Maximum Likelihood

Estimation and Inference in Dynamic Models with Time-Varying Covari-
ances," Econometric Reviews, 11(2), 143-172.

Bonser-Neal, C., G. Brauer, R Neal, and S. Wheatley (1990): "International

Investment Restrictions and Closed-End Country Fund Prices," Journal
of Finance, 45(2), 523-547.

Brauer, G. A. (1984): "'Open-Ending' Closed-End Funds," Journal of Fi-

nancial Economics, 13, 491-507.

Brennan, M. J. (1991): "The Price of Convenience and the Valuation of

Commodity Contingent Claims," pp. 33-71.

Brennan, M. J., and E. S. Schwartz (1979): "A Continuous Time Approach

to the Pricing of Bonds," Journal of Banking and Finance, 3, 133-155.

Brennan, M. J., and E. S. Schwartz (1980): "Conditional Predictions of

Bond Prices and Returns," The Journal of Finance, 35(2), 405-419.

Brennan, M. J., and E. S. Schwartz (1982): "An Equilibrium Model of

Bond Pricing and a Test of Market Efficiency," Journal of Financial and
Quantitative Analysis, 17(3), 301-329.

Brennan, M. J., and E. S. Schwartz (1985): "Evaluating Natural Resource

Investments," Journal of Business, 58(2), 135-157.

Brickley, J. A., and J. S. Schallheim (1985): "Lifting the Lid on Closed-

End Investment Companies: A Case of Abnormal Returns," Journal of
Financial and Quantitative Analysis, 20(1),107-117.

Briys, E., M. Crouhy, and R Schobel (1991): "The Pricing of Default-free

Interest Rate Cap, Floor, and Collar Agreements," Journal of Finance,
46(5), 1879-1892.
234 Bibliography

Brock, W. A., and A. G. Malliaris (1989): Differential Equations, Stability,

and Chaos in Dynamic Eonomics. Amsterdam: North-Holland.

Brockwell, P. A., and R. A. Davies (1987): Time Series: Theory and Models.
New York: Springer.

Brown, R. H., and S. M. Schaefer (1994): "The Term Structure of Real

Interest Rates and the Cox, Ingersoll, and Ross Model," Journal of Fi-
nancial Economics, 35, 3-42.

Brown, S. J., and P. H. Dybvig (1986): "The Empirical Implications of the

Cox, Ingersoll, Ross Theory of the Term Structure of Interest Rates,"
Journal of Finance, 41(3), 616-628.

Broze, L., O. Scaillet, and J.-M. Zakoian (1995): "Testing for Continuous-
Time Models of the Short-Term Interest Rate," Journal of Empirical
Finance, 2, 199-223.

BUhler, W., O. Korn, and R. Schobel (2000): "Pricing and Hedging of Oil
Futures - A Unifying Approach," Discussion paper, University of Tubin-
gen.

Burridge, P., and K. F. Wallis (1988): "Prediction Theory for Autoregressive

Moving Average Processes," Econometric Reviews, 7(1), 65-95.

CalPX (1998): Great Expectations: What Happens When the Markets

Open? Pasadena: California Power Exchange.

CalPX (1999): CalPX Primer. Pasadena: California Power Exchange.

CalPX (2000a): 1998-1999 Market Year Report to Californians. Pasadena:

California Power Exchange.

CalPX (2000b): Providing the Efficient, Competitive Marketplace for Elec-

tricity Trading in California. Pasadena: California Power Exchange.

Campbell, J. Y., A. W. Lo, and A. C. MacKinlay (1997): The Econometrics

of Financial Markets. Princeton: Princeton University Press.

Campbell, J. Y., and R. J. Shiller (1987): "Cointegration and Tests of

Present Value Models," Journal of Political Economy, 95, 1062-1088.
Bibliography 235

Chan, K. C., A. Karolyi, F. A. Longstaff, and A. B. Sanders (1992): "An

Empirical Comparison of Alternative Models of the Short-Term Interest
Rate," Journal of Finance, 47(3), 1209-1225.

Chapman, D. A., J. B. Long, and N. D. Pearson (1999): "Using Proxies for

the Short Rate: When are Months like an Instant?," Review of Financial
Studies, 12(4), 763-806.

Chen, L. (1996a): Interest Rate Dynamics, Derivatives Pricing, and Risk

Management. Berlin: Springer.

Chen, N.-F., R Kan, and M. H. Miller (1993): "Are the Discounts on

Closed-End Funds a Sentiment Index?," Journal of Finance, 48(2), 795-
800.

Chen, R-R (1995): "A Two-Factor, Preference-Free Model for Interest

Rate Sensitive Claims," Journal of Futures Markets, 15(3), 345-372.

Chen, R-R (1996b): Understanding and Managing Interest Rate Risk. Sin-
gapore: World Scientific.

Chen, R-R, and L. Scott (1992): "Pricing Interest Rate Options in a Two-
Factor Cox-Ingersoll-Ross Model of the Term Structure," Review of Fi-
nancial Studies, 5(4), 613-636.

Chen, R-R, and L. Scott (1993): "Maximum Likelihood Estimation for a

Multifactor Equilibrium Model of the Term Structure of Interest Rates,"
Journal of Fixed Income, 3(3), 14-31.

Chen, R-R, and L. Scott (1995): "Multi-Factor Cox-Ingersoll-Ross Models

of the Term Structure: Estimates and Tests from a Kalman Filter Model,"
Working paper, University of Georgia.

Chordia, T., and B. Swaminathan (1996): "Market Segmentation, Imperfect

Information, and Closed-End Fund Discounts," Working paper, Vander-
bilt University.

Clewlow, L., and C. Strickland (1999): "Valuing Energy Options in a One

Factor Model Fitted to Forward Prices," Working paper, Warwick Busi-
ness School.
236 Bibliography

Cortazar, G., and E. S. Schwartz (1994): "The Valuation of Commodity-

Contingent Claims," Journal of Derivatives, 1(4), 27-39.

Cox, J. C., J. E. Ingersoll, and S. A. Ross (1981): "The Relation Between

Forward Prices and Futures Prices," Journal of Financial Economics, 9,
321-346.

Cox, J. C., J. E. Ingersoll, and S. A. Ross (1985a): "An Intertemporal

General Equilibrium Model of Asset Prices," Econometrica, 53(2), 363-
384.

Cox, J. C., J. E. Ingersoll, and S. A. Ross (1985b): "A Theory of the Term
Structure of Interest Rates," Econometrica, 53(2), 385-407.

Crouhy, M., D. Galai, and R. Mark (2000): "A Comparative Analysis of

Current Credit rusk Models," Journal of Banking and Finance, 24, 59-
117.

Dai, Q., and K. J. Singleton (2000): "Specification Analysis of Affine Term

Structure Models," Journal of Finance, 55(5), 1943-1978.

Das, S. (1994): Swap and Derivative Financing. New York: McGraw-Hill,

revedn.

De Jong, F. (1997): "Time-series and Cross-section Information in Affine

Term Structure Models," Discussion paper, Tilburg University.

De Jong, F., and P. Santa-Clara (1999): "The Dynamics of the Forward

Interest Rate Curve: A Formulation with State Variables," Journal of
Financial and Quantitative Analysis, 34(1), 131-157.

De Long, J. B., A. Shleifer, L. H. Summers, and R. J. Waldmann (1990):

"Noise Trader rusk in Financial Markets," Journal of Political Economy,
98(4), 703--738.

Deaton, A., and G. Laroque (1992): "On the Behavior of Commodity

Prices," Review of Economic Studies, 59, 1-23.

Deaton, A., and G. Laroque (1996): "Competitive Storage and Commodity

Price Dynamics," Journal of Political Economy, 104(5), 896-923.
Bibliography 237

DeMunnik, J. F., and P. C. Schotman (1994): "Cross-Sectional versus Time

Series Estimation of Term Structure Models: Empirical Results for the
Dutch Bond Market," Journal of Banking and Finance, 18,997-1025.

Dennis, J. E., and R. B. Schnabel (1996): Numerical Methods for Uncon-

strained Optimization and Nonlinear Equations. Philadelphia: SIAM.

Dimson, E., and C. Mini~Kozerski (1998): "Closed-End FUnds: A Sur-

vey," Working paper, London Business School, forthcoming in: Financial
Markets, Institutions & Instruments.

Diwan, I., V. R. Errunza, and L. W. Senbet (1995): ''The Pricing of Country

Funds from Emerging Markets: Theory and Evidence," Working paper,
University of Maryland.

Dothan, M. U. (1990): Prices in Financial Markets. New York: Oxford

University Press.

Duan, J.-C. (1994): "Maximum Likelihood Estimation Using Price Data of

the Derivative Contract," Mathematical Finance, 4(2), 155-167.

Duan, J.-C., and J.-G. Simonato (1995): "Estimating and Testing

Exponential-Affine Term Structure Models by Kalman Filter," Working
paper, Cirano Montreal.

Duffie, D. (1989): Jilutures Markets. Englewood Cliffs: Prentice HalL

Duffie, D. (1996): Dynamic Asset Pricing Theory. Princeton: Princeton

University Press, 2 edn.

Duffie, D., and S. Gray (1995): "Volatility in Energy Prices," in Managing

Energy Price Risk, ed. by R. Jamson, pp. 39-56. London: Risk Publica-
tions.

Duffie, D., and R. Kan (1996): "A Yield-Factor Model of Interest Rates,"
Mathematical Finance, 6(4), 379-406.

Duffie, D., and K. J. Singleton (1997): "An Econometric Model of the Term
Structure ofInterest-Rate Swap Yields," Journal of Finance, 52(4), 1287-
1321.
238 Bibliography

Engle, R. F., and C. W. J. Granger (1987): "Co-Integration and Error

Correction: Representation, Estimation, and Testing," Econometrica, 55,
251-276.

Eydeland, A., and H. Geman (1998): "Pricing Power Derivatives," Risk, 10,
71-73.

Fama, E. F. (1970): "Efficient Capital Markets: A Review of Theory and

Empirical Work," Journal of Finance, 25, 383-417.

Fama, E. F. (1991): "Efficient Capital Markets: II," Journal of Finance,

46(5), 1575-1617.

Farkas, J. (1902): "Theorie der einfachen Ungleichungen," Journal fUr die

Reine und Angewandte Mathematik, (124), 1-27.

Feynman, R. P. (1948): "Space-Time Approach to Non-Relativistic Quan-

tum Mechanics," Reviews of Modern Physics, 20(2), 367-387.

Gallant, A. R., and H. White (1988): A Unified Theory of Estimation and

Inference for Nonlinear Dynamic Models. Oxford: Blackwell.

Geman, H., N. E. Karoui, and J.-C. Rochet (1995): "Changes of Numeraire,

Changes of Probability Measure, and Option Pricing," Journal of Applied
Probability Trust, 32, 443-458.

Gemmill, G.,. and D. Thomas (2000): "Sentiment, Expenses, and Arbitrage

in Explaining the Discount on Closed-End Funds," Working paper, Lon-
don City University.

Geyer, A. L., and S. Pichler (1997): "A State-Space Approach to Estimate

and Test Multi-Factor Cox-Ingersoll-Ross Models of the Term Structure,"
Working paper, University of Vienna.

Gibbons, M. R., and K. Ramaswamy (1993): "A Test of the Cox, Ingersoll,
and Ross Model of the Term Structure," Review of Financial Studies,
6(3), 619-658.

Gibson, R., and E. S. Schwartz (1990): "Stochastic Convenience Yield and

the Pricing of Oil Contingent Claims," Journal of Finance, 45(3), 959-
976.
Bibliography 239

Girsanov, I. V. (1960): "On Transforming a Certain Class of Stochastic

Processes by Absolutely Continuous Substitution of Measures," Theory
of Probability and its Applications, 5(3), 285-30l.

Golub, G. H., and C. F. van Loan (1996): Matrix Computations. Baltimore:

Johns Hopkins Press, 3 edn.

Gourieroux, C., and A. Monfort (1995): Statistics and Econometric Models,

vol. 2. Cambridge: Cambridge University Press.

Gourieroux, C., and A. Monfort (1997): Time Series and Dynamic Models.
Cambridge: Cambridge University Press.

Gourieroux, C., A. Monfort, E. Renault, and A. Trognon (1984): "Pseudo

Maximum Likelihood Methods: Theory," Econometrica, 52(3), 681-72l.

Grinblatt, M., and F. A. Schwartz (2000): "Financial Innovation and the

Role of Derivative Securities: An Empirical Analysis of the Treasury
STRIPS Program," Journal of Finance, 55(3), 1415-1436.

Hamilton, J. D. (1994a): "State-Space Models," in Handbook of Economet-

rics, ed. by R. Engle, and D. McFadden, vol. 4, pp. 3041-3068. Amster-
dam: North-Holland.

Hamilton, J. D. (1994b): Time Series Analysis. Princeton: University Press.

Hansen, P. R. (1998): "Structural Breaks in the Term Structure of Interest
Rates," Working paper, University of California at San Diego.

Harrison, J. M., and S. R. Pliska (1981): "Martingales and Stochastic In-

tegrals in the Theory of Continuous Trading," Stochastic Processes and
their Applications, 11, 215-260.

Harvey, A. C. (1989): Forecasting, Structural Time Series Models, and the

Kalman Filter. Cambridge: Cambridge University Press.

Harvey, A. C. (1993): Time Series Models. New York: Harvester Wheat-

sheaf, 2 edn.

Heath, D., R. Jarrow, and A. Morton (1992): "Bond Pricing and the Term
Structure of Interest Rates: A New Methodology for Contingent Claims
Valuation," Econometrica, 60(1), 77-105.
240 Bibliography

Heston, S. 1. (1992): "Testing Continuous Time Models of the Term Struc-

ture of Interest Rates," Working paper, Yale University.

Heston, S. L. (1993): "A Closed-Form Solution for Options with Stochastic

Volatility with Applications to Bond and Currency Options," Review of
Financial Studies, 6(2), 327-343.

Hilliard, J. E., and J. Reis (1998): "Valuation of Commodity Futures and

Options under Stochastic Convenience Yields, Interest Rates, and Jump
Diffusions in the Spot," Journal of Financial and Quantitative Analysis,
33(1),61--86.

Hirsch, M. W., and S. Smale (1974): Differential Equations, Dynamical

Systems, and Linear Algebra. San Diego: Academic Press.

Ho, T. S., and S.-B. Lee (1986): "Term Structure Movements and Pricing
Interest Rate Contingent Claims," Journal of Finance, 41(5), 1011-1029.

Ho, T. S. Y. (1992): "Key Rate Durations: Measures ofInterest Rate Risk,"

Journal of Fixed Income, 2(2), 29-44.

Hull, J., and A. White (1987): "The Pricing of Options on Assets with
Stochastic Volatilities," The Journal of Finance, 42(2), 281-300.

Hull, J. C., and A. White (1990): "Pricing Interest-Rate-Derivative Securi-

ties," Review of Financial Studies, 3(4), 573-592.

Ingersoll, J. E. (1976): "A Theoretical and Empirical Investigation of Dual

Purpose FUnds," Journal of Financial Economics, 3, 83-123.

Ingersoll, J. E. (1987): Theory of Financial Decision Making. Savage: Row-

man & Littlefield.

Jamshidian, F. (1989): "An Exact Bond Option Formula," Journal of Fi-

nance, 44(1), 205-209.

Jarrow, R. A., and G. S. Oldfield (1981): "Forward Contracts and Futures

Contracts," Journal of Financial Economics, 9, 373-382.

Jazwinski, A. H. (1970): Stochastic Processes and Filtering Theory. New

York: Academic Press.
Bibliography 241

Johnson, B., and A. Sogomonian (1997): "Electricity Futures," in The US

Power Market, ed. by R. Jameson, pp. 83-98. London: Risk Publications.

Jolliffe, I. T. (1986): Principal Component Analysis. New York: Springer.

Kac, M. (1949): "On the Distribution of Certain Wiener Functionals,"

Transactions of the American Mathematical Society, 65, 1-13.

Kaldor, N. (1939): "Speculation and Economic Stability," Review of Eco-

nomic Studies, 7, 1-27.

Kalman, R. E. (1960): "A New Approach to Linear Filtering and Prediction

Problems," Journal of Basic Engineering, 82, 35-45.

Kalman, R. E. (1963): "New Methods in Wiener Filtering Theory," in Pro-

ceedings of the First Symposium of Engineering Applications of Random
Function Theory and Probability, ed. by J. L. Bogdanoff, and F. Kozin,
pp. 270-388. New York: Whiley.

Kalman, R. E., and R. S. Bucy (1961): "New Results in Linear Filtering

and Prediction Theory," Journal of Basic Engineering, 83, 95-108.

Karatzas, I., and S. E. Shreve (1991): Brownian Motion and Stochastic

Calculus. Berlin: Springer, 2 edn.

Karlin, S., and H. M. Taylor (1975): A First Course in Stochastic Processes.

Boston: Academic Press, 2 edn.

Kellerhals, B. P., and M. Uhrig-Homburg (1998): "Ungleichgewichte auf

Bondmarkten: Aktive Handelsstrategien auf Basis geschatzter Zinsstruk-
turkurven," Finanzmarkt und Portfolio Management, 12(1), 32-45.

Klafiky, T. E., Y. Y. Ma, and A. Nozari (1992): "Managing Yield Curve

Exposure: Introducing Reshaping Durations," Journal of Fixed Income,
2(3), 39-45.

Lee, C. M., A. Shleifer, and R. H. Thaler (1990): "Anomalies: Closed-End

Mutual Funds," Journal of Economic Perspectives, 4(4), 153-164.

Lee, C. M., A. Shleifer, and R. H. Thaler (1991): "Investor Sentiment and

the Closed-End Fund Puzzle," Journal of Finance, 46(1), 75-109.
242 Bibliography

Leong, K. S. (1997): "The Forward Curve in the Electricity Market," in

The US Power Market, ed. by R Jameson, pp. 133-148. London: Risk
Publications.

Litterman, R, and J. Scheinkman (1991): "Common Factors Affecting Bond

Returns," Journal of Fixed Income, pp. 54-61.

Litterman, R, J. Scheinkman, and 1. Weiss (1991): "Volatility and The

Yield Curve," Journal of Fixed Income, pp. 49-53.

Longstaff, F. A., and E. S. Schwartz (1992): "Interest Rate Volatility and

the Term Structure: A Two-Factor General Equilibrium Model," Journal
of Finance, 47(4), 1259-1282.

Love, D. A., and S. F. McShea (1997): "The U.S. Treasury STRlPS: Valua-
tion, Liability, and Asset Allocation Frontier," in Yield Curve Dynamics,
ed. by R J. Ryan, pp. 9-32. Chicago: Glenlake.

Lund, J. (1997): "Five Essays in Financial Econometrics and the Term

Structure ofInterest Rates," Phd dissertation, Aarhus School of Business.

Liitkepohl, H. (1996): Handbook of Matrices. Chichester: John Whiley &

Sons.

Macaulay, F. R (1938): Some Theoretical Problems Suggested by the Move-

ments of Interest Rates, Bond Yields, and Stock Prices in the U.S. Since
1856. New York: National Bureau of Economic Research.

Malkiel, B. G. (1977): "The Valuation of Closed-End Investment-Company

Shares," Journal of Finance, 32(3), 847-859.

Markham, J. W. (1987): The History of Commodity Futures Trading and

Its Regulation. New York: Praeger.

Marshall, J. F., and K. R Kapner (1993): The Swaps Market. Miami: Kolb,
2 edn.

McCulloch, J. H. (1975): "The Tax-Adjusted Yield Curve," Journal of

Finance, 30(3), 811-830.
Bibliography 243

McCulloch, J. H., and H.-C. Kwon (1993): "U.S. Term Structure Data.
1947-1991," Working paper, Ohio State University.

Merton, R C. (1973): "Theory of Rational Option Pricing," Bell Journal

of Economics and Management Science, 4, 141-183.

Merton, R C. (1976): "Option Pricing when Underlying Stock Prices are

Discontinuous," Journal of Financial Economics, 3, 125-144.

Merton, R C. (1992): Continuous-Time Finance. Cambridge: Blackwell,

rev. edn.

Miller, R L., and D. D. VanHoose (1993): Modern Money and Banking.

New York: McGraw-Hill, 3 edn.

Miltersen, K. R, K. Sandmann, and D. Sondermann (1997): "Closed

Form Solutions for Term Structure Derivatives with Log-Normal Interest
Rates," Journal of Finance, 52(1), 409-430.

Minio-Paluello, C. (1998): "The UK Closed-End Fund Discount," Phd the-

sis, London Business School.

Moler, C., and C. van Loan (1978): "Nineteen Dubious Ways to Compute
the Exponential of a Matrix," SIAM Review, 20(4), 801-836.

Morton, A. J. (1988): "Arbitrage and Martingales," Technical report, Cor-

nell University.

Musiela, M., and M. Rutkowski (1997): Martingale Methods in Financial

Modelling. Berlin: Springer.

Nunes, J. P. V., and L. Clewlow (1999): "Kalman Filtering of Gaussian

Exponential-Affine LIB OR-Rate Models Using Caps and Swaptions,"
Working paper, University of Warwick.

Oksendal, B. (1995): Stochastic Differential Equations. Berlin: Springer, 4

edn.

Pearson, N. D., and T.-S. Sun (1994): "Exploiting Conditional Density in

Estimating the Term Structure: An Application to the Cox, Ingersoll,
and Ross Model," Journal of Finance, 49(4), 1279-1304.
244 Bibliography

Peavy, J. W. (1990): "Returns on Initial Public Offerings of Closed-End

Funds," Review of Financial Studies, 3(4), 695-708.

Pennacchi, G. G. (1991): "Identifying the Dynamics of Real Interest Rates

and Inflation: Evidence Using Survey Data," The Review of Financial
Studies, 4(1), 53-86.

Phillips, P. C. B., and S. Ouliaris (1988): "Testing for Cointegration Us-

ing Principal Component Methods," Journal of Economic Dynamics and
Control, 12, 205-230.

Pilipovic, D. (1998): Energy Risk - Valuing and Managing Energy Risk.

New York: McGraw-Hill.

Pirrong, C., and M. Jermakyan (1999): "The Price of Power: The Valua-
tion of Power and Weather Derivatives," Working paper, Olin School of
Business, Washington University.

Pokalsky, J., and J. Robinson (1997): "Integrating Physical and Financial

OTC Contract Portfolios," in The US Power Market, ed. by R. Jameson,
pp. 111-132. London: llisk Publications.

Pontiff, J. (1995): "Closend-End Fund Premia and Returns Implications

for Financial Market Equilibrium," Journal of Financial Economics, 37,
341-370.

Rebonato, R. (1998): Interest-Rate Option Models. Chichester: John

Whiley & Sons, 2 edn.

llichard, S. F., and M. Sundaresan (1981): "A Continuous Time Equilibrium

Model of Forward Prices and Futures Prices in a Multigood Economy,"
Journal of Financial Economics, 9, 347-371.

Roenfeldt, R. L., and D. L. Thttle (1973): "An Examination of the Dis-

counts and Premiums of Closed-End Investment Companies," Journal of
Business Research, 2(1), 129--140.

Rogers, L. (1995): "Which Model for the Term Structure of Interest

Rates Should One Use?," in Mathematical Finance, ed. by M. H. Davis,
D. Duffie, W. H. Fleming, and S. E. Shreve, pp. 93-115. Berlin: Springer
Verlag.
Bibliography 245

Ross, S. A. (1997): "Hedging Long-Run Commitments: Exercises in In-

complete Market Pricing," Economic Notes by Banca Monte die Paschi
di Siena SpA, 26(2), 385-420.

Rothenberg, T. J. (1971): "Identification in Parametric Models," Econo-

metrica, 39(3), 577-59l.

Routledge, B. R, D. J. Seppi, and C. S. Spatt (2000): "Equilibrium Forward

Curves of Commodities," Journal of Finance, 55(3), 1297-1338.

Schaefer, S. M., and E. S. Schwartz (1984): "A Two-Factor Model of the

Term Structure: An Approximate Analytical Solution," Journal of Fi-
nancial and Quantitative Analysis, 19(4),413-424.

SchObel, R (1992): "Arbitrage Pricing in Commodity Futures Markets: Is

It Feasible?," Presentation, Conference on Mathematical Finance, Ober-
wolfach, Germany, 23.08.-29.08.1992.

SchObel, R (1995a): "Inflation and Valuation: A Gaussian Model of the

Term Structure of Interest Rates," Working paper, Universitat Ttibingen.

SchObel, R (1995b): Kapitalmarkt und zeitkontinuierliche Bewertung.

Berlin: Springer.

SchObel, R, and J. Zhu (1999): "Stochastic Volatility with an Ornstein-

Uhlenbeck Process: An Extension," European Finance Review, 3, 23-46.

Schwartz, E., and J. E. Smith (1997): "Short-Term Variations and Long-

Term Dynamics in Commodity Prices," Working paper, University of
California.

Schwartz, E. S. (1997): "The Stochastic Behaviour of Commodity Prices:

Implications for Valuation and Hedging," Journal of Finance, 52(3), 923-
973.

Schweppe, F. C. (1965): "Evaluation of Likelihood Functions for Gaussian

Signals," IEEE Transactions on Information Theory, 11,61-70.

Scott, L. O. (1997): "Pricing Stock Options in a Jump-Diffusion Model with

Stochastic Volatility and Interest Rates: Applications of Fourier Inversion
Methods," Mathematical Finance, 7(4), 413-426.
246 Bi bJiography

Sharpe, W. F., and H. B. Sosin (1975): "Closed-End Investment Companies

in the United States," in European Finance Association: 1974 Proceed-
ings, ed. by B. Jacquillat, pp. 37-63. Amsterdam: North-Holland.

Stein, E. M., and J. C. Stein (1991): "Stock Price Distributions with

Stochastic Volatility: An Analytic Approach," Review of Financial Stud-
ies, 4(4), 727-752.

Stoll, H. R (1969): "The Relationship Between Put and Call Option

Prices," Journal of Finance, 14(5), 801-824.

Stoll, H. R, and R E. Whaley (1993): Futures and Options - Theory and

Applications. Cincinnati: South-Western Publishing.

Sundaresan, S. M. (2000): "Continuous-Time Methods in Finance: A Re-

view and an Assessment," Journal of Finance, 55(4), 1569-1622.

Swaminathan, B. (1996): "Time-Varying Expected Small Firm Returns and

Closed-End Fund Discounts," Review of Financial Studies, 9(3), 845-887.

Tanizaki, H. (1996): Nonlinear Filters. Berlin: Springer, 2 edn.

Theil, H. (1971): Principles of Econometrics. Amsterdam: North-Holland.

Thompson, R (1978): "The Information Content of Discounts and Premi-

ums on Closed-End Fund Shares," Journal of Financial Economics, 6,
151-186.

Titman, S., and W. Torous (1989): "Valuing Commercial Mortgages: An

Empirical Investigation of the Contingent-Claims Approach to Pricing
Risky-Debt," Journal of Finance, 44(2), 345-373.

Uhrig, M. (1996): Bewertung von Zinsoptionen bei stochastischer

Zinsvolatilitiit. Wiesbaden: Gabler.

Uhrig, M., and U. Walter (1996): "A New Numerical Approach for Fitting
the Initial Yield Curve," Journal of Fixed Income, (3), 82-90.

Uhrig-Homburg, M. (1999): "Die Bedeutung der Mean-Reversion von Zin-

sprozessen fur Optionswerte: Das Beispiel der Korridor-Zinsoption," OR
Spektrum, 21, 183-203.
Bibliography 247

UN (1996): "Standard International Trade Classification (Revision 3)," Sta-

tistical Papers 34, New York: United Nations Publications.

Vasicek, O. A. (1977): "An Equilibrium Characterization of the Term Struc-

ture," Journal of Financial Economics, 5(2), 177-188.

Waldman, M. (1992): "Beyond Duration: Risk Dimensions of Mortgage

Securities," Journal of Fixed Income, pp. 5-15.

Weiss, K. (1989): "The Post-Offering Price Performance of Closed-End

FUnds," Financial Management, 18,57-67.

White, H. (1982): "Maximum Likelihood Estimation of Misspecified Mod-

els," Econometrica, 50(1), 1-25.

White, H. (1994): Estimation, Inference, and Specification Analysis. Cam-

bridge: Cambridge University Press.

Wilmott, P., J. Dewynne, and S. Howison (1993): Option Pricing: Mathe-

matical Models and Computation. Oxford: Oxford Financial Press.

Woodley, J. A. C., and H. 1. Hunt (1997): "Reinvesting the Electricity

Industry: Implications of an Electricity Commodity Market," in The US
Power Market, ed. by R. Jameson, pp. 43-55. London: Risk Publications.

Working, H. (1949): "The Theory of the Price of Storage," American Eco-

nomic Review, 39, 1254-1262.

Wu, X. (1996): "A New Stochastic Duration Measure by the Vasicek and
CIR Term Structure Theories," Working paper, City University of Hong
Kong.
Vol. 405: S. Koml6si, T. Rapcsak, S. Schaible (Eds.), Vol. 431: A. Aulin, Causal and Stochastic Elements in Busi-
Generalized Convexity. Proceedings, 1992. VIII, 404 pages. ness Cycles. XI, 116 pages. 1996.
1994. Vol. 432: M. Tamiz (Ed.), Multi-Objective Programming
Vol. 406: N. M. Hung, N. V. Quyen, Dynamic Timing and Goal Programming. VI, 359 pages. 1996.
Decisions Under Uncertainty. X, 194 pages. 1994. Vol. 433: J. Menon, Exchange Rates and Prices. XIV, 313
Vol. 407: M. Ooms, Empirical Vector Autoregressive pages. 1996.
Modeling. XIII, 380 pages. 1994. Vol. 434: M. W. J. Blok, Dynamic Models ofthe Firm. VII,
Vol. 408: K. Haase, Lotsizing and Scheduling for Production 193 pages. 1996.
Planning. vm, 118 pages. 1994. Vol. 435: L. Chen, Interest Rate Dynamics, Derivatives
Vol. 409: A. Sprecher, Resource-Constrained Project Pricing, and Risk Management. XII, 149 pages. 1996.
Scheduling. XII, 142 pages. 1994. Vol. 436: M. Klemisch-Ahlert, Bargaining in Economic and
Vol. 4\0: R. Winkelmann, Count Data Models. XI, 213 Ethical Environments. IX, 155 pages. 1996.
pages. 1994. Vol. 437: C. Jordan, Batching and Scheduling. IX, 178 pages.
Vol. 411: S. Dauz~re-Pl!res, J.-B. Lasserre, An Integrated 1996.
Approach in Production Planning and Scheduling. XVI, 137 Vol. 438: A. Villar, General Equilibrium with Increasing
pages. 1994. Returns. XIII, 164 pages. 1996.
Vol. 412: B. Kuon, Two-Person Bargaining Experiments Vol. 439: M. Zenner, Learning to Become Rational. VII,
with Incomplete Information. IX, 293 pages. 1994. 201 pages. 1996.
Vol. 4\3: R. Fiorito (Ed.), Inventory, Business Cycles and Vol. 440: W. Ryll, Litigation and Settlement in a Game with
Monetary Transmission. VI, 287 pages. 1994. Incomplete Information. VITI, 174 pages. J996.
Vol. 414: Y. Crama, A. Oerlemans, F. Spieksma, Production Vol. 441: H. Dawid, Adaptive Learning by Genetic
Planning in Automated Manufacturing. X, 210 pages. 1994. Algorithms. IX, 166 pages. 1996.
Vol. 415: P. C. Nicola, Imperfect General Equilibrium. XI, Vol. 442: L. Corch6n, Theories ofImperfectly Competitive
167 pages. 1994. Markets. XIII, 163 pages. 1996.
Vol. 416: H. S. J. Cesar, Control and Game Models of the Vol. 443: G. Lang, On Overlapping Generations Models with
Greenhouse Effect. XI, 225 pages. 1994. Productive Capital. X, 98 pages. 1996.
Vol. 417: B. Ran, D. E. Boyce, Dynamic Urban Transpor- Vol. 444: S. J~rgensen, G. Zaccour (Eds.), Dynamic
tation Network Models. XV, 391 pages. 1994. Competitive Analysis in Marketing. X, 285 pages. 1996.
Vol. 418: P. Bogetoft, Non-Cooperative Planning Theory. Vol. 445: A. H. Christer, S. Osaki, L. C. Thomas (Eds.),
XI, 309 pages. 1994. Stochastic Modelling in Innovative Manufactoring. X, 361
Vol. 419: T. Maruyama, W. Takahashi (Eds.), Nonlinear and pages. 1997.
Convex Analysis in Economic Theory. VIII, 306 pages. 1995. Vol. 446: G. Dhaene, Encompassing. X, 160 pages. 1997.
Vol. 420: M. Peeters, Time-To-Build. Interrelated Invest- Vol. 447: A. Artale, Rings in Auctions. X, 172 pages. 1997.
ment and Labour Demand Modelling. With Applications to
Six OECD Countries. IX, 204 pages. 1995. Vol. 448: G. Fandel, T. Gal (Eds.), Multiple Criteria Decision
Making. XII, 678 pages. 1997.
Vol. 421: C. Dang, Triangulations and Simplicial Methods.
IX, 196 pages. 1995. Vol. 449: F. Fang, M. Sanglier (Eds.), Complexity and Self-
Organization in Social and Economic Systems. IX, 317
Vol. 422: D. S. Bridges, G. B. Mehta, Representations of pages, 1997.
Preference Orderings. X, 165 pages. 1995.
Vol. 450: P. M. Pardalos, D. W. Hearn, W. W. Hager, (Eds.),
Vol. 423: K. Marti, P. Kall (Eds.), Stochastic Programming. Network Optimization. VIII, 485 pages, 1997.
Numerical Techniques and Engineering Applications. VIII,
351 pages. 1995. Vol. 451: M. Salge, Rational Bubbles. Theoretical Basis,
Economic Relevance, and Empirical Evidence with a Special
Vol. 424: G. A. Heuer, U. Leopold-Wildburger, Silverman's Emphasis on the German Stock Market.JX, 265 pages. 1997.
Game. X, 283 pages. 1995.
Vol. 452: P. Gritzmann, R. Horst, E. Sachs, R. Tichatschke
Vol. 425: J. Kohlas, P.-A. Monney, A Mathematical Theory (Eds.), Recent Advances in Optimization. VIII, 379 pages.
of Hints. XIII, 419 pages, 1995. 1997.
Vol. 426: B. Finkenstadt, Nonlinear Dynamics in Eco- Vol. 453: A. S. Tangian, J. Gruber (Eds.), Constructing
nomics.IX, 156 pages. 1995. Scalar-Valued Objective Functions. VTII, 298 pages. 1997.
Vol. 427: F. W. van Tongeren, Microsimulation Modelling Vol. 454: H.-M. Krolzig, Markov-Switching Vector Auto-
of the Corporate Firm. XVII, 275 pages. 1995. regressions. XIV, 358 pages. 1997.
Vol. 428: A. A. Powell, Ch. W. Murphy, Inside a Modem Vol. 455: R. Caballero, F. Ruiz, R. E. Steuer (Eds.), Advances
Macroeconometric Model. XVIII, 424 pages. 1995. in Multiple Objective and Goal Programming. VIII, 391
Vol. 429: R. Durier, C. Michelot, Recent Developments in pages. 1997.
Optimization. VIII, 356 pages. 1995. Vol. 456: R. Conte, R. Hegselmann, P. Terna (Eds.), Simu-
Vol. 430: J. R. Daduna,l. Branco, J. M. Pinto Paixilo (Eds.), lating Social Phenomena. VllI, 536 pages. 1997.
Computer-Aided Transit Scheduling. XIV, 374 pages. 1995.
Vol. 457: C. Hsu, Volume and the Nonlinear Dynamics of Vol. 483: M. Konigstein, Equity, Efficiency and Evolutionary
Stock Returns. VIII, 133 pages. 1998. Stability in Bargaining Games with Joint Production. XII,
Vol. 458: K. Marti, P. Kall (Eds.), Stochastic Programming 197 pages. 2000.
Methods and Technical Applications. X, 437 pages. 1998. Vol. 484: D. D. Gatti, M. Gallegati, A. Kirman, Interaction
Vol. 459: H. K. Ryu, D. J. Slottje, Measuring Trends in U.S. and Market Structure. VI, 298 pages. 2000.
Income Inequality. XI, 195 pages. 1998. Vol. 485: A. Garnaev, Search Games and Other Applications
Vol. 460: B. Fleischmann, J. A. E. E. van Nunen, M. G. of Game Theory. VIII, 145 pages. 2000.
Speranza, P. Stahly, Advances in Distribution Logistic. XI, Vol. 486: M. Neugart, Nonlinear Labor Market Dynamics.
535 pages. 1998. X, 175 pages. 2000.
Vol. 461: U. Schmidt, Axiomatic Utility Theory under Risk. Vol. 487: Y. Y. Haimes, R. E. Steuer (Eds.), Research and
XV, 201 pages. 1998. Practice in Multiple Criteria Decision Making. XVII,
Vol. 462: L. von Auer, Dynamic Preferences, Choice 553 pages. 2000.
Mechanisms, and Welfare. XII, 226 pages. 1998. Vol. 488: B. Schmolck, Ommitted Variable Tests and
Vol. 463: G. Abraham-Frois (Ed.), Non-Linear Dynamics Dynamic Specification. X, 144 pages. 2000.
and Endogenous Cycles. VI, 204 pages. 1998. Vol. 489: T. Steger, Transitional Dynamics and Economic
Vol. 464: A. Aulin, The Impact of Science on Economic Growth in Developing Countries. VIII, 151 pages. 2000.
Growth and its Cycles. IX, 204 pages. 1998. Vol. 490: S. Minner, Strategic Safety Stocks in Supply
Vol. 465: T. J. Stewart, R. C. van den Honert (Eds.), Trends Chains. XI, 214 pages. 2000.
in Multicriteria Decision Making. X, 448 pages. 1998. Vol. 491: M. Ehrgott, Multicriteria Optimization. VIII, 242
Vol. 466: A. Sadrieh, The Alternating Double Auction pages. 2000.
Market. VII, 350 pages. 1998. Vol. 492: T. Phan Huy, Constraint Propagation in Flexible
Vol. 467: H. Hennig-Schmidt, Bargaining in a Video Ex- MlIIUIfac(uring. IX, 258. pages. 2000.
periment. Determinants of Boundedly Rational Behavillr. Vo.1A93: .J. Zhu, Modular Pricing of Options. X, 170 pages.
XII, 221 pages. 1999. 2000.
Vol. 468: A. Ziegler, A Game Theory Analysis of Options. Vol. 494: D. "Franzen, Design of Master Agreements for OTC
XIV, 145 pages. 1999. Derivatives. VIII, 175 pages. 2001.
Vol. 469: M. P. Vogel, Environmental Kuznets Curves. xm, VOl. 495: I Konnov, Combined Relaxation Methods for
197 pages. 1999. Variational Inequalities. Xl, 181 pages. 200 I.
Vol. 470: M. Ammann, Pricing Derivative Credit Risk. XII, Vol. 496: P. WeiS, Unemployment in Open Economies. XII,
228 pages. 1999. 226 pages. 200 I.
Vol. 471: N. H. M. Wilson (Ed.), Computer-Aided Transit Vol. 497: J. Inkmann, Conditional Moment Estimation of
Scheduling. Xl, 444 pages. 1999. Nonlinear Equation Systems. VIII, 214 pages. 2001.
Vol. 472: J.-R. Tyran, Money Illusion and Strategic Vol. 498: M. Reutter, A Macroeconomic Model of West
Complementarity as Causes of Monetary Non-Neutrality. German Unemployment. X, 125 pages. 2001.
X, 228 pages. 1999. Vol. 499: A. Casajus, Focal Points in Framed Games. Xl,
Vol. 473: S. Helber, Performance Analysis of Flow Lines 131 pages. 200 I.
with Non-Linear Flow of Material. IX, 280 pages. 1999. Vol. 500: F. Nardini, Technical Progress and Economic
Vol. 474: U. Schwalbe, The Core of Economies with Growth. XVII, 191 pages. 2001.
Asymmetric Information. IX, 141 pages. 1999. Vol. 501: M. Fleischmann, Quantitative Models for Reverse
Vol. 475: L. Kaas, Dynamic Macroeconomics with Imperfect Logistics. XI, 181 pages. 2001.
Competition. XI, 155 pages. 1999. Vol. 502: N. Hadjisavvas, J. E. Martinez-Legaz, J.-P. Penot
Vol. 476: R. Demel, Fiscal Policy, Public Debt and the (Eds.), Generalized Convexity and Generalized Mono-
Term Structure of Interest Rates. X, 279 pages. 1999. tonicity. IX, 410 pages. 2001.
Vol. 477: M. Th~ra, R. Tichatschke (Eds.), III-posed Vol. 503: A. Kirman, J.-B. Zimmermann (Eds.), Economics
Variational Problems and Regularization Techniques. VIII, with Heterogenous,lnteracting Agents. VII, 343 pages. 200 I.
274 pages. 1999. Vol. 504: P.-Y. Moix (Ed.),The Measurement of Market Risk.
Vol. 478: S. Hartmann, Project Scheduling under Limited Xl, 272 pages. 200 I.
Resources. XII, 221 pages. 1999. Vol. 505: S. VoS, J. R. Daduna (Eds.), Computer-Aided
Vol. 479: L. v. Thadden, Money, Inflation, and Capital Scheduling of Public Transport. XI, 466 pages. 2001.
Formation. IX, 192 pages. 1999. Vol. 506: B. P. Kellerhals, Financial Pricing Models in
Vol. 480: M. Grazia Speranza, P. Stlihly (Eds.), New Trends Continuous Time and Kalman Filtering. XIV, 247 pages.
in Distribution Logistics. X, 336 pages. 1999. 2001.
Vol. 481: V. H. Nguyen, J. J. Strodiot, P. Tossings (Eds.).
Optimation. IX, 498 pages. 2000.
Vol. 482: W. B. Zhang, A Theory of International Trade.
XI, 192 pages. 2000.