SPRINGER BRIEFS IN QUANTITATIVE FINANCE

René Aïd

Electricity
Derivatives
SpringerBriefs in Quantitative Finance

Series editors
Peter Bank, Berlin, Germany
Pauline Barrieu, London, UK
Lorenzo Bergomi, Paris, France
Jakša Cvitanic, Nice Cedex 3, France
Matheus Grasselli, Toronto, Canada
Steven Kou, Singapore, Singapore
Mike Ludkovski, Santa Barbara, USA
Rama Cont, London, UK
Nizar Touzi, Palaiseau Cedex, France
Vladimir Piterbarg, London, UK
More information about this series at http://www.springer.com/series/8784
René Aïd

Electricity Derivatives

123
René Aïd
Finance for Energy Market Research Centre
EDF R&D
Clamart
France

ISSN 2192-7006 ISSN 2192-7014 (electronic)

SpringerBriefs in Quantitative Finance
ISBN 978-3-319-08394-0 ISBN 978-3-319-08395-7 (eBook)
DOI 10.1007/978-3-319-08395-7

Library of Congress Control Number: 2014958978

Mathematics Subject Classification: 91G20, 91G80, 91G60

JEL Classification: G12, G13

Springer Cham Heidelberg New York Dordrecht London

© The Author(s) 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
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, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
A mon père, M. Mahand Aïd
et à ma mère, Mme Ferroudja Mohellebi.
Foreword I

The electricity market is currently entering a period of significant changes with the
development of intermittent renewable energies and demand-response mechanisms.
Already quite complex to manage because electricity cannot be stored at rea-
sonable cost and because electricity follows all available paths (according to the
Kirchhoff’s principles), the laws of the market and the pricing system are entering a
new era of development.
The questions dealt with are quite complex on a mathematical standpoint with
high level optimization problems. Do not be afraid: several mathematical formulae
appear in the text.
Nevertheless, these questions are practicable and really usable by traders and the
utilities market. It is not an academic book, it is a book for the industry.
René Aïd succeeds to build that bridge between two worlds.
This book is the result of a brilliant collaboration between the academic world
(Ecole Polytechnique, ENSAE, Dauphine) through the FIME lab, and industrial
world through EDF R&D teams.
It is the result of more than two years of research and the overall understanding
of the evolution of the electricity market.
It is also a landmark of the ambition of the EDF R&D organization, to bring the
best available academic knowledge into the industry.
Thanks and congratulations to René and his team for this performance.

Paris-Saclay, October 2014 Bernard Salha

Senior Vice President
Head of EDF R&D

vii
Foreword II

This monograph is an excellent introduction to the world of electricity markets. The

content is unique within the available literature by the wide spectrum which covers
the subject starting from the managerial aspects of electricity generation, and
arriving at the corresponding financial derivatives.
A special emphasis is put on the various specific aspects of the electricity
financial market as opposed to the stock market, thus justifying the need to develop
related relevant models for the derivatives of hedging and pricing, and the corre-
sponding numerical approximation.
Through his unique positioning as one of the best international experts in the
electricity market R&D, and a researcher strongly connected to the academic
community, René succeeds in delivering the essential messages from electricity
market practitioners. The present valuable presentation of the field will undoubtedly
attract more economists and applied mathematicians, and help them to identify
interesting academic questions with relevant application to the practical electricity
market. The content will also serve in the opposite direction as a reference for the
relevant models that have been developed in the academic literature, and are cur-
rently used by electricity markets R&D practitioners.

Paris-Saclay, October 2014 Nizar Touzi

Professor, Ecole Polytechnique

ix
Preface

The project that led to this book started in August 2011 when Matheus Grasselli
proposed the writing of a monograph on the quantitative financial aspects of energy
markets in a new collection launched by Springer: Springer Briefs. We quickly
defined the scope of the book and the table of contents. But, this process would
certainly have taken much longer without the opportunity given to me by Fred
Espen Benth. Fred invited me to give a short series of lectures at the University of
Oslo in September 2013 on electricity markets and derivatives. This commitment
compelled me to create a large part of the material included in this book.
To fit the requirements of the SpringerBrief series, I chose the field of electricity
derivatives. Electricity markets and prices have drawn the attention of academics
from many different fields: economy, regulations, statistics, finance and mathe-
matical finance. I skipped all of the regulatory aspects which nevertheless involved
first-order economic questions as well as interesting mathematical modelling
problems. I also overlooked the questions of price forecasting because exhaustive
monographs on this subject already exist.
The book ranges from models which allow the tractable computation of futures
prices to the valuation of storage and swing options, which are the most complex
options to be evaluated in this market. My purpose is to give the reader a strong
foundation in this field. Thus, I first provide an explanation of the main properties
of electricity as a commodity and the main characteristics of the electricity market’s
microstructure. With these concepts, the reader is able to go through the whole
zoology of stochastic models that propose to capture the dynamic of the electricity
spot and futures prices. Then, I focus on the most important derivatives: spread
options, tolling contracts, power plants, and storage and swing options. I also
provide the reader with a description of the problems involved with the pricing of
retail contracts and weather derivatives.
This book is intended on the one hand for applied mathematicians, statisticians
and economists looking for a new interesting field of research. And on the other
hand, for practitioners working in energy utilities or on the commodity desks of
financial institutions.

xi
xii Preface

I want to take this opportunity to thank the different institutions and persons that
made this book possible. The first is the EDF group. As an employee of EDF, I was
given the time and the resources to write this book. My successive managers trusted
me, and without this trust I would not have had the chance to finish this book. Thus,
I want to personally thank my manager, Marc Ringeisen, head of the EDF’s Lab
Osiris department. I also want to send special thanks to Bernard Salha, Head of
EDF R&D, who agreed to write a foreword for this book.
Several academic institutions also contributed greatly to the writing of this book:
the University Paris-Dauphine, the Ecole Polytechnique, and the CREST (Centre
for Economic and Statistic Research of the ENSAE). Together with EDF R&D,
they created the Finance for Energy Market Research Centre (the FiME Lab), which
I had the honour to manage from its birth in 2006 to 2012. I want to particularly
thank those three institutions for providing me with document resources and the
University of Paris-Dauphine for providing me with an office (with a priceless view
of the Boulogne wood).
I also want to thank the people who created the FiME Lab in 2006: Nizar Touzi,
professor at the Ecole Polytechnique; Elyès Jouini, Vice-President of the University
Paris-Dauphine; Jean-Michel Lasry, former Scientific Advisor of Credit Agricole
CIB and professor at the University Paris-Dauphine; Pierre-Louis Lions, professor
at the Collège de France; and Patrick Pruvot, former head and founder of the EDF’s
R&D Osiris department. These five people had a significant impact on my life,
particularly Nizar who had a special role. He and I have been collaborating since
2004, and these past ten years literally changed my life. I want to thank him not
only for writing a foreword to this book, but for all I learned from him during these
ten years.
I also want to thank my colleagues at EDF R&D and the FiME members who
helped me substantively improve the manuscript of this book: Nadia Oudjane;
Xavier Warin; Olivier Féron; Clémence Alasseur; Audrey Mahuet, director of
EpexSpot Product Design, who helped me understand some aspects of intra-day
trading and provided me with data; and Matt Davison who as a reviewer gave me
simple and concrete advice on how to improve the book. Further, I want to thank
Nicolas Langrené, my former Ph.D. student, and Corentin Guttierrez and Elias
Daboussi for providing some nice pictures in this book. I also want to thank
Jonathan Moore for transforming this text written in French–English into a book
written in English. I also thank Ute McCrory, my editor at Springer, who helped me
in the design of this book.

Paris-Saclay, October 2014 René Aïd

Contents

1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Electricity Markets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Electricity Features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Storage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Market Microstructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Intraday Market and Balancing Mechanism. . . . . . . . . . . 10
2.2.2 Day-Ahead Market . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.3 Forward Market. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.4 The Diversity of Electricity Markets . . . . . . . . . . . . . . . 19
2.3 Real Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Price Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Preliminary Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 HJM-Style Forward Curve Models. . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Principle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 The Case of Electricity. . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 One-Factor Spot Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.1 Mean-Reversion Process. . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.2 Mean-Reverting Jump-Diffusion Models. . . . . . . . . . . . . 42
3.3.3 Non-Gaussian Mean-Reversion Models . . . . . . . . . . . . . 45
3.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4 Multi-factor Spot Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4.1 An Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4.2 A Review of the Multi-factor Models. . . . . . . . . . . . . . . 50
3.4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

xiii
xiv Contents

3.5 Structural Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.5.1 The Mother of All Structural Models . . . . . . . . . . . . . . . 53
3.5.2 Auxiliary Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.3 Stack Curve Models. . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Derivatives. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Spreads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 Power Plants and Tollings. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Storage and Swings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 Retail Contracts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.5 Weather Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Chapter 1
Introduction

More than 30 years have passed since the first country (Chile 1981) deregulated its
electricity market and quoted an electricity spot price. Nearly 20 years have passed
since the first quotation of a futures contract on electricity (Scandinavian market
1993). Since then, the quantitative finance literature has followed the worldwide
liberalisation of the electricity industry. Amongst the many drivers that can explain
the interest in this field, three of them are worth noting in a monograph dedicated to
electricity derivatives.
First, electricity fails to satisfy the basic hypothesis that sustains modern pricing
theory. Electricity cannot be stored. This fact leads to the extremely spiky behaviour
of spot prices. This phenomenon is now well known and well documented. More
recently, electricity spot prices in Europe have frequently exhibited negative values
due to the joint effects of the operational constraints of power plants and the fact that
electricity cannot be stored. Second, these effects together with the lack of flexibility
in the generation assets lead to complex microstructures in the electricity markets.
They are designed to offer the best possible trading capacities to market operators
while maintaining the security of the electric system. Third, if not specifically, elec-
tricity trading involves the pricing of path-dependent options. They involve not a
unique but a sequence of exercises. They are designed to reflect the constraints in the
generation assets. The pricing of theses derivatives leads to optimal control problems.
They rarely admit an analytical solution. Numerical methods to efficiently provide
a value and its sensitivities are needed.
The purpose of this monograph is twofold. It aims at providing a comprehensive
state of the art investigation into the pricing and hedging problems raised by electricity
derivatives. And it proposes methods to tackle them. Monographs already exist that
provide detailed descriptions of each aspect of the electricity derivatives treated in
this book. Most of them provide a global view on the microstructures in the energy
markets, the main spot and futures price models, some descriptions of standard
derivatives, and some methods to price them. For the reader who is interested in the
field of energy derivatives, I recommend the lectures by Clewlow and Strickland [65],
Eydeland and Wolyniec [85], Geman [93], and Burger et al. [47]. They provide a far
more detailed description of all of the energy derivatives and provide an introduction
© The Author(s) 2015 1
R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7_1
2 1 Introduction

to the main spot and futures price models. Pilipovic’s monograph [143] is more
focused on pricing and hedging in the context of a variety of spot price models
based on the variations in Schwartz’s models of 1997. In this context, the monograph
provides detailed computations of prices and Greeks. For models on spikes, the reader
should consult the reference monograph by Benth et al. [19]. This book covers the
problems in electricity price modelling (estimation, calibration) and the pricing of
standard derivatives (spreads) in the context of general Lévy processes. Just recently,
the collected works edited by Carmona et al. [55] address the question of efficient
numerical methods for specific electricity and gas derivatives (storage and swings).
And even more recently, Swindle’s [155] detailed monograph on the valuation of
energy derivatives offers a practitioners view on all of the aspects of energy derivatives
as well as the methods used by trading desks to price them.
The purpose of this brief book is not to offer an in-depth analysis like those works.
Its objective is to propose a broad view of the features of electricity markets through
the price models and derivatives which are specific to the electricity business, such as
power plants, tolling, swings, and storage. This text could be useful to readers who are
involved in electricity trading or the optimisation of generation assets. The reader
will find here a concise landscape of the academic knowledge on the pricing and
hedging of electricity derivatives. It could also be useful to academics who are look-
ing for new interesting fields to investigate. The book describes the main valuation
problems faced by operators of electric utilities. Crude mathematical formulations
are provided to help the reader understand the mathematical difficulties involved with
these methods. Moreover, I hope this book will encourage new researchers to devote
part of their time to this fascinating field. Indeed, despite two decades of research and
significant progress in the modelling of electricity spot prices, no model has emerged
as a reference—as is the case for equity derivatives with the Black and Scholes model
and the Heath-Jarrow-Morton (HJM) models for interest rates—and a lot is left to be
done to reduce the gap between the operational needs for efficient and robust pricing
models and the available solutions in the literature.
This book is structured in three chapters which deal successively with markets,
price models, and derivatives.
Chapter 2 provides an introduction to electricity markets. This chapter explains the
main features of the electricity commodity: that it cannot be stored and that it is a local
commodity. I provide the reader with a crash course on electrical engineering and
focus on its consequences on the market’s microstructure. I do not address the reactive
power or frequency regulation, but I describe the consequences of Kirchhoff’s laws
on trading. By using the example of the French electricity market, I explain how
the three main markets of the intraday, the day-ahead, and the forward market are
interrelated. Moreover, I explain the main problems faced by the utilities and the
derivatives developed to respond to these needs. In particular, there is a class which
can be called real derivatives by analogy with real options, which consists of physical
generation assets. Indeed, power plants and hydroelectric or gas storage are assets
with embedded options. I leave aside the issues surrounding regulations, even though
the European electricity markets are (still) under construction as the development of
capacity markets shows it. Moreover, the trading activities of electric utilities have
1 Introduction 3

been affected by European financial regulation after the 2008 financial crisis, such
as EMIR (European Market Infrastructure Regulation) and REMIT (Regulation on
Energy Market Integrity and Transparency). But, I do not deal with this topic because
it is a subject all on its own.
Chapter 3 covers the models for electricity prices. Because this monograph is
dedicated to derivative pricing and because futures are the basic tools used to hedge
more complex derivatives, it limits itself to models on the joint dynamic of spot and
futures prices. Thus, the forecast models for the spot prices are not covered here. I
refer the reader to Weron’s (2006) book [165] on this subject. The models presented
here are sorted out in four sections. First, I deal in Sect. 3.2 with the HJM approach.
This approach consists of directly modelling the dynamic of the futures prices. The
spot price is then deduced as a futures price with zero time to maturity. No arbitrage
conditions are encoded in the dynamic of the futures price written directly under
the risk-neutral measure. This is the preferred approach of trading desks in order to
stay away from the spot. The constraints to ensure no arbitrage conditions between
the futures prices and spot prices, which are already known in the case of the yield
curve, still apply for electricity. But, those constraints are considerably increased
in the case of electricity because a futures contract on electricity involves delivery
during a period that makes it closer to a swap contract. Although this approach
is preferred by market operators, there is less literature on it than on the strategy
that consists of modelling the spot price and deducing the futures prices through an
argument of no-arbitrage (the Royal road). The literature which uses this strategy is
important. Therefore, I sort the models into three categories of increasing complexity.
The first class that is presented in Sect. 3.3 consists of a one-dimensional model.
This class offers the simplest models and allows assessing the difficulty involved in
modelling the dynamic of the spot and futures prices. This class includes a Gaussian
mean-reversion model, a jump mean-reversion model, and a one-factor Lévy process,
which is the most popular example of a non-Gaussian process. This class of models
succeeds in capturing the spiky behaviour of electricity spot prices, but generally
fails in getting all the richness of the dynamic of the forward curve. The second class
that is presented in Sect. 3.4 consists of multi-factor models. The typical models in
this class are the hidden Markov chain models. This class of models gives up the
Markovian property of the spot price to be able to capture a realistic dynamic of the
futures prices. They succeed in capturing both the behaviour of the spot and that of the
futures prices but requires filtering procedures for the estimation of the parameters,
because they involve the dynamic of non-observed variables. The third class that is
presented in Sect. 3.5 is the structural models. The principle of these models is to
deduce the spot price—and not necessarily its dynamic—as the result of an over-
simplified relation between generation and consumption. The main advantages of
these models is to propose a direct and consistent answer to the dependence of
the electricity spot price on some other observed variable, which is often of some
interest for utilities (available capacities, fuel prices, consumption). Moreover, the
estimation and calibration processes are simplified by the fact that the models are
set on observable variables. But, this is done at the expense of an increase in the
dimension that penalises the efficiency in the pricing of complex derivatives.
4 1 Introduction

The last Chap. 4 is devoted to derivatives. I concentrate on the products which are,
if not specific, at least of a particular importance to electricity trading. Section 4.1
covers the spread options. These options are not unique to electricity markets, but as
Chap. 2 will make clear, electricity utilities are not interested in the spread options
for themselves but because they are the basic tool to evaluate many other derivatives
such as power plants. These derivatives are seen as fuel spread options or, with
storage facilities, as calendar spread options. Then, Sect. 4.2 deals with the valuation
of power plants and their financial representation, which are tolling contracts. Taking
into account their main constraints leads to the path-dependent options and optimal
switching problems. This section presents an example extracted from the literature
on the different ways the problem can be formulated. In particular, the example
takes explicitly into account the fact that the market is incomplete and the hedge
is not perfect. Section 4.3 deals with one of the most complex derivatives involved
in the electricity industry, namely storage facilities. This chapter also covers swing
contracts. They can be seen as a financial contract modelling a physical storage
facility such as a dam or a gas storage facility. In the case of a physical storage
facility, random inflows, injection or withdrawal costs from the facility make the
optimal control problem more difficult to solve. Section 4.4 is devoted to the contracts
which do not attract much attention from academics but which are essential to the
utilities, namely retail contracts. Depending on the nature of the customer (household,
small business, or industrial sites), the retail contracts can have very different settings
and very different embedded options. Finally, Sect. 4.5 presents weather derivatives.
Although they do not represent an important part of the derivatives traded in electricity
markets, they enjoy a large degree of attention from the professionals that sell them.
Indeed, the activity of electric utilities is climate sensitive: consumption depends on
weather, hydroelectric generation depends on precipitation and snow falls, and wind
generation depends on the wind (!). Thus, it is natural to think of developing hedging
instruments against these risk factors.
The Chap. 5 concludes this monograph with some research perspectives.
Chapter 2
Electricity Markets

This chapter presents the main properties of electricity, the microstructures of the
electricity market and introduces the derivatives which are specific to this market.
Regarding electricity’s properties, I focus only on those that have a direct conse-
quence on pricing and trading. The fact that electricity cannot be stored cannot be
understated. In the same line, the constraints on its transport make electricity a local
commodity. There is no such thing as an international homogeneous electricity mar-
ket or even a regional electricity market on the scales of Europe or the United States
as is the case for gas markets. There are at least as many electricity markets as there
are countries in the world. It is thus not possible to enter into the details of each coun-
try’s specific electricity market. Here, I deal only with the most common features
of their structure. Further, I introduce some of the most specific derivatives of elec-
tricity markets by presenting the context, objective, and constraints of the electricity
utilities whether they are large or small, or whether they hold generation assets or
are purely retailers.

2.1 Electricity Features

All commodities exhibit technical features that make them unique. For example,
trading live cattle on the Chicago Board of Trade (CBOT) might be as technical as
trading electricity. But, electricity presents two main characteristics that raise both
theoretical and practical problems:
1. electricity cannot be stored,
2. the transport of electricity satisfies specific laws.

2.1.1 Storage

The fact that electricity cannot be stored is sometimes tempered by using the
expression that it cannot be stored at reasonable cost. In fact, because electricity
© The Author(s) 2015 5
R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7_2
6 2 Electricity Markets

consumption is a continuous phenomenon, they exist both an energy problem and a

capacity problem, that is, the pace at which energy can be released.
Regarding energy, the most economical way to store a large amount of energy
for electricity generation is still hydroelectric reservoirs. Nevertheless, this is not a
universal solution because it relies on the hydroelectric potential of a country. For
instance, in a country like France where the hydroelectric potential was developed
over the 1950s to the 1970s, it now represents 25 GW of installed capacity of a
total installed capacity of 110 GW. Hydroelectric generation represents 15 % of the
total generation, and its energy storage capacity is around 10 TWh as compared to an
annual consumption of 500 TWh. Moreover, its availability depends on inflows com-
ing from precipitation. Contrary to a thermal power plant whose fuel can be bought,
a hydroelectric plant might be unavailable because its reservoir is empty—and rain
cannot be bought (even now). This point can be illustrated by the Norwegian electric-
ity system. In this country, more than 95 % of the generation is based on hydroelectric.
The total consumption is around 130 TWh, and the hydroelectric generation is about
124 TWh. But, due to the dependence on hydrology, the energy capacity can vary
from 160 TWh during wet years to 100 TWh during dry years which in this last case
leads to a lack of energy. Further, the cost of building a hydroelectric power plant
varies a lot according to the place where it is built. Nevertheless, the investment cost
varies from 1.5 to 2.7 billion euros per GW (source: Report on the development of
hydroelectric generation, French Ministry of Industry, 2006). To make a comparison,
the investment cost for a combined cycle gas turbine is around 450 million euros per
GW.
Regarding capacity, there has to be enough capacity to continuously satisfy the
power demand. However, the energy required to satisfy the power demand over a
certain period of time might exist, but not at the right capacity. To give an order
of magnitude, the total installed capacity of France is around 110 GW while the
annual maximum power demand has reached 100 GW in the last several decades.
In the case of Norway, the installed capacity is around 30 GW with 29 GW of it
hydroelectric. But, the annual maximum power demand has reached 30 GW in the last
several decades too. To overcome the capacity problem, installing enough capacity
to satisfy demand in any situation might be a solution. But, this method requires too
many power plants that would be used for only a few hours in their lifetime. The cost
of this system would be prohibitively expensive.
The fact that power cannot be stored has drastic effects on the way electricity
systems have to be managed. A too long excess of demand compared to generation
might first be resolved by a decrease in frequency and if not properly corrected, in
dramatic blackouts. Thus, this risk has two major implications in terms of genera-
tion management. First, it implies a minute by minute real-time assessment of the
equilibrium between consumption and generation because 15 min of disequilibrium
can result in a major blackout. Second, the transport system operator (TSO) who is
responsible for the electricity system’s security and reliability needs to have at its
disposal operating reserves to be able to cope with the uncertainties which affect
generation and consumption.
2.1 Electricity Features 7

Operating reserves are generation capacities that can be mobilised within a given
notification time. They are sorted according to their response time. Because frequency
is immediately affected by any discrepancy between consumption and generation, a
first reserve consists in the capacities that can be automatically increased or decreased
without any human intervention. This reserve is composed by the primary and the
secondary reserves. They can be mobilized within less than 15 min. Beyond 15 min,
the reserve is manually mobilised through a direct order by the system operator to
the generation plant’s management. This is the tertiary reserve. It consists in two
parts. First, the rapid tertiary reserve provides the generation capacities that can be
mobilised in 15 min for a guaranteed period of at least one hour. It offers a complement
to the secondary reserve. Second, the additional tertiary reserve provides capacities
that can be mobilised in 15–30 min for a guaranteed generation of 6 h. A deferred
reserve is made of power that can be brought in line in more than half an hour. The
volume of each reserve can vary depending on the nature of the uncertainties on a
particular electricity system. To give orders of magnitude, the following minimum
values hold in France where consumption on a given hour varies from 50 GW in
off-peak hours of the summer to 100 GW in the peak hours of winter:
primary reserve ≈700 MW
secondary reserve ≈500 MW
tertiary reserve ≈1.500 MW.
Thus, at any given time t, the system operator can compute his or her operating
margin for the maturity t + h that represents the difference between the demand
forecast at time t + h and the sum of the secondary, tertiary, and deferred reserves.
The margin required for a maturity of 15 min is set by the event of losing the group
with the highest generation capacity. For a longer maturity, it is set by the probability
of using exceptional measures (e.g., load shedding). For instance, in France, this risk
is defined to be lower than a 1 % chance during peak hours. This risk level requires
a margin of 2.3 GW to be available for the next two hours.
The reader with a greater interest in the fields of power system reliability can
consult the Power System Reliability Memento [150] of the French system operator,
which is available on the RTE website.

2.1.2 Transport

Electricity is transported according to Kirchhoff’s laws. Basically, these laws state

that the intensity at each node should be zero and the tension in each loop should
also be zero. An important consequence of these laws is that, in a meshed electricity
network, power goes from one point to another through all available paths. Figure 2.1
presents a three-node electricity network where each line is supposed to have the same
technical properties and the same lengths. Line A–C can handle 180 MW while line
A–B and B–C can handle 90 MW. These limitations hold in both directions. Suppose
that a power operator G1 has a customer in node C whose consumption is 180 MW,
8 2 Electricity Markets

(a) (b)
A A

90 MW 180 MW 60 - 30 120 + 30

B 90 MW C B 60 + 60 C

Fig. 2.1 The effect of Kirchhoff’s law on electricity exchange. a Constraints on capacities, b actual
flows

while a power operator G2 also has a customer in node C whose consumption is

90 MW. Suppose each operator holds enough generation capacity and that there is
no generation cost advantage. If the power flows in lines A–C and B–C like trucks,
there is no conflict. Operator G1 generates 180 MW and transmits it through line
A–C while operator G2 generates 90 MW and transmits it transit through line B–C.
But, because electricity flows through all of the available paths between points, 2/3
of the 180 MW transmitted by G1 will spread from A to C and 1/3 from A to B,
and then from B to C. The same holds for the 90 MW transmitted by operator G2.
These transmissions then lead to the violation of the transit capacity of line B–C by
carrying 120 MW while the line is limited to 90 MW.
It is the main reason why the transfer capacities which are available for trading
between countries require some electricity generation hypothesis before they can
be computed. In Europe, the available net transfer capacities (NTC) are managed
and published by the ENTSOE (European Network System Operator for Electricity)
and are made publicly available on its website (www.entsoe.net). To handle these
constraints, TSOs and the electricity markets use different methods. In the first years
of the European electricity markets, continental Europe chose an explicit auction
mechanism. Those who wanted to sell power from one country and have it delivered
to another country had to buy the transfer capacity between the countries at an auction
that was organised by the TSOs. This method has the benefit of being simple. But, it
induces inefficiencies and often adverse flows (flows going from a high price country
to a low price country) because of the coordination problems between the market
timing of both the power and the transmission capacities.
At the same time, since its early beginning in the 1990s, the NordPool has imple-
mented an implicit auction mechanism between its member countries. In this mech-
anism, the buyers and sellers post their bids without worrying about transport con-
straints. Their bids are assigned to their areas. Then, the market operator performs a
clearing which takes into account the possibilities of exchange between the zones.
While more complex than the former, this method has the benefit of avoiding adverse
flows and simplifying the business of traders. This method was chosen by the differ-
ent actors of the electricity market in continental Europe under the name of market
coupling. From 2006 to now, the coupling of markets has been achieved between
2.1 Electricity Features 9

(in alphabetic order) Belgium, Denmark, Estonia, Finland, France, Germany and
Austria, Great Britain, Latvia, Lithuania, Luxembourg, the Netherlands, Norway,
Poland, Portugal, Spain, and Sweden.
For an introduction on the auction mechanism for electricity transport, I refer the
reader to Stoft’s book on power system economics [154, part 5].
Remark 2.1 The topic of this book is mostly about what happens at the transport
level of electricity because this is the level at which the wholesale markets operate.
Nevertheless, the change that occurs at the distribution level of electricity because of
the introduction of renewable energies cannot be ignored. The distribution network
was historically designed to transfer power from the transport level to the consumers.
Electricity would flow downward from the transport level to the distribution level.
With the connection of renewable sources such as solar panels and wind farms at the
distribution level, now power sometimes flows upward from the distribution network
to the transport network.
This phenomenon requires an adaptation of the way distribution networks are
monitored and operated. The concept of smart grids applies to this adaptation. Smart
Grids correspond to the development of the information technology infrastructure
which allows the control of the distribution network to adapt to the massive intro-
duction of intermittent sources of energy. In this regard, the smart counter is one key
element of the infrastructure. It is the ability to have a fine measure of household
consumption but also the capacity to communicate with the customer to send him
or her price signals. But it is not the only key element. Flexibilities in consumption
(the capacity to defer consumption in time even for a few hours) as well as short-
term storage capacities which are able to cope with important variations in wind
generation become an issue. I will come back to this point in Sect. 2.3.

2.2 Market Microstructure

The design of electricity wholesale markets differs from one country to the next.
However, the properties of electricity have led to common features in the markets’
microstructures. I focus here only on those common factors. In each country, the
electricity market is composed of a series of markets with different time scales and
different scopes. Generally, three different gross markets can be distinguished:
1. the intraday market and/or the balancing mechanism: the balancing mechanism
consists of exchanges between the TSO and the market players to ensure the
real-time assessment of the grid while the intraday market allows the exchanges
between the market players themselves to ensure the equilibrium between their
generation and the consumption of their customers.
2. the day-ahead market: quantities are exchanged one day before delivery for the
next 24 h or 48 half-hours of the next day.
3. the forward market: market players can buy or sell electricity for future delivery.
To illustrate these three markets, I use France and Germany as examples.
10 2 Electricity Markets

2.2.1 Intraday Market and Balancing Mechanism

In a very short time frame (say the next 12 hours), the actors in the electricity market
are most concerned with the precise equilibrium between generation and consump-
tion of their portfolio. Each actor is made financially accountable for all of the dis-
crepancies between the consumption of its customers and its generation. The TSO
is also most concerned about a precise equilibrium between the overall generation
and consumption in the system. For this reason, two systems coexist. The first one is
designed by the TSO and is called the balancing mechanism which adjusts the gen-
eration and consumption. The other one is a market where the actors can exchange
generation to reduce their imbalances.
The first purpose of the balancing mechanism is to ensure the security of the
system. In particular, it aims at maintaining constant the frequency and the voltage to
all points in the networks. To achieve this purpose, France requires some obligations
from all of the market players who own generation assets which can contribute to
the reserves. For instance, coal-fired plants, nuclear power plants, and hydroelectric
power plants can contribute while wind farms cannot because of the way they are
presently connected to the grid. Market players have the obligation to offer all of
their available generation to the TSO so that he or she is able to increase or decrease
generation if needed. These bids can involve a non-negligible level of complexity
because they tend to reproduce precisely the dynamical constraints of the groups.
At this time frame and for the purpose followed by the TSO, it is not possible to
neglect the notification time, ramp constraints, and so on. But, basically, the power
plant owners submit a quantity together with a price and a maturity, and the system
operator selects them according to their economic efficiency.
A second objective of this mechanism is to provide a transparent market price for
the cost of the imbalances in the system. If the TSO has to adjust generation, then
it is because at least one actor is not generating as much power as its customers are
consuming or is generating too much power. Thus, this actor has to incur the costs of
the adjustments made by the TSO. It is done at a price referred to as the imbalanced
price.
To give a concrete example of this mechanism, Table 2.1 shows the way the imbal-
ance price is settled by the RTE, the French TSO, for a given hour h of the day. In
this table, S represents the day-ahead price settled the day before for the hour h, P d
is the weighted average price of the offers used by the TSO to decrease the genera-
tion (or increase the consumption), P u is the average price of the offers used by the
TSO to increase the generation (or decrease the consumption). The table reads in the
following way.
When the network needs to be adjusted upward (lack of generation, first column)
and the actor is itself generating too much (first row), then the actor is paid at the
spot price established the day before for each MWh. In the opposite, if the actor is
itself not producing enough (second row), it pays each MWh of its imbalance at a
price that is always higher than the spot price and the average cost incurred by the
TSO to compensate for its generation deficit.
2.2 Market Microstructure 11

Table 2.1 French TSO imbalance price settlement mechanism as of April 2013
Network Network
adjustment trend adjustment trend
positive negative
 
Pd
Actor imbalance positive: actor is paid S min S,
1+k
Actor imbalance negative: actor pays max (S, P u × (1 + k)) S

When the network needs to be adjusted downward (excess of generation, second

column), and the actor is producing too much (first row), the actor is paid, but at a
price that is less than the average cost incurred by the TSO to decrease the generation
(P d ), and less than the day-ahead spot price.
Although apparently complex, those rules are basic. It is natural to pay those who
produce energy even if the system has too much of it and to make pay those who
do not produce enough even though the system has too much electricity. Moreover,
the system is designed to avoid any arbitrage opportunity between the spot market
and the balancing mechanism. If the payment for having a negative imbalance in the
case where the system is lacking power (first column, second row) is not floored by
the spot price, then the temptation exists to sell power on the spot market at price S
and to not deliver it. This action leads to a negative imbalance position that would
be penalised in some cases at a lower price, leading to a profit.
Moreover, the factor k ensures that the TSO does not have a negative balance
sheet at the end of the year because the TSO always pays those who contribute to
correct the imbalance less than what the TSO pays to correct them.
Besides this mechanism where the TSO is the only buyer or seller, a market exists
where players can adjust their own positions between their generation and the con-
sumption of their customers. In this market, players are constantly exchanging power
to adjust their own perimeter over the next few hours. There is growing interest by
market players in this market because of the increase in short-term uncertainties. This
increase in uncertainty is mainly due to renewable energies. For instance, in France,
these exchanges can be done over-the-counter on the EPEX platform. Electricity can
be bought or sold as much as 32 hours in advance before delivery. The exchanges
performed there are not marginal. For certain hours, the exchanged volumes can be
as large as 5 GW. Moreover, this market is not limited to France. It includes Germany,
Austria, and Switzerland. This market shares some features with forward markets.
The actors can take a commitment for delivery on a future day but reverse it in the
next minute. Figure 2.2 shows the index price of the intraday market on EpexSpot
(weighted average of the intraday transactions) as a function of the day-ahead price
in 2012. The figure shows that there was a spike on the day-ahead market while
nothing happened on the intraday, and, more interesting maybe, there was no spike
on the day-ahead market but one occurred on the intraday. Price caps might differ on
the day-ahead and on the intraday markets. For instance, on EPEX, the prices on the
12 2 Electricity Markets

1200

1000

Epex intraday 2012

800

600

400

200

−200
−500 0 500 1000 1500 2000
Epex dayahead 2012

Fig. 2.2 Relation between the average intraday price for delivery at a given hour and the day-ahead
spot price for the same hour in 2012. Source EpexSpot

day-ahead market range from −500 to +3.000 e/MWh while on the intraday market
they move between −10.000 and +10.000 e/MWh.
Figure 2.3 shows an example of the evolution of the transaction price for the
delivery of a given hour on the German intraday market in December 2010. The price
moves from 70 e/MWh (close to the reference provided by the spot) to 120 e/MWh
before going back to 80.

120

110

100

90

80

70

60
15/22−48 16/00−00 16/01−12 16/02−24 16/03−36 16/04−48 16/06−00

Fig. 2.3 Intraday transaction price for delivery at 7 a.m. on 16 December, 2010, on the German
intraday market. X-axis reads: day/hour-minutes. Source EpexSpot
2.2 Market Microstructure 13

2.2.2 Day-Ahead Market

The day-ahead market is based on a fixed trading auction. Each day the market
participants submit bids before a certain time (around 10:00 a.m.). They can bid
(sales or purchases) for a particular hour of the next day or for a set of hours (order
block). The bids of the market participants for a particular hour form two curves as
Fig. 2.4 shows. One represents the sales curve and one the purchases. Then, around
12:00 p.m., the market organiser clears the market: the organiser fixes a price for
each hour of delivery and determines the sellers and the buyers. The market players
then have enough time to send the generation orders to their power plants and send
their schedule to the TSO.
Moreover, Fig. 2.4 illustrates the fact that it is possible to submit negative prices
for buying and selling. A negative sale price indicates that the seller is ready to pay
to sell, and a negative purchase price indicates that the buyer is ready to be paid to
buy. This phenomenon is the consequence of the lack of flexibility in some thermal
power plants. It can be less expensive to let a coal-fired plant run during hours of the
day when the spot price is below its fuel cost than shutting it down and starting it up
again later. On the demand side, some customers also have the flexibility to increase
their consumption if they are compensated for the cost of changing the schedule
of their production process. The figure shows that Sunday the 16th of June, 2010,
had particularly low demand during the night, the price settled at −154 e/MWh, but
some actors were ready to pay as much as 500 e/MWh with a volume of 4 GWh.
Moreover, because the market participants are allowed to submit block orders, the
clearing process results in a non-convex optimisation problem for which defining a
market price requires caution. For details, see the documentation regarding the market
coupling algorithm named Cosmos available on EpexSpot’s website [78].

Hour 6
200

100
Price (euro/MWh)

−100

−200

−300

−400

−500
0 2000 4000 6000 8000 10000 12000
Volume (MWh)

Fig. 2.4 Day-ahead market volume of sales and purchases on 16 June, 2013. Source EpexSpot
14 2 Electricity Markets

Fig. 2.5 German daily 350

day-ahead electricity price
300
from 2006 to June 2012.
Source EpexSpot 250

200

150

100

50

-50

-100
07/07/2005 19/11/2006 02/04/2008 15/08/2009 28/12/2010 11/05/2012 23/09/2013

A direct consequence of the fact that electricity cannot be stored is the occurrence
of extreme spikes in the day-ahead market. This phenomenon is illustrated in Fig. 2.5
with the daily average of the German spot price during the years 2006 to 2012. The
figure shows that the daily average price reaches values as high as 300 e/MWh, but
it reaches values as high as 2,400 e/MWh during one hour. The figure also shows
occurrences of negative price spikes and their disappearance since 2009 with the
development of solar energy.
As Fig. 2.6 shows, electricity day-ahead prices exhibit all of the seasonal patterns
of the economic activity of a country: daily, weekly, and annual seasonality such
as the moment people wake up to go to work, the moment they get back home and
switch on their home appliances, the moment they go to sleep, and the moment when
offices and some factories close for the weekend. Electricity day-ahead prices also
show fat tails and long memory. They also exhibit correlation with the temperatures
in countries with electric heating or air conditioning.

Fig. 2.6 French daily 350

electricity consumption in
relation to the daily spot 300
prices during the year 2007.
Consumption is scaled to fit 250
the picture. Source EpexSpot
and RTE 200

150

100

50

0
19/11/2006 08/01/2007 27/02/2007 18/04/2007 07/06/2007 27/07/2007 15/09/2007 04/11/2007 24/12/2007 12/02/2008
2.2 Market Microstructure 15

In the European Union, each country has its own electricity day-ahead market
cleared by its own market operator. Without coordination, the resulting quoted prices
might provide the wrong signals when compared to the transit flow between countries.
Indeed, an inquiry performed in October 2004 by the French Ministry of Finance
on electricity prices showed that there was barely no relation between the French-
German price spread and the transit on the French-German interconnection [147,
Sect. I.4.1.1].
Because the quoted day-ahead prices by the market operator have a transparency
function, a mechanism has been developed to ensure a consistent relation between
cross-border transactions and local day-ahead prices. This mechanism consists in a
market coupling process which implements a decentralised implicit auction mech-
anism. It is the same mechanism as in the NordPool market. In each country, the
market participants do not have to care about finding a counterparty in neighbouring
countries. The participant just has to submit its bid in its country (sell or buy). Then,
the market organisers perform a clearing process with transport constraints implied
by the available transfer capacity. If there are no binding transit capacity constraints,
then there is a single price for the clearing area. If there is at least one binding tran-
sit capacity constraint, then two prices emerge. The debate in the literature on the
congestion management mechanisms for cross-border trading is important. For an
introduction to this debate, one can begin with Ehrenmann and Neuhoff’s [82].
Due to the importance of the relation between the spot and futures prices in the
pricing theory, it is important to know what the spot price for power is. Sometimes,
the intraday market is referred to as the spot market. Because the intraday price is the
shortest time to maturity, it might appear as the real spot market, in a sense where the
spot refers to a price for instantaneous delivery. But, the reference price for delivery
in the futures market is the day-ahead market price. For this reason, I refer to the
day-ahead market as the electricity spot market.

2.2.3 Forward Market

Electricity forward markets share many aspects with those of storable commodities
such as oil, coal, or metals. In all of the countries where an electricity market exists,
the organised markets have developed and proposed standardised contracts for future
delivery as well as a margin call mechanism to reduce the counterparty’s risk. As in
any other commodity contract, the standardised contracts for future delivery specify
the currency, the underlying volume, the location of delivery, the trading period, and
the tick size. Because electricity is the same in all of the parts of the network grid,
the question of quality is not relevant. Physical delivery is often preferred. However,
a financial settlement is getting more frequent as markets become more mature.
Moreover, because they are doing it for other commodities or financial products,
organised markets report the prices and quantities of the OTC contracts.
The most important difference in the electricity forward contracts from the storable
commodities comes from their term structure. Because electricity cannot be stored
16 2 Electricity Markets

and because the spot market is a day-ahead market with an hourly time-step, a
power operator who wants to hedge its generation for the next year needs to have
at its disposal the forward contracts for all of the hours in the year. For a non-leap
year, the quantity corresponds to 8,760 hourly forward contracts. Such a forward
structure would result in a huge dispersion of liquidity because the market participants
would spread their needs to all of those contracts. For this reason, electricity forward
contracts aggregate hours during a delivery period. The delivery period specifies all
of the hours during which the electricity should be delivered. The choice is not left
to the seller to pick the time for delivery during the delivery period as, for instance,
is the case for crude oil futures contracts on the New York Mercantile Exchange
(Nymex). The delivery period refers to a month, a quarter, a year, or even a week or a
day. The contract also is precise on the schedule of the delivery: the base-load if the
electricity is to be delivered during all of the hours of the period and the peak hours
or off-peak hours if the electricity is to be delivered only during a special period
which reflects large or low demand.
This aggregation mechanism results in a sparse structure of the electricity forward
curve. The further the maturity, the longer the delivery period. For instance, the
European Energy Market (EEX) offers forward contracts with the following term
structure for delivery in the German market: 6 Calendars, 11 Quarters, 9 Months,
4 Weeks, 2 Weekends and 8 Days.
These contracts can come in three different delivery forms: base-load (every hour
of the delivery period), peak-load (7:00–20:00 Monday to Friday), and off-peak
(base-load minus peak-load). The calendars, quarters, and months come in those
three forms while weeks, weekends, and days come only in base and peak forms.
So, each day, 106 contracts are available, which represents only a very small fraction
of the 525,584 h of the next six years. Nevertheless, despite this concentration of
trades on a small number of contracts, liquidity remains an issue in the electricity
forward market. It is not uncommon to report no more than 20 MW available per day
for two-month-ahead contracts—to be compared with the generation capacity of a
standard coal power plant of 400 MW.
Figure 2.7 gives the evolution of the German month-ahead and year-ahead base-
load contracts from 2006 to mid-2012. The behaviour of their prices is similar to
standard financial products. They do not exhibit strong seasonal patterns with spikes.
The jumps in the month-ahead prices come from the seasonality effect of changing
contracts. Further, the month-ahead contracts present a much higher volatility than
the year-ahead ones. The figure also shows that the effect of the boom on commodities
during the years 2005 to 2008 and the financial crisis of August 2008 are reflected
in the year-ahead price. It rose from 50 e/MWh to 100 in 2 years and went back to
45 in less than 6 months.
Figure 2.8 shows the different possible shapes of the German forward curve. The
figure shows the relative position of the daily spot price in normal backwardation
(left: the spot is slightly above the futures price) as well as in contango (right: the
spot is clearly lower than the futures prices). If the month-ahead and year-ahead
contracts tend to be aligned to form a nice configuration, then the spot price does
not fit the plan and tends to exist independently of the remaining forward curve. The
2.2 Market Microstructure 17

100

90

80

70
Euro/MWh

60

50

40

30

20
23/02/2004 19/11/2006 15/08/2009 11/05/2012 05/02/2015
Time

Fig. 2.7 Daily quotation of German month-ahead (blue solid line) and year-ahead (red dotted line)
futures contracts from 2006 to June 2012. Time in French format dd/mm/yyyy Source EpexSpot

Fig. 2.8 Illustrations of four different dates of the German electricity forward curve with only
year-ahead (red dots) and month-ahead (blue dots) base-load contracts. The black dot corresponds
to the average spot price on the dates. The X-axes are the delivery date, and the Y-axes are the price
in e/MWh
18 2 Electricity Markets

figure at the bottom also shows that the shapes of the term structure can be much
more complex. It gives an illustration of the strong seasonal pattern that can affect
the month-ahead contracts while the year-ahead part of the curve remains perfectly
flat.
Convergence: Another important issue regarding the electricity forward contracts
concerns the settlement mechanism of the delivery period. The electricity futures
contracts are settled against the average day-ahead price of the delivery period. This
mechanism ensures that the electricity futures contracts provide the hedge required
by the market participants.
For example, on 1 March, a power operator owns a power plant of 1 MW. The April
base-load futures price is 40 e/MWh. That price suits the operator’s objectives and
so, he or she immediately sells one contract. On the last business day of March, the
April contract settlement price is 45 e/MWh. Thus, the operator has lost 5 e/MWh
on the futures value but still holds the futures contract during the delivery period.
During the month of April, independently of what is happening in the futures market,
the operator operates the plant on a base-load schedule each hour with full power.
Then, the operator receives exactly the average April day-ahead market price. In this
example, the average April day-ahead price is 30 e/MWh. From the April futures
contract the operator holds, he or she receives for each hour of April, (45 − 30) =
15 e/MWh. In the end, the operator receives −5 + 45 − 30 + 30 = 40 e/MWh,
which was the selling price objective.
This example shows that, contrary to other commodity markets, the operator
cannot rely on the convergence of the futures to the spot price to reverse the operator’s
position in the financial market on the last days of the contract quotation.
In this example, if the operator buys back the April contract on the last business day
of March, believing that the average day-ahead price will settle exactly at 45 e/MWh,
then the result will be completely different. First, the operator still incurs the loss
of 5 e/MWh from the margin call. But, now the operator receives 30 e/MWh on
the spot market with no hedge. With the 5 e/MWh loss on the futures market, the
operator’s selling price is 25 e/MWh. The 15 e difference comes from the basis risk
between the last quotation of the April base-load and the resulting average day-ahead
price.
However, the actors on the market do not hold a year-ahead contract during its
whole delivery period. Instead, as soon as the contracts with shorter maturities appear,
such as quarters-ahead and months-ahead, they sell their longer maturities and buy
the shorter ones. This procedure is automatically implemented in the EEX calendars
and quarters contracts in a mechanism called cascading. A few days before the
delivery period, every open position on a year-ahead futures contract is replaced by
an equal position in the three months from January to March and the three quarters
of April, July, and October.
Options market: There are quotations available for vanilla options on the futures
contracts on the organised markets like the EEX or the NordPool. European calls and
puts are open on the futures contracts with different strike prices and three different
maturities. For example, on the last trading day of 2012, there were three calls and
three puts with the expiration dates of 13 April, 13 July, 13 October and four calls
2.2 Market Microstructure 19

Fig. 2.9 Prices of call options on year-ahead futures base-load contract with expiry on January
2013 as of 20 November, 2012, (left) and their implied volatility (right) as a function of the strike
price. The solid red line indicates the options in the money

and puts with the expiration date 14 January for the year-ahead contracts in 2014.
The same year-ahead contracts were available for 2015 and 2016 with the slight
difference that more calls and puts were available for the last expiration dates. Also,
only the base-load contract was available. Thus, these contracts resulted in only
82 options available for the trade of the year-ahead contracts. The calls and puts
were also available for base-load months and quarters. The prices for those contracts
were provided even though there were no transactions and no open interests. Some
contracts were barely traded while the options with an expiry on N January with
the underlying year N grabbed all of the liquidity. Thus, in 2012, the options on
year-2013 with expiry on 13 January represented all of the open interest for options.
They represented an average of 58 contracts with a maximum of 300, and calls and
puts were equally traded.
Figure 2.9 provides an example of the prices quoted on 20 November, 2012, for
call options on the year-ahead base-load futures contracts for 2013 with expiry in
January 2013. I also provide an estimation of their implied volatility. This figure
depicts a situation where there is some open interest in most of the options with an
available price. It shows that the implied volatility can produce a nice smile with a
volatility of around 30 %.

2.2.4 The Diversity of Electricity Markets

The general presentation above hides a great deal of heterogeneity in the development
of the electricity markets around the globe. Sioshansi and Pfaffenberger’s book [153]
on the reform implementations in the worldwide electricity market shows that these
markets have followed very different paths. The markets in some countries have
important growth both in terms of volume and in the complexity of their products.
An example of such developments is the Pennsylvania, New Jersey, and Maryland
market (PJM) in the United States. The PJM market has existed since 1997 and
20 2 Electricity Markets

now serves more than 60 million customers in the Mid-Atlantic and the District of
Columbia. Its total annual consumption is approximately 800 TWh with more than
180 GW of installed capacity of which 60 % are gas-fired plants. The spot market
is a bid-based mechanism with security constraints which deliver prices for each
node of its operating network (more than 10,000). In addition to offering prices for
energy, PJM also provides its operators financial transmission rights to hedge the
locational risk spread. Since 2007, it has launched a capacity market to ensure long-
term reliability of the system and to provide its investors with a market signal on
the value of the capacity. Its futures and options are also available on the Nymex.
The contracts come in different delivery time schedules (off-peak, peak, day-ahead,
month-ahead) but also in different locations of the PJM network.
In Europe, a similar development has occurred in Scandinavia with the NordPool,
in continental Europe with the EEX, and the market coupling mechanism progres-
sively developed across Western countries.
The NordPool was launched in 1996 and covered only Sweden and Norway. It
now connects Norway, Sweden, Finland, Danemark, Lithuania, Latvia, and Estonia
for an annual generation of approximately 480 TWh generated by 370 companies.
Its hydroelectric generation accounts for 100 TWh and its nuclear for another 100.
The remaining generation comes from coal- and oil-fired plants. The NordPool’s day-
ahead spot price is operated by ElSpot and consists of an implicit auction mechanism
between all of the countries covered by the market. An intraday market also exists
which allows the NordPool to secure the imbalances. The intraday market has seen
important growth thanks to the introduction of wind energy and the increasing need
for intraday rescheduling. The financial market for Scandinavian electricity formerly
known as Eltermin is now operated by Nasdaq OMX Commodities. It offers futures
and options that cover the daily, weekly, monthly, quarterly, and the annual horizons.
Each underlying is given by the hourly day-ahead spot price fixed by ElSpot.
These two markets (PJM and NordPool) disclose not only prices but a large amount
of hour data, such as many time series on generation per technology, transit, planned
outages, inflows, and the states of the reservoirs, which is all of the information
needed by the market participants to precisely assess the equilibrium between offer
and demand.
However, some electricity markets have not developed as much as might have been
expected in terms of financial derivatives. This is the case for Chile. The Chilean
electricity system is basically a one-dimensional network due to the geography of the
country. The market is decomposed into three areas, but the main part consists of the
central interconnected system which covers around 93 % of the Chilean population
and 70 % of the total installed capacity. The generation is a mix of hydroelectric
generation (50 %), natural gas (27 %), coal (10 %), and oil (7 %). The main feature of
the spot market is that it relies on a cost-based economical dispatch performed by the
Economic Load Dispatch Centre. The dispatch minimises the expected discounted
value of the generation cost of serving demand in the next 48 half-hours. The system
price is the running cost of the most expensive unit of generation used to satisfy
the demand. The dispatch is mandatory. A price cap is set by the regulator every six
months according to the value of the load lost. Moreover, each power plant receives a
2.2 Market Microstructure 21

monthly capacity payment based on its availability. Despite the fact that this country
was the first to establish a pool market for electricity in the beginning of the 1980s,
there is not a developed market of standardised futures and options contracts.
This lack does not mean that the development of a financial market is not pos-
sible. New Zealand is an example of the late development of a financial market for
electricity. New Zealand is a country composed of two islands with a population of
4 million people. They have an annual electricity consumption of 36 TWh and an
installed capacity of 9 GW. Deregulation began in 1987, and the spot market began
in 1996. It is similar to the PJM design: bid-based security constraints with nodal
prices providing 48 half-hourly prices for each of the 285 nodes in its network. The
forward contracts are traded on ASX, New Zealand’s electricity futures and options
exchange. The exchange launched in December 2013 trades in base-load monthly
futures, peak-load quarterly futures, and the average rate options on base-load quar-
terly futures.
There is no explanation for the diversity of the financial markets’ development.
Possibly it comes from the generation mix, the network configuration and most of
all, the amount of consumption. For instance, Cyprus is a European state where
an electricity market should be designed. It has 5 TWh of annual consumption that
relies mainly on three fuel power plants for 95 % of its generation. Nevertheless, the
evolution of the wholesale market of England and Wales may indicate that the design
of the spot market can be a key driver in the development of the financial market.
Indeed, the first phase of the English and Wales market in 1992 relied on a mandatory
economical dispatch very similar to the Chilean model. During that period, there was
not to my knowledge a financial market for futures similar to the one in NordPool.
But, now after several changes in the design of the English and Wales power market,
the futures and options can be traded on the Nasdaq OMX Commodities board with
the standard variety of products available (day, week, month, quarter, year, and base-
load and peak-load).

2.3 Real Derivatives

Like any other commodity market, the electricity market has its options markets on
quoted futures. But, the most challenging problems do not come from the valuation
of these standard products but rather from their pricing, hedging, and structuring into
exotic tradable products which are called real derivatives. The electricity operators
as well as the retailers have these options embedded in their portfolios. A power plant
can be seen as a strip of call options. A retail household contract with a curtailment
clause which gives the retailer the right to charge the customer an exceptionally high
price if the customer does not reduce his or her consumption on a sent signal, is a
contract with an embedded put option. Thus, as an analogy with the idea of the real
option for investment decision, these rights can be thought of as real derivatives.
Moreover, because electricity cannot be stored, the TSO always needs some
flexibilities to be able to cope with uncertainties but also to cope with the dynamic
22 2 Electricity Markets

constraints of the generation system. It is sometimes more economical to be able to

reduce consumption or to defer it by a few hours than to satisfy it right away. On the
contrary, it might be better sometimes that the electricity consumption is higher to
avoid shutting down a power plant which is not flexible. An example of these flexibil-
ities can be illustrated by household water heaters. In France, these home appliances
are automatically started up with an electric signal when the consumption is at its
lowest during the night. As Remark 2.1 points out, the development of intermittent
sources of renewable energies has led to a growing interest from all of the market
actors in all of these flexibilities. The flexibilities can be seen as real options bought
by the TSO from the users or by any involved actors from the customers.
The following are the three most important real derivatives that are the daily
concerns of the operators and their financial representation:
1. Power plant and tolling agreement
2. Energy storage and swing contract
3. Retail contract.
Power plant and tolling agreement: The owner of a power plant creates value
by selling power and buying fuel. The owner needs to estimate the value of the
generation of a power plant for investment decisions, long-term contract negotiations,
and risk management applications. A power plant is an industrial facility and thus, is
subject to many technical, environmental, and legal constraints. But, in a first crude
approximation, a power plant can be seen as a strip of calls on the spread between
its fuel price and the electricity price. For a gas-fired plant, one speaks of a spark
spread; while for a coal-fired plant, one speaks of a dark spread. When the carbon
emission price is taken into account, then it is called a clean dark spread and a clean
spark spread. The payoff per MW of a power plant in a period of time [0, T ] is then
given by:
 T
(St − h St − gStc )+ dt,
f
(2.1)
0

where S is the spot price of power, S f is the spot price of its fuel, and h is its heat rate.
The variable S c is the spot price of the emission permit, and g is the emission factor
of the plant. The notation x + refers to max (0, x). The payoff (2.1) makes it clear
that it is possible to identify the value of a power plant with a strip of call options.
Nevertheless, the fact that the instantaneous payoff depends on three assets makes the
problem technically difficult. Moreover, the existence of the operational constraints
drastically reduces the possibility of capturing the successive positive spreads. The
start-up costs, minimum running time, ramp-up and ramp-down constraints, and
the limited number of cycles per day, all limit the flexibility of the power plants.
Thus, the quantity (2.1) overestimates the profit of the power plant. Valuating power
plants with operational constraints gives rise to optimal control problems. One way
to formulate it is:
 T 
f +
sup E qt (St − h St − gSt − κ) dt ,
c
(2.2)
qt ∈A 0
2.3 Real Derivatives 23

where qt is the generation belonging to an admissible set A , and κ is a start-up cost.

However, the problem above is not new in generation management. It is related
to unit commitment problems, that is, scheduling a group of power plants for the
next few hours or days. And it is also related to management problems in mid-term
generation where the assessment of the generation level of a group of power plants is
needed for the next few months. Over the last 40 years, a lot of efforts has been devoted
to the development of numerical optimisation algorithms and software to solve those
problems. These models are based on a context where monopolies try to satisfy a
random demand at the lowest cost. They use mix-integer programming methods or
Lagrangian relaxation methods. The introduction of spot markets has changed the
problem by introducing new phases like the bidding phase where the power operators
have to submit their bids to the market operator. However, the underlying optimisation
methods still apply. For a more in depth description of the constraints in power plant
scheduling and numerical optimisation methods, the reader can consult Wood et al.’s
monograph [169] and the recent survey of Prekopa et al. [146].
The major novelty raised by the development of financial market is the question
of the hedge of the payoff (2.2). Typically, the natural hedging instruments are the
futures on the fuels, carbon emissions, and the futures on power. But, as we seen in
Sect. 2.2.3, the futures contracts available on the electricity market do not have the
fine granularity of the spot price. If T is one year, the above payoff is exposed to
8,760 risk factors whereas only a handful of futures are available. In this regard, the
electricity market is incomplete. As a consequence, even the notion of the value of
the first basic payoff (2.1) is ambiguous. At the present time, it has no consensual
answer. This lack of consensus translates immediately into the difficulty of finding
exchange prices for tolling contracts. These contracts are financial counterparts of
the power plants where the owner concedes an exploitation right in exchange for a
fixed premium. Depending on the price models, the constraints taken into account
and the hedging capacities, there might be a large discrepancy between the prices.
Energy storage and swing contract: Hydroelectric plants are very flexible. They
can provide electricity at very short notice. The valuation problems described for the
thermal power plants are still valid for hydroelectric power plants. But, the existence
of a limited resource of fuel leads to the problem of storage management. The simplest
problem in hydroelectric storage management deals with a single reservoir. The
valuation problem is:
 T 
sup E qt St dt + g(ST , X T ) , (2.3)
qt ∈[0,q],δt 0

where St is the electricity spot price, and X t is the current level of the water in the
reservoir. This level satisfies the following dynamic:

d X st,x = (ast,a − qs − δt )ds, (2.4)

24 2 Electricity Markets

where as are random inflows. Moreover, X is subject to level constraints and should
stay within [x, x]. The control δt is the spilling variable. If X t is at its maximum level
and the inflows exceed the generation capacity q, the operator has no other solution
than to spill the excess water. The function g represents a final value for having a
certain level of water at a final time.
In a more general form, a whole network of hydroelectric plants and reservoirs
influence one another. The constraints on the hydroelectric plants become more com-
plex and the dimension of the system drastically increases. Hydroelectric manage-
ment problems are not new. They were at the origin of the development of dynamic
programming [132]. There is important literature on power systems that is devoted to
these problems. The main difficulty in these problems is the dimension of the state.
The higher the number of reservoirs, the higher the dimension. They are mainly tack-
led by using stochastic dynamic programming methods, such as the dual dynamic
programming method [139], or decomposition methods.
The development of the electricity markets launched a revival of this problem
because of the availability of spot prices. Other methods have then been proposed
based on the optimal switching problem or the stochastic control problem.
As was the case for power plants with tolling contracts, the financial representa-
tion to storage facilities, named swing options, were developed to offer the ability to
store power but without all of the technical constraints involved with a real hydro-
electric system. An important class of swing contracts consists of the demand-side
management contracts. The customers holding this form of contract enjoy a lower
tariff during the year except for a few days when any kWh consumed costs the cus-
tomer a dissuasive price. What is important here is that the days with the higher price
are not known in advance by the customer. The utility warns the customer the day
before. Thus, the utility holds an option on the customer’s consumption. When the
customer is an industrial firm, the result of the signal is deterministic. The firm can-
cels its consumption. But, when the contracts concern a large amount of households,
the resulting total avoided consumption is random.
Retail contract: The retail pricing policy consists in determining the form of
the contracts for customers as well as the price to be charged. As in the insurance
business, retailers first have to decide on the forms of their contracts and their prices;
then, the customers select the contract and consume. When setting the retail pricing
policy, a retailer does not know what is going to be its sourcing cost and precisely
how much the customers are going to consume.
The consumption of the customer drastically differs from a chemical plant to a
household in a building in a big city. The customers are generally sorted into three
main classes: industrial, professional, and household. They differ by their number,
their consumption volumes and patterns, their economic behaviour, their needs, and
by the information the retailers have on their consumption. In a country like France
or Germany, industrial customers are much less numerous than households but they
represent a significant part of the country’s consumption. The industrial customers
can benefit from very precise metering of their consumption whereas households
are still only measured twice a year in many countries. The industrial customers
might not be sensitive to weather whereas households are. The industrial customer
2.3 Real Derivatives 25

can switch from one retailer to another for 1 cEuro/MWh because each cent can
represent thousands of euros on the customer’s electricity bill. Household switching
is more viscous.
For these reasons, the contracts offered to different classes of customers are very
different. The industrial contracts might be tailored to their operational constraints
whereas the households are viewed as a mass market that only needs basic stan-
dardised contracts. The industrial retail contract often includes embedded options
(right to resell to the market the power delivered at a certain price). The household’s
contracts involve fixed premiums proportional to the subscribed capacity plus a price
per kWh. Sometimes they benefit from a two-level tariff depending on the hour of
consumption (peak and off-peak). For a case where a detailed description of a retail
market is given, the reader is referred to the analysis of the Norwegian household
market in von der Fehr and Hansen [161].
This problem has received much less attention in the quantitative energy finance
literature than in the economics literature. Indeed, electricity pricing policy was
addressed as a problem of long-term pricing during the monopoly period of electricity
generation. In that economy, the main concept which dealt with this problem was the
marginal cost pricing theory, deeply illustrated in theory and in practice by Boiteux
[38]. In a nutshell, the marginal cost pricing of electricity insures a maximisation of
the social welfare through a regulated monopoly. With market competition and with
the real gross market prices, nothing insures that selling at the marginal generation
cost is the right policy to maximise the value of the firm. Indeed, the marginal
cost pricing rule requires certain long-term equilibrium conditions on the generation
assets to be fulfilled which are unlikely to hold in actual market conditions.

2.4 Conclusion

The structure of electricity markets can be seen as a multi-layered market with

successive time horizons going from the forward market of up to six years to the
real-time balancing market for the next hour. The electricity forward curve is coarse
and presents a cascading structure. The spot prices are defined by the day-ahead mar-
ket. They present strong seasonality patterns, spiky behavior, and positive as well as
negative prices. Further, power plants, tolling contracts, storage and swing contracts
and retail contracts are some of the most specific derivatives to electricity markets.
Chapter 3
Price Models

This chapter is devoted to the modelling of the electricity spot and futures prices.
Since the electricity market’s deregulation, an important and increasing number of
publications are dedicated to the problem of modelling the electricity spot price. I
exclude from the scope of this chapter the research activity on the forecasting of the
spot prices. The chapter is limited to the models that provide the dynamic of the
futures prices. Indeed, modelling the dynamic of the futures prices is the basic brick
for hedging and pricing more complex products. A model should be able to provide
them at a low computational cost.
But, even with this restriction, an important diversity of models exists for the
spot and futures electricity prices. For pricing and risk management applications, an
electricity price model should have the following properties: realism, consistency,
efficiency, robustness, and generality. Realism concerns the ability of the models to
reproduce the statistical properties of the time series of the spot and forward prices. It
can be assessed by different methods with basic to sophisticated statistical tests. The
consistency concerns the no-arbitrage property between the spot and the forward
prices. Although electricity is non-storable, no-arbitrage conditions appear in the
different models in this chapter. Efficiency is a key element. The models should
lead to fast computation of the forward prices, where fast means milliseconds. Since
forward prices are the basis for any pricing or risk management computation, a single
computation has to be performed in sufficiently efficient way to allow the valuation
of more complex products within a reasonable time (a few minutes). Robustness
deals with the continuity of the estimated parameters with respect to the data and the
continuity of the modelled spot and futures prices. A small modification of either
the entry data or the parameters should not lead to important modifications of the
modelled prices. Further, the model should be general enough to allow its estimation
and calibration for the different electricity markets available. Because there are at
least as many electricity markets as there are countries, the models should be generic
enough to avoid the need for a specific development for each.
The comparison of electricity models within the framework above is a research
topic in of itself. Here, this chapter confines itself to the presentation of the models.
But even with this limited ambition, the large spectrum of models that already exist
© The Author(s) 2015 27
R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7_3
28 3 Price Models

raises the difficulty of finding a proper way to sort them out. No taxonomy is perfect
and neither is the classification used here. The prevailing idea is in a sense to sort
them by increasing complexity.
So, the first class, which is presented in Sect. 3.2, consists of models that follow
the Heath-Jarrow-Morton (HJM) approach. Those models start with the observed
forward prices, model them, and then derive the spot price (if needed) as a future
with a zero maturity limit. This is the preferred method used on trading desks and
in risk management departments because it ensures that the model fits the market
prices every day. This approach is inherited from the modeling of the yield curve
for interest rates. Thus, many observations, results and problems are not particular
to electricity forward curve modeling and can be found in classical text books on
interest rates curve modeling like the book by Musiela and Rutkowski [136] and
the book by Brigo and Mercurio [44]. In particular, the constraints on the volatility
functions used in these models apply to the electricity forward curve. But, in the case
of electricity, the existence of a delivery period for futures adds more constraints if
the setting is to remain no-arbitrage.
The other classes of models take the opposite approach to the HJM method. The
electricity spot price is modelled first, and the futures prices are then deduced as an
expectation through a convenient probability measure from an arbitrage argument.
A first category of electricity price models which follow this procedure is presented
in Sect. 3.3 and consists of one-factor spot models. These models illustrate the main
difficulties when preserving the Markov property of the spot, capturing spikes, and
getting the dynamic of the forward curve. Although rich, this class of models does
not offer enough degrees of freedom to capture the main properties of the electricity
price and therefore lead to a need for more complex classes of models.
The third class of models, which is described in Sect. 3.4, consists of multi-factor
models of the spot. This increase in dimensionality can be done either by adding
factors that are observed or not, or mixing them both. These models show a much
more realistic behaviour in the joint dynamic of the spot and the forward prices. The
cost for this result is the loss of the Markov property of the spot and the need for the
filtering techniques used in the estimation of the parameters of the unobserved state
variables.
The last class, which is presented in Sect. 3.5, is structural models. In these models,
the spot is deduced from an over-simplified version of an equilibrium model between
the production cost curve and the electricity demand. The structural models naturally
implement the dependencies of the spot price with the observed factors such as fuel
prices, consumption, and power plant outages. Relying only on observed factors, they
avoid the use of filtering methods for the estimation. But they are challenging for the
computation of futures prices because the spot is generally a non-linear function of
the observed factors.
But, before I proceed on this journey into price models and because of their
fundamental importance to the pricing theory, I devote Sect. 3.1 to the no-arbitrage
conditions and the convergence of the futures price to the spot price in the case of
electricity.
3 Price Models 29

Remark 3.1 In this chapter, the notations used for the same variable (e.g., the spot
price and the mean-reversion value) might change from one model to another. This is
my choice between being homogeneous or keeping true to the original papers where
the models were presented. I thought it would be better to keep the notations from
the original papers for the simplicity of understanding if one was to consult their
papers for a deeper understanding.
Moreover, we will use from now on indifferently the terms futures and forwards
because their difference is not significant in the context of the price models of this
section.

3.1 Preliminary Remarks

For freely storable assets, the spot price St at time t is linked to the forward price
F(t, T ) for delivery at time T by the relation

F(t, T ) = er (T −t) St , (3.1)

with r the interest rate that is constant. This relation is established with no-arbitrage
arguments, whose details can be found in Musiela and Rutkowski [136, Sects. 1.5
and 1.6]. These arguments rely on the storable property of the asset. It has been
acknowledged that the relation (3.1) does not hold for commodities such as cereals,
copper, or oil. The cost of storage for a commodity is not negligible compared to a
stock. This cost should significantly reduce the expected return from buying the spot.
This remark leads to the thinking that the right relation for the storable commodities
should be:

F(t, T ) = er (T −t)+c(T −t) St , (3.2)

with c(T −t) as the cost of storage that is supposed to be only a function of the time to
maturity. But, even with this correction, the relation still does not hold. In particular,
the forward price could be lower than the spot. In an early paper, Kaldor [108]
introduces the idea of a convenience yield to account for that effect. The convenience
yield represents the extra return that the actors gain from holding the commodity
instead of a forward contract. The relation then writes as:

F(t, T ) = er (T −t)+c(T −t)−y(t,T ) St , (3.3)

with y(t, T ) as the convenience yield.

Remark 3.2 The arbitrage and relations above between forward price and spot price
are precisely described in Eydeland and Wolyniec [85, Chap. 4, pp. 140–143].
However, this construction raises questions even for storable commodities because
the convenience yield is not observable. But, in the case of electricity, it barely holds.
30 3 Price Models

Further, because electricity cannot be stored, the storage argument cannot be used
to justify the relation between the current spot price and the forward price. Thus,
even the concept of the convenience yield cannot be applied to this market, at least
without reconsideration of its underlying.
Moreover, a second hypothesis which is important for the modelling of forward
and spot prices is the fact that as the time to delivery gets closer the forward price
should converge to the spot price:

lim F(t, T ) = ST . (3.4)

t→T

This relation ensures that the basis risk between the forward contract and the spot
market is negligible. When this relation holds, an actor can reverse his or her position
on the forward market and buy or sell on the spot market, while still being sure that
the loss from one market will be offset by the gain from the other. In the case of
electricity, Sect. 2.2 shows that the forward contracts involve a delivery period that
leads the operators to implement a cascading mechanism of switching to contracts
with shorter delivery periods each time they appear.
It is not obvious in this situation that the relation (3.4) occurs here. But, it is possible
to test if this convergence holds or not for the electricity futures, for example, the day-
ahead futures contracts. These contracts are currently traded on the European Energy
Market, EEX, for instance. The underlying of those contracts is the average price
of the spot price on the delivery day of the contract. For a contract with a delivery
on Tuesday to Friday, the forecast of the day-ahead spot price is good enough to
ensure a small discrepancy between the last quotation of the day-ahead futures price
and the average spot price. But, because these contracts are financial, they are only
traded on business days. Thus, the last quotation of the day-ahead futures contract
of a Monday is the Friday before. Thus, the difference between the last quotation of
the day-ahead futures contract and the average realised spot price gives an idea of
the quality of the convergence.
The study of this point is undertaken by Viehmann [160]. The author shows that
even for weekend days the discrepancy between the last quotation of a day-ahead
futures quotation and the realised spot is not statistically different from zero on
average for all of the hours of the day. For some particular hours like peak hours
(6 p.m.), the discrepancy can be close to 5 %. Thus, if the reader is not too demanding,
he or she can consider that there is convergence up to a small premium due to the
fact that even from Friday to Monday events might occur during the weekend that
could change the spot price.
A third point in the spot-forward relation for storable assets is that there is a risk-
neutral measure which ensures that the forward prices can be written as the expected
spot price, namely:

Q
F(t, T ) = Et [ST ] . (3.5)
3.1 Preliminary Remarks 31

For valuation purposes, it is better that the risk-neutral measure used here, Q, is
unique. In the case of electricity, the models presented in this chapter generally lead
to incomplete market models. Thus, the uniqueness of a risk-neutral measure is not
guaranteed, and one has the choice amongst an infinity of them. The next sections will
show the different methods used to pick up one risk-neutral measure. In this regard,
the important point is less the theoretical question of the existence and uniqueness of
this risk-neutral measure but the practical capacity of the model to fit all the observed
futures prices.
Another important part of the literature on electricity forward prices is devoted to
the analysis of the sign of the risk premium, namely:

R(t, T ) = F(t, T ) − EPt [ST ] , (3.6)

where the expectation of the spot price is taken under the historical probability.
This premium can have an important economical meaning on the rationality of the
economic agents in the markets. This premium is an indicator of the relative market
power or the relative risk aversions of the producers and the consumers. The risk
premium might also have an important impact on the generation management process
of the utilities. Some decision-makers might judge that the futures are too expensive
when compared to the expected spot price under the historical probability. But,
the futures prices are not actually expensive. They only imply that the decision-
makers forget that their decision is driven by their own risk aversion. Thus, despite
its important economic interest, the question of the risk premium is not necessary to
develop within the scope of a text dedicated to the pricing of derivatives. However,
for an in-depth analysis of this quantity, the reader can consult amongst other works
Bessembinder and Lemon [33], Longstaff and Wang [126], Diko et al. [77], Benth
et al. [21] and Aïd et al. [4].

3.2 HJM-Style Forward Curve Models

The models described here are inspired by the methodology developed by Heath,
Jarrow, and Morton for the debt market [99]. The intention of the authors was to find
an alternative way to the factor model based on the idea of an instantaneous interest
rate similar to the Vasicek style models [159]. The Vasicek approach suffers from
the difficulty of its fit with the observed yield curve. By proposing to instead directly
model the dynamic of the forward rate, Heath, Jarrow, and Morton (HJM) are able to
overcome this difficulty. The same idea has prevailed in the literature on commodity
prices where the first applications can be traced to Jamshidian [106] for crude oil
and Cortazar [66] for copper. I first present the general principle of this method and
then turn to its implementation in the case of electricity.
32 3 Price Models

3.2.1 Principle

Consider the price F(t, T ) given at time t for delivery at the future instant T > t of
a commodity. I suppose from now on that there exists a risk-neutral measure and I
write the dynamic of F(t, T ) under this probability. The dynamic for F is thus given
by
dF 
(t, T ) = σi (t, T )dWit , (3.7)
F
1≤i≤n

where the σi (t, T ) are n volatility functions, and the W i are n Brownian motions
(potentially correlated). With a storable commodity, the delivery times T are discrete
and basically given by the month steps. Moreover, the initial condition required for
F(0, T ) is given by the observed forward prices for each delivery T , which makes
the fit of the forward price at the initial time a basic procedure.
The number of Brownian motions N can potentially be as many as there are
delivery periods. But, the number of factors needed to reproduce the dynamic of the
term structure sufficiently well is much less. The number of factors needed has been
the subject of many studies based on the principal component analysis (PCA). The
first studies were performed on the money market by Litterman and Scheinkman
[124] and Knez et al. [120] who showed that three factors explained more than 86 %
of the total variance while four grabbed 90 %. The PCA performed on the different
commodity markets shows that three factors are enough to capture more than 90 % of
the variance in the futures returns: for copper more than 97 % [66] and for NYMEX
crude oil and gas more than 98 % [65, Sect. 8.4]. Thus, this approach not only remains
in a complete market setting but it only requires a small number of factors.
Furthermore, if needed, the spot St can be defined as a futures contract with
immediate delivery where St = F(t, t). Its dynamic can then be derived from (3.7)
by first writing the integral form of F(t, T ) as
⎛ ⎞
 t   t 
1
F(t, T ) = F(0, T )exp ⎝− σi2 (s, T )ds + σi (s, T )dWis ⎠ . (3.8)
2 0 1≤i≤n 0 1≤i≤n

And, because St = F(t, t), the use of Itô’s lemma shows that the spot follows the
dynamic:
⎛ ⎞
dSt ⎝   t  t
= ∂2 ln F(0, t) − σi (s, t)∂2 σi (s, t)ds + ∂2 σi (s, t)dWis ⎠ dt
St 0 0
1≤i≤n

+ σi (t, t)dWit , (3.9)
1≤i≤n

where ∂2 denotes the derivatives of a function with respect to the second variable.
3.2 HJM-style Forward Curve Models 33

Remark 3.3 The different steps that lead to therelation (3.9) are the following. Let


t
t
X t = 0 σ (s, t)dWs and St = g(t, X t ) = exp ln f (0, t) − 21 0 σ 2 (s, t)ds + X t .
Thus
1
dS = ∂1 g · dt + ∂2 g · dX + ∂22 g · dXt , dXt .
2
Moreover,
  t
1 2
∂1 g = g × ∂2 ln f (0, t) − σ (s, t)∂2 σ (s, t)dt − σ (t, t)
0 2

It holds that ∂2 g = ∂22 g = g. And, one has

 t
dXt = ∂2 σ (s, t)dWs dt + σ (t, t)dWt
0

and
dXt , dXt  = σ 2 (t, t)dt.

Thus, the term − 21 gσ 2 (t, t) from ∂1 g cancels out with the term + 21 gσ 2 (t, t) coming
from 21 ∂22 g · dXt , dXt .

An example and an important case for application is a single factor with an expo-
nentially decreasing volatility with respect to maturity σ (t, T ) = σ0 e−a(T −t) . By
remarking that
 t  t
1
σ (s, t)dWs = ln F(t, t) − ln F(t, 0) + σ 2 (s, t)ds, (3.10)
0 2 0

this specification leads to a mean-reverting spot price dynamic given by:

dSt
= (μ(t) − a ln(St )) dt + σ0 dWt , (3.11)
St

with
σ02
μ(t) = ∂2 ln F(0, t) + a ln f (0, t) + (1 − e−2at ). (3.12)
4
This example shows the relation between a term structure with decreasing
volatility and a mean-reverting spot price. But, it should not hide that not all of
the volatility functions are admissible. Because, if they are not suitably chosen, the
Markov property of the spot price might be lost. Further, the relation which shows
the dynamic of the spot (3.9) also shows that the drift term depends on the whole
34 3 Price Models

trajectory of the Brownian motions because of the terms:

 t
∂2 σi (s, t)dWis .
0

Conditions on the volatility functions have to be set to ensure the Markov prop-
erty of the spot (see Musiela and Rutkowski [136], Proposition 11.2.2). Without the
Markov property, any simulation of the prices or any computation of the price deriv-
atives becomes cumbersome because the whole past of the Brownian motions has to
be remembered. In the case where volatility is required to depend only on the time to
delivery T − t, then the only possible volatility functions that preserve the Markov
property of the futures prices are of the form σ (t, T ) = σ0 e−a(T −t) with σ0 and a
constants, which are the most commonly used volatility functions in the literature
and in practice.

3.2.2 The Case of Electricity

A first technical difficulty is the existence of a strong seasonal pattern in the electricity
forward curve. This seasonality requires the introduction of a time-dependent drift
term and a procedure to estimate it. The situation is quite different from the case of the
Black and Scholes model for stock markets where it is known that the estimation of the
drift term under the historical probability requires unreasonable historical time-series.
Indeed, the seasonality of the futures prices are connected to the seasonal patterns of
consumption which are well-known. Several methods have been proposed to estimate
the seasonality of futures prices. The first one consists of using a forecast model as
in Fleten et al. [87]. Because the seasonal pattern of the electricity forward curve is
driven by the seasonality of the demand and of the marginal cost of production, a
spot price forecast model can be used to provide a seasonal trend in the observed
futures prices. In their work, Fleten et al. [87] fit the seasonal pattern obtained by
a fundamental model to the forward seasonal pattern by using a constrained least-
square. They also use smoothness constraints for the variations in the futures prices
with respect to maturity to avoid jumps, for example, between the last daily summer
contract and the first daily autumn contract. They compare this procedure with a
truncated two-term Fourier series. They find a slight improvement with the seasonal
trend given by the forecast model.
A second and more important difficulty lies in the number of factors suitable to
model the dynamic of the electricity forward curve. Many studies undertake this
subject, particularly for the NordPool market by the Norwegian School of Financial
Economics. The analyses performed by Frestad et al. [89, 90, 91] and Koekebakker
et al. [121] indicate that the electricity forward curve exhibits a much more volatile
behaviour than any other market. In Frestad [89], the same PCA is performed on
the market price data for the bond, copper, crude oil, and electricity markets. The
analysis of the dynamic of the forward products of each market is performed with a
3.2 HJM-style Forward Curve Models 35

model based on the asset pricing theory. The author shows that three common factors
account for 67 % of the variance, and the unique factor explains 33 % of the variance
in electricity forwards while three factors succeed in explaining 99 % of the variance
in copper, 100 % for crude oil, and 97 % for bonds. This phenomenon is confirmed
to some extent by Koekebakker et al. [121] who use a modelling approach closer
to the standard HJM approach. Applied to NordPool data from September 1995 to
March 2001, they find that two factors have much less explanatory power than in
other commodities markets. One factor accounts for 68 %, two factors for 75 %, three
factors for 80 %, and four factors 83 %. They need to have more than seven factors
to get more than 90 % of the variance.
This need for an increasing number of factors for electricity forward prices comes
from the unrelated dynamic of the weekly contracts with the longer maturity con-
tracts. The behaviour of the next week’s prices has little to do with the electricity
futures price for the next season or the next year. Moreover, the returns of the for-
ward prices even with not so short maturities might not be Gaussian. Indeed, Frestad
et al. [91] show that the hypothesis of the normally distributed return for the for-
ward prices is outperformed by the normal inverse Gaussian distribution even for
long-term maturity contracts. These two remarks are the main justifications for the
development of the models based on the infinite dimensional Lévy processes, or
ambit fields (see Barndorff-Nielsen et al. [12]).
More difficulties come from the fact that the electricity spot price is hourly while
the forward prices involve a delivery period. In the relation (3.7) which gives the
dynamic of the forward prices, it is not clear what is represented by T . In the case
of electricity, it should be clear if it is a given observed maturity (a month, a year) or
if it is an hour in the future. In fact, for the valuation purposes of the real derivatives
presented in Sect. 2.3, the hourly prices are needed which quickly leads to completing
the observed electricity forward curve with a forward curve of hourly contracts.
The construction of an hourly forward curve for the next three or six years when
there are only a hundred available quoted contracts might be difficult. An important
effort has been devoted to the construction of such a forward curve of instantaneous
delivery contracts in Fleten et al. [87], Koekebakker et al. [121], Benth et al. [26],
Frestad [90]. A detailed description of a method can be found in Benth et al.’s book
[19, Chap. 7]. In Benth et al. [23], for instance, their model starts with

1 
N
F(0, Ts , Te ) = f (0, tk ) (3.13)
N
K =1

where f (tk ) is unobserved futures prices for delivery at time tk , Ts is the time of the
beginning of delivery and Te is the time of the end of delivery. The function f is
decomposed into two continuous terms f = s + ε with s the seasonal term and ε
an adjustment term. The constraints are set on the adjustment term to get a smooth
hourly forward curve. Meanwhile, the seasonal term is estimated as the seasonal
component of the historical spot price. Alternative methods might be developed.
36 3 Price Models

But, in any case, clearly there is no unique solution to this initial hourly forward
curve.
However, I assume the function f (0, T ) of the unobserved price at time 0 for
delivery of 1 MW for the hour T can be constructed. I also assume null interest rate.
If I denote by f (t, T ) the unobserved price at t for delivery of 1 MW for the hour T ,
then the observed forward contract F(t, T1 , T2 ) quoted at time t and involving the
delivery of 1 MW each instantly between T1 and T2 is:
 T2
1
F(t, T1 , T2 ) = f (t, s)ds. (3.14)
T2 − T1 T1

The idea then is to set a dynamic on the non-observed instantaneous delivery

forward price f (t, T ), just like for the instantaneous forward rate (see Heath et al.
[99]). Further, f (0, T ) is used as the initial condition of this dynamic.
But, not all volatility functions are admissible if the Markov property of the spot
price is to be preserved. The fact that electricity forward contracts involve a delivery
period somewhat constrains the admissible volatility functions. This remark is raised
by Benth and Koekebakker [25]. Indeed, they show that under mild conditions, the
relation (3.14) implies that the only dynamic for the return of f that preserves the
Markov property of the spot is to have a volatility function that is independent of
the maturity T . This does not leave many alternatives for the dynamic of f except
the volatility functions which depend only on time t, and amongst them the most
popular choice is the Black model: constant volatility (see Black [37]).
Moreover, Benth and Koekebakker [25] obtain similar results with alternative
ways to specify the dynamic of the forward curve. They take one-step backward and
assume that it is possible to specify the dynamic of the forward prices F(t, T1 , T2 )
with the relation of the form:

dF(t, T1 , T2 ) = Σ(t, T1 , T2 )F(t, T1 , T2 )dWt , (3.15)

which holds for any pair of T1 , T2 . One may think that in this setting one would
avoid the integration of the volatility function in relation (3.9) which is responsible
for the loss of the Markov property of the spot. But, the authors show that the only
suitable volatility functions depend only on time t. As a consequence, the only log-
normal volatility structure for the quoted forward contracts that complies with the
no-arbitrage condition that all of the overlapping contracts must satisfy is the Black
model [37].
In both cases—setting a dynamic on an hourly forward price or on all forwards
with a delivery period—one has very little choice for the volatility functions. The
choice is between the realism of the volatility term structure and the relation with
the spot price. For this reason, to satisfy the absence of the arbitrage opportunity and
still be able to take into account the dependence of the volatility function with the
maturity requires limiting the modelling to only the observed contracts and not being
3.2 HJM-style Forward Curve Models 37

able to deduce an instantaneous forward price and a spot price. This is precisely what
is done in Benth and Koekebakker [25] and Kiesel et al. [118].
In Kiesel et al. [118], the authors develop an HJM two-factor model which is one
often used in practice. The authors consider F(t, T ) the price for a monthly futures
contract quoted at date t with a starting delivery time of T , where T corresponds to
the first day of the month. They consider only monthly contracts and consider that
the other contracts are a portfolio of monthly contracts. The observable forward price
is modelled directly under the risk-neutral measure. The initial condition is given by
the observed forward curve. The authors then use Brigo and Mercurio’s [44] method
to approximate the density of the yearly contract by a log-normal distribution that
has the same first two moments. They perform the calibration of the following two-
factor model:

dF(t, T ) = e−κ(T −t) σ1 F(t, T )dW1t + σ2 F(t, T )dW2t ,

with uncorrelated Brownian motions. The volatility for a given month is then:

σ12 −2κ(T −t)

Var(ln F(t, T )) = (e − e−2κ T ) + σ22 t.
2κ
The parameters are estimated by minimising the quadratic difference between the
market variances and the model-implied variances by using at-the-money options
data. This model has the benefit of simplicity even though it does not provide all of
the features required by a price model for derivative valuation.
Remark 3.4 An alternative to the problem of the limited choice of a volatility func-
tion in the HJM-style model for the electricity forward price was recently formulated
in De Franco et al. [71]. Their approach consists in directly writing the dynamic of
F(t, T1 , T2 ) under the historical probability and making the valuation with a quadratic
hedging criterion. In this case, no stochastic differential equation are required for the
forward prices.
I will conclude this section with a word on the joint modelling of electricity for-
ward prices and fuel prices. This point is not often addressed in the context of forward
curve models for electricity. Indeed, in practice, the most basic approach is used which
consists of adding the same style of factor models for fuels and putting correlations
between the factors. Typically, a joint coal-power model takes the form of:

dfe
(t, T ) = σ e (t, T )dWet , (3.16)
fe
dfc
(t, T ) = σ c (t, T )dWct , (3.17)
fc

where f e (t, T ) (resp. f c (t, T )) stands for the forward price of power (resp. coal).
The dependence between the two prices is embedded in the correlation between
the factors dWet and dWct . Although basic, this method assigns a symmetric role to
38 3 Price Models

both variables whereas it is highly unexpected that a shock to the electricity futures
prices will have an impact on the futures prices of coal. This point has an important
impact on risk management because the utilities are exposed to fuel spreads. The
correlation between the returns tends to provide a pessimistic look at the evolution
of the spreads because it fails to capture the fact that if the price of coal increases, the
price of electricity is likely to increase too. Indeed, the phenomenon to be captured
is that the levels of the prices should move together.
A possible way to deal with this form of dependency is to use co-integrated
variables as initiated in Engel and Granger [83]. Co-integrated variables present
the advantage of connecting their levels rather than their returns. This approach is
implemented for the relation between power and gas in Benmenzer et al. [17] and
between gas and oil in Ohana [137]. Benmenzer et al. [17] develop a continuous-
time co-integrated model consistent with no-arbitrage conditions. Their model writes
under the historical probability:

dfe
(t, T ) = σ e (T − t)dXt , (3.18)
fe
df g
(t, T ) = σ g (T − t)dXt , (3.19)
fg
dXt = ηt dt + Π X t dt + ΣdWt . (3.20)

The volatility functions σ e and σ g are vectors of length Ne and N g . Thus, dWt is a
vector of Ne +N g independent Brownian motions. The function ηt is a parameter used
to centre the dynamic around a desired target. Further, the dependencies between the
price levels are obtained because of the matrix Π . The estimation of the parameters
performs well and involves the PCA and linear regressions. The dynamic of the prices
under the risk-neutral measure is obtained because of a suitable market price of risk
λt given by dBt = dWt + λt dt, and λt is λt = Σ −1 (Π X t + ηt ). The simulation
results they report tend to provide more realistic joint trajectories for gas and power
and show considerable difference in the perception of risk exposure for a utility. This
point is also particularly stressed in Döttling and Heider [79].

3.3 One-Factor Spot Model

The purpose of the one-factor spot models is to design an electricity model which
preserves the Markov property of the spot and which allows for the computation of
the whole forward curve by using an expectation of the spot price under a suitable
measure. Although it might appear restrictive, many alternatives were developed in
the last decade. It goes from Gaussian mean-reversion processes where the change in
the measure is given by a constant market price of risk as in Lucia and Schwartz [127]
to Lévy jump process with a time-dependent market price of risk as in Benth et al.
3.3 One-Factor Spot Model 39

Table 3.1 Summary of one-factor spot models discussed in this section

Model Seasonality Process Jumps Change of
measure
Lucia and Schwartz [127] Deterministic Gaussian MR Constant
Cartea and Figueroa [61] Deterministic Gaussian MR Poisson Constant
Benth et al. [22] Deterministic Gaussian MR Lévy Time-
dependant
Benth and Benth [28] Deterministic Non-Gaussian MR – Constant

[22] and Deng and Jiang [75] (Table 3.1). This section illustrates the advantages and
the limits of this approach with four models with increasing complexity (Table 3.1):

• Lucia and Schwartz’s model [127]: deterministic seasonal part with a Gaussian
mean-reversion process and a change of measure given by a constant market price
of risk
• Cartea and Figueroa’s model [61]: deterministic seasonal part with a Gaussian
mean-reversion term, a Poisson jump term, and a change of measure given by a
constant market price of risk
• Benth et al.’s model [22]: deterministic seasonal part with a Gaussian mean-
reversion term plus a Lévy jump term and a time-dependent market price of risk
• Benth and Benth’s model [28]: deterministic seasonal part with a non-Gaussian
mean-reversion process and a constant-time change of measure

3.3.1 Mean-Reversion Process

Because of the characteristics of the electricity spot and futures prices illustrated in
Chap. 2, a Gaussian mean-reversion process with a seasonal part seems very opti-
mistic for modelling their behaviour. Nevertheless, Lucia and Schwartz’s model [127]
is a benchmark for measuring the effect of standard models when being calibrated to
data on the electricity spot price. The authors try different variants but one is enough
to understand the problems involved in this modelling.
I concentrate on the model using the mean-reversion on the log-price of the spot
Pt . Here, Pt is the daily average spot price, and not the hourly spot price, satisfies

ln Pt = f (t) + Yt (3.21)

with f (t), a deterministic known function of the time that represents the seasonal
part, and
dYt = −κYt dt + σ dWt (3.22)
40 3 Price Models

which is a mean-reverting process with a null long-run mean, constant reverting speed
κ, and an initial condition y0 . This model is identical to Schwartz’s 1997 one-factor
model [152]. The law of Pt is well known and satisfies:

−κt σ2
E0 [Pt ] = exp f (t) + (ln P0 − f (0))e + (1 − e−2κt )
4κ
 2 
σ
Var0 [Pt ] = E0 (Pt )2 exp (1 − e−2κt ) − 1
2κ

Although the underlying cannot be stored, the authors suppose that there is a
unique risk-neutral measure under which it is possible to write that the futures prices
are equal to the expected spot price:

F(t, T ) = Et [ST ] .

They suppose also that this measure is given by a constant market price of risk
λ to carry the futures valuation. And, finally, as in Schwartz [152], under this risk-
neutral measure, the dynamic of Y remains mean-reverting and is given by dYt =
κ (α ∗ − Yt ) dt + σ dW∗t with α ∗ = −λσ/κ. As raised by Carmona and Durrleman
[57, Sect. 8, p. 671], the common practice in the literature is to suppose that the
risk-neutral measure preserves the mean-reversion property of the spot, although the
drift might be expected equal to the risk-free rate.
To make this hypothesis convenient, the price F(0, T ) of the future quoted at
time 0 for delivery at the instant T is given by the expectation of the spot under this
measure and can be computed analytically—which is all in the interest of having a
constant market price of risk:

ln F(t, T ) = f (T ) + e−κ(T −t) (ln Pt − f (t))

  σ2  
+ α ∗ 1 − e−κ(T −t) + 1 − e−2κ(T −t) . (3.23)
4κ
These results allow for an estimation and calibration of the model. The deter-
ministic component of the spot f is considered to be given by a sinusoidal function
f (t) = a + bDt + c cos ((t + τ )2π/365) where Dt equals one if t is a holiday
or a weekend. An estimation of the model is thus performed on the NordPool data
covering January 1993 to December 1999, which represents seven years of daily data
(see Fig. 3.1).
The estimation of the parameters κ, σ , a, b, c, τ are performed in a single non-
linear least-square procedure. The estimation leads to a significant non-zero mean-
reversion coefficient of κ = 0.016 × Δt = 5.84 per year with a time step equal to
one day and a volatility of 164 %. The goodness-of-fit of the model can be measured
by a mean absolute percentage error and is lower than 1 %. But, despite the fact that
the model can be made close to the observations, the general aspect of the simulated
3.3 One-Factor Spot Model 41

450

400

350

300

250

200

150

100

50

0
Oct−1991 Mar−1993 Jul−1994 Dec−1995 Apr−1997 Sep−1998 Jan−2000

Fig. 3.1 NordPool historical daily spot prices. source NordPool

Fig. 3.2 Simulated daily 700

spot price trajectory obtained
with Lucia and Schwartz’s 600
[127] one-factor log spot
model 4 with parameter 500
values of κ = 5.84,
σ = 164 %, a = 4.86, 400
b = −0.09, c = 0.306,
τ = 0.836 300

200

100

0
0 50 100 150 200 250 300 350 400
days

prices is quite different from the observed NordPoool daily spot price as Fig. 3.2
shows.
The calibration of the model is thus performed by using the observed forward
prices quoted during one year (end 1998 to end 1999). Denoting the number of days
before the beginning of delivery by T1 and the number of days before the end of
delivery by T2 , the theoretical forward price F (0, T1 , T2 ) is computed as

1 
F (0, T1 , T2 ) = F(0, T ).
T2 − T1
T1 ≤T ≤T2

By using the model above with an out-of-sample parameter estimation and a null
market price of risk, the authors get a root mean square error (RMSE) over all of the
42 3 Price Models

Fig. 3.3 Dynamic of the

forward curve for Lucia and
Schwartz’s [127] one-factor
log spot model 4 with
parameters identical to
Fig. 3.2

sample contracts of ≈9 %, or a monetary unit of 11 NOK. Thus, there are not enough
parameters in the model to allow for a perfect fit of the forward curve.
Moreover, the dynamic of the forward price given by the relation (3.23) does not
depend on the current time for a sufficiently long enough time to maturity because
of the damping effect of the mean-reversion coefficient. Thus, the estimated value of
the long time to maturity forward is quickly the seasonal component of the forward
price. This point is illustrated on Fig. 3.3.
Thus, as a conclusion, it is possible to have a good fit of the historical daily spot
prices with this model, but the simulated prices do not exhibit any spikes. Moreover,
as expected, a constant market price of risk is not enough to succeed in getting
the required perfect match between the predicted futures prices and the currently
quoted ones. A one factor model is not enough to get the dynamic of the forward
curve because Sect. 3.2 shows that at least two are required to get more than 80 % of
the variance in the storable commodities and even more are required for electricity.
A statistical analysis of this model is also performed in Wilkens and Wimschulte
[168] who report poor performance. However, surprisingly despite these aspects,
this model is often used in some utilities with a seasonal market price of risk as the
only modification.

3.3.2 Mean-Reverting Jump-Diffusion Models

The first idea to improve the behaviour of the spot price in the model (3.21–3.22) is
to add a jump component. This is the approach implemented in Cartea and Figueroa
[61] and in Geman and Roncoroni [94]. I am going to focus on Cartea and Figueora’s
(2005) work because they provide formulas for the futures prices. The spot price St
is now assumed to follow:

ln St = g(t) + Yt
3.3 One-Factor Spot Model 43

with g as the deterministic log-seasonality that is supposed to be differentiable, and

Yt as the stochastic process that satisfies

dYt = −αYt dt + σ (t)dWt + J · dqt

with α as the speed of the mean-reversion, σ (t) as the time-dependent volatility, J as

the proportional random jump-size, and dqt as the Poisson process with intensity l.
The jump-size J is supposed to be log-normally distributed with ln J ∼ N (μ J , σ J2 ).
Given this framework, the spot price dynamic satisfies

dSt = α (ρ(t) − ln St ) St dt + σ (t)St dWt + St (J − 1)dqt

with 
1 1 2
ρ(t) = g (t) + σ (t) + g(t).
α 2

The term St (J − 1)dqt comes from the fact that after a shock, St− moves to J St− ,
which makes ΔSt = (J − 1)St− . Moreover, the authors assume that J satisfies
E [J ] = 1 based on the argument that jumps should not lead to an extra return. This
constraint limits the parameters of the model because μ J = −σ J2 /2.
Figure 3.4 illustrates the behaviour of the daily average spot price for the real-
istic parameters estimated on the UK day-ahead market extracted from Cartea and
Figueora’s (2005) paper. Because the parameter values for the seasonal part of the
spot price is not given in the paper, I use the same seasonal function as in Sect. 3.3.1
for Lucia and Schwartz’s [127] model. I observe that the spiky behaviour is now
clearly obtained. This phenomenon is obtained thanks to the strong per annum mean-
reversion value α = 102. It corresponds to a half-life of two days. Thus, although

Fig. 3.4 Simulated daily 50

spot price trajectory obtained
45
with Cartea and Figueroa
[61] with the parameter 40
values of α = 102,
σ J = 0.67, l = 8.5, 35
σ = 1.64. The seasonal part
30
is identical to Lucia and
Schwartz’s [127] model of 25
Fig. 3.2
20

15

10

5
0 50 100 150 200 250 300 350 400
days
44 3 Price Models

quick, this return to a normal level price is still too slow to represent the change in a
price within an hour.
For the valuation of futures prices, the authors follow Lucia and Schwartz’s [127]
approach by using a constant market price of risk λ. Under this measure, x = ln S
satisfies  
dxt = α μ∗ (t) − xt dt + σt dW∗t + ln J dqt

with μ∗ (t) = α1 g (t) + g(t) − λ σ α(t) .

The first interesting result in this setting is that it allows an analytical computation
of the futures prices. Thus, F(t, T ), the price quoted at time t for delivery at the instant
T , is given by:
e−α(T −t)
St
F(t, T ) = G(T ) D(t, T )J (t, T ),
G(t)

where G(t) = exp (g(t)) represents the seasonal part; D(t, T ) is a term which comes
from the diffusion,
 T  
1 2 −2α(T −s) −α(T −s)
D(t, T ) = exp σ (s)e − λσ (s)e ds ,
t 2

and J (t, T ) is the term which comes from the jumps

 T 
J (t, T ) = exp l (ξ(T, s) − 1)ds
t

with  
σ J2  − alpha(T −s) −2α(T −s)

ξ(T, s) = exp − e −e .
2

At this point, some remarks can be made on that formula. First, because the market
price of risk is constant, there is little chance to perfectly fit the forward prices at
the initial time. This point could be somehow addressed by using a seasonal market
price. Second, the jump intensity l and the jump-size σ J both have a negative impact
on the forward price. It is difficult to assess a change in one parameter only because a
modification of the intensity of jump l might have an impact on the estimation of the
volatility σ . However, a blind comparative static leads to a decrease in the F(t, T )

T
that causes an increase in l because the term t (ξ(T, s)−1)ds is negative. Moreover,
a jump of S at time t has an impact on the whole term structure. But, because of
the strong mean-reversion value of α, the effect is strongly dampened. However, this
advantage is also a drawback because the mean-reversion property that is useful to
quickly erase the effect of a jump on the spot market leads to a flat dynamic for the
long-term maturities of the forward prices. As mentioned by the authors themselves
in a footnote, it is necessary to use far lower mean-reversion values to get a realistic
3.3 One-Factor Spot Model 45

dynamic for the forward with long-term maturity and to avoid a forward dynamic
similar to Fig. 3.3.

3.3.3 Non-Gaussian Mean-Reversion Models

An alternative to the strategy with a jump term added to a diffusion process is to

directly represent the spot price with a non-Gaussian process. This alternative is
quickly recognised as a possible strategy in the seminal paper of Benth et al. [22].
More works have followed this track, such as Deng and Jiang [75]. Benth et al.’s
monograph [19] is the main reference in this matter. Even if in Benth et al. [22],
the authors develop a richer version with more than one Lévy factor, in this section,
I present the Benth and Saltyte-Benth’s [28] model, which is the most basic non-
Gaussian model.
In this approach, the electricity spot price St at time t is modelled as

St = Λt e X t , (3.24)

where Λt represents a deterministic seasonal component while the random part X t

follows the dynamic
X t = a(m − X t )dt + d L t , (3.25)

with a as the mean-reversion parameter, m as the long-run mean of X , and L t as

the Lévy process. This method comprises the increments of the process L t which
are independent and stationary in the sense that the distribution of L t − L s depends
only on t − s. Here, I will spare the reader the rather technical conditions that are
raised when using Lévy processes, and I will consider d L t as a random increment
following a given density. Amongst all of the possible densities, the Normal Inverse
Gaussian (NIG) receives the most attention. The density function of the NIG class is
explicitly known and depends on four parameters: the location of the density μ, the
heaviness of the tail α, the skewness β, and the scale δ. This class of Lévy processes
offers a good representation of the heavy tails of the density of the observed returns.
And because the density is known explicitly, it is possible to perform an efficient
maximum likelihood estimation.
Figure 3.5 illustrates the spot price modelled according to a NIG process. The
parameters as well as the seasonality are chosen according to Benth and Saltyte-
Benth’s [28, Sect. 3.2] for the daily average NordPool spot price.
In this incomplete market setting, there are infinitely many equivalent martingale
Q
measures Q such that F(t, T ) = Et [ST ]. But it is possible here to design a change
of measure that will satisfy the former relation and to allow a perfect fit of the initial
forward curve in certain conditions. This change of measure—based on Esscher’s
transform—is parameterised by a deterministic function of the time θ . In the NIG
46 3 Price Models

Fig. 3.5 Simulated daily 50

spot price trajectory obtained 45
with Benth and
Saltyte-Benth [28] with the 40
parameter values of 35
m = −0.3, a = 0.108,
μ = −0.01284, α = 7.91, 30

β = 0.73, δ = 0.139, and 25

Λ(t) = c1 + c2 
cos 2π3 (t − c3 ) with
20

c1 = 19.14, c2 = 3.23, 15
c3 = 10.47
10

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time

model (3.24–3.25), the forward prices are given by:

   St e−α(T −t)
F θ (t, T ) = Λ(T ) exp μa (1 − e−a(T −t) ) Λ(t)

   
T
× exp δ t α 2 − (θ(s) + β)2 − α 2 − (θ(s) + β + e−a(T −s) )2 ds
(3.26)

Thus, if at time zero, there are n forward prices ( f i )i=1...n for delivery at the non-
overlapping maturities 0 < T1 < · · · < Tn and if one chooses a piece-wise constant
function θ (t) = θi on the interval (Ti−1 , Ti ), then (θi )i=1...n satisfies:

f i = qi × Θi (3.27)

with
μ  S e−αTi
−a(Ti ) 0
qi = Λ(Ti ) exp (1 − e )
a Λ(0)

and
⎛ ⎞
⎜  Ti   ⎟
⎜ −a(T −s) ⎟
Θi = exp ⎜δ α − (θ (s) + β) − α − (θ (s) + β + e
2 2 2 i ) ds ⎟ .
2
⎝ 0 ⎠
  
h(s,Ti ,θ(s))

It is possible to recursively solve the system (3.27) starting with θ1 . Once θ1 is

obtained, θ2 comes from:
3.3 One-Factor Spot Model 47

Fig. 3.6 Dynamic of the

forward curve for Benth and
Saltyte-Benth [28] with 35
parameter values identical to
30
Fig. 3.5
25

20

15

10

5
100
80 60
60 50
40
40 30
20 20
Time (days) 10
0 0 Maturity (weeks)

 T1  T2
f 2 = q2 × exp h(s, T2 , θ1 )ds × exp h(s, T2 , θ2 )ds .
0 T1

And so on for the next values of θi , i ≥ 3. This procedure is not describe in Benth
and Slatyte [28] but is in Benth et al. [22]. It is still in a Lévy model setting but with
instances that make the system (3.27) linear.
However, the use of a non-Gaussian process does not circumvent the problem
of the dampening effect of the mean-reversion on the forward curve dynamic. This
phenomenon is illustrated on Fig. 3.6 where I use the forward prices given in the
NIG case by the relation (3.26) and the parameter values given in [28, Sect. 3.2]. In
particular, the mean-reversion value is rather high and equals 0.108. However, the
behaviour is the same as the mean-reversion jump-diffusion process in Sect. 3.3.2.
The short-term maturities present some volatility while the longer maturities are
reduced to the seasonal component of the spot price.

3.3.4 Conclusion

This section shows that it is possible to obtain a realistic model of the spot price
for electricity by using one-factor models. They succeed in reproducing the main
stylised facts of the spot: seasonality, mean-reversion, and spikes. But, the class of
one-factor spot models studied so far fail to reproduce the proper volatility in the
forward price. A one-factor model with volatility functions depending on the spot
could lead to better results. But, possibly not, because the forward prices are being
impacted by information that has no effect on the spot, like the planned outages of
power plants or movements of the forward curve of fuel prices. This point is not
particular to electricity. The dynamic of the forward curve of other commodities also
requires more than one factor to get a realistic modelling of its behaviour. But, the
48 3 Price Models

dynamic is accentuated in the case of electricity because the amount of energy that
can be transported over the maturities through storage is very limited compared to
consumption. This point is clearly illustrated in the study of the information premium
in Benth et al. [20].

3.4 Multi-factor Spot Models

It seems that if the Markov property of the spot is given up, then the possibilities for
modelling become limitless. Thus, before going through a review of the alternatives
already developed in the literature (Sect. 3.4.2), I will start with an illustrative example
to show how the addition of just one new factor avoids the collateral damage of the
mean-reversion on the forward dynamic.

3.4.1 An Illustrative Example

In this example, I consider that the spot price St at time t is given by

St = Λt e X t +Yt (3.28)

where Λt stands for the deterministic seasonal component of the price, X t follows
the mean-reversion process

dXt = λ X (μ X − X t )dt + σ X dWt ,

Yt is the mean-reverting jump process

dYt = −λY Yt dt + h d Nt

with ln h ∼ N (μ J , σ J ), and Nt is a Poisson process with intensity l. The forward

price F(t, T ) is taken equal to the expectation of the spot price at maturity without
any specification on the probability. The point here is to only illustrate the effect of
this modelling on the dynamic of the forward curve.
In this setting, the fact that there are two different mean-reversion speeds allows
me to separate the need for a quick dampening of the spikes occurring because of Y
from the trend given by X . I set the parameters to standard realistic values. For the
diffusion factor I use λ X = 0.03, μ X = 0, and σ X = 1.64 for the jump factors; and
λ J = 200, σ J = 0.3, and μ J = −σ J2 /2 for the intensity of the jumps l = 10. For the
seasonal component, I use the function Λt from Benth and Saltyte [28] presented in
Sect. 3.3.3.
Figure 3.7 provides an illustration of the behaviour of the electricity spot price
while Fig. 3.8 provides the forward curve dynamic. It appears that the spot price
3.4 Multi-factor Spot Models 49

Fig. 3.7 Simulated daily 90

spot price trajectory obtained
with the model (3.28) with 80
the parameter values of
λ X = 0.036, μ X = 1.025, 70
σ X = 3.8, λ J = 200,
μ J = 1, σ J = 1, l = 10 and 60
Λt = c1 + c2 
cos 2π3 (t − c3 ) with 50
c1 = 19.14, c2 = 3.23,
c3 = 10.47 40

30

20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time

Fig. 3.8 Dynamic of the

forward curve for the
two-factor model (3.28) with 100
the same parameters as in 90
Fig. 3.7 80
70
60
50
40
30
20
10
200
150 60
50
100 40
30
50 20
Time 10
0 0 Time to maturity

exhibits the required spikes while the whole forward curve is no longer fixed to a
constant value. The forward curve now floats according to the random variations of
the diffusion factor.
This toy example does not say anything about the possibilities of estimating the
different component of the price. Here, only St is being observed while the estimation
procedure should reconstruct two different processes. Generally, using some kind of
filtering methods cannot be escaped when performing the parameter estimation of a
multi-factor model.
50 3 Price Models

3.4.2 A Review of the Multi-factor Models

As I said in the introduction of this section, due to the important number of

possibilities, it is not easy to find a logical way of sorting the different models
proposed in the literature. A not perfect but feasible way is to sort them according
to the dynamic of the hidden state, whether it is a discrete Markov chain or a contin-
uous process, and according also to the Gaussian or non-Gaussian character of this
dynamic.
I begin with the continuous hidden state model with Gaussian noise. Kaaresen
and Husby [107] and Barlow et al. [11] develop this modelling framework. I focus on
Barlow et al. [11] where the authors investigate different models in the spirit of those
proposed by Schwartz [152] for storable commodities. They perform the parameter
estimations for three models of the spot:
(1)
Model I: St = e X t dXt = λ X (Yt − X t )dt + σ X d Bt
(2)
dYt = μdt + σY d Bt .
(1)
Model II: St = e X t dXt = λ X (Yt − X t )dt + σ X d Bt
dYt = λY (L − Y y )dt + σY d Bt(2)
(1)
Model III: dSt = λ X (Yt − St )dt + σ S St d Bt
(2)
dYt = μYt dt + σY Yt d Bt
In these three models, B (1) and B (2) are independent Brownian motions.
In their work, Barlow et al. [11] focus on a recurrent difficulty in multi-factor
models: the estimation of the parameters of the dynamic of the hidden state. By
using a Kalman filter algorithm, they perform the parameter estimations of these three
models both with simulated data and with real market data from North America. They
conclude that due to the large variation in the spot price and the short length of the
available time series compared to the length needed for the simulated data to perform
a good estimation of the parameters, it is not possible to properly estimate the mean-
reversion values of the models. Nevertheless, these models succeed in producing
more volatile time series than the one-factor mean-reversion models considered in
Sect. 3.3. However, they fail to represent the isolated spikes.
A natural alternative while maintaining the idea of a continuous hidden state
is to look at multi-factor non-Gaussian models. This is the approach developed in
Benth et al. [23]. They define the spot price St as S(t) = μ(t) + X (t) with μ as
the deterministic seasonal component and X as an arithmetic average of the non-
Gaussian mean-reverting process

X (t) = wi Yi (t) (3.29)
i

where
Yi (t) = −λi Yi (t)dt + σi (t)d L i (t). (3.30)
3.4 Multi-factor Spot Models 51

with L i as the pure jump process. The main interest of the arithmetic model (3.29)
is that it provides analytical formulas for the forward prices with variable delivery
periods. In this model, S(t) tends to go back to the seasonal component μ with a
speed controlled by the parameters wi and λi . The processes L i control the variation
in the spot price (both daily volatility and spikes). The functions σi allow for sea-
sonal trends in the volatility. The intensity of the jump is also made seasonal. The
forward prices can be computed with no more complexity than with the one-factor
non-Gaussian model in Sect. 3.3.3. The simulations of the spot price are given for
some instances with three Ornstein-Uhlenbeck processes that capture well the texture
of the NordPool spot prices. The estimation of the parameters in this model is some-
how not obvious. Because it is neither Markovian nor Gaussian, a classical Kalman
filter cannot be relied on. But, there are possible methods that could overcome this
difficulty.
The final alternative consists in using the hidden Markov chains. In these models,
the spikes occur because a hidden variable switches from one state to another. In the
economic literature on modelling a pure electricity spot price, there are a lot of articles
on regime switching models. The works that compute futures prices and perform the
calibration of the model with the forward curve are rarer. The first example of this
kind of approach is by Kholodnyi [116]. In this work and in the references therein,
the author develops a non-Markovian approach with a two-state regime switching
model (one state for the spiky regime and one for the regular regime). The authors
do a computation of the forward prices but have yet to publish the calibration and
the implementation to the markets.
The only known full implementation of a regime switching model with the compu-
tation of the forward prices and the calibration in a electricity market was conducted
in Monfort and Féron [135]. Their approach is based on a discrete time affine model
and a stochastic discount factor. Their work presents more a class of models than
a single one. The computation of the futures prices requires the use of a Laplace
transform and produces semi-explicit expressions. The estimation of their model
uses a Kalman filter. They perform a numerical illustration on the French market
with a regime switching vector autoregressive model. The dynamic of the spot price
depends on a Markov chain with three states and an unobservable continuous latent
variable, which makes their model three dimensional. To provide an example of their
approach, the form of the spot price St = exp st is:


3
st+1 = μ z t+1 + φ1i (st−i+1 − μ z t−i+1 ) + φ21 yt + σ z t+1 εt+1
s

i=1

yt+1 = ψ1 (st − μ z t ) + ψ2 yt + εt+1 ,

y

where μ = [μ1 , μ2 , μ3 ] and σ = [σ1 , σ2 , σ3 ] while z t takes its value from the
columns of the identity matrix of size three, and yt is a hidden continuous state that
helps the fit with the forward prices. One trick of this modelling approach is in the
fact that the dynamic is written as st − μ z t . Thus, if z t is in a spiky state and has a
52 3 Price Models

low probability of staying in that state, then in the next step, st+1 − μ z t+1 forgets the
spike. With this model, the authors succeed in reconstructing the different forward
prices (week, month, quarter, and year) with good precision.
To conclude this section and to make a transition into the next one, I have to
mention that multi-factor models do not exclude the use of observable variables.
Indeed, the model developed by Burger et al. [48] represents a multi-factor model
mixing both observed variables (the load and its ratio with the maximum available
capacity) and two non-observable factors to take into account the long-term trend of
the prices and the short-term volatility that cannot be explained by the load.

3.4.3 Conclusion

Multi-factor models of the spot price supply a great set of possibilities. Although
there are not many papers which present the whole modelling and calibration process
of the spot and forward prices, there is little doubt that they are able to produce a fit
with the observed forward prices. However, with this class of models, the difficulty
moves from the computation of the forward prices to the estimation and calibration
procedures. As indicated, the statistical methods which are involved to perform the
estimation always require some filtering methods. If the use of a Kalmant filter in
finance is not new, some models presented here require filtering methods that are
beyond the common toolbox of the financial engineers.

3.5 Structural Models

Because more than one factor is needed to properly represent the spot price while
having a realistic dynamic of the forward curve, structural models use the available
information on the electric system to get the extra factors needed. For instance, the
approach based on the multi-factor models in Sect. 3.4 is justified for markets where
little information is available on the supply and demand side or on the causes of
the prices. This is not the case for electricity markets. It is the opposite. In each
country, regulators impose the publication of a large amount of data. A quick look at
the NordPool website is convincing: for example, hourly historical consumption per
geographic area, detailed production and planned outages, exchanges, and the levels
of the reservoirs.
The recent survey by Carmona and Coulon [51] clearly presents the advantages
and the drawbacks of this approach compared to reduced-form models, that is, the
models described in Sects. 3.3 and 3.4. The structural models move the modelling dif-
ficulty of electricity prices away from the statistical methods involved in multi-factor
models to a focus on the dependencies and the price formation mechanism. These
models find a certain echo in the electrical engineering community. For example, the
financial engineers in electric utilities generally find themselves more comfortable
3.5 Structural Models 53

using Gaussian models for the observed data and building a price as a function
of those bricks than implementing a customised filtering method for a multi-factor
model.
In this section, I call structural all of the models that use only some observable
variables as extra factors. The depth of this structure can vary a lot amongst the
models. I sort them from the lighter to the stronger structure. I begin first with the
seminal structural model designed by Barlow [10] (Sect. 3.5.1). Although he does
not do the computation of the forward prices in his paper, the idea of this paper
deserves special attention. Then I turn to the spot price models which use the available
information as auxiliary variables (Sect. 3.5.2). In this class, I examine Cartea and
Villaplana’s [64] model which uses demand and capacity; Pirrong and Jermakayan’s
[144] model which uses demand and gas prices; and the models which use co-
integration techniques such as Frikha and Lemaire [92], Dejong and Schneider [72],
and Benmenzer et al. [17]. Then, I review the structural models which implement a
simplified stack curve mechanism (Sect. 3.5.3).

3.5.1 The Mother of All Structural Models

Barlow’s paper [10] shows that the spiky behaviour of electricity spot prices can
be obtained by using an inverse demand function applied to an Ornstein-Uhlenbeck
process. In his model, the spot price St is determined by the equilibrium between an
increasing supply function u t (x) and a decreasing demand function dt (x), u t (St ) =
dt (St ). Assuming a constant supply function u t = g and an inelastic demand dt (x) =
Dt with Dt for the electricity consumption, the spot price is given by St = g −1 (Dt ).
He uses a basic non-linear form for g(x) = a0 − b0 x α with α < 0. The price is
caped if the demand exceeds the maximum capacity a0 . The spot price is thus given
by the relation:
⎧ 
⎨ a0 −Dt 1/α
Dt ≤ a0 − ε0 b0
St = b0 (3.31)
⎩ε1/α Dt ≥ a0 − ε0 b0
0

where ε0 is determined by the level of the price cap A0 .

Remark 3.5 The inverse demand function can be rewritten g −1 (Dt ) ≈ (a0 −D C
t)
by
considering that α ≈ −1 (see the estimation below). In this regard, it appears that
the price is inversely proportional to a quantity that can be interpreted as a residual
capacity if a0 is the maximum available capacity in the system.
After some algebraic simplifications, the spot price dynamic results in a non-linear
Ornstein-Uhlenbeck process that can be expressed as:

(1 + α X t )1/α 1 + α X t ≤ ε0
St = 1/α (3.32)
ε0 1 + α X t ≥ ε0
54 3 Price Models

Fig. 3.9 Simulated daily 1000

spot price trajectory obtained 900
with Barlow’s model [10]
800
with parameter values of
α = −1.08, λ = 172, 700
a = 0.91, and σ = 0.91 600

500

400

300

200

100

0
0 50 100 150 200 250 300 350 400
days

dXt = −λ (X t − a) dt + σ dWt , (3.33)

1/α
where A0 = ε0 and the remaining parameters λ, α, a, and σ have to be estimated.
Figure 3.9 presents an example of a spot price trajectory with the values of the
parameters obtained by the author on Alberta’s market data. Although the modelling
framework makes use only of a continuous process, the trajectory exhibits a striking
spiky behaviour.
Nevertheless, despite this success, there are still some problems left in this model.
As Carmona and Coulon [51] point out, the estimation parameter leads to a negative
α = −1.08 which makes the supplys curve a steep function near the breaking point
given by a0 − ε0 b0 . Thus, the model tends to produce a lot of prices equal to the price
cap. Indeed, the steepness needed to produce spikes leads to a very short range of
demand where prices can explode. Hence, when a spike occurs, it is more likely to
reach the cap. Some improvements have been made to Barlow’s model. In particular,
Kanamura and Ohashi [110] and Kanamura [109] improve the offer curve to make
it closer to a hockey stick to avoid the concentration of prices at the price cap.
But, this model is only a one-factor model. Thus, even if no attempt is made to
compute and analyse the forward price induced by the relation (3.32), it is predictable
that the forward prices do not avoid the drawback of the one-factor model presented
in Sect. 3.3. And the model exhibits a flat dynamic, as shown in Fig. 3.10.

3.5.2 Auxiliary Variables

The first approach in using auxiliary variables is to rely on the statistical relation
between the electricity prices and the variables that cause them, such as fuel prices.
3.5 Structural Models 55

250

200

150

100

50

0
100
80 100
60 80
40 60
40
20 20
Time 0 0 Time to maturity

Fig. 3.10 Dynamic of the forward curve obtained with Barlow’s (2002) model with parameter
values identical to Fig. 3.9 and prices cut at 250 for aesthetic reasons

In particular, co-integration represents the dependence of the gas prices on the oil
prices in Benmenzer et al. [17]. This idea is developed in the case of the relation
between the electricity prices and the gas prices in continental Europe in Dejong and
Schneider [72] and Frikha and Lemaire [92]. Unfortunately, the computation of the
futures prices and the calibration of the forward curve is left for future work.
Close to this approach but with a more practical design is the model by Grine
and Diko [95]. Their main idea is to use a multi-layered structure to reproduce the
intuitive relationship between fuel dependencies. If there are m fuels, then the fuel
price i is given by
⎛ ⎞

S i (t) = ⎝ γ ji S j (t) + ηi (t)⎠ × ω̂i (t) × ex p(bi (t)X i (t)).
j∈D i

The time varying functions ηi and ω̂i represent the yearly and weekly pattern,
while X is an affine jump-diffusion process (drift and volatility are an affine function
of X ), and γ ji , bi (t) are the parameters. Each X i has to be independent of the price
processes of the underlying energies to avoid cycles (X depends on Y which depends
on Z which depends on X ). Thus the summation is made over a set D i that depends
on fuel i. The interest of this modelling is that the relation between the spot prices is
transferred straightforwardly to the futures:
⎛ ⎞
  
Q
F i (t, T ) = ⎝ γ ji F j (t, T ) + ηi (T )⎠ × ω̂i (T ) × Et exp(bi (T )X i (T )) .
j∈D i
56 3 Price Models

In this framework, the authors succeed in providing an analytical approximation

of the forward prices and the vanilla spread options.
Nevertheless, the approach above introduces known dependencies between the
electricity prices and the fuel prices by using only a statistical relation and not a
microeconomic model. This is the solution developed in Cartea and Villaplana [64]
and Lyle and Elliott [129] with demand and capacity, and Pirrong and Jermakyan
[144] with the load and the gas price.
I illustrate the use of the observable auxiliary variables with Cartea and Villa-
plana’s model [64]. In their model, the spot price Pt is defined by an equilibrium
function φ such that
Pt = φ(Dt , Ct )

where D is the electricity consumption and C is the available capacity. The equilib-
rium function φ increases with D and decreases with C. As in Barlow’s model, the
relation between the observed variables and the price is deterministic. The function
φ and the dynamic on the demand Dt and the production capacity Ct are chosen to
allow for analytical computation of the conditional expectations of the spot price:

Pt = β exp (γ Ct + α Dt ) (3.34)

with positive α and β but negative γ . For the demand and capacity dynamic, the
authors use a deterministic seasonal part plus Ornstein-Uhlenbeck noise. They set:

Dt = gtD + X tD dXtD = −κ D X tD dt + σtD dWtD , (3.35)

Ct = gtC + X tC dXC
t = −κ C X tC dt + σtC dWC
t , (3.36)

where W D and W C are independent. The estimation of the parameters of these

dynamics is done by applying the maximum likelihood principle to the historical
data on demand and capacity.
Because of these choices, the forward prices can be analytically computed. Adding
two time-dependent market prices of risk φ D (t) and φ C (t) for each un-hedgeable
factor D and C, the dynamic of X C and X D reads as
 
dXtD = −κ D X tD + θ D (t) dt + σ D (t)dWtD ,
 
dXCt = −κ X tC + θ C (t) dt + σ C (t)dWC
t ,
C

σ D (t)φ D (t) C (t) = σ (t)φ (t) .

C C
with θ D (t) = and θ
κD κC
3.5 Structural Models 57

And thus, the futures write as

ln F(t, T ) = h(T )
 T
+ γ e−κ (T −t) X tC + κ C e−κ
C C (T −s)
θ C (s)ds
t
 T 
1
e−2κ
C (T −s)
+ σC2 (s)ds
2 t
 T
−κ D (T −t) D
e−κ
D (T −s)
+α e Xt + κ D
θ D (s)ds
t
 T 
1
e−2κ (T −s) σ D2 (s)ds
D
+ (3.37)
2 t

Figures 3.11 and 3.12 provide illustrations of the behaviours of the spot price and
of the forward curve dynamic. I use the parameter values provided by the authors for
the English and Wales market. As expected, the spot price does not exhibit any spiky
behaviour. This is because of the choices of a smooth equilibrium function and the
Gaussian factors. But, the forward curve dynamic presents the floating behaviour of
the long-term maturities.

3.5.3 Stack Curve Models

The idea of a structural model based on a stack curve is to simplify the merit order
models used in generation management enough to ensure computational tractability
of the futures prices. As an example, I follow the model develop by Aïd et al. [2, 3].

Fig. 3.11 Simulated daily 46

spot price obtained with
Cartea and Villaplana [64] in 45
the case of constant
volatilities with parameter 44
values of κ D = 0.33,
σ D = 1580, κC = 0.37, 43
σC = 2056 and gtD = 
2π(t+7)
3.5e4 + 1e3 cos 364
42

gtC = 3.7e4
 + 
2π(t+7) 41
1e3 cos 364
40

39
0 50 100 150 200 250 300 350 400
days
58 3 Price Models

46
45
44
43
42
41
40
39
150

100 100
80
60
50 40
20
Time 0 0
Time to maturity

Fig. 3.12 Dynamic of the forward curve obtained with Cartea and Villaplana’s (2008) model with
the same parameter values as in Fig. 3.11

Fig. 3.13 Reconstructed

marginal fuel price on the
French market according to
Aïd et al. model [2]

The idea behind this structural model is to deduce the no-arbitrage condition for the
electricity prices from the no-arbitrage condition for storable fuel prices. The authors
suppose that the electricity demand Dt can be satisfied with n different technologies
by using all of the fuels. Each technology i has an available capacity Cti and its
fuel cost is given by h i Sti where Sti is the fuel price and h i is the heat rate of the
!i−1 k !i
technology. The technology i capacity interval is Iti = k=1 C t ,
k
k=1 C t . Thus,
the marginal fuel of the system is given by:


n
"t =
P h i Sti 1# Dt ∈I i $ (3.38)
t
i=1

The price of the marginal fuel is an (important) approximation of the electricity

spot price. However, it does not take into account the effect of the bidding strategies
that lead to prices disconnected from the marginal fuel price. The marginal fuel
price given by relation (3.38) is illustrated on Fig. 3.13 with French data and an
3.5 Structural Models 59

approximation using only two fuels: coal and oil prices (estimation details can be
found in [2]).
The spiky behaviour of electricity prices can be obtained using a scarcity function.
The high prices are highly correlated with situations where the reserve !margin is low
(as described in Sect. 2.1.1). I denote the overall capacity with C t := nk=1 Ctk and
the reserve margin with Rt = C t − Dt . Statistical analyses show that high spot prices
correspond to low values of Rt . Thus, the authors set:

"t
Pt = g (Rt ) × P (3.39)

with g the scarcity function given by:

γ 
g (x) := min ν
, M 1{x>0} + M1{x 0} . (3.40)
x
This relation says that the spot price is the marginal fuel price multiplied by
a factor that is inversely proportional to the reserve margin. This is an extension
of Barlow’s model. Figure 3.14 presents the effect of the scarcity function on the
modelling of the French electricity spot prices at the 19th hour (the peak hour). This
figure represents a back-testing of the model (3.39). The historical demand, coal
prices, oil prices, and the available capacity are substituted in relation (3.39) after
performing the estimation of the parameters γ and ν. The fit is not perfect. But, the
behaviour is properly reproduced and the phenomenon in Barlow’s model of too
many spikes at the price cap no longer occurs here. This is due to the fact that the
two fuels are represented by the existence of small jumps when the demand changes
from one fuel to another.
The model given by relation (3.39) is an incomplete market model because the
price depends on un-hedgeable factors such as demand and capacity. The definition of
the futures prices requires the choice of a risk-neutral measure. The authors choose to
use the risk-neutral measure given by the local risk minimisation criterion (see Pham
[141]). An alternative criterion that consists in indifference utility pricing with the
exponential utility is studied in Benedetti and Campi [16]. The benefit of the local

Fig. 3.14 Back-tested spot price on the French electricity market at the 19th hour according to Aïd
et al. (2013) model [2]
60 3 Price Models

risk minimisation is that it allows a decomposition of the derivative prices in the

hedgeable and un-hedgeable parts. Moreover, it allows explicit formulas. Further,
the future price F e (t, T ) at time t for delivery at hour T of 1 MWh of power is
given by


n
F e (t, T ) = h i G iT (t, Ct , Dt )F i (t, T ), (3.41)
i=1

where F i (t, T ) is the quoted futures prices for the fuel i at time t for delivery at time
T , and G i is the weights that depend on demand and capacity, but not on fuel prices.
Observe that Fti (T ) are known at time t and no model is needed for them to compute
the electricity forward price at that time. The relation (3.41) states that the electricity
futures price is a linear combination of the fuel prices. However, this relation tells that
in an economy where electricity is produced only by coal-fired plants, the electricity
futures prices are note a basic reproduction of the coal futures prices. The examination
of the weights G i shows a little more of the relative importance of the fuel prices.
The weights are given by:
 
G iT (t, Ct , Dt ) = EPt g(RT )1# DT ∈I i $ , (3.42)
T

where the conditional expectation is taken under the historical probability. The rela-
tion (3.42) states that the weights are not only given
 by the expected
 probability that
P # $
the fuel i will be the marginal fuel, that is, Et 1 DT ∈I i . But, this marginality
T
should be corrected by the tension in the system when it occurs, that is, g(RT ).
However, the calibration of this model to the observed forward curve is left open
in Aïd et al. [2]. Further, once the risk-neutral measure is chosen by using the local
risk minimisation criterion, there are no more parameters left to fit the theoretical
prices to the observed futures prices. They should perfectly fit. This lack is filled
by Féron and Daboussi [86]. I reproduce a small part of their results with their kind
permission.
They reconstruct the French futures prices by using Ornstein-Uhlenbeck models
with seasonal components for the demand and the capacity. They take into account
three capacities (nuclear, coal, and oil). They estimate the parameters of their four
processes based on historical data. The realised availability of the power plants is
published by the French TSO. The weights G i are then computed for all of the
delivery periods involved in the electricity futures prices. The futures prices for coal
and oil are taken from the historical data. The nuclear fuel cost could be neglected
here because nuclear power is rarely marginal. The theoretical futures prices are then
computed by using the relation (3.41) for the delivery period of the quoted futures
prices. A comparison is then made with the observed futures prices.
Figure 3.15 compares the real market quotations and the reconstructed futures
prices for electricity for the year-ahead base-load contract from 2009 to 2012 with
and without the effect of the scarcity function. During all of the study period, the
3.5 Structural Models 61

variations between the observed prices and the reconstructed prices are correct. But,
it is also clear that there are two periods. During the first three years, an important
error is made by the model (10 Euros on a price of 55 Euros) while the fit is quite
good in the remaining years. The effect of the scarcity function is also important.
With the scarcity function, the error is less than 5 % in the remaining years.
But a decision-maker cannot get satisfied with a 5 % misfit between quotations
and modelled prices. However, the framework of the local risk minimisation can be
saved despite the observed error. The error can be attributed to the models of demand
and capacity. They are over-simplified and do not capture the fine structure of their
real dynamic. But, instead of trying to improve their modelling, an estimation of
the implied demand to be added or subtracted from the input factors is computed
to obtain a perfect fit. The value of this implied demand is estimated to be ≈1 GW,
which is less than 1 % of the installed capacity of the French market. Thus, even if
there is no market price of risk to be used to fit the price model with the observed
quotation, it is still possible to correct the models of the input factors with a small
perturbation to obtain this perfect fit.
But, the structural model given by relations (3.39–3.41) suffers some non-
negligible drawbacks.
First, even if it is not reported in their paper, the computation of a year-ahead
contract with the relation (3.41) at one date still consumes too much time. For a year-
ahead contract, this relation should be used 8,760 times. Despite the development
of the Taylor expansion approximation for the weights G i , the Monte-Carlo method
is still better for computing the values (3.42). But, it will still take one minute to
compute one quotation of a year-ahead contract, which is far from an industrial
standard.
Second, the model (3.38) makes the hypothesis that power plant efficiencies can
be sorted out once and for all by looking at their fuel prices. Some old and inefficient
coal-fired plants might be more costly than recent efficient combined gas cycles.
Further, if the carbon price is high enough, then the coal-fired plant becomes less

Market Price Model with scarcity without scarcity

60

50

40

01/09 10/09 08/10 05/11 03/12 12/12

Fig. 3.15 Reconstruction of the French Year-Ahead Base-Load Contract during the period 2009–
2012 using the model (3.41)
62 3 Price Models

efficient than the gas-fired plants, which leads to a switch between the two fuels. Aïd
et al. [3] take this point into account, but at the expense of having to compute the
probability of the switch at future dates.
To deal with these aspects, Carmona et al. [53] develop a structural model not
only based on the modelling of a stack curve, but on a bid curve model. Their model
allows for switches in the merit order between fuels. The spot price is defined by
 %

Pt = max min bi (Sti ), sup p: b̂i−1 ( p, Sti ) < Dt (3.43)
i
i

where Dt is the demand, Sti is the fuel price i, b̂i = bi−1 (., s)( p) is the bid curve for
fuel i, and bi = bi (0, s). In this setting, with two fuels (gas and coal for instance),
Pt is determined by five cases that can then be reduced to three, in the case of a bid
stack curve model as an exponential bi (x, s) = sexp(ki + m i x).
In this model, because the fuels can be mixed, it might be that more than one
generation capacity of the same fuel is not fully used. The model provides a formula
for the spot price of the two fuels and an extension to capture the spikes as well as the
negative prices. The forward prices are obtained in the absence of the hypothesis on
the demand probability distribution. And, with a Gaussian hypothesis, an analytical
formula is provided. Even if those formulas are quite long, they avoid the time
consuming computation of the relation (3.41).

Table 3.2 A classification of the references in this chapter on the electricity price models for pricing
derivatives
HJM style One-factor spot Multi-factor spot Structural
2000 [34]
2001
2002 [127] [107, 127] [10]
2003 [87] [22]
2004 [28] [11, 48, 115]
2005 [101, 121] [61, 75]
2006 [41, 102] [94]
2007 [26] [23, 168] [110]
2008 [25, 89] [46] [134] [64, 144]
2009 [90, 118] [62] [31] [129]
[72] [3, 67]
2010 [91] [84, 95]
2011
2012 [24, 135]
2013 [92] [2, 51]
3.6 Conclusion 63

3.6 Conclusion

More than 40 papers were published in the last 15 years on modelling electricity
prices just for futures valuation purposes. And each time, most of them propose a
new modelling framework. The Table 3.2 tries to illustrate this richness by putting
the papers in the categories used in this chapter. Even if I see that the HJM framework
has some difficulties in tackling the delivery periods of the electricity futures, some
developments based on the ambit fields by the Norwegian School of Mathematics
might reveal themselves as quite efficient in the future. The research on the one-
factor spot model seems to have found its limits, and the trend clearly is towards
multi-factor models. For their part, the structural models should still be considered
in their prime even if their first occurrence goes back to Barlow [10].
This diversity also reflects the lack of a reference model for market players. If
traders accommodate a simple HJM two-factor model because they can avoid making
transactions on the spot market, most of the actors involved with the generation
management of physical assets find themselves in need of joint hourly spot prices
and futures prices. The lack of a reference model has an impact on the perception of
the values of the options by market agents. It leads to divergence in perceptions, higher
bid-ask spreads, and low liquidity. Thus, this lack is an obstacle to the development
of the market. For this reason, even if it might seem that the fields are quite well
explored now, there is still room for the search of a realistic, consistent, efficient,
robust, and generic model for electricity prices that can be adopted by a large portion
of the electricity market players.
Chapter 4
Derivatives

In this chapter, I present the derivatives that can be considered to be particular to

electricity. In this regard, the spread options described in Sect. 4.1 should be con-
sidered as the basis for more complex derivatives such as tolling contracts or swing
options which were first mentioned in Sect. 2.3. Thus, I limit this section on spread
options to a presentation of their different forms and on the cases where there is a
closed-form formula or an analytical approximation. Because the spread options are
also common in other markets, I also refer the reader to the concise and yet complete
survey on spread options both on their taxonomy and their valuation by Carmona
and Durrleman [57].
Sections 4.2 and 4.3 are devoted to what are maybe the most difficult derivatives
to evaluate and hedge on the electricity markets, namely power plants and tolling
contracts on the one hand, and storage and swings on the other hand. As surprising as
it might seem, although particularly difficult, these options receive important atten-
tion from the academic literature. Furthermore, an important panel of sophisticated
numerical methods has been developed which has not been necessarily common
knowledge amongst financial engineers in energy utilities. The description of the
different numerical methods is beyond the scope of this book. For details on them, I
refer the reader to Carmona et al. [53] book. In this chapter, I limit the presentation
of the reported performance of these methods.
Further, I examine more exotic derivatives. Section 4.4 presents the case of retail
contracts. I describe the different approaches that are proposed to put a price to the
load curve of a customer. This problem is far more basic in its standard formulation
than the valuation problems of the real derivatives above. Nevertheless, it is important
because it is the main source of revenues for electricity utilities. Then, I present in
Sect. 4.5 the weather derivatives. These derivatives are supposed to offer a hedge to
firms whose profits depend on weather conditions. I describe how they work and
propose reasons for why they have not yet met the commercial success that they
should have in the energy business, despite its high dependency on temperature,
sunshine, and rainfall.

© The Author(s) 2015 65

R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7_4
66 4 Derivatives

4.1 Spreads

The most commonly used spread options in the electricity markets are fuel-spread
options and locational spread options. A fuel-spread option consists of an owner with
the right to buy electricity and sell a fuel (either gas, coal, or oil). And the holder of
a locational spread has the right to buy power from one zone and to sell it to another
at the cost of its transportation. The fuel-spread options are embedded in thermal
power plants. A power plant owner who wants to hedge the generation value just has
to sell the fuel spread at different future maturities. As Sect. 2.3 shows, the payoff of
the fuel-spread options that are of some interest to power plants are of the form:
 +
f
STe − h ST − gSTc − K

f
with STe the price of electricity, ST the price of fuel, STc the price of carbon at time
T , and K the start-up cost. The parameters h and g are respectively the heat rate of
the power plant and its emission factor. So, there are the three assets involved in this
payoff. It might be that more assets are involved when the power is quoted in euros
while gas is quoted in pounds and coal in US dollars. In this situation, the fuel-spread
option includes four assets (three prices and one exchange rate).
It is not possible to neglect the strike price induced by the start-up cost. Here is
an example with the order of magnitude of the costs and the prices:
• For a coal-fired plant, the fixed start-up cost is ≈50,000 e. Running 12 h/day with
a capacity of 500 MW leads to a K = 8 e/MWh (50.000/(500 × 12)). The dark
spread at the time this book is written is around 4 e/MWh and it was 30 e/MWh
at its recent peak.
• For a gas-fired plant, the fixed start-up cost is ≈15,000 e. Running 6 h/day with a
capacity of 500 MW leads to a K = 5 e/MWh. The crack spread at the time this
book is written is −15 e/MWh.
The situation is not better for the locational spread options. In Europe for instance,
the transportation cost for generation is around 2 e/MWh, and the spread between
France and Germany is around 4 e/MWh.
Thus, the managers of generation assets are interested in the valuation and hedging
of options whose payoff can be summarised as:
 +
p(T, S0a , S0b , K ) = STa − STb − K (4.1)

with S a and S b the price of the two assets, K the strike, and T the maturity.
Since Margrabe’s paper [130] on the options to exchange one asset against the
other, important studies have been written on spread options. Despite the intense
research activity in this field, there is little hope to expect closed-formed formulas
for spread options in situations with a non-zero strike price. Nor is there much hope
4.1 Spreads 67

for more than two assets and for asset prices following a more general dynamic than
a geometric Brownian motion.
I begin here by reporting some of the most commonly used formulas for spread
options, which are Margrabe’s formulas and their derivatives. These formulas provide
a first approximation of the value of the spread options and can be used as bench-
marks for the more complex models. The formulas reported here are not based on an
asymptotic expansion with respect to some parameter that is supposedly small. They
are empirical formulas. I then turn to the valuation formulas in the models which
are closer to the observed dynamic of the commodities and which are applicable to
electricity. Further, I say a word on the papers which provide the valuations of the
spread options for specific electricity price models.  
Black and Scholes assets model. I assume d S i = S i r dt + σ i dW i , i = a, b
with dW a dW b = ρdt. When K = 0, the value of the spread option whose payoff
is given by (4.1) is:
 
p = e−r T S0a N (d1 ) − S0b N (d2 ) (4.2)

with  σ 2T √
log S0a /S0b +
d1 = √ 2
d2 = d1 − σ T
σ T

and

σ 2 = σa2 − 2ρσa σb + σb2 (4.3)

Margrabe’s formula is an application of the Black and Scholes formula for the asset
S a /S b and the equivalent volatility (4.3). Margrabe’s formula ceases to be true for a
non-zero strike price. Only approximations are available. Kirk’s formula (1995) and
Eydeland and Wolynieck [85] are the most commonly cited. Another approximation
formula can be found in the technical report by Bjerksund and Stensland [36]. To
avoid the multiplication of formulas, I limit the chapter to three of them.
Kirk’s formula (1995) reads as
 
p̂ K = e−r T S0a N (d1K ) − (S0b + K )N (d2K ) (4.4)

with
 σ K2 T
log S0a /(S0b + K ) + √
d1K = √ 2
d2K = d1K − σ K T
σK T

and the equivalent volatility is:

2
S0b S0b
σ K2 = σa2 − 2ρσa σb + σb2 (4.5)
S0b + K S0b + K
68 4 Derivatives

Kirk’s formula consists of using Margrabe’s formula with a strike equal to S0b + K
and of changing the volatility for S b accordingly. Here, the reader should note the
non-symmetric role of the two assets in the equivalent volatility σ K . This is not the
case in the variant proposed by Eydeland and Wolynieck’s formula [85, pp. 345–
346]:
 
p̂ E = e−r T S0a N (d1E ) − (S0b + K )N (d2E ) (4.6)

with
 σ E2 T
log S0a /(S0b + K ) + √
d1E = √ 2
d2E = d1E − σ E T
σE T

and
2
S0a S0a
σ E2 = σa2 − 2ρσa σb + σb2 .
S0b + K S0b + K

In this last formula, it is possible to exchange an asset b at a lower price S0b for a
higher strike K and still have the same value for the spread option.
Another basic and yet efficient way to approximate the value of the spread options
is to approximate the density of the spread at maturity with a Gaussian distribution
and by matching its two first moments. This is the method proposed in Carmona and
Durrlemann [57, Proposition 4.1]. This method leads to the following approximation:

 
p̂ M = m − K e−r T N (d1M ) − σ A N  (d1M ) (4.7)

with
m − K e−r T
d1M = m = (S0a − S0b )ex p(−r T )
σM

and
 
= e−2r T (S0a )2 (eσa T − 1) + (S0b )2 (eσb T − 1) − 2S0a S0b (eρσa σb T − 1) . (4.8)
2 2
σM
2

A systematic analysis of the relative precision of these different methods is too

long for this review. But, it is possible to quickly illustrate them with some numerical
computations and at the same time give the reader some reasoning behind the behav-
iour reflected by the value of the spread options. Specifically, I take a null interest
rate, a maturity of one year, and a reference volatility of 10 % for asset a and 15 %
for asset b. Further, I choose S0b and K such that S0b + K = S0a so that the initial
spread value is always zero when K changes. I consider only the case with a strong
correlation of either 80 or −80 %. The resulting value for the spread option for the
different approximation formulas are plotted on Fig. 4.1.
4.1 Spreads 69

σ = 0.1 σb = 0.15 ρ = −0.8 σa = 0.1 σ = 0.15 ρ = 0.8

a b
9.8 4
9.6
3.8
9.4
3.6
9.2
9 3.4
8.8
3.2
8.6
Kirk 3 Kirk
8.4 Eydeland Eydeland
MM MM
Monte−Carlo Monte−Carlo
8.2 2.8
0 5 10 15 20 0 5 10 15 20
K K

Fig. 4.1 Different approximation formulas for the value of a spread option: Kirk, Eydeland and
Wolyniec, Moment Matching (MM) and a Monte-Carlo estimation (MC). Left with strong negative
correlation. Right with strong positive correlation

Several comments can be made about Fig. 4.1. First, the value of the option spread
can be significantly positive although the initial spread is zero. The value decreases
when the correlation between the assets increases: the stronger the correlation the less
probable the spread will increase. Regarding the approximation formulas, Eydeland
and Wolyniec’s formula provides a constant value in this situation because when
Sb + K is constant, so is the equivalent volatility. The moment matching method
provides a precise and constant estimation of the value of the option, while the error
in Kirk’s formula tends to increase as the strike price increases and the initial price of
the asset b decreases. The reason comes from the fact that Kirk’s equivalent volatility
decreases when Sb is substituted with K , which makes the second asset less likely to
reach the threshold Sa + K . Again, the estimation obtained by the moment matching
method is correct for both negative and positive correlations. But, Kirk’s formula
strongly favours positively correlated assets.
Regarding the Greeks, I illustrate the effect of the volatility of the asset a on
the delta. However, I change the conditions of the computations because when the
strike is zero, the volatility of the assets has the same effect. Thus, as before, I take
a null interest rate and an one-year maturity option but now, S0a = 100, S0b = 80,
and K = 10. At time zero, the spread has a positive value. Figure 4.2 presents
the deltas obtained for the different approximation formulas above. For both types
of correlations, the approximations remain good for the volatilities in a reasonable
range. For high volatility, the moment matching method ceases to provide a reliable
estimate.
In the case where more than two assets are taken into account but where the
dynamic of their prices are still driven by the geometric Brownian motions, several
developments are done. The most basic one is the extension of Kirk’s formula to
three assets by Alos et al. [6]. Considering that the asset prices S a , S b , and S c follow
the geometric Brownian motions with the correlations ρab , ρac , and ρbc , the value
of the call option with payoff (STa − STb − STc − K )+ is approximated for small
maturities T by
70 4 Derivatives

σb = 0.1 ρ = −0.8 σb = 0.1 ρ = 0.8

0.9 Kirk
1 Kirk
Eydeland Eydeland
MM
Monte−Carlo
0.95 MM
Monte−Carlo
0.85
0.9
0.8 0.85

0.75 0.8

0.75
0.7
0.7

0.65 0.65
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
σa σa

Fig. 4.2 Delta obtained for the different approximation formulas for the values of the spread
options: Kirk, Eydeland and Wolyniec, Moment Matching (MM) and a Monte-Carlo estimation
(MC). Left with strong negative correlation. Right with strong positive correlation

 
p̂ L = e−r T S0a N (d1L ) − (S0b + S0c + K )N (d2L ) (4.9)

with
 σ L2 T
log S0a /(S0b + S0c + K ) + √
d1L = √ 2
d2L = d1L − σ L T
σL T

and

σ L2 = σa2 + σb2 πb2 + σc2 πc2 − 2ρσa σb πb − 2ρσa σc πc + 2ρσb σc πb πc ,

where πb = S0b /(S0b + S0c + K ) and πc = S0c /(S0b + S0c + K ).

The numerical illustrations reported by the authors show that relation (4.9) can
provide precise results even for the spread options with a maturity of one year.
The creation of approximation formulas for more general spread options (more
assets and more complex price dynamics) is still an active field of research. An
example is provided by Landon’s PhD thesis [122, Chap. 8] which provides formulas
for general spread options with n assets based on asymptotic expansions.
Beyond Black and Scholes. Several authors propose valuation formulas for the
spread options in mean-reversion models. Some formulas are provided in Tsitakis
et al. [157] for two-asset spread options with an Ornstein-Uhlenbeck dynamic and a
zero strike price. A more general situation is addressed in Hikspoors and Jaimungal
[100] where two-factor mean-reversion models are used for each asset. The closed-
form formulas are provided up to a specific change of measure to preserve the mean-
reversion in the risk-neutral measure.
Another approach is to directly model the spread dynamic. Dempster et al. [74]
develop this approach with co-integration, or in Benth and Benth [29] and Cartea
and González-Pedraz [63] with a mean-reverting jump-diffusion process. With this
simplification, the problem is brought back to the valuation of a call option on an asset
following a mean-reverting process, problems for which the closed-form solutions
4.1 Spreads 71

are available when the dynamic has no jump. The introduction of jumps leads all
of the authors to rely on the fast Fourier transform methods to compute the option
price. For an introduction to Fourier methods for pricing derivatives in the context
of jump processes, see Eberlein [80].
The direct modelling of the spread allows the reduction of the dimension of the
problem to one. But this simplification is made at a cost. One loses the capacity to
differentiate the deltas for each asset that compose the spread. When the spread is
modelled directly, there is only one volatility, and the same absolute value has to be
used for the deltas of each asset. It can be misleading in the case of a spread where
the two assets have significantly different volatilities. For instance, consider the case
of the dark spread with the volatility in the electricity spot price at 50 % and the
volatility in the coal price at 10 %. For a spread option with a one-year maturity, the
delta for electricity is around 70 %, while the delta for the coal is closer to 50 %.
In the case of electricity, the valuation of the spread option is also analysed with
structural models. In particular, Carmona et al. [52] offer a comparison between the
valuation of the spread in this setting and other modelling approaches (the exponential
Ornstein-Uhlenbeck). Their comparison includes Margrabe’s benchmark formula
and a formula derived in a co-integration setting. Amongst their findings, they show
that the highly negative correlation between fuel does not imply a high payoff in
the structural models contrary to the prediction of Margrabe’s formula because a
structural link imposes the dependence of the electricity price on its fuel, which
lowers the density of the spreads.

4.2 Power Plants and Tollings

With the development of spot and forward markets, tolling contracts have emerged as
a financial component of physical power plants. Tolling contracts are a risk transfer
mechanism by which the owner of a power plant can transfer the market price risks
(fuel and power) to an investor who is ready to take these risks against a fixed
premium. They are over-the-counter contracts, and many forms are possible. For a
description of tolling contracts, the reader is referred to Eydeland and Wolyniec[85]
and Baldi [7]. But, their general form is to stay close to the operation of a power
plant. In particular, they consist in the daily exercise of the spread option between
the power price and the fuel price during a period of time of at least a year. In this
regard, tolling contracts consist of an approximation of a power plant as a strip of
spread options (hence, the strong and stressed interest of electricity utilities on an
efficient method to value the spread options).
Unfortunately, this proxy neglects several factors. First, power plants are subject
to a list of operational constraints which prevents them from capturing all of the
possible hourly spreads. For example, they have maximum and minimum generation
capacities, minimum stopping time, minimum running time, ramping capacity, and
maximum numbers of start-ups and shutdowns per year. They are also subject to
outages which are both planned for maintenance purposes and unplanned. Unplanned
72 4 Derivatives

outages have the unpleasant effect of leaving the owner who thought he or she was
long in electricity short of the asset. Moreover, the approximation of the value of a
power plant by a strip of spread options neglects the incompleteness of the market.
The possible exercises involved with a gas-fired plant, for instance, are at least daily
while the forward curve provides hedging instruments for only a grossly aggregated
period of time (week, month, quarter, and year).
Taking into account the different constraints of power plants in its valuation leads
to stochastic control problems. The valuation requires more than the computation of
an expectation. It needs to solve an optimisation problem. Moreover, the dimension
of the problem measured by the state of the system might be large. The conditional
expectations of the daily payoff might be three dimensional (power, fuel, and emis-
sion) if the one-factor models are considered for each asset, or six dimensional in
the case of the more realistic two-factor models. The controlled state might itself be
in a dimension higher than one when dynamical constraints are taken into account.
Moreover, this problem only allows for the determination of the power plant’s oper-
ation. Tolling contracts say when the plant should be started up and shut down. They
do not say anything about the actions the operator has to do on the forward market
to hedge the expected value of the power plant. Taking into account the hedging
problem leads to a more complex optimal control problem where both the hedge
and the operation of the power plant have to be determined simultaneously. In this
context, there is no hope to succeed in getting a closed-form formula, which means
having to use numerical methods.
I address here an overview of the different possible formulations of the problems
and of the numerical methods used to solve them. Most papers in the literature deal
with the operation problem, which is already a non-negligibly difficult problem. This
class contains in chronological order Thompson et al. [156], Hlouskova et al. [103],
Deng and Xia [76], Carmona and Ludkovski [58], Pirrong and Jermakayan [144],
Adkins and Paxon [4], and Ryabchenko and Uryasez [151]. On the other hand, there
are few attempts to solve the operation and hedging problems together. In this class,
I find only Ludkovski [128] and Porchet et al. [145].
Formulation of the operation problem: Formulations for the operation problem
can vary according to the constraints taken and to the price dynamic chosen. I follow
the optimal switching formulation provided by Carmona and Ludkovski [58] as a
reference to which the other papers are compared by highlighting the differences.
Indeed, Carmona and Ludkovski [58] offer an exhaustive presentation of the valuation
problem of a power plant which is very close to the problem confronted by the
decision-makers in electricity utilities, and it provides a numerical method to solve it.
The ingredients of their formulation are the following: The vector X t represents
the market prices of electricity and fuel (gas or coal). It is assumed to be Markov.
The power plant can be in M + 1 different states (including zero and maximum
generation). To each state m, a payoff is associated φ(t, X t , m) at time t when the
prices are X t . For example, when the plant is not running, the payoff is φ(t, X t , 0) =
−1 which represents the fixed operating cost; and when the plant is running in a
first state the payoff is φ(t, X t , 1) = Q 1 × (Pt − St ), with X t = (Pt , St ), and Q 1
is the level of generation of the power plant. The decision-maker can switch from
4.2 Power Plants and Tollings 73

state i to state j by incurring a cost Ci, j . The switching cost can depend on X t ,
without changing either the formulation or the numerical scheme. The control of the
decision-maker is given by a discrete sequence u = (ξk , τk ) of stopping times τk
at which the decision-maker makes the power plant switch to state ξk which takes
values in {0, . . . , M}. The control u t which denotes the operation
 mode at time t
should be seen as a piece-wise constant control process u t = τk <T ξk 1[τk ,τk+1 ) .
I denote by U (t) the set of admissible control strategies from t to T . The objective
of the decision-maker is to maximise the expected value of the power plant over the
time horizon [0, T ]:

⎡ ⎤
 T 
J (t, x, i) = sup E ⎣ φ(s, X s , u s )ds − Cu τ − ,u τk ⎦ . (4.10)
u∈U (t) t τk ≤T
k

Moreover, in this context where only the operations of the power plant are looked
at, the expectation is taken under the risk-neutral probability because it is not the
core difficulty. This formulation is named an optimal switching problem because the
objective is to find the optimal strategy to switch from one state to another according
the state of the system. For an introduction to optimal switching problems, the reader
can refer to Pham [142, Chap. 5].
The mainstream numerical method used to solve problem (4.10) is dynamic pro-
gramming. Here, to be able to solve the problem, the authors introduce a series
of problems where the number of switches in (4.10) are limited to k. Denoting
J k (t, x, i) as the value function when the number of switches is limited to k and the
initial condition at time t is (x, i), the algorithm takes the following form:

 T 
J 0 (t, x, i) = E φ(s, X s , i)ds|X t = x (4.11)
t
 τ 
J k (t, x, i) = sup E φ(s, X s , i)ds + Mk,i (τ, X τx )|X t = x (4.12)
τ t
 
Mk,i (t, X tx ) = max −Ci, j + J k−1 (t, x, j) . (4.13)
j=i

The first relation (4.12) is the necessary terminal condition for the recursive
relation (4.13). This relation gives the value one can expect when there is no possible
switch left. The relation (4.13) indicates whether it is worth paying the cost Ci, j
to leave state i at time t while the prices are x. When these values are computed,
the relation (4.13) provides the answer of whether it is interesting to stay in state i
between t and τ (the expectation is greater than the second term) or to leave state i.
The main numerical difficulty in this algorithm is the computation of the con-
ditional expectations E [h(X )|X = x] that appear in Eq. (4.13). Many numerical
schemes have been proposed since Longstaff and Schwartz’s [125] seminal work
74 4 Derivatives

on American options with regression methods. Quantisation [138] and Malliavin

calculus [88] are amongst the most cited alternatives. The schemes based on the
interacting particle models are also proposed and analysed by Del Moral et al. [73].
For a comparison of the efficiency of these different schemes, the reader might be
interested in the results of Bouchard and Warin [43] which shows the increased
performance of the local regression scheme with the dimension of the state. For an
exhaustive presentation of the main numerical schemes with an application to energy,
I refer the reader to Carmona et al. [55].
Alternative methods to dynamic programming: The problem (4.10) admits
alternative representations which are often cited in the quantitative finance literature
and for which intensive mathematical research is ongoing. As in the case of the
classical valuation of a call option in the Black and Scholes model, it is possible to
associate a partial differential equation to problem (4.10). Nevertheless, it takes the
more complex form of a quasi-variational inequality (QVI).
I denote Mv(t, x, i) = max j {−Ci j + v(t, x, j)} as the intervention opera-
tor and L X as the infinitesimal generator of the process X t . In the case when
d X t = μ(X t )dt +σ (X t )dWt , the operator L X is L X v(t, x) = μ(x)vx + 21 σ 2 (x)vx x .
With some not so restrictive regularity conditions, the value function (4.10) is the
solution of:

max {L X v + φ; −ψ + Mv} = 0. (4.14)

Relation (4.14) shows that the maximum of the two terms should always be zero.
Thus, if the first term is zero, the second is negative, and it states that Mψ < ψ,
which means that it is not worth making an intervention. Thus, in this situation, the
evolution of the system is driven by the partial differential equation L X ψ + φ = 0.
Thus, the optimal solution consists in doing nothing as long as the value of X t belongs
to a certain region and to change the state when X t touches a certain boundary.
This is the approach developed in Thompson et al. [156]. In this study, the electric-
ity spot price is modelled with a mean-reversion with jumps. Thus, the operator L X
is a little more complex and takes the form of an integro-differential equation. The
QVI involves then a partial integro-differential equation. It is solved by an adapted
finite difference scheme.
A second formulation can be associated to the problem (4.10): a backward sto-
chastic differential equation (BSDE). By defining the process Ytk,i = J k (t, X tx , i),
the BSDE satisfies the system:
 T  T
Ytk,i = φ(s, X st,x , i)ds + Ak,i
T − Atk,i − Z sk,i dWs , (4.15)
t t

Mk,i (t, X tx ) ≤ YtM,i , (4.16)

4.2 Power Plants and Tollings 75
 T
(Ytk,i − Mk,i (t, X tx ))d Atk,i = 0, A0k,i = 0. (4.17)
0

In the system (4.15), one has to find YtM,i and also the two auxiliary processes
Z tM,i and AtM,i . I will not enter into the interpretation of these processes here and
refer the reader to Pham [142, Chap. 6] for an introduction to the subject. Although
apparently complex, this approach has been used in Hamadène and Jeanblanc [97]
for the valuation of a power plant in the case of a basic on-off regime and by Porchet
et al. [145] to take into account the market’s incompleteness and its hedging problems.
Formulation of the hedging problem: To illustrate the formulation of the valu-
ation of a power plant in an incomplete market but with the ability to hedge, I follow
Ludkovski’s formulation [128]. As above, a finite number I of generation regimes
is possible. The output price is Yt following

dYt = a(Yt , ξt )dt + b(Yt , ξt ) · (ρdWt1 + 1 − ρ 2 dWt2 ).

It leads to an income flow ψ(t, Yt , i). An operational strategy is a sequence (τk , ξk )

of switching times τk to the states ξk . It is possible to switch state at costs Ci, j (t, Yt ).
The revenue of the power plant is thus
 T 
B(ξ ) = e−r t ψ(t, Yt , ξt )dt − e−r τk Cξk−1 ,ξk .
0 τk

The manager is supposed to be able to hedge in continuous time with a storable

asset with price St . The storable asset can be the futures contract or any other assets
which present the highest correlation with the power plant’s revenues. The storable
asset price follows d Su = μSu du + σ Su dWu1 .
When applying the operation strategy ξ and the hedging strategy π , the wealth
t,x,π,ξ
at time t of the owner of the power plant is X t with x the initial wealth. The
wealth satisfies the dynamic:

d Su  
d X ut,x,π,ξ = πu + r X ut,x,π,ξ − πu du + ψ( ξu , u, Yu )du − Cξk−1 ,ξk 1τk =u .
Su
k

The amount πu of money invested in the hedging provides a risky return d S/S
while the remaining part provides a constant risk-free return. Moreover, the wealth
is accrued by the operational income fromthe power plant ψξu (u, Yu )du and is
decreased by the cost of the changed states k Cξk−1 ,ξk 1τk =u .
Because the market is incomplete, it is necessary to provide a way to specify what
the value of an asset is for the agent. The author’s choice relies on the indifference
price. Supposing that the risk preferences of the owner are given by an exponential
utility function U (w) = − exp (−γ w) , the purpose of the owner is to maximise the
expected utility of his or her terminal wealth over all of the admissible operation and
76 4 Derivatives

hedging strategies:
 
t,x,π,ξ
V (t, y, x, i) = sup Et,x,y,i U (X T ) . (4.18)
π,ξ

However, in this case, the value function is not measured in dollars or euros but
in “utils”, a currency that does not exist on corporate balance sheets. The value of
the power plant is thus defined as an indifference price.
Taking the point of view of the owner of the power plant, the indifference price
corresponds to the cash needed to increase the owner’s initial endowment
 forgiving
up the operation of the power plant. I denote U 0 (t, x) = supπ Et,x U (X Tt,x,π ) as the
utility when the owner is not operating the asset. The indifference price pt,T (x, y, i)
which is measured in currency is defined as

V (t, y, x, i) = U 0 (t, x + pt,T (x, y, i)). (4.19)

The indifference price measures the amount of cash that makes the owner indif-
ferent between having a higher initial wealth but no power plant and having a lower
initial wealth but having the power plant. Because of the use of an exponential utility,
the indifference price also corresponds to the value an owner is ready to receive to
give up the operation of the power plant. Note that in this framework, risk manage-
ment (the hedging strategy) has a clear and measurable effect on the value of the
asset.
The complexity of the problem (4.18) and (4.19) cannot be understated. Two
optimal controls, ξ and π , have to be found in a problem with a four-dimensional
state. However, the choice of the utility function as well as the criterion of terminal
wealth maximisation allows to first compute the operational strategy ξ and then the
hedging strategy π . The choice of an exponential utility function also has an important
property. The indifference price is time additive in the sense that for any intermediate
time τ , it can be expressed as a function of the current accumulated income between t
and τ , and the indifference price running from τ to T . This property can be desirable
to the asset owners.
The hedging strategy is given by:

b(Yt ) ∂ p
πt∗ = −ρ (t, Yt , ξt∗ ). (4.20)
σ ∂y

The term ∂∂ py (t, Yt , ξt∗ ) corresponds to the standard delta of the pricing in a
complete market that is modified by a ratio which takes into account the efficiency
of the hedging instrument. With no correlation, that is, the hedging market has no
relation with the local price, the owner does not use it. Because the indifference price
increases with the local price, the more the hedging market is correlated to the local
price, the more the owner sells the hedging asset.
4.2 Power Plants and Tollings 77

The optimal switching times are given by:

 
∗
τk+1 = inf s > τk∗ : p(s, Ys , ξk∗ ) = max∗ p(s, Ys , j) − Cξ ∗k , j
j=ξk

The owner switches as soon as there is some state j where the indifference price
p(s, Ys , j) minus the switching cost again reaches the indifference price p(s, Ys , ξk∗ )
in the current state. In this situation, the operational strategy depends on the hedging
strategy because the switching times are given by the evolution of the indifference
price which depends on the efficiency of the hedging instrument. In the case of the
power plants, this result might seem surprising because usually the scheduling of
the power plant can be determined independently from the hedging strategy. Indeed,
for electricity that might be the case if the available forward contracts present little
relation with the hourly spot price. Therefore, either the hedging instrument has
some efficiency and the hedging and operation strategies have to be coordinated or
the hedging instrument has little efficiency and it is useless to hedge.
To illustrate his model, the author gives a basic but realistic case study of the
operation of an oil platform. The facility has three operating modes: 0, 5 Mbl/y, 10
Mbl/y. The hedging asset’s price is given by a geometric Brownian motion with a 5 %
growth rate and a 40 % volatility. The local price (Y ) is a geometric Brownian motion
with a 5 % growth rate and a 40 % volatility and a 90 % correlation with the hedging
asset. The cash-flows in the different states are ψ0 (y) = 0, ψ1 (y) = 5(y − 50), and
ψ2 (y) = 10(y − 56); and the switching cost is Ci, j = 0.25|i − j| M$. The interest
rate is 5 %, the time horizon is six months and the time discretisation is one decision
per day.
As for American options, the optimal policy is given by switching thresholds:
• Switch from 0 to 1 if Y ≥ 53$/bbl (greater than the break-even price of 50 in
mode 1)
• Switch from 1 to 2 if Y ≥ 65$/bbl (greater than the equal price of 62)
• Switch back from 2 to 1 if Y ≤ 60$/bbl (lower than the break-even price of 60 in
mode 2)
• Switch back from 1 to 0 if Y ≤ 47$/bbl (lower than the break-even price of 50 in
mode 1)
This hedging strategy operates as if the owner holds a portfolio with two con-
tinuously paying calls: Call-1 with pay-off 5(Y − 50)+ and Call-2 with pay-off
5(2Y − 112)+ . When Y is small, then π ∗ ≈ 0. When Y is large, then π ∗ ≈ π Call2 .
When Y is in between, then π ∗ ≈ π Call1 .
The effect of the operational constraints and the incompleteness of the market
can be precisely evaluated in this framework. The value of a strip of options with
daily exercise is 12.4 M$. The indifference price without risk aversion or correlation
but with switching costs is 11.6 M$, which results in an overestimation of 7 %. On
the contrary, neglecting the operational constraints but taking into account the risk
aversion or the market’s incompleteness leads to a value of 9.72 M$, or a 20 %
78 4 Derivatives

overestimation. Taking into account both effects leads to a value of 8.8 M$, which
results in an overestimation of 40 %.
Numerical efficiency: Because the valuation problem of the power plants relies
heavily on the numerical algorithm, I conclude this section with numerical efficiency
considerations. I summarize here the numerical efficiencies obtained by different
authors together with a short description of the difficulty of their problems and the
numerical method used.
In Porchet et al. [145], the valuation of the power plant takes into account the
switching costs, minimum run time, ramp rates, and the non-constant heat rate.
The market is incomplete, and the utility indifference pricing method is used as in
Ludkovski [128]. Both authors solve the problem with two representations: a partial
differential equation (PDE) and a system of forward backward stochastic differential
equations (FBSDE). Each time they use finite difference schemes and compare them
to a Monte-Carlo regression method. They use a two-factor model for electricity and
gas and an un-hedgeable additive noise to the electricity forward price to represent
the non-convergence of that price to the spot price. The authors report that the finite
difference method for the PDE with a six month time horizon in an incomplete market
with three dimensions takes around one week of computation. For the same valuation
with a four-day horizon, the FBSDE takes around 16 min.
In Carmona and Ludkovski [60], by using dynamic programming with a Monte-
Carlo regression for the valuation of a power plant with several constraints on a
three-month horizon and a six-hour time step takes around five minutes on a desktop
computer using Matlab.
In Ryabchenko and Uryasev [151], the authors report the valuation of a tolling
contract with several constraints close to a power plant with a ten-year horizon and
a one-day time step took around 17 s. Their method is based on the approxima-
tion of the optimal stationary switching boundaries which use heuristics and linear
programming.
These three examples show two things. First, the large difference in computation
time that can be obtained with different methods. Second, the difficulty in comparing
those results because each time everything changed: the definition of the problem,
the market model, and the efficiency of the computer itself. Making numerical exper-
iments more standard would considerably help the assessment of the progress in this
field which is of utmost concern to electricity utilities.

4.3 Storage and Swings

Because electricity cannot be stored, utilities have found ways to overcome the
necessity of having flexibility in the generation management system. Energy is stored
in different forms. The main form of storage is a water reservoir. Another form
consists of a demand-side management contract which ensures that the customers will
give up their consumption in certain situations. A typical example of a demand-side
management contract was developed by EDF in the 1980s. The customers holding
4.3 Storage and Swings 79

this special contract enjoy a lower tariff throughout the year, except during a certain
number N of days (22) of the year (1 November to 31 March) where any kWh
consumed cost the customer a dissuasive price. The important thing here is that
the days with the dissuasive price are not known in advance by the customer or the
utility. The utility warns the client the day before. Thus, the electricity utility actually
holds an option contract: it has the right to select 22 days during the year where the
customer’s price is different. This is the first example of a swing option.
Further, with the development of electricity markets, swing options have emerged
as a financial standardised representation for the physical and complex hydroelectric
storage held by utilities. Swing options are contracts where the buyer has the right
to receive a certain amount of energy q between certain fixed bounds 0 and q on
a certain number of days N during a certain period of time T at a price fixed in
advance. But, the contract has a final constraint which states that the entire amount
of energy should be delivered between a lower bound Q and an upper bound Q.
These contracts reproduce the management of an hydroelectric reservoir while
getting rid of most of the operational constraints involved in the management of a real
hydraulic turbine and also getting rid of the randomness of the inflows. Compared
to the demand-side management contract described above, the swing option has no
uncertainty on the volume consumed at each exercise.
As mentioned in Jaillet et al. [105], swing options offer the right level of protection
against high-peaking days during the year. Imagine that there is a particular month
of the year with high demand where the producer might be short. To cover the risk
of the 22 business days of this month, the producer may want to buy a strip of 22 call
options. But, this protection might be excessive because there is little likelihood that
all of the days of the month will be peaking. And 10 American options covering the
period of the 22 days would still be excessive because they have the same optimal
exercise time. Thus, buying a swing contract with 10 dates out of 22 would be the
right level of protection.
As Sect. 2.3 shows, the valuation of an hydroelectric storage system with a single
independent reservoir can be formulated as:
 T 
V (t, s, a, x) = sup E qu Sut,s du + g(ST , X Tt,x ) , (4.21)
qu ∈[0,q,δt ≥0] t

where Sut,s is the electricity spot price—supposed here to be Markov—which takes

the value s at time t, and X tt,x is the level of the reservoir which takes value x at time t.
This level satisfies the following dynamic:

d X st,x = (At,a
s − qs − δs )ds, (4.22)

where At,as is a random Markov inflow which takes the value a at time t. Moreover,
X is subject to level constraints and should stay within [x, x]. The control variable
δs is the spilling variable that allows to throw away a possible excessive amount of
80 4 Derivatives

water. The function g represents a reward function for keeping water at the final time.
Without this reward function, the reservoir will be empty at time T .
For more standard swing options, there is no inflow and, following Barrera-Esteve
[13], the valuation problem writes as
⎡ ⎤

V (t, s, Q) = sup E ⎣ ψ(ti , q(ti ), S(ti ), Q(ti )) + P(T, ST , Q T )⎦
q≤q≤q 0≤i≤N −1
(4.23)

where St is the market spot price—still supposed here to be Markov—which takes

value s at time t. Moreover, the energy
purchased at time t, q(t), has to be between
q and q. The total purchase Q(t) = i q(ti ) has to be between Q m and Q M . The
function P(T, S, Q) represents a penalty function. It can be a basic linear function
of the over- or under-consumption at maturity or a quadratic penalty to get a smooth
function. The reward function ψ can take different forms. The most basic case is
ψ(t, q, S, Q) = q × (S − K ). For gas market applications, the reward functions are
of the form:

ψ(t, q, S, Q) = −q(S + c I )1q≥0 − q(S − cw )1q≤0

where c I is an injection cost, and cw is a withdrawal cost.

Owners are not only interested in the value itself V (0, S0 , 0), but also in the
optimal control strategy q, that is, how to optimally manage the storage or the swing
contract, and how to hedge the value. In the case of a multi-factor price model,
the number of factors increases the dimension of the state. However, there is no
closed-form solutions for these valuation problems. But, some works have explored
the quantitative analysis of swings options without numerical methods, like Bardou
and Bouthemy [8] who show under which conditions swings are bang-bang. In the
cases of linear diffusion and multiple stopping-time problems, Carmona and Dayanik
[54] propose approximation formulas, whereas Keppo [113] and Rodriguez [149]
propose an analysis that states in which cases swing options can be replicated by a
combination of futures and call options. These results offer a better understanding
of the behaviour of the swing options but do not eliminate the need for numerical
methods in the general case.
Literature review. Even more than the problem of valuation and the hedging
of power plants, the case of swing contracts and storage valuation has drawn the
attention of the academic community. The fact that these problems are shared with
the gas market certainly increases the possibility of applying the proposed methods.
Thirty papers have been published in the last decade on this problem. Clearly, a swing
valuation requires different ingredients:
• the asset (constrained, reward function, penalty function)
• the price model
4.3 Storage and Swings 81

Table 4.1 Classification of the research papers on storage and swing valuations
References Asset Price model Formulation Numerical Performance
method
Jaillet (2004) [105] SW 1-MR DP TR –
Dahlgren (2005) [69] SW BS VI FD
Dahlgren (2005) [68]
Zeghal (2006) [170] MS Lévy DP –
Barrera-Esteve (2006) [13] SW MR DP TR, RM,
NN
Carmona (2008) [60]
Carmona (2008) [54]
Wilhelm (2008) [167] SW BS TR, RM, 1 h (RM),
FE 16 s (TR),
25 s (FE)
Kjaer (2008) [119] SW MRJ PIDE FD –
Boogert (2008a) [39] GS MR PDE FD
Pflug (2009) [140]
Hambly (2009) [98] SW 2-MRJ 10 mn
Haarbrucker (2009) [96]
Bardou (2009) [9] GS 2-MR DP Q 1 mn
Wahab (2010) [163] SW RS DP TR
Bardou (2010) [8]
Becker (2010) [15]
Bronstein (2010) [45] SW DP Q 30 s
Kiesel (2010) [117] SW MR PDE, DP FD, RM
Carmona (2010) [59] GS DP RM 30 mn
Wahab (2010) [163] SW RS DP TR 5 mn
Benth (2011) [27] SW MR PDE FD 15 s
Boogert (2011) [40] GS 3-GBM DP RM –
Rodriguez (2011) [149] SW
Marshall (2011) [131] SW 5-GBM DP TR 20 s
Wahab (2011) [162] SW RS DP TR
Bernhart (2012) [32] SW MRJ BSDE RM
Turboult (2012) [158] SW GBM RM
Warin (2012) [164] GS 2-MR DP RM
Wiebauer (2012) [166] SW GBM RM
Edoli (2013) [81] SW MR DP TR 1s
Eriksson (2014) [123] SW MDL PDE FD
Whenever possible, a report of the numerical efficiency is reported. Legend: Asset multiple stopping-
time (MS), swing contract (SW), gas storage (GS), hydroelectric reservoir (HR); Price model
Black and Scholes (BS), Gaussian mean-reversion (MR), mean-reversion with jumps (MRJ), Lévy
(Lévy), regime-switching (RS); Formulation dynamic programming (DP), variational inequality
(VI), partial differential equation (PDE), partial integro-differential equation (PIDE), backward
stochastic differential equation (BSDE); Numerical method finite difference (FD), finite element
(FE), Monte-Carlo regression (RM), neural network (NN), quantisation (Q), Malliavin calculus
(Mal), trees or lattices (TR)
82 4 Derivatives

• the mathematical representation of the solution

• the numerical methods.
This number of ingredients makes it difficult to find a guideline and to summarise
the different findings. Moreover, many alternatives have been developed for each
mainstream method, which makes exposing the different solutions proposed in the
literature more difficult. Here, I only provide a map of this literature and report any
numerical efficiency results with the caveat that no standardised procedure exists to
allow sound comparisons. The resulting map is given in Table 4.1.
As Table 4.1 shows, the literature mainly concentrates on swing contracts with
Gaussian models. The reason is that this case is close to what trading desks are used to
and because it also concentrates on the efficiency of numerical methods. However, it
is not easy to assess the relative efficiency of those different methods. The column per-
formance presents a sample of the efficiencies reported. It is clear that the variations
are quite large ranging from a few seconds to an hour. Because all of these different
works take different specifications of the asset and of the market price model, I rely
on the papers which perform the comparison of different methods applied to the same
problems. The results reported in Wilhelm and Winter [167] are of much interest. For
a swing with five exercises, the authors find a computational time of more than one
hour for a Monte-Carlo regression based on Carmona and Touzi [54] to less than ten
seconds for a finite difference scheme and less than a second for a lattice method. In
the same vein, the comparison performed in Bardou et al. [9] between a quantisation
method and a Monte-Carlo regression exhibits a large efficiency difference in favour
of the quantisation method. However, the test is performed on a one-factor model for
which the quantisation is an advantage that tends to disappear with more dimensions,
as shown in Bouchard and Warin [43]. Thus, despite being an important research
effort, it is still difficult to decide on which numerical methods a pricing and risk
management system for electricity derivatives should be built on.
Moreover, regarding the computation of the hedge on swing options, the literature
is less prolific. For example, Warin [164] is the only paper dealing explicitly on
the problem of the efficient computation of the deltas for the simulation of hedging
strategies. The author proposes a method that reduces by a factor 80 the computational
time of the deltas compared to the most efficient methods.

4.4 Retail Contracts

The complex electricity derivatives described in the preceding sections and exchanged
on the wholesale market constitute only a very small part of the source of the income
of utilities. Most of their revenue comes from the selling of electricity to final cus-
tomers with very different terms than the complex options presented earlier. Further,
retail customers are mainly looking for a fixed price contract for their variable con-
sumption. The problem for the utility is to find a fair price for taking the risks involved
4.4 Retail Contracts 83

with these contracts. The most important uncertainties are the consumption of the
customer and the realised spot prices.
The problem faced by the utilities varies with the segment of the markets, whether
they are industrial, large, professional, small, or households. The load curve of indus-
trial and large customers are generally precisely metered. Industrial customers ask
for a quotation from electricity providers through a call for tender in which they
provide their past load curve. The quotation can be just for the next year or for a
longer term. These contracts are tailor-made and can contain embedded options to
take into account the industrial process of the customers. The offers also come with
an expiry date. Further, during the customer’s decision time, the prices can vary
which puts the providers at greater risk. Either prices drop and the customer asks for
a new quotation, or they go up and the customer exercises the free options embed-
ded in the offer. Thus, there is always a premium added to the price of the contract
to reflect this option value. The competitive pressure is strong in this segment of
customers because it is not very costly—compared to the households—to gain cus-
tomers because their decision is mostly driven by the price. For small firms, things
are a little different. It is no longer possible to design an offer and a price for each
individual firm. Moreover, only some of them have a metered load curve (think of
the office of your dentist or florist). This is even more true for households which only
have two measures per year in countries without smart metering. In this situation,
the prices should apply to a large set of customers for which only aggregated infor-
mation is known. Moreover, the competition for households is not as strong as for
the industrial segment: it is much more difficult and costly to increase one’s market
share of this segment, because it implies spending an important amount of cash on
advertising and maintaining important salesmen.
Although important for utilities and electricity providers, the problem of pric-
ing retail contracts has received little attention from the literature. Indeed, most of
papers addressing the problem concentrate on the industrial customers. This is the
case in Keppo and Räsänen [114], Karandikar et al. [111], Prokopczuk et al. [148],
Karandikar et al. [112], and Burger and Müller [49].
A brief summary of these approaches relies on the allocation of required capital
to cover the price risk and the load risk involved with the customers’ electricity
consumption. If lt is the load of the customer at hour t of the year of delivery, St is
the spot price for the same hour, then the expected cash-flow of the provider from
T
this customer is Π ( p) = E [R( p)] with R( p) = 0 ( p − St )lt dt where p is the
selling price. The criterion chosen in this literature is to find a price that satisfies
a risk adjusted return on capital (RAROC). The RAROC is the ratio between the
expected return of an investment and the capital used. It is fixed ex ante by the utility
as a hurdle rate μ to be achieved by its transactions. In this situation, the return is
identified as the expected cash flow from the contract. To take into account risk,
the definition of the capital used requires a risk measure. In this context, the most
common risk measure proposed is the cash-flow-at-risk which is a quantile of the
distribution of the cash-flow. Thus, noting qα (X ) as the α-quantile of the random
variable X , the problem consists in finding p such that
84 4 Derivatives

Π ( p)
μ= . (4.24)
Π ( p) − qα (R( p))

Variations of this idea are developed in the literature above to take into account
the correlation between the customer and the system loads, or to take into account
that part of the customer’s future consumption that can be hedged. The literature also
develops the modelling of the spot price in the context of spikes.
The fact that this formulation is far less attractive than the nice optimal control
problems arising from the valuation of power plants and swing options has certainly
something to do with the relative lack of interest in this topic. Nevertheless, as
Burger and Müller [49] point out, the price proposed to the customer includes a
margin decided by the management of the utility. Thus, although useful for risk
management purposes, the former approach does not exhaust the problem posed by
the fact that retail prices are the result of a competitive process.
Further, if I consider the UK market at the time this book is written, there is
a significant number of large electricity providers (Centrica, EDF Energy, E.ON,
Npower, SSE, and Scottish Power) involved in a fierce competition process. They
might change the prices in their contracts whenever they want. But, the change in
prices affects their customers who can switch to another provider, leading to a loss
of market share. Knowing this, retailers do not change their prices unless they have
a good reason. And that reason comes from the increase in their sourcing costs on
the wholesale market. Thus, when the wholesale forward prices increase, their retail
margin decreases. At some point in time, some retailers will not take the losses
anymore and will increase their prices. I illustrate the kind of dynamic that can be
observed on the UK retail market on Fig. 4.3 with two retailers. In this picture, one is
the leader who moves first while the other waits for the leader’s move before acting.
It is only a simulation but it reproduces the shape and structure of the observed data.
This situation finds its place in game theory. This game is a mixture between a
revenue management problem and an attrition game. The revenue management is a

Fig. 4.3 Illustration of the 65

dynamic of retail prices.
Blue dotted line wholesale
year-ahead electricity price, 60
Red crossed line retail price
of the leader, Black dotted
line retail price performed by 55
the follower

50

45

40
Jan−2010 Jan−2011 Jan−2012
4.4 Retail Contracts 85

marketing science developed to increase revenues in the airline or the hotel industry
to maximise their profit with a smart pricing policy depending on their booking rate
and on the time left before reservation. It leads to dynamic control problems (see
Bitran and Caldenty’s [35] survey on this topic). But the situation faced by utilities is
in a sense more basic: the product they sell is hardly different from their competitors
and the competition is basically done on the price. Financial and market share losses
are the only motivation for a move on prices. Those situations relate to the attrition
game, which goes back to Maynar [133]. In this game, the players are facing a choice
between staying in the game and enduring a cost or stopping the game and incurring
an income. In the case of an electricity or a commodity market, the study of these
situations with a tractable dynamic stochastic model is yet to be done.

4.5 Weather Derivatives

With weather derivatives, I come to a class of financial products for which the
underlying cannot be held nor even produced. Since 1997, financial products have
been based on the temperature, precipitation, and now the wind. Because a significant
portion of the economy is sensitive to climate conditions (1/7 of the US economy
according to Cao and Wei [50]), it is natural to think of insurance contracts to immu-
nise industries from bad weather conditions. The cash flows of electric utilities are
particularly sensitive to temperature. In southern countries, the hot weather during
summer leads to an increase in air conditioning whereas in countries with electrical
heating, cold waves lead to increases in electricity consumption. And, now, with the
increase in wind generation, electricity utilities are also sensitive to the wind.
But, it is not obvious how to design a standardised financial product whose under-
lying is a non-storable, non-producible asset like a weather condition. And it is even
more challenging to price and hedge such a contract. The typical way a contract
on temperature is structured has four ingredients: a temperature index, a delivery
period, a location, and a tick size. Regarding the index, the most commonly used is
heating degree days (HDD or in short η) and cooling degree days (CDD) on day t.
The calculation for HDD is:
 +
T M + Ttm
ηt = θ − t ,
2

with TtM the maximum temperature of the day, Ttm the minimum temperature of
the day, and θ a temperature threshold often equal to 18 ◦ C. To write a contract, a
definition is needed of the location where the temperature station is located (Chicago
Airport, Paris Orly, Essen…) and a period of time. Generally, the accumulation of
degree days are considered over a month or a season (winter for HDD and summer for
CDD). Further, one has to convert the index quoted in a physical measure (degree,
mm of rain…) into a currency. This is done with the tick size. Each degree day
86 4 Derivatives

500

450

400

350

300

250

200

150

100

50

0
May−2001 Oct−2002 Feb−2004 Jul−2005 Nov−2006 Apr−2008

Fig. 4.4 Heating degree days for the Paris Orly airport from 2002 to 2006 as used by the Chicago
Mercantile Exchange (CME) weather index

corresponds to a certain amount of cash. Thus, if the tick size is 100 e, the buyer of a
forward contract on temperatures
 in December 2014 with HDD as an index, agrees to
pay at maturity 100× 1≤t≤31 ηt . Figure 4.4 shows that for a winter month, a typical
value of the HDD in Paris is approximately 400. Thus, each contract leads to the
payment of 400 × 100 = 40,000 e. The options can be defined as well. For example,
a European call option on the temperature with a strike K in degrees Celsius for a
delivery month has the payoff of
⎛ ⎞+

A⎝ ηt − K ⎠ ,
1≤t≤31

where A is the value of the tick.

The problem with these derivatives is to find a principle to price them. Indeed,
they belong to a frontier between insurance products and financial market products.
But as Bouchard and Elie [42] point out, those derivatives cannot be priced by the
actuarial pricing method where a price is determined by the diversification principal
amongst the customers nor by the principle of temporal diversification (hedging).
For instance, regarding the temperature outcome in winter for electricity utilities in
Europe, the outcome affects all of them in the same way, which seriously limits the
diversification effect for the seller of the option on temperature.
But, the literature does not seem to be afraid by these difficulties and offers several
approaches to tackle this problem. Basically, people either refer to different ways
to maximise utility or to estimate a market price of risk for temperatures by using
data on option prices. Regarding utility maximisation, Cao and Wei [50] propose
to deduce the price of a weather derivative by using Lucas’s representative agent of
inter-temporal utility maximisation. Davis [70] argues that risk preferences towards
4.5 Weather Derivatives 87

weather derivatives are genuinely specific to economic agents, and thus their prices
could not be extracted by using the utility of a representative agent. He proposes as
an alternative to define the price as the marginal value of the substitution for a utility
maximising agent which is between seeing the agent’s initial wealth decrease by the
price of the derivatives and taking the claim. This point of view only deals with one
agent and defines a fair price as relating to that agent’s perception of risk. A simi-
lar approach is undertaken by Carmona and Diko [56] for precipitation derivatives
where the authors use the more classical approach with the valuation method of the
indifference price. However, for a transaction to occur, another agent is needed for
which this price is also admissible. Barrieu and El Karoui [14] performs the study of
this equilibrium by analysing the conditions under which two agents, the writer of the
option and the buyer, will transact on an un-hedgeable claim. This approach helps in
understanding the conditions under which weather derivatives might be exchanged.
But, the utilities in the market might be better off to turn towards the more usual
approach contained in Alaton et al. [5] and Benth et al. [30] for temperature deriva-
tives on the Chicago Mercantile Exchange or Benth et al. [30] for wind futures [18].
This approach is nothing unusual from what was presented in Chap. 3 in fitting the
forward curve by using a market price of risk estimated with option data. Thus, I
will not enter into the detailed implementation here because the problem boils down
to the modelling of the underlying risk factor (either the temperature per station, the
wind, or the precipitation).
I will conclude this section on weather derivatives with comments on the develop-
ment of this market. Despite the efforts of the many actors, in particular the Weather
Risk Management Association, it is difficult to claim that the weather derivatives
market has known the growth expected since its beginning in 1997. In the United
States, weather derivatives are still quoted on the Chicago Mercantile Exchange, but
in Europe the successive attempts to develop such a market has not been successful.
Indeed, the study of the weather derivatives market performed by Huault and Rainelli
[104] clearly shows that the weather derivatives market which represents only a thin
portion of the derivative markets has the capacity to attract attention, but somehow
fails to meet the expectations of the economic agents. The idea of protecting indus-
tries’ incomes from changing weather conditions is not new. But the development
of standardised weather products that will have the required impact on the balance
sheets of a given industry is yet another problem. Most of the deals involving weather
indices are performed on an over-the-counter basis with customised contracts. The
reason is that a temperature measured at some point in the country might not be
representative of the effect on a specific business. For example, the accumulated
temperature in Paris Orly during winter might not tell much about the presence of
snow in the Alps, thus making it difficult to construct an exchange between an actor
positively impacted by a low temperature and one negatively impacted.
Because there is a need in many industries to find protection against bad weather
outcomes for low prices, the market for weather derivatives should see its develop-
ment continue. Nevertheless, at this stage, it seems that research is needed less on
the pricing side than on the design of the contracts. The development of the mar-
ket should not only rely on the exchange between industries looking for protection
88 4 Derivatives

against an undesirable weather outcome and speculators ready to bet on the tem-
perature. An equilibrium between actors having opposite sensitivities to weather
conditions should exist even without speculators. This last class of economic agents
only provides an excess of liquidity.
Chapter 5
Conclusion

After more than 20 years, research has proposed many alternative models or
evaluation methods to address the problems in electricity derivatives. Indeed, from
Gaussian mean-reversion processes to the cutting edge ambit fields, it seems that no
modeling framework has escaped implementation in the electricity markets. But, I
want to use this opportunity to propose some guidelines for future research. I cur-
rently see four major research streams.

1. Despite the large number of existing models, I think that there is still room for
creation and innovation. However, research should not focus on power only, but
should tackle the joint modelling of electricity prices and commodity prices,
including carbon prices. Further, the structural models presented in Chap. 3 are
just some methods to capture this dependence. Other techniques, maybe more
efficient, could be applied with greater success. Moreover, the massive introduc-
tion of renewable energy and its impact has led to the need for price models
which can suitably take into account this increasing phenomenon which leads to
negative prices. Further, little has been done to jointly model prices in intercon-
nected areas. And, new markets have appeared that should attract the attention of
academics as well, such as the intraday market. Those who look at the dynamic
of the 32 h traded forward might find it challenging to model this market. The
intraday market also offers an opportunity to test economic theory on the rela-
tion between spot and forward prices because intensive data are available for this
market and exogenous random events are precisely known. Other markets such
as capacity markets are in their infancy in Europe and should also have their share
of challenges.
2. The ordering and classifying of the zoology of models would be of great help to the
industry. To use the criteria cited in the introduction of Chap. 3, it would be useful
to assess them according to their realism, consistency, efficiency, robustness,
and generality. Amongst these criteria, efficiency and robustness are of major
importance to the development of an operational system of risk management
and valuation. They should be given an unambiguous meaning to allow sound
comparisons. This research would be useful in identifying a model that can be

© The Author(s) 2015 89

R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7_5
90 5 Conclusion

used as a reference or a benchmark in order to form a consensus on the pricing of

electricity derivatives, such as power plants, tolling contracts, and swings. In the
same vein, standardization efforts should be made so that the various numerical
methods developed to evaluate these complex options can be compared.
3. In the coming years, due to the development in households of smart metering,
the questions regarding the pricing of household consumptions will be of major
importance. The effect of competition in the retail market could lead to interesting
applications of option games. Moreover, the value of the flexibilities in household
appliances should also deserves some attention.
4. More research could be devoted to the optimal hedging of physical assets by using
real futures contracts available on the market. Indeed, the rare works presented in
Sect. 4.2 that tackled the joint problem of physical operation of a power plant and
its hedging relied on a simplified version of the electricity forward curve. More
realistic optimal hedging models of physical assets might help to assess the real
efficiency of the hedging strategies.
5. In connection with the preceding point, I wonder if the present structure of the
electricity derivative market is the best suited to achieve its purpose, namely
helping operators to hedge the value of their assets. Indeed, the present structure
with its decomposition into year, quarter, month, week, day and hours results from
a sound decomposition of information and is a good way to avoid a dispersion of
liquidity. It seems quite natural now but, maybe, there are alternative structures
that might allow better hedges. Here, it is not a question of market design, but a
question of the optimal microstructure design: what series of futures and options
contracts might be optimal by offering the best trade-off between liquidity and
hedging efficiency?
6. The problem of valuation and hedging have been addressed at the level of an
individual asset but not often at the level of the electric utility itself. Indeed,
the question of the efficiency of or even the interest in risk management for
a nonfinancial institution is an important field of research in the quantitative
corporate finance literature. In the case of the energy markets, the problem could
be addressed from the perspective of the new European financial regulation that
constrains the use of derivatives by energy firms (REMIT, Regulation on Energy
Market Integrity and Transparency and EMIR, European Market Infrastructure
Regulation).

There is more to be done in the future than just improving the models developed
in the preceding decade. More than 30 years after its birth, the field of electricity
derivatives still needs innovation.
References

1. R. Adkins, D. Paxson, Reciprocal energy-switching options. J. Energy Markets 4(1), 91–120

(2011)
2. R. Aïd, L. Campi, N. Langrené, A structural risk neutral model for pricing and hedging power
derivatives. Math. Financ. 23(3), 387–438 (2013)
3. R. Aïd, L. Campi, A. Nguyen Huu, N. Touzi, A structural risk neutral model of electricity
price. Int. J. Theor. Appl. Financ. 12(7), 925–947 (2009)
4. R. Aïd, G. Chemla, A. Porchet, N. Touzi, Hedging and vertical integration in electricity
markets. Manag. Sci. 57(8), 1438–1452 (2011)
5. P. Alaton, B. Djehiche, D. Stillberger, On modeling and pricing of weather derivatives. Appl.
Math. Financ. 9(1), 1–20 (2002)
6. E. Alos, A. Eydeland, P. Laurence, A Kirk’s and a Bachelier’s formula for three asset spread
options. Energy Risk 9, 52–57 (2011)
7. F. Baldi, Switch, switch, switch! A regime-switching option-based model for valuing a tolling
agreement. Eng. Econ. 55(3), 268–304 (2010)
8. O. Bardou, S. Bouthemy, When are swing options bang-bang? Int. J. Theor. Appl. Financ.
13(6), 867–899 (2010)
9. O. Bardou, S. Bouthemy, G. Pages, Optimal quantization for the pricing of swing options.
Appl. Math. Financ. 16(2), 183–217 (2009)
10. M. Barlow, A diffusion model for electricity prices. Math. Financ. 12(4), 287–298 (2002)
11. M. Barlow, Y. Gusef, M. Lai, Calibration of multifactor models in electricity markets. Int. J.
Theor. Appl. Financ. 7(2), 101–120 (2004)
12. O.E. Barndorff-Nielsen, F.E. Benth, A.E.D. Veraart, Modelling electricity futures by ambit
fields. Adv. Appl. Probab. 46(3), 719–745 (2014)
13. C. Barrera-Esteve, F. Bergeret, C. Dossal, E. Gobet, A. Meziou, R. Munos, D. Reboul-Salze,
Numerical methods for the pricing of swing options: a stochastic control approach. Methodol.
Comput. Appl. Probab. 8(4), 517–540 (2006)
14. P. Barrieu, N. El Karoui, Optimal design of derivatives in illiquid markets. Quant. Financ. 2,
181–188 (2002)
15. M. Becker, Unbiased Monte-Carlo valuation of lookback, swing and barrier options with
continuous monitoring under variance gamma models. J. Comput. Financ. 13(4), 35–61 (2010)
16. G. Benedetti, L. Campi, Utility indifference valuation for non-smooth payoffs with an appli-
cation to power derivatives (2014 submitted)
17. G. Benmenzer, E. Gobet, C. Jérusalem, Arbitrage free cointegrated models in gas and oil
futures markets (2007), http://hal.archives-ouvertes.fr/hal-00200422/fr/

© The Author(s) 2015 91

R. Aïd, Electricity Derivatives, SpringerBriefs in Quantitative Finance,
DOI 10.1007/978-3-319-08395-7
92 References

18. F.E. Benth, J.Ł. Benth, Dynamic pricing of wind futures. Energy Econ. 31(1), 16–24 (2009)
19. F.E. Benth, J.S. Benth, S. Koekebakker, Stochastic Modeling of Electricity and Related Mar-
kets (World Scientific Publishing Company, Singapore, 2008)
20. F.E. Benth, R. Biegler-König, R. Kiesel, An empirical study of the information premium on
electricity markets. Energy Econ. 36, 55–77 (2013)
21. F.E. Benth, A. Cartea, R. Kiesel, Pricing forward contracts in power markets by the certainty
equivalence principle: explaining the sign of the market risk premium. J. Bank. Finance
32(10), 2006–2021 (2008)
22. F.E. Benth, L. Ekeland, R. Hauge, B.F. Nielsen, A note on arbitrage-free pricing of forward
contracts in energy markets. Appl. Math. Financ. 10(4), 325–336 (2003)
23. F.E. Benth, J. Kallsen, T. Meyer-Brandis, A non-Gaussian Ornstein-Uhlenbeck process for
electricity spot price modeling and derivatives pricing. Appl. Math. Financ. 14(2), 153–169
(2007)
24. F.E. Benth, R. Kiesel, A. Nazarova, A critical empirical study of three electricity spot price
models. Energy Econ. 34(5), 1589–1616 (2012)
25. F.E. Benth, S. Koekebakker, Stochastic modeling of financial electricity contracts. Energy
Econ. 30, 1116–1157 (2008)
26. F.E. Benth, S. Koekebakker, F. Ollmar, Extracting and applying smooth forward curves from
average-based commodity contracts with seasonal variation. J. Deriv. 15(1), 52–66 (2007)
27. F.E. Benth, J. Lempa, T.K. Nilssen, On the optimal exercise of swing options in electricity
markets. J. Energy Markets 4(4), 3–28 (2011)
28. F.E. Benth, J. Saltyte-Benth, The normal inverse Gaussian distribution and spot price mod-
elling in energy markets. Int. J. Theor. Appl. Financ. 7(2), 177–192 (2004)
29. F.E. Benth, J. Saltyte-Benth, Analytical approximation for the price dynamics of spark spread
options. Stud. Nonlinear Dyn. Econom. 10(3), 1–28 (2006)
30. F.E. Benth, J. Saltyte-Benth, T. Koekebakker, Putting a price on temperature. Scand. J. Stat.
34(4), 746–767 (2007)
31. F.E. Benth, L. Vos, A multivariate non-Gaussian stochastic volatility model with leverage for
energy markets. Technical report 20, Department of Mathematic CMA (2009)
32. M. Bernhart, H. Pham, P. Tankov, X. Warin, Swing Options Valuation: A BSDE with Con-
strained Jumps Approach, ed. by R. Carmona, P. Del Moral, P. Hu, N. Oudjane. Springer
Proceedings in Mathematics, vol. 12 (Springer, Berlin, 2012), pp. 379–400
33. H. Bessembinder, M.L. Lemmon, Equilibrium pricing and optimal hedging in electricity
forward markets. J. Financ. 23, 1347–1382 (2002)
34. K. Bhanot, Behavior for power prices: implications for the valuation and hedging of financial
contracts. J. Risk 2(3), 43–62 (2000)
35. G. Bitran, R. Caldentey, An overview of pricing models for revenue management. Manuf.
Serv. Oper. Manage. 3(5), 203–229 (2003)
36. P. Bjerksund, G. Stensland, Closed form spread option valuation. Working paper, Department
of Finance NHH, Oct 2006
37. F. Black, The pricing of commodity contracts. J. Financ. Econ. 3(12), 167–179 (1976)
38. M. Boiteux, Peak-load pricing. J. Bus. 33(2), 157–179 (1960)
39. A. Boogert, C. De Jong, Gas storage valuation using a Monte-Carlo method. J. Deriv. 15(3),
81–98 (2008)
40. A. Boogert, C. de Jong, Gas storage valuation using a multifactor price process. J. Energy
Markets 4(4), 29–52 (2011)
41. S. Borovkova, H. Geman, Seasonal and stochastic effects in commodity forward curves. Rev.
Deriv. Res. 9(2), 167–186 (2006)
42. B. Bouchard, R. Elie, L. Moreau, A note on utility based pricing and asymptotic risk diversi-
fication. Math. Financ. Econ. 6(1), 59–74 (2011)
43. B. Bouchard, X. Warin, Monte-Carlo Valuation of American Options: Facts and New Algo-
rithms to Improve Existing Methods, ed. by R. Carmona, P. Del Moral, P. Hu, N. Oudjane.
Springer Proceedings in Mathematics, vol. 12 (Springer, Berlin, 2012), pp. 215–255
44. D. Brigo, F. Mercurio, Interest Rate Models Theory and Practice (Springer, New York, 2001)
References 93

45. A.L. Bronstein, G. Pages, B. Wilbertz, How to speed up the quantization tree algorithm with
an application to swing options? Quant. Financ. 10(9), 995–1007 (2010)
46. E. Broszkiewicz-Suwaj, A. Jurlewicz, Pricing on electricity market based on coupled-
continuous-time-random-walk concept. Phys. A Stat. Mech. Appl. 387(22), 5503–5510
(2008)
47. M. Burger, B. Graeber, G. Schindlmayr, Managing Energy Risk: An Integrated View on Power
and Other Energy Markets (Wiley, Hoboken, 2008)
48. M. Burger, B. Klar, A. Müller, G. Schlindmayr, A spot market model for pricing derivatives
in electricity markets. Quant. Financ. 4, 109–122 (2004)
49. M. Burger, J. Müller, Risk-adequate pricing of retail power contracts. J. Energy Markets 4(4),
53–75 (2011)
50. M. Cao, J. Wei, Weather derivatives valuation and market price of weather risk. J. Futures
Markets 11(24), 1065–1089 (2004)
51. R. Carmona, M. Coulon, A Survey of Commodity Markets and Structural Models for Electric-
ity Prices, ed. by F. E. Benth, V. Kholodny, P. Laurence. Volume Quantitative Energy Finance:
Modeling, Pricing and Hedging in Energy and Commodity Markets (Springer, Berlin, 2013),
pp. 41–83
52. R. Carmona, M. Coulon, D. Schwarz, The valuation of clean spread options: linking electricity,
emissions and fuels. Working paper available on ssrn (2012)
53. R. Carmona, M. Coulon, D. Schwarz, Electricity price modeling and asset valuation: a multi-
fuel structural approach. Math. Financ. Econ. 7(2), 167–202 (2013)
54. R. Carmona, S. Dayanik, Optimal multiple stopping of linear diffusion. Math. Oper. Res.
33(2), 446–460 (2008)
55. R. Carmona, P. Del Moral, N. Oudjane, P. Hu, Numerical Methods in Finance (Springer, New
York, 2012)
56. R. Carmona, P. Diko, Pricing precipitation based derivatives. Int. J. Theor. Appl. Financ. 8(7),
959–988 (2005)
57. R. Carmona, V. Durrleman, Pricing and hedging spread options. SIAM Rev. 45, 627–685
(2003)
58. R. Carmona, M. Ludkovski, Pricing asset scheduling flexibility using optimal switching.
Math. Oper. Res. 15(5–6), 405–447 (2008)
59. R. Carmona, M. Ludkovski, Valuation of energy storage: an optimal switching approach.
Quant. Financ. 10(4), 359–374 (2010)
60. R. Carmona, N. Touzi, Optimal multiple stopping and valuation of swing options. Math.
Financ. 18(2), 239–268 (2008)
61. A. Cartea, M.G. Figueroa, Pricing in electricity markets: a mean reverting jump diffusion
model with seasonality. Appl. Math. Financ. 12(4), 313–335 (2005)
62. A. Cartea, M.G. Figueroa, H. Geman, Modelling electricity prices with forward looking
capacity constraints. Appl. Math. Financ. 16(2), 103–122 (2009)
63. A. Cartea, C. González-Pedraz, How much should we pay for interconnecting electricity
markets? A real options approach. Energy Econ. 34(1), 14–30 (2012)
64. A. Cartea, P. Villaplana, Spot price modeling and the valuation of electricity forward contracts:
the role of demand and capacity. J. Bank. Finance 32(12), 2501–2519 (2008)
65. L. Clewlow, C. Strickland, Energy Derivatives (Lacima Group, London, 2000)
66. G. Cortazar, E. Schwartz, The valuation of commodity contingent claims. J. Deriv. 1, 27–39
(1994)
67. M. Coulon, S. Howison, Stochastic behaviour of the electricity bid stack: from fundamental
drivers to power prices. J. Energy Markets 2(1), 29–69 (2009)
68. M. Dahlgren, A continuous time model to price commodity-based swing options. Rev. Deriv.
Res. 8(1), 27–47 (2005)
69. M. Dahlgren, R. Korn, The swing option on the stock market. Int. J. Theor. Appl. Financ.
8(1), 123–139 (2005)
70. M.H.A. Davis, Pricing weather derivatives by marginal value. Quant. Financ. 1, 305–308
(2001)
94 References

71. C. De Franco, P. Tankov, X. Warin, Numerical methods for the quadratic hedging problem in
Markov models with jumps. J. Comput. Financ. (2015 to appear)
72. C. De Jong, S. Schneider, Cointegration between gas and power spot prices. J. Energy Markets
2(3), 27–46 (2009)
73. P. Del Moral, P. Hu, O. Oudjane, B. Rémillard, On the robustness of the Snell envelope. SIAM
J. Financ. Math. 2, 587–626 (2011)
74. M.A.H. Dempster, E. Medova, K. Tang, Long term spread option valuation and hedging. J.
Bank. Finance 32(12), 2530–2540 (2008)
75. S.-J. Deng, W. Jiang, Levy process-driven mean-reverting electricity price model: the marginal
distribution analysis. Decis. Support Syst. 40(3–4), 483–494 (2005)
76. S.-J. Deng, Z. Xia, A real options approach for pricing electricity tolling agreements. Int. J.
Inf. Technol. Decis. Making 5(3), 421–436 (2006)
77. P. Diko, S. Lawford, V. Limpens, Risk premia in electricity forward prices. Stud. Nonlinear
Dyn. Econ. 10(3), 1–25 (2006) (paper 7)
78. R. Djabali, J. Hoeksema, Y. Langer, COSMOS Description CWE Market Coupling Algorithm
Document (APX Endex, AelPex and EpexSpot, Jan 2011)
79. R. Döttling, P. Heider, Spread volatility of co-integrated commodity pairs. J. Energy markets
55(2), 251–276 (2013)
80. E. Eberlein, Fourier-based Valuation Methods in Mathematical Finance, ed. by F.E. Benth,
V. Kholodnyi, P. Laurence. Quantitative Energy Finance: Modeling, Pricing and Hedging in
Energy and Commodity Markets (Springer, Berlin, 2013), pp. 85–114
81. E. Edoli, S. Fiorenzani, S. Ravelli, T. Vargiolu, Modeling and valuing make-up clauses in gas
swing contracts. Energy Econ. 35, 58–73 (2013)
82. A. Ehrenmann, K. Neuhoff, A comparison of electricity market designs in networks. Oper.
Res. 57(2), 274–286 (2009)
83. R.F. Engle, C.W.J. Granger, Co-integration and error correction: representation, estimation,
and testing. Econometrica 55(2), 251–276 (1987)
84. C. Erlwein, F.E. Benth, R. Mamon, HMM filtering and parameter estimation of an electricity
spot price model. Energy Econ. 32(5), 1034–1043 (2010)
85. A. Eydeland, K. Wolyniec, Energy and Power Risk Management: New Developments in
Modeling, Pricing and Hedging (Wiley, Hoboken, 2002)
86. O. Féron, E. Daboussi, Calibration of Electricity Price Models. Commodities, Energy and
Environmental Finance, ed. by M. Ludkovski, R. Sircar, R. Aïd (Springer, Berlin, 2013)
87. S.E. Fleten, J. Lemming, Constructing forward price curves in electricity markets. Energy
Econ. 25, 409–424 (2003)
88. E. Fournié, J.M. Lasry, J. Lebuchoux, P.-L. Lions, N. Touzi, Some applications of Malliavin
calculus to Monte-Carlo methods in finance. Financ. Stoch. 3, 391–412 (1999)
89. D. Frestad, Common and unique factors influencing daily swap returns in the Nordic electricity
market, 1997–2005. Energy Econ. 30(3), 1081–1097 (2008)
90. D. Frestad, Correlations among forward returns in the Nordic electricity market. Int. J. Theor.
Appl. Financ. 12(5), 589–603 (2009)
91. D. Frestad, F.E. Benth, S. Koekebakker, Modeling term structure dynamics in the Nordic
electricity swap market. Energy J. 31(2), 53–86 (2010)
92. N. Frikha, V. Lemaire, Joint modelling of gas and electricity spot prices. Appl. Math. Financ.
20(1), 69–93 (2013)
93. H. Geman, Commodities and Commodity Derivatives: Modelling and Pricing for Agricultur-
als, Metals and Energy (Wiley, Chichester, 2007)
94. H. Geman, A. Roncoroni, Understanding the fine structure of electricity prices. J. Bus. 79(3),
1225–1262 (2006)
95. S. Grine, P. Diko, Multi-layer model of correlated energy prices. J. Comput. Appl. Math.
233(10), 2590–2610 (2010)
96. G. Haarbrücker, D. Kuhn, Valuation of electricity swing options by multistage stochastic
programming. Automatica 45(4), 889–899 (2009)
References 95

97. S. Hamadéne, M. Jeanblanc, On the starting and stopping problem: application in reversible
investments. Math. Oper. Res. 32(1), 182–192 (2007)
98. B. Hambly, S. Howison, T. Kluge, Modelling spikes and pricing swing options in electricity
markets. Quant. Financ. 9(8), 937–949 (2009)
99. D. Heath, R. Jarrow, A. Morton, Bond pricing and the term structure of interest rates: a new
methodology for contingent claims valuation. Econometrica 60(1), 77–105 (1992)
100. S. Hikspoors, S. Jaimungal, Energy spot price models and spread options pricing. Int. J. Theor.
Appl. Financ. 10(7), 1111–1135 (2007)
101. J. Hinz, L. von Grafenstein, M. Verschuere, M. Wilhelm, Pricing electricity risk by interest
rate methods. Quant. Financ. 5(1), 49–60 (2005)
102. J. Hinz, M. Wilhelm, Pricing flow commodity derivatives using fixed income market tech-
niques. Int. J. Theor. Appl. Financ. 9(8), 1299–1321 (2006)
103. J. Hlouskova, S. Kossmeier, M. Obersteiner, A. Schnabl, Real options and the value of gener-
ation capacity in the German electricity market. Rev. Financ. Econ. 14(3–4), 297–310 (2005)
104. I. Huault, H. Rainelli-Weiss, A market for weather risk? conflicting metrics, attempts at
compromise and limits to commensuration. Organ. Stud. 10(32), 1395–1419 (2011)
105. P. Jaillet, E.I. Ronn, S. Tompaidis, Valuation of commodity-based swing options. Manag. Sci.
50(7), 909–921 (2004)
106. F. Jamshidian, Commodity options valuation in the Gaussian futures term structure model.
Rev. Futures Markets 10(2), 324–346 (1993)
107. K. Kaaresen, E. Husby, A joint state-space model for spot and futures power. Energy Power
Risk Manag. 7(8), 14 (2002)
108. N. Kaldor, A note on the theory of the forward market. Rev. Econ. Stud. 7(3), 196–201 (1940)
109. T. Kanamura, A supply and demand based volatility model for energy prices. Energy Econ.
31(5), 736–747 (2009)
110. T. Kanamura, K. Ohashi, A structural model for electricity prices with spikes: measurement
of spike risk and optimal policies for hydropower plant operation. Energy Econ. 29(5), 1010–
1032 (2007)
111. R.G. Karandikar, S.A. Khaparde, S.V. Kulkarni, Quantifying price risk of electricity retailer
based on CAPM and RAROC methodology. Int. J. Electr. Power Energy Syst. 29(10), 803–809
(2007)
112. R.G. Karandikar, S.A. Khaparde, S.V. Kulkarni, Strategic evaluation of bilateral contract for
electricity retailer in restructured power market. Int. J. Electr. Power Energy Syst. 32(5),
457–463 (2010)
113. J. Keppo, Pricing of electricity swing options. J. Deriv. 11(3), 26–43 (2004)
114. J. Keppo, M. Räsänen, Pricing of electricity tariffs in competitive markets. Energy Econ.
21(3), 213–223 (1999)
115. V. Kholodnyi, Valuation and hedging of European contingent claims on power with spikes: a
non-Markovian approach. J. Eng. Math. 49(3), 233–252 (2004)
116. V. Kholodnyi, Modelling Power Forward Prices for Positive and Negative Power Spot Prices
with Upward and Downward Spikes in the Framework of the Non-Markovian Approach, ed.
by F.E. Benth, V. Kholodny, P. Laurence, Quantitative Energy, Finance: Modeling, Pricing
and Hedging, in Energy and Commodity Markets, (Springer, Berlin, 2013), pp. 189–211
117. R. Kiesel, J. Gernhard, S.-O. Stoll, Valuation of commodity-based swing options. J. Energy
Markets 3(3), 91–112 (2010)
118. R. Kiesel, G. Schindlmayr, R.H. Borger, A two-factor model for the electricity forward market.
Quant. Financ. 9(3), 279–287 (2009)
119. M. Kjaer, Pricing of swing options in a mean reverting model with jumps. Appl. Math. Financ.
15(5–6), 479–502 (2008)
120. P.J. Knez, R. Litterman, J. Scheinkman, Common factors affecting bond returns. J. Financ.
49(5), 1861–1882 (1994)
121. S. Koekebakker, F. Ollmar, Forward curve dynamics in the Nordic electricity market. Manag.
Financ. 31(6), 73–94 (2005)
96 References

122. N. Landon, Almost sure optimal stopping times: theory and applications, Ph.D. thesis, Ecole
Polytechnique, 2013 (Available on Hal)
123. J. Lempa, M. Eriksson, T.K. Nilssen, Swing options in commodity markets: a multidimen-
sional Lévy diffusion model. Math. Methods Oper. Res. 79(1), 31–67 (2014)
124. R. Litterman, J. Scheinkman, Common factors affecting bond returns. J. Fixed Income 1(1),
54–61 (1991)
125. F.A. Longstaff, E.S. Schwartz, Valuing American options by simulation: a simple least-squares
approach. Rev. Financ. 14, 113–147 (2001)
126. F.A. Longstaff, A.W. Wang, Electricity forward prices: a high-frequency empirical analysis.
J. Financ. 59(4), 1877–1900 (2004)
127. J. Lucia, E.S. Schwartz, Electricity prices and power derivatives: evidence from the Nordic
power exchange. Rev. Deriv. Res. 5(1), 5–50 (2002)
128. M. Ludkovski, Financial hedging of operational flexibility. Int. J. Theor. Appl. Financ. 11(8),
799–839 (2008)
129. M.R. Lyle, R.J. Helliott, A “simple” hybrid model for power derivatives. Energy Econ. 31,
757–767 (2009)
130. W. Margrabe, The value of an option to exchange one asset for another. J. Financ. 33(1),
177–186 (1978)
131. T.J. Marshall, R.M. Reesor, Forest of stochastic meshes: a new method for valuing high-
dimensional swing options. Oper. Res. Lett. 39(1), 17–21 (2011)
132. P. Massé, Les réserves et la régulation de l’avenir dans la vie économique (Reservoirs and
regulation for future economic activity) (Herman, Paris, 1947)
133. S. Maynard, The theory of games and the evolution of animal conflicts. J. Theor. Biol. 47,
209–221 (1974)
134. T. Meyer-Brandis, P. Tankov, Multifactor jump-diffusion models of electricity prices. Int. J.
Theor. Appl. Financ. 11(5), 503–528 (2008)
135. A. Monfort, O. Féron, Joint econometric modeling of spot electricity prices, forwards and
options. Rev. Deriv. Res. 15(3), 217–256 (2012)
136. M. Musiela, M. Rutkowski, Martingale Methods in Financial Modelling (Springer, Berlin,
2005)
137. S. Ohana, Modeling global and local dependence in a pair of commodity forward curves with
an application to the US natural gas and heating oil markets. Energy Econ. 32(2), 373–388
(2010)
138. G. Pagés, J. Printemps, Functional quantization for numerics with an application to option
pricing. Monte Carlo Methods Appl. 11(4), 407–446 (2005)
139. M.V.F. Pereira, Optimal stochastic operations scheduling of large hydroelectric systems. Int.
J. Electr. Power Energy Syst. 11, 161–169 (1989)
140. G.C. Pflug, N. Broussev, Electricity swing options: behavioral models and pricing. Eur. J.
Oper. Res. 197(3–16), 1041–1050 (2009)
141. H. Pham, On quadratic hedging in continuous time. Math. Methods Oper. Res. 51, 315–319
(2000)
142. H. Pham, Continuous-time Stochastic Control and Optimization with Financial Applications
(Springer, Berlin, 2011)
143. D. Pilipovic, Energy Risk: Valuing and Managing Energy Derivatives, 2nd edn. (McGraw
Hill, New York, 2007)
144. G. Pirrong, M. Jermakyan, The price of power–the valuation of power and weather derivatives.
J. Bank. Finance 32(12), 2520–2529 (2008)
145. A. Porchet, N. Touzi, X. Warin, Valuation of power plants by utility indifference and numerical
computation. Math. Methods Oper. Res. 70(1), 47–75 (2009)
146. A. Prekopa, J. Mayer, B. Strazicky, I. Deak, J. Hoffer, A. Nemeth, B. Potecz, Scheduling of
Power Generation: A Large-Scale Mixed-Variable Model (Springer, Berlin, 2014)
147. H. Prevot, B. de Juvigny, F. Lehmann, M. Louvot, C. Izart, Rapport d’enquête sur les prix
de l’électricité (report on the prices of electricity) (Document, French Ministry of Economy,
Finance and Industry, 2004)
References 97

148. M. Prokopczuk, S.T. Rachev, G. Schindlmayr, S. Trück, Quantifying risk in the electricity
business: a raroc-based approach. Energy Econ. 29(5), 1033–1049 (2007)
149. J.F. Rodriguez, Hedging swing options. Int. J. Theor. Appl. Financ. 14(2), 295–312 (2011)
150. RTE, Power System Reliability Memento RTE, 2004
151. V. Ryabchenko, S. Uryasev, Pricing energy derivatives by linear programming: tolling agree-
ment contracts. J. Comput. Financ. 14(3), 73–126 (2011)
152. E. Schwartz, The stochastic behavior of commodity prices: implications for valuation and
hedging. J. Financ. 52(3), 923–973 (1997)
153. F.P. Sioshansi, W. Pfaffenberger, Electricity Market Reform: An International Perspective
(Elsevier, Amsterdam, 2006)
154. S. Stoft, Power System Economics: Designing Markets for Electricity (Wiley-IEEE Press,
New York, 2002)
155. G. Swindle, Valuation and Risk Management in Energy Markets (Cambridge University Press,
Cambridge, 2014)
156. M. Thompson, M. Davison, H. Rasmussen, Valuation and optimal operation of electric power
plants in competitive markets. Oper. Res. 52(4), 546–562 (2004)
157. D. Tsitakis, S. Xanthopoulos, A.N. Yannacopoulos, A closed-form solution for the price of
cross-commodity electricity derivatives. Phys. A Stat. Mech. Appl. 371(2), 543–551 (2006)
158. F. Turboult, Y. Youlal, Swing Option Pricing by Optimal Exercise Boundary Estimation, ed.
by R. Carmona, P. Del Moral, P. Hu, N. Oudjane. Springer Proceedings in Mathematics, vol.
12 (Springer, Berlin, 2012), pp. 400–419
159. O.A. Vasicek, An equilibrium characterization of the term structure. J. Financ. Econ. 5, 177–
188 (1977)
160. J. Viehmann, Risk premiums in the German day-ahead electricity market. Energy Policy
39(1), 386–394 (2011)
161. N.-H.M. von der Fehr, P.V. Hansen, Electricity retailing in Norway. Energy J. 13(1), 25–45
(2010)
162. M. Wahab, C.-G. Lee, Pricing swing options with regime switching. Ann. Oper. Res. 185(1),
139–160 (2011)
163. M.I.M. Wahab, Z. Yin, N.C.P. Edirisinghe, Pricing swing options in the electricity markets
under regime-switching uncertainty. Quant. Financ. 10(9), 975–994 (2010)
164. X. Warin, Gas Storage Hedging, ed. by R. Carmona, P. Del Moral, P. Hu, N. Oudjane. Springer
Proceedings in Mathematics, vol. 12 (Springer, Berlin, 2012), pp. 421–445
165. R. Weron, Modeling and Forecasting Electricity Loads and Prices–A Statistical Approach
(Wiley, Chichester, 2006)
166. K. Wiebauer, A Practical View on Valuation of Multi-Exercise American Style Options in
Gas and Electricity Markets, ed. by R. Carmona, P. Del Moral, P. Hu, N. Oudjane, Springer
Proceedings, in Mathematics, vol. 12, (Springer, Berlin, 2012), pp. 353–378
167. M. Wilhelm, C. Winter, Finite element valuation of swing options. J. Comput. Financ. 11(3),
107–132 (2008)
168. S. Wilkens, J. Wimschulte, The pricing of electricity futures: evidence from the European
energy exchange. J. Futures Markets 27(4), 387–410 (2007)
169. A.J. Wood, B.F. Wollenberg, G.B. Sheblé, Power Generation, Operation and Control, 3rd
edn. (Wiley, New York, 2013)
170. A.B. Zeghal, M. Mnif, Optimal multiple stopping and valuation of swing options in Levy
models. Int. J. Theor. Appl. Financ. 9(8), 1267–1297 (2006)