Kobe University Monograph Series in Social Science Research

Tadahiro Nakajima
Shigeyuki Hamori

Energy
Trading
and Risk
Management
Commentary on Arbitrage, Risk
Measurement, and Hedging Strategy
Kobe University Monograph Series in Social
Science Research

Series Editors
Yunfang Hu, Kobe University Graduate School of Economics, Kobe, Japan
Shigeyuki Hamori, Kobe University Graduate School of Economics, Kobe, Japan

Editorial Board
Masahiro Enomoto, Kobe University RIEB, Kobe, Japan
Yoshihide Fujioka, Kobe University Graduate School of Economics, Kobe, Japan
Yuka Kaneko, Kobe University Center for Social Systems Innovation, Kobe, Japan
Kazumi Suzuki, Kobe University Graduate School of Business Administration,
Kobe, Japan
Kenji Yamamoto, Kobe University Graduate School of Law, Kobe, Japan
The Kobe University Monograph Series in Social Science Research is an exciting
interdisciplinary collection of monographs, both authored and edited, that encompass
scholarly research not only in the economics but also in law, political science, busi-
ness and management, accounting, international relations, and other sub-disciplines
within the social sciences. As a national university with a special strength in the
social sciences, Kobe University actively promotes interdisciplinary research. This
series is not limited only to research emerging from Kobe University’s faculties of
social sciences but also welcomes cross-disciplinary research that integrates studies
in the arts and sciences.
Kobe University, founded in 1902, is the second oldest national higher educa-
tion institution for commerce in Japan and is now a preeminent institution for social
science research and education in the country. Currently, the social sciences section
includes four faculties—Law, Economics, Business Administration, and Interna-
tional Cooperation Studies—and the Research Institute for Economics and Business
Administration (RIEB). There are some 230-plus researchers who belong to these
faculties and conduct joint research through the Center for Social Systems Innovation
and the Organization for Advanced and Integrated Research, Kobe University. This
book series comprises academic works by researchers in the social sciences at Kobe
University as well as their collaborators at affiliated institutions, Kobe University
alumni and their colleagues, and renowned scholars from around the world who have
worked with academic staff at Kobe University. Although traditionally the research
of Japanese scholars has been publicized mainly in the Japanese language, Kobe
University strives to promote publication and dissemination of works in English in
order to further contribute to the global academic community.
Tadahiro Nakajima · Shigeyuki Hamori

Energy Trading and Risk

Management
Commentary on Arbitrage, Risk
Measurement, and Hedging Strategy
Tadahiro Nakajima Shigeyuki Hamori
The Kansai Electric Power Company, Graduate School of Economics
Incorporated (Japan) Kobe University
Osaka, Japan Kobe, Japan

ISSN 2524-504X ISSN 2524-5058 (electronic)

Kobe University Monograph Series in Social Science Research
ISBN 978-981-19-5602-7 ISBN 978-981-19-5603-4 (eBook)
https://doi.org/10.1007/978-981-19-5603-4

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Acknowledgements

In publishing this book, we have benefited greatly from various people. First of all,
we would like to express our special thanks to Ms. Juno Kawakami for excellent
editorial guidance. Next, we would like to thank the officers and co-workers at the
Kansai Electric Power Company and the faculty members and graduate students at
Kobe University for many helpful discussions and appropriate comments. Finally,
we would like to thank our family members, Shie, Aiki, and Naoko. Without their
heartwarming supports, we could not have completed this work. This work was
supported by JSPS KAKENHI Grant Number 22K01424.

Osaka, Japan Tadahiro Nakajima

Kobe, Japan Shigeyuki Hamori

v
Contents

1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Arbitrage Trading in Energy Markets and Measuring Its Risk . . . . . . 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Data and Preliminary Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.2 Stationarity and Unit Root Test . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Cointegration Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.4 Long-Term Equilibrium Estimation . . . . . . . . . . . . . . . . . . . . . 20
2.3 Trading Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Arbitrage Between Own Spot Spread and Future
Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Statistical Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 Risk Measurement in Statistical Arbitrage . . . . . . . . . . . . . . . . . . . . . . 28
2.5.1 Value-At-Risk and Expected Shortfall . . . . . . . . . . . . . . . . . . . 29
2.5.2 Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.3 Copula Estimation and Risk Measurement . . . . . . . . . . . . . . . 41
2.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3 Fuel Market Connectedness and Fuel Portfolio Risk . . . . . . . . . . . . . . . 53
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.1 Crude Oil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.2 Natural Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.1 Connectedness Index (Diebold and Yilmaz [4]) . . . . . . . . . . . 58
3.3.2 Spectral Decomposition (Baruník and Křehlík [1]) . . . . . . . . 60
3.3.3 EGARCH Volatility Series Estimation . . . . . . . . . . . . . . . . . . 61
3.4 Analysis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

vii
viii Contents

3.4.1 Crude Oil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.4.2 Natural Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4 Hedging Strategy with Futures Contracts . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3 Optimal Hedge Ratio and Hedge Effectiveness . . . . . . . . . . . . . . . . . . 89
4.4 Multivariate GARCH Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.4.1 Diagonal VECH Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4.2 Diagonal BEKK Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.4.3 CCC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5 Analysis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5.1 HH Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5.2 NBP Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5 Market Risk of a Power Generation Business . . . . . . . . . . . . . . . . . . . . . 105
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3 Data and Preliminary Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.1 Price Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.2 Return Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.3.3 Volatility Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4 Analysis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4.1 Return Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4.2 Volatility Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4.3 Risk Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6 Alternative to Postface: Market Risk Transfer in Power
Companies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
About the Authors

Dr. Tadahiro Nakajima is Senior Researcher at the Kansai Electric Power Company,
Incorporated. He received a Ph.D. from Kobe University. He received the Highly
Commended Paper of the Studies in Economics and Finance—Literati Award
(Emerald Publishing). His main research interests are applied time series anal-
ysis, energy economics, and energy markets. He has published about 20 articles
in international peer-reviewed journals and a book from Springer.

Dr. Shigeyuki Hamori is Professor of Economics at Kobe University in Japan. He

received a Ph.D. from Duke University. He is President of the International Research
Institute for Economics and Management, Distinguished Fellow of the International
Engineering and Technology Institute (DFIETI), and Distinguished Fellow of the
Institute of Data Science and Artificial Intelligence (DFIDSAI). His main research
interests are applied time series analysis, empirical finance, data science, and interna-
tional finance. He is Co-editor of the Singapore Economic Review, Associate Editor
of the International Review of Financial Analysis, and Associated Editor of the
Eurasian Economic Review. He served as Editor of special issues of various journals
such as Frontiers in Environmental Science, Energies, Emerging Market Finance
and Trade and Journal of Risk and Financial Management. He has published about
250 articles in international peer-reviewed journals and 20 books from Springer,
Routledge, World Scientific, etc.

ix
List of Figures

Fig. 2.1 Mean and median illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Fig. 2.2 Distribution examples and their skewness . . . . . . . . . . . . . . . . . . . 10
Fig. 2.3 Distribution examples and their kurtosis . . . . . . . . . . . . . . . . . . . . 11
Fig. 2.4 Natural gas and electricity futures prices . . . . . . . . . . . . . . . . . . . . 12
Fig. 2.5 Natural gas and electricity spot prices . . . . . . . . . . . . . . . . . . . . . . 13
Fig. 2.6 Spurious regression example (series plot) . . . . . . . . . . . . . . . . . . . 15
Fig. 2.7 Spurious regression example (scatter plot) . . . . . . . . . . . . . . . . . . 15
Fig. 2.8 Trading model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Fig. 2.9 Power generation cost and long-term equilibrium
relationship in the spot market . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Fig. 2.10 Coefficients of the equilibrium equation for statistical
arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Fig. 2.11 Time-series plots of the profit from each trade . . . . . . . . . . . . . . . 28
Fig. 2.12 Time-series plots of the profit from each combination
of the two trading strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Fig. 2.13 One-day 99% VaR and the expected shortfall in a normal
distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Fig. 2.14 One-day 99% VaR and the expected shortfall in a normal
distribution with a deformed left tail . . . . . . . . . . . . . . . . . . . . . . . 31
Fig. 2.15 Procedure to measure portfolio risk using various copulas . . . . . 42
Fig. 2.16 Degree of freedom and log-likelihood in estimating the t
copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Fig. 2.17 Scatter plots of random numbers following the bivariate
Gaussian copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Fig. 2.18 Scatter plots of random numbers following the bivariate t
copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Fig. 2.19 Scatter plots of random numbers following the bivariate
Clayton copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Fig. 2.20 Scatter plots of random numbers following the bivariate
Gumbel copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

xi
xii List of Figures

Fig. 2.21 Scatter plots of random numbers following the bivariate

Frank copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Fig. 2.22 Simplified diagram of an electric power business model . . . . . . . 49
Fig. 3.1 Crude oil prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Fig. 3.2 Natural gas prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Fig. 3.3 Crude oil return series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Fig. 3.4 Spillover effects between return series (crude oil) . . . . . . . . . . . . 65
Fig. 3.5 Crude oil volatility series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Fig. 3.6 Spillover effects between volatility series (crude oil) . . . . . . . . . . 69
Fig. 3.7 Degree of freedom and log-likelihood, t copula (crude oil
portfolio) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Fig. 3.8 Natural gas return series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Fig. 3.9 Spillover effects between return series (natural gas) . . . . . . . . . . . 75
Fig. 3.10 Natural gas volatility series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Fig. 3.11 Spillover effects between volatility series (natural gas) . . . . . . . . 78
Fig. 3.12 Degree of freedom and log-likelihood in estimating the t
copula (natural gas portfolio) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Fig. 4.1 HH spot and futures prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Fig. 4.2 NBP spot and futures prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Fig. 4.3 HH spot and futures return series . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Fig. 4.4 Covariance series between futures and spot prices (HH) . . . . . . . 96
Fig. 4.5 OHR calculated using the estimated diagonal VECH
model for the HH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Fig. 4.6 OHR calculated using the estimated diagonal BEKK
model for the HH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Fig. 4.7 OHR calculated using the estimated CCC model for the HH . . . 98
Fig. 4.8 NBP spot and futures return series . . . . . . . . . . . . . . . . . . . . . . . . . 100
Fig. 4.9 Covariance series between futures and spot prices (NBP) . . . . . . 101
Fig. 4.10 OHR calculated using the estimated diagonal VECH
model for the NBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Fig. 4.11 OHR calculated using the estimated diagonal BEKK
model for the NBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Fig. 4.12 OHR calculated using the estimated CCC model
for the NBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Fig. 5.1 Commodity price series, Europe . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Fig. 5.2 Commodity return series, Europe . . . . . . . . . . . . . . . . . . . . . . . . . 111
Fig. 5.3 Commodity volatility series, Europe . . . . . . . . . . . . . . . . . . . . . . . 113
Fig. 5.4 Spillover effects between return series . . . . . . . . . . . . . . . . . . . . . 115
Fig. 5.5 Spillover effects between volatility series . . . . . . . . . . . . . . . . . . . 117
Fig. 5.6 Degree of freedom and log-likelihood in the t copula
estimates (European commodity portfolio) . . . . . . . . . . . . . . . . . . 120
Fig. 6.1 LNG price formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
List of Figures xiii

Fig. 6.2 Impact of LNG procurement with a lower threshold price

on the business balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Fig. 6.3 Power price formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Fig. 6.4 Impact of power sales with an upper threshold price
on the business balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
List of Tables

Table 2.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Table 2.2 ADF unit root tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Table 2.3 Johansen cointegration test: Futures market . . . . . . . . . . . . . . . . 19
Table 2.4 Johansen cointegration test: Spot market . . . . . . . . . . . . . . . . . . 19
Table 2.5 Power generation business specifications . . . . . . . . . . . . . . . . . . 24
Table 2.6 Long-term equilibrium estimation . . . . . . . . . . . . . . . . . . . . . . . . 26
Table 2.7 Profit from each trade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Table 2.8 Profit from each combination of the two trading strategies . . . . 28
Table 2.9 Maximum unrealized loss and total profit from statistical
arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Table 2.10 Parameters of each marginal distribution function . . . . . . . . . . . 42
Table 2.11 Risk measurement using the estimated copulas . . . . . . . . . . . . . 44
Table 3.1 Descriptive statistics (crude oil prices) . . . . . . . . . . . . . . . . . . . . 55
Table 3.2 Descriptive statistics (natural gas prices) . . . . . . . . . . . . . . . . . . 57
Table 3.3 Descriptive statistics (crude oil return series) . . . . . . . . . . . . . . . 63
Table 3.4 ADF unit-root test results (crude oil return series) . . . . . . . . . . . 64
Table 3.5 VAR model (crude oil return series) . . . . . . . . . . . . . . . . . . . . . . 65
Table 3.6 Residual variance–covariance matrix in the VAR model
(crude oil return series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Table 3.7 Spillover index and spectral analysis (crude oil return
series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Table 3.8 Estimated AR-EGARCH model (crude oil return series) . . . . . 67
Table 3.9 Descriptive statistics (crude oil volatility series) . . . . . . . . . . . . 67
Table 3.10 ADF unit-root test results (crude oil volatility series) . . . . . . . . 68
Table 3.11 VAR model (crude oil volatility series) . . . . . . . . . . . . . . . . . . . . 68
Table 3.12 Residual variance–covariance matrix in the VAR model
(crude oil volatility series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Table 3.13 Spillover index and spectral analysis (crude oil volatilities) . . . 70
Table 3.14 Parameters of each marginal distribution function (crude
oil returns) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

xv
xvi List of Tables

Table 3.15 Risk measurement using the estimated copulas (crude oil
portfolio) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Table 3.16 Risk of each crude oil market . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Table 3.17 Descriptive statistics (natural gas return series) . . . . . . . . . . . . . 73
Table 3.18 ADF unit-root test results (natural gas return series) . . . . . . . . . 74
Table 3.19 VAR model (natural gas returns) . . . . . . . . . . . . . . . . . . . . . . . . . 74
Table 3.20 Residual variance–covariance matrix in the VAR model
(natural gas returns) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Table 3.21 Spillover index and spectral analysis (natural gas returns) . . . . 75
Table 3.22 Estimated AR-EGARCH model (Natural gas return
series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Table 3.23 Descriptive statistics (natural gas volatility series) . . . . . . . . . . . 77
Table 3.24 ADF unit-root test results (natural gas volatility series) . . . . . . 77
Table 3.25 VAR model (natural gas volatility series) . . . . . . . . . . . . . . . . . . 78
Table 3.26 Residual variance–covariance matrix in the VAR model
(natural gas volatility series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Table 3.27 Spillover index and spectral analysis (natural gas
volatility series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Table 3.28 Parameters of each marginal distribution function
(natural gas return series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Table 3.29 Risk measurement using the estimated copulas (natural
gas portfolio) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Table 3.30 Risk of each natural gas market . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Table 4.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Table 4.2 Descriptive statistics (HH return series) . . . . . . . . . . . . . . . . . . . 94
Table 4.3 Estimated multivariate GARCH model (HH) . . . . . . . . . . . . . . . 95
Table 4.4 Average OHR and HE (HH) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Table 4.5 Descriptive statistics (NBP return series) . . . . . . . . . . . . . . . . . . 99
Table 4.6 Estimated multivariate GARCH model (NBP) . . . . . . . . . . . . . . 100
Table 4.7 Average OHR and HE (NBP) . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Table 5.1 Descriptive statistics (price series) . . . . . . . . . . . . . . . . . . . . . . . 108
Table 5.2 Descriptive statistics (return series) . . . . . . . . . . . . . . . . . . . . . . . 109
Table 5.3 ADF unit root test results (return series) . . . . . . . . . . . . . . . . . . . 112
Table 5.4 Estimated AR-EGARCH model . . . . . . . . . . . . . . . . . . . . . . . . . 112
Table 5.5 Descriptive statistics (volatility series) . . . . . . . . . . . . . . . . . . . . 112
Table 5.6 ADF unit root test results (volatility series) . . . . . . . . . . . . . . . . 114
Table 5.7 VAR model (return series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Table 5.8 Residual variance–covariance matrix, VAR model (return
series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Table 5.9 Spillover index and spectral analysis (return series) . . . . . . . . . 116
Table 5.10 VAR model (volatility series) . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Table 5.11 Residual variance–covariance matrix, VAR model
(volatility series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
List of Tables xvii

Table 5.12 Spillover index and spectral analysis (volatility series) . . . . . . . 118
Table 5.13 Parameters of each marginal distribution function (return
series) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Table 5.14 Risk measurement using estimated copulas (European
commodity portfolio) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Chapter 1
Preface

This book introduces empirical methods for researchers and graduate students to
analyze energy markets empirically. In addition, it is helpful for those aiming to be
energy traders. To provide beginners and practitioners with the tools necessary to
analyze energy markets appropriately, it focuses on how to use analytical methods
and interpret the results rather than a comprehensive explanation of theory. Moreover,
it begins with basic statistics and econometrics, so it is understandable for aspiring
undergraduate students. We next review the latest academic and discussion papers.
We often see that explanations are omitted or insufficient, despite the analysis of
relatively new studies, and that the established knowledge is used without mentioning
the applicable conditions. Therefore, in many cases, we cannot keep up with the logic
of the papers, and thus cannot correctly evaluate and utilize the findings. To prevent
such problems, this book provides brief explanations of the procedures adopted.
Most of the latest empirical analysis approaches have been, and will continue to be,
developed in studies of traditional financial markets. They made not only academic
contributions to clarifying the truth, but also practical contributions to decision-
making by governments and central banks, as well as the risk and fund management
of financial institutions. Energy commodities include crude oil, natural gas, coal,
and electricity. These have typical commodity characteristics. They are economic
goods with full or partial, but substantial, fungibilities. They are treated as equivalents
regardless of who produces them. Each price is determined by the function of the
market as a whole. Their spot and derivative markets are well-established and highly
liquid, and energy markets are already financialized.
Energy traders sell and/or buy non-differentiated commodities in highly liquid
markets with the same aims as traditional financial traders. Power generation compa-
nies procure various fuels in highly liquid fuel markets, convert them into electricity,
and sell the output in highly liquid electricity markets. Petroleum refiners buy crude
oil in highly liquid crude oil markets, convert it to petroleum products such as gaso-
line, heating oil, and kerosene, and then sell these in highly liquid markets. Then, to
maximize profits and minimize risk, they trade derivatives whose underlying assets

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_1
2 1 Preface

are the commodities that they buy and sell in highly liquid energy derivatives markets.
Their operations are very similar to those of financial traders who raise funds in the
market and manage financial securities on their own accounts and those of forex
traders exchanging currencies in the forex market.
Unlike traditional financial securities, each type of energy certainly has its own
unique properties (e.g., essentiality for the real economy, high storage costs, deteri-
oration during storage, and a constant perfect match between supply and demand).
However, if we consider these properties in the analysis or measurement, or when
interpreting the results, it is possible to apply approaches and techniques from
traditional financial markets to energy markets.
Just as empirical studies in traditional financial markets contribute both academ-
ically and to practical and industrial applications, studies of energy markets offer
significant contributions. Certainly, many empirical studies already contributed to a
wide range of fields. Therefore, we assume that neither lazy regulation nor loopy
business management, which relies on thoughtless intuition and shallow experience,
exist in the energy industry.
This book offers the opportunity to acquire advanced empirical approaches and
techniques, and discusses some examples of their application to actual energy
markets. Specifically, Chap. 2 presents two potential trading strategies that utilize the
long-term equilibrium relationship between natural gas prices and electricity prices
via simulation using historical data. Moreover, it measures the risk of a portfolio
consisting of natural gas futures short positions and wholesale electricity futures
long positions, which are held during the statistical arbitrage between gas and power
futures. In this context, this book introduces descriptive statistics and the augmented
Dickey-Fuller (ADF) unit root test developed using the approach proposed by Dickey
and Fuller [4], the cointegration test proposed by Johansen [7], the dynamic ordi-
nary least squares method to estimate the long-term equilibrium equation, arbitrage
simulation, and the concept of copula first published by Sklar [9], which functionally
expresses the connection of stochastic variables, value-at-risk (VaR), and expected
shortfall as risk indicators. Chapter 3 measures the connectedness in the return and
volatility for crude oil and natural gas between North American, European, and Asian
markets. Furthermore, it spectrally decomposes the connectedness indicators. These
results reveal the spillover effects between these regions and suggest the approxi-
mate potential of each fuel portfolio consisting of the three regional markets. More-
over, it estimates some types of copulas for each fuel and calculates the VaR and
expected shortfall for each fuel portfolio. In this context, this book explains the vector
autoregressive (VAR) model, vector moving average (VMA) model, the Diebold
and Yilmaz [5] approach to measuring spillover effects, the Baruník and Křehlík [1]
technique for spectral decomposition of Diebold and Yilmaz’s [5] indicators based
on the Fourier transform, and the exponential generalized autoregressive conditional
heteroscedasticity (EGARCH) model by Nelson [8] to generate each volatility series.
Chapter 4 estimates the dynamic correlation between spot returns and future returns
in the natural gas markets in the United States and the United Kingdom. It then calcu-
lates the optimal hedge ratio, which is the future position ratio to the spot position,
References 3

to minimize portfolio risk. In this context, we introduce three multivariate general-

ized autoregressive conditional heteroscedasticity (GARCH) models: the constant
conditional correlation (CCC) model by Bollerslev [2], diagonal VECH model by
Bollerslev et al. [3], and diagonal BEKK model by Engle and Kroner [6]. Chapter 5
examines and considers the relationship between the returns and volatility of crude
oil, natural gas, coal, electricity, and carbon credits in the European futures market.
It then measures the market risk of a power generation business by interpreting that
business as a portfolio consisting of long positions in fuel futures and short positions
in electricity futures. Chapter 6, as an alternative to postface, introduces two trans-
action cases that include market risk transfers. We break down the energy prices into
the energy and risk values by interpreting the energy transaction contract from the
business balance.

References

1. Baruník, J., & Křehlík, T. (2018). Measuring the frequency dynamics of financial connectedness
and systemic risk. Journal of Financial Econometrics, 16(2), 271–296.
2. Bollerslev, T. (1990). Modelling the coherence in short-run nominal exchange rates: A
multivariate generalized ARCH model. Review of Economics and Statistics, 72(3), 498–505.
3. Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1998). A capital asset pricing model with
time-varying covarianves. Journal of Political Economy, 96(1), 116–131.
4. Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time
series with a unit root. Journal of the American Statistical Association, 74(366a), 427–431.
5. Diebold, F. X., & Yilmaz, K. (2012). Better to give than to receive: Predictive directional
measurement of volatility spillovers. International Journal of forecasting, 28(1), 57–66.
6. Engle, R. F., & Kroner, K. F. (1995). Multivariate simultaneous generalized ARCH. Econometric
Theory, 11(1), 122–150.
7. Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic Dynamics
and Control, 12(2–3), 231–254.
8. Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.
Econometrica, 59(2), 347–370.
9. Sklar, A. (1959). Distribution functions of n dimensions and margins (pp. 229–231). Publications
of the Institute of Statistics at the University of Paris.
Chapter 2
Arbitrage Trading in Energy Markets
and Measuring Its Risk

2.1 Introduction

Power generation companies procure fuel, process it into electricity, and sell the
electricity unless they use fuel-free power generation methods such as hydropower,
solar power, and wind power.
Typical fuels include coal, natural gas, petroleum gas, crude and refined prod-
ucts, and uranium. Many of their spot and financial derivatives are bought and sold
through commodity exchanges. Furthermore, even unlisted commodities and deriva-
tives related to their listed commodities and securities are traded at prices linked to
their spot prices. Fuel spots are often traded on long-term fixed-price contracts to
promote investment in each upstream fuel business. However, we can interpret these
transaction prices as being linked to their related markets from a long-term perspec-
tive. On the other hand, various types of electricity in each region are traded, while
many derivatives whose underlying assets are electric power are listed on exchanges.
Like fuel, electricity is also traded as long-term fixed-price contracts; however, it is
extremely common to buy and sell electricity at market prices or their linked prices.
The various fuels power generation companies procure and the electricity they
generate are commodities that cannot be differentiated in terms besides their prices.
Moreover, these markets, including their derivatives markets, are liquid. Therefore,
both fuel and power prices can be considered efficiently formed. In other words, no
one can buy fuel continuously at a lower price than others can, regardless of how much
they investigate the fuel market. Similarly, no one can sell electricity continuously
at a higher price than others can, regardless of how much they examine the power
market. Of course, a power generation company invests in a fuel production and/or
distribution business to ensure superior fuel procurement relative to other companies.
However, the investment is not considered to be part of the power generation business.
The power generation business can only be differentiated by how efficiently it
converts fuel to electricity. However, the efficiency of a power generation facility
depends on the plant manufacturer and not on the power generation operator.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 5
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_2
6 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Certainly, if a firm sells so much electricity that it requires many generators, it can
aim to build a generator portfolio that minimizes the expected future cash output
by considering the fixed and variable costs of various power generation methods.
Regretfully, few businesses in one market can take such measures. In addition,
rational maintenance and an efficient organizational structure can reduce operating
costs depending on the reliability of equipment. However, the effect is not large and
other firms easily catch up. Unfortunately, the price of commoditized goods gener-
ally approaches their marginal cost, so the power generation business tends to have
low profits, despite the burden of large-scale investment risk. If a power generation
company has only inferior production equipment relative to the market, it could go
out of business. Given these conditions, it is difficult for power companies that earn
gross profit only by the difference between the procurement price of fuel and the
selling price of electricity to improve their profit margins unless they take advantage
of this price difference.
Therefore, in this chapter, we propose arbitrage between the spot and futures
price differences and statistical arbitrage between fuel and power prices as trading
strategies that utilize the unit price difference between natural gas and electricity.
Moreover, our simulation with actual historical data shows that even if a company
owns an inferior power generator, it might be able to profit by adopting these arbitrage
strategies. We then estimate the five types of copulas to grasp the joint distribution of
gas and power price returns. The simulation with random numbers that follow those
copulas measures the risk of a portfolio consisting of gas futures short positions and
power futures long positions, which are held during statistical arbitrage.
Many studies analyze arbitrage. Alexakis [1] discusses the implications of statis-
tical arbitrage strategies based on the cointegration relationship between global stock
indexes. Mayordomo et al. [25] study statistical arbitrage strategies in credit deriva-
tive markets. Focardi et al. [10] introduce a new arbitrage approach based on dynamic
factor models of prices rather than returns, and demonstrate their performance in the
stock market. Hain et al. [15] examine the profits of a cointegration-based statistical
arbitrage strategy in European energy markets and confirm the statistically signifi-
cant risk-adjusted excess returns. Baviera and Baldi [2] introduce a new statistical
arbitrage strategy with stop-loss and leverage in high-frequency trading and apply
the method to the spread on heating oil and gas oil futures. Liu and Su [22] analyze
the dynamic causality between gold and silver returns and provide the implications
for statistical arbitrage strategies. Nakajima [26] investigates statistical arbitrage
using the cointegration relationship between wholesale electricity futures and natural
gas futures. Stübinger and Schneider [35] propose an integrated statistical arbitrage
strategy based on overnight price gaps and applied it to high-frequency represen-
tative stock data. Sánchez-Granero et al. [32] propose a novel approach based on
a statistical arbitrage technique to test the efficiency of the Latin American stock
markets. Keilbar and Zhang [19] result a great performance by a simple statistical
arbitrage trading strategy using the cointegration spreads between cryptocurrencies.
The remainder of this chapter is organized as follows. Section 2.2 explains the data
we use in this chapter. Section 2.3 proposes the two trading strategies. Section 2.4
2.2 Data and Preliminary Analyses 7

provides the simulation results. Section 2.5 measures the risk in statistical arbitrage.
Section 2.6 provides the overall concluding remarks and considerations.

2.2 Data and Preliminary Analyses

Here, we employ the PJM Western Hub Real-Time Peak to represent wholesale
power and Henry Hub Natural Gas to represent natural gas. We use futures prices
from January 2, 2015 to December 30, 2020 and spot prices from January 2, 2015 to
January 29, 2021. Henry Hub and PJM prices are expressed in USD per mmbtu and
per MWh, respectively. We obtain these daily data from Bloomberg. Henry Hub and
PJM are among the most representative natural gas and wholesale electricity sources
in the United States. In addition, we use the most recent daily data.

2.2.1 Descriptive Statistics

Before conducting various analyses and simulations, it is extremely important to

interpret the representative statistics of the data. Table 2.1 provides the summary
statistics of the Henry Hub and the PJM.
Considering that each future has a maturity of one month, we set each spot price
to January 29, 2021 and each future to December 30, 2020 to simulate the spot-future
arbitrage described later in Sect. 2.3.1. Because we extract only the days when both

Table 2.1 Descriptive statistics

Henry Hub PJM Western Hub peak
Futures Spot Futures Spot
Period From 1/2/2015 1/2/2015 1/2/2015 1/2/2015
To 12/30/2020 1/29/2021 12/30/2020 1/29/2021
Observations 1511 1477 1511 1477
Mean 2.65 2.64 36.49 35.18
Median 2.72 2.71 35.10 32.50
Maximum 4.84 7.13 87.60 366.91
Minimum 1.48 1.33 19.20 15.63
Standard deviation 0.50 0.56 8.75 17.24
Skewness 0.26 0.93 1.75 8.48
Kurtosis 4.26 8.84 8.95 124.26
Jarque–Bera 118 (0.00) 2313 (0.00) 3002 (0.00) 922,552 (0.00)
Note Values in parentheses indicate p-values
8 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

the Henry Hub and PJM data are available, we have 1511 and 1477 observations for
the futures and spot prices, respectively.
The mean and median are numerical values located in the center of the economic
variables. The mean x of the series (xi |i = 1, 2, . . . , N ) is calculated as

1 
N
x= xi . (2.1)
N i=1

On the other hand, the median is a value located in the center of each series
arranged in descending order. The medians of these futures and spot series are at
the 756th and 739th values, respectively. If the number of observations is even, then
the median is the average of the two data points in the center. Thus, the median is a
more stable index expressing the middle than the mean because outlier values have
less effect. Figure 2.1 shows three distribution examples with the same mean, but
different medians. Table 2.1 indicates that both the mean and median of each future
are higher than those of each spot. In other words, both Henry Hub and PJM tend
to be contango. We can infer that the supply and demand are not very tight during
this period. We can express the relationship between the future price p f and its spot
price ps as

p f = ps eCc ΔT , (2.2)

Fig. 2.1 Mean and median illustration

2.2 Data and Preliminary Analyses 9

where Cc is the cost of carry expressed in terms of yield and ΔT is the period from
the present to maturity. The cost of carry is the sum of the risk-free interest rate
and holding cost, expressed as yield minus the convenience yield. Therefore, if their
supply and demand remained tight during the period, then the utility of holding
their spots would be increasing. Thus, their costs of carry should become negative,
and their futures should become lower than their spots. In addition, the medians of
both the Henry Hub future and spot prices are higher than their respective means.
Therefore, we can expect to find many outliers in the left tail of each distribution.
On the contrary, the medians of both the PJM future and spot prices are lower than
their respective means. Therefore, we can expect to find many outliers in the right
tail of each distribution.
The next topic is the dispersion of each variable, which indicates the spread of
values from the mean. The simplest indicator is the range, calculated by subtracting
the minimum value from the maximum value. The ranges of the PJM and Henry
Hub futures are narrower than the range of each spot price. Furthermore, for all four
variables, the difference between the maximum and mean is much larger than that
between the minimum and mean. This indicates a long distribution in the right tail.
As mentioned above, the median provides an interpretation of dispersion by
comparing it with the mean. The percentile is a statistic developed based on the
median concept. The α-th percentile is the value located at the first α percentage in
ascending order. For example, the 65th percentile is the value located at 65% of the
sample size, counting from the minimum value. The 25th, 50th, and 75th percentiles
are called the lower quartile, median, and upper quartile, respectively. The percentile
is an important concept in risk measurement; however, we omit this statistic in Table
2.1, because we discuss risk measurement in Sect. 2.5.
One statistic that can express the dispersion of a distribution is the mean absolute
deviation, calculated as

1 
N
Mean absolute deviation = |xi − x|. (2.3)
N i=1

However, this statistic is rarely used, so we omit it from Table 2.1.

The most representative statistic for dispersion is the standard deviation σ ,
calculated using the following formula as the corrected sample standard deviation:
/
1 N
σ = (xi − x)2 . (2.4)
N −1 i=1

Table 2.1 shows that the standard deviation of each spot is larger than that of
each future. It turns out that the dispersion for their spots are larger than that of their
futures. The main reason is that it is technically difficult or too expensive to store
natural gas and power.
The skewness μ3 is a measure of the asymmetry of the distribution about its mean,
defined as
10 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

N ( )
1  xi − x 3
μ3 =  , (2.5)
N i=1 σ


where σ is the standard deviation of the observations, calculated using the following
equation:
/
 1 N
σ = (xi − x)2 . (2.6)
N i=1

The σ includes bias because it is calculated using the estimated mean x. The
standard deviation corrected for this bias is the σ defined in Eq. (2.4). However,


there is little difference between σ and σ in practice because we often have an

extremely large set of observations. If a distribution is symmetric, its skewness is 0.
Distributions with a long right tail have positive skewness, whereas those with a long
left tail have negative skewness. Figure 2.2 plots three distribution samples with the
same mean but different skewness values. Table 2.1 indicates that all variables have
positive skewness, meaning that the distribution has a long right tail. In particular,
the distortion to the right of the PJM spot distribution is clearly large.
The kurtosis μ4 is a measure of the sharpness and flatness of the distribution,
defined as
N ( )
1  xi − x 4
μ4 =  . (2.7)
N i=1 σ

Fig. 2.2 Distribution examples and their skewness

2.2 Data and Preliminary Analyses 11

The kurtosis of the normal distribution is 3. If a distribution has a sharp peak and
long fat-tails, then its kurtosis exceeds 3. Conversely, if a distribution has a rounded
peak and short lean tails, its kurtosis is less than 3. Figure 2.3 plots three distribution
samples with the same mean but different kurtosis values. Table 2.1 indicates that
the kurtosis of each series is larger than 3. Each variable had its own long fat-tail and
a peak distribution. In particular, the PJM spot price series is remarkable.
Jarque and Bera’s [16] test is a goodness-of-fit test for whether the skewness
and kurtosis of a distribution follow a normal distribution. In the null hypothesis
that observations are normally distributed, the Jarque–Bera statistic, defined by the
following equation, asymptotically has an χ 2 distribution with degree of freedom 2.
( )
N 1
J arque-Bera = μ3 2 + (μ4 − 3)2 . (2.8)
6 4

The Jarque–Bera statistics in Table 2.1 reject the normal distribution hypothesis
for all the variables. Electric power prices are more fat-tailed than gas prices and
their spot prices are more fat-tailed than their futures prices.
Finally, Figs. 2.4 and 2.5 provide the time plots of these future prices and spot
prices, respectively. Each variable has properties estimated from Table 2.1. We can
see that both Henry Hub and PJM spot prices fluctuate in the same way as their respec-
tive future prices. However, these spot prices spike frequently, and this tendency is
remarkable, especially for the PJM. In energy spot markets, price spikes and jumps
often occur because of a sudden increase in demand and shortage in supply capacity.
This is because equal amounts of supply and demand are simultaneously required.

Fig. 2.3 Distribution examples and their kurtosis

12 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Fig. 2.4 Natural gas and electricity futures prices

Unfortunately, it is extremely difficult or virtually impossible to store natural gas

and electricity. Moreover, we note that there may be a certain relationship between
the Henry Hub prices and the PJM prices. We can observe that the prices of gas and
power are synchronized, regardless of their spots and futures.

2.2.2 Stationarity and Unit Root Test

We must not forget the concept of “stationarity” in analyzing the time-series data of
economic variables. A stochastic process {xt } is defined as a stationary process if it
satisfies the following three conditions:
Condition 1. The expected value E(xt ) does not depend on time t. In other words,
all expected values are equal at any t.
Condition 2. The variance V(xt ) does not depend on time t. In other words, all
variances are equal at any t.
Condition 3. The autocovariance Cov(xt , xt−i ) does not depend on time t, but only
on time difference i (> 0).
As the most representative stationary process, we introduce white noise, which
satisfies the following three conditions:
Condition 1. The expected value E(xt ) is 0 at any t.
2.2 Data and Preliminary Analyses 13

Fig. 2.5 Natural gas and electricity spot prices

Condition 2. The variance V(xt ) is a constant at any t.

Condition 3. The autocovariance Cov(xt , xt−i ) is 0 if i is not 0.
Most market price series that energy market analysts frequently deal with are
nonstationary processes, like many other economic variable series.
To consider the characteristics of stochastic processes, suppose the p-th order
autoregression process of the following equation.


p
xt = θi xt−i + u t , (2.9)
i=1

where u t is the white noise. We can express the characteristic polynomial of Eq. (2.9)
as

m p − m p−1 θ1 − m p−2 θ2 − · · · − θ p = 0. (2.10)

If one of the solutions of this equation is less than one, then this {xt } is a stationary
process. If one of the roots is one, then {xt } is a unit root process. If one root is larger
than 1, then {xt } is called an explosive process. Both unit root and explosive processes
are nonstationary processes.
Next, we discuss spurious regressions as one of the most important points to keep
in mind when analyzing a unit root process. Suppose two unit root processes xt and
yt that are independent of each other. We estimate the following regression model
14 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

using ordinary least squares (OLS).

yt = α + βxt + εt . (2.11)

Theoretically, xt and yt are irrelevant. Therefore, β should be 0. However, the

estimated model has the following characteristics.
Characteristic 1. The correlation coefficient is large, meaning that the model fits
well.
Characteristic 2. The t-value is large, meaning that β is significant.
Characteristic 3. The Durbin-Watson ratio is small, meaning that εt has a positive
autocorrelation.
Granger and Newbold [13] discovered the phenomenon called “spurious regres-
sion” using Monte Carlo simulations. Phillips [29] proves this analytically. We must
study economic variables with great care because most of them have unit roots. We
provide an example below. We generate 100,000 data points for each series using
the following formula:

xt = xt−1 + εxt
yt = yt−1 + ε yt , (2.12)

where x1 = y1 = 0, and εxt and ε yt are random numbers that follow an independent
standard normal distribution. Figures 2.6 and 2.7 show the time series and scatter
plots, respectively. Although xt and yt have no theoretical relationship, they seem to
be in harmony. The estimated result of Eq. (2.11) using OLS is

Correlation coefficient = 0.903

β = −144(t-value = −518, p-value = 0)
Durbin-Watson ratio = 0.000 (2.13)

This result thus represents a spurious regression.

The augmented Dickey–Fuller (ADF) unit root test is based on the approach
proposed by Dickey and Fuller [5]. The ADF test is one of the most representative
unit root tests, with a unit root null hypothesis and stationary alternative hypothesis.
We transform Eq. (2.9) into


p
Δxt = ξ xt−1 + ηi Δxt−i + u t . (2.14)
i=1

The null hypothesis is that the unit root process is ξ = 0, and the alternative
hypothesis is that the stationary process is ξ < 0. In the null hypothesis, the t-value
of ξ does not have a Student’s t distribution, but has a specific distribution known
as the Dickey–Fuller (DF) table. This table tests the unit root process. A common
2.2 Data and Preliminary Analyses 15

Fig. 2.6 Spurious regression example (series plot)

Fig. 2.7 Spurious regression example (scatter plot)

16 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

assumption is that a constant term is added to Eq. (2.9). Furthermore, it is also

frequently supposed that a time trend term is added. Each model has the following
equations:


p
Δxt = δ + ξ xt−1 + ηi Δxt−i + u t (2.15)
i=1


p
Δxt = δ + ζ t + ξ xt−1 + ηi Δxt−i + u t , (2.16)
i=1

where δ is a constant term and ζ t is a time trend term. Regarding the selection of
autoregression lag order p, it is reasonable to extend the order until the test for the
residual term no longer accepts autocorrelation. The possibility of autocorrelation
of u t extends the model. However, we often simply select the order based on infor-
mation criteria such as the Akaike information criterion (AIC), Schwarz Bayesian
information criterion (SBIC), and Hannan-Quinn information criterion. In addition,
we can select the order by referring to previous studies on the same economic vari-
able. The Dickey and Fuller [5] test is an ADF test that uses a model without lagged
terms.
Phillips et al. [31] propose an ADF test for rational bubbles using the right tail
of the DF distribution. In other words, the unit root process null hypothesis remains
ξ = 0, whereas the explosive process alternative hypothesis is ξ > 0.
Other typical unit root tests include the PP test developed by Phillips and Perron
[30], and the KPSS test by Kwiatkowski et al. [20]. The PP test, similar to the ADF
test, is based on a unit root null hypothesis and a stationary alternative hypothesis,
whereas the KPSS test presupposes a stationary null hypothesis and a unit root
alternative hypothesis. We should select the test type in consideration of the power
and purpose of the unit root test.
Table 2.2 shows the ADF unit root test results for each energy price and its
first difference. Equation (2.14) is the model adopted. For all variables, their levels
accept the unit root hypothesis and their first difference rejects the null hypothesis.
In general, most price series are unit root processes and their first difference series
are stationary processes. These results are consistent with previous studies.

2.2.3 Cointegration Test

Figures 2.1 and 2.2 bring to mind the long-term equilibrium relationship between
Henry Hub and the PJM in both futures and spot markets. However, as all four
variables accept the unit root hypothesis, we must suspect a spurious regression.
Engle and Granger [7] introduced the concept of “cointegration,” which connects
multiple unpredictable stochastic variables with a unit root. If a linear combination
of multiple unit root processes is stationary, then these variables have a cointegrated
2.2 Data and Preliminary Analyses 17

Table 2.2 ADF unit root

Commodity Variable ADF-t statistic
tests
Henry Hub futures Level –0.80 (0.37)
First difference –30.48* (0.00)
PJM Western Hub peak Level –1.34 (0.17)
futures First difference –18.83* (0.00)
Henry Hub spot Level –0.88 (0.34)
First difference –13.75* (0.00)
PJM Western Hub peak Level –2.46 (0.01)
spot First difference –19.43* (0.00)
Note * indicates that the unit root hypothesis is rejected at the 1%
significance level. The values in parentheses indicate p-values

relationship. In other words: suppose that the following vector consists of ν variables
in a unit root process:

Xt = T (x1t , x2t , . . . , xνt ). (2.17)

The following linear combination is derived from the inner product of the ν
dimensional coefficient vector and Xt :

βXt = (β1 , β2 , . . . , βν ) T (x1t , x2t , . . . , xνt ). (2.18)

If β X t is a stationary process, then x1t , x2t , . . . , xνt have a cointegrated relation-

ship. Additionally,

β = (β1 , β2 , . . . , βν ) (2.19)

is the cointegrating vector. If there is cointegration between some variables, then

the deviation of the observed values from their long-term equilibrium is a stable
stochastic process. Because many economic variables have unit roots, this concept
is very often applied in a wide range of fields to examine the relationships between
economic variables.
Therefore, we test whether the Henry Hub and PJM prices are cointegrated and
expect to use this cointegrated relationship in the trading strategies.
Engle and Granger’s [7] proposed test for cointegration has limitations. First, it
does not expect a system with three or more variables to have two or more coin-
tegration relationships. Second, the test results may change when the variables are
interchanged.
Section 2.2.3 adopts the test proposed by Johansen [17] and Johansen and Juselius
[18]. We assume the following model:
18 2 Arbitrage Trading in Energy Markets and Measuring Its Risk


p−1
ΔXt = AX t−1 + B i ΔX t−i + ut . (2.20)
i=1

Because the cointegration condition is that Eq. (2.18) is a stationary process, we

can add a constant term vector or time trend vector to Eq. (2.20). The rank of A
represents the number of independent cointegrating vectors; therefore, we have the
following three cases:
Case 1. rank( A) = ν: All elements of X t are a stationary process.
Case 2. rank( A) = ρ(0 < ρ < ν): There are ρ cointegrating vectors.
Case 3. rank( A) = 0: β X t is a nonstationary process.
Therefore, after estimating Eq. (2.20), we perform the cointegration test based on
the eigenvalues of A.
Section 2.2.3 tests for the relationship between Henry Hub prices and PJM prices
in the futures and spot markets. Let the eigenvalues of A be k1 , k2 (k1 > k2 ). The
acceptance of the unit root hypotheses should not hold rank( A) = 2. Therefore, if
there is a cointegrating vector; that is, if rank( A) = 1, then the following equations
hold:

log(1 − k1 ) < 0
log(1 − k2 ) = 0. (2.21)

Alternatively, if there is no cointegrating vector; that is, if rank( A) = 0, then

log(1 − k1 ) = 0
log(1 − k2 ) = 0. (2.22)

In general, the null hypothesis of the trace test is the existence of ρ or fewer
cointegrating vectors, and the alternative hypothesis is that there are more than ρ
cointegrating vectors. The test statistic ktrace is
ν
 (  )
ktrace = −T log 1 − k i , (2.23)
i=ρ+1

where k i is the estimated eigenvalue and T is the number of observations. We should

test the relationship for each value of ρ = 0, 1, . . . , υ − 1. Since Sect.2.2.3 adopts
a two-variable model, the null hypothesis is one or fewer cointegrating vectors,
and the alternative hypothesis is two or more cointegrating vectors. The statistics
ktrace,ν=2,ρ=0 and ktrace,ν=2,ρ=1 are
( ) ( )
ktrace,ν=2,ρ=0 = −T log 1 − k̂1 − T log 1 − k̂2
2.2 Data and Preliminary Analyses 19

Table 2.3 Johansen

Hypothesized number Trace test Maximum eigenvalue
cointegration test: Futures
of coefficients test
market
None 31.18* (0.00) 30.72* (0.00)
One 0.47 (0.56) 0.47 (0.56)
Note * indicates rejection of the hypothesis at the 1% significance
level. The values in parentheses indicate p-values

( )
ktrace,ν=2,ρ=1 = −T log 1 − k̂2 . (2.24)

In general, the null hypothesis of the maximum eigenvalue test is the existence
of ρ cointegrating vectors, and the alternative hypothesis is that ρ + 1 cointegrating
vectors exist. The test statistic kmax is
( ) 

kmax = −T log 1 − k ρ+1 . (2.25)

Since Sect. 2.2.3 uses a two-variable model, the null hypothesis is one coin-
tegrating vector, and the alternative hypothesis is two cointegrating vectors. The
statistics kmax,ν=2,ρ=0 and kmax,ν=2,ρ=1 are
( )
kmax,ν=2,ρ=0 = −T log 1 − k̂1
( )
kmax,ν=2,ρ=1 = −T log 1 − k̂2 . (2.26)

Tables 2.3 and 2.4 present the Johansen cointegration test results for Henry Hub
prices and PJM prices in the futures and spot markets, respectively. Both the trace test
and the maximum eigenvalue test reject the hypothesis of no cointegrating vector and
accept the hypothesis of one cointegrating vector. We have the same result in both the
futures and spot markets. We can conclude that there is a cointegration relationship
between Henry Hub prices and PJM prices.

Table 2.4 Johansen

Hypothesized number Trace test Maximum eigenvalue
cointegration test: Spot
of coefficients test
market
None 147.97* (0.00) 146.79* (0.00)
One 1.18 (0.32) 1.18 (0.32)
Note * indicates rejection of the hypothesis at the 1% significance
level. The values in parentheses indicate p-values
20 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

2.2.4 Long-Term Equilibrium Estimation

We can estimate the cointegrating vectors by using dynamic OLS (DOLS). OLS
estimates the following equation with lag terms for the explanatory variables to
eliminate autocorrelation:
⎛ ⎞
ν−1
 
K
xν,t = ϕ0 + ⎝βi ϕi,t + φi, j Δxi,t− j ⎠. (2.27)
i=1 j=−K

Since Sect. 2.2.4 utilizes a two-variable model, the model for estimating the long-
term equilibrium is


K
P J M t = ϕ0 + ϕ1 H enr y H ubt + φ j ΔH enr y H ubt− j . (2.28)
j=−K

The lag order K was determined using SBIC. The long-term equilibrium equation
for future prices is

P J M f utur e = 10.673 × H enr y H ub f utur e + 8.159. (2.29)

The long-term equilibrium equation for the spot prices is

P J M spot = 11.142 × H enr y H ubspot + 5.732. (2.30)

2.3 Trading Strategies

The only way to profit by trading goods is to “buy at a lower price and sell at a
higher price.” If we trade only one item, then price forecasting is the most important
matter. Is this realistically possible? A market is efficient if the information that
affects the market price is comprehensive, constant, and has a timely effect on the
price. Markets for securities and commodities listed on exchanges are almost efficient
and depend on liquidity. In other words, we cannot forecast the price because the
price already reflects all the currently available information, and any information that
affects the price will occur independently of the price. Unfortunately, it is impossible
for market participants to earn returns above the market average. Certainly, a “fully
efficient market” is theoretical or virtual. Therefore, some investors and speculators
try to collect information before it is reflected in the price. However, these actions
make the market more efficient. Because the stationary hypothesis for most energy
prices is rejected by the unit root test using daily data, energy companies should
consider energy markets as efficient, and energy prices as unpredictable.
2.3 Trading Strategies 21

In general, power companies procure various types of fuels from various markets,
produce electricity using various power generation methods, and sell the power
through various sales channels. Section 2.3 assumes a simple model of purchasing
natural gas at the Henry Hub price and selling electricity at the PJM price, as Fig. 2.8
illustrates. We propose two trading strategies. Section 2.4 will simulate these methods
using actual historical data. Both focus not on these prices but on the price difference
between Henry Hub prices and PJM prices. We cannot expect profit owing to market
efficiency, even if we analyze each price in detail. On the other hand, we demonstrate
the potential to make a profit by investigating price differences, which is a stationary
process. When buying the gas required to produce one unit of electricity and selling
it, the gross margin is often called the spark spread.
The trading strategy introduced in Sect. 2.3.1 is the arbitrage between the futures
market spreads and a company’s spreads expected from its power generation effi-
ciency. This takes advantage of the spread of futures as a stochastic process. All we
have to do is take the Henry Hub long position and the PJM short position to secure
profits when a favorable futures spread occurs stochastically. The strategy proposed
in Sect. 2.3.2 is statistical arbitrage utilizing the cointegration relationship between
Henry Hub prices and PJM prices in the futures market. Making use of the long-
term equilibrium equation in the futures market that expresses the futures spread, the
lower PJM long positions and the higher Henry Hub short positions are expected to
yield profit in the narrower spreads than the market when the spread approaches the
long-term equilibrium. Of course, the higher PJM short positions and lower Henry
Hub long positions, which are taken in wider spreads than the market, are expected

Fig. 2.8 Trading model

22 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

to make a profit. However, this is almost always included in the arbitrage between
futures spreads and the company’s spot spreads. Therefore, we did not simulate this
method.

2.3.1 Arbitrage Between Own Spot Spread and Future Spread

Power generation costs consist of fixed costs (e.g., equipment depreciation costs,
labor costs, and maintenance costs) and variable costs (e.g., fuel costs and exhaust
processing costs). However, the variable cost is almost the fuel cost. Therefore, the
unit cost of gas-fired generation expressed as Cost and the corresponding natural
gas procurement cost expressed as Gas satisfy the following equation:

Cost = α0 × Gas + α1 , (2.31)

where α0 and α1 are the coefficients. Although the Henry Hub futures price
H enr y H ub f,t and the PJM futures price P J M f,t , both of which are unit root
processes, are cointegrated, the long-term equilibrium equation, which is a stochastic
process, can have large outliers. Then, when the own spot spread, that is, the differ-
ence between the power generation unit cost and the corresponding gas procurement
unit price, is smaller than the future spread, that is, the future price difference between
the PJM and Henry Hub, we swap the spot spread, Spr ead s and the future spread,
Spr ead f , which we express as

Spr ead s = α0 × H enr y H ub f,t + α1 − H enr y H ub f,t (2.32)

Spr ead f = P J M f,t − H enr y H ub f,t . (2.33)

Therefore, the difference between these spreads is

Spr ead s − Spr ead f = α0 × H enr y H ub f,t + α1 − P J M f,t . (2.34)

In the following equation:

α0 × H enr y H ub f,t + α1 − P J M f,t < 0. (2.35)

If we take the Henry Hub long position and the PJM short position corresponding
to the electric energy planned for generation, we can lock in profit.
2.3 Trading Strategies 23

2.3.2 Statistical Arbitrage

By estimating the long-term equilibrium equation of H enr y H ub f,t and P J M f,t in

a cointegration relationship, we can determine whether the futures spread on a candi-
date trading date is wider or narrower than the expected spread. This determination
enables statistical arbitrage trading between Henry Hub and PJM.
Since the prices in period t are not available for trading in period t, we estimate the
following long-term equilibrium equation using the price series up to period t − 1:

P J M f,t = β f,0 × H enr y H ub f,t + β f,1 . (2.36)

If the futures spread is higher than the expected value, then we express it as

P J M f,t > β f,0 × H enr y H ub f,t + β f,1 . (2.37)

We can consider that the PJM price is higher and the Henry Hub price is lower;
therefore, we take the PJM short position and Henry Hub long position. Then, the
condition for closing these arbitrage positions is

P J M f,t − avgShor t P J M f + avgLong H enr y H ub f

− H enr y H ub f,t > 0, (2.38)

where avgShor t P J M f is the average price of the PJM futures short positions taken,
and avgLong H enr y H ub f is the average price of the Henry Hub futures long posi-
tions taken. The clearance of all these futures positions under this condition leads to
profit.
Conversely, if the futures spread is below the expected value, then we express it
as

P J M f,t < β f,0 × H enr y H ub f,t + β f,1 . (2.39)

We determine that the PJM price is lower and the Henry Hub price is higher;
therefore, we take the PJM long position and Henry Hub short position. Then, the
condition for closing these arbitrage positions is

avgLong P J M f − P J M f,t + H enr y H ub f,t

− avgShor t H enr y H ub f > 0, (2.40)

where avgLong P J M f is the average price of the PJM futures long positions and
avgShor t H enr y H ub f is the average price of the Henry Hub futures short positions.
If this condition is satisfied, we can lock in profits by clearing all these positions.
As mentioned above, there can be two patterns of statistical arbitrage trading:
one is when the spread is larger than the expected value, and the other when the
spread is smaller than the expected value. However, Sect. 2.4 simulates only the case
24 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

of Eq. (2.39) where the spread is narrower because the case of Eq. (2.37), which
indicates a wider spread, often includes the case of Eq. (2.35).
Statistical arbitrage trading can take each position of any size in a single transac-
tion. In Sect. 2.4, we take the Henry Hub short position and the PJM long position
corresponding to the planned electric energy generation.

2.4 Simulation Results

This section simulates the trading strategy in Sect. 2.3 using one year of historical
Henry Hub and PJM futures and spot data. We adopt futures with a one-month
maturity. Thus, the futures trading period is from January 1, 2020 to December 31,
2020, and the spot trading period is from February 1, 2020 to January 31, 2021.
Table 2.5 presents the specifications of the assumed power generation business
used in the simulation. We calculate the fixed unit cost by allocating the annual fixed
cost to the peak hours, and dividing it by the amount of electric energy sold in one day.
Calculating each coefficient of Eq. (2.31) from these yields the following equation:

Cost[USD/MWh] = 7.590 × Gas[USD/mmbtu] + 10.464. (2.41)

As a reference, we estimate the cointegrating vector by the DOLS using the

historical data of these spot prices from February 1, 2020 to January 31, 2021.
Consequently, the long-term equilibrium equation between the Henry Hub spot price
H enr y H ubs,t and the PJM spot price P J M s,t during this period is

P J Ms,t [USD/MWh] = 6.551

× H enr y H ubs,t [USD/mmbtu] + 11.576 (2.42)

Figure 2.9 plots Eqs. (2.41) and (2.42), which express the relationship between
the unit prices of natural gas and electricity. Since the lowest price of the Henry Hub

Table 2.5 Power generation business specifications

(A) Electric Power (Power transmission end) 600 MW
(B) Power selling time 19 h/day
(C) Electric energy to sell (= A × B) 11,400 MWh/day
(D) Thermal efficiency 45 %
(E) Natural gas energy required (= C/D) 86,528 mmbtu/day
(F) Fixed cost 55,000,000 USD/year
(G) α0 (= E/C) 7.590 mmbtu/MWh
(H ) α1 (= F/C/365 × B/24) 10.464 USD/MWh
2.4 Simulation Results 25

spot during the trading period is 1.33 [USD/mmbtu], the power generation cost is
always higher than the expected PJM spot price. In other words, the competitive-
ness of the power generation company is less than the spot market. Regrettably, the
expected value of profit from spot trading alone is negative. Substituting the mean
value of the Henry Hub spot prices during the simulation period, which is about 2.05
[USD/mmbtu], in Eq. (2.41), the mean value of costs is about 26.05 [USD/MWh]. On
the other hand, the mean value of the PJM spot prices is about 25.36 [USD/MWh].
Thus, the expected loss in spot trading is

E x pected Loss = (26.05 − 25.36) × 11400 × 365

 
USD
= 2, 873, 803 . (2.43)
year

As we can see, this simulation sets conditions that are disadvantageous to the
business operator.
To reflect changes in the market environment (e.g., economic conditions, season-
ality, climate), this statistical arbitrage simulation estimates the long-term equilib-
rium equation of these futures prices monthly. For example, statistical arbitrage in
January 2020 uses the equilibrium equation estimated by DOLS using observations
from January 2015 to December 2019. We estimate the long-term equilibrium equa-
tions used in the subsequent trading months by the moving window method. Table
2.6 summarizes the results.

Fig. 2.9 Power generation cost and long-term equilibrium relationship in the spot market
26 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Table 2.6 Long-term equilibrium estimation

Trading period Long-term equilibrium Observation period
P J M f,t =
β f,0 × H H f,t + β f,1
β f,0 β f,1
January 2020 9.18 13.12 January 2015 to December 2019
February 2020 9.16 13.00 February 2015 to January 2020
March 2020 8.81 13.44 March 2015 to February 2020
April 2020 9.31 11.78 April 2015 to March 2020
May 2020 9.76 10.40 May 2015 to April 2020
June 2020 10.40 8.50 June 2015 to May 2020
July 2020 10.31 8.67 July 2015 to June 2020
August 2020 10.18 8.98 August 2015 to July 2020
September 2020 10.20 8.84 September 2015 to August 2020
October 2020 10.32 8.38 October 2015 to September 2020
November 2020 10.27 8.27 November 2015 to October 2020
December 2020 10.21 8.20 December 2015 to November 2020

Figure 2.10 plots the coefficients of the monthly long-term equilibrium equations.
These coefficients capture changes in the surrounding market conditions. This tech-
nique, which is able to recognize changes in the market structure, can be expected
to generate profits as long as we do not hold a position for a long time.
The simulation trades these spots from February 2020 to January 2021, and futures
from January 2020 to December 2020. The spot-future arbitrage and statistical arbi-
trage methods take positions corresponding to the electric energy planned to be
sold.
The simulation results are as follows. Table 2.7 indicates the profit from each
trade. The unrealized loss is due to futures that are not cleared on the last trading
day. Spot-future arbitrage makes a profit of 6,446,848 [USD], which can cover the
spot trading loss of 2,881,676 [USD]. The realized profit and unrealized loss from
statistical arbitrage are 1,086,396 [USD] and 211,463 [USD], respectively. Therefore,
the total profit is 874,933 [USD].
Figure 2.11 provides the time plots of each trading strategy’s profit. The black
line representing spot trading started in February 2020, because there was no spot
trading in January 2020. The red, blue, and green lines express each arbitrage ending
in December 2020 because the simulation does not trade futures in January 2021.
Table 2.8 provides the results for each combination of spot and futures trades at
the end of the simulation period. If we trade only spots, we make a loss of 2,881,676
[USD], which is almost the same as the loss of 2,873,803 [USD] estimated from
Eq. (2.43). By adopting both trading strategies, we earn a profit of 4,440,105 [USD].
Applying only the spot-future arbitrage results in a profit of 3,565,172 [USD], while
2.4 Simulation Results 27

Fig. 2.10 Coefficients of the equilibrium equation for statistical arbitrage

Table 2.7 Profit from each

Realized profit Unrealized Total profit
trade
profit
Spot trade –2,881,676 0 –2,881,676
Spot-future 6,446,848 0 6,446,848
arbitrage
Statistical 1,086,396 –211,463 874,933
arbitrage

applying only the statistical arbitrage results in a loss of 2,006,743 [USD] which is
less than the loss of 2,881,676 [USD] when only spot trading occurs.
Figure 2.12 presents the time-series plots of the profit from each combination of
trading strategies. We can observe that the effectiveness of spot-future arbitrage is
larger than that of statistical arbitrage throughout the period. However, the larger the
leverage of statistical arbitrage, which is defined as the ratio of futures positions to
spot positions, the greater the effectiveness of statistical arbitrage.
Therefore, to increase profits from statistical arbitrage trading, we consider
increasing leverage. Table 2.9 shows the maximum unrealized loss during the trading
period and total profit at the end of the trading period by leverage size. Clearly, both
losses and profits are proportional to leverage. Therefore, if we take futures positions
three times as great as spots, then statistical arbitrage can cover the loss from spot
trading. Companies want to maximize their profits, but the risk they can take depends
28 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Fig. 2.11 Time-series plots of the profit from each trade

Table 2.8 Profit from each

Statistical arbitrage No statistical
combination of the two
arbitrage
trading strategies
Spot-future 4,440,105 3,565,172
arbitrage
No spot-future –2,006,743 –2,881,676
arbitrage

mainly on their capital. We must rationally manage the maximum unrealized loss
during the trading period to maximize the profit from statistical arbitrage.

2.5 Risk Measurement in Statistical Arbitrage

If the expected return of a statistical arbitrage is positive, then we should maximize

the futures positions to maximize profits. However, the trading technique expects
a long-term equilibrium in the future, so we should manage risk, which means a
possible future loss. This is because excessive unrealized losses may hinder business
operations. Risks that occur during trading mainly include credit risk caused by the
default of business partners, operational risk caused by failure in daily business activ-
ities and procedures, liquidity risk caused by an extreme decrease in trading volume
that forces us to settle at a significantly unfavorable price, and market risk caused
2.5 Risk Measurement in Statistical Arbitrage 29

Fig. 2.12 Time-series plots of the profit from each combination of the two trading strategies

Table 2.9 Maximum unrealized loss and total profit from statistical arbitrage
Leverage Maximum unrealized loss during the Total profit at the end of the trading
trading period period
1 1,362,795 874,933
2 2,725,589 1,749,866
3.3 4,497,223 2,887,279

by fluctuations in market prices. Chapter 2 covers only the market risk. We quanti-
tatively measure the risk of a portfolio held during statistical arbitrage consisting of
Henry Hub futures short positions and PJM futures long positions. This risk is due
to fluctuations in futures prices.

2.5.1 Value-At-Risk and Expected Shortfall

This section introduces value-at-risk (VaR) and expected shortfall, which are the two
most representative risk indicators that summarize portfolio risk into one numerical
value.
VaR is the predicted maximum loss of a portfolio currently held in a given period
within a given confidence interval, and is expressed in monetary amounts or yields.
For example, if the maximum loss rate of a portfolio that can occur with a 99%
30 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

probability after one day is 10%, then the VaR at a confidence level of 99% with
a one-day holding period is 10%, or the 1-day 99% VaR is called 10%. We can
also interpret this condition as a 1% chance of a return worse than − 10% after one
day. Figure 2.13 expresses the 1-day 99% VaR when the return on the portfolio is
normally distributed.
VaR is common and is one of the most attractive risk indicators because it is
relatively easy to measure and helps us understand the worst risk. However, VaR
can be misleading about risk, depending on the shape of the left tail of the return
distribution. Supposing the distribution of the return of a portfolio in Fig. 2.14, which
is a normal distribution with the left tail less than –10% partially inverted, we can
understand that the VaR in Fig. 2.14 is same as in Fig. 2.13, while the risk in Fig. 2.14
is greater than in Fig. 2.13. VaR is not appropriate as a single numerical indicator of
portfolio risk, depending on the shape of the left tail. Therefore, expected shortfall
is the conditional expected loss in the case of loss exceeding VaR; in other words,
expected shortfall can be understood as the mean value of loss in the worst case, and
is consistent with intuition. We see that although the VaRs in Figs. 2.13 and 2.14 are
at the same point, the expected shortfall in Fig. 2.14 is to the left of the expected
shortfall in Fig. 2.13.

Fig. 2.13 One-day 99% VaR and the expected shortfall in a normal distribution
2.5 Risk Measurement in Statistical Arbitrage 31

Fig. 2.14 One-day 99% VaR and the expected shortfall in a normal distribution with a deformed
left tail

2.5.2 Copula

There are historical and parametric methods for measuring VaR and expected short-
falls. The historical method does not mathematically assume the distribution of value
fluctuations but finds the percentile or conditional mean from the actual historical
data as VaR or expected shortfall. Although this method can easily measure these
indicators even in a complicated distribution, the measurement results are extremely
dependent on the observations. The parametric method models the distribution of
value fluctuations to calculate VaR or expected shortfall. In some cases, we may math-
ematically calculate the risk values easily by assuming a normal or lognormal distri-
bution. In other cases, Monte Carlo simulations often derive risk values by assuming
a complicated distribution. This measurement result depends on the assumed distri-
bution model and estimated parameters. Therefore, the selection of a distribution
model and estimation of its parameters are very important. As a parametric method,
this section introduces a method using copulas, which convert from multiple marginal
distributions into joint distributions. Sklar [33] proposed the copula concept.
Assuming m returns as risk factors, which are expressed by the stochastic variables
R1 , R2 , . . . , Rm , let their distribution functions be F1 , F2 , . . . , Fm . To capture the
stochastic value fluctuations of the portfolio consisting of m securities, we must esti-
mate the joint distribution function F of the stochastic variables R1 , R2 , . . . , Rm from
the marginal distribution functions F1 , F2 , . . . , Fm . In this case, a unique function C
exists that satisfies the following equation:
( )
Probability R1  ρ1 , R2  ρ2 , . . . , Rm  ρm
32 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

= F(ρ1 , ρ2 , . . . , ρm ) = C(F1 (ρ1 ), F2 (ρ2 ), . . . , Fm (ρm )). (2.44)

This is known as Sklar’s [33] theorem. Thus, for any u = (u 1 , u 2 , . . . , u m ), the

following equation holds:
( )
C(u 1 , u 2 , · · · , u m ) = F F1−1 (u 1 ), F2−1 (u 2 ), . . . , Fm−1 (u m ) , (2.45)

where 0  u 1 , u 2 , . . . , u m  1 is satisfied. Further, let c, f 1 , f 2 , . . . , f m , f be

the density function of C, F1 , F2 , . . . , Fm , F, respectively. We obtain the following
equation from Eq. (2.44):


m
f (r1 , r2 , . . . , rm ) = c(F1 (r1 ), F2 (r2 ), . . . , Fm (rm )) f i (ri ). (2.46)
i=1

Therefore, the log-likelihood function ll(π 1 , π 2 , . . . , π m ; π ) is

ll(π 1 , π 2 , . . . , π m ; π )

N
( ( ) ( ) ( ) )
= log c F1 r1, j ; π 1 , F2 r2, j ; π 2 , . . . , Fm rm, j ; π m ; π
j=1


N 
m
( )
+ log f i ri, j ; π i , (2.47)
j=1 i=1

where ri, j is the j-th observation of the i-th stochastic variable (i = 1, 2, . . . , m; j =

1, 2, . . . , N ), π i is the parameters of the marginal distribution function Fi (·; π i )
(i = 1, 2, . . . , m), and π is the copula parameters. Estimating Eq. (2.47) by the
maximum likelihood means estimating all parameters, namely copula and marginal
distribution function parameters, simultaneously. The copula is estimated after esti-
mating the marginal distributions considering the computational load. We adopt a
double exponential distribution, expressed as follows, for the probability density
function f i of the marginal distribution function Fi :
(  )
1  x − μi 
f i (x) = √ exp − √  ,
 (2.48)
2σi σi / 2

where μi and σi are the mean and standard deviation, respectively. Integrating
Eq. (2.48), we obtain the following equation for the marginal distribution function
Fi :
( )
1 x − μi
Fi (x) = 1 − exp − √ , if x > μ
2 σi / 2
( )
1 x − μi
Fi (x) = exp √ , if x  μ (2.49)
2 σi / 2
2.5 Risk Measurement in Statistical Arbitrage 33

From Eq. (2.49), we obtain the inverse function:

σi 1
x = μi − √ log(2 − 2Fi (x)), if Fi (x) >
2 2
σi 1
x = μi + √ log(2Fi (x)), if Fi (x)  (2.50)
2 2

The most common copulas are broadly divided into two families: the elliptical
copula and Archimedean copula. First, the two most common elliptical copula func-
tions are the Gaussian copula, introduced by Lee [21] and generalized by Ophem [28],
and the t copula, proposed by Embrechts et al. [6] and Fang et al. [8]. These copulas
have elliptical distributions. These represent the correlation structure between the
stochastic variables as a matrix. The three most popular Archimedean copulas are
Gumbel [14], Clayton [4], and Frank [11] copulas. These copulas express a corre-
lation structure with a single parameter. Here, we provide a brief description of the
Gaussian, t, Gumbel, Clayton, and Frank copulas, although we refer the reader to
Nelsen [27] for the theory and precise properties.
Gaussian Copula
Given that m stochastic variables follow the m variate standard normal distribution
m with the correlation matrix Σ, Gaussian copula C Gaussian (u 1 , u 2 , . . . , u m ; Σ) is

C Gaussian (u 1 , u 2 , . . . , u m ; Σ)
( )
= m −1 (u 1 ), −1 (u 2 ), . . . , −1 (u m ); Σ , (2.51)

where −1 is the inverse function of the univariate standard normal distribution. The
density function cGaussian (u 1 , u 2 , . . . , u m ; Σ) is

∂C Gaussian
m
(u 1 , u 2 , . . . , u m ; Σ)
cGaussian (u 1 , u 2 , . . . , u m ; Σ) =
∂u 1 ∂u 2 . . . ∂u m
( )
1 1 ( )
=√ exp − T ω Σ −1 − E ω , (2.52)
|Σ| 2

where E denotes the identity matrix, and

ω = T (ω1 , ω2 , . . . , ωm )
( )
= T −1 (u 1 ), −1 (u 2 ), . . . , −1 (u m )
= −1 (u). (2.53)

Thus, using the N observations of each stochastic variable, the log-likelihood

ll Gaussian is
34 2 Arbitrage Trading in Energy Markets and Measuring Its Risk


N
∂C m Gaussian (u 1 , u 2 , . . . , u m ; Σ)
ll Gaussian (Σ) = log
j=1
∂u 1 ∂u 2 . . . ∂u m

1  T ( −1 )
N
N
=− log|Σ| − ωj Σ − E ωj (2.54)
2 2 j=1

where
( )
ω j = −1 u j . (2.55)


Therefore, the maximum likelihood estimator Σ is given by the following

equation:

1  T

N
Σ =− ωj ωj. (2.56)
N j=1

t Copula
Given that m stochastic variables follow the m variate t distribution tdmf, with degree
of freedom d f ( 3) and the correlation matrix Σ, the t copula Ct (u 1 , u 2 , . . . , u m )
is
( )
Ct (u 1 , u 2 , . . . , u m ; Σ, d f ) = tdmf, td−1
f (u 1 ), t −1
df (u 2 ), . . . , t −1
df (u m ) , (2.57)

where td−1
f is the inverse function of the univariate t distribution with degree of
freedom d f . The density function ct (u 1 , u 2 , . . . , u m ; Σ, d f ) is

∂Ctm (u 1 , u 2 , . . . , u m ; Σ, d f )
ct (u 1 , u 2 , . . . , u m ; Σ, d f ) =
∂u 1 ∂u 2 , . . . , ∂u m
( )( ( ))m ( )− d f 2+m
ωΣ −1 ω
 d f 2+m  d2f
T
1 + df
= ( )( ( ))m  ( ) d f +1 , (2.58)
√ d f +1 N ω2j − 2
|Σ| 2  2
df
j=1 1 + d f

where

ω = T (ω1 , ω2 , . . . , ωm )
( )
= T td−1
f (u 1 ), t −1
df (u 2 ), . . . , t −1
df (u m ) . (2.59)

Thus, using the N observations of each stochastic variable, the log-likelihood ll t

is
2.5 Risk Measurement in Statistical Arbitrage 35


N
∂C m (u 1 , u 2 , . . . , u m ; Σ, d f )
llt (Σ, d f ) = log t

j=1
∂u 1 ∂u 2 . . . ∂u m
⎛ ( d f +m ) ⎞ ⎛ (df ) ⎞
 2  2 N
= N log⎝ ( ) ⎠ + m N log⎝ ( ) ⎠ − log|Σ|
d f +1 2
 2 df
 2
( )
df +m 
N T
ω j Σ −1 ω j
− log 1 +
2 j=1
df

df + 1  ( ( )2 )
N m
+ log 1 + ω j i , (2.60)
2 j=1 i=1

where
( )
ω j = td−1
f uj . (2.61)


Therefore, the maximum likelihood estimator Σ satisfies

df +m 
N

ωjTω
Σ= −1
.  (2.62)
N j=1 d f + ωΣ ω j
T

Given various degrees of freedom d f , the maximum likelihood estimator Σ is

obtained by iterative convergence calculation. Then, we calculate the log-likelihood


ll t (Σ, d f ). Let d f and Σ , which maximize ll t (Σ, d f ), be the parameters of the t

copula.
Clayton Copula
We can express the Clayton copula CClayton (u 1 , u 2 , . . . , u m ; πc ) with the following
equation using a parameter πc (> 0):
( m )− π1
 c

CClayton (u 1 , u 2 , . . . , u m ; πc ) = u i−πc −m+1 . (2.63)

i=1

The density function cClayton (u 1 , u 2 , . . . , u m ; πc ) is then

∂CClayton
m
(u 1 , u 2 , . . . , u m ; πc )
cClayton (u 1 , u 2 , . . . , u m ; πc ) =
∂u 1 ∂u 2 . . . ∂u m
(m−1 )( )( )− π1 −m
 
m m c

= (1 + i πc ) u i−πc −1 u i−πc − m + 1 . (2.64)

i=1 i=1 i=1
36 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Thus, using the N observations of each stochastic variable, the log-likelihood

ll Clayton is


N
∂CClayton
m
(u 1 , u 2 , . . . , u m ; πc )
llClayton (πc ) = log
j=1
∂u 1 ∂u 2 . . . ∂u m


m−1
=N log(1 + i πc )
i=1
⎛ (m ) ⎞
 j
N
⎜ (1 + πc ) log ui ⎟
− ⎜ ( ) ( ( i=1
) )⎟ (2.65)
⎝ m
j
−πc ⎠
j=1 + + m log −m+1
1
πc
ui
i=1

We may calculate the πc that maximizes the log-likelihood ll Clayton numerically.

Gumbel Copula
( )
We can express the Gumbel copula C Gumbel u 1 , u 2 , . . . , u m ; πg with the following
equation using a parameter πg (> 1):
⎛ ⎛ ⎞ π1 ⎞

g
( ) ⎜
m
( )π ⎟
C Gumbel u 1 , u 2 , . . . , u m ; πg = ex p ⎝−⎝ −logu j g ⎠ ⎠. (2.66)
j=1

( )
The density function cGumbel u 1 , u 2 , . . . , u m ; πg is
( )
( ) ∂C Gumbel
m
u 1 , u 2 , . . . , u m ; πg
cGumbel u 1 , u 2 , . . . , u m ; πg = . (2.67)
∂u 1 ∂u 2 . . . ∂u m

Thus, using the N observations of each stochastic variable, the log-likelihood

ll Gumbel is
( )
( ) N
∂C Gumbel
m
u 1 , u 2 , . . . , u m ; πg
ll Gumbel πg = log . (2.68)
j=1
∂u 1 ∂u 2 . . . ∂u m

We can numerically calculate the πg that maximizes the log-likelihood ll Gumbel .

For simplicity of notation, we define the following equations:

vi = − log u i (i = 1, 2, . . . , m)
( m )1
j
( j )πg πg
ψi = vi (2.69)
i=1
2.5 Risk Measurement in Statistical Arbitrage 37

Then, the log-likelihoods m ll Gumbel of the m(= 2, 3, 4, 5) variate Gumbel copula

can be described by Eqs. (2.70), (2.71), (2.72), and (2.73).
( )
( )  N
∂C Gumbel
2
u 1 , u 2 ; πg
ll Gumbel πg = log
2

j=1
∂u 1 ∂u 2
⎛ (( )πg ( )πg ) π1 ⎞
j j
− v1 + v2
g

⎜ ( ( )πg −1 )⎟
N ⎜ (( )πg ( )πg ) π1 −2 ⎟
⎜ j j
v1 v2 ⎟
= ⎜ + log v1
j
+ v2
j g
⎟
⎜ j j
u1 u2 ⎟
j=1 ⎜ ( (( )πg ( )πg ) π1 ) ⎟
⎝ ⎠
j j
+ log πg − 1 + v1 + v2
g

⎛ j
⎞
( −ψ2 )
( )
N ⎜ ( )1−2πg ⎟
π −1

g
⎜ ⎟
j j
v1 v2 j
= ⎜ + log ψ2 ⎟ (2.70)
⎜ j j
u1 u2 ⎟
j=1 ⎝ ( ) ⎠
j
+ log πg − 1 + ψ2
( )
( ) N
∂C Gumbel
3
u 1 , u 2 , u 3 ; πg
3
ll Gumbel πg = log
j=1
∂u 1 ∂u 2 ∂u 3
⎛ j ⎞
(( −ψ3 )
)πg −1
⎜ j j j
v1 v2 v3 ( )1−3πg ⎟
 N ⎜ + log ψ3
j ⎟
⎜ ⎟
⎜ ⎟
j j j
u1 u2 u3
= ⎜ ( ( )( ) )⎟ (2.71)
⎜ 2πg − 1 πg − 1 ⎟
j=1
⎝ ⎠
+ log ( ) j ( j )2
+3 πg − 1 ψ3 + ψ3
( )
( ) N
∂C Gumbel
4
u 1 , u 2 , u 3 , u 4 ; πg
4
ll Gumbel πg = log
j=1
∂u 1 ∂u 2 ∂u 3 ∂u 4
⎛ j
⎞
( −ψ4 )
⎜ ( )πg −1
( )1−4πg ⎟
⎜ j j j j
v1 v2 v3 v4 ⎟
N ⎜
⎜ + log ψ4
j
⎟

j j j j
u1 u2 u3 u4 ⎟
⎜ ⎛( )( )( )⎞⎟
= ⎜ ⎟ (2.72)
⎜ 3πg − 1 2πg − 1 πg − 1 ⎟
j=1 ⎜ ( )( ) j ⎟
⎜ + log⎜
⎝ + 11π − 7 π − 1 ψ 4) ⎠ ⎟
⎟
⎝ (
g
(
) j )g
2 ( 3 ⎠
j
+6 πg − 1 ψ4 + ψ4
( )
( ) N
∂C Gumbel
5
u 1 , u 2 , u 3 , u 4 , u 5 ; πg
5
ll Gumbel πg = log
j=1
∂u 1 ∂u 2 ∂u 3 ∂u 4 ∂u 5
38 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

⎛ j ⎞
−ψ5
⎜ ⎛( )πg −1 ⎞ ⎟
⎜ v
j j j j j
v v v v ( ) ⎟
⎜ ⎟
⎜ + log⎜ ⎟
1−5π
⎝
1 2 3 4 5
ψ5
j g
⎠ ⎟
⎜ j j j j j ⎟
N ⎜
⎜ u u u u u ⎟
 1 2 3 4 5 ⎟
⎜ ⎛( )( )( )( ) ⎞⎟
= ⎜ ⎟ (2.73)
⎜ 4πg − 1 3πg − 1 2πg − 1 πg − 1 ⎟
j=1 ⎜ ⎜ ( )( )( ) j ⎟⎟
⎜ ⎜ +5 5πg − 3 2πg − 1 π(g − )1 ψ5 ⎟⎟
⎜ ⎟⎟
⎜ + log⎜
⎜ ( )( ) j 2 ⎟⎟
⎜ ⎜ +5 7πg − 5 πg − 1 ψ5 ⎟⎟
⎝ ⎝ ⎠⎠
( )( j )3 ( j )4
+10 πg − 1 ψ5 + ψ5

Frank Copula
( )
We can express the Frank copula C Frank u 1 , u 2 , . . . , u m ; π f with the following
equation using parameter π f (> 0):
( m ( −π f u i ))
( ) 1 i=1 e −1
C Frank u 1 , u 2 , . . . , u m ; π f = − log 1 + ( )m−1 . (2.74)
πf e−π f − 1
( )
The density function c Frank u 1 , u 2 , . . . , u m ; π f is
( )
( ) ∂C Frank
m
u1, u2, . . . , um ; π f
c Frank u 1 , u 2 , . . . , u m ; π f = . (2.75)
∂u 1 ∂u 2 . . . ∂u m

Thus, using the N observations of each stochastic variable, the log-likelihood

ll Frank is
( )
( ) N
∂C Frank
m
u1, u2, . . . , um ; π f
ll Frank π f = log . (2.76)
j=1
∂u 1 ∂u 2 . . . ∂u m

We can numerically calculate the π f that maximizes the log-likelihood ll Frank .

For simplicity of notation, we define the following equation:

wi = e−δu i − 1(i = 1, 2, . . . , m). (2.77)

Then, the log-likelihoods m ll Frank of the m(= 2, 3, 4, 5) variate Frank copula can
be described by Eqs. (2.78), (2.79), (2.80), and (2.81), respectively.
( )
( ) N
∂C Frank
2
u1, u2; π f
2
ll Frank π f = log
j=1
∂u 1 ∂u 2
2.5 Risk Measurement in Statistical Arbitrage 39
( )( )( )
−δ w1 + 1 w2 + 1 e−δ − 1
j j

N
= log ( )2 (2.78)
j j
j=1 e−δ − 1 + w1 w2
( )
( )  N
∂C Frank
3
u1, u2, u3; π f
3
ll Frank π f = log
j=1
∂u 1 ∂u 2 ∂u 3
⎛ ( )( )( )( )2 ⎞
j j j −δ
 ⎜
N −δ 2
w + 1 w + 1 w + 1 e − 1
1 2 3
⎟
= log⎝ ( e−δ −1) −w1 w2 w3
2 j j j
⎠ (2.79)
× ( )3
j=1 (e −1) +w1 w2 w3
−δ 2 j j j

( )
( )  N
∂C Frank
4
u1, u2, u3, u4; π f
4
ll Frank π f = log
j=1
∂u 1 ∂u 2 ∂u 3 ∂u 4
⎛ ( )( )( )( )( )3 ⎞
−δ 3 w1 + 1 w2 + 1 w3 + 1 w4 + 1 e−δ − 1
j j j j
⎜ ⎛( )6 ( )3 j j j j ⎞ ⎟
N ⎜ −δ
− 1 − 4 e−δ − 1 w1 w2 w3 w4 ⎟ ⎟
⎜ ⎜ e
( )2 ( )2 ( )2 ( )2 ⎟
= log⎜ ⎜
⎝
⎟
⎠ ⎟
⎜ + w
j
w
j
w
j
w
j ⎟
j=1 ⎝ 1 2 3 4 ⎠
× ( )4
(e −1) +w1 w2 w3 w4
−δ 3 j j j j

(2.80)
( )
( )  N
∂C Frank
5
u1, u2, u3, u4, u5; π f
5
ll Frank π f = log
j=1
∂u 1 ∂u 2 ∂u 3 ∂u 4 ∂u 5
⎛ ( )( )( )( )( ) ⎞
j j j j j
−δ 4 w1 + 1 w2 + 1 w3 + 1 w4 + 1 w5 + 1
⎜ ( )4 ⎟
⎜ ⎟
⎜ × e−δ − 1 ⎟
⎜ ⎛ ( ) ( ) ⎞ ⎟
⎜ 12 8 j j j j j ⎟ (2.81)
N ⎜ e−δ − 1 − 11 e−δ − 1 w1 w2 w3 w4 w5 ⎟
⎜ ⎜ ( −δ ( ) (
)8 j 2 j 2 j 2 j 2 j 2 ⎟ ) ( ) ( ) ( ) ⎟
= log⎜ ⎜ ⎟ ⎟
⎜ ⎜ +11 e − 1 w1 w2 w3 w4 w5 ⎟ ⎟
j=1 ⎜ ⎝ ( ) ( ) ( ) ( ) ( ) ⎠ ⎟
⎜ 3 3 3 3 3 ⎟
⎜ − w
j
w
j
w
j
w
j
w
j ⎟
⎜ 1 2 3 4 5 ⎟
⎝× (( )4 ) 5
⎠
j j j j j
e−δ − 1 + w1 w2 w3 w4 w5

After estimating all the parameters using the above, we should select the copula to
measure the risk indicators. Although it is desirable to adopt a copula that minimizes
the difference between the joint distribution obtained from historical data and the
joint distribution estimated by each copula, it is rational to select the copula based on
the information criterion, considering that the number of parameters and the degrees
of freedom differ for each type of copula. As in Breymann et al. [3], we select the
copula based on the Akaike information criterion (AIC).
We now generate the uniform random numbers U1 , U2 , . . . , Um , which follow
the estimated copula C(u 1 , u 2 , . . . , u m ). Then, we can find the VaR and expected
40 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

−1
shortfall from R̃1 = F1−1 (U1 ), R̃2 = F2−1 (U2 ), . . . , R̃m = F m (Um ), where
F1−1 , F2−1 , . . . , Fm−1 are the inverse functions of the marginal distribution functions.
These allow us to assess the portfolio risk. The generation of random numbers based
on each copula type is as follows.
Gaussian Copula
We obtain the lower triangular matrix using the Cholesky decomposition of the esti-


mated Σ . By multiplying the matrix by a vector T (s1 , s2 , . . . , sm ) whose elements are

random numbers that follow an independent standard normal distribution, we may
generate random numbers S1 , S2 , . . . , Sm that follow a multivariate normal distribu-


tion with a correlation matrix Σ . Using the standard normal distribution function ,
we can generate random numbers U1 = (S1 ), U2 = (S2 ), . . . , Um = (Sm ) that
follow the Gaussian copula.
t Copula
We obtain the lower triangular matrix using the Cholesky decomposition of the esti-


mated Σ . By multiplying the matrix by a vector T (s1 , s2 , . . . , sm ) whose elements are

random numbers that follow an independent standard normal distribution, we may
generate random numbers S1 , S2 , . . . , Sm that follow a multivariate normal distribu-


tion with a correlation matrix Σ . Generating random numbers Z 1 , Z 2 , . . . , Z m that

√ the χ distri-
2
follow an independent standard normal distribution, which follows
bution with degree of freedom
√ √ d f , we may calculate T1 = d f S1 /Z 1 , T2 =
d f S2 /Z 2 , . . . , Tm = d f Sm /Z m , which itself follows a multivariate t distri-


bution with degree of freedom d f and a correlation matrix Σ . Using the t distri-
bution function td f with degree of freedom d f , we can generate random numbers
U1 = td f (T1 ), U2 = td f (T2 ), . . . , Um = td f (Tm ) that follow the t copula.
Clayton Copula
Generating uniform random numbers I1 , I2 , . . . , Im (0  I1 , I2 , . . . , Im  1) and
random numbers γ that follow the standard gamma distribution G(1/πc ), we can
calculate random numbers U1 , U2 , . . . , Um that follow the Clayton copula, as follows:

U1 = (1 − (1/γ ) log I1 )−1/πc , U2 = (1 − (1/γ ) log I2 )−1/πc ,

· · · , Um = (1 − (1/γ ) log Im )−1/πc . (2.82)

Gumbel Copula
Generating uniform random numbers V (0  V  cir cle ratio π ) and random
numbers W that follow the standard exponential distribution, we calculate the
following equation:
( (( ) ) )πg −1
sin πg − 1 V /πg
≡ θ. (2.83)
W
2.5 Risk Measurement in Statistical Arbitrage 41

Using uniform random numbers I1 , I2 , . . . , Im (0  I1 , I2 , . . . , Im  1), we

can calculate random numbers U1 , U2 , . . . , Um that follow the Gumbel copula, as
follows:
( )
U1 = exp −(−(1/θ ) log I1 )1/πg ,
U2 = (−(1/θ ) log I2 )1/πg , . . . , Um = (−(1/θ ) log Im )1/πg . (2.84)

Frank Copula
We use only a bivariate Frank copula in this book, and thus omit the generation of
random numbers that follow trivariates or more Frank copulas. Generating uniform
random numbers I1 , I2 (0  I1 , I2  1), we can set random numbers U1 and U2 that
follow the Frank copula, as follows:

U1 = I 1
( ( ) )
1 I2 1 − e−π f
U2 = − log ( ) +1 . (2.85)
πf I2 e−π f U1 − 1 − e−π f U1

While countless studies apply copulas to financial markets, other studies analyze
energy markets using copulas. Ghorbel and Trabelsi [12] adopt the concept of copulas
to study the dependence structure between the crude oil, natural gas, and heating
oil markets. Lu et al. [23] construct the conditional joint distribution using copula
functions with GARCH-type models to calculate the VaR of a portfolio composed
of crude oil futures and natural gas futures. Zhang [36] examines the impact of
changes in crude oil prices on bunker prices and tanker freight rates using copula
models. Soliman and Nasir [34] employ a time-varying copula connection function
to analyze the risk dependency relationship between the energy and carbon markets
in Europe. Fernandes et al. [9] propose a stochastic-copula approach that considers
seasonality, mean reversion, and jumps to examine the dependence between European
electricity and natural gas day-ahead prices. Ly et al. [24] reveal the dependencies
and co-movements between European electricity markets based on copula models.

2.5.3 Copula Estimation and Risk Measurement

In this section, we aim to measure the risk of the spread between the Henry Hub
futures prices and the PJM futures prices because the statistical arbitrage simulated
in Sect. 2.4 holds a portfolio consisting of Henry Hub futures short positions and
PJM futures long positions.
First, we calculate each return series of the Henry Hub futures prices and PJM
futures prices. We also estimate the mean and standard deviation of each return
42 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Table 2.10 Parameters of

Henry Hub futures PJM futures
each marginal distribution
function Mean 3.52 × 10–4 6.52 × 10–4
Standard deviation 3.16 × 10–2 4.64 × 10–2

series. Table 2.10 provides these values, which are the parameters of each marginal
distribution function.
We then estimate each copula and generate random numbers to determine the VaR
and expected shortfall according to the procedure in Sect. 2.5.2. However, we should
note the following. For long positions, the smaller the return, the larger the loss.
Therefore, we should measure the risk in the range of the left tail of the distribution.
However, for short positions, the larger the return, the larger the loss; therefore, we
should measure the risk in the range of the right tail of the distribution.
Figure 2.15 presents a simplified diagram of the procedure in this section.
Figure 2.16 plots the relationship between the degree of freedom and the log- 

likelihood in the estimated t copula. The t copula parameter is the Σ estimated

with the degree of freedom 3 that maximizes the log-likelihood. We generate the
distribution of the daily rate of return by simulating each estimated copula 10,000
times to calculate the VaR and expected shortfall. Table 2.11 provides the estimated
copulas and the VaRs and expected shortfalls calculated based on these copulas.

Fig. 2.15 Procedure to measure portfolio risk using various copulas

2.5 Risk Measurement in Statistical Arbitrage 43

Fig. 2.16 Degree of freedom and log-likelihood in estimating the t copula

The VaR and expected shortfall measured based on the Frank copula, which has
the smallest AIC, can be regarded as the risk of statistical arbitrage. Accordingly, the
possibility of losing more than 4.79% over a day is less than 1%, and the expected loss
is 6.53% if the loss is more than 4.79%. These values aid in risk management, such
as portfolio restructuring, position reductions, clearance before maturity, increasing
the allowance, and capital expansion.
Figures 2.17, 2.18, 2.19, 2.20, and 2.21 provide scatter plots of the random
numbers u P J M and u H H that follow the Gaussian, t, Clayton, Gumbel, and Frank
copulas, which we estimate in this section
In adapting to the risk measurement of the portfolio consisting of the Henry Hub
futures short positions and the PJM futures long positions, the generation of random
numbers is reversed. In other words, u P J M → 0, u H H → 1 expresses the left tail of
each marginal distribution function, and u P J M → 1, u H H → 0 expresses the right
tail of each marginal distribution function.
We can observe a difference in the dependency structure between these variables
for each type of copula. Although the dots are relatively concentrated in backslash-
shaped copulas, these variations are larger in the Gaussian and t copulas. From
another point of view, the dots were also generated near the lower left and upper
right. This indicates that when one price fluctuates significantly, the other price may
fluctuate in the opposite direction. In the Clayton copula, we can see that mutual
dependence is extremely strong when the PJM return is positive and the Henry Hub
return is negative. In the Gumbel copula, the interconnection is very strong when
the PJM return is negative and the Henry Hub return is positive. The Frank copula
44 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Table 2.11 Risk

Gaussian AIC –7.687
measurement using the 

estimated copulas Σ 1.000 0.493

0.493 1.000
Value-at-riska 5.46%
Expected shortfalla 6.64%
t AIC – 6.279
Degree of freedom 3


Σ 1.000 0.583
0.583 1.000
Value-at-riska 5.56%
Expected shortfalla 6.83%
Clayton AIC – 10.372
πc 3.163
Value-at-riska 3.24%
Expected shortfalla 4.05%
Gumbel AIC – 10.230
πg 2.689
Value-at-riska 3.37%
Expected shortfalla 4.40%
Frank AIC –10.651
πf 9.342
Value-at-riska 4.79%
Expected shortfalla 6.53%
a
Note indicates the value at the 99% confidence level
Bold indicates the minimum AIC and the risk measures at that time

has no distinctive features in the strength of the interdependence. If we know the

dependency of each return before estimating some types of copulas, it is rational to
adopt the copula that fits the facts instead of selecting a copula type by the AIC.

2.6 Concluding Remarks

Many power generation companies purchase primary energy as fuel and produce
and sell electricity as secondary energy. Energy is generally a good that cannot be
differentiated except for price. Moreover, both fuel and power are traded in highly
liquid, efficient markets. Therefore, we do not interpret power generation companies
as electric power manufacturers, but reinterpret them as business operators that add
value to energy economics by converting fuel currency to power currency. Like tradi-
tional financial firms, they make efficient use of energy derivatives that correspond
to their physical position.
2.6 Concluding Remarks 45

Fig. 2.17 Scatter plots of

random numbers following
the bivariate Gaussian copula

Fig. 2.18 Scatter plots of

random numbers following
the bivariate t copula
46 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

Fig. 2.19 Scatter plots of

random numbers following
the bivariate Clayton copula

Fig. 2.20 Scatter plots of

random numbers following
the bivariate Gumbel copula
2.6 Concluding Remarks 47

Fig. 2.21 Scatter plots of

random numbers following
the bivariate Frank copula

This chapter proposes arbitrage between the spot and futures of the price difference
between fuel and electricity, and the statistical arbitrage between fuel futures and
power futures. In the simulation and risk measurement, we adopt the most primitive
model that buys the Henry Hub, one of the most representative natural gas price
indicators, and sells the PJM, one of the most representative wholesale electricity
indicators.
First, we clarify the characteristics of futures and spot prices by examining their
descriptive statistics. The mean and median for each future were higher than those
for each spot. Because supply and demand are not so tight during the period, their
medians are higher than their respective means in the Henry Hub futures and spot
markets. Many outliers are present in the left tail. By contrast, the medians are lower
than their respective means in the PJM futures and spot markets. Several outliers may
be present in the right tail. Additionally, the ranges of both the Henry Hub futures
and PJM futures are narrower than the ranges of each spot market. Moreover, the
standard deviation of each spot is larger than that of each future. This is because it
is technically difficult or too expensive to store natural gas and electricity. For all
four prices, the difference between the maximum and mean is much larger than the
difference between the minimum and mean. Moreover, all price series have positive
skewness. These distributions have a long right tail. Each variable’s kurtosis is larger
than 3. All distributions have long fat-tails and a sharp peak.
The time plots of these prices indicate that each spot fluctuates in the same manner
as its respective future. However, the spots spike frequently because a simultaneous,
equal amount of the supply and demand is needed. Moreover, we can see that the
Henry Hub and PJM are synchronized, regardless of their spots and futures.
48 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

The Jarque-Berra test rejects the hypothesis that each price series has kurtosis
and skewness following a normal distribution. According to the ADF test results, all
price series have a unit root and the first difference series does not. The Johansen
tests between the Henry Hub and PJM both in the futures and spot markets accept the
cointegration hypothesis. Therefore, we expect various trading strategies to utilize
their long-term equilibrium relationships.
In this chapter, we propose two trading strategies. The first is arbitrage between
spot and futures spreads, selling the PJM futures and buying the Henry Hub futures
if the PJM futures price is higher than the power generation cost estimated from the
Henry Hub futures price. The other is statistical arbitrage between the Henry Hub
futures and PJM futures, buying the PJM futures and selling the Henry Hub futures
if the PJM price is estimated by substituting the Henry Hub futures price into the
long-term equilibrium equation between the Henry Hub futures price and the PJM
futures price is higher than the PJM futures price.
To demonstrate their effectiveness, we simulate both trading strategies using
historical data, assuming that a power generation company is inferior to the spot
market. According to the simulation, although spot trading alone results in an
expected loss of its market subordinated level, adopting both trading strategies can
make a large profit instead of only offsetting the expected loss.
Moreover, we measure the risk of a portfolio consisting of Henry Hub future short
positions and PJM future long positions held in statistical arbitrage. We estimated
five types of copulas that represent the mutual dependency of these two return series
after estimating their marginal distributions. Using Monte Carlo simulations with
the Frank copula, which is the most suitable for historical data, we calculate the VaR
and expected shortfall.
This chapter provides knowledge on the interpretation and handling of time-series
data, proposes trading strategies with confirmation of their effectiveness, and teaches
how to measure portfolio risk. However, the trading simulation in this section ignores
the market impact. Usually, a buy order raises the market price and a sell order
lowers the market price. In addition, the impact of energy companies on commodity
exchanges is often significant. Therefore, they may not be able to execute trades as
expected. It is worth re-examining trading strategies by simulating artificial markets
or using a market impact model.
More importantly, there is no limit to the development of trading algorithms
aimed at maximizing profits and minimizing risk. The following are five examples
of development directions:
• Add futures with different contract months to our portfolio in the spot-future
arbitrage.
• Adjust the leverage of statistical arbitrage considering the size of equity capital.
• Testing the cointegration on a moving window sample, which is similar to esti-
mating the cointegrating vector and incorporating the test statistics into the
leverage in statistical arbitrage.
• Estimate the volatility of the long-term equilibrium and utilize it for statistical
arbitrage and/or clearance conditions.
References 49

Fig. 2.22 Simplified diagram of an electric power business model

• Change the update frequency of the cointegrating vector estimation in the

statistical arbitrage strategy.
Moreover, establishing a more realistic business model is challenging. Figure 2.22
presents a simple and realistic model, but it is still not simple in practice. Actual util-
ities hold a variety of fuel-type power plants with various power generation efficien-
cies and trade fuel and power in spot markets and futures with varying maturity and
various indicators, including special bilateral contracts. Therefore, we can expect
many opportunities for market distortion. In other words, there should be many
opportunities to make money at low risk. To maximize profits and minimize risk, we
must aim to build rational models and develop algorithms.

References

1. Alexakis, C. (2010). Long-run relations among equity indices under different market condi-
tions: Implications on the implementation of statistical arbitrage strategies. Journal of
International Financial Markets, Institutions and Money, 20(4), 389–403.
2. Baviera, R., & Baldi, T. S. (2019). Stop-loss and leverage in optimal statistical arbitrage with
an application to energy market. Energy Economics, 79, 130–143.
3. Breymann, W., Dias, A., & Embrechts, P. (2003). Dependence structures for multivariate high-
frequency data in finance. Quantitative Finance, 3(1), 1–14.
4. Clayton, D. G. (1978). A model for association in bivariate life tables and its application in
epidemiological studies of familial tendency in chronic disease incidence. Biometrika, 65(1),
141–151.
50 2 Arbitrage Trading in Energy Markets and Measuring Its Risk

5. Dickey, A. D., & Fuller, W. A. (1979). Distribution of the estimates for autoregressive time
series with a unit root. Journal of the American Statistical Society, 74, 427–431.
6. Embrechts, P., McNeil, A. J., & Straumann, D. (2002). Correlation and dependence in risk
management: Properties and pitfalls. In M. A. H. Dempster (Ed.), Value at risk and beyond
(pp. 176–223). Cambridge University Press.
7. Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error correction: Representation,
estimation, and testing. Econometrica, 55(2), 251–276.
8. Fang, H. B., Fang, K. T., & Kotz, S. (2002). The meta-elliptical distributions with given
marginals. Journal of Multivariate Analysis, 82(1), 1–16.
9. Fernandes, M. C., Diasa, J. C., & Nunes, J. P. V. (2021). Modeling energy prices under energy
transition: A novel stochastic-copula approach. Economic Modelling, 105, 105671.
10. Focardi, S. M., Fabozzi, F. J., & Mitov, I. K. (2016). A new approach to statistical arbi-
trage: Strategies based on dynamic factor models of prices and their performance. Journal of
Banking & Finance, 65, 134–155.
11. Frank, M. J. (1979). On the simultaneous associativity of F(x, y) and x+y−F(x, y). Aequationes
Mathematicae, 19, 194–226.
12. Ghorbel, A., & Trabelsi, A. (2014). Energy portfolio risk management using time-varying
extreme value copula methods. Economic Modelling, 38, 470–485.
13. Granger, C. W., & Newbold, P. (1974). Spurious regressions in econometrics. Journal of
Econometrics, 14, 114–120.
14. Gumbel, E. J. (1960). Bivariate exponential distributions. Journal of the American Statistical
Association, 55(292), 698–707.
15. Hain, M., Hess, J., & Uhrig-Homburg, M. (2018) Relative value arbitrage in European
commodity markets. Energy Economics, 69, 140–154.
16. Jarque, C. M., & Bera, A. K. (1987). A test for normality of observations and regression
residuals. International Statistical Review, 55(2), 163–172.
17. Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian
vector autoregressive models. Econometrica, 59(6), 1551–1580.
18. Johansen, S., & Juselius, K. (1990). Maximum likelihood estimation and inferences on cointe-
gration with application to the demand for money. Oxford Bulletin of Economics and Statistics,
52(2), 169–210.
19. Keilbar, G., & Zhang, Y. (2021). On cointegration and cryptocurrency dynamics. Digital
Finance, 3, 1–23.
20. Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis
of stationary against the alternative of a unit root. Journal of Econometrics, 54, 159–178.
21. Lee, L. F. (1983). Generalized econometric models with selectivity. Econometrica, 51(2),
507–512.
22. Liu, G. D., & Su, C. W. (2019). The dynamic causality between gold and silver prices in China
market: A rolling window bootstrap approach. Finance Research Letters, 28, 101–106.
23. Lu, X. F., Lai, K. K., & Liang, L. (2014). Portfolio value-at-risk estimation in energy futures
markets with time-varying copula-GARCH model. Annals of Operations Research, 219(1),
333–357.
24. Ly, S., Sriboonchitta, S., Tang, J., & Wong, W. K. (2022). Exploring dependence structures
among European electricity markets: Static and dynamic copula-GARCH and dynamic state-
space approaches. Energy Reports, 8, 3827–3846.
25. Mayordomo, S., Peña, J. I., & Romo, J. (2014). Testing for statistical arbitrage in credit
derivatives markets. Journal of Empirical Finance, 26, 59–75.
26. Nakajima, T. (2019). Expectations for statistical arbitrage in energy futures markets. Journal
of Risk and Financial Management, 12(1), 14.
27. Nelsen, R. B. (2006). An introduction to copulas. Springer Science & Business Media.
28. Ophem, H. V. (1999). A general method to estimate correlated discrete random variables.
Econometric Theory, 15(2), 228–237.
29. Phillips, P. C. B. (1986). Understanding spurious regressions in econometrics. Journal of
Econometrics, 33(3), 311–340.
References 51

30. Phillips, P. C. B., & Perron, P. (1988). Testing for a unit root in time series regression.
Biometrika, 75(2), 335–346.
31. Phillips, P. C. B., Wu, Y., & Yu, J. (2011). Explosive behavior in the 1990s NASDAQ: When
did exuberance escalate asset values? International Economic Review, 52(1), 201–226.
32. Sánchez-Granero, M. A., Balladares, K. A., Ramos-Requena, J. P., & Trinidad-Segovia, J. E.
(2020). Testing the efficient market hypothesis in Latin American stock markets. Physica A:
Statistical Mechanics and its Applications, 540, 123082.
33. Sklar, A. (1959). Fonctions de Répartition à n Dimensions et Leurs Marges. Publications de
l’Institut Statistique de l’Université de Paris, 8, 229–231.
34. Soliman, A. M., & Nasir, M. A. (2019). Association between the energy and emission prices:
An analysis of EU emission trading system. Resources Policy, 61, 369–374.
35. Stübinger, J., & Schneider, L. (2019). Statistical arbitrage with mean-reverting overnight price
gaps on high-frequency data of the S&P 500. Journal of Risk and Financial Management,
12(2), 51.
36. Zhang, Y. (2018). Investigating dependencies among oil price and tanker market variables by
copula-based multivariate models. Energy, 161, 435–446.
Chapter 3
Fuel Market Connectedness and Fuel
Portfolio Risk

3.1 Introduction

In energy importing countries, many trading companies, oil refineries, electric power
companies, and city gas companies procure crude oil and natural gas globally. They
must construct and adjust their energy portfolios while considering the spillover
effect between each price index because they adopt various price indexes and special
pricing formulas. Furthermore, they must measure and manage portfolio risk.
Even for portfolios with many components, its market integration neutralizes the
diversified procurement. If the total spillover effect of returns is large, then the port-
folio may not bring fuel procurement to stable price levels. The strong connectivity
of volatility may insulate a portfolio against any risks incurred in one component.
Even under market integration, if the rate of spillover between markets is slow, then
monitoring each market may allow for some action.
As these facts illustrate, it is extremely meaningful to analyze the relationships
between portfolio component candidates from multiple perspectives. Specifically,
measuring the spillover effect of volatility and returns, the connectedness of an entire
portfolio as well as between its components, and the spectrum of the spillover effect
can give us a rough idea of the portfolio’s potential. It is reasonable to narrow down
a few probable portfolio candidates from a large number of candidates by measuring
the spillover effect with the lowest calculation cost possible; while measuring risk
using Monte Carlo simulation is highly accurate, it is very computationally intensive.
A considerable number of studies discuss whether oil markets are integrated.
Gülen [6,7] accepts the global crude oil market integration hypothesis using cointe-
gration analysis. Kleit [11] adopts an approach based on arbitrage theory and demon-
strates the convergence of the global crude oil market. Reboredo [19] accepts the
hypothesis that the global crude oil market is “one great pool” by estimating several
copula models with different conditional dependence structures and time-varying
dependence parameters. Ji and Fan [9] confirm market integration by constructing a
minimal spanning tree for the global crude oil market. Using a threshold unit-root

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 53
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_3
54 3 Fuel Market Connectedness and Fuel Portfolio Risk

approach, Fattouh [5] argues that oil markets are not necessarily integrated in every
period, despite the generally integrated global crude oil market. Using a threshold
cointegration analysis, Hammoudeh et al. [8] find an asymmetric adjustment process,
despite the long-run equilibrium relationships between different crude oil bench-
marks. Jia et al. [10] analyze the evolutionary features of global crude oil market
integration using a proposed novel wavelet-based complex network. Zhang [23]
examines the integration between the Chinese and North American crude oil markets
by testing for return and volatility spillovers. Nakajima [12] examines the risk trans-
mission between crude oil and petroleum product prices in Japan’s oil futures market
using the realized volatility and exponential generalized autoregressive conditional
heteroscedasticity (EGARCH) model.
Several other studies analyze information efficiency in natural gas markets. Silver-
stovs et al. [22] and Nakajima and Toyoshima [13] examine the global natural gas
market integration between North America, Europe, and Japan. Chai et al. [3] study
the relationship between the Chinese and global natural gas markets. Olsen et al.
[16], Scarcioffolo and Etienne [20], and Ren et al. [18] investigate integration in
the North American natural gas market. Nick [15], Osička et al. [17], and Bastianin
et al. [2] discuss natural gas market integration in Europe. Shi et al. [21] analyze the
interrelationships of liquefied natural gas (LNG) markets in Asia.
This chapter examines portfolios of crude oil and natural gas as case studies. We
cover three representative price indices for each energy type in the North American,
European, and Asian markets. First, we provide an overview of each energy market
using descriptive statistics and time plots for price, returns, and volatility. Second,
we measure the total connectedness of the crude oil and natural gas markets, and
the spillover effect between each price index in each market. We then spectrally
analyzed the result. In this context, this chapter describes the vector autoregression
(VAR) model, vector moving average (VMA) model, connectedness index proposed
by Diebold and Yilmaz [4], spectral analysis approach proposed by Baruník and
Křehlík [1], and EGARCH model by Nelson [14]. Finally, we measure the risk of
each energy portfolio using the copulas introduced in Chap. 2.
The remainder of this chapter is organized as follows. Section 3.2 describes the
data used in this chapter. Section 3.3 introduces and explains the methodologies.
Section 3.4 provides the results of the analyses and measurements. Section 3.5
discusses the overall concluding remarks and considerations.

3.2 Data

In this chapter, we analyze data representing the European, North American, and
Asian markets. For crude oil, we use Brent futures in Europe, West Texas Intermediate
(WTI) futures in North America, and Dubai-Oman futures in Asia. For natural gas,
we utilize the Title Transfer Facility (TTF) futures in Europe, Henry Hub (HH)
futures in North America, and Japan Korea Marker (JKM) futures in Asia. We take
advantage of the daily data from January 1, 2018 to December 31, 2019, though we
3.2 Data 55

extract only days when all three crude oil (natural gas) prices are available. The Brent
and WTI prices are expressed in USD per barrel. Dubai-Oman prices are given in
USD per kiloliter. All natural gas prices were in USD per mmbtu. We obtain all the
data from Bloomberg.

3.2.1 Crude Oil

First, we must interpret the descriptive statistics to review the data. Table 3.1 reports
the summary statistics of the Brent, WTI, and Dubai-Oman futures prices. Since the
mean values for these price series are larger than their respective medians and all
price series have positive skewness, we can assume that these distributions have a
long right tail, which includes many outliers. The mean ratio of the ranges is 50%
for Brent, 52% for WTI, and 45% for Dubai-Oman, with standard deviations of
9.9%, 11%, and 9.6%, respectively. Therefore, while the WTI prices are the most
volatile and the Dubai-Oman prices are the most stable, we can conclude that these
three price series have almost the same level of dispersion. Because the kurtosis for
all variables is less than 3, these distributions have rounded peaks and short, thin
tails. The Jarque–Bera test rejects the normal distribution hypothesis for these three
variables.
Figure 3.1 shows time plots of these prices. These three prices seem to be
completely synchronized and fluctuating, as the descriptive statistics suggest. These
three markets can be integrated. Therefore, a portfolio comprising these commodi-
ties may be worthless. However, this is merely a prejudice that relies on intuition.
Mathematical verification is essential for understanding these facts.

Table 3.1 Descriptive

Brent WTI Dubai-Oman
statistics (crude oil prices)
Period From 1/4/2018 to 12/30/2019
Observations 470 470 470
Mean 68.01 61.06 417.88
Median 67.02 60.95 409.99
Maximum 86.29 76.41 532.16
Minimum 52.16 44.61 346.15
Standard deviation 6.75 6.59 40.15
Skewness 0.33 0.10 0.55
Kurtosis 2.32 2.20 2.48
Jarque–Bera 17.89 (0.00) 13.47 (0.00) 28.98 (0.00)
Note p-values are in parentheses
56 3 Fuel Market Connectedness and Fuel Portfolio Risk

Fig. 3.1 Crude oil prices

3.2.2 Natural Gas

As in Sect.3.2.1, we calculate the representative statistics to provide an overview of

the data. Table 3.2 lists the summary statistics of the TTF, HH, and JKM futures
prices. Because the TTF’s mean price is smaller than its median and its skewness
is 0.00, the distribution has many outliers in the left tail, while its tails have nearly
symmetrical lengths. For the HH and JKM prices, because the means are larger than
their medians and their skewness is positive, the distribution has many outliers in the
right tail, which is longer than the left tail. The ranges have a mean ratio of 112% for
the TTF, 99% for the HH, and 98% for the JKM, with standard deviations of 28, 17,
and 31%, respectively. Therefore, although the TTF prices fluctuate the most widely,
the TTF and JKM prices have the same level of dispersion. Because the kurtosis for
the TTF and JKM is less than 3, their distributions have a rounded peak and short,
thin tails. However, the HH price kurtosis is larger than 3, indicating that it has a
sharp peak and a thick, long-tailed distribution. The Jarque–Bera test did not support
a normal distribution for any variable. Unlike the distribution of the three crude oil
price indices, we assume that each index has a characteristic distribution.
Figure 3.2 shows the time plots of the natural gas prices. It can be observed
that the price of the TTF swings greatly, while the price of the HH hardly swings.
Although these three variables seem to be temporarily synchronized, they seem to
fluctuate in three ways. This is consistent with the interpretation of the descriptive
3.3 Methodology 57

Table 3.2 Descriptive

TTF HH JKM
statistics (natural gas prices)
Period From 1/2/2018 to 12/31/2019
Observations 504 504 504
Mean 21.29 2.80 7.68
Median 22.29 2.74 7.49
Maximum 34.41 4.84 11.81
Minimum 10.66 2.07 4.28
Standard deviation 5.97 0.48 2.42
Skewness 0.00 1.67 0.11
Kurtosis 1.92 6.78 1.56
Jarque–Bera 24.31 (0.00) 535.86 (0.00) 44.85 (0.00)

Fig. 3.2 Natural gas prices

statistics. Because these markets do not appear to be integrated, the potential value
of a portfolio consisting of these three natural gas indices might be expected.

3.3 Methodology

In this chapter, we employ the connectedness index proposed by Diebold and Yilmaz
[4] to examine the spillover effects between the three markets and their integration.
58 3 Fuel Market Connectedness and Fuel Portfolio Risk

Moreover, to capture the spillover speed between these markets, we apply the spectral
analysis approach proposed by Baruník and Křehlík [1] to the connectedness index.
We also measure the spillover effect of market risk for crude oil and natural gas by
analyzing these volatility series, which we estimate using Nelson’s [14] EGARCH
model. Finally, we measure the risk of the crude oil and natural gas portfolios using
four types of copulas (i.e., Gaussian, t, Clayton, and Gumbel). We calculate the value-
at-risk (VaR) and expected shortfall using the daily rate of return distribution of the
portfolio generated from 100,000 random numbers for each estimated copula. In this
section, we introduce the connectedness index (Diebold and Yilmaz [4]), spectral
analysis approach (Baruník and Křehlík [1]), and EGARCH model (Nelson [14]).
For the copulas, we refer the reader to Sect. 2.5.2.

3.3.1 Connectedness Index (Diebold and Yilmaz [4])

If x t is defined as a stochastic process of N variables, we can describe a general

VAR( p) model containing values up to the past time p as follows:


p
x t = Φ0 + Φ j x t− j + ε t , (3.1)
j=1

where Φ0 is an N -dimensional constant vector, Φ j is an N × N coefficient matrix, p

is the lag length, and ε t is an N -dimensional vector with the following characteristics:

E(ε t ) = O
 
V(ε t ) = E ε t T ε t = Σ
E(ε t ε t−s ) = O, if s > 0, (3.2)

where εt N is the independently and identically distributed sequence of dimensional

random vectors with zero mean and covariance matrix Σ . x t is a covariance stationary
process if an N-variable stochastic process satisfies the following three conditions:
Condition 1. The expected value E(x t ) does not depend on time t.
 
Condition 2. The variance V(x t ) = E (x t − E(x t ))T (x t − E(x t )) does not depend
on time t.
Condition 3. The autocovariance Cov(x t , x t−s ) does not depend on time t, but only
on the time difference s(> 0).
We can estimate the coefficient matrices of the VAR( p) model using ordinary least
squares (OLS) as a consistent estimator. It is theoretically correct to select lag order
p under the condition that the autocorrelation of the estimated residuals is rejected.
However, considering the calculation cost, it is more reasonable to select lag order
p based on the information criterion.
3.3 Methodology 59

Diebold and Yilmaz [4] assumes Φ0 = O. We set the lag length p based on the
Schwarz Bayesian information criterion (SBIC).


p
xt = Φ j x t− j + ε t . (3.3)
j=1

If x t is a covariance stationary vector, then we can represent Eq. (3.3) in the

following VMA model to examine the dynamic interdependencies between the vector
elements:
∞

xt = A j ε t− j , (3.4)
j=0


p
where Ai = Φ j At− j , A0 is the N × N identity matrix and A j = O, if j < 0.
j=1
Using the H -step-ahead forecast error variance decompositions, the spillover effect
from the l-th to the k-th variable up to H -step-ahead is
 H −1 T 2
1 h=0 ek Ah el
θkl =  H −1 T , (3.5)
σll h=0 ek Ah Σ T Ah ek

where σll is the standard deviation of the error term for the l-th equation, and ek and
el are the selection vectors whose k-th and l-th elements are 1, and the others are 0,
respectively. Each element of the variance decomposition matrix is normalized by
∼
the row sum N as pairwise connectedness θ kl :

∼ θkl
θ kl = . (3.6)
N
Total connectedness S is
N N ∼
k=1 l=1,k/=l θ kl
S= . (3.7)
N
The numerator of Eq. (3.7) is the sum of the spillover effects, excluding its effect
on itself. The total connectedness indicates the sum of the relative proportion of the
portfolio’s response to the shock of a variable. Furthermore, we can measure the
directional spillover effects Sk· received by the k-th variable from all other variables
as
N ∼
k=1,k/=l θ kl
Sk· = . (3.8)
N
60 3 Fuel Market Connectedness and Fuel Portfolio Risk

Similarly, we can measure the directional spillover effects S·l transmitted by the
l-th variable to all other variables as
N ∼
l=1,k/=l θ kl
S·l = . (3.9)
N

3.3.2 Spectral Decomposition (Baruník and Křehlík [1])

Adopting the approach introduced by Baruník and Křehlík [1], we spectrally decom-
pose the connectedness indices by Diebold and Yilmaz [4] to grasp when the spillover
effect factors occur.
We define the map from the time-domain function f (t) to the frequency domain
function F(ω) by the following equation, which is called the discrete Fourier
transform (DFT):

N −1

F(ω) = e−iωt f (t), for 0  ω < 2π. (3.10)
t=0

The DFT of Eq. (3.5) is

   2
 H −1
e−i ωh Ah Σ
h=0
1 jk
ϕ(ω) jk =     . (3.11)
σkk H −1 −iωh A Σ H −1 iωh A '
h=0 e h h=0 e h
jj

This equation indicates the spillover effect from the k-th variable to the j-th vari-
able up to H -steps-ahead, expressed by the angular frequency ω. The ω component
ratio to all the frequency components concerning the spillover to the j-th variable,
which is defined as the weighting function B j (ω), is calculated as
   
H −1 −iωh H −1 iωh '
h=0 e A h Σ h=0 e A h
jj
B j (ω) =     . (3.12)
2π H −1 −i λh H −1 iλh '
1
2π 0 h=0 e A h Σ h=0 e A h dλ
jj

We express the spillover effect from the k-th variable to the j-th variable in all
the bands as
2π
1
(θ∞ ) jk = B j (ω)ϕ(ω) jk dω. (3.13)
2π
0
3.3 Methodology 61

We calculate this spillover effect in band D is as

1
(θ D ) jk = B j (ω)ϕ(ω) jk dω. (3.14)
2π
D

We convert (θ D ) jk to the relative contribution as follows:

∼  (θ D ) jk
θD = N . (3.15)
j=1 (θ∞ ) jk
jk

The connectedness in band D is

N N ∼ 
j=1 θD
k=1, j/=k
jk
CD = N N ∼  . (3.16)
j=1 k=1 θ D
jk

We decompose connectedness into short-term (1–5 business days), medium-term

(6–20 business days), and long-term (over 21 business days) factors. In other words,
Eq. (3.14) becomes
9π
5
1
(θ Shor t ) jk = B j (ω)ϕ(ω) jk dω
2π
π
5
π 39π
5 20
1 1
(θ Medium ) jk = B j (ω)ϕ(ω) jk dω + B j (ω)ϕ(ω) jk dω
2π 2π
π 9π
20 5
π
20 2π
  1 1
θ Long jk = B j (ω)ϕ(ω) jk dω + B j (ω)ϕ(ω) jk dω. (3.17)
2π 2π
0 39π
20

3.3.3 EGARCH Volatility Series Estimation

To measure the risk spillover between markets, we examine the volatility series that
represents the second moment of the return series. We can categorize volatility series
estimations into three types: autoregressive conditional heteroskedasticity (ARCH),
stochastic volatility (SV), and realized volatility (RV) models. The ARCH model
is a simultaneous model of the return and latent volatility series. The SV model
is a formulation of the logarithmic fluctuation of volatility using an autoregressive
62 3 Fuel Market Connectedness and Fuel Portfolio Risk

moving average model. The RV series is not estimated by a certain model but is the
sum of the squares of the returns in a very short time calculated using high-frequency
data over the measurement period. We adopt Nelson’s [14] EGARCH model, an
ARCH model, here. The EGARCH model is described by the autoregressive (AR)-
EGARCH model:


k
Rt = c0 + ci Rt−i + εt
i=1
√
εt = ht ut

ν 1 || z ||ν
ut = exp − | |
λ21+1/ν Γ(1/ν) 2 λ
/
2−2/ν Γ(1/ν)
λ=
Γ(3/ν)
| |
p
| εt−i |  q
r
εt−i
ln(h t ) = θ + αi || √ |+
| β i ln(h t−i ) + γi √ , (3.18)
i=1
h t−i i=1 i=1
h t−i

where Rt is the conditional mean of the return series; h t is the conditional variance
of the return series; εt is the prediction error series;ci ,θ ,αi ,βi , and γi are the AR-
EGARCH model parameters; u t is the generalized error distribution (GED); Γ(·)
is the Gamma function; and ν(> 0) is the shape parameter. We can estimate all
parameters by the maximum likelihood method. We selected the lag length of each
term based on the SBIC. After determining the lag order of the AR term, ARCH
term, generalized autoregressive conditional heteroscedasticity (GARCH) term, and
asymmetric term, we have the order of the terms. This model offers two advantages.
First, the estimated volatility always meets the non-negative constraint regardless
of the estimated parameters because we use the logarithmic value of volatility as
the dependent variable. Second, this model captures the asymmetry of volatility
fluctuations due to positive and negative returns.

3.4 Analysis Results

This section reports the results of the analysis of the crude oil and natural gas portfo-
lios consisting of representative indexes in Europe, North America, and Asia. After
the necessary pre-analysis and model estimation, we measure the connectedness
indices and perform the spectral decomposition. We examine the spillover effects of
both the return and volatility series between these three markets. Finally, we measure
the VaR and expected shortfall for each portfolio.
3.4 Analysis Results 63

3.4.1 Crude Oil

Diebold and Yilmaz’s [4]method requires VMA representation. That is, all variables
must be stationary. As described in Chap. 2, ordinary price series are often non-
stationary processes. Thus, we apply this technique to the crude oil return series
rather than their price series to examine the price spillover effect between crude oil
markets. Here, we obtain the return series using the following equation:

(Price)t − (Price)t−1
(Retur n)t = . (3.19)
(Price)t−1

However, in general, the following equation is also common:

(Retur n)t = log(Price)t − log(Price)t−1 . (3.20)

Table 3.3 shows the basic statistics for the crude oil return series. The mean and
median values for Brent and WTI are significantly higher than those for Dubai-
Oman. For all three variables, the median value was greater than the mean. Thus,
each distribution appeared to have a considerable number of outliers in the left tail.
These characteristics are very strong, especially for Brent and WTI. The range of
the Brent return series, which is approximately the same as the range of the WTI
return series, is narrower than that of the Dubai-Oman return series. Conversely, the
standard deviation for Brent, which is the same level as that for WTI, is larger than that
for Dubai-Oman. The dispersions for Brent and WTI tend to be similar. In summary,
the Dubai-Oman return series are less variable than the Brent and WTI, although the
range is wider because of the extremely small minimum value. As all three variables
have negative skewness, their distributions have a long left tail. Because all three
return series have kurtosis over 3, their distributions have a sharp peak and long, fat
tails. The Jarque–Bera test rejects the hypothesis of normally distributed skewness
and kurtosis for each crude oil return series.

Table 3.3 Descriptive

Brent WTI Dubai-Oman
statistics (crude oil return
series) Observations 469 469 469
Mean 0.019% 0.018% 0.011%
Median 0.190% 0.152% 0.021%
Maximum 9.31% 8.19% 7.76%
Minimum – 7.17% – 7.90% – 14.41%
Standard deviation 0.0186 0.0197 0.0154
Skewness – 0.26 – 0.32 – 1.39
Kurtosis 5.79 4.83 23.16
Jarque–Bera 158 (0.00) 73 (0.00) 8094 (0.00)
Note p-values are in parentheses
64 3 Fuel Market Connectedness and Fuel Portfolio Risk

Figure 3.3 plots the return series. As Table 3.3 implies, this figure shows that the
Dubai-Oman return series has some outliers that differ from the other two variables.
However, we observe that all three variables tended to be similar.
Table 3.4 provides the results of the unit-root test for the crude oil return series.
As the augmented Dickey-Fuller (ADF) test without a constant term and time-trend
term rejects the unit-root hypothesis for all variables, we can confirm that all return
series are stationary processes. In other words, we can represent the VAR model in
the VMA.
We estimate the VAR model of the return series by selecting the lag order based
on the SBIC. Table 3.5 summarizes the results. The lag order is 1. We present the
variance–covariance matrix of the residuals of the VAR model in Table 3.6.
We depict the spillover indices obtained from Tables 3.5 and 3.6 in Fig.3.4. It is no
exaggeration to say that the Brent and WTI return series are almost linked because
their mutual spillover effects are extremely strong, at about 45%. Both Brent and WTI

Fig. 3.3 Crude oil return series

Table 3.4 ADF unit-root test

Return series ADF-t statistics
results (crude oil return
series) Brent –22.55* (0.00)
WTI –22.48* (0.00)
Dubai-Oman –23.49* (0.00)
Note * indicates rejection of the unit-root hypothesis at the 1%
significance level. p-values are in parentheses
3.4 Analysis Results 65

Table 3.5 VAR model (crude

Brent (t) WTI (t) Dubai-Oman (t)
oil return series)
Brent (t − 1) –0.231 –0.105 0.207
WTI (t − 1) 0.203 0.056 0.188
Dubai-Oman (t − 1) –0.090 –0.070 –0.130

Table 3.6 Residual

Brent WTI Dubai-Oman
variance–covariance matrix in
the VAR model (crude oil Brent 3.44 × 10–4 3.34 × 10–4 3.16 × 10–5
return series) WTI 3.34 × 10–4 3.89 × 10–4 2.97 × 10–5
Dubai-Oman 3.16 × 10–5 2.97 × 10–5 1.83 × 10–4

have a moderately strong impact on Dubai-Oman, at around 18%. On the other hand,
the spillover effect from Dubai-Oman to Brent and WTI is barely observable. The
total connectedness for these crude oil markets was 42.95%. We can thus conclude
that market integration has progressed to some extent. In terms of returns, the value
of a portfolio consisting of these three crude oil indexes might not be so high and
might not be much different from a portfolio consisting of Dubai-Oman plus Brent
or Dubai-Oman plus WTI.
We present the spectral analysis of the connectedness indices in Table 3.7. The
total connectedness from 1 to 5 business days, from 6 to 20 business days, and over
21 business days is 34.97, 6.01, and 1.97%, respectively. Considering that the total

Fig. 3.4 Spillover effects between return series (crude oil)

66 3 Fuel Market Connectedness and Fuel Portfolio Risk

Table 3.7 Spillover index and spectral analysis (crude oil return series)
To From Bandwidth
Brent (%) WTI (%) Dubai-Oman (%) Others (%)
Brent 44.38 36.60 1.07 12.56 0  date  5
WTI 37.44 44.31 0.78 12.74
Dubai-Oman 14.64 14.37 52.56 9.67
Others 17.36 16.99 0.62 34.97
Brent 7.14 6.34 0.04 2.13 6  date  20
WTI 5.87 7.26 0.04 1.97
Dubai-Oman 2.89 2.86 8.12 1.92
Others 2.92 3.07 0.03 6.01
Brent 2.33 2.08 0.01 0.70 21  date
WTI 1.92 2.38 0.01 0.64
Dubai-Oman 0.95 0.94 2.65 0.63
Others 0.96 1.01 0.01 1.97
Brent 53.85 45.02 1.12 15.38 Total
WTI 45.22 53.95 0.83 15.35
Dubai-Oman 18.48 18.18 63.34 12.22
Others 21.24 21.07 0.65 42.95

connectedness for the entire period is 42.95%, the short-term factors contribute the
most to the return spillover. The spillover effect depends largely on events occurring
within approximately one week. Each spillover effect between these variables tends
to be almost the same as total connectedness. The short-, medium- and long-term
factors had the largest, second largest, and smallest effects, respectively.
To investigate risk spillover, we measure volatility connectedness. Table 3.8 shows
the estimated AR-EGARCH model for generating volatility series. We determine
the lag length of each term based on the SBIC. We select the lag order of the AR
term, ARCH term, GARCH term, and asymmetric term in this order. Comparing the
estimated coefficients, each model for the three variables is similar. The ARCH term
(α1 ) is statistically significant at the 1% level for Dubai-Oman, but not for Brent
and WTI. The GARCH term (β1 ) and asymmetric parameter (γ1 ) are statistically
significant at the 1% level for all variables. The GED parameter is also statistically
significant for all variables and is less than 2. This means that these error terms have
a fat-tailed distribution.
Table 3.9 lists the descriptive statistics of the crude oil volatility series generated
by the estimated EGARCH model. For these crude oil volatility series, the mean is
larger than the median; that is, the right tail has a large number of outliers. The Dubai-
Oman volatility series has a smaller mean, median, maximum, and minimum than do
the Brent and WTI volatility series. This is consistent with the standard deviation of
the return series in Table 3.3. Moreover, the standard deviation of the Dubai-Oman
3.4 Analysis Results 67

Table 3.8 Estimated AR-EGARCH model (crude oil return series)

Brent WTI Dubai-Oman
Mean equation c0 0.001 (0.03) 0.001 (0.11) 0.000 (0.90)
c1 –0.061 (0.10) –0.049 (0.23) –0.057 (0.00)
Variance equation θ –0.320 (0.00) –0.213 (0.03) –0.267 (0.00)
α1 –0.065 (0.05) –0.022 (0.47) –0.148 (0.00)
β1 0.956 (0.00) 0.972 (0.00) 0.958 (0.00)
γ1 –0.152 (0.00) –0.116 (0.00) –0.121 (0.00)
GED ν 1.173 (0.00) 1.273 (0.00) 0.813 (0.00)
Note p-values are in parentheses

volatility series is smaller than the other two series. Skewness is positive for the
volatility series for all these indexes. That is, each distribution has a long right tail.
Kurtosis is over 3 for Brent and WTI. That is, each distribution has a sharp peak and
long, fat tails. The Dubai-Oman kurtosis is less than 3. Its distribution has a rounded
peak and short, thin tails. According to the Jarque–Bera statistics, no volatility series
is significantly normally distributed in terms of skewness and kurtosis.
Figure 3.5 provides the time plot of the volatility series for the three crude oil
indices. The Brent and WTI volatilities fluctuate synchronously at approximately
the same level. The Dubai-Oman volatility level is lower than that of the other two
series. However, these three volatility series appear synchronized.
Table 3.10 presents the ADF test results for these volatility series. All tests adopted
a model without a time-trend term and with a constant term, as in Eq. (2.15). We
cannot reject that the WTI and Dubai-Oman volatility series have a unit root at the 1%
significance level. However, we reject the unit-root hypothesis for all volatility series
at the 10% significance level. These volatility series are considered to be stationary
processes. Thus, we apply Diebold and Yilmaz’s [4] approach.

Table 3.9 Descriptive

Brent WTI Dubai-Oman
statistics (crude oil volatility
series) Observations 468 468 468
Mean 3.45 × 10–4 3.96 × 10–4 1.94 × 10–4
Median 2.98 × 10–4 3.52 × 10–4 1.76 × 10–4
Maximum 1.10 × 10–3 1.16 × 10–3 5.27 × 10–4
Minimum 8.28 × 10–5 1.50 × 10–4 1.45 × 10–5
Standard deviation 1.78 × 10–4 1.93 × 10–4 1.13 × 10–4
Skewness 1.69 1.81 0.55
Kurtosis 6.35 6.47 2.57
Jarque–Bera 440 (0.00) 491 (0.00) 27 (0.00)
Note p-values are in parentheses
68 3 Fuel Market Connectedness and Fuel Portfolio Risk

Fig. 3.5 Crude oil volatility series

Table 3.10 ADF unit-root

Volatility series ADF-t statistics
test results (crude oil
volatility series) Brent –3.44* (0.01)
WTI –2.79*** (0.06)
Dubai-Oman –2.88** (0.05)
Note *, **, and *** indicate rejection of the unit-root hypothesis
at the 1, 5, and 10% significance level, respectively. p-values are
in parentheses

We estimate the VAR model for the volatility series after selecting the lag order
using the SBIC. We provide the estimates in Table 3.11 and the variance–covariance
matrix of the residuals of this VAR model in Table 3.12. We require this matrix to
apply Diebold and Yilmaz’s [4] approach.
We illustrate the spillover effects calculated using Tables 3.11 and 3.12, in Fig. 3.6.
Similar to return spillover, Brent and WTI have a very strong mutual spillover effect.
The spillover effect on Dubai-Oman volatility is extremely strong from both Brent

Table 3.11 VAR model

Brent (t) WTI (t) Dubai-Oman (t)
(crude oil volatility series)
Brent (t − 1) 0.904 –0.015 0.126
WTI (t − 1) 0.068 0.999 –0.062
Dubai-Oman (t − 1) 0.019 0.018 0.901
3.4 Analysis Results 69

Table 3.12 Residual

Brent WTI Dubai-Oman
variance–covariance matrix in
the VAR model (crude oil Brent 2.97 × 10–9 2.42 × 10–9 1.53 × 10–10
volatility series) WTI 2.42 × 10–9 2.39 × 10–9 1.37 × 10–10
Dubai-Oman 1.53 × 10–10 1.37 × 10–10 8.15 × 10–10

and WTI volatility. On the other hand, the spillover from Dubai-Oman to the other
markets is less than that from either Brent or WTI. The total connectedness of these
volatilities is 62.00%. This result implies that these markets are integrated at a consid-
erable level in terms of risk. However, it does not completely deny the value of this
portfolio, which consists of the three indexes of risk diversification.
Table 3.13 reports the results of the spectral decomposition of the connectedness
indexes shown in Fig. 3.6. The total connectedness of the crude oil volatility series
is 62.00%, which is 0.21% for short-term, 0.69% for medium-term, and 61.10% for
long-term factors. In contrast to the return series, long-term factors contribute the
most to volatility spillover; that is, most of the spillover effect is caused by events
that occurred more than one month previously. Each spillover effect between these
variables tends to be almost the same as total connectedness. Events within a week
have the least impact, events that occurred between a month and two weeks prior
have the second lowest impact, and events that occurred more than one month prior
had the largest impact.

Fig. 3.6 Spillover effects between volatility series (crude oil)

70 3 Fuel Market Connectedness and Fuel Portfolio Risk

Table 3.13 Spillover index and spectral analysis (crude oil volatilities)
To From Bandwidth
Brent (%) WTI (%) Dubai-Oman (%) Others (%)
Brent 0.27 0.16 0.00 0.05 0  date  5
WTI 0.05 0.07 0.00 0.02
Dubai-Oman 0.20 0.21 0.69 0.14
Others 0.09 0.12 0.00 0.21
Brent 0.82 0.49 0.00 0.16 6  date  20
WTI 0.17 0.25 0.01 0.06
Dubai-Oman 0.76 0.63 2.03 0.47
Others 0.31 0.37 0.00 0.69
Brent 46.95 49.49 1.81 17.10 21  date
WTI 44.56 53.03 1.86 15.47
Dubai-Oman 44.96 40.62 9.89 28.53
Others 29.84 30.04 1.22 61.10
Brent 48.05 50.14 1.81 17.32 Total
WTI 44.79 53.35 1.86 15.55
Dubai-Oman 45.93 41.46 12.61 29.13
Others 30.24 30.53 1.23 62.00

Various petroleum products and fuel oils used to generate power are often traded at
prices linked to the international crude oil index. Therefore, we measure the risk of a
portfolio consisting of these crude oil indexes (Brent, WTI, and Dubai-Oman). First,
we estimate the Gaussian, t, Clayton, and Gumbel copulas introduced in Chap. 2.
Second, we select the most appropriate copula. Finally, we determine the VaR and
expected shortfall using random numbers following the copula.
Table 3.14 presents the estimated means and standard deviations, which are the
parameters of the marginal distribution function of each return series.
We estimate each copula according to the procedure described in Sect.2.5.2. While
Chap. 2 estimates bivariate copulas, in this chapter, we estimate trivariate copulas.
Figure 3.7 plots the relationship between the degree of freedom and the log-
likelihood when estimating the t copula. The maximum log-likelihood occurs when
the degree of freedom is 4. Therefore, we estimate the t copula parameters; that is,


Σ , with degree of freedom 4. We report all estimated copulas in Table 3.15.

Table 3.14 Parameters of

Brent WTI Dubai-Oman
each marginal distribution
function (crude oil returns) Mean 1.85 × 10–4 1.84 × 10–4 1.09 × 10–4
Standard deviation 1.86 × 10–2 1.97 × 10–2 1.54 × 10–2
3.4 Analysis Results 71

Fig. 3.7 Degree of freedom and log-likelihood, t copula (crude oil portfolio)

We generated the distribution of the daily rate of return for the portfolio by simu-
lating each copula 100,000 times. We then calculate the VaR and expected short-
fall. Table 3.15 provides the measured VaRs and expected shortfalls based on these
copulas. According to the Gumbel copula with the minimum Akaike information
criterion (AIC), the VaR and expected shortfall are 4.81 and 6.07%, respectively.
This portfolio has a 1% chance of losing more than 4.81% one day later, with an
average loss of 6.07% and a probability of 1%.
Table 3.16 lists the risk for each crude oil index alone. There is no diversification
effect that decreases the VaR and expected shortfall of the portfolio to below that of
Dubai-Oman its volatility is much lower (about 50%) than that of Brent and WTI, as
Table 3.9 indicates. However, the risk is smaller when we include it in the portfolio
than when procuring fuel oil at the price linked to Brent or WTI alone. Firms that
have only long positions in Brent or WTI, should consider the risk of such a portfolio.

3.4.2 Natural Gas

Similar to the analysis of the crude oil markets in Sect. 3.4.1, we examine the spillover
effect between three representative natural gas price indices in the European, North
American, and Asian markets. Moreover, we measure the risk of such a natural gas
portfolio.
72 3 Fuel Market Connectedness and Fuel Portfolio Risk

Table 3.15 Risk measurement using the estimated copulas (crude oil portfolio)
Gaussian AIC –1.144


Σ 1.000 0.914 0.078

0.914 1.000 0.094
0.078 0.094 1.000
Value-at-risk* 4.41%
Expected shortfall* 5.47%
t AIC 0.647
Degree of freedom 4


Σ 1.000 0.927 0.073

0.927 1.000 0.079
0.073 0.079 1.000
Value-at-risk* 3.61%
Expected shortfall* 4.21%
Clayton AIC – 8.455
πc 1.954
Value-at-risk* 4.79%
Expected shortfall* 6.03%
Gumbel AIC –9.044
πg 2.326
Value-at-risk* 4.81%
Expected shortfall* 6.07%
Note * indicates the value at the 99% confidence level
Bold indicates the minimum AIC and the risk measures at that time

Table 3.16 Risk of each

Brent (%) WTI (%) Dubai-Oman (%)
crude oil market
Value-at-risk* 6.34 5.44 4.29
Expected shortfall* 7.87 6.82 5.39
Note * indicates the value at the 99% confidence level

Table 3.17 shows the descriptive statistics for the TTF return series in Europe,
the HH return series in North America, and the JKM return series in Asia. All mean
values are negative, and all medians are negative, except for JKM, which had a median
of 0.00%. We can conclude that all prices have a downward trend during this period.
The descending order of the mean value is –0.023% for HH, –0.049% for TTF, and
–0.100% for JKM; the descending order of the median value is 0.000% for JKM, –
0.070% for HH, and –0.204% for TTF; the descending order of the maximum values
is 42.5% for JKM, 36.8% for TTF, and 17.9% for HH; the descending order of the
minimum values is –11.8% for TTF, –16.5% for HH, and –23.4% for JKM; and the
descending order of the standard deviation is 0.0294 for TTF, 0.0322 for JKM, and
3.4 Analysis Results 73

0.0356 for HH. We assume that these three return series have different distributions.
Skewness had a positive value for all three variables, indicating a long right tail of
each distribution. Kurtosis exceeded 3 for all variables. Their distributions have sharp
peaks and long, fat tails. The Jarque–Bera test rejects the normality hypothesis for
each distribution.
We plot each return series in Fig. 3.8. Similar to the impression Table 3.17 gives,
we can observe that these variables fluctuate without synchronization.
Table 3.18 shows the ADF unit-root test results for the natural gas volatility series.
All tests adopt a model that has neither a constant term nor a time-trend term. We
can reject the unit-root hypothesis at the 1% significance level in all tests. We can
apply the Diebold and Yilmaz [4] approach to these natural gas return series because
we confirm the stationarity of all series.
Table 3.19 presents the estimated coefficients of the VAR model for these natural
gas return series. The lag order is 1, which we select based on SBIC. Table 3.20
shows the variance–covariance matrix.
Figure 3.9 illustrates the spillover effects, calculated by applying Tables 3.19 and
3.20 to Eq. (3.5). The strongest spillover effect is 2.91% from the JKM return series
to the TTF return series, and the second-strongest spillover effect is 1.50% from the
TTF return series to the JKM return series. All the other spillover effects are less
than 1%. Moreover, total connectedness has a small value, of 1.72%. These markets
are not integrated in terms of returns. The value of a portfolio consisting of these
three natural gas indices may be more attractive than holding each index alone. This
is because we can expect a diversification effect from this procurement approach.
Table 3.21 lists the spectral analysis results. The total connectedness is 1.72%,
which consists of 1.56% for short-term factors, 0.12% for medium-term factors, and
0.04% for long-term factors. Hence, the short-term factors explain total connected-
ness. Each spillover effect has the same tendency. The largest is due to short-term
factors, and the weakest is due to long-term factors. However, the significance of the
spectral decomposition might be low because the total connectedness is extremely
small.

Table 3.17 Descriptive

TTF HH JKM
statistics (natural gas return
series) Observations 503 503 503
Mean – 0.049% – 0.023% – 0.100%
Median – 0.204% – 0.070% 0.000%
Maximum 36.8% 17.9% 42.5%
Minimum – 11.8% – 16.5% – 23.4%
Standard 0.0356 0.0294 0.0322
deviation
Skewness 2.96 0.13 4.86
Kurtosis 29.33 10.92 76.63
Jarque–Bera 15,264 (0.00) 1316 (0.00) 115,608 (0.00)
Note p-values are in parentheses
74 3 Fuel Market Connectedness and Fuel Portfolio Risk

Fig. 3.8 Natural gas return series

Table 3.18 ADF unit-root

Return ADF-t statistics
test results (natural gas return
series) TTF –21.53* (0.00)
HH –12.32* (0.00)
JKM –23.28* (0.00)
Note * indicates rejection of the unit-root hypothesis at the 1%
significance level. p-values are in parentheses

Table 3.19 VAR model

TTF (t) HH (t) JKM (t)
(natural gas returns)
TTF (t − 1) 0.055 0.056 0.008
HH (t − 1) 0.059 –0.049 0.025
JKM (t − 1) –0.141 –0.010 –0.040

Table 3.20 Residual

TTF HH JKM
variance–covariance matrix in
the VAR model (natural gas TTF 1.25 × 10–3 –9.95 × 10–6 1.41 × 10–4
returns) HH –9.95 × 10–6 8.64 × 10–4 –1.16 × 10–7
JKM 1.41 × 10–4 –1.16 × 10–7 1.04 × 10–3
3.4 Analysis Results 75

Fig. 3.9 Spillover effects between return series (natural gas)

Table 3.21 Spillover index and spectral analysis (natural gas returns)
To From Bandwidth
TTF (%) HH (%) JKM (%) Others (%)
TTF 75.92 0.21 2.87 1.03 0  date  5
HH 0.38 81.35 0.00 0.13
JKM 1.18 0.04 80.19 0.41
Others 0.52 0.08 0.96 1.56
TTF 15.66 0.02 0.04 0.02 6  date  20
HH 0.05 13.67 0.00 0.02
JKM 0.24 0.01 13.72 0.08
Others 0.10 0.01 0.01 0.12
TTF 5.27 0.01 0.00 0.00 21  date
HH 0.02 4.53 0.00 0.01
JKM 0.08 0.00 4.54 0.03
Others 0.03 0.00 0.00 0.04
TTF 96.85 0.24 2.91 1.05 Total
HH 0.44 99.55 0.01 0.15
JKM 1.50 0.05 98.45 0.52
Others 0.65 0.10 0.97 1.72
76 3 Fuel Market Connectedness and Fuel Portfolio Risk

Table 3.22 Estimated AR-EGARCH model (Natural gas return series)

TTF HH JKM
Mean equation c0 –0.001 (0.51) 0.000 (0.65) 0.000 (1.00)
c1 0.017 (0.67) –0.043 (0.35) 0.000 (0.21)
c2 –0.059 (0.18)
c3 0.043 (0.34)
Variance equation θ –0.277 (0.00) –0.247 (0.01) –4.655 (0.02)
α1 0.200 (0.00) 0.200 (0.00) 2.051 (0.25)
β1 0.981 (0.00) 0.988 (0.00) –0.207 (0.63)
β2 –0.229 (0.45)
γ1 –0.020 (0.59) 0.078 (0.00) –0.509 (0.75)
GED ν 1.154 (0.00) 1.447 (0.00) 0.103 (0.00)
Note p-values are in parentheses

Next, we examine the spread of risk. We generate the volatility series for each
natural gas index, which we require to measure the connectedness of the volatility.
We estimate the AR-EGARCH model of natural gas returns and select the lag order
of the model based on the SBIC. We select the lag order of the AR term, ARCH term,
GARCH term, and asymmetric term in this order. The estimated GED parameter is
less than 2. Table 3.22 presents the estimated parameters of the AR-EGARCH model.
Table 3.23 provides the basic statistics of these natural gas volatility series, which
we generated using the estimated EGARCH models. The mean, median, maximum,
and minimum for the JKM is two orders of magnitude larger than those of the TTF
and HH. These results express the instability of the JKM return series. Skewness
for all variables is positive. That means that all distributions have a long right tail.
Kurtosis for all variables is over 3, indicating that all distributions have a sharp peak
and long, fat tails. According to the Jarque–Bera test, none of the volatility series are
significantly normally distributed.
Figure 3.10 displays a time plot of the volatility series. The TTF series had small
peaks in March 2018 and April 2019, and increased from June to October 2019.
The HH series exhibited a large peak at the end of 2018. The JKM series has some
very large spikes, whereas it barely fluctuated from approximately 0.04. All three
variables have different, though characteristic, fluctuations.
We test whether Diebold and Yilmaz’s [4] approach is applicable to these volatility
series. The ADF test, which uses a model with a constant term and without a time-
trend term, rejects the unit-root hypothesis for all volatility series. Table 3.24 shows
the ADF unit-root test results. The vector of these volatility series is a stationary
covariance process. Therefore, we can represent the VAR model as a VMA model.
We estimate a VAR model for the natural gas volatility series using the lag order
chosen based on the SBIC. We summarize the estimated coefficients in Table 3.25
and provide the variance–covariance matrix of the residuals of this VAR model in
Table 3.26. We require these residuals to apply the Diebold and Yilmaz [4] approach.
3.4 Analysis Results 77

Table 3.23 Descriptive

TTF HH JKM
statistics (natural gas
volatility series) Observations 502 500 502
Mean 1.10 × 10–3 7.87 × 10–4 4.81 × 10–2
Median 7.65 × 10–4 4.61 × 10–4 4.00 × 10–2
Maximum 7.61 × 10–3 6.56 × 10–3 1.05
Minimum 1.67 × 10–4 1.63 × 10–4 1.83 × 10–2
Standard 9.13 × 10–4 8.83 × 10–4 6.32 × 10–2
deviation
Skewness 2.54 3.08 12.63
Kurtosis 13.07 14.81 177.74
Jarque–Bera 2662 (0.00) 3694 (0.00) 6.52 × 105
(0.00)
Note p-values are in parentheses

Fig. 3.10 Natural gas volatility series

Table 3.24 ADF unit-root

Volatility series ADF-t statistics
test results (natural gas
volatility series) TTF –4.12* (0.00)
HH –3.64* (0.01)
JKM –24.47* (0.00)
Note * indicates rejection of the unit-root hypothesis at the 1%
significance level. p-values are in parentheses
78 3 Fuel Market Connectedness and Fuel Portfolio Risk

Table 3.25 VAR model

TTF (t) HH (t) JKM (t)
(natural gas volatility series)
TTF (t − 1) 0.948 0.010 18.547
HH (t − 1) 0.008 0.972 12.675
JKM (t − 1) 0.001 0.000 0.109

Table 3.26 Residual

TTF HH JKM
variance–covariance matrix in
the VAR model (natural gas TTF 1.03 × 10–7 1.36 × 10–9 5.55 × 10–7
volatility series) HH 1.36 × 10–9 6.10 × 10–8 5.42 × 10–7
JKM 5.55 × 10–7 5.42 × 10–7 4.04 × 10–3

Figure 3.11 depicts the connectedness indexes calculated from Tables 3.25 and
3.26. The total connectedness is 16.90%, which is stronger than the total connect-
edness of the return series. The risk impact of the HH on the other two indexes is
significant, at 10.66% for the TTF and 10.17% for the JKM. The volatility spillover
from the TTF to the JKM is the strongest, at 16.59%, while the volatility spillover
from the TTF to the HH is not large, at 7.49%. The volatility spillover from the JKM
to the TTF is 4.85%, and the volatility spillover from the JKM to the HH is 0.96%.
The JKM does not have a large risk impact on the natural gas market.
Table 3.27 reports the spectral analysis of the connectedness indexes of the natural
gas volatility series. The total connectedness during the full period is 16.90%, while

Fig. 3.11 Spillover effects between volatility series (natural gas)

3.4 Analysis Results 79

the total connectedness from 1 to 5 business days, from 6 to 20 business day, and
over 21 business days is 0.14, 0.56, and 16.21%, respectively. The volatility spillover
effect depends mostly on events before 21 business days. The spectral decomposition
of each spillover tends to be the same as of total connectedness. The short-term factor
is the smallest, followed by the medium- and long-term factors.
LNG transported by ships and natural gas supplied in pipelines is often traded
at prices linked to the international natural gas index. Similar to how we measured
the risk of the crude oil portfolio in Sect.3.4.1, we measure the risk of a portfolio
consisting of the TTF, HH, and JKM. We obtain the VaR and expected shortfall of this
natural gas portfolio by simulation using random numbers following the Gaussian,
t, Clayton, and Gumbel copulas. We select the VaR and expected shortfall obtained
by simulation using the copula with the smallest AIC.
Table 3.28 presents the means and standard deviations, which are the parameters
of the marginal distribution function of each return series.

Table 3.27 Spillover index and spectral analysis (natural gas volatility series)
To From Bandwidth
TTF (%) HH (%) JKM (%) Others (%)
TTF 1.70 0.07 0.09 0.05 0  date  5
HH 0.05 1.36 0.00 0.02
JKM 0.19 0.02 53.02 0.07
Others 0.08 0.03 0.03 0.14
TTF 6.23 0.30 0.30 0.20 6  date  20
HH 0.19 5.19 0.00 0.07
JKM 0.79 0.10 12.39 0.30
Others 0.33 0.13 0.10 0.56
TTF 76.56 10.29 4.46 4.92 21  date
HH 7.25 85.01 0.95 2.73
JKM 15.61 10.06 7.83 8.56
Others 7.62 6.78 1.80 16.21
TTF 84.49 10.66 4.85 5.17 Total
HH 7.49 91.56 0.96 2.81
JKM 16.59 10.17 73.24 8.92
Others 8.02 6.94 1.93 16.90

Table 3.28 Parameters of

TTF HH JKM
each marginal distribution
function (natural gas return Mean –4.92 × 10–4 –2.30 × 10–4 –1.00 × 10–3
series) Standard 3.56 × 10–2 2.94 × 10–2 3.22 × 10–2
deviation
80 3 Fuel Market Connectedness and Fuel Portfolio Risk

First, we estimate four trivariate copulas: the Gaussian, t, Clayton, and Gumbel
copulas. Figure 3.12 provides a graph of the relationship between the degree of
freedom and the log-likelihood for the t copula estimate. The maximum log-
likelihood occurs when the degree of freedom is 3. Therefore, the t copula parameters


are the elements of the correlation matrix Σ estimated with degree of freedom 3.
We summarize all estimated copulas and measured risk values in Table 3.29. The
VaR and expected shortfall are 8.39 and 10.65%, respectively. We measure these
using random numbers following the Gumbel copula with the smallest AIC. This
natural gas portfolio has a 1% chance of losing more than 8.39% one day later, with
an average loss of 10.65% and a probability of 1%. Compared to the VaR of 4.81%
and expected shortfall of 6.87% of the crude oil portfolio, this natural gas portfolio
carries higher risk. The risk of each natural gas index in Table 3.30 is much higher
than the risk of each crude oil index in Table 3.16. The reason that the natural gas
portfolio is of higher risk than the crude oil portfolio does not seem to be that the
diversification of natural gas procurement is ineffective. Rather, although the risk of
TTF or JKM alone is high, it is possible to reduce the HH risk level by forming this
natural gas portfolio.

Fig. 3.12 Degree of freedom and log-likelihood in estimating the t copula (natural gas portfolio)
3.5 Concluding Remarks 81

Table 3.29 Risk measurement using the estimated copulas (natural gas portfolio)
Gaussian AIC – 0.180


Σ 1.000 0.004 –0.002

0.004 1.000 0.218
–0.002 0.218 1.000
Value-at-risk* 5.56%
Expected shortfall* 6.71%
t AIC 1.175
Degree of freedom 3


Σ 1.000 0.021 0.037

0.021 1.000 0.285
0.037 0.285 1.000
Value-at-risk* 5.28%
Expected shortfall* 6.34%
Clayton AIC – 7.301
πc 1.530
Value-at-risk* 8.64%
Expected shortfall* 10.91%
Gumbel AIC –8.106
πg 2.090
Value-at-risk* 8.39%
Expected shortfall* 10.65%
Note * indicates the value at the 99% confidence level
Bold indicates the minimum AIC and the risk measures at that time

Table 3.30 Risk of each

TTF (%) HH (%) JKM (%)
natural gas market
Value-at-risk* 9.88 8.15 9.21
Expected shortfall* 12.28 10.24 11.51
Note * indicates the value at the 99% confidence level

3.5 Concluding Remarks

When we procure different fuels at prices linked to different price indicators, we

should understand the spillover effect of the price level and the risk between these
indicators. We should determine the diversification effect of procurement based on
the total connectedness of an energy portfolio. We must monitor the connectedness of
the portfolio, including potential price indicators and grasp the impact of excluding
components already in our portfolio. Quantitative measures of risk of such a potential
portfolio is extremely important. It is beneficial to consider what components we
82 3 Fuel Market Connectedness and Fuel Portfolio Risk

should have and how much we should hold to reconstruct the portfolio. Moreover,
the measurement results may affect the financial strategy.
This chapter describes the connectedness index proposed by Diebold and Yilmaz
[4], which can capture the spillover effect between markets. We then introduce
Baruník and Křehlík’s [1] spectral decomposition method, which determines when
spillover factors occur. Moreover, we explain the EGARCH model, which can
generate each volatility series needed to examine the risk spillover effects.
We illustrate two cases: a crude oil portfolio and a natural gas portfolio. We
examine each portfolio, which consists of representative markets for Europe, North
America, and Asia; specifically, Brent, WTI, and Dubai-Oman to represent crude oil
markets and the TTF, HH, and JKM to represent natural gas markets, respectively.
First, we describe the crude oil markets. While Dubai-Oman’s returns and
volatility are smaller than those of Brent or WTI, the price series, return series,
and volatility series of these three indicators appear to fluctuate synchronously. The
total connectedness of the return and volatilities series are 42.95 and 62.00%, respec-
tively. These crude oil markets appear to be integrated at a relatively high level. The
spillover effects of returns and volatility between Brent and WTI are mutually strong.
Dubai-Oman receives considerable risk from both Brent and WTI.
We next describe the natural gas markets. The prices, returns, and volatility for
the TTF, HH, and JKM fluctuate in three different ways. Although the total connect-
edness of volatility is 16.90%, that of returns is 1.72%. This result indicates that
intercontinental natural gas market liquidity may still be low. Any spillover effect
of returns is less than 1%, except between TTF and JKM. The volatility spillover
effects larger than 10%, are from HH to TTF, TTF to JKM, and HH to JKM.
Third, we describe the results of the spectral decomposition of connectedness.
Both the crude oil and natural gas analysis results indicate that the spillover effects
of the return series depend mostly on short-term factors (i.e., events within 5 business
days), while the spillover effects of the volatility series depend mostly on long-term
factors (i.e., events more than a month previously).
Finally, we describe the risk measurement results. The crude oil portfolio has a
VaR of 4.81% and an expected shortfall of 6.07%, whereas the natural gas portfolio
has a VaR of 8.39% and an expected shortfall of 10.65%. The risk of the natural
gas portfolio is higher than that of the crude oil portfolio because the procurement
diversification effect of natural gas is less than that of crude oil as natural gas has
much higher risk than does crude oil.
Readers should understand how to monitor markets and measure portfolio risk
when designing a portfolio. In practice, we should handle more variables and monitor
connectedness more frequently, that is, with the observation period as a moving
window. While we analyze only a fuel procurement portfolio, we can also measure
the risk of an electricity sales portfolio by keeping in mind that the right tail of the
return distribution is a risk. Electricity retailers often sell power using a variety of
price formulas, including fixed prices. The approaches and exercises introduced in
this chapter are meaningful for such practitioners.
References 83

References

1. Baruník, J., & Křehlík, T. (2018). Measuring the frequency dynamics of financial connectedness
and systemic risk. Journal of Financial Econometrics, 16(2), 271–296.
2. Bastianin, A., Galeotti, M., & Polo, M. (2019). Convergence of European natural gas prices.
Energy Economics, 81, 793–811.
3. Chai, J., Wei, Z., Hu, Y., Su, S., & Zhang, Z. G. (2019). Is China’s natural gas market globally
connected? Energy Policy, 132, 940–949.
4. Diebold, F. X., & Yilmaz, K. (2012). Better to give than to receive: Predictive directional
measurement of volatility spillovers. International Journal of Forecasting, 28(1), 57–66.
5. Fattouh, B. (2010). The dynamics of crude oil price differentials. Energy Economics, 32(2),
334–342.
6. Gülen, S. G. (1997). Regionalization in the world crude oil market. Energy Journal, 18(2),
109–126.
7. Gülen, S. G. (1999). Regionalization in the world crude oil market: Further results. Energy
Journal, 20(1), 125–139.
8. Hammoudeh, S., Thompson, M., & Ewing, B. (2008). Threshold cointegration analysis of
crude oil benchmarks. Energy Journal, 29(4), 79–95.
9. Ji, Q., & Fan, Y. (2016). Evolution of world crude oil market integration: A graph theory
analysis. Energy Economics, 53, 90–100.
10. Jia, X., An, H., Sun, X., Huang, X., & Wang, L. (2017). Evolution of world crude oil market
integration and diversification: A wavelet-based complex network perspective. Applied Energy,
185(2), 1788–1798.
11. Kleit, A. N. (2001). Are regional oil markets growing closer together? An arbitrage cost
approach. Energy Journal, 22(2), 1–15.
12. Nakajima, T. (2019). Test for volatility spillover effects in Japan’s oil futures markets by a
realized variance approach. Studies in Economics and Finance, 36(2), 224–239.
13. Nakajima, T., & Toyoshima, Y. (2019). Measurement of connectedness and frequency dynamics
in global natural gas markets. Energies, 12(20), 3927.
14. Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach.
Econometrica, 59(2), 347–370.
15. Nick, S. (2016). The informational efficiency of European natural gas hubs: Price formation
and intertemporal arbitrage. Energy Journal, 37(2), 1–30.
16. Olsen, K. K., Mjelde, J. W., & Bessler, D. A. (2015). Price formulation and the law of one
price in internationally linked markets: An examination of the natural gas markets in the USA
and Canada. Annals of Regional Science, 54(1), 117–142.
17. Osička, J., Lehotský, L., Zapletalová, V., Černoch, F., & Dančák, B. (2018). Natural gas market
integration in the Visegrad 4 region: An example to follow or to avoid? Energy Policy, 112,
184–197.
18. Ren X, Lu Z, Cheng C, Shi Y, Shen J (2019) On dynamic linkages of the state natural gas
markets in the USA: Evidence from an empirical spatio-temporal network quantile analysis.
Energy Economics, 80, 234–245.
19. Reboredo, J. C. (2011). How do crude oil prices co-move? A copula approach. Energy
Economics, 33(5), 948–955.
20. Scarcioffolo, A. R., & Etienne, X. L. (2019). How connected are the U.S. regional natural
gas markets in the post-deregulation era? Evidence from time-varying connectedness analysis.
Journal of Commodity Markets, 15, 100076.
21. Shi, X., Shen, Y., & Wu, Y. (2019). Energy market financialization: Empirical evidence and
implications from East Asian LNG markets. Finance Research Letters, 30, 414–419.
22. Silverstovs, B., L’Hégaret, G., Neumann, A., & Hirschhausen, C. (2005). International market
integration for natural gas? A cointegration analysis of prices in Europe, North America and
Japan. Energy Economics, 27(4), 603–615.
23. Zhang, B. (2019). Are Chinese and international oil markets integrated? International Review
of Economics and Finance, 62, 41–52.
Chapter 4
Hedging Strategy with Futures Contracts

4.1 Introduction

A futures contract is a promise to buy or sell a security at a currently agreed upon

price at a predetermined time in the future. Futures are one of the most representative
derivatives along with options and swaps. Commodity futures are those whose under-
lying assets are specific standardized commodities (e.g., precious metals, agricultural
products, energy, etc.) and are often listed on commodity exchanges. Although they
can be sold and bought within their maturity, they will automatically settle at matu-
rity. The main purposes of trading commodity futures are risk hedging, speculation
(e.g., diversified investment beyond traditional financial securities and trading based
on market forecasts), arbitrage (e.g., pair trading between highly correlated securi-
ties and trading between securities with different maturities considering risks and
interest rates), and procurement only in the case of physical settlement.
For many non-financial companies, commodity futures represent a means to
hedge risk. Trading futures can eliminate uncertainties arising from price fluctua-
tions because they determine future cash flows. We can not only determine profits
in advance but also avoid unacceptable losses by trading futures. For example, firms
often buy crude oil futures to avoid losses caused by future spot price increases when
procuring crude oil. If the crude oil spot price actually rises, the profit obtained by
liquidating the futures can cover the loss from the price increase in the spot market. As
a completely opposite example, when selling electricity, firms often sell electricity
futures to avoid losses due to future spot price declines. If the electricity spot price
actually falls, the profit gained by counter-trading futures can cover the loss from
the actual price drop. Many futures trades have the advantage of lowering hedging
costs because of margin trading, which itself often reduces hedging costs in futures
trades.
If we trade futures to hedge the risk of spot price fluctuations, we must be careful
to curb the diversification of the portfolio return consisting of a spot and its futures.
The ratio of future positions to the spot position that minimizes this variance is the

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 85
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_4
86 4 Hedging Strategy with Futures Contracts

optimal hedge ratio (OHR). The OHR is obtained by dividing the covariance of the
spot and futures returns by the variance of the futures return. Various multivariate
generalized autoregressive conditional heteroscedasticity (GARCH) models have
been proposed to capture the conditional covariance between multiple asset returns
and the conditional variance of each return. Traders can monitor the OHRs calculated
from the estimated multivariate GARCH model to design an optimal portfolio.
Many studies investigate hedging portfolio strategies in the oil and/or gas markets.
Knill et al. [13] indicate that oligopolists with superior knowledge of supply and
demand, such as oil and gas companies, hedge only when they expect unfavorable
events. Ripple and Moosa [17] examine the difference in the hedging effect of crude
oil futures on maturity and reveal that futures hedging is more effective in near-
month contracts. Chang et al. [5] evaluate the OHR and optimal portfolio weights of
crude oil portfolios using a multivariate GARCH model. Chang et al. [6] compare
the performance of five types of multivariate volatility models for the crude oil spot
and futures return series by calculating the OHR and hedging effectiveness (HE)
index. Toyoshima et al. [18] estimate three types of multivariate GARCH models for
crude oil spot and futures markets. The strategy built by the asymmetric dynamic
conditional correlation (ADCC) model has the best variance reduction performance.
Wang et al. [20] examine the performance of an unhedged strategy and the minimum
variance hedging strategies, for which they estimate required parameters from five
types of models that generate constant hedge ratios, six types of bivariate GARCH
models, two types of copula models, and five types of regime-switching models for
18 commodities (e.g., crude oil and natural gas), three currencies, and three stock
indices. Ghoddusi and Emamzadehfard [10] test multiple features of hedging perfor-
mance using upstream and downstream prices in the United States (US) natural gas
market. Hanly [11] reports significant differences between the OHRs and hedging
performance from an analysis of crude oil, petroleum product, and natural gas bench-
marks. They argue that the hedging performance of natural gas futures is less effective
than that of crude oil and heating oil. Wang et al. [19] evaluate hedging perfor-
mance not by the ability to minimize the variance of the hedged portfolio but by the
ability to minimize risk. They employ three constant and seven time-varying hedging
models for the crude oil index. Lv et al. [15] investigate the performance of port-
folios consisting of Chinese petrochemical-related stocks and crude oil indices in
terms of risk and return. Furió and Torró [9] design optimal hedging strategies with
natural gas futures based on the expected utility maximization approach. Their study
reveals that United Kingdom (UK) natural gas futures have a significant seasonally
variable futures bias. Li et al. [14] reveal that the hedging effect is high due to the
strong correlations between Chinese crude oil spots and futures, and confirm that the
Chinese crude oil futures market, launched in 2018, realizes its main function and
launching aim.
This chapter considers the procurement of natural gas in the US and the UK. We
adopt the Henry Hub (HH) in the US and the national balancing point (NBP) in the
UK. We estimate the three types of bivariate GARCH models of spot returns and
future returns for each natural gas market and calculate the OHR and HE indices.
We use the diagonal VECH model proposed by Bollerslev et al. [3], diagonal BEKK
4.2 Data 87

model proposed by Baba et al. [1] and Engle and Kroner [8], and constant condi-
tional correlation (CCC) model proposed by Bollerslev [2], which are the basic
representative multivariate GARCH models. A comparison of each HE shows that
the portfolio constructed on the OHR calculated by the diagonal BEKK model has
the largest hedging effect for both the HH and NBP.
The remainder of this chapter is organized as follows. Section 4.2 provides the
price series of natural gas spots and futures in the US and UK markets. Section 4.3
describes the OHR and HE. Section 4.4 explains the multivariate GARCH model.
Section 4.5 presents the results of the analysis. Section 4.6 expresses the overall
concluding remarks and considerations.

4.2 Data

We use the HH and NBP as natural gas price indices to represent the US and UK
markets, respectively. We use the spot and futures prices from January 2, 2015 to
December 30, 2020. HH and NBP prices are expressed in USD and GBP per mmbtu,
respectively. We obtain these daily data from Bloomberg.
Before the analysis, we confirm the representative statistics to provide an overview
of the data. Table 4.1 lists the descriptive statistics of the price series. Because we
extract only the days for which both futures and spot prices are available, we have
1484 and 1517 observations for the HH and NBP, respectively. Both the mean and
median of each future are higher than those of each spot. Contangos occur in both
the HH and NBP markets. We can infer that supply and demand were not very tight
during this period. The medians of both the HH futures and spot were higher than
their respective means. Therefore, we can expect many outliers in the left tail of
each distribution. On the other hand, the means of both the NBP futures and spot are
almost the same as their respective medians, indicating that we can expect a roughly
equivalent number of outliers on the left and right tails. Although their spot price
series means are lower than their futures price series means, their spot maximum
values are much larger than their futures maximum values. We can see common
price spikes in the energy spot markets. For both the HH and NBP, the futures price
series have a narrower range and a smaller standard deviation than the spot price
series. That is, their futures price series are more scattered than their spot price series
are. All the variables have positive skewness and a distribution with a long right tail.
All variables had kurtosis greater than 3 and had a distribution with a sharp peak and
long fat-tails. The Jarque–Bera test rejects the normal distribution null hypothesis for
the HH future, HH spot, and NBP spot price series, while accepting the hypothesis
for the NBP future price series at the 10% significance level.
Figure 4.1 shows the time plots of the HH futures and spot prices. There is almost
no separation between the futures and spot prices. However, only the spot price series
has large price spikes that occurred in January 2018 and March 2019. In November
2016 and September 2020, spot prices were slightly lower than futures prices.
88 4 Hedging Strategy with Futures Contracts

Table 4.1 Descriptive statistics

HH NBP
Futures Spot Futures Spot
Period From 1/2/2015 to 12/30/2020
Observations 1484 1484 1517 1517
Mean 2.656 2.638 0.410 0.404
Median 2.718 2.717 0.410 0.402
Maximum 4.837 7.135 0.777 2.300
Minimum 1.482 1.330 0.083 0.073
Standard deviation 0.501 0.562 0.132 0.146
Skewness 0.253 0.904 0.127 1.652
Kurtosis 4.280 8.639 3.078 21.777
Jarque–Bera 117 (0.00) 2168 (0.00) 4.428 (0.12) 22,976 (0.00)
Note p-values are in parentheses

Fig. 4.1 HH spot and futures prices

Figure 4.2 shows the time plots of the NBP futures and spot prices. The price
difference between the futures and its spot tends to be smaller than that of the HH. We
observe a large spike in spot prices in March 2018. Between August and December
2019, its spot prices have a downward trend separate from their future prices.
4.3 Optimal Hedge Ratio and Hedge Effectiveness 89

Fig. 4.2 NBP spot and futures prices

4.3 Optimal Hedge Ratio and Hedge Effectiveness

To hedge the spot price fluctuation from t − 1 to t, suppose that we hold θt units of
its future for one unit of its spot during that period. The return series for this hedged
portfolio is then

Rt = st − θt f t , (4.1)

where st and f t are the spot and future return series, respectively.
According to Johnson [12], the conditional variance of the return series is

var(Rt |It−1 ) = var(st |It−1 ) − 2θt cov(st , f t |It−1 ) + θt2 var( f t |It−1 ), (4.2)

where It−1 is the information set available at time t −1, var(st |It−1 ) is the conditional
variance of the spot return series, var( f t |It−1 ) is the conditional variance of the future
return series, and cov(st , f t |It−1 ) are the conditional covariance of the spot and future
return series.
By partially differentiating Eq. (4.2) with respect to θt , we obtain

∂var(Rt |It−1 )
= −2cov(st , f t |It−1 ) + 2θt var( f t |It−1 ). (4.3)
∂θt
90 4 Hedging Strategy with Futures Contracts

If this partial derivative is equal to 0, then the conditional variance of the portfolio
is minimal. Setting Eq. (4.3) equal to zero and solving for θt , the O H R t conditional
on It−1 is

cov(st , f t |It−1 )
O H R t = (θt |It−1 ) = . (4.4)
var( f t |It−1 )

Because both cov(st , f t |It−1 ) and var( f t |It−1 ) are conditional, we cannot easily
calculate the OHR using ordinary least squares (OLS). We must obtain the O H R t by
estimating a multivariate GARCH model. Then, we calculate the following average
OHR:

1 
T
average O H R = O H Rt , (4.5)
T t=1

where T denotes sample size. To compare the performance of the OHRs calculated
using different techniques, Ku et al. [16] propose the HE index. The H E t conditional
on It−1 is

var(st |It−1 ) − var(Rt |It−1 )

H Et = . (4.6)
var(st |It−1 )

We calculate the following average HE:

1 
T
average H E = H Et. (4.7)
T t=1

A higher HE indicates higher hedge effectiveness and larger risk reduction. We

must thus select a hedging strategy that maximizes HE.

4.4 Multivariate GARCH Model

In general, the multivariate GARCH model is expressed as

r t = μt + e t
1/2
et = H t ζ t , (4.8)

where r t is the return vector, μt is the mean vector of the return, H t is the
covariance matrix, and each element of ζ t is generated from an independently and
identically distributed random number following the standard normal distribution
(iid N (0, 1)). Researchers proposed various multivariate GARCH models depending
on the formulation of H t .
4.4 Multivariate GARCH Model 91

In this chapter, we examine a portfolio consisting of two variables. Therefore,

we describe each multivariate GARCH as a bivariate GARCH model. We write the
mean equation as follows, assuming that the lag order is 0:

f t = m f + ε f,t
st = m s + εs,t
 
ε f,t = h f,t z f,t + h f s,t z s,t
 
εs,t = h f s,t z f,t + h s,t z s,t
z f,t ∼ iid N (0, 1)
z s,t ∼ iid N (0, 1), (4.9)

where h f,t is the conditional variance of f t , h s,t is the conditional variance of st , and
h f s,t is the conditional covariance of f t and st .
In the following explanation of each bivariate GARCH model, both the lag order
of the ARCH term and that of the GARCH term in the variance equation are set to
1 for simplicity.

4.4.1 Diagonal VECH Model

Bollerslev et al. [3] propose the following VECH model:

⎡ ⎤ ⎡ f ⎤ ⎡ f ⎤⎡ ⎤
h f,t c a f f a f s f as ε2f,t−1
⎣ h f s,t ⎦ = ⎣ f s c ⎦ + ⎢ ⎥
⎣ f s a f f s a f s f s as ⎦⎣ ε f,t−1 εs,t−1 ⎦
h s,t s
c s
a f s a f s s as εs,t−1
2
⎡ ⎤⎡ ⎤
f
g f f g f s f gs h f,t−1
⎢ fs ⎥
+ ⎣ g f f s g f s f s gs ⎦⎣ h f s,t−1 ⎦. (4.10)
s
g f s g f s s gs h s,t−1

We can understand this model as a simple extension of a univariate GARCH model

to a bivariate GARCH model. However, it has a disadvantage in that the number of
parameters to estimate increases sharply as the number of variables and lag order
increase. Therefore, Bollerslev et al. [3] propose a diagonal VECH model in which
both the ARCH and GARCH matrices are diagonal matrices. We can express this as
follows:
⎡ ⎤ ⎡ f ⎤ ⎡f ⎤⎡ ⎤
h f,t c af 0 0 ε2f,t−1
⎣ h f s,t ⎦ = ⎣ f s c ⎦ + ⎢ ⎥
⎣ 0 f s a f s 0 ⎦⎣ ε f,t−1 εs,t−1 ⎦
h s,t s
c 0 0 s as εs,t−1
2
92 4 Hedging Strategy with Futures Contracts
⎡ ⎤⎡ ⎤
f
g 0 0 h f,t−1
⎢ f ⎥
+⎣ 0 fs
g f s 0 ⎦⎣ h f s,t−1 ⎦. (4.11)
0 0 s gs h s,t−1

The expression as simultaneous equations is

h f,t = f c + f a f ε2f,t−1 + h f,t−1

h s,t = s c + s as εs,t−1
2
+ s gs h s,t−1
h f s,t = fs
c+ fs
a f s ε f,t−1 εs,t−1 + fs
g f s h f s,t−1 . (4.12)

Although the bivariate VECH model requires us to estimate 21 parameters, the

bivariate diagonal VECH model requires us to estimate only 9 parameters: f c, f a f ,
f
g f , s c, s a s , s g s , f s c, f s a f s , and f s g f s .

4.4.2 Diagonal BEKK Model

Baba et al. [1] and Engle and Kroner [8] propose the following BEKK model:

f
h f,t h f s,t c f sc
= fs
h f s,t h s,t c sc
f
af fs
as ε2f,t−1 ε f,t−1 εs,t−1 f
af fs
as
+ T
fs
af s
as ε f,t−1 εs,t−1 εs,t−1
2 fs
af s
as
f fs f fs
gf gs h f,t−1 h f s,t−1 gf gs
+ fs s
T
fs s . (4.13)
gf gs h f s,t−1 h s,t−1 gf gs

The BEKK model has fewer parameters to estimate than the VECH model does.
We must estimate as many as 21 parameters in Eq. (4.10). However, Eq. (4.13)
requires only 11 parameters. This effect becomes very large as the number of variables
increases. Assuming that the number of variables is three and the lag order of both the
ARCH and GARCH matrices are 1, the number of parameters to estimate is 78 and
24 in the VECH and BEKK models, respectively. Moreover, to reduce the number of
parameters to estimate, Baba et al. [1] and Engle and Kroner [8] propose a diagonal
BEKK model in which both the ARCH and GARCH matrices are diagonal matrices.
We can express this model as

f
h f,t h f s,t c f sc
= fs
h f s,t h s,t c sc
f
af 0 ε2f,t−1 ε f,t−1 εs,t−1 f
af 0
+ T
0 s as ε f,t−1 εs,t−1 εs,t−1
2
0 s as
4.4 Multivariate GARCH Model 93

f f
gf 0 h f,t−1 h f s,t−1 T gf 0
+
0 s gs h f s,t−1 h s,t−1 0 s gs
⎡  2 ⎤
c c f fs
⎣
f
a ε f,t−1a f as ε f,t−1 εs,t−1 ⎦
f s
= fs + f
s 2
c sc f
a f as ε f,t−1 εs,t−1
s
as εs,t−1
⎡ 2 ⎤
f f s
g f h f,t−1 g f gs h f s,t−1 ⎦
+⎣  2 . (4.14)
f
g f s gs h f s,t−1 s gs h s,t−1

We calculate Eq. (4.14) and express it using the following simultaneous equations:

h f,t = f c + f a 2f ε2f,t−1 + f g 2f h f,t−1

h s,t = s c + s as2 εs,t−1
2
+ s gs2 h s,t−1
h f s,t = fs
c + f a f s as ε f,t−1 εs,t−1 + f a f s as ε f,t−1 εs,t−1 . (4.15)

Although the bivariate BEKK model requires us to estimate 11 parameters, the

bivariate diagonal VECH model requires only 7 parameters: f c, f a f , f g f , s c, s a s ,
s
g s , and f s c.

4.4.3 CCC Model

Bollerslev [2] proposes a CCC model in which the correlation matrix is constant,
although each diagonal component of H t , that is, the variance of each variable,
is conditional. The off-diagonal components of H t are 0. When we represent the
variance of each variable using a univariate GARCH model with all lag orders of
1, we can write the CCC model as the following equations using the correlation
coefficient f s c between f t and st .:

h f,t = f c + f a f ε2f,t−1 + f g f h f,t−1

h s,t = s c + s as εs,t−1
2
+ s gs h s,t−1

h f s,t = f s c h f,t h s,t . (4.16)

The bivariate CCC model requires the estimation of only 7 parameters: f c, f a f ,

f
g f , s c, s a s , s g s , and f s c.
94 4 Hedging Strategy with Futures Contracts

4.5 Analysis Results

We estimate the three types of multivariate GARCH models (the diagonal VECH
model, diagonal BEKK, and CCC model) for the HH and NBP, and calculate the
average OHR and HE.

4.5.1 HH Market

First, we generate the HH futures and spot return series. Table 4.2 shows the descrip-
tive statistics for each variable. For these two variables, the mean is positive, whereas
the median is negative. This result implies that these distributions have a larger
number of positive outliers than negative ones. As Fig. 4.1 shows, because the spot
price series has large spikes, its return series has an extremely large maximum value
and a very small minimum value. These ranges and standard deviations indicate that
the spot return series tends to fluctuate more than the future series. These distribu-
tions have a positive skewness, implying that these distributions have a long right
tail. These distributions have kurtosis values greater than 3, meaning that these distri-
butions have sharp peaks and fat-tails. According to the Jarque–Bera statistics, both
series are not significantly normally distributed in terms of skewness and kurtosis.
Figure 4.3 plots the HH spot and futures return series. Referring to Fig.4.1, we can
observe that when the spot price series spikes, the spot return series also spikes and
then plummets. Throughout this period, the fluctuations in the futures return series
are not as large as in the spot return series.
Table 4.3 shows the estimation results for each multivariate GARCH model for the
HH. Each parameter of the variance equations is significant at the 1% level, whereas
the parameters of these mean equations are not significant, even at the 10% level.

Table 4.2 Descriptive

Futures Spot
statistics (HH return series)
Observations 1483 1483
Mean 0.036% 0.124%
Median – 0.042% – 0.055%
Maximum 21.9% 101.6%
Minimum – 16.5% – 35.5%
Standard deviation 0.032 0.056
Skewness 0.625 5.119
Kurtosis 8.311 90.092
Jarque–Bera 1839 (0.00) 475,171 (0.00)
Note p-values are in parentheses
4.5 Analysis Results 95

Fig. 4.3 HH spot and futures return series

Table 4.3 Estimated multivariate GARCH model (HH)

Diagonal VECH Diagonal BEKK CCC
Mean equation mf 6.11 × 10–4 (0.31) 6.75 × 10–4 (0.28) 4.64 × 10–4 (0.42)
ms 3.94 × 10–4 (0.53) 1.73 × 10–4 (0.78) 2.26 × 10–4 (0.71)
Variance equation fc 2.62 × 10–5 (0.00) 1.32 × 10–5 (0.00) 1.07 × 10–5 (0.00)
fa
f 1.09 × 10–1 (0.00) 3.20 × 10–1 (0.00) 1.01 × 10–1 (0.00)
f gf 8.71 × 10–1 (0.00) 9.49 × 10–1 (0.00) 8.96 × 10–1 (0.00)
sc 8.82 × 10–5 (0.00) 5.76 × 10–5 (0.00) 4.97 × 10–5 (0.00)
sa
s 3.21 × 10–1 (0.00) 5.36 × 10–1 (0.00) 3.76 × 10–1 (0.00)
sg
s 7.02 × 10–1 (0.00) 8.68 × 10–1 (0.00) 7.01 × 10–1 (0.00)
f sc 4.05 × 10–5 (0.00) 1.75 × 10–5 (0.00) 4.45 × 10–1 (0.00)
f sa
fs 7.65 × 10–2 (0.00)
fsg
fs 8.32 × 10–1 (0.00)
Note p-values re in parentheses

Figure 4.4 plots the covariance series calculated using each multivariate GARCH
model. Any model can represent the conditional covariance between these return
series. Moreover, these three covariance series fluctuate almost in tandem.
Figure 4.5 plots the OHR series calculated using the estimated diagonal VECH
model. The OHR series is about 50% throughout this period. Referring to Fig. 4.1,
we can see that the OHR series soars in December 2015 when both the futures and
96 4 Hedging Strategy with Futures Contracts

Fig. 4.4 Covariance series between futures and spot prices (HH)

spot price series spike, January 2018 when only the spot price series spikes, and
October 2020 when only the futures price series spikes. The OHR series plummets
to the negative in September 2020 when the spot price series separates and moves
downward relative to the futures price series. Referring to Fig. 4.4, we can see that
the OHR series is synchronized with the covariance series.
Figure 4.6 depicts the time plots of the OHR series calculated using the estimated
diagonal BEKK model. The OHR series was approximately 50% during this period.
Referring to Fig. 4.1, we can observe that the OHR series spikes in December 2015
when both the futures and spot price series spike, January 2018 when only the spot
price series spikes, and October 2020 when only the futures price series spikes. The
OHR series dips into the negative in September 2020 when the spot price series
separates and moves downward relative to the futures price series. This result is
almost the same as that of the diagonal VECH model. However, the OHR in the
above four periods is approximately twice that of the diagonal VECH. Figures 4.4
and 4.6 indicate that the OHR series is linked to the covariance series.
Figure 4.7 displays the OHR series calculated using the estimated CCC model.
The OHR series was approximately 60% throughout this period, although some
spikes occurred. These spikes are even larger than in the diagonal BEKK model.
There are no negative OHR values, which appear in the OHR series calculated by
the diagonal VECH and diagonal BEKK models. Referring to Fig. 4.1, we see that
the OHR series soars in December 2015 when both the futures and spot price series
spike, January 2018 when only the spot price series spikes, and October 2020 when
only the futures price series spikes. This is the same as the OHR series calculated
4.5 Analysis Results 97

Fig. 4.5 OHR calculated using the estimated diagonal VECH model for the HH

Fig. 4.6 OHR calculated using the estimated diagonal BEKK model for the HH
98 4 Hedging Strategy with Futures Contracts

Fig. 4.7 OHR calculated using the estimated CCC model for the HH

Table 4.4 Average OHR and

Average OHR (%) Average HE (%)
HE (HH)
Diagonal VECH 56.7 21.4
Diagonal BEKK 57.7 22.7
CCC 65.9 19.8
Note Bold indicates the average HE for the optimal hedging
strategy

using the diagonal VECH and BEKK models. In addition, the OHR series spikes in
March 2015 and March 2019, when only the spot price series spikes were small, and
in November 2020, when only the spot price series plummets. From Fig. 4.4, we can
see that the OHR series is synchronized with the covariance series.
Table 4.4 lists the average OHR and HE calculated using each multivariate
GARCH model. A portfolio built using the OHR obtained by the diagonal BEKK
model has the largest hedging effect. Therefore, we can conclude that it is the best
hedging strategy among these three. However, the average HE is 22.7%, which is
not so large.

4.5.2 NBP Market

Table 4.5 shows the descriptive statistics for the NBP futures and spot return series.
For these two return series, the means are larger than the medians, implying that these
4.5 Analysis Results 99

Table 4.5 Descriptive

Futures Spot
statistics (NBP return series)
Observations 1516 1516
Mean 0.065% 0.168%
Median – 0.027% 0.000%
Maximum 40.9% 91.7%
Minimum – 14.0% – 66.1%
Standard deviation 0.034 0.056
Skewness 1.921 2.551
Kurtosis 21.694 68.214
Jarque–Bera 23,007 (0.00) 270,282 (0.00)
Note p-values are in parentheses

distributions have a larger number of outliers in the right tail than in the left tail. As
Fig. 4.2 illustrates, because the spot price series has a large spike, its return series
has extreme minimum and maximum values. The range of the spot return series is
wider than that of the futures return series and the standard deviation of the spot
return series is larger than that of the futures return series. This result implies that
the spot return series tends to be more volatile than the futures return series. Both
distributions had positive skewness, implying that each distribution has a long right
tail. These distributions had kurtosis values of over 3, indicating that each distribution
has a sharp peak and fat-tail. The Jarque–Bera test rejects the normal distribution
hypothesis for both series.
Figure 4.8 plots the NBP spot and futures return series. In Fig. 4.2, we can observe
that when the spot price series spikes, the spot return series also spikes.
Table 4.6 shows the estimation results for each multivariate GARCH model for
the NBP return series. All the parameters of these variance equations are statistically
significant at the 1% level, whereas all the parameters of the mean equations are not
statistically significant, even at the 10% level.
Figure 4.9 plots the covariance series calculated using each multivariate GARCH
model. Each model can represent the conditional covariance between these return
series, and these covariance series are synchronized.
Figure 4.10 plots the OHR series calculated using the estimated diagonal VECH
model. The OHR series is about 75% throughout this period. Referring to Fig. 4.2,
we can see that the OHR series plummets to about –80% in March 2018 when only
the spot price series spikes. The OHR series is negative in August 2016 and October
2019 when the spot price series separates and moves downward relative to the futures
price series. The OHR series spikes to nearly 200% in June 2017 when only the spot
price series swoops down.
Figure 4.11 plots the OHR series calculated using the estimated diagonal BEKK
model. The OHR series is about 75% throughout this period. In Fig. 4.2, we can see
that the OHR series plummets to about –200% in March 2018 when only the spot
price series spikes. The OHR series plummets to the negative in August 2016 and
100 4 Hedging Strategy with Futures Contracts

Fig. 4.8 NBP spot and futures return series

Table 4.6 Estimated multivariate GARCH model (NBP)

Diagonal VECH Diagonal BEKK CCC
Mean equation mf 4.82 × 10–4 (0.37) 5.25 × 10–4 (0.34) 3.48 × 10–4 (0.48)
ms 3.17 × 10–4 (0.59) 5.70 × 10–4 (0.34) 3.79 × 10–4 (0.53)
Variance equation fc 8.65 × 10–6 (0.00) 1.31 × 10–5 (0.00) 4.86 × 10–6 (0.00)
fa
f 8.09 × 10–2 (0.00) 3.03 × 10–1 (0.00) 9.31 × 10–2 (0.00)
f gf 9.16 × 10–1 (0.00) 9.52 × 10–1 (0.00) 9.11 × 10–1 (0.00)
sc 4.88 × 10–5 (0.00) 7.68 × 10–5 (0.00) 4.51 × 10–5 (0.00)
sa
s 2.11 × 10–1 (0.00) 5.71 × 10–1 (0.00) 2.82 × 10–1 (0.00)
sg
s 7.82 × 10–1 (0.00) 8.27 × 10–1 (0.00) 7.33 × 10–1 (0.00)
f sc 1.50 × 10–5 (0.00) 3.55 × 10–5 (0.00) 6.13 × 10–1 (0.00)
f sa
fs 8.01 × 10–2 (0.00)
fsg
fs 8.95 × 10–1 (0.00)
Note p-values are in parentheses

August and October 2019 when the spot price series separates and moves downward
relative to the futures price series. The OHR series spikes to over 200% four times.
Comparing Fig. 4.10 with Fig. 4.11, we can see that the OHR series calculated using
the diagonal BEKK model oscillates more frequently and greatly than that from the
diagonal BEKK model.
4.5 Analysis Results 101

Fig. 4.9 Covariance series between futures and spot prices (NBP)

Fig. 4.10 OHR calculated using the estimated diagonal VECH model for the NBP
102 4 Hedging Strategy with Futures Contracts

Fig. 4.11 OHR calculated using the estimated diagonal BEKK model for the NBP

Figure 4.12 plots the OHR series calculated using the estimated CCC model.
Throughout this period, the OHR series is about 80% and is not negative. From
Fig. 4.2, we can see that the OHR series spikes to over 300% in June 2017, when
only the spot price series plummets, and in March 2018, when only the spot price
series spikes. Besides these events, the OHR series spikes to over 200% four times.
Table 4.7 shows the average OHR and HE calculated using these three multi-
variate GARCH models. If we construct a portfolio using the OHR obtained from
the diagonal BEKK model, we obtain the highest average HE, of 42.4%. Therefore,
we can conclude that this is the best hedging strategy among these three.

4.6 Concluding Remarks

When firms procure energy with a high price fluctuation risk, they often trade futures
to hedge the risk. Such firms must work to curb the volatility of the portfolio return,
which consists of a spot and its futures. By dividing the covariance of the spot and
futures return series by the variance of the future return series, we can obtain the ratio
of futures positions to spot positions that minimizes volatility, which is defined as the
OHR. In other words, by estimating the multivariate GARCH model that formulates
the conditional covariance and variance, we can obtain the conditional OHR series,
which helps us construct a timely optimal portfolio.
4.6 Concluding Remarks 103

Fig. 4.12 OHR calculated using the estimated CCC model for the NBP

Table 4.7 Average OHR and

Average OHR (%) Average HE (%)
HE (NBP)
Diagonal VECH 77.2 41.0
Diagonal BEKK 77.9 42.4
CCC 84.5 37.6
Note Bold indicates the average HE for the optimal hedging
strategy

This chapter estimates three types of bivariate GARCH models consisting of the
spot and futures return series in the US and UK natural gas markets to calculate
OHR and HE. We adopt the diagonal VECH, diagonal BEKK, and CCC models as
multivariate GARCH models. The OHR series fluctuates drastically depending on
the spot and futures market conditions. All the multivariate GARCH models here
can capture the time dependence of the covariance and variance in the same way.
However, comparing the average HE values reveals that constructing the portfolio
using the diagonal BEKK model is the best hedging strategy for both the HH and
NBP markets.
While this chapter adopts primitive multivariate models (i.e., the diagonal VECH
model, diagonal BEKK model, and CCC model), various multivariate GARCH
models exist. Section 4.4 explains the VECH and BEKK models. Advancing the
CCC model, Engle [7] proposes a dynamic conditional correlation (DCC) model that
assumes that the correlation coefficient is conditional. Moreover, Cappiello et al. [4]
104 4 Hedging Strategy with Futures Contracts

propose the ADCC model, which incorporates the asymmetry in which the corre-
lation tends to be stronger after a negative return than after a positive return. We
recommend that readers consider the optimal hedging strategy by applying these
various methods to compare the HE values for the securities and commodities of
interest.

References

1. Baba, Y., Engle, R. F., Kraft, D., & Kroner, K. F. (1987). Multivariate simultaneous generalized
ARCH. Unpublished manuscript, Department of Economics, University of California.
2. Bollerslev, T. (1990). Modelling the coherence in short-run nominal exchange rates: A
multivariate generalized ARCH model. Review of Economics and Statistics, 72(3), 498–505.
3. Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1988). A capital asset pricing model with
timevarying covariances. Journal of Political Economy, 96(1), 116–131.
4. Cappiello, L., Engle, R. F., & Sheppard, K. (2006). A symmetric dynamics in the correlations
of global equity and bomd returns. Journal of Financial Econometrics, 4(4), 537–572.
5. Chang, C. L., McAleer, M., & Tansuchat, R. (2010). Analyzing and forecasting volatility
spillovers, asymmetries and hedging in major oil markets. Energy Economics, 32(6), 1445–
1455.
6. Chang, C. L., McAleer, M., & Tansuchat, R. (2011). Crude oil hedging strategies using dynamic
multivariate GARCH. Energy Economics, 33(5), 912–923.
7. Engle, R. F. (2002). Dynamic conditional correlation: A simple class of multivariate general-
ized autoregressive conditional heteroskedasticity models. Journal of Business & Economics
Statistics, 20(3), 339–350.
8. Engle, R. F., & Kroner, K. F. (1995). Multivariate simultaneous generalized ARCH. Economic
Theory, 11(1), 122–150.
9. Furió, D., & Torró, H. (2020). Optimal hedging under biased energy futures markets. Energy
Economics, 88, 104750.
10. Ghoddusi, H., & Emamzadehfard, S. (2017). Optimal hedging in the US natural gas market:
The effect of maturity and cointegration. Energy Economics, 63, 92–105.
11. Hanly, J. (2017). Managing energy price risk using futures contracts: A comparative analysis.
Energy Journal, 38(3), 93–112.
12. Johnson, L. L. (1960). The theory of hedging and speculation in commodity futures. Review
of Economic Studies, 27(3), 139–151.
13. Knill, A. M., Minnick, K., & Nejadmalayeri, A. (2006). Selective hedging, information
asymmetry, and futures prices. Journal of Business, 79(3), 1475–1501.
14. Li, J., Huang, L., & Li, P. (2021). Are Chinese crude oil futures good hedging tools? Finance
Research Letters, 38, 101514.
15. Lv, F., Yang, C., & Fang, L. (2020). Do the crude oil futures of the Shanghai International Energy
Exchange improve asset allocation of Chinese petrochemical-related stocks? International
Review of Financial Analysis, 71, 1–1537.
16. Ku, Y. H. H., Chen, H. C., & Chen, K. H. (2007). On the application of the dynamic conditional
correlation model in the estimating optimal time-varying hedge ratios. Applied Economics
Letters, 14(7), 503–509.
17. Ripple, R. D., & Moosa, I. A. (2007). Hedging effectiveness and futures contract maturity: The
case of NYMEX crude oil futures. Applied Financial Economics, 17(9), 683–689.
18. Toyoshima, Y., Nakajima, T., & Hamori, S. (2013). Hedging strategy: New evidence from the
data of the financial crisis. Applied Financial Economics, 23(12), 1033–1041.
19. Wang, Y., Geng, Q., & Meng, F. (2019). Futures hedging in crude oil markets: A comparison
between minimum-variance and minimum-risk frameworks. Energy, 181(15), 815–826.
20. Wang, Y., Wu, C., & Yang, L. (2015). Hedging with futures: Does anything beat the naïve
hedging strategy? Management Science, 61(12), 2870–2889.
Chapter 5
Market Risk of a Power Generation
Business

5.1 Introduction

Power generation businesses have unstable profit environments. While the volatility
of fuel and electricity markets is extremely high, companies must make large-scale
capital investments that take a long time from decision making to the start of operation
and long-term mass fuel procurement under take-or-pay contracts. Therefore, the
daily marginal profit (total electricity sales minus total fuel costs) is unstable, and
can even be negative. In addition, the amount of electricity sold, that is, the amount
of fuel required, is uncertain. It is extremely difficult to maximize profits while
minimizing risk.
In the long run, such firms should focus on optimizing power generation equip-
ment, long-term contract fuel procurement, and long-term contract power sales port-
folios. Assuming some future cases that consider both energy indicators (e.g., price,
demand, and supply) and general economic indicators (e.g., inflation rate, interest
rate, and exchange rate), we should aim to design an optimum portfolio by adjusting
business parameters (e.g., the ratio of each fuel type, performance of each gener-
ator, ratio of long-term contracts, and price determination formula). In the short
term, as a prerequisite for existing equipment and long-term contracts, measures
should be taken (e.g., strengthening equity capital, increasing accounting allowances,
and restructuring portfolios through short-term trading contracts) that consider only
normal market price fluctuations as a risk factor. For this purpose, it is necessary to
monitor the spillover between markets related to the power generation business and
measure market risk in business operations appropriately.
Several existing studies analyze the relationship between the value and/or risk of
energy-related business companies and other indexes. Reboredo [8] examines the
systemic risk and dependence between oil prices and six clean energy stock price
indexes using copulas to express the dependence structure and measure the condi-
tional value-at-risk (VaR). Siburg et al. [9] propose forecasting the VaR of bivariate
portfolios using copulas calibrated using nonparametric sample estimates of the

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 105
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_5
106 5 Market Risk of a Power Generation Business

coefficient of lower tail dependence. We demonstrate the superiority of the proposed

model over the conventional parametric model by analyzing the benchmarks of elec-
tricity, crude oil, natural gas, and coal, and the equities of five global electric utility
service providers. Boubaker and Sghaier [2] propose a Markov-switching copula
model to investigate the presence of regime changes in the time-varying dependence
structure between crude oil benchmark price returns and stock price index returns
in six Gulf Cooperation Council countries. Zhang et al. [11] introduce the volatility
threshold dynamic conditional correlations approach, which is an extension of the
dynamic conditional correlation (DCC) and asymmetric DCC (ADCC), to investi-
gate the spillover of stock market volatility indexes on oil and gas markets. Ji et al. [5]
introduce six time-varying copulas to measure four types of delta conditional VaRs
and examine the impact of uncertainty on crude oil, natural gas, and clean energy
companies. This study considers three proxies for economic policy uncertainty, finan-
cial market uncertainty, and energy market uncertainty and reveals the magnitude
and asymmetric effects of their influence. Hanif et al. [4] examine the dependence
structure of major energy firm equities in the United States and the European Union
(EU). This study discusses the effect on the sensitivity of energy equity portfolios to
crude oil prices to the selection of bivariate or multivariate copulas. Zhang et al. [12]
analyze both the return and volatility spillover between energy and electricity utility
stock markets and conclude that investors should monitor current economic events to
hedge their risks through proper portfolio diversification. Wu et al. [10] construct a
total systemic risk index for global energy companies and investigate whether stock
market volatility, energy market risks, and exchange rate risks are factors driving
this index. Liu et al. [6] calculate the spillover index between a newspaper-based
index that reflects uncertainty in the stock market caused by infectious diseases and
three renewable energy stock indices. Mzoughi et al. [7] study the dependence struc-
ture and risk spillover between green financial securities and the energy commodity
index.
This chapter focuses on the market risk in the power generation business for short-
term risk management. As an analysis case, we take an electric power company in
the European market. The company procures three types of fuel at their respective
prices linked to representative crude oil, natural gas, and coal futures prices and sells
electricity at prices linked to typical wholesale electricity futures. Furthermore, the
company buys carbon credit futures. First, we provide an overview of the descriptive
statistics and time plots of each futures price. Second, we measure the spillover
effects of both the return and volatility series between these price indices as well as
the total connectedness. Third, we decompose them spectrally. Finally, we measure
the market risk of this business using these futures markets. Although this example is
extremely simple, it can clarify the market risk of actual power generation companies
as an extension of this study.
The remainder of this chapter is organized as follows. Section 5.2 explains
the methodologies. Section 5.3 presents the data and the preliminary analyses.
Section 5.4 provides the results of the main analyses and measurements. Section 5.5
discusses the overall concluding remarks and considerations.
5.3 Data and Preliminary Analyses 107

5.2 Methodology

We examine the spillover effect among five markets, namely the crude oil, natural
gas, coal, electricity, and carbon credit markets. Then, we measure the risk of a
portfolio consisting of these five indices.
First, we confirm the descriptive statistics and time plots for these price series as an
overview. Then, we generate the return (see Sect. 3.4.1) and volatility series using the
exponential generalized autoregressive conditional heteroscedasticity (EGARCH)
model (see Sect. 3.3.3) and confirm the descriptive statistics and time plots for these
series. We test the stationarity of these variables using the augmented Dickey Fuller
(ADF) unit root test (see Sect. 2.2.2) to confirm whether we can represent these
variables by the vector moving average (VMA) model.
Second, by measuring the connectedness proposed by Diebold and Yilmaz [3]
(see Sect. 3.3.1) and spectrally decomposing the connectedness as in Baruník and
Křehlík [1] (see Sect. 3.3.2), we grasp the spillover effects of returns and volatility
among these markets.
Finally, estimating four types of copulas, (the Gaussian, t, Clayton, and Gumbel
copulas, see Sect. 2.5.2), we measure the risk of the portfolio. This portfolio consists
of long positions in crude oil, natural gas, coal, and carbon credits, and short positions
in electricity. The risk of a long position is in the left tail of the return distribution, and
the risk of a short position is in the right tail of the return distribution. We generate
the distribution of daily returns for the portfolio by simulating each estimated copula
500,000 times. We then calculate the VaR and expected shortfall.

5.3 Data and Preliminary Analyses

This chapter uses an index to represent each European commodities market. Specif-
ically, we adopt Brent futures as the crude oil index, Title Transfer Facility (TTF)
futures as the natural gas index, Rotterdam futures as the coal index, French baseload
(FrenchBL) futures as the electricity index, and EU allowance (EUA) futures as the
carbon credit index. We use daily data from January 1, 2015 to December 31, 2021,
though we extract only the days on which all five prices are available. The Brent, TTF,
Rotterdam, FrenchBL, and EUA prices are expressed in EUR per barrel, mmbtu, ton,
MWh, and tonneCO2 , respectively. We obtain all the data from Bloomberg.

5.3.1 Price Series

We confirm the representative statistics for all the price series. Table 5.1 summarizes
the descriptive statistics. For four indexes (TTF, Rotterdam, FrenchBL, and EUA)
other than Brent, the mean is larger than the median. These distributions contain
108 5 Market Risk of a Power Generation Business

Table 5.1 Descriptive statistics (price series)

Brent TTF Rotterdam FrenchBL EUA
Period From 1/2/2015 to 12/31/2021
Observations 502 502 502 502 502
Mean 50.53 20.81 65.64 57.24 19.79
Median 52.40 17.52 55.69 44.55 15.98
Maximum 74.94 180.27 236.54 772.11 88.88
Minimum 17.83 3.51 34.69 18.58 3.93
Standard deviation 10.78 17.56 26.60 57.09 16.79
Skewness – 0.23 4.12 2.48 5.91 1.48
Kurtosis 2.64 23.42 11.88 50.29 4.92
Jarque–Bera 25 (0.00) 36,139 (0.00) 7702 (0.00) 177,117 (0.00) 927 (0.00)
Note p-values are in parentheses

many outliers in the right tail. Although the range of these four variables was wide,
the standard deviation was not large. The outliers in the right tail may be accidental.
These four distributions have a positive skewness and a long right tail, consistent
with the fact that these four variables have maximum values much larger than their
respective means. These four distributions had kurtosis values greater than 3, meaning
that their distributions have sharp peaks and long fat-tails. Brent has the opposite
characteristics. Its median is larger than its mean and its skewness is negative. Its
distribution contained many outliers in the long left tail. This is consistent with the
difference between its maximum and mean being smaller than the difference between
its mean and minimum. The distribution had a kurtosis of less than 3, indicating
that the distribution has a rounded peak and short, thin tails. For all variables, the
Jarque–Bera test rejected the normal distribution hypothesis.
Figure 5.1 shows the time plots of the five prices. In the short period after the
spring of 2021, the four markets besides Brent rose sharply. These time plots are
consistent with the descriptive statistics.

5.3.2 Return Series

Table 5.2 shows the basic statistics for these five return series. Similar to the price
series results, only Brent is characteristic. For TTF, Rotterdam, FrenchBL, and EUA,
the mean was larger than the median. This result implies that these distributions have
more outliers in the right tail than in the left tail. On the other hand, the distribution
of the Brent return series had a smaller mean than the median and more outliers
in the left tail than in the right tail. The standard deviation of the FrenchBL return
series was relatively large and the range was relatively wide. We can conclude that
FrenchBL was the riskiest commodity during this period. The distributions of the
5.3 Data and Preliminary Analyses 109

Fig. 5.1 Commodity price series, Europe

Brent, Rotterdam, and EUA return series have a negative skewness, implying a long
left tail. In contrast, the skewness for TTF and FrenchBL is positive, indicating a
long right tail. Each return series had a kurtosis larger than 3. This result means
that each distribution had a sharp peak and a long fat-tail. The Jarque–Bera test
rejects the hypothesis that each series has skewness and kurtosis following a normal
distribution.

Table 5.2 Descriptive statistics (return series)

Brent TTF Rotterdam FrenchBL EUA
Observations 1788 1788 1788 1788 1788
Mean 0.057% 0.136% 0.068% 0.196% 0.178%
Median 0.115% − 0.037% 0.043% 0.000% 0.134%
Maximum 21.9% 37.3% 19.1% 71.9% 13.6%
Minimum − 25.1% − 23.3% − 41.8% − 50.9% − 17.7%
Standard 0.0267 0.0371 0.0210 0.0458 0.0289
deviation
Skewness − 0.24 1.36 − 3.83 3.57 − 0.24
Kurtosis 17.21 17.19 108.29 78.85 6.64
Jarque–Bera 15,066 (0.00) 15,555 (0.00) 830,296 (0.00) 432,403 (0.00) 1001 (0.00)
Note p-values are in parentheses
110 5 Market Risk of a Power Generation Business

Figure 5.2 displays the time plots of these return series. As each standard deviation
of the return series indicates, we observe a highly volatile FrenchBL return series and
less volatile Brent, Rotterdam, and EUA return series. The Brent return series is more
volatile during January 2015 to December 2016 and April to June 2020 than during
the other periods. For the TTF, the return series fluctuates more from May 2019 to
December 2021 than in the other periods. The Rotterdam and FrenchBL return series
are stable for some time; however, they often spike. The spikes for FrenchBL were
larger and more frequent than those for Rotterdam. The EUA return series has almost
the same fluctuation range over the entire period. As this figure illustrates, these five
return series do not seem synchronized.
Table 5.3 shows the results of the ADF unit root test for all return series. All
tests adopt a model that has neither a constant term nor time-trend time. All unit
root hypotheses were rejected at the 1% significance level and all return series are
stationary. Therefore, we can represent these vector autoregressive (VAR) model
using the VMA model.

5.3.3 Volatility Series

To measure risk spillover effects, we generate a volatility series for these commodi-
ties. Table 5.4 shows the estimated autoregressive (AR)-EGARCH model. The lag
length of each term was selected based on the Schwarz Bayesian information crite-
rion (SBIC). We determine the lag order of the AR, autoregressive conditional
heteroskedasticity (ARCH), generalized ARCH (GARCH), and asymmetric terms
in this order. The ARCH term (α1 ) and GARCH term (β1 ) are statistically signifi-
cant at the 1% level for all commodities. Each generalized error distribution (GED)
parameter estimate is less than 2, meaning that each error term has a fat-tailed
distribution.
Table 5.5 shows the descriptive statistics of these volatility series. For all these
variables, the mean is larger than the median; that is, these volatility distributions
have more outliers in the right tail than in the left tail. For TTF and FrenchBL,
the mean is relatively large, the maximum is extremely large, and the minimum is
relatively small. These two commodities are riskier than the other three commodities.
All volatility series have a positive skewness; that is, these distributions have a long
right tail. All volatility series have kurtosis greater than 3; that is, these distributions
have a sharp peak and fat-tails. The Jarque–Bera test results reject the hypothesis
that each series has skewness and kurtosis following a normal distribution.
Figure 5.3 displays the time plots of these volatility series. The spikes for the
TTF and FrenchBL volatility series are very high. Although the volatility for Brent,
Rotterdam, and EUA is relatively small, we see some spikes in their series. We cannot
deny the synchronization between these spikes.
We confirm whether Diebold and Yilmaz’s [3] approach is applicable to these
volatility series to measure the spillover effects between these volatility series. The
ADF test, which uses a model with a constant term and without a time-trend term,
5.3 Data and Preliminary Analyses 111

Fig. 5.2 Commodity return series, Europe

112 5 Market Risk of a Power Generation Business

Table 5.3 ADF unit root test

Return series ADF-t statics
results (return series)
Brent − 41.48* (0.00)
TTF − 38.53* (0.00)
Rotterdam − 40.15* (0.00)
FrenchBL − 20.67* (0.00)
EUA − 45.22* (0.00)
Note * indicates rejection of the unit-root hypothesis at the 1%
significance level. p-values are in parentheses

Table 5.4 Estimated AR-EGARCH model

Brent TTF Rotterdam FrenchBL EUA
Mean c0 0.001 (0.01) 0.000 (0.97) 0.000 (0.02) 0.000 (0.75) 0.002 (0.00)
equation c − 0.042 (0.05) 0.051 (0.03) 0.074 (0.00) 0.081 (0.00) − 0.079 (0.00)
1
Variance θ − 0.278 (0.00) − 0.262 (0.00) − 0.124 (0.02) − 0.709 (0.00) − 0.340 (0.00)
equation α 0.147 (0.00) 0.263 (0.00) 0.062 (0.00) 0.191 (0.00) 0.207 (0.00)
1
β1 0.978 (0.00) 0.991 (0.00) 0.990 (0.00) 0.917 (0.00) 0.975 (0.00)
γ1 − 0.095 (0.00) 0.006 (0.71) 0.010 (0.21) − 0.011 (0.58) − 0.010 (0.50)
GED ν 1.235 (0.00) 1.353 (0.00) 0.722 (0.00) 0.687 (0.00) 1.340 (0.00)
Note p-values are in parentheses

Table 5.5 Descriptive statistics (volatility series)

Brent TTF Rotterdam FrenchBL EUA
Observations 1787 1787 1787 1787 1787
Mean 6.42 × 10−4 1.33 × 10−3 2.64 × 10−4 1.17 × 10−3 8.39 × 10−4
Median 4.53 × 10−4 6.89 × 10−4 2.00 × 10−4 7.00 × 10−4 7.03 × 10−4
Maximum 8.84 × 10−3 1.80 × 10−2 1.75 × 10−3 7.33 × 10−2 3.92 × 10−3
Minimum 1.10 × 10−4 7.12 × 10−5 1.11 × 10−4 8.00 × 10−5 1.39 × 10−4
Standard 7.94 × 10−4 1.92 × 10−3 2.12 × 10−4 2.64 × 10−3 5.57 × 10−4
deviation
Skewness 5.31 3.73 4.26 16.74 1.95
Kurtosis 37.6 21.5 23.7 383.6 8.0
Jarque–Bera 97,639 (0.00) 29,595 (0.00) 37,359 (0.00) 10,869,514 2970 (0.00)
(0.00)
Note p-values are in parentheses

rejects the unit root hypothesis for all volatility series. Table 5.6 provides the ADF
unit root test results. The vector of these volatility series is a stationary covariance
process. Therefore, we can represent the VAR model as the VMA model.
5.3 Data and Preliminary Analyses 113

Fig. 5.3 Commodity volatility series, Europe

114 5 Market Risk of a Power Generation Business

Table 5.6 ADF unit root test

Volatility series ADF-t statics
results (volatility series)
Brent − 5.71* (0.00)
TTF − 5.43* (0.00)
Rotterdam − 3.55* (0.01)
FrenchBL − 18.03* (0.00)
EUA − 5.89* (0.00)
Note * indicates rejection of the unit-root hypothesis at the 1%
significance level. p-values are in parentheses

5.4 Analysis Results

This section provides the spillover effects of the return series and the volatility series
between the Brent, TTF, Rotterdam, FrenchBL, and EUA markets. We then present
the results of the spectral decomposition of these effects. Moreover, after estimating
four types of quinquevariate copulas, we measure the market risk of the example
power generation business.

5.4.1 Return Series

Table 5.7 presents the estimated coefficients of the VAR model for the Brent, TTF,
Rotterdam, FrenchBL, and EUA returns series. The lag order is 1, which we selected
based on the SBIC.
Table 5.8 lists the variance–covariance matrix of the residuals of this VAR model.
Figure 5.4 illustrates the spillover indexes, which we obtain from Tables 5.7 and
5.8. Only indexes of more than 5% were extracted. We observe spillover effects
among the natural gas, electricity, and carbon credit markets. Furthermore, the
spillover effects from TTF to Rotterdam and from the EUA to Brent are 5.85 and
5.10%, respectively. Finally, the total connectedness is 18.71%. We can conclude
that the level of market integration is not very high.
We present the spectral analysis of these connectedness indices in Table 5.9. The
total connectedness from 1 to 5 business days, from 6 to 20 business days, and over 21

Table 5.7 VAR model (return series)

Brent (t) TTF (t) Rotterdam (t) FrenchBL (t) EUA (t)
Brent (t-1) 0.023 − 0.022 − 0.001 − 0.049 − 0.039
TTF (t-1) 0.014 0.048 0.079 0.125 − 0.053
Rotterdam (t-1) 0.017 0.076 0.024 0.068 − 0.095
FrenchBL (t-1) − 0.004 0.077 0.008 0.032 − 0.012
EUA (t-1) − 0.029 − 0.032 − 0.040 0.093 − 0.021
5.4 Analysis Results 115

Table 5.8 Residual variance–covariance matrix, VAR model (return series)

Brent TTF Rotterdam FrenchBL EUA
Brent 7.12 × 10−4 1.22 × 10−4 6.53 × 10−5 7.47 × 10−5 1.80 × 10−4
TTF 1.22 × 10−4 1.36 × 10−3 1.71 × 10−4 6.13 × 10−4 3.63 × 10−4
Rotterdam 6.53 × 10−5 1.71 × 10−4 4.23 × 10−4 2.02 × 10−4 1.05 × 10−4
FrenchBL 7.47 × 10−5 6.13 × 10−4 2.02 × 10−4 2.05 × 10−3 3.63 × 10−4
EUA 1.80 × 10−4 3.63 × 10−4 1.05 × 10−4 3.63 × 10−4 8.23 × 10−4

Fig. 5.4 Spillover effects between return series

business days is 14.06, 3.46, and 1.19%, respectively. Since the total connectedness
for the entire period is 18.71%, the short-term factors contribute the most to the
return spillover; the spillover effect is largely dependent on events occurring within
approximately one week. Each spillover effect between these return series tends to
be almost the same as total connectedness. The short-term factor is the largest, the
medium-term factor is the second largest, and the long-term factor is the smallest.

5.4.2 Volatility Series

We estimate the VAR model for the volatility series using a lag order of 1 based on
the SBIC. Table 5.10 reports the estimated results.
116 5 Market Risk of a Power Generation Business

Table 5.9 Spillover index and spectral analysis (return series)

To From Bandwidth
Brent TTF Rotterdam FrenchBL EUA Others (%)
(%) (%) (%) (%) (%)
Brent 72.73 1.06 0.96 0.29 4.22 1.31 0  date  5
TTF 0.93 56.98 2.73 7.42 6.72 3.56
Rotterdam 0.92 3.46 66.71 2.83 2.07 1.85
FrenchBL 0.24 7.84 2.65 58.86 4.58 3.06
EUA 3.96 8.89 2.93 5.62 63.20 4.28
Others 1.21 4.25 1.85 3.23 3.52 14.06
Brent 14.31 0.27 0.25 0.05 0.66 0.25 6  date  20
TTF 0.16 13.37 0.99 2.67 1.55 1.08
Rotterdam 0.21 1.76 14.40 1.07 0.44 0.70
FrenchBL 0.04 2.93 0.91 13.73 1.63 1.10
EUA 0.38 0.69 0.10 0.52 9.97 0.34
Others 0.16 1.13 0.45 0.86 0.86 3.46
Brent 4.79 0.09 0.09 0.02 0.22 0.08 21  date
TTF 0.05 4.59 0.35 0.94 0.54 0.38
Rotterdam 0.07 0.63 4.90 0.39 0.15 0.25
FrenchBL 0.01 1.02 0.32 4.68 0.56 0.38
EUA 0.12 0.19 0.02 0.15 3.27 0.10
Others 0.05 0.39 0.15 0.30 0.29 1.19
Brent 91.83 1.42 1.29 0.35 5.10 1.63 Total
TTF 1.14 74.94 4.07 11.04 8.81 5.01
Rotterdam 1.20 5.85 86.00 4.29 2.66 2.80
FrenchBL 0.29 11.79 3.88 77.27 6.77 4.55
EUA 4.46 9.77 3.06 6.29 76.43 4.71
Others 1.42 5.77 2.46 4.39 4.67 18.71

Table 5.10 VAR model (volatility series)

Brent TTF Rotterdam FrenchBL EUA
Brent (t − 1) 0.968 − 0.000 0.002 0.046 0.015
TTF (t − 1) − 0.001 0.945 0.001 0.060 − 0.002
Rotterdam (t − 1) 0.030 0.285 0.990 0.678 0.069
FrenchBL (t − 1) − 0.000 0.001 − 0.000 0.658 0.001
EUA (t − 1) 0.009 0.002 − 0.001 0.125 0.964
5.4 Analysis Results 117

The variance–covariance matrix of the residuals of this VAR model, which we

require for the Diebold and Yilmaz [3] approach, is shown in Table 5.11.
We depict the spillover indices, which we calculate using Tables 5.10 and 5.11, in
Fig. 5.5. Only indexes of more than 5% are extracted. We can observe spillover effects
from TTF to the EUA and Rotterdam, and from the EUA to Brent. These results are
similar to those for the return series. However, besides these effects, Fig. 5.5 looks
completely different from Fig. 5.4. We should note that return management and risk
management differ. The spillover from Brent to the EUA and from Rotterdam to
TTF is very strong. The connectedness from Rotterdam to the EUA and FrenchBL is
14.91 and 7.17%, respectively. Finally, total connectedness is 20.14%. From a risk
perspective, we can conclude that the market integration is at a medium level.
Table 5.12 provides the results of the spectral decomposition of these connected-
ness indices. The total connectedness of these volatility series is 20.14%; specifically,

Table 5.11 Residual variance–covariance matrix, VAR model (volatility series)

Brent TTF Rotterdam FrenchBL EUA
Brent 4.47 × 10−8 5.90 × 10−11 1.62 × 10−10 5.41 × 10−10 4.47 × 10−9
TTF 5.90 × 10−11 3.15 × 10−7 7.92 × 10−10 1.26 × 10−7 2.51 × 10−9
Rotterdam 1.62 × 10−10 7.92 × 10−10 9.03 × 10−10 6.26 × 10−9 2.65 × 10−10
FrenchBL 5.41 × 10−10 1.26 × 10−7 6.26 × 10−9 3.75 × 10−6 1.62 × 10−8
EUA 4.47 × 10−9 2.51 × 10−9 2.65 × 10−10 1.62 × 10−8 2.53 × 10−8

Fig. 5.5 Spillover effects between volatility series

118 5 Market Risk of a Power Generation Business

0.20% for short-term, 0.56% for medium-term, and 19.38% for long-term factors.
The spillover effects of volatility can be explained primarily by long-term factors.
Most spillover is caused by events that occurred more than one month previously.
Each spillover effect between these volatility series tends to be almost the same as
total connectedness. Events within a week have the least impact, events that occurred
between a month and two weeks prior have the second least impact, and events that
occurred more than one month prior have the largest impact.

Table 5.12 Spillover index and spectral analysis (volatility series)

To From Bandwidth
Brent TTF Rotterdam FrenchBL EUA Others
(%) (%) (%) (%) (%) (%)
Brent 1.70 0.00 0.03 0.00 0.01 0.01 0  date  5
TTF 0.02 2.86 0.13 0.02 0.00 0.03
Rotterdam 0.05 0.12 0.04 0.00 0.00 0.03
FrenchBL 0.01 0.24 0.12 30.06 0.06 0.09
EUA 0.04 0.01 0.13 0.00 1.38 0.03
Others 0.02 0.07 0.08 0.01 0.01 0.20
Brent 6.47 0.02 0.11 0.00 0.05 0.04 6  date  20
TTF 0.08 9.98 0.44 0.10 0.00 0.13
Rotterdam 0.17 0.40 0.13 0.00 0.00 0.12
FrenchBL 0.02 0.40 0.13 34.01 0.13 0.13
EUA 0.17 0.09 0.47 0.01 5.28 0.15
Others 0.09 0.18 0.23 0.02 0.04 0.56
Brent 83.44 0.05 3.00 0.05 5.06 1.63 21  date
TTF 0.80 65.64 17.99 1.74 0.21 4.15
Rotterdam 4.06 12.84 80.62 1.44 0.14 3.69
FrenchBL 1.50 3.60 6.92 21.57 1.23 2.65
EUA 20.74 0.40 14.31 0.84 56.13 7.26
Others 5.42 3.38 8.45 0.81 1.33 19.38
Brent 91.61 0.08 3.14 0.05 5.12 1.68 Total
TTF 0.89 78.48 18.56 1.86 0.21 4.30
Rotterdam 4.27 13.35 80.79 1.44 0.14 3.84
FrenchBL 1.53 4.25 7.17 85.64 1.41 2.87
EUA 20.95 0.50 14.91 0.85 62.79 7.44
Others 5.53 3.64 8.76 0.84 1.38 20.14
5.5 Concluding Remarks 119

Table 5.13 Parameters of each marginal distribution function (return series)

Brent TTF Rotterdam FrenchBL EUA
Mean 5.69 × 10−4 1.36 × 10−3 6.83 × 10−4 1.96 × 10−3 1.78 × 10−3
Standard deviation 2.67 × 10−2 3.71 × 10−2 2.10 × 10−2 4.58 × 10−2 2.89 × 10−2

5.4.3 Risk Measurement

We measure the market risk of a power company that procures crude oil, natural gas,
coal, and carbon credits and sells electricity. Regarding this company as a portfolio
consisting of long positions in fuels and carbon credit futures and short positions
in electricity futures, we calculate the VaR and expected shortfall of this company
through a simulation using random numbers following the four types of quinque-
variate copulas: the Gaussian, t, Clayton, and Gumbel copulas. We selected the VaR
and expected shortfall obtained by the simulation using the copula with the smallest
Akaike information criterion (AIC).
Table 5.13 presents the means and standard deviations, which are the parameters
of the marginal distribution function of each return series.
We estimate each copula using the procedure described in Sect. 2.5.2. Chapters 2,
3, and 5 estimate bivariate, trivariate, and quinquevariate copulas, respectively.
Figure 5.6 plots the relationship between the degree of freedom and the log-
likelihood for the t copula estimates. We find the maximum log-likelihood when the 

degree of freedom is 5. Therefore, we estimate the t copula parameters; that is, Σ ,

with degree of freedom 5.
Table 5.14 provides all the estimated copulas. We generate the distribution of daily
returns for the portfolio by simulating each estimated copula 500,000 times. Then,
we calculate the VaR and expected shortfall. Table 5.14 provides the measured VaRs
and expected shortfalls based on these copulas. According to the Gumbel copula
with the minimum AIC, the VaR is 3.87% and the expected shortfall is 4.83%. This
portfolio has a 1% chance of losing more than 3.87% one day later, with an average
loss of 4.83% at a probability of 1%.

5.5 Concluding Remarks

This chapter investigates the market risk of a power generation business by defining
a power generation company as an institutional investor who manages an energy
portfolio consisting of long positions in fuel futures and short positions in electricity
futures. If a company trades fuels and electricity at prices pegged to the prices of
highly liquid commodities, we can easily provide an overview and measure the
market risk of that business.
Electricity is a major source of economic activity. Power companies broadly
accommodate economic activity, making power-generation businesses essential.
120 5 Market Risk of a Power Generation Business

Fig. 5.6 Degree of freedom and log-likelihood in the t copula estimates (European commodity
portfolio)

Even if a physically stable supply of electricity is guaranteed, the sudden withdrawal

of power-generation companies would have a significant impact on the economy
and must be avoided. Because the market risk of the power business is measurable,
the government should strictly define the method of measuring the market risk of
the power generation business and regulate the capital adequacy and liquidity ratios
of power generation companies based on the measured risk. Each power genera-
tion company should disclose its market risk and how it measures market risk to
shareholders, creditors, and customers as accountability for business continuity.
This chapter analyzes a power generation business in Europe as a case study.
This hypothetical business buys fuel oil, natural gas, fuel coal, and carbon credits at
prices linked to the futures prices of the Brent, TTF, Rotterdam, and EUA, and sells
electricity at a price linked to the futures price of the FrenchBL.
We measure the intermarket spillover effects of the volatility and return series
using Diebold and Yilmaz’s [3] approach. We then spectrally decompose the results to
determine the period during which the spillover factor occurred. Finally, we calculate
the VaR and expected shortfall of the generation business by using quinquevariate
copulas. The total connectedness of the return series and volatility series are 18.71
and 20.14%, respectively. The level of market integration in terms of the spillover
effects of the return series is about the same as that in terms of the spillover effects of
the volatility series, however, the breakdown is different. The return spillover effects
among the TTF, FrenchBL, and EUA are strong, while the volatility spillover effects
from Rotterdam to the EUA, from Brent to the EUA, and bidirectional between TTF
5.5 Concluding Remarks 121

Table 5.14 Risk measurement using estimated copulas (European commodity portfolio)
Gaussian AIC 34.134


Σ 1.000 0.151 0.156 0.098 0.226

0.151 1.000 0.311 0.486 0.371
0.156 0.311 1.000 0.293 0.182
0.098 0.486 0.293 1.000 0.343
0.226 0.371 0.182 0.343 1.000
Value-at-risk* 3.50%
Expected shortfall* 4.21%
t AIC 15.551
Degree of freedom 5


Σ 1.000 0.165 0.179 0.122 0.204

0.165 1.000 0.328 0.581 0.382
0.179 0.328 1.000 0.276 0.163
0.122 0.581 0.276 1.000 0.403
0.204 0.382 0.163 0.403 1.000
Value-at-risk* 3.35%
Expected shortfall* 4.01%
Clayton AIC − 11.966
πc 1.478
Value-at-risk* 3.78%
Expected shortfall* 4.74%
Gumbel AIC − 13.095
πg 2.237
Value-at-risk* 3.87%
Expected shortfall* 4.83%
Note * indicates the value at the 99% confidence level
Bold indicates the minimum AIC and the risk measures at that time

and Rotterdam are strong. The return spillover effects depend mostly on events within
five business days, whereas the volatility spillover effects depend mostly on factors
before 21 business days. The VaR was 3.87% in the simulation using the estimated
Gumbel copula. This business had a 1% chance of losing more than 3.87% one day
later. The expected shortfall was 4.83%. If this business loses more than 3.87% one
day later, then its expected loss will be 4.83%.
It may be possible to measure the market risk of the oil refining business, which
procures crude oil and sells petroleum products (i.e., gasoline, kerosene, and heating
oil) if it trades raw materials and products at prices pegged to the prices of highly
liquid commodity futures. We might measure the market risk of a business that
purchases corn, potatoes, and sugar cane and manufactures and sells ethanol. In
addition, if readers find companies that earn their operating revenue by trading goods
122 5 Market Risk of a Power Generation Business

in highly liquid markets, they may be able to measure market risk as an extension of
this study.

References

1. Baruník, J., & Křehlík, T. (2018). Measuring the frequency dynamics of financial connectedness
and systemic risk. Journal of Financial Econometrics, 16(2), 271–296.
2. Boubaker, H., & Sghaier, N. (2016). Markov-switching time-varying copula modeling of
dependence structure between oil and GCC stock markets. Open Journal of Statistics, 6(5),
565–589.
3. Diebold, F. X., & Yilmaz, K. (2012). Better to give than to receive: Predictive directional
measurement of volatility spillovers. International Journal of Forecasting, 28(1), 57–66.
4. Hanif, W., Arreola-Hernandez, J., Shahzad, S. J. H., Hoang, T. H. V., & Yoon, S. M. (2020).
Regional and copula estimation effects on EU and US energy equity portfolios. Applied
Economics, 52(49), 5311–5342.
5. Ji, Q., Liuc, B. Y., Nehler, H., & Uddin, G. S. (2018). Uncertainties and extreme risk spillover
in the energy markets: A timevarying copula-based CoVaR approach. Energy Economics, 76,
115–126.
6. Liu, T., Nakajima, T., & Hamori, S. (2022). The impact of economic uncertainty caused by
COVID-19 on renewable energy stocks. Empirical Economics, 62(4), 1495–1515.
7. Mzoughi, H., Urom, C., & Guesmi, K. (2022). Downside and upside risk spillovers between
green finance and energy markets. Finance Research Letters, 47, 102612.
8. Reboredo, J. C. (2015). Is there dependence and systemic risk between oil and renewable
energy stock prices? Energy Economics, 48, 32–45.
9. Siburg, K. F., Stoimenov, P., & Weiß, G. N. F. (2015). Forecasting portfolio-Value-at-Risk with
nonparametric lower tail dependence estimates. Journal of Banking & Finance, 54, 129–140.
10. Wu, F., Zhang, D., & Ji, Q. (2021). Systemic risk and financial contagion across top global
energy companies. Energy Economics, 97, 105221.
11. Zhang, Y. J., Chevallier, J., & Guesmi, K. (2017). “De-financialization” of commodities?
Evidence from stock, crude oil and natural gas markets. Energy Economics, 68, 228–239.
12. Zhang, W., He, X., Nakajima, T., & Hamori, S. (2020). How does the spillover among natural
gas, crude oil, and electricity utility stocks change over time? Evidence from North America
and Europe. Energies, 13(3), 727.
Chapter 6
Alternative to Postface: Market Risk
Transfer in Power Companies

Electric power companies often trade different types of energy at prices linked to
futures prices on commodity exchanges. They understand what is traded under the
contract and quantify its value to appropriately manage the company. In other words,
they often trade both the value of energy and the value of market risk in one contract;
therefore, they determine the value of energy quantitatively in terms of profitability,
and the value of market risk in terms of profitability and business continuity. Ng
et al. [11] adopt an option pricing methodology to compare the cost of a long-term
contract with a price cap to that of spot purchases. The following two examples are
inspired mainly by Ng et al. [11].
First, we introduce an example fuel procurement contract. Long-term liquefied
natural gas (LNG) trading contracts with price formulas, which are a function of
benchmark crude oil prices, are not uncommon, especially in the Asia–Pacific region.
We discuss the case in which an electric power company continues to buy LNG at
the price series PL N G,t , which is determined by the following equation:

PL N G,t = C0 + C1 Poil,t , (6.1)

where Poil,t is the price series of crude oil and C0 and C1 are the coefficients. We
can interpret C0 and C1 as the fixed costs and the variable costs to exchange crude
oil for LNG, respectively. Additionally, C0 and C1 include the seller’s profit. We can
think of exchanging the currency of crude oil for the currency of LNG. If C1 properly
represents the value of LNG for crude oil, then the risk of crude oil price fluctuations
due to the external environment is completely passed through to the buyer; that is,
the electric power company. The buyer considers the following measures against this
risk:
Measure 1. Passing the risk on to power generation costs and ultimately to selling
prices.
Measure 2. Hedging the risk with energy derivatives.

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 123
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4_6
124 6 Alternative to Postface: Market Risk Transfer in Power Companies

Measure 3. Creating accounting allowances that match the risk amount. In some
cases, the capital strategy changes.
It is of utmost importance for electric power companies to secure a stable spread
between the selling price and power generation cost. Therefore, we base this study on
the measures that depend on selling prices. For simplicity, suppose that the company
sells electricity at a price consistent with procurement in Eq. (6.1). In other words, we
assume that power sales contracts are limited to those that pass the risk of crude oil
price fluctuations to the selling price. Next, suppose that the company procures addi-
tional LNG with the following price formula to respond to the increase in electricity
sales:
ad
PL N G,t = C0 + C1 Poil,t , if Poil,t  Ps
ad
PL N G,t = C0 + C1 Ps , if Poil,t < Ps , (6.2)

where ad PL N G,t is the price of the additionally procured LNG. In reality, there is
probably no extreme price formula. However, it is not uncommon for the gradient
for Poil,t < Ps to be smaller than C1 to reduce the risk that the seller does not recover
its investment in LNG manufacturing equipment. We illustrate Eqs. (6.1) and (6.2)
in Fig. 6.1. The power company risks that the spread between the selling price and
power generation cost is narrower when Poil,t < Ps . Assuming that the business
balance in Eq. (6.1) is neutral, we can write the business balance in Eq. (6.2) as in
Eq. (6.3) and Fig. 6.2:

B L = 0, if Poil,t > Ps
( )
B L = C1 Poil,t − Ps , if Poil,t  Ps , (6.3)

where B L is the business balance in Eq. (6.2) assuming that the business balance in
Eq. (6.1) is neutral.
We can interpret this as a short position in a crude oil put option according to
Kawamoto and Tsuzaki [7]. In practice, taking such risks is unusual for electric
power companies. Therefore, the buyer can avoid this risk by taking a long position
in a crude oil put option, which we can calculate using the LNG procurement amount
in Eq. (6.2) and the value of C1 . If we do not use derivatives to hedge the risk, it is
rational to book an allowance that matches the risk amount. The risk amount is

Vt Put t dt, (6.4)
T

where T is the contract period, Vt is the position of the put option, and Put t is the price
of the put option. If the put option is listed on an exchange, then we can utilize the
price. Otherwise, we must price it theoretically based on market conditions (e.g., the
underlying asset price, its volatility, and government bond interest). Equation (6.4)
is the swap price of Eq. (6.3) for LNG procured in Eq. (6.2).
6 Alternative to Postface: Market Risk Transfer in Power Companies 125

Fig. 6.1 LNG price formulas

Fig. 6.2 Impact of LNG procurement with a lower threshold price on the business balance
126 6 Alternative to Postface: Market Risk Transfer in Power Companies

Second, we consider an example power sale contract. We discuss the case in

which an electric power company continues to sell power in the price series Ppower,t ,
determined as follows:

Ppower,t = α Poil,t + β Pgas,t + γ Pcoal,t + δ, (6.5)

where Poil,t , Pgas,t , and Pcoal,t are the price series for crude oil, natural gas, and coal,
respectively. If α, β, γ , and δ appropriately represent fuel power generation efficiency,
fuel power generation amount ratio, and fixed cost for each option, then the company
can pass on the risk of fuel price fluctuations to the selling price. However, power
sales contracts with an upper threshold price are common, as Eq. (6.6) demonstrates.
( )
N
Ppower,t = Max Pmax , Ppower,t
( )
= Max Pmax , α Poil,t + β Pgas,t + γ Pcoal,t + δ , (6.6)

where N Ppower,t is the power price with an upper threshold price, Pmax . We illustrate
Eqs. (6.5) and (6.6) in Fig. 6.3. The power company takes the risk that the spread
between the selling price and power generation cost is narrower when Ppower,t >
Pmax . Assuming that the business balance in Eq. (6.5) is neutral, we can present the
business balance in Eq. (6.6) as in Eq. (6.7) and Fig. 6.4.

B p = 0, if Ppower,t  Pmax
( )
B p = Pmax − α Poil,t + β Pgas,t + γ Pcoal,t + δ , if Ppower,t > Pmax (6.7)

where B p is the business balance in Eq. (6.6) assuming that the business balance in
Eq. (6.5) is neutral.
Just as Asakawa [1] quantifies the gas distribution price with an upper threshold,
we regard Eq. (6.7) as a short position in an electricity call option. This call option
is not listed on any commodity exchanges. If we design a portfolio equivalent to
Eq. (6.7), then we can pass the risk to the energy derivatives market. Although full
hedging is impossible, we may gain partial coverage from holding a portfolio of
oil, natural gas, and coal call options. Instead, we can clearly see why some rational
management theoretically calculates the option value to book an allowance consistent
with the risk amount. If each fuel price in Eq. (6.5) that defines the underlying asset
price is efficient, then we can reasonably evaluate the option using historical data.
We can now easily estimate the call option price using Black and Scholes’s
[4] model. The call option price series is

Ct = Ppower,t Φ(d1 ) − Pmax e−r (Ti −t) Φ(d2 ), (6.8)

where Φ is the standard normal distribution function, Ti is the time to sell power, r
is the risk-free rate, and
6 Alternative to Postface: Market Risk Transfer in Power Companies 127

Fig. 6.3 Power price formulas

Fig. 6.4 Impact of power sales with an upper threshold price on the business balance
128 6 Alternative to Postface: Market Risk Transfer in Power Companies
( ) ( )
ln Ppower,t /Pmax + r + σ 2 /2 (Ti − t)
d1 = √
σ Ti − t
√
d2 = d1 − σ Ti − t, (6.9)

where σ is the volatility of Ppower,t . From Eq. (6.5), we can calculate σ as

( )
σ 2 = Var Ppower,t
( )
= Var α Poil,t + β Pgas,t + γ Pcoal,t + δ
( ) ( ) ( )
= α 2 Var Poil,t + β 2 Var Pgas,t + γ 2 Var Pcoal,t
( ) ( )
+ 2αβCov Poil,t , Pgas,t + 2βγ Cov Pgas,t , Pcoal,t
( )
+ 2γ αCov Pcoal,t , Poil,t . (6.10)

Suppose that an M-month power sale contract, whose price is determined by

Eq. (6.6), concludes at t = 0. In addition, the electric energy of Wi is expected to be
sold in month i. In this case, we can calculate the swap value of risk at the time of
the contract using the following formula:


M  
Wi Ppower,0 Φ(d1 ) − Pmax e−r 12 Φ(d2 ) ,
i
(6.11)
i=1

where

Ppower,0 = α Poil,0 + β Pgas,0 + γ Pcoal,0 + δ

( ) ( )
log Ppower,0 /Pmax + r + σ 2 /2 12i
d1 = /
σ 12i
/
i
d2 = d1 − σ . (6.12)
12

However, it is difficult to say that this risk value is appropriate because it assumes
that the underlying asset price follows geometric Brownian motion and that the vari-
ance and covariance are constant. However, in practice, it is rational from the perspec-
tive of calculation cost to adjust the numerical value calculated using Eq. (6.11) based
on the deep insight of professionals. If higher accuracy is required, one method is as
follows. First, we estimate a stochastic process model for the underlying asset price
series from historical data. Second, we generate the option cash flow distribution by
Monte Carlo simulation using the estimated stochastic process model. Finally, we
obtain the option value by calculating the expected value.
A multivariate derivative is a derivative whose future payoff depends on multiple
underlying assets. The basket option, whose underlying asset is a portfolio of multiple
assets, is a typical multivariate derivative asset. In Eq. (6.7), because the multivariate
derivative price depends on the correlation between the underlying asset prices and
References 129

their probability distributions, the derivative price is generally estimated using an

approximate expression or a Monte Carlo simulation. While several studies examine
multivariate derivative assets, none examine power prices. Björk [3] and Dhaene et al.
[5] extend the univariate Black and Scholes [4] model to a multivariate case using
correlated Brownian motions. However, this model assumes a lognormal distribution
for each underlying asset price and a Gaussian dependence structure, which is unre-
alistic. Luciano and Schoutens [10] discuss a Lévy multivariate model for assets that
incorporates jumps, skewness, kurtosis, and stochastic volatility. Their proposed
model has the strengths of the univariate variance-gamma process and introduces
a non-Gaussian dependence structure. Linders and Schoutens [8] use a one-factor
Lévy model to value basket options by replacing the distribution of the portfolios
with a reasonable approximation. Linders and Stassen [9] propose a methodology for
pricing basket options using a multivariate variance-gamma model. By modeling the
underlying assets using time-changed geometric Brownian motions with a common
gamma subordinator, they express the basket option price as a linear combination of
the Black and Scholes models. Bayer et al. [2] consider basket option pricing in a
multivariate Black–Scholes and Variance-Gamma model, and propose using a Monte
Carlo simulation designed specifically to solve numerical integration problems with
non-smooth integrands. Hanbali and Linders [6] adopt the standard least-squares
Monte Carlo approach to generate American basket option prices and discuss the
method in terms of calculation time. We expect the readers to advance the study on
multivariate derivative assets in energy markets by referring to many past studies
cultivated in traditional financial markets. Furthermore, we hope that they contribute
to literature for pricing the electricity call option as well as for designing the optimal
portfolio consisting of fuel call options, which are discussed in this chapter.

References

1. Asakawa, H. (2010). Real option analysis of price adjustment system for gas distribution.
Journal of Real Options and Strategy, 3(1), 63–76.
2. Bayer, C., Siebenmorgen, M., & Tempone, R. (2018). Smoothing the payoff for efficient
computation of basket option prices. Quantitative Finance, 18(3), 491–505.
3. Björk, T. (1998). Arbitrage theory in continuous time. Oxford University Press.
4. Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of
Political Economy, 81(3), 637–654.
5. Dhaene, J., Kukush, A., & Linders, D. (2013). The multivariate Black & Scholes market: Condi-
tions for completeness and no-arbitrage. Theory of Probability and Mathematical Statistics,
88, 1–14.
6. Hanbali, H., & Linders, D. (2019). American-type basket option pricing: A simple two-
dimensional partial differential equation. Quantitative Finance, 19(10), 1689–1704.
7. Kawamoto, K., & Tsuzaki, K. (2008). Market valuation of LNG price formulas. Journal of
Japan Society of Energy and Resources, 29(2), 1–7.
8. Linders, D., & Schoutens, W. (2016). Basket option pricing and implied correlation in a one-
factor Lévy model. In Innovations in derivatives markets (pp. 335–367). Springer-Open
9. Linders, D., & Stassen, B. (2016). The multivariate variance gamma model: Basket option
pricing and calibration. Quantitative Finance, 16(4), 555–572.
130 6 Alternative to Postface: Market Risk Transfer in Power Companies

10. Luciano, E., & Schoutens, W. (2006). A multivariate jump-driven financial asset model.
Quantitative Finance, 6(5), 385–402.
11. Ng, F., Björnsson, H. C., & Chiu, S. S. (2004). Valuing a price cap contract for material
procurement as a real option. Construction Management and Economics, 22(2), 141–150.
Index

A Conditional covariance, 86, 89, 91, 95, 99,

Akaike Information Criterion (AIC), 16, 102
39, 43, 44, 71, 72, 79–81, 119, 121 Conditional variance, 62, 86, 89–91
Arbitrage, 2, 6, 21–23, 26, 28, 47, 48, 53, 85 Connectedness, 2, 53, 54, 57–62, 65, 66,
Archimedean copula, 33 69, 73, 76, 78, 79, 81, 82, 106, 107,
Asymmetric term, 62, 66, 76, 110 114, 115, 117, 118, 120
Augmented Dickey-Fuller (ADF), 2, 14, Constant Conditional Correlation (CCC), 3,
16, 17, 48, 64, 67, 68, 73, 74, 76, 77, 87, 93–96, 98, 100, 102, 103
107, 110, 112, 114
Contango, 8, 87
Autoregression term (AR term), 62, 66, 76
Autoregressive Conditional Convenience yield, 9
Heteroskedasticity term (ARCH Copula, 2, 6, 31–48, 53, 54, 58, 70–72,
term), 62, 66, 76, 91, 110 79–81, 86, 105–107, 114, 119–121
Crude oil, 1–3, 41, 53–56, 58, 62–72, 80,
82, 85, 86, 106, 107, 119, 121, 123,
B 124, 126
Baruník and Křehlík, 2, 54, 58, 60, 82, 107 Crude oil portfolio, 71, 72, 79, 80, 82, 86
Basket option, 128, 129
BEKK, 92, 93, 98, 103
Black and Scholes, 126, 129
Brent, 54, 55, 63–72, 82, 107–110, 112,
114–120 D
Descriptive statistics, 2, 7, 47, 54, 55, 57,
63, 66, 67, 72, 73, 77, 87, 88, 94, 98,
C 99, 106–110, 112
Call option, 126, 129 Diagonal BEKK, 3, 86, 87, 92, 94–100,
Carbon credit, 3, 106, 107, 114, 119, 120 102, 103
Clayton copula, 35, 40, 43, 46 Diagonal VECH, 3, 86, 91–101, 103
Coal, 1, 3, 5, 106, 107, 119, 120, 126
Diebold and Yilmaz, 2, 54, 57–60, 63, 67,
Cointegrating vector, 17–20, 24, 48, 49
73, 76, 82, 107, 110, 117, 120
Cointegration (Cointegrating,
Cointegrated), 6, 16–19, 21–23, 48, Discrete Fourier Transform (DFT), 60
53, 54 Dubai-Oman, 54, 55, 63–72, 82
Cointegration test, 2, 16, 18, 19 Dynamic correlation, 2
Commodity future, 85, 121 Dynamic ordinary least squares, 2
© The Editor(s) (if applicable) and The Author(s), under exclusive license 131
to Springer Nature Singapore Pte Ltd. 2022
T. Nakajima and S. Hamori, Energy Trading and Risk Management,
Kobe University Monograph Series in Social Science Research,
https://doi.org/10.1007/978-981-19-5603-4
132 Index

E Jarque-Bera, 11, 55, 56, 63, 67, 73, 76, 87,

Electricity, 1–3, 5–7, 12, 13, 21, 24, 41, 44, 94, 99, 108–110
47, 82, 85, 105–107, 114, 119, 120, Johansen, 2, 17, 19, 48
124, 126, 129 Joint distribution, 6, 31, 39, 41
Electric power, 5, 11, 24, 44
Electric power business, 49
Electric power company, 106, 123, 126 K
Elliptical copula, 33 Kurtosis, 10, 11, 47, 48, 55, 56, 63, 67, 73,
EU allowance (EUA), 107, 108, 110, 114, 76, 87, 94, 99, 108–110, 112, 129
117, 120
Expected shortfall, 2, 30, 31, 40, 42, 43, 48,
58, 62, 70, 71, 79, 80, 82, 107,
119–121 L
Explosive process, 13, 16 Liquefied Natural Gas (LNG), 54, 79,
Exponential generalized autoregressive 123–125
conditional heteroscedasticity, Long-term equilibrium, 2, 16, 17, 20–26, 48
Exponential, GARCH, EGARCH, 2,
54, 107
M
Marginal distribution, 31, 32, 40, 42, 43,
F 48, 70, 79, 119
Frank copula, 33, 38, 41, 43, 47, 48 Market integration, 53, 54, 65, 114, 117,
French baseload (FrenchBL), 107 120
Fuel portfolio, 2 Market risk, 3, 58, 105, 106, 114, 119–121,
123
Mean, 8–11, 25, 30–32, 41, 47, 55, 56, 62,
G
63, 66, 70, 72, 76, 79, 85, 87, 90, 94,
Gaussian copula, 33, 40, 45
98, 99, 107–110, 119
Generalized Autoregressive Conditional
Mean absolute deviation, 9
Heteroscedasticity term (GARCH
Median, 8, 9, 47, 55, 56, 63, 66, 72, 76, 87,
term), 3, 62, 86
94, 98, 107, 108, 110
Generalized Error Distribution (GED), 62,
Multivariate derivative, 128, 129
66, 76, 110
Gumbel copula, 36, 37, 40, 41, 43, 46, 70, Multivariate generalized autoregressive
71, 79, 80, 107, 119, 121 conditional heteroscedasticity
(Multivariate GARCH), 3, 86

H
Hedge Effectiveness (HE), 89, 90 N
Hedge, Hedging strategy, 90, 98, 102–104 National Balancing Point (NBP), 86–89,
Henry Hub (HH), 7–9, 11, 12, 17–19, 94, 98–103
21–25, 41, 43, 47, 48, 54, 56, 72, 73, Natural gas, 1, 2, 5–7, 9, 12, 13, 21, 22, 24,
76, 78, 82, 86–88, 94, 95, 103 41, 47, 53–57, 71, 73–78, 82, 86, 87,
106, 107, 114, 119, 120, 126
Natural gas portfolio, 58, 62, 71, 79–82
I Nonstationary process, 13, 18
Independently and identically distributed
Normal distribution, 11, 14, 30, 31, 33, 40,
random number following the
48, 55, 56, 87, 99, 108–110, 126
standard normal distribution, iid
N(0, 1), 90

O
J Optimal Hedge Ratio (OHR), 2, 86, 87, 89,
Japan Korea Marker (JKM), 54, 56, 72, 73, 90, 94–96, 98–100, 102, 103
76, 78–80, 82 Optimal portfolio, 86, 102, 129
Index 133

P Spurious regression, 13–16

Percentile, 9, 31 Standard deviation, 9, 10, 32, 41, 47, 55,
PJM, 7–12, 16–19, 21–25, 41, 43, 47, 48 56, 59, 63, 66, 67, 70, 72, 77, 79, 87,
Portfolio, 2, 3, 6, 29–31, 40–43, 48, 53–55, 94, 99, 108–110, 112, 119
57, 59, 65, 69–71, 73, 81, 82, 85, 86, Stationarity, 12, 73, 107
89–91, 98, 102, 103, 105–107, 119, Stationary process, 12–14, 16–18, 21, 58,
126, 128 64, 67
Power, 1, 2, 5–7, 9, 11, 16, 21, 22, 24, 25, Statistical arbitrage, 2, 6, 7, 21, 23–29, 41,
44, 48, 49, 53, 70, 82, 105, 106, 119, 43, 47–49
123, 124, 126–129
Power generation business, 3, 5, 6, 24, 105,
106, 114, 119, 120 T
Put option, 124 T copula, 33–35, 40, 42, 43, 45, 70, 71, 80,
119, 120
Title Transfer Facility (TTF), 54, 56,
R 72–74, 76, 78–80, 82, 107, 109, 110,
Range, 2, 9, 17, 42, 47, 55, 56, 63, 87, 94, 114, 117, 120
99, 108, 110 Trading strategy, 6, 21, 24, 26
Return, 2, 6, 20, 30, 41–43, 48, 53, 54, 58,
61–66, 68, 70–76, 78, 79, 82, 85, 86,
89, 90, 94, 95, 98, 99, 102, 104,
U
106–110, 114, 115, 117, 119, 120
Unit root, 14, 16–18, 48, 67, 112
Risk management, 43, 106, 117
Unit root process, 13, 14, 16, 17, 22
Risk measurement, 9, 41, 43, 44, 47, 72, 81,
Unit root test, 2, 12, 14, 16, 17, 20, 107,
82, 119, 121
110, 112, 114
Risk, Portfolio risk, 29, 30, 48, 53, 82
Risk transfer, 123
Rotterdam, 107, 108, 110, 112, 114, 117,
120, 121 V
Value-at-Risk (VaR), 2, 29, 44, 58, 81, 105,
121
S Variance-covariance matrix, 64, 65, 68, 69,
Schwarz Bayesian Information Criterion 73, 74, 76, 114, 117
(SBIC), 16, 20, 59, 62, 66, 68, 73, VECH, 91, 92
76, 110, 114, 115 Vector Autoregressive (VAR), 2, 110
Skewness, 9, 10, 47, 55–57, 63, 67, 73, 76, Vector Moving Average (VMA), 2, 54, 59,
87, 94, 99, 108–110, 112, 129 63, 64, 76, 107, 110, 112
Spectral analysis, 54, 58, 65, 73, 78, 114, Volatility, 2, 3, 48, 53, 54, 58, 61, 62,
116, 118 66–69, 71, 73, 76–78, 82, 102,
Spectral decomposition, 2, 60, 62, 69, 73, 105–107, 110, 112–115, 117, 118,
79, 82, 114, 117 120, 124, 128, 129
Spillover, 2, 53, 54, 57, 59–66, 68, 70, 71,
73, 75, 78, 79, 81, 82, 105–107, 110,
114, 115, 117, 118, 120, 121 W
Spot-future arbitrage, 7, 26–28, 48 West Texas Intermediate (WTI), 54, 55,
Spread, 6, 9, 21–24, 41, 48, 76, 124, 126 63–68, 70, 71, 82