Multi-Factor Energy Price Models
and Exotic Derivatives Pricing
by
Samuel Hikspoors
A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Graduate Department of Statistics University of Toronto
c Copyright by Samuel Hikspoors 2008 Multi-Factor Energy Price Models and Exotic Derivatives Pricing
Samuel Hikspoors
Doctor of Philosophy, Department of Statistics University of Toronto, May 2008
Abstract
The high pace at which many of the world’s energy markets have gradually been opened to competition have generated a significant amount of new financial activity. Both academi- cians and practitioners alike recently started to develop the tools of energy derivatives pricing/hedging as a quantitative topic of its own. The energy contract structures as well as their underlying asset properties set the energy risk management industry apart from its more standard equity and fixed income counterparts. This thesis naturaly contributes to these broad market developments in participating to the advances of the mathematical tools aiming at a better theory of energy contingent claim pricing/hedging. We propose many realistic two-factor and three-factor models for spot and forward price processes that generalize some well known and standard modeling assumptions. We develop the associated pricing methodologies and propose stable calibration algorithms that motivate the application of the relevant modeling schemes.
ii A Monique, Jean, Anna et Jean-Paul
Pour tout l’amour et le support...
iii Acknowledgements
This work would hardly have been possible without the support and collaboration of my supervisor and friend, Sebastian Jaimungal. These last few years have been a great learning experience; thanks Seb!
I would also like to thank Kenneth R. Jackson, Luis Seco and Sheldon Lin for having been part of my supervisory committee, and Matt Davison (UWO) for having been my external examinator; Thank you all for your guidance.
Finally, a few friends who very much influenced me over the years: Gu, JP, Tone, Louis, PO, Kam and Lanna; wish you all the best!
iv Contents
1 Introduction 1
2 Background Material 5
2.1 Derivatives Pricing/Hedging ...... 6
2.2 Asymptotic Tools and Terminology ...... 7
3 First Paper: Energy Spot Price Models and Spread Options Pricing 12
3.1 Abstract ...... 12
3.2 Introduction ...... 13
3.3 Real World Dynamics and Pricing ...... 16
3.3.1 Model Specifications ...... 16
3.3.2 Spot Spread Valuation : An Actuarial Approach ...... 19
3.4 Risk-Neutral Dynamics and Pricing ...... 24
3.4.1 Measure Change ...... 25
3.4.2 Forward Prices ...... 27
3.4.3 Spot Spread Valuation ...... 28
3.4.4 Forward Spread Valuation ...... 29
3.5 Spot Prices with Jumps ...... 31
3.5.1 Model Specification ...... 31
3.5.2 Forward Prices ...... 33
v 3.5.3 Spark Spread Valuation ...... 34
3.6 Model Calibration ...... 39
3.6.1 Methodology ...... 40
3.6.2 Some Results: Crude Oil ...... 42
3.7 Conclusions ...... 46
3.8 Acknowledgements ...... 47
4 Second Paper: Asymptotic Pricing of Commodity Derivatives using Stochastic Volatility Spot Models 48
4.1 Abstract ...... 48
4.2 Introduction ...... 49
4.3 Spot Price Models and Main Properties ...... 53
4.3.1 Constant Volatility Models ...... 53
4.3.2 Stochastic Volatility Extensions ...... 56
4.4 Forward Price Approximation ...... 59
4.4.1 One-Factor Model + SV ...... 60
4.4.2 Two-Factor Model + SV ...... 67
4.5 European Single-Name Options ...... 68
4.5.1 Smooth Payoff Function ...... 68
4.5.2 Nonsmooth Payoff: Calls and Puts ...... 73
4.6 European Two-Name Options ...... 74
4.6.1 Smooth Payoff Function ...... 74
4.6.2 Nonsmooth Payoff: Forward Spread ...... 85
4.7 Conclusions and Future Work ...... 86
4.8 Appendix ...... 88
5 Unspanned Stochastic Volatility for Energy Derivatives: A Forward
vi Price Approach 90 5.1 Abstract ...... 90 5.2 Introduction ...... 91 5.3 Model Description ...... 94 5.4 Single-Name Options Valuation ...... 97 5.5 Calendar Spreads Valuation ...... 103
5.5.1 Case I: A Constrained Correlation Structure (ρ12 := 0) ...... 104
5.5.2 Case II: A Constrained Volatility Structure (f := σX ) ...... 107 5.6 Calibration ...... 110 5.7 Conclusions ...... 113 5.8 Appendix I: Sketch of Proof of Theorem 5.4.1 ...... 114 5.9 Appendix II: Expansion Coefficients ...... 116
6 Conclusion 122
vii List of Figures
3.1 The NYMEX Light Sweet Crude Oil spot prices and simulated spot prices based on the calibration in Table 3.1...... 43 3.2 The relative root-mean squared-error of each forward curve based on the calibration in Table 3.1...... 44 3.3 This diagram depicts the evolution of the implied market prices of risk using the calibrated real-world and risk-neutral parameters...... 45
4.1 The annualized running five-day moving volatility of the NYMEX sweet crude oil spot price for the period 10/07/03 to 07/03/06...... 50 4.2 This diagram depicts typical forward curves implied the model for three choices of V . The long-run forward price is set at 61 in the left panel and
59 in the right panel. The spot is 60, β = 0.5 and σX = 0.2...... 67
5.1 Implied Volatility Curves on 4-Jan-2006, for three increasing term. . . . . 111 5.2 Implied Volatility Curves on 10-Jan-2006, for three increasing term. . . . 111
viii List of Tables
3.1 The calibrated real-world and risk-neutral model parameters using the NYMEX Light Sweet Crude Oil spot and futures data for the period 1/10/2003 − 25/07/2006...... 42 3.2 This table shows the evolution of the estimated risk-neutral parameters through time as more recent data is added to the calibration procedure. The average and standard deviation are reported using 176 days onwards. 44
∗ 5.1 The calibrated macro-parameters (volatility expansion method), with Di := √ Di, using the NYMEX Light Sweet Crude Oil implied volatilities on 04/01/2006...... 112
∗ 5.2 The calibrated macro-parameters (volatility expansion method), with Di := √ Di, using the NYMEX Light Sweet Crude Oil implied volatilities on 10/01/2006...... 113
ix Chapter 1
Introduction
The last few years have seen a tremendous increase of activity within the energy risk management industry. The various level of deregulation of many energy markets greatly increased the price volatility. This deregulation process offered the possibility to benefit from market price fluctuations, but it also created the likelihood of large adverse price movements. Most energy companies and major energy consumers have therefore increased the variety of risk management tools and techniques they use. Various financial institu- tions have also developed significant resources to address the hedging and speculative needs of other industry players as well as of their own energy portfolios.
These last financial developments generated a need for new pricing/hedging tech- niques adapted to the special contract structures found within energy markets. Both the fundamentals of the energy price processes and the structures of derivatives contracts significantly differ from their equity and interest rate counterparts. Many practitioners and academics alike have been working on both theoretical and practical issues of this relatively new branch of financial engineering.
The work presented in this thesis can be seen within this financial engineering context. It focuses mainly on the developments of new stochastic models and mathematical tech-
1 1 Introduction 2 niques toward a more complete theory of pricing/hedging of complex energy derivatives. The methods used are rooted in the now classical Martingale theory of contingent claim pricing, with a very particular flavour coming from the type of price processes and con- tract structures found within the energy markets. A distinctive source of difficulty within the present context is the lack of liquidity in spot markets, which explains the fact that most derivative contracts are written on forwards/futures. These futures contracts are typically heavily traded and therefore of a greater interest for risk management purposes. A concept that plays a crucial role in understanding illiquidity issues for spot prices is the cost of carry associated to most physical assets. Commodities are physical assets (in contrast to purely financial assets) that need to be produced, stored and redistributed, making the trading in futures contracts the fundamental price discovery process for market participants. In some case the commodity cannot even be stored (electricity), making its market price behaviour even harder to track and predict. An equally important concept related to illiquidity issues is the convenience yield, a yield which is naturally embedded into the cost of carry and that is meant to measure the benefit of holding the commodities.
The thesis is divided in four core chapters. Chapter 2 contains a concise overview of the minimal background needed in order to make sense of the mathematical and financial developments of the subsequent chapters. This include what the author considers to be a list of essential readings in mathematical finance and energy markets. References to both monographs and journal publications on basic topics are given. This list of readings is far from being complete and is not meant to be anything else than a strong recommendation for whoever would want to learn about the foundations of quantitative finance, with a particular interest in energy derivatives. Each of the subsequent three chapters are based on independent but overall strongly related projects.
Chapter 3 is based on the refereed journal publication Hikspoors-Jaimungal [HJ07]. In this work the authors propose a two-factor generalization of the classical one-factor 1 Introduction 3 exponential-OU model for spot price evolution. They also provide a simple methodology to price both forward contracts and spread options on forward prices. In doing so, the authors give the tools for a more flexible calibration of a mean-reversion spot model to fu- tures price as well as European one-name contracts. An efficient calibration methodology is explained and performed in the last section of the paper.
Chapter 4 is based on the forthcoming refereed journal publication Hikspoors-Jaimungal [HJ08a]. In this paper the authors extend the applications of the asymptotic PDE theory of singular perturbations to the field of commodity derivatives pricing. Based on realistic stochastic volatility (SV) spot price model assumptions, the authors provide a rigorous development of the theory for forward prices as well as single-name European options. The model assumptions include a stochastic volatility improvement of the two-factor spot model proposed in Chapter 3 (or [HJ07]). The authors then proceed to the development of a closed form approximation of the price of any two-name European contracts written on forward prices, a development that significantly extend all previous work on singular perturbation results in finance.
Chapter 5 is based on the paper Hikspoors-Jaimungal [HJ08b]. In this work the authors explore a stochastic volatility forward curve modeling approach in which the unhedgeable ”spot risk” is naturally embedded. The employed methodology generalize the one proposed in [ST08] by including a general stochastic volatility component based on a fast mean-reverting OU-process. This naturally links the concept of unspanned SV to the theory of singular perturbations, as initially applied to commodity derivatives in [HJ08a]. We build a closed form asymptotic approximation to the prices of a large subset of single- name and two-name European contracts on forwards. Interestingly enough, the price of any calendar spread options is completely determined once the model has been calibrated to related liquid single-name contracts.
A review of the relevant literature is provided within the introduction of each of the 1 Introduction 4 three above mentioned chapters. Finally, Chapter 6 concludes with a quick overview of the achieved results and some comments on ongoing and forthcoming work. Chapter 2
Background Material
The mathematical foundations and computational tools of financial engineering have been developed to an extent where most of participants are now specializing in one of the many sub branches of the global field of research. Hundreds of researchers have worked on these developments over the last four decades, making mathematical finance one of the most mature and sophisticated fields of applied mathematics, both from a computational and a theoretical point of view. This body of literature is obviously far too large to be summarized in any useful way. Here we deal only with what we consider to be a solid, yet minimal, background required in order to fully understand the technical developments proposed in this thesis.
We first give a quick overview of some general references where the relevant basic mathematical tools of financial engineering can be found. We do not explicitly review these results as it is assumed that the reader of the present document has already studied the foundations of mathematical finance. We then move to an exposition of what can be seen as an essential primer on both the basic results and terminology necessary for a full appreciation of the theory of singular perturbations in finance.
5 2 Background Material 6
2.1 Derivatives Pricing/Hedging
Many books contain an exposition of the theory and applications of quantitative finance, both from the probabilistic and the PDE approach. The so called bible of financial engineering is John Hull’s practical exposition of the subject [Hul05]. Although there is no need to read Hull’s book in order to understand most of the academic literature, it is nevertheless a must read for anyone who wants to aquire a more complete background than what the purely mathematical structures might suggest. There exist other references that offer a similar background, but we recommend to stick to Hull’s classic introduction, both for the quality of the exposition and its common usage in industry.
For a more mathematically sound foundation, but still very readable, one may consult both of Bjork [Bjo05] and Shreve [Shr04] textbooks. Both of these books treat the basics of Martingale derivatives pricing theory and succeed at succinctly developing the necessary mathematical background. The former is essentially a pedagogically improved version of the monograph [MR05] and is particularly strong in its exposition of interest rates (IR) models and fixed income derivatives pricing. This last reference offers one of the finest exposition of the change of numeraire technique and of its applications to IR derivatives pricing. Shreve’s book has a more complete exposition of the basic mathematical tools of stochastic calculus, and is particularly strong in its exposition of simple equity derivatives pricing. It also contains a clever and original primer on jump processes for financial applications. Oksendal [Oks03] is a more complete reference on stochastic differential equations.
The crucial topic of stochastic volatility models is introduced in Chapter 7 of [MR05] and Chapter 2 of [FPS00a]. Possibly the two most influential papers on the topic are Hull-White [HW87] and Heston [Hes93]; These two references laid the foundations for a very large section of modern mathematical finance.
Just as fundamental as stochastic volatility modeling is the study of non-Gaussian 2 Background Material 7 based models of returns. These often rely on the use of jump-diffusion processes (more generally Levy processes, or even semimartingales) and are extensively studied in numer- ous publications. The first to explicitly tackle the modeling of discontinuous asset price behaviour was Robert Merton in [Mer76]. A large number of researchers have subse- quently contributed to the advances of these techniques in finance, a solid and up to date account of which can be found in the monograph Cont-Tankov [CT04].
Finally and most importantly within the present context, there is an increasing number of books, monographs and articles published on energy derivatives pricing theory and its applications. A simple introduction to the basic modeling and pricing issues can again be found in [Hul05]. We nevertheless recommend a more specialized treatment of the topic such as the one offered in Clewlow-Strickland [CS00] and in the more recent monograph Eydeland-Wolyniec [EW03]. These last references constitute an excellent source of information on energy risk management; there, one can find the essential topics of the field, both from a practical and a theoretical perspective.
2.2 Asymptotic Tools and Terminology
This section presents the minimal background necessary in order to understand the ma- chinery of singular perturbation theory, especially as it is applied to contingent claims pricing/hedging. After having read this section, the work contained in Chapter 4 and Chapter 5 should be fairly self-contained. For a more complete review on the topic, the reader is referred to the monograph [FPS00a].
We start with a quick review of the main properties of the most widely used stochastic process in this thesis; the Ornstein-Uhlenbeck (OU) process, otherwise known as Vasicek’s model for short rates:
dYt = α (φ − Yt) dt + σ dWt , (2.1) 2 Background Material 8
where Wt is a Wiener process (Brownian motion) and all parameters are constants. It is easily seen that this process tends to be dragged back to its long run mean φ at a speed controlled by the parameter α. The magnitude of the random fluctuations is determined by the constant volatility parameter σ.
The solution of the SDE (2.1) with initial condition Y0 := y is well known and is expressed as
Z t −αt −α(t−s) Yt = φ + (y − φ)e + σ e dWs . (2.2) 0
Notice that this process is normally distributed, with mean φ + (y − φ)e−αt and variance
σ2 −αt 2α (1 − e ). A very important distribution related to the solution of a OU process is its long run distribution, that is, its distribution when the time parameter goes to infinity
σ2 (t → ∞). One can directly check that Y∞ ∼ N (φ, 2α ).
The so called infinitesimal generator of the OU process (2.1) is the following dif- ferential operator
∂ σ2 ∂2 L := α (φ − y) + . (2.3) ∂y 2 ∂2y
We make use of this last operator in order to define the invariant distribution of the OU process (2.1) as the distribution of the random variable Y such that
E [Lh(Y )] = 0 , (2.4) for all smooth and bounded functions h(·). For the rest of the present section we will denote the density of this last invariant distribution as Φ(·). Using this notation, we rewrite equation (2.4) as
Z ∞ Φ(y)Lh(y)dy = 0 , (2.5) −∞ 2 Background Material 9 which in turn can be rewritten in terms of the adjoint L∗ of the operator L in (2.4) as
Z ∞ h(y)L∗Φ(y)dy = 0 , (2.6) −∞ with the new operator defined by
∂ 1 ∂2 L∗ := −α ((φ − y) ·) + σ2 . (2.7) ∂y 2 ∂2y
Using the fact that equation (2.6) has to hold for all smooth and positive functions h(y), one can show that Y actually has the same distribution as Y∞, that is, the invariant
2 2 σ2 distribution of the OU process Yt is a normal N (φ, ν ), with ν := 2α .
In what follows as well as in the rest of the thesis, the symbol hgi will denote the aver- aging of the function g(·) with respect to the invariant distribution under consideration. Within the present context, that would be
Z ∞ hgi := E [g(Y∞)] = g(y)Φ(y)dy = 0 . (2.8) −∞
We call this averaging operation centering. As we are about to show, this notation is particularly handy to express concisely some necessary conditions naturally inherited by all solutions of Poisson equations.
We now review a few facts about Poisson equations that will be most useful, espe- cially in the developments of the various proofs contained in Chapter 4 and Chapter 5. In particular, we study some important characteristics of the following Poisson equation
L0X (y) + f(y) = 0 , (2.9)
where the operator L0 is given by
∂ ∂2 L := (φ − y) + ν2 . (2.10) 0 ∂y ∂2y 2 Background Material 10
Notice that this last operator is simply the infinitesimal generator of the OU process
−1 divided by α: L0 = α L. This operator will be at the center of many subsequent investigations. We apply the previously defined centering bracket on both side of equation (2.9) and make use of integration by parts as well as obvious vanishing boundary values to get
hfi = −hL0X i (2.11) Z ∞ = − (L0X (y)) Φ(y) dy (2.12) −∞ Z ∞ ∗ = X (y)(L0Φ(y)) dy (2.13) −∞ = 0 . (2.14)
This last computation clearly shows that solutions to the Poisson equation (2.9) must satisfy the so called centering condition: hfi = 0. This necessary condition will be heavily used in subsequent chapters.
Some growth properties of the solutions of Poisson equations (2.9) will also be of a crucial importance. We simply state the two most relevant such results without proof. First assume that the centering condition hfi = 0 is satisfied and that the function f(·) also has the following growth property
n |f(y)| ≤ C1 (1 + |y| ) , (2.15)
for some n ∈ N and an arbitrary constant C1. If f(y) is bounded (n := 0) then the growth of X (y) is at most logarithmic:
|X (y)| ≤ C2 [1 + log(1 + |y|)] , (2.16)
for some constant C2. On the other hand, if the growth of f(y) is as in (2.15) for n ≥ 1, 2 Background Material 11 then X (y) preserve the same growth structure:
n |X (y)| ≤ C3 (1 + |y| ) , (2.17)
for some constant C3. These last inequalities will provide us with powerful tools for analysing the properties of the various Poisson equations we shall encounter in Chapter 4 and 5. Chapter 3
First Paper: Energy Spot Price Models and Spread Options Pricing
3.1 Abstract
In this article, we construct forward price curves and value a class of two asset exchange options for energy commodities. We model the spot prices using an affine two-factor mean- reverting process with and without jumps. Within this modeling framework, we obtain closed form results for the forward prices in terms of elementary functions. Through measure changes induced by the forward price process, we further obtain closed form pricing equations for spread options on the forward prices. For completeness, we address both an actuarial and a risk-neutral approach to the valuation problem. Finally, we provide a calibration procedure and calibrate our model to the NYMEX Light Sweet Crude Oil spot and futures data, allowing us to extract the implied market prices of risk for this commodity.
12 3 Energy Spot Price Models and Spread Options Pricing 13
3.2 Introduction
Energy commodity markets are fundamentally different from traditional financial security markets in several ways: Firstly, these markets lack the same level of liquidity that the majority of financial markets enjoy. Secondly, the storage costs of most commodities translate into peculiar price behavior; some commodities are extremely difficult to store or cannot be stored at all – electricity being a prime example. Thirdly, partly due to the structural issues surrounding energy price determination, electricity prices are typically exposed to very high volatility and to large shocks. Finally, commodity prices tend to have strong mean reverting trends. These stylized empirical facts are well documented in, for example, [CS00], [CD03], [EW03] and [Hul05].
The world wide energy commodities markets have created a need for a deeper quan- titative understanding of energy derivatives pricing and hedging. We contribute to this program firstly by proposing a two-factor mean-reverting spot price process, both with and without a jump component, and secondly by carrying out the explicit valuation of spread options written on two forward prices. The spot price model is similar in spirit to the two-factor model proposed in [Pil97]; however, in that work the second factor follows a geometric Brownian motion and, therefore, in the long run, the targeted mean blows up. Instead, we chose the mean-reverting level of the first factor to mean-revert to a second long-run mean. Our modeling framework can then be viewed as a perturbation on the standard one-factor mean-reverting approach. This is an appealing approach as the one-factor model has been extensively studied and approximately fits forward price curves. Adding a perturbation on top of this first order model allows us to correct some of the deficiencies of the one-factor model while maintaining tractability. In particular, the second factor perturbation does not change the stationary behavior that the one-factor model enjoys. We delegate the details of our purely diffusive model, and its relation and differences to the classical [Pil97] model, to Section 3.3.1 and of our jump-diffusion model 3 Energy Spot Price Models and Spread Options Pricing 14 to Section 3.5.1.
Much like the financial markets, energy markets are rife with derivative products. However, one product stands out among the many that are traded over the counter: spread options. These options provide the owner with the right to exchange a prespecified quantity of one asset for another, at a fixed cost. An even more popular option is the spread option on forward prices which allows the holder to exchange two forward contracts, possibly with differing maturity dates, rather than the commodity. The holder of such an option receives at maturity T a payoff of
ϕ(F (1) ,F (2) ) := max F (1) − α F (2) − K, 0 . (3.1) T,T1 T,T2 T,T1 T,T2
It is well known that even when the commodity prices are modeled as geometric Brownian motions (GBMs), no closed form solution exists for K 6= 0. As such, we focus on the zero exchange cost case K = 0 which we succeed in valuing in closed form under our two-factor mean-reversion modeling assumptions. Given our closed form solutions, the general case K 6= 0 can be valued using either Monte Carlo or PDE methods with our K = 0 result acting as a control variate.
In a financial markets context, before proceeding to the valuation of derivatives, the objective measure is transformed to an equivalent risk-neutral measure. However, in the context of energy derivatives, due to the illiquidity issue, such a measure change is by no means necessary. Therefore, rather than immediately moving to a risk-neutral valuation procedure, we first present a simple actuarial valuation approach for pricing exchange options in Section 3.3.2. This approach has been adopted by some industry participants and is justified by assuming that the risks associated with the energy prices are non- diversifiable (see for example [Hul05]). [Mar78] first valued exchange options assuming asset prices are GBMs under the risk-neutral measure and by utilizing a measure change induced by treating one of the assets as numeraire. However, since commodities are not 3 Energy Spot Price Models and Spread Options Pricing 15 liquid, their spot prices cannot act as a numeraire. Nonetheless, we show that there is a closely related measure change which renders the valuation in closed form even under the actuarial approach.
Although some industry participants adopt an actuarial valuation procedure, risk- neutral approaches are still very popular. In Section 3.4, we specify a class of equivalent risk-neutral measures which maintains the structure of the real-world dynamics. This allows us to reuse much of the valuation technology developed in Section 3.3.2. Once again, we show that an equivalent measure provides closed form pricing equations for spread options.
Most energy commodities are adequately modeled by diffusive processes; however, electricity prices themselves contain several jumps . Section 3.5 is devoted to a jump- diffusion generalization of our previous results appropriate for spark-spread options – exchange options in which a fuel commodity is exchanged for electricity. Using the affine structure of our two-factor model with jumps, we obtain the forward prices as a solution to a system of coupled ODEs which we explicitly solve. Furthermore, through measure changes and Fourier transform methods, `ala [DPS00] and [CM99], we obtain closed form formula for the price of spark-spread options on forwards.
We complete the paper in Section 3.6 with a calibration procedure that fits the model to spot and forward prices simultaneously. Calibrating to both spot and forward prices allows us to further extract the market prices of risk implied by the data. We apply the calibration procedure to the NYMEX Light Sweet Crude Oil spot and futures data and report on the stability of the implied model parameters as well as on the implied market prices of risk. Interestingly, the real-world mean-reversion rates are found to be higher than the risk-neutral mean-reversion rates. Furthermore, the rate of mean-reversion of the stochastic long-run mean level is lower than the mean-reversion rate of the log-spot price process. We find that these features are reflected in the market prices of risk themselves. 3 Energy Spot Price Models and Spread Options Pricing 16
3.3 Real World Dynamics and Pricing
3.3.1 Model Specifications
A quick glance at historical spot prices for energy markets shows that traditional geometric Brownian motion models, even as a first order model, are inadequate. A successful model must include mean reversion as an essential feature. For early use of such models see the papers by [GS90] and [CS94]. These early one-factor models are a good first order model; however, as energy derivatives will often have maturities extending into months, or even years, such first order models require improvement. They invariably cannot match the term structure of forward prices for example. To this end, [Pil97] first suggested the two-factor mean-reverting model:
(1) dSt = β(θt − St) dt + σS St dWt , (3.2)
(2) dθt = α θtdt + σθ θt dWt , (3.3)
(1) (2) where the two Brownian risk factors are correlated: d[W ,W ]t = ρ dt. In this model,
θt represents the stochastic long-run mean to which spot prices St revert. This additional degree of stochasticity helps to correct some of the biases that a fixed long-run mean pro- duces. In this parametrization, the mean reverting level is a geometric Brownian motion and can therefore grow without bound leading to non-stationary spot price processes. To circumvent this problem, and to assist with obtaining closed form formulae for spread options, we propose to model the log spot-price with an affine two-factor mean-reverting process. Much like Pilipovic’s model, the first factor mean-reverts to stochastic level; however, we ensure that the stochastic mean-reverting level also mean-reverts to a second long-run mean. With such a parametrization, the distribution of spot-prices is stationary, prices do not grow without bound, and the model remains within the Affine modeling 3 Energy Spot Price Models and Spread Options Pricing 17 class.
If the individual assets are driven by a two-factor model, then four driving factors are
(i) (i) required to value spread options – two for each asset. Let {Wt }0≤t≤T and {Zt }0≤t≤T , with i = 1 or 2, denote these four Brownian risk factors and F = {Ft}0≤t≤T denote the natural filtration generated by these processes. The measure P will denote the real-world probability distribution and {Ω, F, P} is used to denote the complete stochastic basis for (i) the probability space. The spot-prices {St }0≤t≤T , with i = 1 or 2, are obtained via an exponentiation of the driving risk-factors. More specifically, the spot-prices are defined as follows:
(i) (i) (i) St := exp gt + Xt , i = 1, 2 . (3.4)
Seasonality is an important feature of some commodity prices; we therefore include the
(i) seasonality term gt using the following popular ans¨atz:
n (i) (i) X (i) (i) gt = A0 t + Ak sin(2π k t) + Bk cos(2π k t) . (3.5) k=1
For calibration stability, n is typically kept small: n = 1 or 2. In our subsequent calcula-
(i) tions we leave gt general, assuming only smoothness and differentiability.
To complete the specification of the two-factor model which drives the spot-prices,
(i) Xt is assumed to satisfy the following coupled SDEs:
(i) (i) (i) (i) (i) dXt = βi Yt − Xt dt + σX dWt , (3.6)
(i) (i) (i) (i) dYt = αi φi − Yt dt + σY dZt . (3.7)
(i) (i) Here, βi controls the speed of mean-reversion of Xt to the stochastic long-run level Yt ; (i) αi controls the speed of mean-reversion of the long-run level Yt to the target long-run (i) (i) mean φi; σX and σY control the size of the fluctuations around these means. 3 Energy Spot Price Models and Spread Options Pricing 18
To reduce the parameter space, the measure P is chosen such that the following simple correlation structure is imposed on the Brownian motions:
(1) (2) d[W ,W ]t = ρ12 dt , (3.8)
(i) (i) d[W ,Z ]t = ρi dt , i = 1, 2 , (3.9) and all other cross correlations are zero. This structure allows the main driving factors
(i) Xt to be correlated to one another and their own idiosyncratic long-run mean-reverting (i) (1) (2) processes Yt , while this structure forces the long-run reverting factors Yt and Yt to have an instantaneous correlation of zero. It is a straightforward, albeit tedious, matter to generalize this correlation structure.
(i) The coupled SDEs (3.6)-(3.7) can be solved by (i) first solving (3.7) for Yt – which is the standard mean-reverting Ornstein-Uhlenbeck process – to obtain
Z t (i) (i) −αi(t−s) (i) −αi(t−u) (i) Yt = φi + Ys − φi e + σY e dZu ; (3.10) s
(i) and then (ii) substituting this result into (3.6) to solve for Xt while accounting for the (i) (i) (i) correlation and feedback of Yt into Xt . After some tedious calculations Xt can be represented as
(i) (i) −βi(t−s) (i) (i) (i) Xt = Gs,t + e Xs + Ms,t Ys Z t Z t (i) −βi(t−u) (i) (i) (i) (i) + σX e dWu + σY Mu,t dZu , (3.11) s s
βi (i) (i) where γi := , and G and M are the deterministic functions αi−βi s,t s,t
(i) −βi(t−s) −αi(t−s) Ms,t := γi e − e , (3.12)
(i) −βi(t−s) (i) Gs,t := φi 1 − e − φi Ms,t . (3.13) 3 Energy Spot Price Models and Spread Options Pricing 19
Armed with our two-factor model and the solutions (3.10)-(3.11), we now focus our at- tention on the valuation of spot spread options and defer the valuation of spreads on forwards and model calibration to sections 3.4.4 and 3.6 respectively.
3.3.2 Spot Spread Valuation : An Actuarial Approach
Much like the financial markets, energy markets are rife with derivative products. How- ever, one product stands out among the many that are traded over the counter: spread options. These options provide the owner with the right to exchange a specified quantity of one asset for another, at a fixed cost. The holder of such an option receives a maturity payoff of
(1) (2) (1) (2) ϕ(ST ,ST ) := max ST − αST − K, 0 . (3.14)
When the cost of exchanging K is set to zero, the option is known as a Margrabe or exchange option [Mar78]. Various approximations for the general (spot) spread option, under simple diffusion processes, have been studied in the literature and the reader is referred to [CD03] for an excellent overview and further references. In the context of electricity markets, this option is known as the spark-spread option and α represents the heat rate of a given plant. The heat rate encapsulates the plant’s profitability by specifying the number of units of the underlying commodity (such as oil or natural gas) which produces one unit of power – this product is studied in Section 3.5. If the exchange is between crude oil and a refined product (such as gasoline) the option is known as a crack-spread option. Many other specific examples of exchange options exist in the energy sector. In all cases, the exchange option can be used to hedge against market movements of spot prices or, alternatively, to speculate on those moves. In either case, a valuation framework is required.
It is difficult and sometimes not viable to store electricity and energy commodities; 3 Energy Spot Price Models and Spread Options Pricing 20 this results in an illiquid spot market. [HP81] demonstrated that the absence of arbitrage is equivalent to the existence of a measure, not necessarily unique, under which the rela- tive price process of tradable assets to the money market account are martingales. Such measures are known as a risk-neutral measures. However, this conclusion has one im- portant assumption – unrestricted and liquid trading of the underlying asset. In illiquid (spot) energy markets, it may be dubious to adopt a risk-neutral pricing framework, and although we ultimately proceed with that program, we first take an actuarial approach. By assuming that the risks associated with the energy prices are non-diversifiable, it is possible to justify an actuarial approach to pricing derivatives [Hul05] which values an option as its discounted real-world expected payoff.
Definition. The actuarial valuation principle assigns the following price to a T -maturity
(1) (2) contingent claim with payoff ϕ(ST ,ST ):
h i P (1) (2) Πt,T := P (t, T ) Et ϕ ST ,ST . (3.15)
The notation Et[A] represents the expectation of A conditional on the filtration Ft.
Throughout the article we assume the (possibly random) interest rates to be independent from other risk factors and write the price at time t of a zero-coupon bond maturing at T as P (t, T ).
In the traditional valuation procedure, expectations are taken under the risk-neutral measure Q; here, however, the relevant measure is the real-world one P. This complicates the problem somewhat. When the asset and the derivative are tradable, it is possible to use a numeraire change to value the Margrabe option; in the present context the asset cannot be used as numeraire and the relevant measure is not the risk-neutral one. Nonetheless, it is possible to adopt a similar strategy; to this end, define the auxiliary 3 Energy Spot Price Models and Spread Options Pricing 21 process
h i (i) P (i) Ht,T := Et ST . (3.16)
(i) If the expectation in (3.16) is computed under a risk-neutral measure, then Ht,T represents (i) the T -maturity forward price, which motivates us to coin Ht,T the T -maturity pseudo- forward price process. At maturity this “price” process coincides with the spot-price
(i) (i) HT,T = ST allowing the actuarial value of the exchange option to be written:
P (1) (2) Πt,T = P (t, T ) Et HT,T − α HT,T . (3.17) +
The pseudo-forward price process has two other notable properties: (i) its expectation
h (i) i (i) P is bounded at all finite times E |Ht,T | = H0,T < +∞ for all t < T , and (ii) it is (i) a P-martingale. These two properties allow a normalized version of Ht,T to assist in transforming the measure P into a particularly convenient measure for pricing. This measure change can, in some sense, be interpreted as being induced by using the pseudo- forward price process as a numeraire asset. The following Theorem contains one of our main tools.
Theorem. 3.3.1 Let {ηt}0≤t≤T denote the Radon-Nikodym process
! (2) de Ht,T η := P := . (3.18) t d (2) P t H0,T
Then, for any A ∈ FT
P Pe(A) = E [I(A) ηT ] , (3.19)
i i and in particular Wft and Zet (i = 1, 2) defined by
Z t h i (2) (2) (2) −β2(T −u) (2) (2) Wft = Wt − σX e + ρ2σY Mu,T du (3.20) 0 3 Energy Spot Price Models and Spread Options Pricing 22
Z t h i (2) (2) (2) (2) (2) −β2(T −u) Zet = Zt − σY Mu,T + ρ2σX e du (3.21) 0 Z t h i (1) (1) (2) −β2(T −u) Wft = Wt − ρ12σX e du (3.22) 0 Z t h i (1) (1) (2) −β2(T −u) Zet = Zt − ρ1ρ12σX e du (3.23) 0 are Pe−Wiener processes with correlation structure
(1) (2) d[Wf , Wf ]t = ρ12 dt , (3.24)
(i) (i) d[Wf , Ze ]t = ρi dt , i = 1, 2 , (3.25) and all other cross correlations zero.
Proof. Given properties (i) and (ii) above and η0 = 1, it is clear that ηt is a Radon- Nikodym derivative process and equation (3.19) immediately follows. To demonstrate
(i) (i) (i) that Wft and Zet are Pe-Wiener processes substitute (3.11) into Ht,T and then compute the expectation explicitly as follows:
h n (i) P (i) (i) −βi(T −t) (i) (i) (i) Ht,T = Et exp gT + Gt,T + e Xt + Mt,T Yt Z T Z T (i) −βi(T −u) (i) (i) (i) (i) +σX e dWu + σY Mu,T dZu t t
(i) (i) (i) −βi(T −t) (i) (i) (i) = exp gT + Gt,T + Rt,T + e Xt + Mt,T Yt (3.26)
Here,
h(t, T ; 2β ) 2 2 R(i) := i σ(i) + γ σ(i) + 2ρ γ σ(i)σ(i) t,T 2 X i Y i i X Y 2 h(t, T ; 2α ) 2 −h(t, T ; α + β ) γ σ(i) + ρ γ σ(i)σ(i) + i γ σ(i) (3.27) i i i Y i i X Y 2 i Y h(t, T ; a) := (1 − e−a(T −t))/a . (3.28) 3 Energy Spot Price Models and Spread Options Pricing 23
(i) The Girsanov kernel can now be read off directly from (3.26) and the solution for Xt (i) and Yt given in (3.10)-(3.11). Through Girsanov’s theorem we find that (3.20)-(3.23) are Pe-Wiener processes. 2
Corollary 3.3.2 The actuarial valuation formula (3.15) can be transformed to
h i P (1) (2) Πt,T := P (t, T ) Et ϕ HT,T ,HT,T (1) (2) ϕ HT,T ,HT,T (2) Pe = P (t, T ) Ht,T Et (2) . (3.29) HT,T
Proof. Use ηt to change the measure. 2
It now remains to compute the expectation appearing in (3.29) under the trans- formed measure Pe. Recall that a process Mt is a Pe-martingale if and only if the process (1) (2) Mt (dPe/dP)t is a P-martingale. Consequently, both Ht,T := Ht,T /Ht,T and (dP/dPe)t are Pe-martingales. This, together with Corollary 3.3.2, reduces the actuarial price of the Margrabe spread option to
(2) Pe Πt,T = P (t, T ) Ht,T Et [(HT,T − α)+] (3.30)
Since Ht,T is a Pe-martingale, its drift under the Pe-measure is zero. Putting this together with equation (3.26), we find that Ht,T satisfies the SDE:
dH t,T (1) −β1(T −t) (1) (2) −β2(T −t) (2) = σX e dWft − σX e dWft Ht,T (1) (1) (1) (2) (2) (2) + σY Mt,T dZet − σY Mt,T dZet (3.31)
This expression clearly shows that Ht,T is a geometric Brownian motion with time de- pendent (but deterministic) volatility; consequently, its terminal value can be expressed in terms of its initial value via HT,T = Ht,T exp (Ut,T ), where Ut,T is a normal random 3 Energy Spot Price Models and Spread Options Pricing 24
1 2 2 variable with mean equal to − 2 (σt,T ) and variance equal to (σt,T ) . Here,
Z T Z T 2 Pe (1) −β1(T −s) (1) (2) −β2(T −s) (2) (σt,T ) = Et σX e dWfs − σX e dWfs t t Z T Z T 2# (1) (1) (1) (2) (2) (2) +σY Ms,T dZes − σY Ms,T dZes t t
(1) (2) (1) (2) = 2 Rt,T + 2 Rt,T − 2 ρ12 σX σX h(t, T ; β1 + β2) . (3.32)
(i) (i) The deterministic functions Ms,T and Rt,T are as in (3.12) and (3.27) respectively. It is now a straightforward matter to recover the final result of this section – a Black-Scholes like expression for the actuarial price of the exchange option.
Proposition 3.3.3 The actuarial value at time t of the T -maturity exchange option is
h (1) (2) i Πt,T = P (t, T ) Ht,T Φ(d + σt,T ) − α · Ht,T Φ(d) (3.33)
with σt,T as in (3.32) and d defined as
log Ht,T − 1 (σ )2 d := α 2 t,T . (3.34) σt,T
3.4 Risk-Neutral Dynamics and Pricing
In complete market settings, there exists a unique equivalent measure which induces the relative price process of tradable assets to be martingales. This measure is known as the risk-neutral measure Q. In the present context, the underlying asset is not tradable in the usual sense due to the illiquidity issue and potentially large storage costs. In the previous section we dealt with this issue by resorting to an actuarial valuation procedure and assigned a price equal to the discounted expectation of the terminal payoff under the real-world measure. However, one can in principle still utilize risk-neutral methodologies 3 Energy Spot Price Models and Spread Options Pricing 25 adjusting for the incompleteness of the market settings. Within such incomplete markets there may exists many equivalent risk-neutral measures; it is the job of the market as a whole, via trading of derivatives, to decide which measure prevails at any one given point in time. In this next section, we provide a class of equivalent martingale measures that maintains the structure of our real-world dynamics for asset prices. These measures are then used to obtain forward prices and value spread options.
3.4.1 Measure Change
In this section, we introduce a class of risk-neutral measure changes which maintains the real-world structure of the asset dynamics. The following Lemma introduces the new measure induced by a four dimensional market price of risk vector.
Lemma 3.4.1 Let {ζt}0≤t≤T denote the Radon-Nikodym process,
Z t dQ (1) (1) (1) (1) (2) (2) (2) (2) ζt := = E λu dWu + ψu dZu + λu dWu + ψu dZu , (3.35) dP t 0 where E(At) is the Dolean-Dade’s exponential of the process At. Then for any A ∈ FT we have,
P Q(A) = E [A ζT ] . (3.36)
In particular the following are Q-Wiener processes:
Z t (1) (1) (1) (2) (1) W t = Wt − λu + ρ12λu + ρ1ψu du (3.37) 0 Z t (1) (1) (1) (1) Zt = Zt − ρ1λu + ψu du (3.38) 0 Z t (2) (2) (1) (2) (2) W t = Wt − ρ12λu + λu + ρ2ψu du (3.39) 0 Z t (2) (2) (2) (2) Zt = Zt − ρ2λu + ψu du (3.40) 0 3 Energy Spot Price Models and Spread Options Pricing 26 with correlation structure,
h (1) (2)i d W , W = ρ12 dt , (3.41) t h (i) (i)i d W , Z = ρi dt , i = 1, 2 , (3.42) t and all other cross-correlations zero.
Proof. Decompose the correlated processes into uncorrelated processes and apply Gir- sanov’s Theorem. 2
Notice that there are no restrictions on the form of the market prices of risk other than the usual integrability ones. In particular, the drifts under the risk-neutral measure Q are not constrained to the risk-free rate. This is precisely the effect of incompleteness in the present context. The following Theorem applies constraints on the market prices of risk such that the risk-neutral dynamics and the real-world one are of the same form.
Theorem. 3.4.2 If the market price of risk processes are chosen as follows:
(i) (i) (i) (i) (i) (i) λt = λ + λX Xt + λY Yt , (3.43)
(i) (i) (i) (i) (i) (i) ψt = ψ + ψX Xt + ψY Yt , (3.44) subject to the constraints ( (i, j) ∈ {(1, 2), (2, 1)} )
(i) (i) ψX = −ρi λX , (3.45)
(i) (j) (i) λ + ρ12λ + ρiψ = 0 , (3.46)
(i) (j) (i) (i) (j) (i) λX + ρ12λX + ρiψX = − λY + ρ12λY + ρiψY , (3.47)
(i) (i) (i) αi = αi − σY ρiλY + ψY , (3.48)
(i) (i) (i) αi φi = αiφi + σY ρiλ + ψ , (3.49)
(i) (i) (j) (i) βi = βi + σX λY + ρ12λY + ρiψY , (3.50) 3 Energy Spot Price Models and Spread Options Pricing 27
(i) (i) then the risk-neutral dynamics of Xt and Yt remain within the same class as (3.6)- (3.7). In particular,
(i) (i) (i) (i) (i) dXt = βi (Yt − Xt ) dt + σX dW t (3.51)
(i) (i) (i) (i) dYt = αi (φi − Yt ) dt + σY dZt . (3.52)
(i) (i) Proof. Insert the expressions for the Q-Wiener processes W t and Zt into (3.6)-(3.7). Collect similar terms and equate coefficients. 2
This ans¨atzmay seem restrictive; however, even though the risk-neutral dynamics remains within the same class as the real-world one, the coefficients driving that dynamics may be significantly different. This flexibility is sufficient for the simultaneous calibration of the risk-neutral and real-world model parameters, while remaining parsimonious.
3.4.2 Forward Prices
Since the risk-neutral dynamics of the driving diffusion processes (3.51)-(3.52) are of the same form as they were under the objective measure (3.6)-(3.7), the forward price curves can easily be extracted from equation (3.26). This is because, within a risk-neutral framework, the forward prices are defined as
h i (i) Q (i) Ft,T := Et ST , (3.53) the precise risk-neutral analog of the pseudo-forward price defined in (3.16). All that remains is to change the P-parameters for the Q-parameters.
Proposition 3.4.3 The forward prices associated to commodity i = 1, 2 are given by
(i) (i) (i) (i) (i) −βi(T −t) (i) (i) F (t, T ) = exp gT + Rt,T + Gt,T + e Xt + M t,T Yt (3.54) 3 Energy Spot Price Models and Spread Options Pricing 28
(i) (i) (i) where the expressions for M t,T , Gt,T and Rt,T are supplied in equations (3.12)-(3.13) and
(3.27) respectively – with the appropriate change of parameters (αi → αi and so on...).
These results can be viewed as an extension of the one-factor model [CF05] study, albeit without jumps. In section 3.5, we address the two factor model with jumps.
3.4.3 Spot Spread Valuation
To value the option under a risk-neutral measure, we follow along the same lines as in Section 3.3.2. In the present context, the measure change is the one induced by using the forward price to drive the measure change. To this end, define a new measure Qe via the Radon-Nikodym derivative process
h i ! Q S(2) (2) dQe Et T Ft,T = h i = (2) . (3.55) dQ Q (2) F t E0 ST 0,T
All steps leading to Proposition 3.3.3 remain valid in this new context, and rather than repeating them, we simply quote our final risk-neutral pricing result.
Proposition 3.4.4 The risk neutral value at time t of the exchange option with maturity T is:
h (1) (2) i Πt,T = P (t, T ) Ft,T Φ(d + σt,T ) − α Ft,T Φ(d) (3.56) where d defined as
2 log Ft,T − (σt,T ) d := α 2 (3.57) σt,T and σt,T as in (3.32) – with all P-parameters replaced by Q-parameters (i.e., αi → αi and so on...). 3 Energy Spot Price Models and Spread Options Pricing 29
(i) It is important to note that the market provides quotes for the forward curve Ft,T , i = 1, 2 for a set of maturities T = {T1,...,Tn}. These curves can be used to calibrate the risk- neutral parameters. Once the model has been calibrated to market data, the resulting pricing rules are just as simple to use as the Black-Scholes formula for a European option
(i) on a single asset. Although the explicit expressions for the forward prices Ft,T and the effective volatility σt,T are somewhat bulky, they involve nothing more complex than exponentiation and are therefore very efficient to calculate.
3.4.4 Forward Spread Valuation
In the previous sections we focused on valuing a spread option based on the future spot prices; however, a more popular derivative product involves the spread between the for- ward prices of the two assets (possibly with differing maturities). Such spreads on forwards pay
(1) (2) (1) (2) ϕ FT,T ,FT,T = FT,T − α FT,T (3.58) 1 2 1 2 +
at the maturity date T where it is implicit that T1 ,T2 ≥ T . We can once again use a measure change to simplify the calculations, this time it is convenient to use the T2- maturity forward price of asset 2, i.e. F (2) , to induce a measure change. In particular we t,T2 define the Radon-Nikodym derivative process
! (2) de Ft,T Q := 2 . (3.59) dQ F (2) t 0,T2
The time t price of the forward spread option is therefore
F Q (1) (2) Πt,T := P (t, T ) Et FT,T − α FT,T 1 2 + = P (t, T ) F (2) Qe (F − α) . (3.60) t,T2 Et T ;T1,T2 + 3 Energy Spot Price Models and Spread Options Pricing 30
Here, F := F (1) /F (2) is the ratio of the two relevant forward prices. In analogy with t;T1,T2 t,T1 t,T2 our earlier calculations, the relative process Ft;T1,T2 is a Qe-martingale and therefore its Qe-dynamics is driftless. Following along the same arguments as in Section 3.3.2, it is easy
∗ ∗ to show that FT ;T1,T2 = Ft;T1,T2 exp{Ut;T,T1,T2 } where Ut;T,T1,T2 is a normal random variable with mean equal to − 1 (σ∗ )2 and variance equal to (σ∗ )2. The explicit form 2 t;T,T1,T2 t;T,T1,T2 for the variance is
2 ∗ 2 (1) (σt;T,T1,T2 ) := γ1σY [h(t, T1; 2α1) − h(T,T1; 2α1)] 2 (2) + γ2σY [h(t, T2; 2α2) − h(T,T2; 2α2)] 2 2 (1) (1) (1) (1) + σX + γ1σY + 2ρ1γ1σX σY h(t, T1; 2β1) − h(T,T1; 2β1)
2 2 (2) (2) (2) (2) + σX + γ2σY + 2ρ2γ2σX σY h(t, T2; 2β2) − h(T,T2; 2β2)
2 (1) (1) (1) − 2 γ1σY + 2ρ1γ1σX σY h(t, T1; α1 + β1) − h(T,T1; α1 + β1)
2 (2) (2) (2) − 2 γ2σY + 2ρ2γ2σX σY h(t, T2; α2 + β2) − h(T,T2; α2 + β2) (1) (2) −2ρ12σX σX exp −β1(T1 − T ) − β2(T2 − T ) h(t, T ; β1 + β2) (3.61) where, h(t, T ; a) is defined in (3.28). The pricing equation (3.60) now reduces to a Black- Scholes like pricing result.
Proposition 3.4.5 The risk neutral value at time t of the forward spread option (3.58) is
h i ΠF = P (t, T ) F (1) Φ(d∗ + σ∗ ) − α F (2) Φ(d∗) (3.62) t,T t,T1 t,T t,T2
∗ 2 ∗ with (σt,T ) as in (3.61) and d defined as
Ft;T1,T2 1 ∗ 2 ∗ log α − 2 (σt,T ) d := ∗ . (3.63) σt,T 3 Energy Spot Price Models and Spread Options Pricing 31
Not surprisingly, the pricing result is very similar to the one in Proposition 3.4.4 and reduces to it when T = T1 = T2.
3.5 Spot Prices with Jumps
The two factor diffusion model captures the main characteristics of most energy spot prices, however, it cannot account for the possibility of sudden jumps in the price data. Such behavior is particularly important for modeling electricity prices and various spreads contingent on electricity and other hedging assets. The most important example of such
(1) (2) option is the so-called spark-spread option which pays FT,T − αFT,T − K at the 1 2 + maturity date T . Here F (1) := QS(1) represents the electricity forward price, F (2) rep- t,T1 Et T1 t,T2 resents the forward price of the commodity used to generate electricity, and α represents the heat rate which encapsulates the number of units of energy that the plant produces per unit of commodity. Notice that the structure of this option allows forward prices of differing maturities to be used as the underlying. As commented earlier on, closed form solutions, even for the purely diffusion case, are not accessible for general strike levels; consequently, we limit ourselves to the exchange option with a strike of zero.
3.5.1 Model Specification
For brevity, we now focus solely on the risk-neutral valuation procedure, and provide model specifications directly under the risk-neutral measure. Typically, when electric- ity prices jump they revert back to normal levels very quickly. A widely used model specification incorporates jumps and diffusions simultaneously as follows:
d ln(St) = α(θ − ln(St−)) dt + σ dWt + dQt . (3.64) 3 Energy Spot Price Models and Spread Options Pricing 32
Regardless of the specification of the jump process Qt, such models suffer from unre- alistically large diffusive volatilities and mean-reversion rates. This occurs because the process must revert very quickly to normal levels after a large jump, implying a high mean-reversion rate α. This in turn induces an artificially large diffusive volatility, since otherwise all diffusive components would revert to the mean extremely quickly and, ex- cluding the jumps, the paths would appear essentially deterministic.
We avoid these problems by splitting the jump component from the diffusion compo- nent and modeling them separately. Specifically, define the (power) spot price by
(1) n (1) (1) o St := exp gt + Xt + Jt , (3.65)
(i) (i) where Xt and Yt satisfy the usual two-factor SDEs (3.51)-(3.52), and the new jump component Jt is defined via
dJt = −κ Jt− dt + dQt , (3.66)
PNt with Qt a compound Poisson process: Qt := 1 li, where Nt is a time inhomogeneous
Poisson process with activity rate λ(s), and {li} the set of i.i.d. jump sizes with distri- bution function Fl(u). Furthermore, Jt− denotes the value of Jt prior to any jump at time t. The jump component Jt mean-reverts to zero with rate κ; typically, κ will be quite large because electricity prices revert back to normal very quickly after a jump. This empirical fact has no direct bearing on the valuation procedure, however, it does attribute to the manner in which we have split the jump component from the diffusion component. We allow the activity rate to vary with time to permit seasonality effects in the rate of jump arrivals; however, we restrict it to be deterministic – it is possible to generalize to stochastic activity rates; however, the additional modeling flexibility renders the calibration process unstable. Finally, it is well known that diffusions and jump pro- 3 Energy Spot Price Models and Spread Options Pricing 33 cesses cannot have any instantaneous correlations, while this does not preclude the jump size from depending on the Brownian risk factors we make the natural assumption that
Nt and {li} are independent of all the Q-Brownian processes.
3.5.2 Forward Prices
(1) Equipped with this jump-diffusion model, we now derive the forward price Ft,T associated (1) with the spot St . As usual, the forward price is the risk-neutral expectation of the asset price at the maturity
h n oi (1) (1) (1) Q (1) Q (1) (1) f(t, Xt ,Yt ,Jt) := Ft,T := Et ST = Et exp gT + XT + JT . (3.67)
Rather than computing this expectation directly, we make use of the affine form of the processes along the lines of [DPS00]. Since f is a Q-martingale, it satisfies the zero-drift (1) (1) condition Af = 0 where A is the generator of the process t, Xt ,Yt ,Jt . The affine ans¨atz:
(1) (1) n (1) (1) o f(t, Xt ,Yt ,Jt) = exp A(t, T ) + B(t, T ) Xt + C(t, T ) Yt + D(t, T ) Jt , (3.68)
(1) with terminal conditions A(T,T ) = gT , B(T,T ) = 1, C(T,T ) = 0 and D(T,T ) = 1, reduces the PDE Af = 0 to the equivalent system of coupled ODEs:
Bt − β1B = 0 , (3.69)
Ct + β1B − α1C = 0 , (3.70)
Dt − κD = 0 , (3.71) (σ(1))2 A + α φ C + X B2 t 1 1 2 (1) 2 Z ∞ (σY ) 2 (1) (1) D·u + C + ρ1σX σY BC = − λ(u) e − 1 dFl(u) (3.72) 2 −∞ 3 Energy Spot Price Models and Spread Options Pricing 34
Although rather tedious, standard methods can be used to solve this system and obtain the forward price.
(1) Proposition 3.5.1 The forward price for the two-factor jump-diffusion spot process St is
n (1) o (1) (1) −β1(T −t) (1) (1) −κ(T −t) Ft,T = exp At,T + e Xt + M t,T Yt + e Jt , (3.73)
(1) where the deterministic function M t,T is provided in equation (3.12),
Z T (1) (1) −κ(T −s) At,T = gT + λ(s) ϕl e − 1 ds t − α1γ1φ1 h(t, T ; α1) − h(t, T ; β1) 1 2 + γ σ(1) h(t, T ; 2α ) + h(t, T ; 2β ) − 2h(t, T ; α + β ) 2 1 Y 1 1 1 1 1 2 + σ(1) h(t, T ; 2β ) 2 X 1 (1) (1) + ρ1γ1 σX σY h(t, T ; 2β1) − h(t, T ; α1 + β1) , (3.74)
and ϕl(u) is the m.g.f. of the individual jump sizes li,
Z ∞ Q u l1 u z ϕl(u) := E e = e dFl(z) . (3.75) −∞
These results can be viewed as an extension of the one-factor model [CF05] study.
3.5.3 Spark Spread Valuation
We now turn to the pricing of the (exchange) spark spread option with T -terminal payoff
(1) (2) (i) FT,T − αFT,T and T ≤ T1,T2 . As usual, the forward prices are expressed as Ft,T := 1 2 + h i n o Q (i) (1) (1) (1) Et ST where St := exp gt + Xt + Jt is the two-factor jump-diffusion spot price (2) n (2) (2)o presented in the previous section and St := exp gt + Xt is the pure diffusion process of Section 3.4. We begin our analysis by rewriting the risk-neutral pricing formula 3 Energy Spot Price Models and Spread Options Pricing 35 in terms of an equivalent measure induced by the forward price process of the purely diffusive asset. In particular,
F Q (1) (2) Πt,T := P (t, T ) Et FT,T − α FT,T 1 2 + = P (t, T ) F (2) Qe [(F − α) ] , (3.76) t,T2 Et T ;T1,T2 + where the measure Qe is induced by h i ! Q (2) (2) Et ST F dQe 2 t,T2 := h i = (2) , (3.77) dQ Q S(2) F t E0 T2 0,T2 and we introduced the ratio process F := F (1) /F (2) . The process F is once t;T1,T2 t,T1 t,T2 t;T1,T2 again a Qe-martingale; however, because of the presence of the jump component, this fact alone does not allow us to extract its distribution. Instead, we make use of transform methods. [CM99] were among the first to illustrate that Fast Fourier transform (FFT) methods can be used to efficiently value European options. The reader is referred to their work for implementation details and other efficiency tricks.
The FFT methods require the m.g.f. of the logarithm of the effective stochastic process
– in our case the process Ft;T1,T2 . To this end, define ZT := ln FT ;T1,T2 so that
(1) (2) (1) (2) (1) (1) Z = A − A + e−β1(T1−T ) X − e−β2(T2−T ) X + M Y T T,T1 T,T2 T T T,T1 T (2) (2) −κ(T1−T ) −M T,T2 YT + e JT , (3.78) and define the corresponding m.g.f. process
ZT Qe u ZT Ψt (u) := Et e . (3.79)
ZT The process Ψt (u) is clearly a Qe-martingale; consequently, it satisfies the zero drift
ZT condition AΨt (u) = 0 (for every u where it is defined) where A is the generator of the 3 Energy Spot Price Models and Spread Options Pricing 36
(1) (2) (1) (2) process t, Xt ,Xt ,Yt ,Yt ,Jt under Qe. Furthermore, since our modeling framework is affine, we employ the ans¨atz
n ZT (1) (1) (2) (2) (1) (1) Ψt (u) := exp A(t, T ) + B (t, T ) Xt + B (t, T ) Xt + C (t, T ) Yt
(2) (2) o + C (t, T ) Yt + D(t, T ) Jt (3.80)
ZT ΨT (u) := exp {u ZT } (3.81)
Here, A(t, T ),B(1)(t, T ),B(2)(t, T ),C(1)(t, T ),C(2)(t, T ), and D(t, T ) are all deterministic functions of time. Note that T1 and T2 have been removed from the arguments for easier readability. Since the boundary condition (3.81) must hold for all terminal values of the
(1) (2) (1) (2) auxiliary processes Xt ,Xt ,Yt ,Yt ,Jt , the deterministic functions must satisfy the induced boundary conditions
(1) (2) A(T,T ) = u A − A ,B(1)(T,T ) = ue−β1(T1−T ), T,T1 T,T2 1 (1) (2) −β2(T2−T ) (1) (3.82) B (T,T2) = −ue ,C (T,T1) = uM T,T1 , (2) (2) −κ(T1−T ) C (T,T2) = −uM T,T2 ,D(T,T1) = ue .
Expanding the PDE Af = 0, rewriting it in terms of an equivalent system of coupled ODEs and solving that system (similar to the analysis in Section 3.5.2) provides the final result.
ZT Qe u ZT Proposition 3.5.2 The transform Ψt (u) := Et e is given by
n ZT −β1(T1−t) (1) −β2(T2−t) (2) Ψt (u) = exp At,T + u e Xt − e Xt
(1) (2) o (1) (2) −κ(T1−t) + M t,T1 Yt − M t,T2 Yt + e Jt (3.83) 3 Energy Spot Price Models and Spread Options Pricing 37
(i) where M t,T is defined in (3.12),
Z T A = u A(1) − A(2) + λ(s) ϕ ue−κ(T1−s) − 1 ds t,T T,T1 T,T2 l t −α1(T1−T ) −α2(T2−T ) +u −α1φ1γ1e h(t, T ; α1) + α2φ2γ2e h(t, T ; α2) 2 (2) −2α2(T2−T ) −β1(T1−T ) − γ2σY e h(t, T ; 2α2) + α1φ1γ1e h(t, T ; β1) 2 2 (2) (2) (2) (2) −2β2(T2−T ) − σX + 2ρ2γ2σX σY + γ2σY e h(t, T ; 2β2)
−β2(T2−T ) − α2φ2γ2e h(t, T ; β2)
(2) (1) −α1(T1−T )−β2(T2−T ) − ρ1ρ12γ1σX σY e h(t, T ; α1 + β2) 2 (2) (2) (2) −(α2+β2)(T2−T ) + 2ρ2γ2σX σY + 2 γ2σY e h(t, T ; α2 + β2) n o i (1) (2) (2) (1) −β1(T1−T )−β2(T2−T ) + ρ12σX σX + ρ1ρ12γ1σX σY e h(t, T ; β1 + β2) 1 2 +u2 γ σ(1) e−2α1(T1−T )h(t, T ; 2α ) 2 1 Y 1 1 2 + γ σ(2) e−2α2(T2−T )h(t, T ; 2α ) 2 2 Y 2 2 2 1 (1) 1 (1) (1) (1) + γ σ + σ + ρ γ σ σ e−2β1(T1−T )h(t, T ; 2β ) 2 1 Y 2 X 1 1 X Y 1 2 2 1 (2) 1 (2) (2) (2) + γ σ + σ + ρ γ σ σ e−2β2(T2−T )h(t, T ; 2β ) 2 2 Y 2 X 2 2 X Y 2 2 (1) (1) (1) −(α1+β1)(T1−T ) − γ1σY + ρ1γ1σX σY e h(t, T ; α1 + β1) 2 (2) (2) (2) −(α2+β2)(T2−T ) − ρ2γ2σX σY + γ2σY e h(t, T ; α2 + β2) i (1) (2) −β1(T1−T )−β2(T2−T ) − ρ12σX σX e h(t, T ; β1 + β2) , (3.84)
ϕl(u) is the MGF of the individual jump sizes (see (3.75)), and the function h(t, T ; ·) is given in equation (3.28).
Now that the transform is explicit, it is possible to use standard Fourier analysis techniques to value the spread option. Under some mild assumptions on the m.g.f. of jump sizes, it is possible to analytically continue the m.g.f. to the entire complex plane. 3 Energy Spot Price Models and Spread Options Pricing 38
For completeness in the exposition, we remind the reader how the pricing equation (3.76) appears in Fourier transformed variables. Firstly,
ΠF = P (t, T ) F (2) Qe [(F − α) ] t,T t,T2 Et T ;T1,T2 +
(2) = P (t, T ) eα F Qe (eZT −α − 1) , (3.85) t,T2 Et +
x where α := log(α). By introducing η(x) := (e − 1)+, the expectation in equation (3.85) reduces to the product of Fourier transforms
1 Z ∞ Qe ZT −α ˜ Et (e − 1)+ = η˜(−p) fZT −α(p) dp , (3.86) 2π −∞
˜ whereη ˜(p) and fZT −α(p) are the Fourier transforms of η(x) and the probability density of ZT − α, respectively. It is well known that
Z ∞ 1 η˜(p) := eipxη(x) dx = (3.87) −∞ p(i − p) whenever =(p) > 1. A simple change of variables reveals that
Z ∞ ˜ ipx Qe ip(ZT −α) −iαp ZT fZT −α(p) := e fZT −α(x) dx = Et e = e Ψt (ip) . (3.88) −∞
Putting these results together leads to our final pricing equation – up to a numerical integration.
Proposition 3.5.3 The price at time t of the exchange option is
Z ∞ e−iαp ΨZT (ip) dp Π = P (t, T ) eα F (2) t , (3.89) t,T t,T2 −∞ −p(p + i) 2π
ZT with Ψt (·) as in Proposition 3.5.2.
Some final remarks are crucial at this point:
1. In our framework, the price process of a spot exchange option is simply given by 3 Energy Spot Price Models and Spread Options Pricing 39
setting T = T1 = T2 in equation (3.89).
2. The integral part of equation (3.89) seems formidable; however, the coefficients are nothing more complicated than exponentials and there exists very efficient numerical methods, such as FFT, for performing the integrals. Therefore, we do not pursue this further, and instead refer the reader to the monograph by [CT04] for further information and references on these topics.
3. The market reveals the entire forward curve and, of course, the risk-free zero coupon bond prices. Before using the valuation formula, the model must be calibrated to these market prices. Once the parameters are calibrated, then the pricing equation (3.89) will provide consistent no-arbitrage prices to the various spread options.
4. It is possible to repeat this analysis when both assets contain jumps. Needless to say, the resulting equations will be bulkier (but not fundamentally more complicated), and although the change of measure will be more subtle, it posses no real problems. However, in real applications, both assets typically do not contain sudden jumps, as one is usually the raw commodity used to produce electricity.
5. Some care must be taken to ensure the integration path in (3.89) remains in the intersection of the regions =(p) > 1 and where the complex continuation of the
ZT function Ψt (z) is analytic in z. However, for typical jump distributions, such as
ZT double exponential and normal, Ψt (z) will be analytic in the region =(p) > 1, and any simple path lying in =(p) > 1 will do.
3.6 Model Calibration
In this section, we finally address the issue of parameter estimation. We perform this last step in two stages. Firstly, in Section 3.6.1 we provide a detailed review of an effi- 3 Energy Spot Price Models and Spread Options Pricing 40 cient method for calibrating the pure diffusion two-factor model to market futures prices, resulting in the risk-neutral model parameters. We also describe how jump parameters can be simultaneously estimated from market spot prices. Secondly, we describe how a method borrowed from interest rate model calibration can be used to estimate the real- world model parameters from a knowledge of spot and future prices. This simultaneous calibration of futures and spot prices to the risk-neutral and real-world measures further allows us to extract the implied market prices of risk. An alternative approach to real- world calibration is to use a well known Kalman Filter approach. Such approaches do not utilize futures prices data and can be quite useful. For more details on the calibration of various two-factor models to spot data and further references on the topic we refer to the work of [BGL04]. Section 3.6.2 concludes with concrete applications of our statistical methodology to the NYMEX Light Sweet Crude Oil data and some further comments.
3.6.1 Methodology
Before proceeding to the calibration process, recall that the log of the forward price
(i) n (i) (i)o associated with the spot St := exp gt + Xt is given by (Section 3.4.2):
(i) (i) (i) (i) (i) −βi(T −t) (i) (i) log Ft,T = gT + Gt,T + Rt,T + e Xt + M t,T Yt (3.90)
(i) (i) (i) (i) −βi(T −t) (i) (i) (i) = gT + Gt,T + Rt,T + e log St − gt + M t,T Yt (3.91)
(i) (i) (i) := U t,T + M t,T Yt . (3.92)
(i) Here, the function U t,T is introduced to simplify notation. Given the spot price data at (i) (i) (i) time t, U t,T is completely determined, while the last term M t,T Yt depends on the hidden long-run mean level Yt. Therefore, a standard nonlinear least-squares optimization cannot be applied directly. Instead, we will express the hidden factor in terms of the remaining model parameters and obtain an optimal fit to the observed futures curve at various time 3 Energy Spot Price Models and Spread Options Pricing 41 points.
(i)∗ Let F p denote the observed futures prices at tp ∈ {t1, ..., tm} with delivery time tp,Tq p p p Tq ∈ {T1 , ..., Tnp } and denote by Θ a point in the (risk-neutral) parameter space Ω of (i) our model. For each given quoted time tp, obtain Ytp (Θ) (as a function of the remaining parameters) such that it minimizes the following sum of squares:
n p 2 X h (i) (i)∗ i Sum(tp, Θ) := log F p − log F p . (3.93) tp,Tq tp,Tq q=1
(i) The optimal Ytp (Θ) is easily found to be
h (i) (i) i Pnp (i)∗ M p log F p − U p #(i) q=1 tp,Tq tp,Tq tp,Tq Yt (Θ) = . (3.94) p h (i) i2 Pnp M p q=1 tp,Tq
Substituting this optimal value into the initial sum of squares (3.93), summing over the range of initial times {tp} and performing a nonlinear least-squares optimization as follows:
m n m 2 ∗ X X h (i) (i) #(i) (i)∗ i Θ := ArgMin U p + M p · Y (Θ) − log F p , (3.95) Θ∈Ω tp,Tq tp,Tq tp tp,Tq p=1 q=1 provides an “optimal fit” of the model to futures prices, therefore obtaining our risk- neutral model parameters (β, α, φ, σX , σY , ρ). An implementation of this methodology naturally requires both futures prices and spot prices at the corresponding futures quote times.
It is worth mentioning that this method does not directly extend to jump-diffusion
(spot) models since the coefficients of the Xt and Jt terms in the forward price (3.73) are unequal. This prevents a simple factorization into functions that are known given
(1) the spot prices and the hidden process Yt . To circumvent that problem a standard alternative methodology is to extract the jump parameters from the spot price data only. Such a calibration can be carried out in two ways: (i) by cutting off all data points 3 Energy Spot Price Models and Spread Options Pricing 42 lying below a given level, so that only spikes remain. From these data points one can then infer the value of the various jump parameters (see for example the discussion in [CS00]); or (ii) by utilizing particle-filter approaches which generalizes the Kalman filter to non-normal innovations (see for example [ABT05]). The standard assumption that the calibrated jump parameters are unchanged when moving to the risk-neutral world is then invoked. Given, the jump parameters it is now possible to repeat the previous futures price calibration process to obtain the risk-neutral diffusive components.
We now turn to the real-world P-parameters (β, α, φ, σX , σY , ρ) estimation problem. Since under any diffusive model for spot prices, a change of measure from the real-world to risk-neutral cannot alter the volatility structure of the model, from equation (3.95) we obtain σX , σY and ρ under P. The remaining set of P-parameters (β, α, φ) are relatively straightforward to obtain. Firstly, we obtain β and φ via linear regression on the spot price data assuming a mean reverting one-factor model for Xt as a proxy to our two-factor model. The one-factor mean-reversion level φ becomes, in our model, the stochastic long- run mean level Yt. Secondly, we perform a similar regression on the estimated hidden
# process Yt which was obtained by minimizing the error on an individual futures curve basis (see equation (3.94)). Equation (3.94) provides a data set which we can use as an input in a regression to find α. We find this procedure to be very stable and, as shown in the next section, leads to reliable parameter estimation.
3.6.2 Some Results: Crude Oil
α β φ α β φ σX σY ρ 0.15 0.31 3.27 0.73 1.07 4.21 33% 63% -0.96
Table 3.1: The calibrated real-world and risk-neutral model parameters using the NYMEX Light Sweet Crude Oil spot and futures data for the period 1/10/2003 − 25/07/2006. 3 Energy Spot Price Models and Spread Options Pricing 43
In this section, we present the calibration results of our two-factor pure diffusion model (Section 3.4) to the NYMEX Light Sweet Crude Oil spot and futures data for the period 1/10/2003 to 25/07/2006. In Table 3.1, we report the calibration results for the real-world and risk-neutral parameters. There are a few notable observations: (i) both real-world mean-reversion rates α and β are significantly larger than the risk-neutral mean-reversion rates α and β, (ii) The real-world long-run mean-reversion level φ is larger than the risk-neutral long-run mean φ, and (iii) in both the real-world and risk-neutral measures, the mean-reversion rates (α and α) of the long-run mean Yt are smaller than the mean- reversion rates (β and β) of the log-spot Xt.
$100
$80
$60 Price $40 Market Spot
$20 Simulation Long Run Mean $0 08/17/03 03/04/04 09/20/04 04/08/05 10/25/05 05/13/06 Date
Figure 3.1: The NYMEX Light Sweet Crude Oil spot prices and simulated spot prices based on the calibration in Table 3.1.
In Figure 3.1, we plot the spot price data together with the stochastic long-run mean
# level Yt implied by the futures prices. For comparison, we also include one simulated sample path based on a simulation of the prices using the real-world model parameters in Table 3.1. Figure 3.2 illustrates the relative root-mean squared-error (RMSE) for each forward curve using the model parameters reported in Table 3.1. The average RMSE per 3 Energy Spot Price Models and Spread Options Pricing 44 curve is 0.7% with only a few dates having relative errors larger than 1%. Recall that the model parameters are fixed over all curves, and are not adjusted on a curve by curve basis. With this in mind, we believe the fit is excellent.
1.8% 1.6% 1.4% 1.2% 1.0% 0.8% 0.6% Relative RMSE 0.4% 0.2% 0.0% 08/17/03 03/04/04 09/20/04 04/08/05 10/25/05 05/13/06 Date
Figure 3.2: The relative root-mean squared-error of each forward curve based on the calibration in Table 3.1.
# Days β α φ σX σY ρ 88 0.38 0.26 3.34 33% 19% -0.97 176 0.52 0.21 3.06 33% 54% -0.79 264 0.62 0.10 2.36 33% 56% -0.73 352 0.61 0.08 1.97 35% 60% -0.64 440 0.52 0.09 2.33 35% 58% -0.71 528 0.43 0.10 2.98 34% 52% -0.95 616 0.34 0.13 3.24 34% 58% -0.96 704 0.31 0.15 3.27 33% 63% -0.96 Average: 0.48 0.12 2.74 34% 57% -0.82 Stdev: 0.13 0.05 0.52 1% 4% 0.14
Table 3.2: This table shows the evolution of the estimated risk-neutral parameters through time as more recent data is added to the calibration procedure. The average and standard deviation are reported using 176 days onwards. 3 Energy Spot Price Models and Spread Options Pricing 45
3
1
-1
-3
-5
-7 Market Price of Risk -9 λ ψ
-11 8/17/03 3/4/04 9/20/04 4/8/05 10/25/05 5/13/06 Date
Figure 3.3: This diagram depicts the evolution of the implied market prices of risk using the calibrated real-world and risk-neutral parameters.
We also investigated the stability of our estimation procedure through time. We calibrated the model to the first T1 calender days and then to the first T2 calender days and so on. The time periods are approximately equally spaced at 88 days from 1/10/2003 to 25/07/2006. We report these calibration results in Table 3.2. The most stable parameters are the volatility σX of the Xt process, the volatility σY of the stochastic long-run mean level Yt, the mean-reversion level α of the stochastic long-run mean Yt and the correlation coefficient ρ. The remaining parameters, although not as unvarying as the previous four, are well behaved. None of the parameters suddenly explode or tend to zero, and always remain realistic.
Finally, since we were successful in extracting the real-world and risk-neutral param- eters, we further extract the implied market prices of risk through Theorem 3.4.2. The evolution of the implied market prices of risk λ and Ψ are displayed in Figure 3.3. Interest- ingly, they are very strongly correlated to one another, becoming almost indistinguishable after one and a half years. This is due to the high correlation coefficient of ρ = −0.96. 3 Energy Spot Price Models and Spread Options Pricing 46
Also, both market prices of risk are negative for essentially the entire time period. This is a reflection of the real-world mean-reversion rates (α and β) and real-world long run mean-reversion level (φ) being higher than the risk-neutral ones ((α, β and φ)). The mar- ket therefore attaches slower reversion rates and lower long run levels than the implied historical levels.
3.7 Conclusions
We introduced a diffusive two-factor mean-reverting process for modeling spot prices of energy commodities. The two-factor diffusive model extends the one-factor mean- reverting model by making the long-run mean a stochastic degree of freedom which itself mean-reverts to a specified level. We also generalized the model to incorporate jumps in the price process such as those observed in electricity prices. To maintain realistic mean- reversion rates and diffusive volatilities we decoupled the jump and diffusive processes. Given our affine modeling framework, we were successful in obtaining expressions for the forward price curves in terms of elementary functions. Through a measure changed induced by the forward price process, our modeling framework allows us to obtain closed form pricing equations for various spread options. We obtained pricing equations under both an actuarial and risk-neutral valuation procedures.
Finally, we provided a method for calibrating both the diffusion and jump-diffusion models to spot and forward prices simultaneously. This simultaneous calibration proce- dure further allowed us to extract the implied market prices of risk. Using the NYMEX light sweet crude oil data set, we demonstrated that the calibration procedure produces realistic and stable implied risk-neutral and real-world model parameters. 3 Energy Spot Price Models and Spread Options Pricing 47
3.8 Acknowledgements
The authors would like to thank Bill Bobey for assistance with acquiring the NYMEX data and Hans Tuenter, OPG Energy Markets, for fruitful discussions on various aspects of the energy markets. Chapter 4
Second Paper: Asymptotic Pricing of Commodity Derivatives using Stochastic Volatility Spot Models
4.1 Abstract
It is well known that stochastic volatility is an essential feature of commodity spot prices. By using methods of singular perturbation theory, we obtain approximate but explicit closed form pricing equations for forward contracts and options on single- and two-name forward prices. The expansion methodology is based on a fast mean-reverting stochastic volatility driving factor and leads to pricing results in terms of constant volatility prices, their Delta’s and their Delta-Gamma’s. Both the standard single factor mean-reverting spot model and a two-factor generalization, in which the long-run mean is itself mean- reverting, are extended to include stochastic volatility and each is analyzed in detail. The stochastic volatility corrections can be used to efficiently calibrate option prices and compute sensitivities.
48 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 49
4.2 Introduction
A quick glance at any commodities price data will reveal the obvious fact that volatility is a stochastic quantity. A now classical and extremely popular model for incorporating this stochasticity of volatility is the [Hes93] model, in which the instantaneous price variance follows a [CIR85] (CIR) like process. [EG98] were among the first to utilize the Heston model in the context of energy derivatives. More recently, [RS06] introduce a stochastic convenience yield model with one underlying stochastic volatility factor in the same spirit of Heston. They make an extensive case study on soybean futures and options data and demonstrate that stochastic volatility is a significant factor. Since Heston inspired stochastic volatility models lead to affine structures, they appear natural; however, the resulting pricing equations are in terms of inverse Fourier transforms rather than explicitly in terms of elementary, or even special, functions. This is not a substantial disadvantage when valuing only a few options; however, in a calibration and trading environment many contracts are involved and consistently calibrating all instruments to market prices would be difficult and time consuming. Furthermore, determining hedge ratios will require computations of the sensitivities of the price to various parameters – the so-called “Greeks” – which, if computed using Fourier methods, may result in further speed reduction. Finally, none of the previous models are tractable enough to provide pricing results, or even approximations, for options on two forward contracts. To circumvent all of these issues, we transport singular perturbation theory techniques, first developed for equity derivatives by [FPS00b] and then for interest rate derivatives in [CFPS04], into the context of commodities and commodities derivatives.
Asymptotic methods have three main advantages over traditional approaches: (i) they naturally lead to efficient calibration across a set of forward contracts; (ii) they lead to approximate, but explicit, closed form pricing equations for a wide class of contingent claims; and (iii) the resulting approximate prices are independent of the specific underlying 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 50
50%
40%
30%
20%
10%
0% 10/07/03 04/24/04 11/10/04 05/29/05 12/15/05 07/03/06 Dates Figure 4.1: The annualized running five-day moving volatility of the NYMEX sweet crude oil spot price for the period 10/07/03 to 07/03/06.
volatility model. Notably, these prices are exact when the mean-reversion rate is large – serendipitously, this is precisely the manner in which the market prices seem to behave. In addition, many of the salient features of option prices – most strikingly the implied volatility smile or smirk – are captured by these methods. [FPS00b] were the first to introduce the use of asymptotic methods in the context of derivative pricing and, together with their collaborators, have written several articles on the application of these techniques to the equity and interest rate markets. To this date, none of these techniques have been applied to the commodities markets where a unique set of challenges arise.
To motivate the validity of asymptotic methods for commodities, we plot the run- ning five-day realized volatility for the NYMEX sweet crude oil spot price for the period 10/07/03 to 07/03/06 in Figure 4.1 which clearly demonstrates the fast mean-reversion of volatility. We therefore model the underlying commodity spot price volatility as a function
σX (Zt) of a fast-mean reverting hidden process Zt. As is well known, commodities, unlike equities, tend to have strong mean-reversion effects in the prices themselves. Secondly, the long-run mean-reversion is not constant through time, rather it is stochastic. These and many other stylized empirical facts are well documented in, for example, [CS00], [EW03] and [Gem05]. Correctly accounting for such behavior together with stochastic volatility and using such models to price derivatives on one and two forward contracts is the main 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 51 contribution of this article.
[HJ07] introduced tractable two-factor mean-reverting models (with and without jumps) and priced forward and spread options on forward contracts. In this article, we successfully determine the asymptotic corrections for forward prices based on stochastic volatility extensions of the one- and two-factor mean-reverting diffusive spot price mod- els. To this end, we quickly review the one- and two-factor spot price models, together with the resulting forward and option prices, in Section 4.3.1. The stochastic volatility extended one- and two-factor mean-reverting models are introduced in Section 4.3.2 and we illustrate that such a model does not provide closed form forward prices. Section 4.4 contains two of our main asymptotic expansion results: the forward prices for the stochas- tic volatility extended one- and two-factor mean-reverting models are shown to be well approximated by adjusted constant volatility results (see equation (4.39) and (4.45)). By calibrating to existing forward prices, the volatility function σX (z) is rendered irrelevant; instead, a new effective pseudo-parameter arises as a smoothed version of the stochas- tic volatility. This pseudo-parameter appears again in the pricing of contingent claims, allowing a consistent calibration between forward and options prices.
It is important to point out that the one-factor commodity model is closely related the IR model examined in [FPS00a]; however, the relevant pricing object is the futures price rather than bond prices. This subtle modification induces a distinct non-vanishing boundary condition into the pricing problem and later on induces modified boundary con- ditions into the option pricing problem. This has the effect of introducing non-commuting operators into the resulting pricing PDEs which we solve using commutator relations (see the solution to equations (4.55) and (4.82)). In addition to the modified boundary condi- tion, we extend the one-factor model to a more realistic two-factor model for commodities. The two-factor model requires its own delicate balancing and analysis to ensure that the asymptotic expansions remain valid. 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 52
Given that the model is calibrated to forward prices, the next task is to determine the price corrections to contingent claims. Since typical single-name contingent claims are written on the forward prices, which we have already approximated, the asymptotic analysis relies on a consistent layering of approximations. In Section 4.5, these asymptotic price corrections to single-name contingent claims are explored. Interestingly, we demon- strate that the corrections depend solely on the Delta’s and Delta-Gamma’s of the option using the constant volatility model (see equations (4.57) and (4.61)). Furthermore, once the free pseudo-parameter arising in the forward price approximation is calibrated to mar- ket prices, the option price corrections are uniquely determined. Section 4.6 contains the extension of these methods to contingent claims written on two forward prices. There are several subtle issues associated with the expansion; nonetheless, we pleasantly find that the resulting price corrections are once again in terms of the Delta’s and Delta-Gamma’s of the constant volatility price (see equations (4.84), (4.85) and (4.99), (4.100)).
Although the techniques employed are quite similar to the single-name option case, the generalization of singular perturbation techniques to multi-name options is quite novel. The analysis requires balancing two fast mean-reverting processes which drive coupled spot price models. The implied forward processes then both factor into the option payoff, requiring a layering of the expansion parameters. It is important to prove that the cor- rection terms are indeed small enough to ignore. Furthermore, due to the non-vanishing boundary condition in the one and two-factor forward prices, commutation relations are once again necessary to solve the resulting PDEs.
We close the paper with conclusions and some comments on ongoing and future work in Section 4.7. 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 53
4.3 Spot Price Models and Main Properties
This section first provides an overview of the standard one- and two-factor constant volatil- ity models for commodity spot price dynamics. For early uses of the one-factor models see [GS90] and [CS94]. The forward prices, call and exchange option prices are also re- viewed. Given these constant volatility models, the stochastic volatility (SV) extensions are then introduced and we briefly demonstrate that the SV extensions lack an affine structure.
We explain why and where asymptotic methods constitute a very useful set of tools in energy markets, as they already have been shown to be for their stocks and interest rate counterparts.
4.3.1 Constant Volatility Models
The One-Factor Model
For completeness, this section provides a quick review of a well known one-factor commod- ity spot price model and its use in derivatives pricing. Let St denote the spot dynamics defined under the risk-neutral measure Q. The standard model assumes
St := exp {gt + Xt} , (4.1)
(1) dXt = β (φ − Xt) dt + σX dWt , (4.2)
(1) where σX is the constant volatility, gt is a deterministic seasonality factor and W is a Q- Wiener process. An important traded commodity instrument is the futures contract with futures price Ft,T . In a no-arbitrage, deterministic interest rate, environment the futures
Q and forward price coincides and the forward price must be given by Ft,T := Et [ST ], where
Q Et [R] represents the expectation of R conditional on the natural filtration Ft generated by the underlying Wiener process(es). The forward price process being a martingale, must 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 54 satisfy the following PDE
AF (t, x) = 0 , (4.3) F (T, x) = egT +x ,
where A is the infinitesimal generator of (t, Xt). Within the present context, a straight- forward calculation provides the following result
σ2 F = exp g + φ 1 − e−β(T −t) + X h(t, T ; 2β) + e−β(T −t) (log(S ) − g ) (4.4) t,T T 2 t t
Here, and in the sequel,
h(t, T ; a) := (1 − e−a(T −t))/a . (4.5)
Turning to the valuation of European contingent claims, let ϕ(FT0,T ) denote the termi- nal payoff at time T0 of a European option written on a forward price. The no-arbitrage price Πt,T0 is the discounted expectation under the risk-neutral measure Q. Specifically,
h R T0 i Q − t rs ds Q Πt,T0 := Et e ϕ (FT0,T ) = P (t, T0) Et [ϕ (FT0,T )] . (4.6)
Here, and in the remainder of this article, interest rates are deterministic, and we denote the T0-maturity zero-coupon bond price contracted at time t by P (t, T0). Following the
martingale techniques employed in [HJ07] Section 3.4, the price Ct;T0,T at time t of a
T0-expiry call option with strike K written on the forward FT0,T can be expressed in the following Black-Scholes like form:
h R T0 i Q − t rs ds Ct;T0,T = Et e (FT0,T − K)+
∗ ∗ ∗ = P (t, T0) Ft,T Φ(d + σt;T0 ) − K Φ(d ) . (4.7)
∗ ∗ Here, d and σt;T0 are functions of the model parameters and time only, and Φ(·) is the 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 55 standard gaussian cdf. A similar result follows for forward exchange option prices:
R T0 (1) (2) F Q − t rs ds Πt;T ,T ,T = Et e FT ,T − αFT ,T 0 1 2 0 1 0 2 + h i = P (t, T ) F (1) Φ(d + σ ) − αF (2) Φ(d) . (4.8) 0 t,T1 t;T0 t,T2
The interested reader is referred to the original article for the precise form of the various coefficients.
The Two-Factor Model : Mean-Reverting Long Run Mean
[HJ07] utilize a two-factor mean-reverting model, in which the long-run mean of the previous one-factor model is itself stochastic and mean-reverts to a second long-run mean. In that work, the authors study the valuation of forward contracts and exchange options and also include jumps into the spot price dynamics. In this article, we focus on the jump-free model; however, much of the results can be extended to the jump case with little additional complication.
In this two-factor model, the Q-dynamics of the spot St is
St = exp {gt + Xt} , (4.9)
(1) dXt = β (Yt − Xt) dt + σX dWt , (4.10)
(2) dYt = α (φ − Yt) dt + σY dWt , (4.11) with correlation structure,
(1) (2) d W ,W t = ρ1 dt . (4.12)
Here, β controls the speed of mean-reversion of Xt to the stochastic long-run level Yt; α controls the speed of mean-reversion of the long-run level Yt to the target long-run mean
φ; σX and σY control the size of the fluctuations around these means. The forward price 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 56 process can be shown to be
−β(T −t) Ft,T = exp gT + Rt,T + Gt,T + e Xt + Mt,T Yt (4.13)
where the expressions for Mt,T , Gt,T and Rt,T are functions of time and the model param-
eters. Even within this more general setting, the call option price Ct;T0,T on a forward as
F well as the exchange option price Πt;T0,T1,T2 on forwards have similar forms to (4.7) and ∗ ∗ (4.8) respectively. More complicated expression for d , σt;T0,T , d and σt;T0,T arise, yet they remain explicit functions only of the model parameters and time. The interested reader is once again referred to [HJ07] for details.
4.3.2 Stochastic Volatility Extensions
The SV Extended One-Factor Model
In this section, the stochastic volatility (SV) extended one-factor model is explored in detail; in particular, the volatility σX is now assumed to be driven by a fast mean- reverting stochastic process. Explicitly, the spot is now modeled under the risk-neutral measure Q as
St = exp {gt + Xt} , (4.14)
(1) dXt = β (φ − Xt) dt + σX (Zt) dWt , (4.15)
(3) dZt = α (m − Zt) dt + σZ dWt , (4.16)
where σX (·) is a strictly positive smooth function bounded above and below by posi- tive constants and with bounded derivatives. We also specify the following correlation structure
(1) (3) d W ,W t = ρ2 dt . (4.17) 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 57
The smoothness and boundedness assumptions on the volatility function σX (·) may ap- pear overly restrictive at first; however, as we later demonstrate, singular perturbation methods remarkably lead to pricing results that are completely independent of its detailed specification.
It is not possible to solve the system of SDEs (4.14)-(4.16) explicitly; nonetheless, we now explore its implications for forward prices. As usual, the forward price is F (t, x, z) =
Q Et,x,z [ST ]. Equivalently, F (t, x, z) can be characterized as the solution of the following PDE: ∂F ∂F ∂F ∂t + β(φ − x) ∂x + α(m − z) ∂z + 1 2 ∂2F 1 2 ∂2F ∂2F (4.18) 2 σX (z) ∂x2 + 2 σZ ∂z2 + ρ2σZ σX (z) ∂x∂z = 0 F (T, x, z) = egT +x
As we now show, a solution to (4.18) can be decomposed into two independent parts; one having a log-affine structure in x and the other being independent of x. First, let Wt be
(1) (3) a Q-Wiener process independent of Wt ,Wt and define the following
−β(T −t) dZet := α(m − Zet) + ρ2σZ σX (Zet)e dt + dWt , (4.19) 1 c(t, z) := σ2 (z)e−2β(T −t) + βφe−β(T −t) , (4.20) 2 X Z T Q M(t, z) := Et,z exp c(s, Zes)ds . (4.21) t
Then, by smoothness and boundedness of c(·, ·) and of the coefficients of dZet, M(t, z) is finite and satisfies the following PDE (see [DPS00])
∂M + α(m − z) + ρ σ σ (z)e−β(T −t) ∂M + 1 σ2 ∂2M + c(t, z) M = 0 ∂t 2 Z X ∂z 2 Z ∂z2 (4.22) M(T, z) = 1 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 58
−β(T −t) By direct, tedious, computations exp gT + e x M(t, z) is seen to satisfy the PDE
−β(T −t) (4.18); consequently, the forward price F (t, x, z) = exp gT + e x M(t, z).
Given the form of M(t, z), the forward prices clearly do not share the natural affine structure that other models often possess (e.g., compare with the constant volatility two- factor model (4.13)). It is also doubtful that an explicit (closed form) solution of the PDE (4.22) exists. Hence, this model appears to suffer from the deficiencies of Heston-like models which require either solving a PDE numerically or resorting to Fourier methods, rendering the models less useful for calibration purposes. Surprisingly, it is possible to partially overcome these difficulties if we accept to limit the range of applicability of our SV model to commodities having fast mean-reverting volatility (α 1). This is indeed the approach we pursue in the rest of this work.
The SV Extended Two-Factor Model
In this section, the stochastic volatility (SV) extended two-factor model is recorded for completeness. Starting with the two-factor model of Section 4.3.1, we make the volatility
σX a function of a fast mean-reverting stochastic process – analogous to the SV extended one-factor model. The spot is now modeled under a Q-measure as
(1) dXt = β (Yt − Xt) dt + σX (Zt) dWt , (4.23)
(2) dYt = αY (φ − Yt) dt + σY dWt , (4.24)
(3) dZt = α (m − Zt) dt + σZ dWt , (4.25) with correlation structure,
(1) (2) (1) (3) (2) (3) d W ,W t = ρ1 dt , d W ,W t = ρ2 dt , d W ,W t = 0 , (4.26)
and restrictions on σX (·) parallel to the previous section. 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 59
Rather than repeating the analysis of the previous subsection, we instead point out that resulting forward prices are not of the affine form. Nevertheless, asymptotic methods will lead to approximate, but explicit, closed form forward and option prices.
4.4 Forward Price Approximation
It is well known that the invariant distribution of the volatility driving factor Zt is Gaus-
2 2 sian with a variance of ν := σZ /2α. The asymptotic expansion revolves around assuming that α 1 and simultaneously holding the variance ν2 of the invariant distribution finite and fixed. As such, our developments are primarily parameterized by the small parameter := α−1. The ultimate goal of this section is to obtain a sound approximation (in a sense to be defined shortly) to the forward price, and in tandem eliminate the dependency of the approximate forward curve on the non-observable Zt.
Such closed form forward price approximations will allow efficient statistical estimation of the model parameters, and lead to tractable pricing of derivatives written on these forward curves. We use the methodology originally applied in [FPS00b] and [CFPS04] for stock and IR options respectively. For detailed discussions on the fundamentals of these asymptotic techniques we refer to the monograph [FPS00a].
Although the work in this section follows the methodology in [FPS00a], there are sev- eral distinctions. Firstly, and most simply, the model is motivated by commodities market behavior. Secondly, the relevant pricing objects are forward prices and not bond prices. This induces distinct boundary conditions which later on play an important distinguishing role and will require commutator relations to solve the resulting pricing PDEs. Finally, we extend the techniques to the more realistic two-factor model. 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 60
4.4.1 One-Factor Model + SV
In this section, we assume that the spot price dynamics is driven by the SV extended one-factor model in section 4.3.2. Recall that
Q F (t, x, z) := Et,x,z [ST ] , (4.27) where the dependence on (:= α−1) is made explicit. Rewriting the PDE (4.18) as
−1 − 1 A F = A + 2 A + A F (t, x, z) = 0 , 0 1 2 (4.28) F (T, x, z) = egT +x , with the three new operators defined as
∂ ∂2 A := (m − z) + ν2 , (4.29) 0 ∂z ∂z2 √ ∂2 A := 2ρ νσ (z) , (4.30) 1 2 X ∂x∂z ∂ ∂ 1 ∂2 A := + β(φ − x) + σ2 (z) , (4.31) 2 ∂t ∂x 2 X ∂x2
highlights the various scales of the individual operators. Note that A0 is the infinitesimal generator of a simple Vasicek (OU) process; A2 is the infinitesimal generator of the process
(t, Xt); while the A1 operator accounts for the correlation between the log spot price Xt and the volatility driver Zt processes. √ Expanding F in powers of
(0) √ (1) (2) 3 (3) F = F + F + F + 2 F + ... (4.32) where we impose the boundary conditions F (0)(T, x, z) := F (T, x, z) := egT +x and F (1)(T, x, z) := 0. We have explicitly assumed that the zeroth order term matches the payoff at maturity, while the first correction term vanishes at maturity. This terminal 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 61 splitting is not required, however it is natural, leading to explicit closed form approxima- tions, and allowing us to prove that the remaining corrections terms are O(). Inserting this last expansion into the PDE (4.28) and collecting terms with like powers √ of gives
1 1 0 = A F (0) + √ A F (0) + A F (1) + A F (0) + A F (1) + A F (2) 0 1 0 2 1 0 √ (1) (2) (3) + A2F + A1F + A0F + ... . (4.33)
√ From this last equation, the coefficients of the various powers of must vanish individu- ally. In the subsequent analysis we investigate these resulting equations and deduce from them the main properties of F (i)(t, x, z) for i = 0, 1, 2 and 3 explicitly.
−1 (0) • –Order Equation : A0F = 0 This holds for all z; therefore F (0) must be independent of z: F (0) = F (0)(t, x).
− 1 (0) (1) • 2 –Order Equation : A1F + A0F = 0
(0) (1) (1) Since F is independent of z, this implies A0F = 0. This further implies F is also independent of z; that is, F (1) = F (1)(t, x).
0 (0) (1) (2) • –Order Equation : A2F + A1F + A0F = 0
(1) (0) (2) Since F is independent of z, this implies the Poisson equation A2F +A0F = 0
(0) and the resulting centering equation hA2F i = 0 is a necessary condition for a solu- tion to exist. Here, and in the remainder of the article, the bracket notation hf(z)i
denotes the expectation of f(Z) where Z ∼ N(m, ν2), the invariant distribution of
(0) the Q-process Zt, as defined in (4.16). Since F is independent of z, the center-
(0) ing equation becomes hA2iF = 0. Remarkably, this is the PDE (4.3) satisfied by the forward price based on the one-factor spot model with constant volatility
p 2 (0) σX := hσX (z)i. Enforcing the boundary condition F (T, x) = exp (gT + x),
(0) implies that F is the one-factor forward price (4.4) with constant volatility σX . 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 62
Up to this order, it is also possible to extract properties of F (2) which will prove
(0) useful in the subsequent analysis. Due to the centering equation hA2iF = 0, notice that
1 A F (0) = (A − hA i) F (0) = σ2 (z) − hσ2 i F (0) , (4.34) 2 2 2 2 X X xx
(0) (2) which allows the zero-order equation A2F + A0F = 0 to be rewritten as
1 1 F (2) = − A−1 σ2 (z) − hσ2 i F (0) = − (ψ(z) + c(t, x)) F (0) , (4.35) 2 0 X X xx 2 xx
where the function ψ is define as the solution of
2 2 A0ψ = σX − hσX i , (4.36)
and c(t, x) is an arbitrary constant of integration. A straightforward calculation also shows that
Z z 0 1 2 2 2 ψ := ∂zψ = 2 2 σX (u) − hσX i Φ(u; m, ν ) du , (4.37) ν Φ(z; m, ν ) −∞
2 2 where Φ(·; m, ν ) is the cdf of N(m, ν ), the invariant distribution of Zt.
1 (1) (2) (3) • 2 –Order Equation : A2F + A1F + A0F = 0 This is a second Poisson equation, but now for F (3). Its centering equation is
(1) (2) (1) (2) hA2F + A1F i = hA2iF + hA1F i = 0 which is easily shown to transform
1 (0) √ 1 (1) − 2 0 (1) (1) 2 0 into hA2iF = 2 ρ2νhσX ψ iFxxx. Define Fe := F and V := ( 2 ) ρ2νhσX ψ i, 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 63
the centering equation is then
hA iF (1)(t, x) = VF (0) , 2 e xxx (4.38) Fe(1)(T, x) = 0 .
Equation (4.38) is the zero boundary version of the usual one-factor forward price √ PDE (4.3) with constant volatility σX and an additional source term of order . Using the previous result that F (0) has the form of the one-factor forward price (4.4),
direct computations show that Fe(1) = −V h(t, T ; 3β) F (0) is a solution to equation (4.38).
Piecing together all of the above partial results, the price approximation based on the first two terms of the expansion (4.32) is succinctly written as
F (t, x, z) ' F (0)(t, x) + Fe(1)(t, x) := ( 1 − V h(t, T ; 3β)) F (0)(t, x) . (4.39)
Intriguingly, the right hand side of (4.39) is independent of the unobservable Zt process. This is an extremely convenient consequence of asymptotic derivative valuation results. It is also worth noting that for calibration purposes, the constant V can, and should, be used as a parameter in its own right. All of the details of the mapping from Zt to the volatility process (σX (Zt)) is averaged out and embedded in the constant V . Rather than specifying the “micro-structure” in the model, it is perfectly valid to specify the “macro-structure” in V as implied from futures price data.
We now state one of our main results on the validity of the approximation (4.39).
2 Theorem. 4.4.1 For any fixed (T, x, z) ∈ R+ × R and all t ∈ [0,T ], we have
(0) (1) F (t, x, z) − F (t, x) + Fe (t, x) = O() , 4 Asymptotic Pricing of Commodity Derivatives using SV Spot Models 64 where the approximation F (0)(t, x) + Fe(1)(t, x) is defined in (4.39) and F (0)(t, x) as in
p 2 (4.4) with σX replaced by hσX (z)i.
Proof. Define the function Υ(t, x, z) as the error terms of order 2 and higher. Explicitly,
(0) √ (1) (2) 3 (3) Υ := F + F + F + 2 F − F . (4.40)
We first aim at proving that |Υ| = O(). Applying the infinitesimal generator A of
(t, Xt,Zt) on Υ and cancelling vanishing terms, based on our previous analysis of the F (i) functions, we find
−1 − 1 (0) √ (1) (2) 3 (3) A Υ = A0 + 2 A1 + A2 F + F + F + 2 F − F