Carnegie Mellon University

The David A. Tepper School of Business

Stochastic Behavior of Spot and Futures Commodity Prices: Theory and Evidence

A Thesis in
Financial Economics
by
Jaime Casassus

Dissertation Committee:
Professor Pierre Collin-Dufresne (Chair)
Professor Bryan Routledge
Professor Chester Spatt
Professor Stanley Zin
Professor Daniele Coen-Pirani (External Reader)

Submitted in Partial Fulfillment

of the Requirements
for the Degree of
Doctor of Philosophy

April 2004
 Copyright 2004
c

Jaime Casassus

All Rights Reserved

To Francisca, Pascale, Nico and Max.
Abstract

This dissertation examines the stochastic behavior of commodity prices from two alternative ap-

proaches. The first chapter develops a three-factor Gaussian model of commodity spot prices, con-

venience yields and interest rates. Our model allows convenience yields to be a function of the

spot price and interest rates. Also, it allows for time-varying risk-premia. For crude oil and cop-

per we find strong evidence for spot-price level dependence in convenience yields, which implies

mean-reversion in spot prices under the risk-neutral measure, and is consistent with the “theory of

storage.” For silver, gold and copper we find evidence for time-varying risk-premia, which implies

mean-reversion of commodity prices under the physical measure albeit with different strength and

long-term mean. The model thus disentangles the different sources of mean-reversion in spot com-

modity prices. The spot-price level dependence in convenience yields has a substantial impact for

option prices, while the time-varying risk-premia affect risk management decisions. The second

chapter examines commodity prices in a structural framework. In our model production of the con-

sumption good requires two inputs: the consumption good and a oil. Oil is produced by wells whose

flow rate is costly to adjust. Investment in new Oil wells is costly and irreversible. As a result in

equilibrium, investment in Oil wells is infrequent and lumpy. The resulting equilibrium oil price

exhibits mean-reversion and heteroscedasticity. The spot oil price is not Markov (in itself) and is

best described as a regime-switching process. The futures curve exhibits backwardation as a result

of a convenience yield, which arises endogenously due to the productive value of oil as an input for

production. The model is capable of matching the first two moments of the futures curves and the

average consumption of oil-output and output-consumption ratios from macroeconomic data.

i
Acknowledgements

I would like to express all my gratitude to my advisor Professor Pierre Collin-Dufresne for his

guidance and encouragement. Pierre has been a super advisor. I thank him for all those challenging

and inspiring discussions we had, for his contribution to this dissertation and for cheering me up

when I was intellectually frustrated.

I am thankful to the Finance faculty at Carnegie Mellon University and in particular to Profes-

sors Bryan Routledge, Rick Green, Chris Telmer, Chester Spatt, and Stanley Zin for their valuable

suggestions.

I would like to thank my finance friends and colleagues Stephan Dieckmann, Ozge Gokbayrak,

Iulian Obreja, Anastasiya Ostrovnaya, Bernhard Paasche, Cheryl Telmer and Chris Telmer for their

support, energy, friendship and, of course, the beers we had together. Thanks also to the members

of the 312 Project and the T2 Trailer for years of inspiration and endless generation of new ideas.

The financial support from the Pontificia Universidad Catolica de Chile, The Chilean Govern-

ment, Carnegie Mellon University and The Carnegie-Bosch Institute is also gratefully acknowl-

edged.

Last but not least, I would like to thanks my family. There are no words to express my gratitude

to Francisca and our kids Nico, Pascale and Max. Thanks Fran for unconditional support since

the very first minute I decided to become a Finance Professor many years ago. Thanks to my kids

for showing me that life is much more than numbers and papers. Thanks family for your love,

encouragement, optimism, youth, patience and creating the warm environment that enabled me to

finish this dissertation with success.

ii
Table of Contents

1 ‘Maximal’ Model of Commodity Futures with Stochastic Convenience Yield and Inter-
est Rates 2
1.1 The ‘Maximal’ Convenience Yield Model . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Specification of Risk Premia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Empirical implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.1 Description of the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.2 Empirical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3.4 Estimation of the Jump component in commodity spot prices . . . . . . . . 24
1.3.5 Restrictions on common pricing kernel dynamics . . . . . . . . . . . . . . 26
1.3.6 Summary of the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4 Implication of mean-reversion for option pricing and value at risk. . . . . . . . . . 30
1.4.1 Option Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.4.2 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.6 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.7 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Appendix 1.A Closed-form solution for futures prices . . . . . . . . . . . . . . . . . . 51
Appendix 1.B Closed-form solution for zero-coupon bonds . . . . . . . . . . . . . . . 52
Appendix 1.C The {r, δ, X} representation . . . . . . . . . . . . . . . . . . . . . . . . . 52
Appendix 1.D The {r, bδ, X} representation . . . . . . . . . . . . . . . . . . . . . . . . . 54
Appendix 1.E MLE of the Maximal Model . . . . . . . . . . . . . . . . . . . . . . . . 56
Appendix 1.F MLE of the Triple-Jump Model . . . . . . . . . . . . . . . . . . . . . . 58
Appendix 1.G Closed-form solution for futures prices with jumps . . . . . . . . . . . . 59
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2 Equilibrium Commodity Prices with Irreversible Investment and Non-Linear Tech-

nologies 65
2.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.1.1 Representative Agent Characterization . . . . . . . . . . . . . . . . . . . . 71

iii
TABLE OF CONTENTS 1

2.1.2 Sufficient conditions for existence of a solution . . . . . . . . . . . . . . . 74

2.1.3 Optimal consumption investment with fixed costs and irreversibility . . . . 78
2.2 Equilibrium Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.2.1 Asset Prices and the Pricing Kernel . . . . . . . . . . . . . . . . . . . . . 85
2.2.2 Oil Spot Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.2.3 Oil Futures Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2.3 Model Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.3.1 Data and Calibration Criterium . . . . . . . . . . . . . . . . . . . . . . . . 90
2.3.2 Commodity Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.3.3 Regime-Switching Estimation . . . . . . . . . . . . . . . . . . . . . . . . 98
2.4 Extensions - Flexible production with adjustment cost . . . . . . . . . . . . . . . . 102
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
2.6 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
2.7 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Appendix 2.A Numerical Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Chapter 1

‘Maximal’ Model of Commodity Futures

with Stochastic Convenience Yield and
Interest Rates1

Commodity derivatives markets have witnessed a tremendous growth in recent years. A variety of

models have been proposed for pricing commodity derivatives such as futures and options.2 In his

presidential address, Schwartz (1997) selects and empirically compares three models. His empirical

results suggest that three factors, driving spot prices, interest rates and convenience yields are nec-

essary to capture the dynamics of futures prices. Further, models accommodating mean-reversion

in spot prices under the risk-neutral measure seem desirable, although in that case, Schwartz ar-

gues commodities cannot be seen as “an asset in the usual sense,”3 because they do not satisfy the

standard no-arbitrage condition for traded assets.

Below, we develop a three factor Gaussian model of commodity futures prices which nests the

three specifications analyzed by Schwartz (1997) as well as Brennan (1991), Gibson and Schwartz

(1990), Ross (1997) and Smith and Schwartz (1998). Instead of modeling separately the dynamics
1
An updated version of this chapter can be found at www.ing.puc.cl/˜jcasassu as the working paper “Max-
imal Gaussian Jump-diffusion Model of Commodity Futures with Stochastic Convenience Yield and Interest Rates”
co-authored with Pierre Collin-Dufresne.
2
Early description of the market can be found in the collection of papers “Managing Energy Price Risk” (1995). Seppi
(2003) offers a more up to date survey.
3
See Schwartz (1997) p. 926. See also Ross (1997).

2
 3

of spot, interest rates and convenience yield process, we start by directly specifying the most gen-

eral identifiable three (latent) factor Gaussian model of futures prices.4 Assuming the term structure

of risk-free interest rates is driven by a single factor and imposing a restriction on the drift of spot

prices which amounts to the standard no-arbitrage condition, we identify the convenience yield im-

plied by our general model of futures prices. We show that it allows for a richer unconditional

covariance structure of convenience yields, commodity prices and interest rates than previous mod-

els. In particular, the convenience yield may depend both on the spot price and the risk-free rate

itself.

One simple insight of our framework is that the models by Ross (1997) and Schwartz (1997),

which allow for mean-reversion under the risk-neutral measure of spot prices, can simply be inter-

preted as arbitrage-free models of commodity spot prices, where the convenience yield is a function

of the spot price. Spot price level dependence in convenience yields leads to mean-reversion of spot

prices under the risk-neutral measure. The latter feature seems to be empirically desirable to fit the

cross-section of futures prices.

Several papers (Working (1949), Brennan (1958), Deaton and Laroque (1992), Routledge, Seppi

and Spatt (2000)) have shown that convenience yields arise endogenously as a result of the inter-

action between supply, demand and storage decisions. In particular, Routledge, Seppi and Spatt

(RSS 2000) show that, in a competitive rational expectations model of storage, when storage in the

economy is driven to its lower bound, e.g. in periods of relative scarcity of the commodity available

for trading, convenience yields should be high. This provides some economic rational for allowing

the convenience yield to depend on spot prices as in our model. Indeed, assuming that periods of

scarcity, e.g., low inventory, correspond to high spot prices, this theory predicts a positive relation

between the convenience yield and spot prices.

Further, RSS 2000 note that the correlation structure between spot prices and convenience yields

should be time-varying, in contrast to the prediction of standard commodity derivatives pricing mod-

els such as Brennan (1991), Gibson and Schwartz (1990), Amin, Ng and Pirrong (1995), Schwartz

(1997) and Hilliard and Reis (1998). While the model we develop has a constant instantaneous

correlation structure (since it is Gaussian) it allows for a more general unconditional correlation
4
In Dai and Singleton’s (2000) terminology we use the ‘maximal’ A0 (3) model.
 4

structure of spot price and convenience yields than previous papers.

Most theoretical models of convenience yields (such as RSS 2000), assume that interest rates

are zero, and thus do not deliver predictions about how interest rates should affect the convenience

yield. However, to the extent that inventory and interest rates are correlated, it seems consistent with

the theory to find a relation between interest rates and convenience yields. Further, interest rates in

general proxy (at least partially) for economic activity, which in turn may affect convenience yields.

To empirically implement the model and estimate the significance of the previously imposed

over-identifying restrictions, we need a specification of risk-premia. Following Duffee (2002), we

allow risk-premia to be affine in the state variables. This specification nests the constant risk-

premium assumption made in previous empirical analysis of commodity futures (e.g., Schwartz

(1997)). Existing theoretical models of commodity prices based on the theory of storage (such

as RSS 2000) assume risk-neutrality, and thus make no prediction about risk-premia. However,

allowing for time-varying risk-premia is important since, as argued by Fama and French (1987,

1988), negative correlation between risk-premia and spot prices may generate mean-reversion in

spot prices. In the context of our affine model, allowing for risk-premia to be level dependent

implies that state variables have different strength of mean-reversion under the historical and risk-

neutral measures. Mean-reversion under the risk-neutral measure is due to convenience yields,

whereas mean-reversion under the historical measure results from both the convenience yield and

the time-variation in risk-premia. The former is important to capture the cross-section of futures

prices, whereas the latter affects the time series properties of spot and futures prices.

We use weekly data on crude oil, copper, gold and silver futures contracts and U.S. treasury

bills, from 1/2/1990 to 8/25/2003. We estimate the model using maximum-likelihood, since it takes

full advantage of the Gaussian-affine structure of our model.5 Using standard pricing results on

Gaussian-affine models (Langetieg (1980), Duffie, Pan and Singleton (2000)) we obtain closed-

form solutions for futures, zero-coupon bond prices and the transition density of the state vector.

Results indicate that the maximal convenience yield model improves over all (nested) specifications

previously investigated. Three factors are needed to capture the dynamics of futures prices. Allow-
5
The same approach has been widely used in the literature: Chen and Scott (1993), Pearson and Sun (1994), Duffie
and Singleton (1997).
 5

ing convenience yields and risk-premia to be a function of the level of spot commodity prices as

well as interest rates is an important feature of the data. For crude oil and copper we find conve-

nience yields are significantly increasing in spot commodity prices, in line with predictions of the

theory of storage. For silver this dependence is much lower and for gold it is negligible. For these

two metals the level of convenience yield is much lower and not very variable. For all commodities

the sign of the dependence of convenience yields on interest rates is positive and significant. For

all commodities we find economically significant negative correlation between risk-premia and spot

prices. The point estimates further suggest that the contribution of time variation in risk-premia to

the total mean-reversion strength under the historical measure is increasing in the degree to which

an asset may serve as a store of value, e.g. as a financial asset. Related, the level of convenience

yields is increasing in the degree to which an asset serves for production purposes (high for oil and

copper, and low for gold and silver).

These results are robust to the inclusion of jumps in the spot dynamics. Specifically, we de-

compose the jump component of spot commodity prices into three parts. We find evidence for a

high-intensity jump with stochastic jump size with approximately zero mean, and two lower in-

tensity jumps with constant jump sizes. The estimates of the risk-neutral drift parameters of the

state vector are almost unchanged. Including jumps mainly affects the estimates of the volatility

coefficients and the risk-premia parameters. Indeed, we show that jumps in the spot price have little

impact on the predicted cross-section of futures prices.6 However, accounting for jumps helps better

capture the historical measure dynamics of futures prices.

Bessembinder et al. (1995) also find evidence for mean-reversion in commodity prices by com-

paring the sensitivity of long-maturity futures prices to changes in spot prices (or, effectively, short

maturity futures prices). Since their test uses only information from the cross-section of futures

prices, it cannot detect mean-reversion resulting from “movements in the risk-premium component”

(see their discussion p.362). Consequently, their test cannot determine whether historical time se-
6
As we show in Appendix 1.G, jumps impact futures prices only when the convenience yield depends on the spot
price. The intuition is that futures prices are martingales under the risk-neutral measure. Combined with the martingale
restriction on the drift of the spot price process, this implies that jumps in the spot price can only ‘matter’ if there is a
common jump in the convenience yield (or the interest rate). Hilliard and Reis (1998) for example, find that, in their
model, jumps have no impact on futures prices. Their convenience yield model is not maximal however.
 6

ries of commodity prices actually exhibit mean-reversion.7 In contrast to their paper, our model

allows to disentangle the various sources of mean-reversion: level dependence in convenience yield

vs. time-variation in risk-premia. Fama and French (1988) study the importance of time-variation

in risk-premia for mean-reversion in commodity prices using simple univariate linear-regressions

of changes in spot prices and forward premium on the basis (similar to Fama (1984)). Their re-

sults are inconclusive for most commodities (and in particular for the metals studied here), mainly,

they argue, because the basis exhibits too little volatility for regressions to reliably identify time-

variation in risk-premia. In contrast, viewing commodity futures through the ‘filter’ of affine models

potentially allows us to obtain more reliable estimates of time-variation in risk-premia.8

Finally, we document the economic importance of disentangling the two sources of mean-

reversion, by studying two applications: option pricing and value at risk computations. Ignoring

spot price dependence of convenience yields results in a mis-specification of the risk-neutral dy-

namics of the spot price and can result in gross mis-valuation of options. Mean-reversion under

the risk-neutral measure effectively reduces the term volatility of the spot price and the expected

convenience yield (which acts as a stochastic dividend) which tend to reduce option values. This

is especially true for oil and copper, where an important fraction of the total mean-reversion is due

to the positive relation between spot prices and convenience yields. Comparing option prices using

our parameter estimates with those obtained using a restricted model (with parameters estimated

imposing that convenience yield be linearly independent of the spot price) results in sizable errors

of about 30% for in and at the money options. An implication is that for crude oil and copper in-

vestments the naive model will predict much higher real-option values and tend to differ investment

more than the more realistic ‘maximal’ model.

Similarly, ignoring time-variation in risk-premia may lead to severe over-estimation of the value

at risk of real, commodity-related investments. Comparing the value at risk of an investment in one

unit of the asset obtained when estimating the naive restricted model vs. the ‘maximal’ model we

find that the tails of the distribution of the naive model tend to be fatter the longer the maturity of the
7
Indeed, the risk-premia could, in principle, be time-varying in a way to offset the ‘risk-neutral’ mean-reversion
induced by convenience yields.
8
Piazzesi (2002) offers further discussion of the advantages of the affine framework, which explicitly imposes cross-
sectional no-arbitrage restriction, over an unrestricted VAR for example.
1.1. The ‘Maximal’ Convenience Yield Model 7

investment considered. For copper, gold and silver, we find that for a five year horizon investment

the loss implied by a 5% value at risk more than doubles when computed with models which ignore

time-variation in risk-premia.

These examples illustrate that disentangling the sources of mean-reversion in risk-premia can

have a substantial impact on valuation, investment decision9 and risk-management.

The rest of the chapter is structured as follows. Section 2.1 presents the model. Section 1.2

discusses the specification of risk premia. Section 1.3 describes the empirical analysis and discusses

the results. Section 1.4 shows the economic implications of the model and Section 1.5 concludes.

1.1 The ‘Maximal’ Convenience Yield Model

In this section we develop a general three-factor Gaussian model of (log) futures prices. Following

Duffie and Kan (DK 1996), Duffie, Pan and Singleton (DPS 2000) and Dai and Singleton (DS 2000),

we first introduce a ‘canonical’ representation of a three-factor Gaussian state vector driving futures

prices.10 We assume that the spot commodity price S (t) is defined by:

X(t) := log S (t) = φ0 + φ>

Y
Y(t) (1.1)

φ0 is a constant, φY is a 3 × 1 vector, and Y > (t) = Y1 (t), Y2 (t), Y3 (t) is a vector of state variables that

follows a Gaussian diffusion process under the risk-neutral measure Q:11

dY(t) = −κ Q Y(t)dt + dZ Q (t) (1.2)

9
The impact of various assumptions about the dynamics of the convenience yield on real option valuation and invest-
ment decisions is also discussed in the last section of Schwartz (1997).
10
The model is in the A0 (3) family using the terminology of DS 2000. They show that N-factor affine model can
be classified into N+1 families of models denoted AM (N) depending on the number, M, of state variables entering the
conditional variance covariance structure of the state vector. In Gaussian models the conditional covariance structure is
constant, M = 0.
11
We assume the existence of such a risk-neutral measure. See Duffie (1996) for conditions under which the existence
of such a measure is equivalent to the absence of arbitrage. If sufficient number of futures contracts are traded, then
in general, with such an affine structure markets are complete and this martingale measure is unique. Collin-Dufresne
and Goldstein (2000) however build finite dimensional affine models with a continuum of traded derivatives that yield
incomplete markets.
1.1. The ‘Maximal’ Convenience Yield Model 8

where κ Q is a 3×3 lower triangular matrix that reflects the degree of mean reversion of the processes,

and dZ Q is a 3 × 1 vector of independent Brownian motions. It is well-known (e.g., Duffie (1996))

that the futures price F T (t) at time t for purchase of one unit of commodity S (T ) at time T is

simply the expected future spot price under the risk-neutral measure. Using standard results on

pricing within the affine framework (e.g., Langetieg (1980), DK (1996), DPS (2000)), we obtain the

following expression:
h i >
F T (t) = EtQ eX(T ) = eAF (T −t)+BF (T −t) Y(t) (1.3)

where AF (τ) and BF (τ) are the solution to the following system of ODEs:

dAF (τ) 1
= B (τ)> BF (τ)
dτ 2 F
dBF (τ) >
= −κ Q BF (τ)
dτ

with boundary conditions AF (0) = φ0 and BF (0) = φY which can be solved in closed form (see

Appendix 1.A).

Such a model is maximal in the sense that, conditional on observing only futures prices (and

not the state variables Y1 , Y2 , Y3 themselves), it has the maximum number of identifiable parameters.

This result follows directly from the analysis in DS 2000. However, unlike in DS 2000 where bonds

are derivatives of the non-traded short rate, in our framework, the underlying process S (t) is a traded

commodity. We emphasize that the assumption that we observe all futures prices implies that the

spot price, which is but one particular futures price, is ‘observable.’12 Absence of arbitrage therefore

implies:

EtQ [dS (t)] = (r(t) − δ(t))S (t)dt (1.4)

where r(t) is the instantaneous risk-free rate and δ(t) is the instantaneous convenience yield. The

latter has the standard interpretation of a dividend flow, net of storage costs, which accrues to the

holder of the commodity in return for immediate ownership (e.g., Hull (1997)). As discussed in the

introduction, convenience yields also arise endogenously in models based on the “theory of storage”
12
All that is really needed is that it can be ‘inverted’ from the cross-section of futures prices, which implies that at least
three futures prices be observed in our model. This is similar to the special role played by the short rate for identification
of parameters in affine term structure models (e.g., Collin-Dufresne, Goldstein and Jones (2002)).
1.1. The ‘Maximal’ Convenience Yield Model 9

(e.g., RSS 2000) as a result of the interaction between supply, demand and storage decisions. Aug-

menting the data set with bond prices and making an identifying assumption about the short-rate

model driving the term structure of interest rates, we can recover the process for the convenience

yield from equation (1.4), effectively viewing the latter as defining the convenience yield.13 Fol-

lowing previous empirical papers on commodity futures, we assume the risk-free rate follows a

one-factor Gaussian process:14

r(t) = ψ0 + ψ1 Y1 (t) (1.5)

RT
Zero-coupon bond prices may be computed explicitly by solving for PT (t) = EtQ [e− t
r(s)ds
] as in

Vasicek (1977) (see Appendix 1.B).

Using the definitions for X(t) and r(t) given in equations (1.1) and (1.5), the arbitrage restric-

tion (1.4), and applying Itô’s lemma, we obtain the following expression for the maximal conve-

nience yield model implied by our model:

EtQ [dX(t)] + 12 VtQ [dX(t)] 1

δ(t) = r(t) − = ψ0 − φ> φ + ψ1 Y1 (t) + φ> κ Q Y(t) (1.6)
dt 2 Y Y Y

Noting that equations for X, r, δ given in (1.1), (1.5) and (1.6) above specify a unique transformation

from the latent variables {Y1 , Y2 , Y3 } to {r, δ, X} we may derive the dynamics of the convenience yield

implied by the model. We summarize the results in the following proposition:

Proposition 1.1 Assume the risk-free interest rate follows an autonomous one-factor Ornstein-

Uhlenbeck process as in equation (1.5), then the ‘maximal’ model of futures prices and convenience

yields defined in equations (1.1-1.6) can equivalently be represented by:

 
dr(t) = κrQ θrQ − r(t) dt + σr dZrQ (t) (1.7)
 
dδ(t) = κδ0Q + κδrQ r(t) + κδQ δ(t) + κδX
Q X(t) dt + σ dZ Q (t)
δ δ
(1.8)
!
1
dX(t) = r(t) − δ(t) − σ2X dt + σX dZXQ (t) (1.9)
2
13
If the spot price is actually not a traded asset (as would be the case for electricity futures for example), then the
process δ defined by equation 1.4 is still of interest, as it reflects, per definition, how much the spot price dynamics differ
from that of a traded asset.
14
This model is maximal in the A0 (1) family, i.e., conditional on observing only bond prices, it has the maximum
number of parameters identifiable for a one-factor Gaussian model.
1.1. The ‘Maximal’ Convenience Yield Model 10

where ZXQ , ZδQ , ZrQ are standard correlated Brownian motions.

Proof: The proof follows immediately from applying Itô’s Lemma to X, r, δ defined in equations

equations (1.1), (1.5) and (1.6) above and noting that these equations specify a unique transfor-

mation from the latent variables {Y1 , Y2 , Y3 } to {r, δ, X}. In the appendix we provide the relation

between the parameters of the latent model and the parameters of the {r, δ, X} representation. For

future reference we define the correlation coefficients:

dZXQ (t)dZδQ (t) = ρXδ dt dZXQ (t)dZrQ (t) = ρXr dt dZδQ (t)dZrQ (t) = ρδr dt (1.10)

In the class of three-factor Gaussian models of futures (and spot) commodity prices, where the

short rate is driven by one factor, this is the most general specification of the convenience yield

that is also identifiable. By analogy to the terminology of DS (2000), we call it the ‘maximal’

convenience yield model.15

The proposition shows that the drift of the convenience yield process in general may depend on

both the interest rate and the spot rate. This contrasts with the specifications analyzed in the existing

literature which, in general, assume that the convenience yield follows an autonomous process, i.e.,

that the highlighted coefficients in equation (1.8) are zero. The following proposition provides a

better understanding for the significance of imposing restrictions on the parameters κδrQ , κδX
Q.

Proposition 1.2 The maximal convenience yield of proposition 1.1 can be decomposed as

δ(t) = b
δ(t) + αr r(t) + αX X(t) (1.11)

where b
δ follows an autonomous Ornstein-Uhlenbeck process:

 
db
δ(t) = κbQ θbQ − b
δ(t) dt + σbδ dZbQ (t) (1.12)
δ δ δ

15
Note that unlike in DS (2000), we have a three state variable model of two types of securities, bond and futures
prices. Even though the two models are separately maximal, one may wonder if together they form a maximal model,
as the joint observation of the two securities may allow the empiricist to recover more information about the state vector
than observing the two separately. It turns out that in the Gaussian case the joint observation of two types of securities
does not help identify more parameters. The model above is thus maximal, conditional on observing bond and futures
prices, and restricting the term structure to be driven by only one-factor.
1.1. The ‘Maximal’ Convenience Yield Model 11




 Q
 αr = 0 ⇐⇒ κδr = 0

There is a unique such decomposition such that 



 αX = 0 ⇐⇒ κδX Q =0

Using that decomposition the dynamics of the spot price process become:

 
dX(t) = αX (θXQ − X(t)) + (αr − 1)(θrQ − r(t)) + θbQ − b
δ(t) dt + σX dZXQ (t) (1.13)
δ

 
where the long-term mean of the log spot price is given by θXQ = 1
αX (1 − αr )θrQ − θbδQ − 12 σ2X .

Proof: Applying Itô’s lemma to the right hand side of equation (1.11) and equating drift and dif-

fusion of the resulting process with those of equation (1.8) shows that there exist two possible

proposed decompositions given by:

q !
1
α±X = −κδQ ± (κδQ )2 − 4κδX
Q
(1.14)
2
α±X − κδrQ
α±r = (1.15)
α±X + κrQ + κδQ
±
κbQ = −κδQ − α±X (1.16)
δ

± ±
κbQ θbQ = κδ0Q + α±X σ2X /2 − α±r κrQ θrQ (1.17)
δ δ

σb± dZbQ± = σδ dZδQ − α±X σX dZXQ − α±r σr dZrQ (1.18)

δ δ

(σb± )2 = σ2δ + (α±X )2 σ2X + (α±r )2 σ2r (1.19)

δ

−2ρδX σδ σX α±X + 2ρrX α±r α±X σr σX − 2ρrδ σδ α±r σr (1.20)

ς ς
Defining ς = sign(κδQ ) we see that only the solution αςX , αςr , κbδQ , θbδQ satisfies the condition αX =
Q = 0 and α = 0 ⇐⇒ κ Q = 0. Further, we note that α , α are real if and only if
0 ⇐⇒ κδX r δr r X

2
κδQ − 4κδX
Q ≥ 0 which corresponds to the condition that eigenvalues of the mean-reversion matrix be

real.

Finally, we note that ZbδQ defined by equations (1.18) and (1.20) is a standard Brownian motion

which is correlated with ZXQ , ZrQ . For future reference we define the correlation coefficients as:

dZXQ (t)dZbQ (t) = ρbδX dt dZbQ (t)dZrQ (t) = ρrbδ dt (1.21)

δ δ
1.1. The ‘Maximal’ Convenience Yield Model 12

The two propositions above show that once we assume the short rate follows an autonomous

one-factor process, then the arbitrage restriction (1.4) delivers a convenience yield process which

has its own specific stochastic component b

δ but is also linearly affected by the short rate and the log

spot price.16 Proposition 1.2 also makes apparent that the maximal model nests the three models

analyzed in Schwartz (1997), as well as the models of Ross (1997), Brennan and Schwartz (1985),

Gibson and Schwartz (1990) and Schwartz and Smith (2000).17 For example, Schwartz’s model

1 corresponds to a one factor (X) model with αr = 0. Schwartz’s model 2 corresponds to a two-

factor model (X, b

δ) with αr = αX = 0. Schwartz’s model 3 corresponds to a three factor model with

αr = αX = 0.

One simple insight of the maximal convenience yield model is that the one-factor models of

Ross (1997) and Schwartz (1997), which allow for mean-reversion under the risk-neutral measure

of spot prices, can simply be interpreted as arbitrage-free models of commodity spot prices, where

the convenience yield is a function of the log-spot price. A positive relation between the conve-

nience yield and the (log) spot price, i.e. a positive αX , leads to a mean-reverting spot price under

the risk-neutral measure. The latter feature seems to be empirically desirable to fit the cross-section

of futures prices. A positive relation between convenience yield and spot price also seems consis-

tent with the predictions of theoretical models. Several papers (Working (1949), Brennan (1958),

Deaton and Laroque (1992), Routledge, Seppi and Spatt (2000)) have shown that convenience yields

arise endogenously as a result of the interaction between supply, demand and storage decisions. In

particular, Routledge, Seppi and Spatt (RSS 2000) show that, in a competitive rational expectations

model of storage, when storage in the economy is driven to its lower bound, e.g. in periods of

relative scarcity of the commodity available for trading, convenience yields should be high. This

provides some economic rational for allowing the convenience yield to depend on spot prices as

in our maximal model. In fact, assuming that periods of low inventory and relative scarcity of the
16
Note that it seems economically sensible to assume that a market wide variable such as the short rate follows an
autonomous process, i.e. is not driven by the convenience yield or the spot price of a specific commodity. The model
could easily be extended to allow for multi-factor term structure models. However, to maintain the assumption that
interest rate risk is ‘autonomous’ and at the same time have convenience yield and spot price be specific sources of risk
would require a four-factor model.
17
Some of these are actually nested in the models analyzed by Schwartz (1997).
1.2. Specification of Risk Premia 13

commodity coincide with high spot prices, the theory of storage predicts a positive relation between

the convenience yield and spot prices.

Further, RSS 2000 note that the correlation structure between spot prices and convenience yields

should be time-varying, in contrast to the prediction of standard commodity derivatives pricing

models such as Brennan (1991), Gibson and Schwartz (1997), Amin, Ng and Pirrong (1995) and

Schwartz (1997). Since it is a Gaussian model, the maximal convenience yield has a constant in-

stantaneous correlation structure. However, since all state variables enter the drift of convenience

yield and spot price, it allows for a richer unconditional correlation structure than previous specifi-

cations.18

Finally, note that previous models restrict αr to be zero, i.e. convenience yields to be indepen-

dent of the level of interest rates. While most theoretical models assume zero interest rates (e.g.,

RSS 2000) and thus do not deliver empirical predictions about that coefficient, relaxing this as-

sumption seems desirable. If we expect interest rates and inventory to be correlated, then, following

the “theory of storage” argument, we may expect a significant non-zero coefficient. In fact, to the

extent that holding inventory becomes more costly in periods of high interest rates, we may expect

a negative correlation between interest rates and inventory and thus a positive αr .

Of course, our model is a reduced-form model which makes no predictions about these relations.

However, it is the natural framework to investigate empirically these questions. In the next sections

we discuss the specification of risk-premia and empirical implementation.

1.2 Specification of Risk Premia

Our discussion above is entirely cast in terms of the risk-neutral dynamics of state variables. These

are useful to price the cross-section of futures prices. To explain the historical time-series dynamics

of prices and subject our model to empirical scrutiny we need a specification of risk-premia. In

contrast to previous empirical research (e.g., Schwartz (1997)) which assumes constant risk-premia,

we allow risk-premia to be a linear function of the state variables following Duffee (2002) and Dai
18
In RSS 2000, the correlation is derived endogenously and is a function of the level of inventory. To the extent that
spot prices proxy for inventories, the maximal convenience yield model may be able to capture that feature. It is a reduced
form model, however, and inventory is not an explicit state variable of the model.
1.2. Specification of Risk Premia 14

and Singleton (2002).

In terms of the canonical representation, this amounts to defining the relation between the his-

torical and risk-neutral measure using the following specification of the Girsanov factor:

dZ Q (t) = dZ(t) + β0Y + β1Y Y(t) dt (1.22)

Here dZ is a 3 × 1 vector of independent Brownian motions under the physical martingale measure,

β0Y is a 3 × 1 vector of constants and β1Y is a 3 × 3 matrix of constants. With this specification the

dynamics of the state variables is Gaussian under both the historical and risk-neutral measure. The

process under the physical measure is given by:

dY(t) = (β0Y − (κ Q − β1Y )Y(t))dt + dZ(t) (1.23)

Note that both the mean reversion coefficient and the long-run mean differ under both measures.

Under the physical measure the mean reversion matrix is (κ Q − β1Y ) and the long-run mean vector is

(κ Q − β1Y )−1 β0Y . Traditional models assume that β1Y = 0.19

For ease of economic interpretation we prefer to study the {r, b

δ, X} representation obtained in

proposition 1.2.20 The risk premium specification of equation (1.22) can equivalently be rewritten

in terms of the rotated Brownian motion basis (see Appendix 1.D) as:

        
 Q        
 Zr   r  Z  β0r   βrr βrbδ βrX   r(t) 
         
d  ZbQ  = d  Z  + Σ−1  β  +  β βbδbδ

βbδX  

δ(t)  dt
b (1.24)
 δ   bδ   0bδ   bδr
 Q         
ZX   Z   β   β βXbδ βXX   X(t) 
X 0X Xr

19
Duffee (2002) shows that the more general, “essentially affine” specification, improves the ability of term structure
models at capturing the predictability of bond price returns under the historical measure, while retaining their ability at
pricing the cross-section of bonds (i.e. fitting the shape of the term structure).
20
As is apparent from the proof of proposition 1.2, studying this particular decomposition of the ‘maximal’ convenience
yield model of proposition 1.1 effectively restricts the model to the parameter set for which the eigenvalues of the mean-
reversion matrix are real. We checked empirically (by directly estimating the model of proposition 1.1) that this restriction
was never binding for our data.
1.2. Specification of Risk Premia 15

where  
 
 σr 0 0 
 

Σ =  0 σbδ 0  (1.25)
 
 
0 0 σX 

Given the representation adopted it seems natural to impose the following restrictions on the risk

premia: 



 βrbδ = βrX = 0




(1.26)

 βb = βb = 0
δr δX

The first set of restrictions basically guarantees that the risk-free interest rates term premia do not

depend on the level of convenience yield or commodity spot price. This insures that the short rate

follows an autonomous process under both measures. It is also consistent with our applying this

model to different commodities which all share the same interest rate model. The second set of

restrictions simply guarantees that the component of the convenience yield (b

δ) which is linearly

independent of interest rate and spot price level under the risk-neutral measure, remains so under

the historical measure. With these restrictions, the dynamics of the state variables {r, b
δ, X} under

the historical measure have the same form as under the risk-neutral measure, but with different

risk-adjusted drift coefficients.

Proposition 1.3 If risk-premia are given by equations (1.24)-(1.26) then the state variables {r, b
δ, X}

introduced in proposition 1.2 have the following dynamics under the physical measure:

 
dr(t) κrP θrP − r(t) dt + σr dZr (t)
= (1.27)
 
db
δ(t) = κbP θbP − b δ(t) dt + σbδ dZbδ (t) (1.28)
δ δ
!
1 2
dX(t) = µ(t) − δ(t) − σX dt + σX dZX (t) (1.29)
2
 
:= κXr (θr − r(t)) + κ Pb(θbP − b
P P
δ(t)) + κXP (θXP − X(t)) dt + σX dZX (t) (1.30)
Xδ δ

where δ is as defined in equation (1.11) and ZX , Zbδ , Zr are standard Brownian Motions defined in

equation (1.24).
1.2. Specification of Risk Premia 16

The relation between the P and Q parameters expressed in terms of the risk-premia is:

κrQ θrQ +β0r

κrP = κrQ − βrr θrP = (1.31)
κrQ −βrr

Q κbQ θbQ +β
0b
κP = κbδ − βbδbδ θP = δ δ δ
(1.32)
b
δ b
δ κbQ −βbb
δ δδ
 
µ(t) = r(t) + β0X + βXr r(t) + βXbδb
δ(t) + βXX X(t) (1.33)

P = κQ − β
κXr κ Pb = 1 − βXbδ κXP = κXQ − βXX (1.34)
Xr Xr Xδ

Q θQ − κP θP + θQ − κP θP
κXr
P
θXQ κXQ + β0X r Xr r b
δ b b Xδ δ
θX = + (1.35)
κXQ − βXX κXQ − βXX

Proposition 1.3 above shows that allowing for essentially affine risk-premia allows to disentangle

the level of mean-reversion in spot commodity prices under the risk-neutral measure from the level

of mean-reversion under the historical measure. The former is essential to capture the term struc-

ture of futures prices (i.e. the cross section), whereas the latter captures the time-series properties

of spot commodity prices. Fama and French (1987, 1988) argue that negative correlation between

risk-premia and spot prices can generate mean-reversion in spot prices. Our model captures this

feature as is apparent from equations (1.29) and (1.33). A negative βXX implies negative correlation

between risk-premia and spot prices and generates mean-reversion in spot prices. Thus, our model

has the ability to distinguish two sources of mean-reversion. First, mean-reversion in (log) spot

prices can be due to level dependence in convenience yield (a positive αX ) which is consistent with

the theory of storage. Second, mean-reversion can appear as a result of negative correlation between

risk-premia and spot prices (a negative βXX ). Only the convenience yield component affects the cross

section of futures prices, i.e. enters the risk-neutral measure dynamics. Both drive the time-series

of commodity prices, i.e., enter the historical measure price dynamics. In addition, the instanta-

neous correlation of the spot price with interest rate and b

δ combined with the signs of respectively

κXr and κXbδ may contribute to ‘mean-reversion like’ behavior in commodity prices. Distinguishing

between the various sources (if any) of mean-reversion may have important consequences for valu-

ation and investment decision, as well as risk-management, as we document below. We first turn to

the empirical estimation of the model.

1.3. Empirical implementation 17

1.3 Empirical implementation

We estimate our model for four types of commodity futures using maximum likelihood. We first

describe the data, then the empirical methodology and discuss the results.

1.3.1 Description of the Data

Our data set consists of futures contracts on crude oil, copper, gold and silver and zero-coupon bond

prices.21 For all commodities we use weekly data from 1/2/1990 to 8/25/2003. Table 1.1 contains

the summary statistics of the four commodities. The maturities of the contracts studied differ across

commodity. We use short-term contracts with maturities 1, 3, 6, 9, 12, 15 and 18 months (labeled

from F01 to F18), and depending on availability we also include longer maturity contracts. For

crude oil, copper, gold and silver we use long-term contracts with maturities up to 36, 24, 48 and

48 months, respectively. This long-term data is not fully available for the whole period studied (713

weeks) since many of these contracts where not available in 1990. If an specific contract is missing

we select the one with the nearest maturity. A special characteristic of futures contracts is that the

last trading day is a specific day of each month, implying that the maturity of the contracts varies

over time.22 For interest rates we use constant maturity Treasury yields to build zero-coupon bonds

with maturities of 0.5, 1, 2, 3, 5, 7 and 10 years.

Figure 1.1 shows the price of the F01 and F18 contracts for crude oil, copper, gold and silver.

We can see a decreasing tendency on copper and gold prices during the period analyzed. Also,

copper, oil and gold reached their lowest price in the period during the first half of 1999. Finally,

if we “casually” compare the F01 and F18 contracts, there appears to be mean-reversion (under the

risk-neutral measure) in copper and crude oil prices. Indeed the difference between the F18 and

F01 futures prices alternates signs.23 Because of convergence, the F01 futures price should be close
21
The data for the commodities is from the New York Mercantile Exchange. The crude oil data is from the NYMEX
Division, while copper, gold and silver data is from the COMEX Division. The interest rate data is from The Federal
Reserve Board.
22
The last trading day is different across commodities. For copper, gold and silver the last trading day is the close of
the third last business day of the maturing delivery month, while for crude oil it is the close of the third business day prior
to the 25th calendar day of the month preceding the delivery month.
23
Suppose that d ln S t = (r − δ − κ ln S t − 12 σ2 )dt + σdZtQ where all coefficients are constant. Then simple calculations
 2

show that F T (t) = EtQ [S T ] = exp θ + (ln S t − θ)e−κ(T −t) + σ2 B2κ (T − t) where κθ = r − δ − 21 σ2 and Bκ (τ) = (1 − e−κτ )/κ.
−2κT
σ2 (e−2κT 1 −e 18 )
It is thus clear that F18 > F01 ⇔ ln S t < θ + −κT . Further in the absence of mean-reversion under the
4κ(e−κT 1 −e 18 )
1.3. Empirical implementation 18

to the spot price. Thus alternating signs in F18 − F01 suggests periods of strong backwardation

in oil and copper markets as documented in Litzenberger and Rabinowitz (1995). Gold and silver

exhibit fewer episodes of strong backwardation. The ‘basis’ estimated by F18 − F01 appears to

be more stable and mostly positive. Figure 1.2 plots the term structures for each commodity and

confirms these findings. It seems that oil and copper has higher degrees of mean-reversion (under

the risk-neutral measure) than gold and silver. Figure 1.3 presents the historical evolution of the

6-month and 60-month interest rates used for the estimation.

1.3.2 Empirical Methodology

We use maximum-likelihood estimation using both time-series and cross-sectional data in the spirit

of Chen and Scott (1993) and Pearson and Sun (1994).24 Since the three state variables {r, b
δ, X} are

not directly observed in our data set, their approach consists in arbitrarily choosing three securities

to pin down the state variables. Instead, we follow Collin-Dufresne, Goldstein and Jones (2002)

and choose to fit the first principal component of the term structure of interest rates and the first

two principal components of the futures curve. Since the principal components remain affine in the

state variables, they can easily be inverted for the state variables using the closed form formulas

given in appendices Appendix 1.A and Appendix 1.B which depend on the risk-neutral parameters.

The remaining principal components of the term structure and of futures prices, which, at any point

in time, are also deterministic functions of the state variables are then over identified.25 Following

Chen and Scott (1993), we assume they are priced or measured with ‘measurement errors,’ which

we assume follow an AR(1) process. For simplicity, we assume that measurement errors in the fu-

tures prices principal components have the same auto-correlation coefficient. Similarly, we estimate

only one auto-correlation coefficient for risk-free term structure errors. Given the known Gaussian

transition density for the state variables and the distribution for the error terms, the likelihood can be
risk-neutral measure (κ = 0) we observe that F18 − F01 has the same (constant) sign as r − δ.
24
Duffie and Singleton (1997), Collin-Dufresne and Solnik (2001), Duffee (2002) use a similar method. Schwartz
(1997) uses a Kalman-filter approach.
25
The principal components can be thought of as portfolios of contracts with different maturities. The first principal
component is in general an equally weighted portfolio of contracts, while the second principal component is a portfolio
with weights that are linearly decreasing with maturity. See Collin-Dufresne, Goldstein and Jones (2002) for further
details on the procedure.
1.3. Empirical implementation 19

derived.26 We note that the transition density depends on the historical measure parameters. Apart

from the likelihood value itself, the resulting properties of the “measurement errors” provide direct

(mis-)specification tests for the model. Since long-term futures contracts are not always available,

we back out the factors from the principal components of the data that is fully available for the

whole period studied (713 weeks). For crude oil and copper we use the PCs of the contracts with

maturity up to 18 months while for gold and silver we use the contracts with maturity up to 24

months. The remaining long-term contracts are assumed to be observed with measurement errors.

1.3.3 Empirical Results

Table 1.2 presents the maximum-likelihood estimates of the ‘maximal’ convenience model pre-

sented in propositions 1.2 and 1.3. For each commodity we present the risk-neutral parameters

which affect the drift of spot price, convenience yield and interest rate processes under the risk-

neutral measure, the risk-premia parameters, the volatility and correlation parameters, and the auto-

correlation coefficients of the measurement errors of futures (ρF ) and Treasury rates (ρP ). Table 1.3

presents the likelihood-ratio test results for three different sets of restrictions compared to the max-

imal model. In table 1.4 we also report the point estimates of the drift parameters of the various

processes under the historical measure. As shown in proposition 1.3, these are simple transforma-

tions of the risk-neutral and risk-premia parameters given in table 1.2 (for example, κXP = αX − βXX ).

Finally, table 1.5 reports point estimates for the unconditional first and second moments (long-term

mean and covariances) of convenience yield and log spot prices. In the same table, we also present

the long-term spot prices.27

Table 1.2 shows that all risk-neutral parameters are significant except for some of the correlation

coefficients ρrbδ , ρrX . This suggests that three factors are indeed necessary to explain the dynamics

of each of the four commodities and, further, that innovations in the risk-free interest rates are

uncorrelated with innovations in commodity spot prices and convenience yields (i.e, the assumption

that the risk-free rate is an autonomous process seems appropriate). The coefficient αX is significant

across all commodities. It is high and positive for oil and copper which is consistent with the theory
26
Further details are provided in Appendix 1.E.
27
Given the Gaussian nature of our model it is straightforward
 to calculate the exact moments for the state variables
 
{r, b
δ, X}. For the long-term spot price we use E exp(X) = exp E [X] + 21 VAR[X] .
1.3. Empirical implementation 20

of storage and indicates mean-reversion in spot prices under the risk-neutral measure. The estimated

αX is lower for silver and negligible for gold, which is evidence against this type of mean-reversion

in these commodities. The sensitivity of convenience yields to interest rates αr is significant and

positive across commodities, which is consistent with the theory of storage.28 Interestingly it is

higher for crude oil and copper than for gold and silver. Performing a likelihood ratio test to jointly

test for the significance of αr and αX , we find that they are highly significant for most commodities

(barely for gold at the 5% level - see table 1.3).

The significance of the risk-premia parameters varies across commodities, but there are some

consistent patterns.29 Most risk-premia coefficients related to the spot price (i.e., β0X , βXX ) are sig-

nificant. In contrast, risk-premia related to the interest rate dynamics are barely (or not) significant.

Further, βXX is always negative implying that risk-premia are time-varying and, in fact, negatively

correlated with the spot price. All spot commodity prices exhibit mean reversion under the physical

measure as evidenced by the positive coefficient of mean-reversion αX − βXX . When performing

a likelihood ratio test for the significance of time-variation in risk-premia (i.e., a joint test that all

coefficients in the β1Y matrix are zero) we find that they are jointly significant (see table 1.3).

Overall the results show that the maximal model, which allows convenience yields to be a func-

tion of the interest rate and spot price, associated with the more flexible time varying risk-premia

specification is a significant improvement over nested models proposed in the literature. The joint

likelihood ratio tests of table 1.3 suggest that allowing more general dynamics of the convenience

yield is the more important feature. This may be due to the fact that spot price dependence in conve-

nience yield results in mean-reversion under both the risk-neutral and historical measures, whereas

the time-variation in risk-premia only affects the strength of mean-reversion under the physical

measure.

We first provide more detailed discussions of the individual commodities, then summarize the

implications for the dynamics of convenience yields and the sources of mean-reversion in commod-
28
As described in the introduction, most theoretical models do not allow for stochastic interest rates. However, as-
suming that costs of holding inventory increase with interest rates suggests a negative correlation between inventory and
interest rates. We thus expect a positive relation between interest rates and convenience yield.
29
For simplicity, in the estimation results presented we dropped the time-varying risk-premia parameters that had a
t-ratio less than 1.0, which corresponds to a level of significance of 31.7%. This was the case of βXr for oil, copper and
silver and βXbδ for oil and gold.
1.3. Empirical implementation 21

ity spot prices as well as the evidence on model (mis-)specification.

Crude Oil

We find that the oil price has a significant positive effect on the convenience yield (αX = 0.248).

This implies strong mean reversion of log spot prices under the risk-neutral measure. Also, there is

evidence of negative correlation between risk-premia and spot prices. The parameter βXX is −0.498,

implying that the mean reversion under the physical measure is higher from the mean reversion

under the risk-neutral measure (κXP > κXQ ).30 The (historical) mean reversion in oil prices is due

to both, the convenience yield and the time-variation in risk-premia. The relation between the

convenience yield and interest rates is significant and positive (αr = 1.764) which is consistent with

the ‘prediction’ of theory of storage. All risk-neutral coefficients are significant for oil, except for

some correlations, indicating that three factors are necessary to capture the dynamics of oil futures

prices. The ‘idiosyncratic’ component of the convenience yield b

δ has high volatility σbδ = 0.384,

low persistence κbδ = 1.191 and is positively correlated with the spot price ρbδX = 0.795. While

this third factor is clearly a significant component of the convenience yield, it seems to be driven

by innovations that are correlated with the spot market and are short lived. The long-term maximal

convenience yield is 0.109 which is the highest among the commodities studied (see table 1.5). Also

from this table the estimate for the long-term spot price is 23.45 dollars per barrel.

Copper

Copper has a similar behaviour than crude oil. This is not surprising since both commodities share

the characteristic of being input for productive processes. We find a statistically significant positive

relation between the spot price of copper and its convenience yield (αX = 0.150). This implies

mean reversion in spot prices under the risk-neutral measure. We find a significant negative corre-

lation between risk-premia and spot prices (βXX = −0.859). This implies that the mean-reversion is

stronger under the historical measure than under the risk-neutral measure. Table 1.4 give the point

estimates of κXP = 1.009 vs. κXQ = 0.150 (in Table 1.2). The relation between convenience yields and

interest rates is positive and statistically significant as before. The idiosyncratic component of the
30
Recall from proposition 1.2 that κXQ = αX .
1.3. Empirical implementation 22

convenience yield b
δ is quite volatile σbδ = 0.178, not persistent κbδ = 1.048 and positively correlated

with the spot price ρbδX = 0.588. As for oil, convenience yield in the copper market is primarily

driven by the spot price itself and economic factors that are correlated with spot price innovation

and are short lived.31 Finally, table 1.5 gives a long-term mean for the convenience yield of 0.063

and long-term average copper price of 91.45 cents per pound.

Gold

We find that there is a negligible relation between the convenience yield and gold spot prices (αX =

0.000). This suggests that gold tends to be mean-averting under the risk-neutral measure. However,

the price of gold exhibits mean-reversion under the historical measure (κXP = 0.301), because of the

negative correlation between the time-varying risk-premia and the spot price (βXX = −0.301). Inter-

est rates seem to be more important in driving the convenience yield of gold than spot prices (αr =

0.332). The idiosyncratic factor has a small effect on the convenience yield. Its long-term mean is

small (θbδQ = −0.009), its volatility is very low σbδ = 0.015, somewhat persistent κbδQ = 0.392, and

not highly correlated with spot prices and interest rates (ρbδX = 0.295, ρrbδ = −0.047). Overall the

convenience yield of gold is quite small, not very variable and mainly driven by the interest rate.

Table 1.5 shows that the convenience yield has a long-term mean of 0.009 and an unconditional

standard deviation of only 0.010. The long-term price of gold is 390.91 dollars per troy ounce.

Silver

The dynamics of silver share some characteristics with the behaviour of gold. We find that silver

has a low mean-reversion degree under the risk-neutral measure (αX = 0.085). Silver prices exhibit

mean-reversion under the historical measure due to the negative correlation between spot prices and

risk-premia (βXX = −1.503). Interest rates have an effect on convenience yield similar to gold (αr =

0.326). The low volatility of the idiosyncratic factor (σbδ = 0.067) suggests that the convenience

yield of silver is mainly driven by spot prices and interest rates. The idiosyncratic factor b
δ follows a

mean-averting process under the risk neutral measure which, due to its magnitude (|κbδ | > |αX |), also
31
Of course the convenience yield is also affected by the interest rate through the parameter αr but to a lesser extent.
Even though, αr is greater than αX , recall from equation (1.11) that the effect in the convenience yields are through the
magnitude of αr r(t) and αX X(t).
1.3. Empirical implementation 23

induces mean-aversion in the convenience yield. Table 1.5 shows that the long-term convenience

yield is 0.002 and the unconditional standard deviation is 0.018. Finally, table 1.5 gives a long-term

average silver price of 477.99 cents per troy ounce.

Mis-specification

Since we estimate the parameters for each commodity separately, we obtain four different estimates

for interest rate parameters. In general, the estimates seem reasonable (e.g., in line with estimation

of single factor models found in the literature) and do not vary significantly across estimation except

for gold, which is weak evidence that the model correctly captures the relation between interest rates

and convenience yield and commodity process. Not surprisingly, the auto-correlation coefficient

for the term structure ‘measurement errors’ is quite high ≈ 0.99 indicating that at least a second

factor is needed to capture the dynamics of the term structure. This is well-known (Litterman

and Scheinkman (1991)), but our primary focus is to analyze the term structure of commodity

futures, and we expect an additional term structure factor to have only limited explanatory power for

commodity prices. More important for our study are the ‘measurement errors’ for the commodity

futures. The auto-correlation coefficients are lower than for interest rates (around 0.7), but very

significant. Figure 1.4 graphs time series of the pricing errors of some futures contract for the

four commodities. In table 1.6 we present some summary statistics about these pricing errors for

the maximal model. There does not seem to be a systematic bias in the fit of the model. Not

surprisingly, the analysis of the unconditional pricing errors (ut ) show that the model performs (in

terms of MSE) slightly less well with the two commodities that exhibit higher volatility (i.e., oil and

copper).

Inspection of the time series of future prices indicates that perhaps some of these errors are

attributable to the inability of the pure diffusion model to accommodate jumps. For example, gold

prices experienced a +25% jump in prices during September-October 1999. This jump followed

an announcement made by the European central banks, in response to increased pressures of gold

producers, to cut sales of gold reserves. Further, demand for gold at that time may have been fueled

by the Y2K uncertainty. Since our Gaussian model does not allow for jumps, some of the empirical

findings may be due to the model trying to accommodate for the presence of jumps in spot prices.
1.3. Empirical implementation 24

To make sure the presence of jumps in the spot time series does not affect our conclusions, we

re-estimate the model by allowing for jumps in the underlying spot price dynamics.

1.3.4 Estimation of the Jump component in commodity spot prices

We allow for jumps in commodity prices by considering the model introduced in proposition 1.2

where the dynamics of X(t) are modified as follows:

 
 1 X3  X3

dX(t) = r(t) − δ(t) − σX −
2 
(ϕi − 1)λi  dt + σX dZX (t) +
Q
νi (t)dNi (t) (1.36)
2 i=1 i=1

P
where Ni (t) = j 1τ j ≤t is the counting process associated with a sequence of stopping times
i

τ1i , τ2i , . . . generated by a standard Poisson process with Q-measure intensity λiQ (see Bremaud
j
(1981) for a rigorous exposition of point processes). The νi (τi ) ∀ j = 1, 2 . . . are i.i.d. random vari-

ables that are independent of the Poisson process and the Brownian motions. Further we assume ν1

is Gaussian with mean jump size m1 and standard deviation v1 , while ν2 and ν3 have constant jump

sizes m2 and m3 , respectively (i.e., v2 = v3 = 0).32 We denote the Laplace transform of the random
v2
i
variable νi by ϕi = emi + 2 i = 1, .., 3.

Applying Itô’s lemma to the spot price defined as before by S (t) = eX(t) we obtain:

dS (t) X
Q
= (r(t) − δ(t))dt + σ dZ (t) + dMiQ (t) (1.37)
S (t− ) X X
i

Rt
where MiQ (t) := 0
(eνi (s) − 1)dNi (s) − (ϕi − 1)λiQ t is a Q-Martingale. Thus

" #
Q dS (t)
E = (r(t) − δ(t))dt
S (t− )

and as before δ retains the interpretation of a ‘convenience’ yield that accrues to the holder of the

commodity similarly to a dividend yield.

To empirically implement the model we need a specification of risk-premia for both, Brownian

motion and Jump risk. We use the same ‘essentially affine’ risk-premium structure for Brownian
32
We found that allowing for more than one jump to have a stochastic jump size did not improve the likelihood and
thus choose to report only the constant jump size case.
1.3. Empirical implementation 25

motions as in equation (1.24). We also studied the risk-premia for the jump intensities. If jump

risk is systematic (e.g., if there is a common jump in the pricing kernel) then intensities need to be

risk-adjusted. If jump risk is non-systematic (for example because it is conditionally diversifiable as

in Jarrow, Lando and Yu (2000) or ‘extraneous’ as in Collin-Dufresne and Hugonnier (1999)) then

intensities are not risk-adjusted and remain the same under both measures. We empirically found

that allowing intensities to change did not improve the fit of the model significantly.33 We report the

case where the jump intensities are not risk-adjusted, i.e. Ni are Poisson processes with the same

intensities under both measures. Further, we assume jump size risk is not priced, i.e. that the jump

distribution is the same under both measures.34

Following Duffie and Kan (1996), futures prices may be computed in closed form for this Gaus-

sian jump-diffusion model. We report the closed form formulas in Appendix 1.G. We estimate the

model using maximum likelihood as exposed in Section 1.3.2. The only change is that the transi-

tion density of the log-spot price is no longer Gaussian. Following Ball and Torous (1983), Jorion

(1988), and Das (2002) we approximate the transition density by a mixture of Gaussian (the ap-

proximation would be exact if the time interval was infinitesimal). Several problems arise when

implementing this approach. Mainly, the likelihood function is unbounded if the model is estimated

without any restrictions. We use Honore’s (1998) approach to obtain consistent estimates of the pa-

rameters. More details about the estimation procedure and approximation to the likelihood function

are presented in Appendix 1.F.

Results for the parameter estimates are reported in table 1.7.35 For all commodities we find

significant evidence for the presence of a frequent stochastic jump component with mean m1 close

to zero. This reflects small variable frequent jumps in the spot price that are unaccounted for by the

pure diffusion model.36 For all commodities we also find evidence for the presence of a positive

and a negative less frequent jumps, except for gold where instead of a negative jump we find a

highly frequent small positive jump.37 For copper and silver the positive jump has a mean between
33
In fact, futures prices are very insensitive to the presence of these jumps (because they have almost zero mean and
futures prices are Q-expectations of the future spot price). As a result, most improvement in the likelihood is due to the
improvements in the P-measure distribution for this jump component. See the discussion below and Appendix 1.G.
34
Obviously, for the constant jump size cases this is a requirement since both measures must be equivalent.
35
For the results presented we keep jumps with parameter estimates that have a t-ratio greater than 1.
36
Small and frequent jumps may be suggestive of Levy processes, see e.g. Bakshi and Madan (2000).
37
Note however, that the average mean is negative and larger in magnitude than for the other commodities m1 =
1.3. Empirical implementation 26

7.3% and 10%, and occurs twice every two years on average. For gold this jump has a similar

mean m2 = 0.096, but occurs only once every six years. For crude oil this jump is not significant.

The negative jumps have different means and intensities across commodities. For crude oil the

jump size is m3 = −0.176 and it is very infrequent (once every six years on average). For copper

and silver these jumps have means of -8.3% and -11.5%, and they occur, on average, every five and

three years, respectively. In table 1.8 we carry out a likelihood ratio test for the hypothesis of having

no jumps which is clearly rejected. The inclusion of jumps appears especially significant for crude

oil and silver.

Overall however, our previous results seem mostly robust to the inclusion of jumps. When

we compare the parameter estimates to those obtained without jumps in table 1.2 we see that the

estimates of risk-neutral drift parameters are almost unchanged. Including jumps mainly affects

the estimates of the volatility coefficients and the risk-premia parameters. These results can be

explained by the fact that jumps in the spot price have little impact on the predicted cross-section

of futures prices. Indeed, we show in Appendix 1.G that futures prices are unchanged if αX = 0.

The intuition is that futures prices are martingales under the risk-neutral measure. The no-arbitrage

restriction on the risk-neutral drift of the spot price (i.e., equation 1.4) implies that jumps in the spot

price can only ‘matter’ if there is a common jump in the spot rate and/or the convenience yield.38

Further, we also show in the appendix that for the estimated jump intensity and jump distribution,

the impact of jumps on futures prices is negligible. However, accounting for jumps helps better

capture the historical measure dynamics of futures prices.

1.3.5 Restrictions on common pricing kernel dynamics

In perfectly integrated markets all commodities should be priced by the same pricing kernel. One

may thus wonder whether the specification of risk-premia proposed in Section 1.2 is consistent with

some arbitrage-free dynamics of a (common) pricing kernel, and whether these implied dynamics

are economically ‘reasonable.’ An alternative to our implementation might be to propose a joint

−0.017. The larger positive jump found for gold may be related to the special September-November 1999 period discussed
previously.
38
Hilliard and Reis (1998) for example, find that, in their model, jumps have no impact on futures prices. Their
convenience yield model is not maximal however.
1.3. Empirical implementation 27

specification of the dynamics of the pricing kernel and all commodity prices and perform a joint

estimation.39 The latter would seem particularly appropriate if commodity prices are largely driven

by a small set of common factors. A simple look at the correlation structure (table 1.9) shows that,

except for gold and silver, innovations in commodity prices exhibit low correlation. A principal

component analysis reveals that the factors driving oil and interest rates are distinct from factors

driving the metal prices (see table 1.10). Indeed, the first eigenvector loads almost exclusively on

oil, and the last eigenvector exclusively on the risk-free rate. Further, the eigenvalues corresponding

to the three metal factors are of the same order of magnitude (the first two metal factors repre-

sent around 16% of the total variance and the third about 4%), which suggests that there is not a

predominant common factor in the metal market.

This suggests that little is lost in performing the estimation separately for oil and metals. How-

ever, there may be some scope to perform a common estimation of the pricing kernel dynamics

and all three metal prices. Fortunately, it is possible to test whether the restrictions imposed by our

specification of risk-premia are significant.

Indeed, assume that there exists a filtered probability space (Ω, F , P) where the filtration F is

the natural filtration generated by two n-dimensional vectors of Brownian motions BP (t) and Z P (t).

Consider the following dynamics of log-spot commodity prices (Xti i = 1, . . . n) and the common

pricing kernel Mt .

dM(t) Xn
1
= −rdt − (β0i + β1i Xi (t))dBiP (t) (1.38)
M(t) i=1
σi

dXi (t) = µi (t)dt + σi dZiP (t) (1.39)

σ2
= (r − δi − i )dt + σi dZiQ (t) (1.40)
2

The last equality is the standard absence of arbitrage restriction, where the ZiQ (t) i = 1 . . . n are

Brownian motions under the risk-neutral measure Q, which is equivalent to P and defined by

dQ 
dP  = erT M(T ) 40
M(0) . Note that for simplicity of notation we assume here that the risk-free rate
FT
and the convenience yields are constant, but our argument extends straightforwardly to the more
39
We thank a referee for making this point.
40
To insure that this is an appropriate change of measure, we need to verify some additional regularity condition. In
this Gaussian framework, this follows straightforwardly from theorem 7.15 p. 279 in Liptser and Shiryaev (1974).
1.3. Empirical implementation 28

general case. By Girsanov’s theorem we have (defining dBi (t)dZ j (t) = ηi j dt) that

n Z
X t
1
ZiQ (t) = ZiP (t) + ηi j (β0 j + β1 j X j (s))ds ∀i = 1, . . . n (1.41)
j=1 0 σj

and thus the expected change in log-spot prices is given by:

σ2i X n
µi (t) = r − δi − + ηi j (β0 j + β1 j X j (t)) (1.42)
2 j=1

This framework allows us to see the implicit restrictions put by our specification of risk-premia on

the joint dynamics of the pricing kernel and spot prices. We are basically restricting the correlation

structure of the vector of Brownian motions. Specifically, comparing equations (1.24) and (1.41), we

see that our risk-premium specification of Section 1.2 is consistent with the common pricing kernel

model above if and only if ηi j = 0 ∀i , j. Fortunately we can test this restriction without resorting

to full-fledged joint estimation (which would be highly computationally intensive). Comparing

equations (1.39), (1.40) and (1.42) with proposition 1.2 and ??prop:rwdelxP) we see that the only

implication of the more general model specification (i.e., with ηi j , 0 for some i , j) for our data

set is that the expected return of commodity i (say oil) can depend on the log-price of commodity j

(say gold).41 We can easily test this by performing a Vector Auto Regression and doing a likelihood

ratio test to see whether allowing for cross-dependence in the expected changes of commodities are

significant.

In tables 1.11 we present the results for the following VAR: Xt = c + φXt−1 + µt where µt follows

an AR(1) process µt = ρ µt−1 + t and t ∼ N(0, Σ). The components of Xi (t) for i = 1, .., 5 are the

(log) prices for crude oil, copper, gold, silver, and the six-month interest rate. Table 1.12 presents

the results for the same auto-regression but keeping only copper, gold and silver, since the factor

analysis above suggests that this is where a joint estimation should benefit most. The results show

that almost all of the off-diagonal terms φi j i , j are not significant. Further, in both cases the

likelihood ratio test cannot reject that all of the φi j with i , j are zero. In other words, the results

from the VAR strongly support our specification assumption that ηi j = 0 ∀i , j. Given the overall
41
Indeed, futures prices are unchanged since our specification does not restrict the risk-neutral dynamics relative to the
more general model.
1.3. Empirical implementation 29

low correlations between various commodities (i.e., dZi dZ j ) documented in table 1.9 this suggests

that there would be little gains to performing a joint estimation with all commodity prices at once

(this seems especially true, given the size of our data set).

1.3.6 Summary of the results

Implied convenience yields

In figure 1.5 we present the implied convenience yields for the four commodities. These graphs

were obtained using the estimated {r, b

δ, X} state variables and then calculating the implied conve-

nience yield for each time-series observation.42 The figure clearly distinguishes oil and copper

which have highly volatile implied convenience yields from silver and gold whose convenience

yields are close to zero and exhibit little variability. This is in part attributable to a higher standard

deviation of the spot commodity prices for oil and copper, as well as a higher volatility of the resid-

ual third factor, σbδ (see table 1.2). Table 1.5 confirms these results. Gold and silver have implied

convenience yields of about 1.7% and 0.2%, whereas copper and oil have convenience yields of

respectively 6.3% and 10.9%.

Sources of Mean-reversion: convenience yield and time-varying risk-premia

Overall our results suggest that the maximal convenience yield model improves upon all nested

specifications tested in the literature (such as the models studied by Schwartz (1997)). We find that

for all commodities level-dependence in convenience yield is significant, but it is higher for assets

that tend to be used as inputs to production, such as oil and copper. Time variation in risk-premia,

on the other hand, seems to be highest for assets which also may serve as a store of value and

thus, perhaps, resemble more financial assets, such as gold and silver. Our results show that both

convenience yields as justified by the the option/storage theoretic models (Litzenberger and Rabi-

nowitz (1995), Deaton and Laroque (1988), RSS 2000, Casassus, Collin-Dufresne and Routledge

(2003)), and time-varying risk-premia (e.g., Fama and French (1987, 1988)) contribute to explain-

ing mean-reversion in commodity prices with more or less impact depending on the nature of the

commodity.

Aside from their econometric interest, these results have also economic implications. In the
42
We have presented the four implied convenience yields with the same scale for comparison purposes.
1.4. Implication of mean-reversion for option pricing and value at risk. 30

following section we offer two simple applications that demonstrate the impact on valuation and

risk-management of ignoring the various sources of mean-reversion in commodity prices.

1.4 Implication of mean-reversion for option pricing and value at risk.

Schwartz (1997) states that the stochastic behavior of commodity prices may have important im-

plications for valuation of commodity related securities. We have documented that allowing con-

venience yields to be a function of spot prices and interest rates, and allowing risk-premia to be

time-varying better captures dynamics of commodity futures prices. Both features have largely

been ignored by previous pricing models. We focus on two simple examples to document how

significant the implications are for economic applications, namely (i) valuation of options, and (ii)

computation of VAR.

1.4.1 Option Pricing

As discussed previously, allowing convenience yields to be a function of the spot price, effectively

induces mean-reversion under the risk-neutral measure. Since the latter ‘matters’ for valuation, we

expect this to affect the cross section of option prices. Using the Fourier inversion approach in-

troduced by Heston (1993) we can compute in closed-form (up to a Fourier transform inversion)

European option values within our three-factor affine framework.43 We compute the option value

using two sets of parameters. First, we use the parameters corresponding to our ‘maximal’ conve-

nience yield model as given in table 1.2. Second, we re-estimate the parameters assuming one were

to ignore the level dependence in convenience yields (i.e., setting αX = αr = 0 in our model). This

corresponds basically to estimating model 3 of Schwartz (1997), but with a more flexible specifica-

tion of risk-premia and an AR(1) representation of the measurement errors.

The option prices obtained for each commodity with the two sets of parameters are shown in

figure 1.6. For each commodity we value European call options written on a unit of the asset with

a maturity of two years, and strike prices of $25 per barrel for oil, 100 cents per pound for copper,

$350 per troy ounce for gold and 550 cents per troy ounce for silver. The figure shows that the
43
See Duffie, Pan and Singleton (2000) for a thorough exposition of this option valuation approach.
1.4. Implication of mean-reversion for option pricing and value at risk. 31

difference can become quite important for oil, copper and silver, especially for options that are at

and in the money. For gold the difference in option values is small indicating that the coefficient αX

while statistically significant has a small economic impact.

For commodities with a positive relation between convenience yields and spot prices (i.e. crude

oil, copper and silver) ignoring the level dependence in convenience yields, leads to overestimation

of call option values. The direction of the bias for a positive αX is expected for two reasons. First,

the ‘maximal’ convenience yield model effectively introduces mean-reversion under the risk-neutral

measure and thus leads to reduced term volatility which reduces option prices. Second, a positive αX

implies a convenience yield which is stochastic and increasing in the spot price. This contributes to

decreasing call option prices for in the money options.44 For commodities with a negative relation

between convenience yields and spot prices (gold), we find the opposite bias, i.e., in the money call

options are underestimated with existing models.

The size of the error for the estimated parameters is quite dramatic. For example, the error is

close to 30% for in the money options written on crude oil. This suggests that appropriately model-

ing the dynamics of convenience yields may have important consequences on investment decisions

within real-option models. For natural resource investments related to commodities like crude oil

and copper, our results suggest that in a typical ‘waiting to invest’ (Majd and Pindyck (1987))

framework that ignores the spot price dependence in convenience yields, the optimal investment

rule would have a tendency to postpone investment sub-optimally.

1.4.2 Value at Risk

Our results indicate that for metals such as gold, silver and copper a substantial part of the mean-

reversion in spot prices is due to negative covariation between spot prices and risk-premia. This

implies that ignoring the time-variation of risk-premia during estimation will lead to mis-estimation

of the holding period return distribution of commodities. To illustrate the latter for each commodity

we compute the Value at Risk (VAR) of the five year return on a portfolio invested in one unit of the

commodity. By definition, the VAR is computed from the total (i.e., ‘cum dividend’) return under
44
This intuition is analogue to a call option on a dividend-paying stock in the Black-Scholes world. The higher the
dividend rate, the lower the price of the option.
1.5. Conclusions 32

the historical measure. It thus allows us to focus on the effect of time-varying risk-premia.

As before we compute the VAR for two different sets of parameters. One corresponds to our

maximal model, i.e. table 1.2. For the other we re-estimate the three factor model, but constrain-

ing risk-premia to be constant (i.e., setting βXX = βrr = βbδbδ = βXr = βXbδ = 0). We note that in

both cases the return on the commodity in our maximal model of proposition 1.1 and 1.2 is Gaus-

sian, which is consistent with the usual assumptions of the VAR framework. Figure 1.7 shows

the five year holding period return distributions corresponding to the two sets of parameters, and

graphs the corresponding 5% VAR.45 The figure clearly shows that accounting for time-variation in

risk-premia has a substantial impact on the dispersion of the holding period return, especially for

copper, gold and silver. The return distribution is more spread out for the case with a constant risk-

premia. Consequently, the VAR (the potential loss corresponding to a 5% tail event) for the three

metals more than doubles when the distribution is estimated without accounting for time variation

in risk-premia. This suggests that economic capital required to cover holdings in precious metals

are significantly reduced when appropriately taking into account the dynamics of risk-premia. The

same figure shows that the VAR for oil is less significantly affected.

1.5 Conclusions

We develop a three-factor model of commodity spot prices, convenience yields and interest rates,

which extends previous research in two ways. First, the model nests several (e.g., Brennan (1991),

Gibson and Schwartz (1990), Schwartz (1997), Ross (1997), Schwartz and Smith (2000)) proposed

specifications. Second, it allows for time-varying risk-premia. We show that previous models have

implicitly imposed unnecessary restrictions on the unconditional correlation structure of commodity

prices, convenience yields and interest rates. In particular, the present model allows for convenience

yields to be a function of spot commodity prices, which leads to mean-reversion in spot prices.

Mean-reversion in spot prices can also be generated by negative correlation between risk-premia and

spot prices. The former affects the risk-neutral dynamics of commodity prices, i.e. the cross-section

of futures prices. The latter affects only the historical measure dynamics of prices, i.e. the time series
45
As before, to facilitate the comparison across commodities, we keep the same scale for all distributions.
1.5. Conclusions 33

of futures prices. Both components can thus be identified with panel data on futures prices. Using

data on crude oil, copper, gold and silver commodity futures, we empirically estimate the model

using maximum likelihood. We find both features of the model to be economically and empirically

significant. In particular, we find strong evidence for spot-price level dependence in convenience

yields of crude oil and copper, which implies mean-reversion in spot prices under the risk-neutral

measure, and is consistent with the “theory of storage.” We find evidence for time-varying risk-

premia, which implies mean-reversion of commodity prices under the physical measure albeit with

different strength and long-term mean. We also document the presence of a jump component in

commodity prices.

The results suggest that the relative contribution of both effects to mean reversion (level depen-

dent convenience yield vs. time-varying risk-premia) depends on the nature of the commodity, and,

in particular, on the extent to which the commodity may serve as an input to production (e.g., a

consumption good) versus as a store of value (e.g., a financial asset). We find that for metals like

gold and silver, negative correlation between risk-premia and spot prices explains most of the mean

reversion, whereas for oil and copper some of the mean-reversion in spot prices is attributable also

to convenience yields.46 The analysis of various examples suggests that disentangling the sources

of mean-reversion and careful modeling of the dynamics of the convenience yield can have a sub-

stantial impact on (real) option valuation, investment decisions and risk-management.

46
Of course, it is not clear that the primary use of gold and silver is as a ‘store of value.’ Indeed, both have many
industrial uses. Further theoretical research seems warranted to better understand these cross-sectional differences across
commodities. Casassus, Collin-Dufresne and Routledge (2003) provide a first step in that direction.
1.6. Tables 34

1.6 Tables
1.6. Tables 35

Table 1.1: Statistics of Crude Oil, Copper, Gold and Silver Futures Contracts
Statistics for weekly observations of crude oil, copper, gold and silver futures contracts between 1/2/1990 and 8/25/2003.
Oil prices are in dollars per barrel, copper prices are in cents per pound, gold prices are in dollars per troy ounce, silver
prices are in cents per troy ounce and maturities are in years.

Crude Oil Data Copper Data

Contract Obs. Mean Price Maturity Contract Obs. Mean Price Maturity
(Std. Error) (Std. Error) (Std. Error) (Std. Error)
F01 713 21.98 0.037 F01 713 94.32 0.037
(5.47) (0.024) (21.05) (0.024)
F03 713 21.60 0.203 F03 713 93.65 0.203
(4.87) (0.025) (19.66) (0.024)
F06 713 21.10 0.454 F06 713 92.67 0.454
(4.16) (0.025) (17.66) (0.025)
F09 713 20.72 0.704 F09 713 91.84 0.704
(3.65) (0.025) (15.98) (0.025)
F12 713 20.45 0.954 F12 713 91.16 0.954
(3.26) (0.027) (14.60) (0.028)
F15 713 20.25 1.205 F15 713 90.64 1.202
(2.96) (0.027) (13.54) (0.045)
F18 713 20.11 1.451 F18 713 90.32 1.450
(2.71) (0.036) (12.76) (0.048)
F24 581 19.82 1.946 F24 223 79.51 1.952
(2.51) (0.072) (5.19) (0.024)
F30 581 19.76 2.434
(2.27) (0.127)
F36 581 19.74 2.830
(2.11) (0.198)
Gold Data Silver Data
Contract Obs. Mean Price Maturity Contract Obs. Mean Price Maturity
(Std. Error) (Std. Error) (Std. Error) (Std. Error)
F01 713 336.96 0.038 F01 713 478.02 0.038
(44.24) (0.025) (57.12) (0.026)
F03 713 338.73 0.204 F03 713 480.96 0.203
(44.57) (0.031) (57.46) (0.028)
F06 713 342.08 0.456 F06 713 486.16 0.445
(45.47) (0.062) (57.54) (0.063)
F09 713 345.40 0.705 F09 713 491.13 0.697
(46.47) (0.063) (57.57) (0.064)
F12 713 348.77 0.954 F12 713 495.96 0.944
(47.55) (0.064) (57.82) (0.062)
F15 713 352.29 1.206 F15 713 500.96 1.196
(48.60) (0.066) (58.24) (0.064)
F18 713 355.91 1.454 F18 713 506.15 1.444
(49.84) (0.068) (58.70) (0.069)
F24 713 362.57 1.895 F24 713 515.00 1.877
(51.85) (0.110) (59.96) (0.109)
F30 666 365.50 2.427 F30 565 532.06 2.433
(51.89) (0.152) (60.41) (0.174)
F36 666 373.25 2.917 F36 565 541.06 2.919
(54.23) (0.157) (65.19) (0.196)
F48 666 389.43 3.901 F48 565 561.50 3.908
(59.06) (0.161) (77.37) (0.188)
1.6. Tables 36

Table 1.2: Maximum-Likelihood Parameter Estimates for the Maximal Model

Maximum-likelihood parameter estimates for the maximal model for crude oil, copper, gold and silver weekly prices and
interest rate data between 1/2/1990 to 8/25/2003.

Crude Oil Copper Gold Silver

Parameter Estimate Estimate Estimate Estimate
(Std. Error) (Std. Error) (Std. Error) (Std. Error)
κrQ 0.027 0.035 0.032 0.033
(0.007) (0.007) (0.007) (0.007)
κbQ 1.191 1.048 0.392 -0.157
δ
(0.023) (0.038) (0.035) (0.008)
αr 1.764 0.829 0.332 0.326
(0.083) (0.097) (0.046) (0.101)
αX 0.248 0.150 0.000 0.085
(0.010) (0.015) (0.000) (0.007)
θrQ 0.057 0.118 0.095 0.111
(0.030) (0.015) (0.017) (0.016)
θbQ -0.839 -0.673 -0.009 -0.530
δ
(0.033) (0.063) (0.003) (0.043)
β0r 0.003 0.002 0.002 0.000
(0.009) (0.012) (0.009) (0.012)
β0bδ -1.047 -0.435 0.004 -0.510
(0.367) (0.348) (0.005) (0.178)
β0X 1.711 5.142 1.858 12.466
(0.964) (1.956) (1.539) (3.381)
βrr -0.137 -0.173 -0.140 -0.113
(0.165) (0.165) (0.162) (0.226)
βbδbδ -1.660 -0.749 -1.143 -0.962
(0.480) (0.531) (0.455) (0.322)
βXr -2.857
(2.452)
βXbδ 1.919 6.051
(0.929) (3.733)
βXX -0.498 -0.859 -0.301 -1.503
(0.313) (0.338) (0.271) (0.471)
σr 0.009 0.009 0.009 0.009
(0.000) (0.000) (0.000) (0.000)
σbδ 0.384 0.178 0.015 0.019
(0.013) (0.006) (0.001) (0.001)
σX 0.397 0.228 0.132 0.223
(0.012) (0.006) (0.004) (0.006)
ρbδX 0.795 0.588 0.295 -0.422
(0.015) (0.035) (0.034) (0.061)
ρrbδ -0.009 0.107 -0.047 0.019
(0.031) (0.038) (0.051) (0.087)
ρrX 0.051 0.143 -0.061 0.065
(0.033) (0.037) (0.037) (0.044)
ρF 0.796 0.699 0.813 0.800
(0.011) (0.013) (0.011) (0.010)
ρP 0.993 0.986 0.989 0.987
(0.003) (0.003) (0.003) (0.003)
Log-likelihood 57153.5 51761.5 72899.8 66114.9
1.6. Tables 37

Table 1.3: Likelihood Ratio Tests contracts

Likelihood ratios for the maximal model and (a) a model where the convenience yield is not affected by interest rates and
spot prices, (b) a model with constant risk premia, and (c) both constraints together. The 5% significance level for these
constraints are given by Prob{χ22 ≥ 5.99} = 0.05, Prob{χ25 ≥ 11.07} = 0.05 and Prob{χ27 ≥ 14.07} = 0.05, respectively.

Restriction Crude Oil Copper Gold Silver

αr = αX = 0 a
1047.20 312.78 5.60 66.13
β1Y = 0 b 13.16 12.06 20.01 21.27
αr = αX = 0 and β1Y = 0 c 1057.92 328.02 23.96 87.81

Table 1.4: Maximal Model Estimates of Historical Parameters

Maximal model estimates of historical parameters for crude oil, copper, gold and silver using weekly prices and interest
rate data between 1/2/1990 to 8/25/2003.

Parameter Crude Oil Copper Gold Silver

Estimate Estimate Estimate Estimate
κrP 0.165 0.208 0.172 0.147
κbP 2.850 1.797 1.535 0.805
δ
κXr
P
0.764 -0.171 2.189 -0.674
κPb 1.000 -0.919 1.000 -5.051
Xδ
κXP 0.746 1.009 0.301 1.588
θrP 0.029 0.030 0.028 0.023
θbP -0.718 -0.635 0.000 -0.529
δ
θXP 3.120 4.498 5.946 6.159
1.6. Tables 38

Table 1.5: Unconditional First and Second Moments with the Maximal Model Estimates
Unconditional first and second moment estimates from the maximal model of δ(t) and X(t) for crude oil, copper, gold and
silver using weekly prices and interest rate data between 1/2/1990 to 8/25/2003.

Unconditional Crude Oil Copper Gold Silver

Moments
E [δ] 0.109 0.063 0.009 0.002
Stdev (δ) 0.210 0.116 0.010 0.018
E [X] 3.120 4.498 5.946 6.159
Stdev (X) 0.264 0.190 0.212 0.122
Corr (δ, X) 0.790 0.776 -0.246 0.475
h i
E eX 23.45 91.45 390.91 476.46

Table 1.6: Statistics of Pricing Errors using the Maximal Model

Mis-specification statistics for the error terms of the F01 futures contract using the maximal model. The error term ut is
g where F01
log(F01) − log(F01), g is the estimated F01 futures contract. The statistics are for the four commodities using
data between 1/2/1990 and 8/25/2003.

Error Crude Oil Copper Gold Silver

Meanut 0.000 -0.001 0.000 0.000
Stdevut 0.012 0.008 0.001 0.003
Maxut 0.063 0.051 0.005 0.008
1.6. Tables 39

Table 1.7: Maximum-Likelihood Parameter Estimates for the Triple-Jump Model

Maximum-likelihood parameter estimates for the triple-jump model for crude oil, copper, gold and silver weekly prices
and interest rate data between 1/2/1990 to 8/25/2003.

Crude Oil Copper Gold Silver

Parameter Estimate Estimate Estimate Estimate
(Std. Error) (Std. Error) (Std. Error) (Std. Error)
κrQ 0.027 0.035 0.031 0.034
(0.007) (0.007) (0.007) (0.007)
κbQ
1.190 1.046 0.384 -0.173
δ
(0.023) (0.038) (0.035) (0.007)
αr 1.772 0.845 0.324 0.264
(0.083) (0.095) (0.049) (0.069)
αX 0.250 0.145 0.003 0.131
(0.010) (0.015) (0.003) (0.006)
θrQ
0.056 0.118 0.096 0.116
(0.028) (0.015) (0.017) (0.016)
θbQ
-0.840 -0.656 -0.025 -0.815
δ
(0.033) (0.062) (0.015) (0.037)
β0r 0.004 0.002 0.002 -0.001
(0.009) (0.009) (0.009) (0.011)
β0bδ -1.245 -0.411 -0.010 -1.029
(0.324) (0.333) (0.016) (0.525)
β0X 0.024 3.063 -1.100 12.070
(0.489) (1.920) (1.309) (6.272)
βrr -0.146 -0.170 -0.137 -0.111
(0.169) (0.167) (0.172) (0.209)
βbδbδ -1.914 -0.733 -1.003 -1.265
(0.405) (0.522) (0.449) (0.671)
βXr -5.677
(2.413)
βXbδ 1.365 6.767
(0.846) (5.960)
βXX 0.054 -0.483 0.233 -1.065
(0.165) (0.340) (0.232) (0.395)
σr 0.009 0.009 0.009 0.009
(0.000) (0.000) (0.000) (0.000)
σbδ 0.384 0.178 0.015 0.024
(0.013) (0.006) (0.001) (0.001)
σX 0.319 0.173 0.080 0.208
(0.014) (0.019) (0.007) (0.008)
ρbδX 0.957 0.810 0.363 -0.957
(0.017) (0.081) (0.067) (0.024)
ρrbδ -0.009 0.106 -0.039 0.007
(0.039) (0.038) (0.054) (0.041)
ρrX 0.060 0.184 -0.120 0.046
(0.043) (0.048) (0.053) (0.041)
m1 0.001 -0.002 -0.017 0.000
(0.001) (0.003) (0.014) (0.002)
v1 0.025 0.017 0.019 0.016
(0.006) (0.012) (0.012) (0.006)
λ1 65.126 63.829 8.800 49.869
(19.231) (51.522) (6.115) (21.485)
m2 0.073 0.096 0.100
(0.013) (0.009) (0.010)
λ2 0.674 0.156 0.755
(0.541) (0.113) (0.295)
m3 -0.176 -0.083 0.022 -0.115
(0.015) (0.021) (0.003) (0.008)
λ3 0.154 0.219 8.063 0.369
(0.093) (0.317) (3.184) (0.187)
ρF 0.796 0.699 0.813 0.811
(0.011) (0.013) (0.011) (0.010)
ρP 0.993 0.986 0.989 0.986
(0.003) (0.003) (0.003) (0.003)
Log-likelihood 57210.5 51782.4 72952.2 66216.7
1.6. Tables 40

Table 1.8: Likelihood Ratio Tests for the Triple-Jump Model

Likelihood ratio for the triple-jump model and the maximal model. The 5% significance level for the jumps constraints
is given by Prob{χ27 ≥ 14.07} = 0.05.

Restriction Crude Oil Copper Gold Silver

mi = v i = λ i = 0 114.08 41.902 104.898 203.721

Table 1.9: Correlation Matrix of Log-Commodity Prices and Interest Rates

Correlation matrix for the weekly changes in log-commodity prices and interest rate data between 1/2/1990 to 8/25/2003.

crude oil copper gold silver interest rate

crude oil 1.00 0.06 0.18 0.14 0.04
copper 0.06 1.00 0.15 0.20 0.12
gold 0.18 0.15 1.00 0.60 -0.04
silver 0.14 0.20 0.60 1.00 -0.01
interest rate 0.04 0.12 -0.04 -0.01 1.00

Table 1.10: Principal Component Analysis of Log-Commodity Prices and Interest Rates
Principal component decomposition for the weekly changes in log-prices and interest rate data between 1/2/1990 to
8/25/2003.

PC(1) PC(2) PC(3) PC(4) PC(5)

crude oil 0.99 -0.16 -0.05 0.03 0.00
copper 0.07 0.63 -0.77 0.00 0.00
gold 0.08 0.30 0.25 -0.92 0.00
silver 0.13 0.70 0.58 0.40 0.00
interest rate 0.00 0.00 0.00 0.00 -1.00
Eigenvalue 0.0030 0.0012 0.0008 0.0002 1.05E-06
% explained 58.11% 22.93% 15.43% 3.51% 0.02%
1.6. Tables 41

Table 1.11: Vector Autoregression to Test Restrictions in Common Pricing Kernel

Vector autoregression (VAR) to test the restrictions imposed by our risk-premium specification on the correlation structure
(dBi (t)dB j (t)). The VAR is Xt = c + φXt−1 + µt where µt follows an AR(1) process µt = ρ µt−1 + t and t ∼ N(0, Σ).
Here Xi (t) for i = 1, .., 4 are the (log) prices for crude oil, copper, gold and silver, respectively. X5 (t) is the time series for
the six-month interest rate. The results are for weekly prices and interest rate data between 1/2/1990 to 8/25/2003. The
likelihood ratio tests the hypothesis that all off-diagonal terms of the φ matrix are zero. The 5% significance level for this
constraint is given by Prob{χ220 ≥ 31.41} = 0.05.

Unrestricted Correlation Structure Restricted Correlation Structure

Parameter Estimate (Std. Error) Estimate (Std. Error)
c1 0.123 (0.076) 0.046 (0.023)
c2 -0.058 (0.111) 0.038 (0.032)
c3 0.141 (0.085) 0.038 (0.025)
c4 0.341 (0.072) 0.140 (0.048)
c5 -0.011 (0.003) 0.000 (0.000)
φ11 0.982 (0.008) 0.985 (0.007)
φ12 0.017 (0.018)
φ13 -0.031 (0.023)
φ14 0.007 (0.011)
φ15 -0.121 (0.247)
φ21 0.000 (0.011)
φ22 0.961 (0.012) 0.992 (0.007)
φ23 0.043 (0.014)
φ24 -0.005 (0.016)
φ25 0.229 (0.166)
φ31 -0.003 (0.008)
φ32 0.007 (0.013)
φ33 0.983 (0.013) 0.993 (0.004)
φ34 -0.010 (0.011)
φ35 -0.086 (0.164)
φ41 -0.011 (0.009)
φ42 -0.004 (0.006)
φ43 -0.007 (0.008)
φ44 0.960 (0.011) 0.977 (0.008)
φ45 0.015 (0.036)
φ51 0.000 (0.000)
φ52 0.000 (0.000)
φ53 0.001 (0.000)
φ54 0.001 (0.000)
φ55 0.995 (0.003) 0.998 (0.002)
ρ1 -0.155 (0.046) -0.154 (0.038)
ρ2 -0.031 (0.048) -0.040 (0.061)
ρ3 0.035 (0.066) 0.032 (0.027)
ρ4 0.025 (0.041) 0.020 (0.032)
ρ5 -0.021 (0.031) 0.013 (0.021)
Log-likelihood 18.57 18.53
LR test 0.08
1.6. Tables 42

Table 1.12: VAR to Test Restrictions in Common Pricing Kernel of Copper, Gold and Silver
Vector autoregression (VAR) to test the restrictions imposed by our risk-premium specification on the correlation structure
(dBi (t)dB j (t)). The VAR is Xt = c + φXt−1 + µt where µt follows an AR(1) process µt = ρ µt−1 + t and t ∼ N(0, Σ). Here
Xi (t) for i = 2, .., 4 are the (log) prices for copper, gold and silver, respectively. The results are for weekly prices and
interest rate data between 1/2/1990 to 8/25/2003. The likelihood ratio tests the hypothesis that all off-diagonal terms of
the φ matrix are zero. The 5% significance level for this constraint is given by Prob{χ26 ≥ 12.59} = 0.05.

Unrestricted Correlation Structure Restricted Correlation Structure

Parameter Estimate (Std. Error) Estimate (Std. Error)
c2 -0.100 (0.083) 0.037 (0.033)
c3 0.138 (0.057) 0.044 (0.023)
c4 0.247 (0.088) 0.136 (0.048)
φ22 0.978 (0.008) 0.992 (0.007)
φ23 0.030 (0.013)
φ24 0.005 (0.009)
φ32 0.001 (0.005)
φ33 0.989 (0.008) 0.992 (0.004)
φ34 -0.012 (0.007)
φ42 -0.004 (0.007)
φ43 -0.004 (0.012)
φ44 0.967 (0.010) 0.978 (0.008)
ρ2 -0.043 (0.039) -0.043 (0.038)
ρ3 0.039 (0.034) 0.031 (0.031)
ρ4 0.022 (0.027) 0.019 (0.059)
Log-likelihood 9.72 9.71
LR test 0.02
1.7. Figures 43

1.7 Figures
 45 150
1.7. Figures

35 125

F01 F01
25 100
F18 F18

15 75

Futures Prices (Copper)

Futures Prices (Crude Oil)

5 50
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

500 750

650

400
550
F01 F01
F18 F18
450
300

Futures Prices (Gold)

350
Futures Prices (Silver)

200 250
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

Figure 1.1: Futures prices. F01 and F18 futures contracts on crude oil, copper, gold and silver between 1/2/1990 to 8/25/2003. Oil prices are in
dollars per barrel, copper prices are in cents per pound, gold prices are in dollars per troy ounce and silver prices are in cents per troy ounce.
44
 45 150
1.7. Figures

35 125

25 100

15 75

Futures Curves (Copper)

Futures Curves (Crude Oil)

5 50
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2
Maturity (years) Maturity (years)

500 800

700

400
600

500
300

400

Futures Curves (Gold)

Futures Curves (Silver)

200 300
0 1 2 3 4 0 1 2 3 4
Maturity (years) Maturity (years)

Figure 1.2: Futures curves. Monthly term structures of futures prices on crude oil, copper, gold and silver between 1/2/1990 to 8/25/2003. Oil
prices are in dollars per barrel, copper prices are in cents per pound, gold prices are in dollars per troy ounce and silver prices are in cents per troy
45

ounce.
 1.7. Figures

10.0

8.0

6.0

6-month
60-month

4.0

Interest Rates
2.0

0.0
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

Figure 1.3: Interest rates. 6-month and 60-month interest rates from constant maturity treasury bills between 1/2/1990 to 8/25/2003.
46
 0.10 0.10
1.7. Figures

0.05 0.05

0.00 0.00

-0.05 -0.05

Measurement Errors (Copper)

Measurement Errors (Crude Oil)

-0.10 -0.10
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

0.10 0.10

0.05 0.05

0.00 0.00

-0.05 -0.05

Measurement Errors (Gold)

Measurement Errors (Silver)

-0.10 -0.10
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

Figure 1.4: Pricing errors. Difference between true and estimated (log) futures prices using the maximal model for the period between 1/2/1990
to 8/25/2003. The thick line and the continuous thin line correspond to the F01 and F18 contracts, respectively. The discontinuous line corresponds
47

to the F36, F12, F42 and F24 for crude oil, copper, gold and silver, respectively.
 0.75 0.75
1.7. Figures

0.50 0.50

0.25 0.25

0.00 0.00

-0.25 -0.25

Convenience Yield (Oil)

Convenience Yield (Copper)
-0.50 -0.50
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

0.75 0.75

0.50 0.50

0.25 0.25

0.00 0.00

-0.25 -0.25

Convenience Yield (Gold)

Convenience Yield (Silver)

-0.50 -0.50
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

Figure 1.5: Maximal convenience yield. Implied convenience yield from the maximal model for crude oil, copper, gold and silver between
1/2/1990 to 8/25/2003.
48
 20 100
1.7. Figures

80
15

60
αr = αX = 0 αr = αX = 0
10
Maximal Maximal
40

Call Option (Copper)

20

Call Option (Crude Oil)

0 0
0 10 20 30 40 50 0 50 100 150 200
Spot Price Spot Price

500 600

500
400

400
300
αr = αX = 0 αr = αX = 0
300
Maximal Maximal
200
200

Call Option (Gold)

Call Option (Silver)

100
100

0 0
0 200 400 600 800 0 200 400 600 800 1000
Spot Price Spot Price

Figure 1.6: Option prices. Two-year maturity European Call Option prices using the maximal model. The strike price for oil is $25 per barrel,
for copper is 100 cents per pound, for gold is $350 per troy ounce and for silver is 550 cents per troy ounce. Spot and options prices are in the
49

same unit as strike prices. Each line corresponds to a different set of parameters. The bold line is corresponds to the maximal model, while the
thin line is assuming αr = αX = 0 (and re-estimation of parameters).
 1.7. Figures

β1Y = 0 β1Y = 0
Maximal Maximal

Likelihood (Copper)

Likelihood (Crude Oil)

-30% -20% -10% VAR 0% 10%
VAR 20% 30% -30% -20% VAR
-10% VAR
0% 10% 20% 30%

β1Y = 0 β1Y = 0
Maximal Maximal

Likelihood (Gold)
Likelihood (Silver)

-30% -20% VAR

-10% VAR
0% 10% 20% 30% -30% -20% VAR -10% VAR
0% 10% 20% 30%

Figure 1.7: Value at Risk. Distribution of returns and Value at Risk for holding one unit of commodity for 5 years using the maximal model.
Value at Risk is calculated at a 5% significance level. The two distributions correspond to different set of parameters. The bold line is using
50

copper estimates from the maximal model, while the thin line is assuming constant risk premia (and re-estimation of parameters).
Appendix 1.A. Closed-form solution for futures prices 51

Appendix 1.A Closed-form solution for futures prices

The value of a futures contract with maturity τ is given by:

h i
F(Y, τ) = exp AF (τ) + BF (τ)T Y (1.A1)

where the closed-form solution for AF (τ) and BF (τ) are:

Q Q
1 2 1 − e−2τκ11 1  2 
2 1−e
−2τκ22
AF (τ) = φ0 + M11 Q
+ M 12
+ M22 Q
2 2κ11 2 2κ22
Q
1 2 2

2 1−e
−2τκ33
+ M13 + M23 + M33 Q
2 2κ33
 Q Q
  Q Q

1 − e−τ κ11 +κ22 1 − e−τ κ11 +κ33
+M11 M12 Q Q
+ M11 M13 Q Q
κ11 + κ22 κ11 + κ33
 Q Q
−τ κ22 +κ33

1−e
+ M12 M13 + M22 M23 Q Q
(1.A2)
κ22 + κ33

  
  
M13   e−τκ11
Q
 M11 M12
  
  
BF (τ) =  0 M23   e−τκ22 
Q
M22  (1.A3)
   
 
M33   e−τκ33 
Q
0 0

with

M11 = φ1 + α1 φ2 + α2 φ3 + α3 φ3 (1.A4)


M12 = −α1 φ2 + α2 φ3 (1.A5)

M13 = −α3 φ3 (1.A6)

M22 = φ2 + α2 φ3 (1.A7)

M23 = −α2 φ3 (1.A8)

M33 = φ3 (1.A9)
Appendix 1.B. Closed-form solution for zero-coupon bonds 52

and

Q
κ21
α1 = Q Q
(1.A10)
κ11 − κ22
Q
κ32
α2 = Q Q
(1.A11)
κ22 − κ33
Q Q Q
κ31 κ21 κ32
α3 = Q Q
− Q Q Q Q
(1.A12)
κ11 − κ33 κ11 − κ33 κ22 − κ33

Appendix 1.B Closed-form solution for zero-coupon bonds

The value of a zero-coupon bond with maturity τ is given by:

 
P(Y1 , τ) = exp AP (τ) + BP (τ)Y1 (1.B1)

where the closed-form solution for AP (τ) and BP (τ) are:

  2   2 Q
 1  ψ1    ψ1  1 − e−τκ11
AP (τ) = − ψ0 −  Q   τ −  Q  Q
2 κ11 κ11 κ11
 2 Q
1  ψ1  1 − e−2τκ11
+  Q  Q
(1.B2)
2 κ11 2κ11
Q
1 − e−τκ11
BP (τ) = −ψ1 Q
(1.B3)
κ11

Appendix 1.C The {r, δ, X} representation

We apply an invariant transformation to the canonical base to get the economic representation

{r, δ, X} (see Dai and Singleton (2000)). This transformation rotates the state variables, but all the

initial properties of the model all maintained, i.e. the resulting model is a three-factor Gaussian

model that is maximal. We have the transformations for X(t), r(t) and δ(t) from equations (1.1), (1.5)

and (1.6), respectively.

δ(t) = η0 + η>
Y
Y(t) (1.C1)
Appendix 1.C. The {r, δ, X} representation 53

where

1 >
η0 = ψ0 − φ φ (1.C2)
2 Y Y
   
   Q Q Q 
 η1   ψ1 + κ11 φ1 + κ21 φ2 + κ31 φ3 

  
ηY =  η2  =  Q
κ22 Q
φ2 + κ32 φ3 
 (1.C3)
   
   
η3 κQ φ33 3

We define the transformed state vector W > (t) = (r(t), δ(t), X(t)). The linear transformation in matrix

form is

W(t) = ϑ + LY(t) (1.C4)

where Y(t) follows the process in (1.2). The matrices for the linear transformations are:

   
   
 ψ0   ψ1 0 0 

   
ϑ =  η0  and L =  η1 η2 η3  (1.C5)
   
   
φ0  φ1 φ2 φ3 

From equation (1.C4) and Ito’s lemma we have

dW(t) = Lκ Q L−1 (ϑ − W(t))dt + LdZ Q (t) (1.C6)

The mean-reversion and long-run parameters under the equivalent martingale measure are given by
h i
Q
κrQ = κ11 , κδrQ = − Lκ Q L−1 , κδQ = −κ22
Q Q
− κ33 Q = κ Q κ Q , θ Q = ψ and κ Q = η> κ Q L−1 ϑ. This
, κδX 22 33 r 0 δ0 Y
21

transformation was done under the equivalent martingale measure. Using the specification of the

risk premia in equation (1.22) and equation (1.C4) we get the rotation under the physical measure

dW(t) = Lκ Q L−1 (ϑ − W(t))dt + LdZ(t) + L(β0Y + β1Y Y(t))dt

= Lκ Q L−1 (ϑ − W(t))dt + LdZ(t) + (Lβ0Y − Lβ1Y L−1 ϑ + Lβ1Y L−1 W(t))dt

Appendix 1.D. The {r, b
δ, X} representation 54

The risk-premia parameters for the {r, δ, X} representation are

   
   
 β0r   βrr βrδ βrX 
   
 β  = Lβ − Lβ L−1 ϑ  β 
βδX  = Lβ1Y L
−1
 0δ  0Y 1Y and  δr βδδ (1.C7)
   
 β   β 
0X Xr βXδ βXX 

To get the covariance matrix we match the instantaneous covariance matrices of the state variables

from the model equation (1.C6) and the model in Proposition 1.1

 
 
 σ2r ρrδ σr σδ ρrX σr σX 
 

LLT =  ρrδ σr σδ σ2δ ρδX σδ σX  (1.C8)
 
 
ρrX σr σX ρδX σδ σX σ2X

η1 φ1
From equation (1.C8) we get that σ2r = ψ21 , σ2δ = η> η , σ2X = φ>
Y Y
φ , ρrδ =
Y Y σδ , ρrX = σX and
η> φ
ρδX = Y Y
σδ σX .

Appendix 1.D The {r, b

δ, X} representation

We follow the same approach as in Appendix 1.C. We have the transformations for X(t) and r(t)

from equations (1.1) and (1.5), respectively. From equation (1.11) in Proposition 1.2 and equa-

tion (1.6) we can back out the idiosyncratic component of the convenience yield, b
δ(t), as a function

of the canonical state variables47

b
δ(t) = b η>
η0 + b Y(t) (1.D1)
Y

where

1
b
η0 = − φ> φ + (1 − αr )ψ0 − αX φ0 (1.D2)
2 Y Y
   
  
Q Q Q Q
 Q κ −κ33 κ21 κ32
 b η1   −φ2 κ21 κ22 Q − φ3 Q 
    Q
11 −κ22 κ 11 −κ Q
22 
 
 =  
b
ηY =  b Q Q Q
 η2   φ2 (κ22 − κ33 ) + φ3 κ32  (1.D3)
   

η3  
b 0
47
There are two possible decompositions for this representation equivalent to the ones in the proof of Proposition 1.2.
We present the unique decomposition that satisfies the conditions of that proposition.
Appendix 1.D. The {r, b
δ, X} representation 55

and

Q Q Q Q
φ1 (κ11 − κ33 ) + φ2 κ21 + φ3 κ31 −b
η1
αr = 1+ (1.D4)
ψ1
Q
αX = κ33 (1.D5)

 
b > (t) = r(t), b
We define the transformed state vector W δ(t), X(t) . The linear transformation in matrix

form is
b =b
W(t) ϑ+b
LY(t) (1.D6)

where Y(t) follows the process in (1.2). The matrices for the linear transformations are:

   
   
 0 
ψ  1 ψ 0 0 
   
ϑ =  b
b η 
 and b
L =  b  η b
η b
η 
 (1.D7)
 0
  1 2 3

 
φ0  φ1 φ2 φ3 

From equation (1.D6) and Ito’s lemma we have

b =b
dW(t) L−1 (b
Lκ Qb b
ϑ − W(t))dt +b
LdZ Q (t) (1.D8)

The remaining mean-reversion and long-run parameters under the equivalent martingale measure
Q
are given by κrQ = κ11 , κbδQ = κ22
Q
, θrQ = ψ0 and θbδQ = b
η0 . This transformation was done under the

equivalent martingale measure. Using the specification of the risk premia in equation (1.22) and

equation (1.D6) we get the rotation under the physical measure

b
dW(t) = b L−1 (b
Lκ Qb b
ϑ − W(t))dt +b
LdZ(t) + b
L(β0Y + β1Y Y(t))dt

= b L−1 (b
Lκ Qb b
ϑ − W(t))dt +b
LdZ(t) + (b
Lβ0Y − b L−1b
Lβ1Y b ϑ+b
Lβ1Y b b
L−1 W(t))dt
Appendix 1.E. MLE of the Maximal Model 56

The risk-premia parameters for the {r, b

δ, X} representation are

   
   
 β0r   βrr βrbδ βrX 
   
 β  = b  β 
 0bδ
b b−1b
 Lβ0Y − Lβ1Y L ϑ and  bδr βbδbδ βbδX  = b Lβ1Y b
L−1 (1.D9)
   
 β   β 
0X Xr βXbδ βXX 

To get the covariance matrix we match the instantaneous covariance matrices of the state variables

from the model equation (1.D8) and the model in Proposition 1.2

 
 
 σ2r ρrbδ σr σbδ ρrX σr σX 
 

bLT =  ρ bσr σb
Lb σb2 ρbδX σbδ σX  (1.D10)
 rδ δ δ 
 2 
ρrX σr σX ρbδX σbδ σX σX

b
η1 φ1
η>
From equation (1.D10) we get that σ2r = ψ21 , σb2 = b bη , σ2X = φ>
Y Y
φ , ρrbδ =
Y Y σb , ρrX = σX and
δ δ
η>
b φ
ρbδX = Y Y
σbσX .
δ

Appendix 1.E MLE of the Maximal Model

We follow the maximum-likelihood approach of Collin-Dufresne, Goldstein and Jones (2002) which

extends Chen and Scott (1993) and Pearson and Sun (1994) in the following way: instead of assum-

ing that we observe without error some futures contracts and bond prices, we choose to fit the

principal components (PCs) of futures and bonds.48 From the perfectly observed data and using

the closed-form solutions for futures and bonds in Appendices Appendix 1.A and Appendix 1.B

we can invert for the state variables Y(t). We assume that the first two PCs of the futures curve

and the first PC of the yield curve are observed without error. The rest of the PCs are assumed

to be observed with measurement errors that are jointly normally distributed and follow an AR(1)

process. Using the principal component approach instead of single contracts has some advantages.

First, it guarantees (by construction) that we fit perfectly the first two PCs of futures and the first PC

of the yield curve. Second, it orthogonalizes the matrix of measurement errors. Finally, dispenses

the arbitrariness of what contracts are perfectly observed.

48
These principal components are linear in the state variables and can be thought of being portfolios of single contracts.
Appendix 1.E. MLE of the Maximal Model 57

Our cross-sectional data set is composed by m futures contracts and n zero-coupon bonds. The

ith principal components of the futures curve is

> 
PC Fi (b
Y; Θ) = ωiF AF (Θ) + BF (Θ)> b
Y (1.E1)

where ωiF is an m × 1 eigenvector corresponding to the ith principal component, AF (Θ) is an m ×

1 vector and BF (Θ) is an 3 × m matrix that determine the theoretical value of the log futures
 
prices for different maturity contracts, i.e. AF (Θ)> = AF (τ1F ; Θ), . . . , AF (τm F
; Θ) and BF (Θ) =
 
BF (τ1F ; Θ), . . . , BF (τm
F
; Θ) (see Appendix 1.A). Θ is the parameter space of the model.

In the same way we obtain the ith principal component of the yield curve

> 
PC Pi (b
Y1 ; Θ) = ωiP AP (Θ) + BP (Θ)b
Y1 (1.E2)

To invert the state variables b

Y we perfectly observe the first two principal components of the futures

curve and the first principal component of the yield curve. The relation between the data and the

state variables is

G(t) = A (Θ) + B (Θ) b

Y(t) (1.E3)

 > > > 

where G(t)> = ω1P LnP(t), ω1F LnF(t), ω2F LnF(t) is the 1 × 3 vector of perfectly observed PCs
 > > > 
at time t,49 A (Θ)> = ω1P AP (Θ), ω1F AF (Θ), ω2F AF (Θ) is a 1 × 3 vector and B is a 3 × 3 matrix

given by
 
 
 ωP BP (Θ) 0 0 
1 >
 
B (Θ) =  ω 1 > B (Θ)> 
 (1.E4)
 F F

 > 
2
ωF BF (Θ) >

At any given point in time t, we can invert equation (1.E3) to back out the state variables b
Y(t)> =

(b
Y1 (t), b
Y2 (t), b
Y3 (t)):
b
Y(t) = B (Θ)−1 (G(t) − A (Θ)) (1.E5)
49
LnP(t) and LnF(t) are the vectors of the logarithm of the observed bonds and futures contracts at time t, respectively.
Appendix 1.F. MLE of the Triple-Jump Model 58

The other bonds and futures principal components are priced with error

> > 
ωiP LnP(t) = ωiP AP (Θ) + BP (Θ)b
Y1 (t) + uiP (t) for i = 2, .., n
> > 
ωiF LnF(t) = ωiF AF (Θ) + BF (Θ)> b
Y(t) + uiF (t) for i = 3, .., m

We assume that the measurement errors uij (t) follow and AR(1) process, i.e. uij (t) = ρ j uij (t −1)+eij (t)

for j ∈ {P, F}, and the errors eij (t) are jointly normally distributed with zero mean and covariance
h >
i
matrix E eij eij .

The conditional likelihood function for every time t will be given by the likelihood function

of G(t) times the likelihood function of the measurement errors, fu (u(t) | u(t − 1)) = fe (e(t)). We

don’t know the conditional density function of G(t), but since it is an affine function of the state

vector b
Y(t) and we know the conditional distribution of b
Y(t)50 , we can get it using the relation in

equation (1.E5):
 
fG (G(t) | G(t − 1)) = abs(JbY ) fb b
Y(t) | b
Y(t − 1) (1.E6)
Y

 
where JbY is the Jacobian of the transformation from G(t) to b
Y(t), i.e. JbY = det B−1 .

The estimated parameters will be the ones that maximize the log-likelihood function:

X
T
maxL (Θ) = fG (G(t) | G(t − 1)) + fe (e(t)) (1.E7)
Θ
t=1

where fG (G(1) | G(0)) is the unconditional density function.

Appendix 1.F MLE of the Triple-Jump Model

For the case with a triple-jump component in spot prices we use a similar approach than the one in

Appendix 1.E. Since the jumps are in the spot price it is easier to work with the economic represen-

tation {r, b
δ, X}, than with the canonical form {Y1 , Y2 , Y3 }. Using equation (1.C4) from Appendix 1.D

and equation (1.E5), the conditional likelihood of the perfectly observed principal components is

 
fG (G(t) | G(t − 1)) = abs(JWb ) fW b b
b W(t) | W(t − 1) (1.F1)
50
In our Gaussian model we can calculate the exact moments for the distribution of b
Y(t).
Appendix 1.G. Closed-form solution for futures prices with jumps 59

b is the vector of the economic state variables {r, b

where W δ, X} implied from the perfectly observed
b
data and JWb is the Jacobian of the transformation from G(t) to W(t). Equation (1.F1) is similar to
b instead of the canonical vector b
equation (1.E6), but depends on the economic state variables W, Y.

We also use the close-form solution for futures prices with jumps from Appendix 1.G instead of the

one in Appendix 1.A.

Since the transition density with jumps is no longer Gaussian, we follow Ball and Torous (1983),

Jorion (1988), and Das (2002) and approximate it by a mixture of Gaussian. This approximation is

as follows

  ∞ X
X ∞ X
∞
fW b b
b W(t) | W(t − 1) = p1 (N1 (∆t) = k1 )p2 (N2 (∆t) = k2 )p3 (N3 (∆t) = k3 ) ×
k1 =0 k2 =0 k3 =0
 
b b
b W(t) | W(t − 1), N1 (∆t) = k1 , N2 (∆t) = k2 , N3 (∆t) = k3
fW

where the last term is the likelihood function conditional on a fixed number of jumps k1 and k2

which is Gaussian, and the Poisson probabilities are

(λi ∆t)ki
pi (Ni (∆t) = ki ) = e−λi ∆t (1.F2)
ki

This approximation would be exact if the time interval was infinitesimal.

Appendix 1.G Closed-form solution for futures prices with jumps

The model is given by

δ(t) = αr r(t) + b
δ(t) + αX X(t) (1.G1)
Appendix 1.G. Closed-form solution for futures prices with jumps 60

where the state variables {r, b

δ, X} have the following risk-neutral dynamics:

 
dr(t) = κrQ θrQ − r(t) dt + σr dZrQ (t) (1.G2)
 
db
δ(t) = κbQ θbQ − b δ(t) dt + σbδ dZbQ (t) (1.G3)
δ δ δ
 
 1 X 3 
dX(t) = r(t) − δ(t) − σX −
2
(ϕi − 1)λi  dt
Q
2 i=1
X
3
Q
+σX dZX (t) + νi (t)dNi (t) (1.G4)
i=1

The futures price F T (t) = EtQ [S (T )] = EtQ [eX(T ) ]. We show that the expectation has the following

exponential affine form:

 
F T (t) = exp A0 (T − t) + BX (T − t)X(t) + Br (T − t)r(t) + Bbδ (T − t)b
δ(t) (1.G5)

where the functions A0 , BX , Br , Bbδ are given by:

BX (τ) = e−αX τ (1.G6)

1  −α τ −κ τ

Bbδ (τ) = e X − e bδ (1.G7)
αX − κbδ
αr − 1 −α τ

Br (τ) = e X − e−κr τ (1.G8)
α −κ
Z Xτ ( r
1 2
A0 (τ) = (BX (s)σ2X + Bb2 (s)σb2 + B2r (s)σ2r ) (1.G9)
0 2 δ δ
X



− λi BX (s) ϕi − 1 − ϕi (BX (s)) − 1 + Bbδ (s)κbδ θbδ + Br (s)κr θr
i
o
+ρrbδ σr σbδ Br (s)Bbδ (s) + ρrX σr σX Br (s)BX (s) + ρbδX σX σbδ BX (s)Bbδ (s) ds

 α2 σ2i

where we define ϕi (α) = exp αmi + 2 . The proof consists in verifying that the candidate solu-

tion given in equation (1.G5) is a Q-martingale. Indeed applying itô’s lemma to F defined in (1.G5)

and using equations (1.G6) to (1.G9) we see that

T n Z o
T T
F (t) = F (T ) − BX (T − s)σX dZX (s) + Bbδ (T − s)σbδ dZbδ (s) + Br (T − s)σr dZr (s)
t
XZ T n
o
+ (eBX (s)νi (s) − 1)dNi (s) − ϕi (BX (s)) − 1 λiQ ds (1.G10)
i t
Appendix 1.G. Closed-form solution for futures prices with jumps 61

Thus

F T (t) = EtQ [F T (T )] = EtQ [eX(T ) ]

where for the second equality we have used the fact that BX (0) = 1 and A0 (0) = Bbδ (0) = Br (0) = 0.

Inspecting the solution we see that the only impact of jumps on the prices of futures is through

the term


Ji := λiQ BX (s)(ϕi − 1) − ϕi (BX (s)) − 1

in the expression for A0 .

Note that if αX = 0 then BX (s) = 1 and thus we have Ji = 0. We conclude that If αX = 0 then

futures prices are not affected by the presence of jumps.

Next we show that if αX , 0 then the impact of jumps is likely to be small if the jump intensity,

and jump size volatility are ‘small.’ Indeed if a Taylor series approximation is appropriate we have
σ2i v2i B2X
ϕi ≈ 1 + (mi + 2 ) and ϕi (BX ) ≈ 1 + (mi BX + 2 ). Substituting in the expression for Ji we obtain:

v2i
Q
Ji ≈ λi B (1 − BX )
2 X

We thus see that the impact of jumps for the cross section of futures prices is minimal if the jump

intensity and jump size volatility are small. Thus, jumps are mainly helpful in capturing time series

properties of futures prices. Of course, jumps would have a significant impact for the cross-section

of option prices.
BIBLIOGRAPHY 62

Bibliography
Amin, K., Ng, V., and Pirrong, G. (1995). Valuing energy derivatives. in Robert James ed.: Man-
aging Energy Price Risk (Risk Publications, London.

Bakshi, G. and Madan, D. (2000). Spanning and derivative security valuation. Journal of Financial
Economics, 55 no.2:205–238.

Ball, C. A. and Torous, W. N. (1983). A simplified jump process for common stock returns. Journal
of Financial and Quantitative Analysis, v18n1:53–65.

Bessembinder, H., Coughenour, J. F., Paul J, S., and Smoller, M. M. (1995). Mean-reversion in
equilibrium asset prices: Evidence from the futures term structure. Journal of Finance, VOL. L
NO.1 (March):361–375.

Brémaud, P. (1981). Point Processes and Queues. Springer Verlag.

Brennan, M. J. (1958). The supply of storage. American Economic Review, 48:50–72.

Brennan, M. J. (1991). The price of convenience and the valuation of commodity contingent claims.
in D. Lund and B. Oksendal, Eds.: Stochastic Models and Option Values, North Holland.

Brennan, M. J. and Schwartz, E. S. (1985). Evaluating natural resource investments. Journal of

Business, VOL. 58:135–157.

Casassus, J., Collin-Dufresne, P., and Routledge, B. R. (2003). Equilibrium commodity prices with
irreversible investment and non-linear technologies. Working Paper Carnegie Mellon University.

Chen, R. R. and Scott, L. (1993). Maximum likelihood estimation for a multifactor equilibrium
model of the term structure of interest rates. Journal of Fixed Income, December,Vol 3 no 3:14–
31.

Collin-Dufresne, P., Goldstein, R., and Jones, C. (2002). Identification and estimation of ‘maximal’
affine term structure models: An application to stochastic volatility. Carnegie Mellon Working
paper.

Collin-Dufresne, P. and Goldstein, R. S. (2002). Unspanned stochastic volatility: Empirical evi-

dence and its affine representation. forthcoming The Journal of Finance.

Collin-Dufresne, P. and Hugonnier, J. (1999). On the pricing and hedging of contingent claims in
the presence of extraneous risks. Working paper Carnegie Mellon University.

Collin-Dufresne, P. and Solnik, B. (2001). On the term structure of default premia in the swap and
libor markets. Journal of Finance, VOL. LVI NO. 3:1095–1115.

Dai, Q. and Singleton, K. (2002). Expectation puzzles, time-varying risk premia, and affine models
of the term structure. Journal of Financial Economics, 63(3):415–441.

Dai, Q. and Singleton, K. J. (2000). Specification analysis of affine term structure models. Journal
of Finance, 55:1943–1978.

Das, S. R. (2002). The surprise element: Jumps in interest rates. Journal of Econometrics, v106:27–
65.
BIBLIOGRAPHY 63

Deaton, A. and Laroque, G. (1992). Competitive storage and commodity price dynamics. Journal
of Political Economy, 104:896–923.

Duffee, G. R. (2002). Term premia and interest rate forecasts in affine models. Journal of Finance,
57 no 1.

Duffie, D. (1996). Dynamic Asset Pricing Theory. Princeton: University Press.

Duffie, D., Pan, J., and Singleton, K. (2000). Transform analysis and option pricing for affine
jump-diffusions. Econometrica, 68:1343–1376.

Duffie, D. and Singleton, K. (1997). An econometric model of the term structure of interest-rate
swap yields. Journal of Finance, LII.no.4:1287–1381.

Fama, E. F. (1984). Forward and spot exchange rates. Journal of Monetary Economics, VOL.
14:319–38.

Fama, E. F. and French, K. R. (1987). Commodity futures prices: Some evidence on forecast power,
premiums and the theory of storage. Journal of Business, VOL. 60 No. 1:55–73.

Fama, E. F. and French, K. R. (1988). Permanent and temporary components of stock prices.
Journal of Political Economy, 96:246–273.

Gibson, R. and Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil contin-
gent claims. Journal of Finance, v45n3 (July):959–76.

Heston, S. L. (1993). A closed form solution for options with stochastic volatility. Review of
financial studies, 6:327–343.

Hilliard, J. E. and Reis, J. (1998). Valuation of commodity futures and options under stochastic
convenience yields, interest rates, and jump diffusions in the spot. Journal of Financial and
Quantitative Analysis, v33n1:61–86.

Honoré, P. (1998). Pitfalls in estimating jump diffusion models. Working Paper Aarhus University.

Jarrow, R., Lando, D., and Yu, F. (2000). Default risk and diversification: Theory and application.
Working Paper Cornell University.

Jorion, P. (1988). On jump processes in the foreign exchange and stock markets. The Review of
Financial Studies, v1n4:427–445.

Langetieg, T. C. (1980). A multivariate model of the term structure. Journal of Finance, 35:71–97.

Liptser, R. S. and Shiryaev, A. N. (1974). Statistics of Random Processes. Springer-Verlag.

Litterman, R. and Scheinkman, J. (1991). Common factors affecting bond returns. Journal of Fixed
Income, june:54–61.

Litzenberger, R. H. and Rabinowitz, N. (1995). Backwardation in oil futures markets: Theory and
empirical evidence. Journal of Finance, VOL. L No. 5:15217–1545.

Majd, S. and Pindyck, R. S. (1987). Time to build, option value, and investment decisions. Journal
of Financial Economics, VOL. 18:7–27.
BIBLIOGRAPHY 64

Pearson, N. D. and Sun, T.-S. (1994). Exploiting the condititional density in estimating the term
structure: An application to the Cox, Ingersoll and Ross model. Journal of Finance, XLIX, no
4:1279–1304.

Piazzesi, M. (2002). Affine term structure models. Working Paper, Forthcoming Handbook of
Econometrics.

Robert James Ed: Risk Publications, L. (1995). Managing energy price risk.

Ross, S. A. (1997). Hedging long run commitments: Exercises in incomplete market pricing. Eco-
nomic Notes by Banca Monte, 26:99–132.

Routledge, B., Seppi, D., and Spatt, C. (2000). Equilibrium forward curves for commodities. Jour-
nal of Finance, LV n 3:1297–1338.

Schwartz, E. S. (1997). The stochastic behavior of commodity prices: Implications for valuation
and hedging. Journal of Finance, v52n3 (July):923–73.

Seppi, D. J. (2003). Risk-neutral stochastic processes for commodity derivative pricing: An intro-
duction and survey. in Ehud I. Ronn ed.: Real Options and Energy Management Using Options
Methodology to Enhance Capital Budgeting Decisions, Risk Publications, London.

Smith, J. and Schwartz, E. S. (2000). Short-term variations and long-term dynamics in commodity
prices. Management Science, v47n2 (July):893–911.

Vasicek, O. (1977). An equilibrium characterization of the term structure. Journal of Financial

Economics, 5:177–188.

Working, H. (1949). The theory of price of storage. American Economic Review, 39:1254–62.
Chapter 2

Equilibrium Commodity Prices with

Irreversible Investment and Non-Linear
Technologies1

Empirical evidence suggests that commodity prices behave differently than standard financial asset

prices. The evidence also suggests that there are marked differences across types of commodity

prices. This chapter presents an equilibrium model of commodity spot and futures prices for a

commodity whose primary use is as an input to production, such as oil. The model captures many

stylized facts of the data.

Robust features exhibited by time series of commodity spot and futures prices are mean-reversion

and heteroscedasticity. Further, combining time series and cross-sectional data on futures prices

provides evidence of time-variation in risk-premia as well as existence of a ‘convenience yield’

(Fama and French (1987), Bessembinder et al. (1995), Casassus and Collin-Dufresne (CC 2002)).

Interestingly, the empirical evidence also suggests that there are marked differences across different

types of commodity prices (e.g., Fama and French (1987)). Casassus and Collin-Dufresne (2002)

use panel data (cross-section and time series) of futures prices to disentangle the importance of con-
1
An updated version of this chapter can be found at www.ing.puc.cl/˜jcasassu as the working paper “Equilibrium
Commodity Prices with Irreversible Investment and Non-Linear Technologies” co-authored with Bryan Routledge and
Pierre Collin-Dufresne.

65
 66

venience yield versus time-variation in risk-premia for various commodities. Their results suggest

that ‘convenience yields’ are much larger and more volatile for commodities that serve as an input

to production, such as copper and oil, as opposed to commodities that may also serve as a store of

value, such as gold and silver. A casual look at a sample of futures curve for various commodities

(reproduced in figure 2.2 below) clearly shows the differences in futures price behavior. Gold and

silver markets exhibit mostly upward sloping futures curve with little variation in slope, whereas

copper and especially oil futures curve exhibit more volatility. In particular, oil future curves are

mostly downward-sloping (i.e., in backwardation), which, given the non-negligible storage costs2

indicates the presence of a sizable ‘convenience yield.’ Further, casual empiricism suggests that the

oil futures curves are not Markov in the spot oil price (as highlighted in figure 3, which shows that

for the same oil spot price one can observe increasing or decreasing futures curves). Lastly, the

volatility of oil futures prices tends to decrease with maturity much more dramatically than that of

gold futures prices.

The commodity literature can be mainly divided into two approaches. The equilibrium (or struc-

tural) models of commodity prices focus on the implications of possible stockouts, which affects

the no-arbitrage valuation because of the impossibility of carrying negative inventories (Gustafson

(1958), Newbery and Stiglitz (1981), Wright and Williams (1982), Scheinkman and Schechtman

(1983), Williams and Wright (1991), Deaton and Laroque (1992), Chambers and Bailey (1996),

and Bobenrieth, Bobenrieth and Wright (2002)). These papers predict that in the presence of stock-

outs, prices may rise above expected future spot prices net of cost of carry. The implications for

futures prices have been studied in Routledge, Seppi and Spatt (2002). One of the drawbacks of

this literature is that the models are highly stylized and thus cannot be used to make quantitative

predictions about the dynamics of spot and futures prices. For example, these papers assume risk-

neutrality which forces futures prices to equal expected future spot prices and thus rule out the

existence of a risk premium. Further, these models in general predict that strong backwardation

can occur only concurrently with stock-outs. Both seem contradicted by the data. Fama and French

(1988), Casassus and Collin-Dufresne (2002) document the presence of substantial time variation in

risk-premia for various commodities. Litzenberger and Rabinowitz (1995) find that strong backwar-
2
The annual storage cost are estimated to be around 20% of the spot price by Ross (1997).
 67

dation occurs 77% of the time3 in oil futures markets, whereas stock-outs are the exception rather

than the rule.

In contrast, reduced-form models exogenously specify the dynamics of the commodity spot

price process, the convenience yield and interest rates to price futures contracts as derivatives

following standard contingent claim pricing techniques (e.g., Gibson and Schwartz (1990), Bren-

nan (1991), Ross (1997), Schwartz (1997), Schwartz and Smith (2000) and Casassus and Collin-

Dufresne (2002)). The convenience yield is defined as an implicit dividend that accrues to the holder

of the commodity (but not to the holder of the futures contract). This definition builds loosely on

the insights of the original ‘theory of storage’ (Kaldor (1939), Working (1948, 1949), Telser (1958),

Brennan (1958)) which argues that there are benefits for producers associated with holding inven-

tories due to the flexibility in meeting unexpected demand and supply shocks without having to

modify the production schedule. The reduced-form approach has gained widespread acceptance be-

cause of its analytical tractability (the models may be used to value sophisticated derivatives) as well

as its flexibility in coping with the statistical properties of commodity processes (mean-reversion,

heteroscedasticity, jumps). However, reduced-form models are by nature statistical and make no

predictions about what are the appropriate specifications of the joint dynamics of spot, convenience

yield and interest rates. The choices are mostly dictated by analytical convenience and data.

In this chapter we propose a general equilibrium model of spot and futures prices of a commod-

ity whose main use is as an input to production. Henceforth we assume that the commodity modeled

is oil.

Three features distinguish our model from the equilibrium ‘stock-out’ models mentioned above.

First, we consider that the primary use of the commodity is as an input to production. Commodity is

valued because it is a necessary input to produce the (numeraire) consumption good. We assume a

risky two-input constant returns to scale technology. Second, we assume that agents are risk-averse.

This allows us to focus on the risk-premium associated with holding the commodity versus futures

contracts. Finally, we assume that building oil wells and extracting oil out of the ground is a costly

process. We assume these costs are irreversible in the sense that once built an oil well can hardly be
3
And in fact, weak backwardation, when futures prices are less than the spot plus cost of carry, occurs 94% of the
times.
 68

used for anything else but producing oil. This last feature allows us to focus on the ‘precautionary’

benefits to holding enough commodity to avoid disruption in production.

We derive the equilibrium consumption and production of the numeraire good, as well as the

demand for the commodity. Investment in oil wells is infrequent and ‘lumpy’ as a result of fixed

adjustment costs and irreversibility. As a result there is a demand for a security ‘buffer’ of commod-

ity. Further, the model generates mean-reversion and heteroscedasticity in spot commodity prices,

a feature shared by real data. One of the main implications of our model is that even though un-

certainty can be described by one single state variable (the ratio of capital to commodity stock),

the spot commodity price is not a one-factor Markov process. Instead, the equilibrium commodity

price process resembles a jump-diffusion regime switching process, where expected return (drift)

and variance (diffusion) switch as the economy moves from the ‘near-to-investment’ region to the

‘far-from-investment’ region. The equilibrium spot prices may also experience a jump when the

switch occurs. The model generates an endogenous convenience yield which has two components,

an absolutely continuous component in the no-investment region and a singular component in the

investment region. This convenience yield reflects the benefit to smoothing the flow of oil used in

production. It is decreasing in the outstanding stock of oil wells.

When the economy is in the investment region, the fixed costs incurred induce a wealth effect

which leads all security prices to jump. Since the investment time is perfectly predictable, all

financial asset prices must jump by the same amount to rule out arbitrage. However, we find that in

equilibrium, oil prices do not satisfy this no-arbitrage condition. Of course, the apparent ‘arbitrage

opportunity’ which arises at investment dates, subsists in equilibrium, because oil is not a traded

asset, but instead valued as an input to production. We further find that the futures curves can be

in contango or in backwardation depending on the state of the economy. As observed in real data

the frequency of backwardation dominates (for reasonable parameters) that of contango. The two-

regimes which characterize the spot price also determine the shape of the futures curve. We find

that futures curve reflect a high degree of mean-reversion (i.e., are more convex) when the economy

is in the ‘near-to-investment’ region. This is partly due to the increased probability of an investment

which announces a drop in the spot price.

In a sense our model formalizes many of the insights of the ‘theory of storage’ as presented in,
 69

for example, Brennan (1958). Interestingly, the model makes many predictions that are consistent

with observed spot and futures data and that are consistent with the qualitative predictions made in

the earlier papers on the theory of storage, and on which reduced-form models are based. Thus our

model can provide a theoretical benchmark for functional form assumptions made in reduced-form

models about the joint dynamics of spot and convenience yields.

Such a benchmark seems important for at least two reasons. First, it is well-known that most

of the predictions of the real options literature hinge crucially on the specification of a convenience

yield (e.g., Dixit and Pindyck (1994)).4 Indeed, following the standard intuition about the sub-

optimality of early exercise of call options in the absence of dividends, if the convenience yield is

negligible compared to storage costs, it may be optimal to not exercise real options. More generally,

the functional form of the convenience yield can have important consequences on the valuation of

real options (Schwartz (1997), Casassus and Collin-Dufresne (2002)). Second, equilibrium models

deliver economically consistent long-term predictions. This may be a great advantage compared

to reduced from models, which, due to the non-availability of data, may be hard to calibrate for

long-term investment horizons.

With this in mind we estimate our model. The price follows a highly non-linear dynamics whose

moments need to be calculated numerically. For this reason, we consider a linear approximation for

the price process described above. The approximated model is desirable as well because, once

estimated, it can be used straightforwardly for financial applications, like valuation or risk manage-

ment. Our model has regime switching between the near-investment and the far-from-investment

regions. The linearization implies that the price process is exponentially affine conditional on the

regime. Under this representation is it straightforward to calculate a good approximation of the

likelihood. Therefore, we use the quasi-maximum likelihood technique of Hamilton (1989) to esti-

mate our model with crude oil data from 1990 to 2003. We find that most parameters are significant

for both regimes, which validates our model. There is an infrequent state that is characterized

by high prices and negative return and a more frequent that has lower average prices and exhibits
4
Real Option Theory emphasizes the option-like characteristics of investment opportunities by including, in a natural
way, managerial flexibilities such as postponement of investments, abandonment of ongoing projects, or expansions of
production capacities (e.g. see the classical models of Brennan and Schwartz (1985), McDonald and Siegel (1986) and
Paddock, Siegel and Smith (1988)).
 70

mean-reversion. To further test the model we estimate the smoothed inference about the state of

the economy (Kim (1993)), i.e., we back out the inferred probability of being in one state or the

other. We compare the shape of futures curves in both states of the economy and find that, as pre-

dicted by the theoretical model, futures curves are mostly convex in the near-to-investment region

but concave in the far-from investment region, reflecting the high degree of mean-reversion when

investment and a drop in prices is imminent. This provides some validation for our equilibrium

model and also suggests that a regime switching model may be a useful alternative to the standard

reduced-form models studied in the literature.

The model presented here is related to existing literature and, in particular, builds upon Cox,

Ingersoll and Ross (1985).5 Dumas (1992) follows CIR and sets up the grounds for analyzing

dynamic GE models in two-sector economies with real frictions. He studies the real-exchange rate

across two countries in the presence of shipping cost for transfers of capital.6 Recent applications of

two-sector CIR economies along the lines of Dumas (1992) have been proposed by Kogan (2001) for

studying irreversible investments and Mamaysky (2001) who studies interest rates in a durable and

non-durable consumption goods economy. Similar non-linear production technologies to the one

we use here have been proposed by Merton (1975) and Sundaresan (1984). Merton (1975) solves

a one-sector stochastic growth model similar to the neoclassical Solow model where the two inputs

are capital stock and labor force, while Sundaresan (1984) studies equilibrium interest rates with

multiple consumption goods that are produced by technology that uses the consumption good and a

capital good as inputs.7 Fixed adjustment costs have been used in multiple research areas since the

seminal (S,s) model of Scarf (1960) on inventory decisions. In the asset pricing literature, Grossman

and Laroque (1990) uses fixed transaction costs to study prices and allocations in the presence of

a durable consumption good. In the investments literature, Caballero and Engel (1999) explains

aggregate investment dynamics in a model that builds from the lumpy microeconomic behavior of

firms facing stochastic fixed adjustment costs.

5
In fact, our model converges to a one -factor CIR production economy when oil is not relevant for the numeraire
technology.
6
Uppal (1993) presents a decentralized version of Dumas’s economy.
7
Surprisingly, there are not many models that use these type of production technologies in continuous time. Recently,
Hartley and Rogers (2003) has extended the Arrow and Kurz (1970) two-sector model to an stochastic framework and
use this type of production technology with private and government capital as inputs.
2.1. The model 71

Our study is also related to the work of Carlson, Khokher and Titman (2002), who propose

an equilibrium model of natural resources. However, in contrast to our model, they assume risk-

neutrality, an exogenous demand function for commodity, and (the main friction in their model) that

commodity is exhaustible, whereas in our model commodity is essentially present in the ground in

infinite supply but is costly to extract.

Section 2 presents the model. Section 3 characterizes equilibrium commodity prices in our

benchmark model with irreversibility and costly oil production. Section 4 considers the special

case, where the oil flow rate of each well is flexible with adjustment costs for this type of flexibility.

Section 5 presents the empirical estimation of the model and discusses its economic implications.

Finally, Section 6 concludes.

2.1 The model

We consider an infinite horizon production economy with two goods. The model extends the Cox,

Ingersoll and Ross (CIR 1985a) production economy to the case where the production technology

requires two inputs, which are complementary.

2.1.1 Representative Agent Characterization

There is a continuum of identical agents (i.e., a representative agent) which maximize their expected

utility of intertemporal consumption, and have time separable constant relative risk-aversion utility

given by 


 −ρt C 1−γ

 e 1−γ if γ > 0, γ , 1
U(t, C) = 
 (2.1)


 e−ρt log (C) if γ = 1

Their is a single consumption good in our economy. Agents can consume the consumption good

or invest it in a production technology. The production technology requires an additional input, the

commodity, which is produced by a stock of oil wells. The dynamics of the stock of oil wells (Qt )

and the stock of consumption good (Kt ) are described in equation (2.2) and equation (2.3) below:
2.1. The model 72

dQt = −(it + δ)Qt dt + Xt dIt (2.2)

dKt = ( f (Kt , it Qt ) − Ct ) dt + σKt dwt − β(Xt ; Qt , Kt )dIt . (2.3)

The oil ‘industry’ produces a flow of oil at rate it and depreciates at rate δ. The representative

agent can decide when and how many additional oil wells to build. We denote by It the investment

time indicator, i.e. dIt = 1 if investment occurs at date t and 0 else. Investment is assumed to

be irreversible (Xt ≥ 0) and costly in the sense that to build Xt new wells at t, the representative

agents incurs a cost of β(Xt ; Qt , Kt ) of the numeraire good. We assume that the cost function has the

following form:

β(Xt ; Qt , Kt ) = βK Kt + βQ Qt + βX Xt (2.4)

βX is a variable cost paid per new oil well. βK K + βQ Q represent the fixed costs incurred when

investing. As is well-known, fixed costs (βK , βQ > 0) lead to an ‘impulse control’ optimization

problem, where the optimal investment decision is likely to be lumpy (i.e., occurring at discrete

dates).8 In contrast if only variable costs are present (βX > 0 and βK = βQ = 0) then the opti-

mal investment decision is an ‘instantaneous control’ which leads to a ‘local time,’ i.e., singular

continuous, investment policy (e.g., Dumas (1991), Harrison (1990)). Below we assume that

βK , βQ , βX > 0.

The case where βK = βQ = 0 can be recovered by taking the appropriate limit as shown in Jeanblanc-

Picque and Shiryaev (1995) and we discuss it in the appendix. Further, to insure that investment is

feasible we assume that:

βK + βQ < 1.

We note that, while in our model investment immediately creates new oil wells (i.e., there is no

time-to-build frictions in our model), one could potentially interpret the costs as a proxy for this
8
The assumption that the fixed component of the investment cost is scaled by the size of the economy, Kt and Qt ,
ensures that the fixed cost does not vanish as the economy grows.
2.1. The model 73

friction.

For simplicity we assume in this section that the extraction rate per unit time of each oil well

is fixed at it = ī. This is meant to capture the fact that it is very costly to increase or decrease the

production flow of oil wells. In practice this is true within certain limits. We thus reconsider the

model with an optimally chosen extraction rate in the presence of adjustment costs in Section 2.4.

The numeraire-good industry, equation (2.3), has a production technology that requires both the

numeraire good and oil. Output is produced continuously at the mean rate

f (k, q) = αk1−η qη .

As in Merton (1975) and Sundaresan (1984) we use the Cobb-Douglas production function (ho-

mogeneous of degree one and constant returns to scale). The parameter η represents the marginal

productivity of oil in the economy. The output of this industry is allocated to consumption (Ct ≥ 0),

reinvested in numeraire good production, or used for investment to create more oil.9 The creation of

Xt new oil costs β(Xt ; Qt , Kt ) of the numeraire good. This cost is borne only when investment occurs

(dIt = 1). Uncertainty in our economy is captured by the Brownian motion wt which drives the dif-

fusion term of the return of the production technology in equation (2.3). We assume that there exists

an underlying probability space (Ω, F, P) satisfying the usual conditions, and where F = {F }t≥0 is

the natural filtration generated by the Brownian Motion wt .

Given our previous discussion it is natural to seek an investment policy of the form {(XT i , T i )}i=0,1,...

where {T i }i=0,... are a sequence of stopping times of the filtration F such that It = 1{T ≤t} and the XT i
i

are FT i -measurable random variables. Let us define the set of admissible strategies A, as such

strategies that lead to strictly positive consumption good stock process (Kt > 0 a.s.). Further, we

restrict the set of allowable consumption policies C to positive integrable F adapted processes. Then

the optimal consumption-investment policy of the representative agent is summarized by:

"Z ∞ #
−ρs
sup E0 e U(C s )ds (2.5)
C∈ C; {(T i ,XT )}i=0,... ∈A 0
i

9
There is no storage of the numeraire good. Output that is not consumed, used in oil investment, or further production
of the numeraire good depreciates fully.
2.1. The model 74

R∞
Let us denote by J(t, K, Q) = supC;A Et [ t e−ρs U(C s )ds] the value function associated with

this problem.

2.1.2 Sufficient conditions for existence of a solution

Before characterizing the full problem 2.5 we establish sufficient conditions on the parameters for

a solution to the problem to exists. We note that this is slightly different than in traditional models

with fixed costs such as Dumas (1992) or Kogan (2002). Indeed, unlike in these models the no-

transaction cost problem does not provide for a natural upper bound. Indeed, in our case, if we

set βK = βQ = βX = 0 the value function becomes infinite, since it is then optimal to build an

infinite number of oil wells (at no cost). Thus unlike in these papers, it is natural to expect that

sufficient conditions on the parameters for existence of the solution should depend on the marginal

cost of building an oil well (as well as other parameters). Indeed, intuitively, if the marginal costs

of an additional oil well is too low relative to the marginal productivity of oil in the K-technology

one would expect the number of oil wells built (and thus the value function) to be unbounded. To

establish reasonable conditions on the parameters we consider the case where there are only variable

costs (βK = βQ = 0 and βX > 0), but where the investment decision is perfectly reversible. Let us

denote Ju (t, K, Q) the value function of the perfectly reversible investment/consumption problem.

Clearly, the solution to that problem will be an upper bound to the value function of (2.5).

When the investment decision is perfectly reversible then it becomes optimal to adjust the stock
Ju Q
of oil wells continuously so as to keep Ju K = βX . This suggests that one can reduce the dimension-

ality of the problem, and consider as the unique state variable Wt = Kt + βX Qt the ‘total wealth’ of

the representative agent (at every point in time the agent can freely transform Q oil wells into βX Q

units of consumption good and vice-versa). Indeed, the dynamics of W are:

dWt = (α(īQt )η Kt
1−η
− Ct − βX (i + δ)Qt )dt + σKt dwt (2.6)

Qt
Since along each path, the agent can freely choose the ratio of oil to capital stock Zt = Kt , the above

suggests that she should optimally do so to maximize point-wise the expected return of total wealth,
2.1. The model 75

h i
i.e., such as to maxQ α(īQ)η K 1−η − βX (i + δ)Q , which gives:

! 1−η
1
Qt αiη η
= ≡ Z∗ (2.7)
Kt βX (ī + δ)

This suggests that it is optimal to maintain a constant ratio of oil wells to consumption good stock

point-wise. It also gives the optimal investment policy, which should satisfy:

dQt = Z ∗ dKt (2.8)

Using equations (2.7) and (2.8) we may rewrite the dynamics of Wt as

dWt   σ
= α(1 − η)(īZ ∗ )η − cut dt + dwt (2.9)
Wt 1 + Z ∗ βX

where we define

Ct = cut Wt . (2.10)

The proposition below verifies that if

( !)
1 1−γ ∗ η γ σ2
au := ρ− α(1 − η)( īZ ) − >0 (2.11)
γ 1 + Z ∗ βX 1 + Z ∗ βX 2

then the optimal strategy is indeed to consume a constant fraction of total wealth cut = au and to

invest continuously so as to keep Qt /Kt = Z ∗ .

Proposition 2.4 Assume that there are no fixed costs (βK = βQ = 0), and that investment is costly

(βX > 0), but fully reversible. If condition (2.11) holds then the optimal value function is given by

(au )−γ (K + βX Q)1−γ

Ju (t, K, Q) = e−ρt (2.12)
1−γ

The optimal consumption policy is

Ct∗ = au (Kt∗ + βX Q∗t ) (2.13)

2.1. The model 76

and the investment policy is characterized by:

Q∗t
= Z∗ (2.14)
Kt∗

where Z ∗ is the constant defined in equation (2.7).

Proof Applying Itô’s lemma to the candidate value function we have:

dJu (t, Kt , Qt ) + U(t, Ct )dt n o

= (au )γ (cut )1−γ − (1 − γ)cut dt (2.15)
Ju (t, Kt , Qt )
( ! )
1−γ η γ σ2 1−γ
+ α(īZt ) − βX (ī + δ)Zt − − ρ dt + σdwt
1 + Zt β X 1 + Zt β X 2 1 + Zt β X

where we have defined Ct = cut (Kt + βX Qt ). Using the definition of Z ∗ and au in respectively (2.7)

and (2.11) we have:

Z T Z T
1−γ
Ju (T, KT , QT ) + U(t, Ct )dt ≤ Ju (0, K0 , Q0 ) + Ju (t, Kt , Qt ) σdwt (2.16)
0 0 1 + Zt βX

Taking expectations on both sides (and assuming that the stochastic integral is a martingale) we

obtain:
" Z T #
E Ju (T, KT , QT ) + U(t, Ct )dt ≤ Ju (0, K0 , Q0 ) (2.17)
0

with equality when we choose the controls cut = au and Zt = Z ∗ . Further we note that for this choice

of controls, we have:
dJu 1−γ
= −au dt + σdwt (2.18)
Ju 1 + Z ∗ βX

which implies that

uT
lim E[Ju (T, KT , QT )] = lim Ju (0, K0 , Q0 )e−a =0
T →∞ T →∞

under the assumption (2.11). It also shows that the stochastic integral above is a square integrable

martingale for this choice of control. Letting T → ∞ in (2.17) shows that our candidate value

function indeed is the optimal value function and confirms that the chosen controls are optimal.
2.1. The model 77

We note that in the case where η = 0, then Oil has no impact on the optimal decisions of the

agent and the value function Ju is the typical solution one obtains in a standard Merton (1976) or

Cox-Ingersoll-Ross (1985a) economy. In that case the condition on the coefficient au becomes:

( )
1 σ2
a0 = ρ − (1 − γ)(α − γ ) > 0 (2.19)
γ 2

which we assume below for simplicity.

A lower bound to the value function is easily derived by choosing to never invest in oil wells

(i.e., setting dIt = 0 ∀t) and by choosing an arbitrary feasible consumption policy Ctl = α(īZt )η Kt .

Indeed, in that case we have:

dZt
= −(ī + δ)dt + σdwt (2.20)
Zt
dKt
= σdwt (2.21)
Kt

It follows that if the following condition holds:

!
l σ2
a := ρ + (1 − γ) η(ī + δ) + (1 − η)(η + γ(1 − η)) >0 (2.22)
2

then, we have
Z 1−η
∞ l 1−γ
−ρt (C t )
(αK0 (īQ0 )η )1−γ
Jl (0, K0 , Q0 ) := E[ e dt] = (2.23)
0 1−γ (1 − γ)al

We collect the two previous results and a few simple properties of the the value function in the

following proposition.

Proposition 2.5 If al , au > 0, the value function of problem (2.5) has the following properties.

1. Jl (t, K, Q) ≤ J(t, K, Q) ≤ Ju (t, K, Q).

2. J(t, K, Q) is increasing in K, Q.

3. J(t, K, Q) is concave homogeneous of degree (1 − γ) in Q and K.

2.1. The model 78

For the following we shall assume conditions (2.11) and (2.22) are satisfied, i.e., that al , au > 0.

2.1.3 Optimal consumption investment with fixed costs and irreversibility

We first derive the HJB equation and appropriate boundary conditions, as well as the optimal con-

sumption/investment policy based on a heuristic arguments due to the nature of the optimization

problem faced. Then we give a more formal verification argument.

First, since the solution depends on the time variable t only through the discounting effect in the

expected utility function, we define the ‘discounted’ value function J(K, Q), such that J(K, Q, t) =

e−ρt J(K, Q). Given that investment in new oil is irreversible (Xt ≥ 0) and the presence of fixed

costs, it is natural to expect that the optimal investment will be infrequent and ‘lumpy’ (e.g., Dumas

(1991)) and defined by two zones of the state space {Kt , Qt }: A no-investment region where dIt = 0

and an investment region where dIt = 1. This is analogous to the shipping cone in Dumas (1992),

but with only one boundary because investment is irreversible.

Optimal Consumption Strategy in the No-Investment Region

When the state variables {Kt , Qt } are in the no-investment region, the numeraire good K can be

consumed or invested in numeraire-good production. In this region, it is never transformed into new

oil (dIt = 0). That is; J(Kt − β(Xt ), Qt + X) < J(Kt , Qt ) and it is not optimal to make any new

investment in oil. The solution of the problem in equation (2.5) is determined by the following the

Hamilton-Jacobi-Bellman (HJB) equation:

sup {−ρJ + U(C) + DJ} = 0 (2.24)

{C≥0}

where D is the Itô operator

  1
DJ(K, Q) ≡ f (K, īQ) − C JK + σ2 K 2 JKK − (ī + δ)QJQ (2.25)
2

with JK and JQ representing the marginal value of an additional unit of numeraire good and oil

respectively. JKK is the second derivative with respect to K.

2.1. The model 79

The first order conditions for equation (2.24) characterize optimal consumption. At the opti-

mum, the marginal value of consumption is equal to the marginal value of an additional unit of the

numeraire good; that is

−1
Ct∗ = JK γ . (2.26)

Similarly, at the optimum, the marginal value of an additional unit of oil determines the representa-

tive agent’s shadow price for that unit and we denote S t as the the equilibrium oil price. Define the

marginal price of oil, S t . That is, S t solves J(Kt , Qt ) = J(Kt + S t , Qt − ). With a Taylor expansion,

this implies
JQ
St = (2.27)
JK

Optimal Investment Strategy

We assume in equation (2.4) that there is a fixed cost when investing in new oil. This increasing-

returns-to-scale technology implies that the investment in new oil decision faced by the represen-

tative agent is an Impulse Control problem (see Harrison et al. (1983)). As is well known, these

problems have the characteristic that whenever investment is optimal, the optimal size of the invest-

ment is non-infinitesimal and the state variables jump back into the no-investment region. Optimal

investment is infrequent and lumpy.

The investment region is defined by J(Kt − β(Xt ), Qt + Xt ) ≥ J(Kt , Qt ); that is when the value of

additional oil exceeds its cost. Of course, along the optimal path, the only time when this inequality

could be strict is at the initial date t = 0 with stocks {K0 , Q0 }.10 Without loss of generality we

assume that the initial capital stocks {K0 , Q0 } are in the no-investment region. Let J1 = J(Kt∗ , Q∗t )

be the value function before investment and J2 = J(Kt∗ − β(Xt∗ ), Q∗t + Xt∗ ) be the value function right

after the investment is made. The investment zone is defined by the value matching condition.

J1 = J2 (2.28)

There are three optimality conditions that determine the level of numeraire good Kt∗ , the amount

of oil Q∗t , and the size of the optimal oil investment Xt∗ at the investment boundary. We follow
10
If this is the case, there is an initial lumpy investment that takes the state variables into the no-investment zone.
2.1. The model 80

Dumas (1991) to determine these super-contact (smooth pasting) conditions.11

J1K = (1 − βK )J2K (2.29)

J1Q = −βQ J2K + J2Q (2.30)

0 = −βX J2K + J2Q (2.31)

These equations imply that

(βX − βQ )J1K − (1 − βK )J1Q = 0. (2.32)

Reduction of number of state variables

Because the numeraire good production function is homogeneous of degree one ( f (k, q) = αk1−η qη )

and the utility function is homogeneous of degree (1 − γ), the value function inherits that property.

This implies that the ratio of oil to the numeraire good is sufficient to characterize the economy.

Indeed, let us define j(z) as

−γ K 1−γ
J(K, Q) = a0 j(z) (2.33)
1−γ

where z is the log of the oil wells to numeraire-good ratio

Q
z = log (2.34)
K

and a0 is a constant.12 The dynamic process for zt is obtained using a generalized version of Itô’s

Lemma.

dzt = µzt dt − σdwt + Λz dIt∗ (2.35)

11
For a discussion of value-matching and super-contact (smooth-pasting) conditions, see Dumas (1991), Dixit (1991)
and Dixit (1993). If βK = βQ = 0 in equation (2.4) then we face an Infinitesimal Control problem. In this case, the optimal
investment is a continuous regulator (Harrison (1990)), so that oil stock before and after investment are the same. In this
case, equations (2.29) to (2.32) result directly from equation (2.28) as can be checked via a Taylor series expansion (as
shown in Dumas (1991)). To solve this case we consider two additional ‘super-contact’ conditions −J1QK + βX J1KK = 0
and −J1QQ + βX J1KQ = 0.
12
Simply for convenience, we set introduce the coefficient a0 > 0 so that as noted in the previous section in the special
case η = 0 (e.g., oil is not used for production), j(z) = 1.
2.1. The model 81

where

!
1
µzt = − f (1, īe ) − zt
c∗t + ī + δ − σ2 (2.36)
2
Λz = z2 − z1 (2.37)

and the consumption rate, c∗t = Ct∗ /Kt∗ , is a function of zt .

The no-investment and investment regions are also characterized solely by zt . Using the same

subscripts as in equation (2.28), define z1 = log(Q∗t ) − log(Kt∗ ) as the log oil to numeraire-good

ratio just prior to investment. Similarly, define z2 = log(Q∗t + Xt∗ ) − log(Kt∗ − β(Xt∗ )) as the log ratio

immediately after the optimal investment in oil occurs. z1 defines the no-investment and investment

region. When zt > z1 it is optimal to postpone investment in new oil. If the state variable zt reaches

z1 , an investment to increase oil stocks by Xt∗ is made. The result is that the state variable jumps to

z2 which is inside the no-investment region. Given the investment cost structure in equation (2.4),

the proportional addition to oil, xt , is just a function of z1 and z2 .

Xt∗ e−z1 − e−z2 − (βK e−z1 + βQ )

xt∗ = = (2.38)
Q∗t e−z2 + βX

The jump in oil wells is

Q2
= 1 + x∗ (2.39)
Q1

and, we can express the jump in the consumption good stock simply as:

K2 1 − βK + ez1 (βX − βQ )
= (2.40)
K1 1 + βX ez2

Finally, the optimal consumption from (2.26) can be rewritten in terms of j as:

!− γ1
C∗ j0 (zt )
c∗t = t∗ = a0 j(zt ) − (2.41)
Kt (1 − γ)

Plugging this into the Hamilton-Jacobi-Bellman in equation (2.24) we obtain one-dimensional

2.1. The model 82

ODE for the function j.

!1− γ1
0 00 j0 (z)
θ0 j(z) + θ1 j (z) + θ2 j (z) + a0 γ j(z) −
1−γ


+α(ī ez )η (1 − γ) j(z) − j0 (z) = 0 (2.42)

where
1 1 1
θ0 = −ρ − γ(1 − γ)σ2 , θ1 = −ī − δ − (1 − 2γ)σ2 , θ2 = σ2 (2.43)
2 2 2

To determine the investment policy, {z1 , z2 }, the value-matching condition of equation (2.28) be-

comes:

 1−γ
(1 + ez2 βX )1−γ j(z1 ) − 1 − βK + ez1 (βX − βQ ) j(z2 ) = 0 (2.44)

Lastly, using the homogeneity there are only two super-contact conditions to determine that capture

equations (2.29), (2.30), and (2.31).13 They are

 
(1 − γ)ez1 (βX − βQ ) j(z1 ) − 1 − βK + ez1 (βX − βQ ) j0 (z1 ) = 0 (2.45)

(1 − γ)ez2 βX j(z2 ) − (1 + ez2 βX ) j0 (z2 ) = 0 (2.46)

The following proposition summarizes the above discussion and offers a verification argument.

Proposition 2.6 Suppose that we can find two constants z1 , z2 (0 ≤ z1 ≤ z2 ) and a C 2 (z1 , ∞)

function j(·), which solve the ODE given in equation (2.42) with boundary conditions (2.44), (2.45),

and (2.46), such that it satisfies the following conditions:

j0 (z)
j(z) ≥ ≥0 (2.47)
(1 − γ)
!
j(x) 1 − βK + e x (βX − βQ ) 1−γ j(y)
≥ ∀y ≥ x ≥ z1 (2.48)
1−γ 1 + β X ey 1−γ
13
In a similar way, if βK = βQ = 0 the two super-contact conditions presented in footnote (11) become the same
condition (1 + (1 − γ)ez1 βX ) j0 (z1 ) − (1 + ez1 βX ) j00 (z1 ) = 0.
2.1. The model 83

then the value function is given by

−γ K 1−γ
J(t, K, Q) = e−ρt a0 j(z) (2.49)
1−γ

where z = log Q
K . Further the optimal consumption policy is given in equation (2.41). The optimal

investment policy consists of a sequence of stopping times and investment amounts, {(T i , XTi )}i=0,2...

given by T 0 = 0 and:

• If z0 ≤ z1 then invest (to move z0 to z2 ):

e−z0 (1 − βK ) − e−z2 − βQ
X0∗ = Q0 (2.50)
e−z2 + βX

Then start anew with new initial values for the stock of consumption good K0 − β(X0∗ , K0 , Q0 )

and stock of oil wells Q0 + X0∗ .

• If z0 > z1 then set X0∗ = 0 and define the sequence of F-stopping times:

T i = inf {t > T i−1 : zt− = z1 } i = 1, 2, . . . (2.51)

and corresponding FT i -measurable investments in oil wells:

e−z1 (1 − βK ) − e−z2 − βQ
XT∗ i = QT i (2.52)
e−z2 + βX

Proof Applying the generalized Itô’s lemma to our candidate value function we find:

−γ 1−γ
e−ρt a0 Kt− nh
dJ(t, Kt , Qt ) + U(t, Ct )dt = θ̂0 (zt ) j(zt ) + θ̂1 (zt ) j0 (zt ) + θ2 j00 (zt ) (2.53)
1−γ
γ
i

+ a0 (ct )1−γ − ct (1 − γ) j(zt ) − j0 (zt ) dt + (1 − γ) j(zt ) − j0 (zt ) σdwt
  
X  1 − βK + ezTi− (βX − βQ ) 1−γ 



 
 
+  
 j(z ) − j(z T i− 
)
1 + βX e i
zT Ti


T i ≤t

where for simplicity we have defined θ̂0 (z) = θ0 + (1 − γ)α(īez )η and θ̂1 (z) = θ2 − α(īez )η and

Ct = ct Kt . Suppose we can find a function j(·) defined on some closed domain D, such that for any
2.2. Equilibrium Prices 84

y > x (with y, x ∈ D) we have

!1−γ
1 − βK + e x (βX − βQ ) j(y) j(x)
− ≤0
1 + βX ey 1−γ 1−γ

and
" !#
θ̂0 (z) j(z) + θ̂1 (z) j0 (z) + θ2 j00 (z) γ (c)
1−γ j0 (z)
+ sup a0 − c j(z) − ≤0
1−γ c 1−γ 1−γ

then we have

Z Z !
T T
−γ 1−γ j0 (zt )
J(T, KT , QT ) + U(t, Ct )dt ≤ J(0, K0 , Q0 ) + e−ρt a0 Kt− j(zt ) − σdwt (2.54)
0 0 1−γ

Under the assumption of the proposition j is such a function. Furthermore the Bellman equa-

tion (2.42) guarantees that for the candidate choice of control for consumption (given in (2.41)) the

drift is zero, and the value matching condition (2.44) insures that at the optimum the jump is zero.

Thus taking expectation (and assuming that the stochastic integral is a martingale) we get

" Z T #
E J(T, KT , QT ) + U(t, Ct )dt ≤ J(0, K0 , Q0 ) (2.55)
0

with equality for our choices of optimal controls. It remains to show that limT →∞ E [J(T, KT , QT )] =

0 and that the stochastic integral is indeed a true martingale. To be completed... 

The Hamilton-Jacobi-Bellman equation with boundary conditions does not have a closed-form so-

lution. In Appendix 2.A we sketch the numerical technique used to solve this system of equations.

In the following we characterize the equilibrium asset prices and oil prices.

2.2 Equilibrium Prices

The solution to the representative agent’s problem of equation (2.5) is used to characterize equi-

librium prices. We first describe the pricing kernel and financial asset prices. Next, we use the

marginal value of a unit of oil, as in equation (2.27), to characterize the equilibrium spot-price of

oil. Finally, we characterize the structure of oil futures’ prices. Interestingly, with only a single
2.2. Equilibrium Prices 85

source of diffusion risk, the model produces prices that can have both jumps and a regime-shift

pattern.

2.2.1 Asset Prices and the Pricing Kernel

Since in our model the markets are dynamically complete, the pricing kernel is characterized by the

representative agent’s optimal solution (see Duffie (1996)). First, define the risk-free money-market

account whose price is Bt . The process for the money market price is

dBt
= rt dt + ΛB dIt (2.56)
Bt

where rt is the instantaneous risk-free rate in the no-investment region. ΛB is a jump in financial

market prices that can occur when the lumpy investment in the oil industry occurs. Note that the

jumps, ΛB dIt , occur at stochastic times, but since they occur based on the oil-investment decision,

they are predictable. In equilibrium, the ΛB is a constant.

The pricing kernel for our economy satisfies

dξt dBt
=− − λt dwt (2.57)
ξt Bt

with ξ0 = 1. In the no-investment region (dIt = 0), the pricing kernel is standard. However, when

investment occurs (dIt = 1), there is a singularity in the pricing kernel (through the ΛB dIt term

in dBt ). This is consistent with Karatzas and Shreve (1998), who show that in order to rule out

arbitrage opportunities, all financial assets in the economy must jump by the same amount ΛB .14

14
The oil commodity price, S t , is not a financial asset and may, as is described later, jump by a different amount at the
point of oil-industry investment.
2.2. Equilibrium Prices 86

Proposition 2.7 In equilibrium, financial assets are characterized by:

JK (Kt , Qt )
ξt = e−ρt (2.58)
JK (K0 , Q0 )
rt = fK (Kt , īQt ) − σλt (2.59)
Kt JKK
λt = −σ (2.60)
JK
βK
ΛB = − (2.61)
1 − βK

where fK (., .) is the first derivative of the production function with respect its first argument. More-

over, the equilibrium interest rate and market price of risk are only functions of the state variable

zt , i.e. rt = r(zt ) and λt = λ(zt ). 15

The interest rate in the no-investment region is the marginal productivity of the numeraire good

adjusted by the risk of the technology as in Cox, Ingersoll and Ross (1985) (CIR). The only dif-

ference in our model is the effect of the non-linear technology f (k, q). Similarly, the price of risk

in equation (2.60) is driven by the shape of the productivity of the numeraire good. Interestingly,

there can be a jump (predictable) in asset prices that occurs each time investment in oil is optimal

(dIt = 1). From equation (2.29) we can calculate the size of the jump in the stochastic discount fac-

tor and note that it depends on the oil investment cost structure. In particular, recall from (2.4) that

creating Xt new oil wells costs β(Xt , Qt , Kt ) units of the consumption good. Equilibrium financial

prices will jump if βK > 0 where βK determines how the cost function is related to the size of the

numeraire industry.

2.2.2 Oil Spot Prices

The market-clearing spot price of oil is determined by the marginal value of a unit of oil along the

representative agent’s optimal path. This shadow price, from equation (2.27), is a function of the

ratio of oil to numeraire good state variable, zt .

JQ e−zt j0 (zt )
St = = (2.62)
JK (1 − γ) j(zt ) − j0 (zt )
15
We decide to present these variables under {Kt , Qt } rather than under zt to show that these expressions are similar to
the standard results in a CIR economy.
2.2. Equilibrium Prices 87

To characterize the oil spot price behavior, consider the spot price at the investment boundary,

z1 . From the smooth-pasting condition in equation (2.31), the oil price immediately after new

investment is

S 2,t = βX (2.63)

That is, oil’s value is equal to the marginal cost of new oil at the time of investment. Immediately

prior to new investment, the condition in equation (2.32) implies that

βX − βQ
S 1,t = (2.64)
1 − βK

which depends on both the fixed and marginal cost of acquiring new oil. Therefore, at the point of

investment, the oil price jumps by the constant

βQ − βK βX
ΛS = (2.65)
1 − βK

Since oil is not a traded financial asset, the jump in the price of oil can be different that the ΛB

jump in financial prices. The only situation that produces both asset and oil prices that have no

jumps is when there is no fixed cost to investing in oil (βK = βQ = 0), hence investment is not

lumpy. However, it is also possible to generate continuous asset prices and discontinuous oil prices

(βK = 0, βQ > 0). Alternatively, if βQ = βK βX , then oil prices have no jump. In this case, the cost

of oil investment from equation (2.4) is β(Xt ; Qt , Kt ) = βK (Kt + βX Qt ) + βX Xt . Since S 2,t = βX ,

this implies that the fixed cost component of investing in new oil wells is proportional to aggregate

wealth in the economy. The simulations that follow illustrate this case.

2.2.3 Oil Futures Prices

Given the equilibrium processes for spot prices and the pricing kernel, we can characterize the

behavior of oil futures prices in our model. Define F(z, t, T ) as the date-t futures contract that

delivers one unit of oil at date T given that the state of the economy is z.16 The stochastic process
16
Since the futures contracts are continuously market-to-market, the value of the futures contract is zero.
2.2. Equilibrium Prices 88

for the futures price is

dFt
= µF,t dt + σF,t dwt + ΛF dIt (2.66)
Ft

where µF,t , σF,t and ΛF are determined in equilibrium following Cox, Ingersoll and Ross (1985).

Proposition 2.8 The equilibrium futures price F(z, t, T ) in equation (2.66) satisfies µF,t = σF,t λt

and F(z1 , t, T ) = F(z2 , t, T ), implying ΛF = 0 and the following partial differential equation

1 2
σ Fzz + (µz − σλt )Fz + Ft = 0 (2.67)
2

with boundary condition

F(z, T, T ) = S (z) (2.68)

In many commodity pricing models the second factor used to describe futures prices is the net

convenience yield (see Gibson and Schwartz (1990)). Typically, this assumption is motivated as a

benefit for holding stocks (net of any storage or depreciation costs). In these models, backwardation

(downward slopped forward curve) is implied by the convenience yield. For example, Casassus and

Collin-Dufresne (2002) present a reduced-form model with mean reversion in commodity prices.

When the spot price is high, the convenience yield is high and pushes the spot price back toward a

long-term mean. In our model, we can determine the convenience yield implicitly from equilibrium

prices using the no-arbitrage condition for tradable assets

" #
dS t dBt dYt
Et∗ = − (2.69)
St Bt St

where Et∗ is the expectation under the equivalent martingale measure. The convenience yield is

defined as the implicit return to the holder of the commodity, but not to the owner of a futures

contract. If the commodity S t were a financial asset the convenience yield would be the dividend

flow that implies no arbitrage. This is analogous to calculating the implicit convenience yield from

the “cost-of-carry” and the slope of the futures curve as in Routledge, Seppi and Spatt (2000). The

implicit cumulative net convenience yield Yt has the following dynamics:

dYt = yt S t dt + ΛY S t dIt (2.70)

2.3. Model Calibration 89

If we compare the risk-adjusted drift of the price in equation (2.74) with the one from (2.69) we

conclude that 17

ī
yt = ( fq (Kt , īQt ) − S t ) − δ (2.71)
St
ΛY = ΛB − ΛS (2.72)

In our setting, there are two components to the convenience yield. The first is the continuous com-

ponent yt which accrues continuously. It depends on the marginal productivity of oil in production.

The endogenous convenience yield is increasing in fq and, hence, is increasing in the oil’s impor-

tance as a productive input, η. Also, yt is decreasing in the commodity inventories, Qt . This implies

that the convenience yield is higher near the investment region. Interestingly, in Section 2.4, where

we allow for optimal oil extraction, i∗t , this term vanishes in the case that the adjustment costs for

substituting inputs are zero, i.e. yt = −δ.

The second component of convenience yield is the predictable jump that occurs in prices at

the time of oil investment. ΛY < 0 represents the singular component, which represents arbitrage

profits that agents could make were they able to buy the commodity in the investment region. Note

that if one could short-sell the money-market fund and buy the commodity, one would lock a risk-

free profit of −ΛY > 0. Of course, the commodity is not a financial assets, and its ‘price’ is the

shadow value to the consumers of using it as an input to production, which is very high just prior to

investing.

2.3 Model Calibration

In this section we want to understand the empirical properties of the model in Section 2.1. First,

we use economic aggregates and derivatives data to calibrate our model. Then we discuss the

implications of the model for commodity prices. In particular, we find that two regimes arise in our

economy due to the fixed cost components of the investment. Finally, we do a simple estimation of

a regime-switching model that supports our findings.

17
The continuous component of the convenience yield yt is a function only of zt , but as before, we prefer to present
this variable under {Kt , Qt } rather than under zt to deliver better economic intuition from the result. In fact, the variable
fq would be expressed in terms of fz which has a less clear economic meaning.
2.3. Model Calibration 90

2.3.1 Data and Calibration Criterium

For the calibration we use economic aggregates and crude oil futures prices. The basic idea is

to match some moments of this dataset. In order to do this, we solve the Hamilton-Jacobi-Bellman

equation, the optimal consumption/investment strategies, and futures prices with the numerical tech-

nique described in the Appendix. We are interested in the numerical solution for the dynamics of the

state variable zt . We simulate the state variable to obtain its simulated probability density function

and then calculate the moments of the different variables.18

Due to the high computational burden of the simulation approach, we calibrate some parameters

and leave other parameters free to match the moments of the data. We set the oil share of income η

to 0.04 which is consistent with recent RBC studies that include energy as a production factor (see

Finn (1995), Finn (2000) and Wei (2003)). We set the depreciation rate of the commodity stock δ to

0.02.19 We choose the patience factor ρ to be 0.05 and try different degrees of risk-aversion γ. Here

we present the calibration results when γ is 1.8. The remaining parameters (i.e. the productivity

factor α, the input ratio ī, the volatility of capital σ and the investment cost structure βK , βQ and

βX ) are set to reproduce the following moments of the data: (i) the mean and volatility of oil futures

prices for different maturities, (ii) the average aggregate consumption of oil-output ratio, and (iii)

the average aggregate output-consumption of capital ratio. Since we are more interested in the

asset-pricing implication of our capital-oil economy we give a higher weight to the first two sets of

moments. It is important to note that in the futures data we are trying to match a full term structure

of means and volatilities (across maturities). We pursue a “visual” inspection of the parameter space

subject to the constrains that guarantee the existence of the value function (see equations (2.11) and

(2.22)).

Crude oil futures prices were obtained from the New York Mercantile Exchange (NYMEX). We

use daily prices from 1/2/97 to 8/29/03 and contracts with maturities of 1, 6, 12, 18, 24, 36, 48, 60,

72 months. We consider this period of time in order to include contracts with longer maturities. The

contracts with the higher number of observations are the 1-, 6- and 12-months maturity contracts
18
To generate a simulated probability density function that converges, we first discretize the state space of zt in a grid
of 40,000 points and then simulate monthly samples of the state variable for 107 years.
19
Since we calibrate the fixed input ratio version of the model (i.e. i∗t = ī), the depreciation parameter δ is redundant.
To see this we could set î = ī + δ and α̂ = α īη /(ī + δ)η and absorb the δ parameter. In other words, the parameters that
really matters for the calibration are î and α̂.
2.3. Model Calibration 91

with 1653 observations, while the one with the lowest number of observations is the 72-months con-

tract with 1457 observations. For the period considered the means and volatilities of futures prices

decrease with the maturity of the contract. This implies a high degree of backwardation in crude

oil prices (in our dataset 66% of the times the 6-months maturity contract is below the 1-month

maturity contract). Consumption of petroleum and annual crude oil prices from 1949 to 2002 are

from the Energy Information Administration (EIA) from the US Department of Energy. Aggregate

consumption of capital and output from 1949 to 2002 were obtained from the US National Income

and Product Accounts (NIPA) annual data. As a proxy for aggregate consumption we consider con-

sumption of durable goods and services, and for aggregate output we use Gross Domestic Product

(GDP) account. The average annual consumption of oil-output ratio in the data is 2.16% and its

volatility is 0.01. This ratio is very stable most of the time (around 2%), but it peeked in 1981 at

almost 6%. The average annual output-consumption of capital ratio in the data is 1.8 and is stable

with a slightly decreasing tendency (its standard deviation is 0.08).

In our model the instantaneous consumption of oil in terms of the numeraire good is ī Qt S t dt

and the instantaneous output is f (Kt , ī Qt )dt. The ratio between these two flows is only a function

of the state variable zt . Its expectation under the distribution of zt is a good proxy for the average

consumption of oil-output ratio. All parameters of the model affect this ratio, but we note that for

calibration purposes it is extremely sensitive to the input rate of oil ī. Thus, matching this expected

ratio with the data is a good condition for pinning down the parameter ī. In a similar way, the proxy

for the average annual output-consumption of capital ratio in our model is E z [ f (Kt , ī Qt )/Ct ]. This

ratio is also affected by all parameters in the model, but it is more sensitive to γ and to a lesser extent

it is also sensitive to σ and α.

The parameters that we choose to match the calibration criterium are presented in Table 2.1.

The historical moments and their implied value with the calibrated parameters are shown in Table

2.2. Figure 2.4 shows the plots for the mean and volatility of futures prices. We can see that

for the calibrated parameters the model makes a good job when trying to match the unconditional

moments of the futures data. Also, the expected consumption of oil-output ratio with the calibrated

parameters is 2.1% with a volatility of 0.01. The expected output-consumption of capital ratio is 2.2

with a volatility of 0.01. In order to match the latter ratio with the data we would need to increase
2.3. Model Calibration 92

the risk aversion degree, but as we will see later, this is difficult to achieve because the existence

conditions are binding. Also, since η is low, it is difficult to match the volatility of this ratio.20 Other

important figures that our model is capable of reproducing are an expected interest rate of 3.5% and

an expected convenience yield for oil of 6.1%, which is in line with the 10.8% reported in Casassus

and Collin-Dufresne (2002).

The investment costs parameters play an important role in the economy. The marginal invest-

ment cost βX produces a shift in the whole distribution of commodity prices, thus it is a good

parameter to match the average spot price. The effect of the fixed cost components βK and βQ is

more important for higher moments of commodity prices. As we mention before, fixed costs imply

that investment is lumpy, meaning that the state variable jumps from the investment trigger to the

optimal oil-capital ratio. This creates a lot of variability in the state variable and thus in prices.

Figure 2.5 makes a comparison of the probability density function of zt and S t for the cases without

fixed costs and with large fixed costs (20% of capital). The upper left plot shows the PDFs of zt and

the commodity spot price S t as a function of zt . Without fixed costs investment is infinitesimal and

the state variable stays most of the time near the boundary. For this reason prices are typically high

and have low volatility.21 This creates a lot of negative skewness in the distribution of the prices

(see the upper right plot in this figure). The implication of this for futures prices is that backwar-

dation is present almost 100% of the times. On the other hand, if we consider high fixed costs, the

state variable stays in the no-investment region most of the times. In this example, the fixed costs

are so high that the volatility is low for both, high prices and when the price is the marginal costs

(returning point after investment is made).22 . Fixed investment costs increase the volatility of the

price and also could potentially generate positive skewness (see the lower right plot in this figure).

Finally, the volatility of return in capital σ, is also one of the important parameters. It accounts

for the only source uncertainty in our model. This volatility affects the decision of the agents whose

objective is to smooth consumption, thus, it modifies the whole dynamics of the state variable.23

It directly affects the rate of consumption through the interest rate (negative effect) and through
20
In the limit, when η = 0 both output and consumption are proportional to the stock of capital, so this ratio is constant.
21
As we show later, the volatility of prices is related to the drift of the price function, which is zero at the boundary.
22
We discuss the non-monotonic price function in the next subsection
23
For example, it is well known that agents postpone irreversible investments with higher degrees of uncertainty. This
affects the investment trigger and also the size of the investment.
2.3. Model Calibration 93

the volatility of consumption (precautionary savings effect). Also, σ plays a crucial role when

matching the mean and volatility futures curves. A higher volatility combined with positive degrees

of risk aversion, increases the (positive) gap between expected spot prices and futures prices. This

implies higher degrees of backwardation in the futures curves. Finally, the short and long ends of

the term structure of volatilities increase with higher levels of uncertainty. For these reasons we are

seeking for higher degrees of uncertainty in our calibration experiment. Unfortunately, the existence

conditions become extremely tight when σ is high and agents are more risk averse than log. For

example, consider the condition for the lower bound in equation (2.22). Since η is relatively small

there is a trade off between having higher degrees of uncertainty or higher degrees of risk aversion,

but it is impossible to have both.24

2.3.2 Commodity Prices

Figure 2.6 plots the equilibrium oil price as a function of the state variable, zt , the log ratio of oil

stocks to the numeraire good. The oil price is driven by both current and anticipated oil stocks. In

the no-investment region, the supply of oil depletes as oil is used in the production of the numeraire

good. Far from the investment trigger, the decreased supply of oil increases the price. The marginal

cost of adding new oil is βX (equation (2.4)). The fixed cost involved in adding new oil implies that

it is not optimal to make a new investment as soon as the spot price (marginal value of oil) reaches

βX . Therefore the spot price rises above βX as oil is depleted. However, closer to the investment

threshold, the oil price reflects the expected lumpy investment in new oil (i.e., the probability of

hitting the investment threshold is high) and the price decreases. The parameters in this example are

such that ΛS = 0 so the price is continuous at the investment threshold; that is S (z1 ) = S (z2 ).

We use the maximum price S max to partition the state space into two regimes. On the right in

Figure 2.6 with zt ≥ zS max is the far-from-investment zone. In this region, investment is new oil in

the short term is sufficiently unlikely, and the oil price is decreasing in zt . On the left in Figure 2.6

with z1 < zt ≤ zS max is the near-investment zone. In this region, the likelihood of investment in

new oil dominates and a decrease in the stock of oil, zt declines, reduces the price in anticipation

of the increased future oil stocks. Figure 2.7 shows the probability of investing at least one time
24
In the limiting case when η = 0 this condition is ρ > γ(γ − 1)σ2 /2 which is difficult to satisfy given that ρ is 0.05.
2.3. Model Calibration 94

for different horizons. Since the state variable is continuous inside the no-investment region, the

probability in the near-investment zone is higher than the one in the far-from-investment region. Of

course, the likelihood of investment (at least once) is increasing in the horizon.

The fact that the oil price S t is a non-monotonic function of the state variable zt is an important

feature of our model. Since the inverse function z(S ) does not exist, the oil price process is non-

Markov in S t . This is a feature found in the data. Typically, more than one factor is required to

match oil futures prices (see, for example, Schwartz (1997)). Note in Figure 2.3 that two futures

curves with the same spot price are not identical. In our model, the “second factor” that is needed

in addition to the current spot price is whether the economy is in the near-investment or far-from-

investment region.

We state the equilibrium process for the oil price in terms of S t and εt where εt is an indicator

that is one if zt is in the far-from-investment region, and two if zt is in the near-investment region.

Note that there is a one-to-one mapping between {S t , εt } and zt .

Proposition 2.9 The oil price in equation (2.62) is governed by the following two-regime stochastic

process

dS t = µS (S t , εt )S t dt + σS (S t , εt )S t dwt + ΛS dIt (2.73)

µS (S t , εt ) = r(S t , εt ) − y(S t , εt ) + σS (S t , εt )λ(S t , εt ) (2.74)

(S t + e−z(S t ,εt ) )Λ(S t , εt ) − e−z(S t ,εt ) γσ
σS (S t , εt ) = (2.75)
St
βQ − βK βX
ΛS = (2.76)
1 − βK

where





 1 if z > zS max
ε = 
 (2.77)


 2 if z1 < z ≤ zS max

and where r(S t , εt ) = r(zt ) and λ(S t , εt ) = λ(zt ) as in Proposition 2.7, z(S t , εt ) = zt and y(S t , εt ) = yt

is the convenience yield defined later in equation (2.71).

Figure 2.8 shows a typical path for the state variable zt (bottom plot) and the oil price S t (top
2.3. Model Calibration 95

plot). The horizontal lines below show the optimal investment strategy (z1 , z2 ) and the boundary

between the two regimes z Max . Whenever zt hits the investment boundary z1 , it jumps back to z2

inside the no-investment region. The process for zt is only bounded by below and shows some

degrees of mean reversion. When zt is far from the investment trigger (zt is high) the drift of zt is

negative, because the production function f (k, q) uses a lot of oil to produce capital, i.e., Q decreases

quickly while K increases. The simulated oil price is shown in the upper part of the figure. The

price is non-negative, bounded at S max , and mean reverting.

Central to commodity derivative pricing are the conditional moments for the spot-price pro-

cess. Figure 2.9 plots the conditional instantaneous return and conditional instantaneous volatility

of return as a function of S t . The second factor εt , indicating if zt is in the far-from-investment or

near-investment region, is one above the dashed-line and two below this line. From the conditional

drift, note that the oil price is mean-reverting however, the rate of mean reversion (negative drift) is

much higher in the near-investment region. Similarly, the conditional volatility behaves differently

across the two regions. The sign of the volatility in the figure measures the correlation of the oil

price with the shocks in numeraire good production (see equation (2.3)). A positive shock to Kt

means a negative change in zt (less oil relative to the numeraire good). Recall from Figure 2.6,

the decrease in zt implies an increase in the spot price in the far-from-investment , hence a positive

correlation. However, in the near-investment region the spot price decreases implying a negative

correlation. At the endogenously determined maximum price, S max , the volatility is zero and the

drift is negative, which means that the price will decrease almost surely. The volatility of zS max is

non-zero, so there is uncertainty to which direction is the state variable moving after being at this

point.

In order for the regime shifting behavior of the spot price to be detectable (and economically im-

portant), the unconditional distribution for the state variable, zt needs to place some weight near the

boundary of the near-investment and far-from-investment regions. Figure 2.10 plots the probability

density function (simulated) for the state variable zt . This variable is bounded from below by z1 .

The distribution has positive skewness. For our calibration, 84.5% of the time the oil price is above

the marginal cost (that is z1 < zt < z2 ) and 22.5% of the time the economy is in the near-investment
2.3. Model Calibration 96

region (zt < ZS max ).25

From the previous discussion, the non-monotonicity in the relationship between the state vari-

able, zt and the spot price, S t , is crucial for the regime shifting behavior of the spot price. The

size of the hump in Figure 2.6 is determined by the optimal investment policy z1 and z2 . In order

for the hump to be large, investment in new oil wells needs to be large; that is the size of z2 − z1 .

To understand how investment policy is affected by our model parameters, Figure 2.11 shows the

investment strategy defined z1 and z2 under various parameters. The graph on the upper-left corner

shows the effect of economies of scale in the strategy {z1 , z2 }. The bigger is the fixed cost component

βK , the more is the investment delay (z1 is decreasing in βK ). In these charts the difference z2 − z1

gives an idea of the optimal number of oil wells to be built from equation (2.38). When the fixed

cost component is small, the number of new oil wells is low (in the limiting case, investment is

infinitesimal). For higher fixed cost, the investment increases because of higher levels of economies

of scale. The graph to the right shows that a higher marginal cost delays investments. The lower-

left graph of figure 2.11 shows the investment strategy as a function of the oil share η. If the oil

share is very low, then investment is postponed indefinitely. As long as oil becomes relevant for the

production function, the investment trigger increases, which means that investment is made earlier.

The graph on the lower-right corner shows the investment sensitivity to the risk aversion degree of

the individuals. The higher the degree of risk aversion the earlier is the investment undertaken (z1 is

increasing with γ). The intuition for this is that agents care more about smoothing consumption, so

they make investment decisions to stabilize the state variable zt . These decisions are to invest a less

amount more frequently.

Figure 2.12 shows the futures prices for different spot prices and maturities. As with the process

for spot prices in Proposition 2.9, we can use the {S t , εt } characterization of the state variable zt

with futures prices. The thick futures curves are for spot prices in the far-from-investment region

while the thin lines are for spot prices in the near-investment region. The mean-reversion in futures

prices is inherited from the bounded equilibrium oil price. When the oil price is low, the state

variable is far from the investment trigger. This means that the supply of oil can only decrease, so
25
Recall that for this example, we are assuming that the price is continuous, so S 1 = S 2 = βX . This implies that S t is
above βX when z1 < zt < z2 .
2.3. Model Calibration 97

the expected price in the future is above the current price. In these situations the futures curves are

upward-sloping or in contango (for example, see the curve when S t = 15 in figure 2.12). When

the price is near the maximum price the futures curves are downward-sloping, i.e., backwardation

(see the curves when S t = 25). The expected price is below the current price, because of a high

probability of an increase in oil supply (or a high net convenience yield). Figure 2.12 also shows

that the spot price is not sufficient to characterize the futures curve. For higher prices there are two

different futures curves that share the same spot price. One for the case of S t in far-from-investment

and one for S t in the near-investment region. In general, the futures curve are steeper when the spot

price is in the near-investment region. This is a direct implication of a higher convenience yield in

this region (see figure 2.13). This can also be interpreted as a likely sooner investment to create new

oil. Our model also generates non-monotonic curves (see the humped curve when S t = 20 and the

economy is in the far-from-investment region). In these situations, there is an expected shortage of

oil in the short-run, but in the medium-run some new oil will likely be created through investment.

The case when S t = 20 and the economy is in the near-investment region has the opposite situation.

Today the price is above the marginal cost, but with a high probability there will be new investments,

which drops the expected price in the short-run and price is likely to rise in the medium range.

Recall from equations (2.61) and (2.65) that both asset prices and Oil spot prices may jump

at the (predictable) investment in oil. However, as shown in Proposition 2.8, futures prices are

continuous and ΛF = 0. This is not surprising since a futures price is a martingale (expectations

under the equivalent measure of the future spot price) and perfectly anticipate the spot price jump.

The volatility of the futures contract are shown in figure 2.14. To compare the futures volatil-

ity for different oil spot prices we show the relative volatility which we define as σF (S t , εt ; T −

t)/σS (S t , ε). This ratio corresponds to the inverse of the optimal hedge ratio, which is the number

of futures contracts in a portfolio that minimizes the risk exposure of one unit of oil. This ratio is

1 when t = T , because the futures price with zero maturity is the spot price. The thick lines show

the relative volatility for oil spot prices in the far-from-investment region and the thin lines when the

spot is in the near-investment zone. In general, the volatilities are much lower for higher maturi-

ties, which is a consequence from the mean reverting behavior of risk-adjusted prices (often called

the Samuelson Effect). The figure also demonstrates the heteroscedasticity in equilibrium futures
2.3. Model Calibration 98

prices. First, the volatility curves depend on the spot price. In most affine reduce-form models for

commodity prices (see for example Schwartz (1997)), the futures price is proportional to the spot

price, thus relative volatility ratio is assumed to be constant for any given spot price and maturity

date T which does not occur in our model. Second, the curves are non-monotonic in the maturity

horizon. For high prices, the expected investment in oil (rise in supply) is reflected in the futures

contract and also in the volatility. For short maturities and very high prices the relative volatility has

an abrupt behavior because the volatility of the spot price is very low (recall that σS (S max , ε, t) = 0).

As in figure 2.9, a negative volatility implies a negative correlation between the spot price and the

futures price. For example, if the spot price is very high and is in the far-from-investment region it

has a very low volatility and is positively correlated with shocks in capital. In the near future, the

price is expected to be in the near-investment region and to be negatively correlated with shocks in

capital. This implies that the spot and futures price have negative correlation, which is shown with

negative relative volatility values in the figure.

2.3.3 Regime-Switching Estimation

In this section we estimate a linear approximation version of the commodity pricing model in

Proposition 2.9. This model has two regimes that corresponds to the near-investment and far-from-

investment regions. The model for the price is exponentially affine conditional on any given regime.

Despite the fact that we are linearizing the conditional moments with our approximation, the model

is non-linear because of its regime switching characteristic. Estimating the linear approximation

version of the model has several advantages. First, the estimation is much simpler because we can

get an approximation of the likelihood in closed form, while in the “exact” model everything has

to be calculated numerically. Second, it is easier to extend the exponentially affine model with

regime shifts for derivative pricing and risk-management applications. Finally, structural estima-

tions typically need information about the state variables, which in our case is difficult to observe.

By considering the approximated model we can base our estimation solely on observed oil prices.

The main prediction of our model is that there are two different regions in the economy, i.e. the

near-investment and the far-from-investment zones. We consider these two regimes in the approxi-

mated model. Figures 2.6 and 2.9 shows that the price behaves differently depending on the active
2.3. Model Calibration 99

region in the economy. The linear approximation of the structural model in Section 2.1 is

dS t = µS (S t , εt )S t dt + σS (S t , εt )S t dwt (2.78)

where

µS (S , ε) = α + κε (log[S Max ] − log[S ]) (2.79)

p
σS (S , ε) = σε log[S Max ] − log[S ] (2.80)

and εt is a two-state Markov chain with transition (Poisson) probabilities

 
 1 − λ dt λ1 dt 
 
Pt = 
1
 (2.81)
 λ dt 1 − λ2 dt 
2

The process in equations (2.78)-(2.80) is exponentially affine conditional on being in a regime,

i.e. the process for the logarithm of the price has a linear drift term and volatility. The lineariza-

tion of these terms is a first order approximation of the “exact” process for the oil price in equa-

tions (2.73) to (2.75). Equation (2.81) is the transition matrix for the regime variable εt . Here, λi

can be interpreted as the intensity of a jump process for moving out of state εt = i. A second, less

important approximation is that these λ’s are constant, something that is not true in the exact model

since they depend in the price S t (or in the state variable zt in a similar way than the probability of

investment presented in figure 2.7). We set εt = 1 in the far-from-investment region and εt = 2 in

the near-investment region.

Data Description and Estimation Method Our data set consists of weekly Brent crude oil prices

between Jan-1982 and Aug-2003 deflated by the US Consumer Price Index. The average price is

15.41 dollars per barrel in Jan-1982 prices (or 30.16 dollars per barrel in Aug-2003 prices). The

annualized standard deviation of weekly returns is 38.1%. The skewness in crude oil prices for this

period is 1.26 and the excess kurtosis is 0.62.

The parameter space for the approximated model in equations (2.78)-(2.80) is given by Θ = {α,

κ1 , κ2 , σ1 , σ2 , S Max , λ1 , λ2 }. We use the maximum likelihood estimator for regime-switching

2.3. Model Calibration 100

models proposed by Hamilton (1989). We are doing a quasi-maximum likelihood estimation by

considering only the first two moment of the distribution. This should not have a significant impact

on the estimates because we are working with weekly data. The Hamilton’s estimators accounts for

the non-linearities due to the regime-shift characteristic of our model. A by-product of the estima-

tion technique are the smoothed inferences for each regime. We follow Kim’s (1993) algorithm,

which is a backward iterative process that starts from the smoothed probability of the last observa-

tion. The smoothed probabilities are important because they give information about the true regime

that was active any given day.

Results The parameter estimates and standard errors of our model are given in Table 2.3. In

general, most parameters are significant implying that there are clearly two regimes in the data for

the period studied. The parameters vary across regimes implying that these regimes are significantly

different. The economy stays on average one year in the first regime, λ1 = 0.984, before switching

to the second regime. Moreover, the first regime is the most frequent one, since the economy

stays approximately 83.7% of the time in it (λ2 /(λ1 + λ2 ) = 0.837). The economy stays in the

second regime on average a couple of months before jumping back to regime 1 (λ2 = 5.059). The

parameter α is negative and significant implying that the process for the price has an upper bound

at S Max . Also, the estimate for S Max is a reasonable upper bound given the historical path of crude

oil prices (S Max = 39.8). Figure 2.15 displays a graphical representation of the estimates of the

drift and volatility of returns of oil. The figure shows that under the most frequent regime (thick

line), the crude oil price follows a strong mean-reverting process (κ1 = 0.319), i.e. the drift is

positive for low spot prices and negative for high prices. The infrequent regime is different (thin

line), since for reasonable prices the drift is significantly negative and almost constant (α = −0.248

and κ2 = 0.055). Also, the second regime is characterized to be more volatile than the first regime

(σ2 > σ1 ).

Figure 2.16 shows the crude oil price and the inferred probability of being in the near-investment

state (regime 2). We can see that most of the time this probability is low (thin line), implying that the

economy stays mainly in the far-from-investment regime. Also, when the probability is high, most

of the times the price decreases very sharply, which is a characteristic of the near-investment regime.
2.3. Model Calibration 101

In the far-from-investment periods, the price seems to have a mean reverting behavior. Many of these

results are reflected also in the estimates of table 2.3. Figure 2.16 shows that the near-investment

regime is generally for high prices (like in figure 2.6), but sometimes it can be for low spot prices as

well. This implies that in the exact model the fixed cost components of the irreversible investment

are high enough such that the average price is above the marginal price. This allows to generate

both, high and low prices in the near-investment state.

The smoothed probabilities from the maximum likelihood estimation are also important to val-

idate the predictions about the futures prices. For this we do a simple exercise. First, we use the

smoothed probabilities to detect the periods of time where the economy was under one regime or

the other. Second, we group the futures curve in different regimes according to the backed out

dates.26 Third, we sort the curves for both regimes by the price of the shortest maturity contract

(typically the one-month futures contract with price F1 ) and group them according to this price.27

Finally, we compare the behavior of the futures curves under both regimes with the predictions

from our model. We follow a very simple approach for this comparison by calculating the sample

mean of the shortest maturity contract (F1 ) and the average short-term curvature of the futures curve

(F1 − 2F6 + F12 ).28

Table 2.4 shows the results. There are three important results that validate our model. First, for

each regime the column “Nobs” shows the number of observations in every bin (range of F1 prices).

Just by comparing these columns for both regimes we see that the median in the near-investment

regime is higher than the one in the far-from-investment regime. This confirms that on average

the prices are higher in the near-investment regime. Second, we can see that in both regimes the

curvature is positive for high prices and negative for low prices, implying mean reversion under

the equivalent martingale measure. This is one of the main predictions for the futures prices in

our model. Finally, we see that for high spot prices (i.e. the first three bins {“30-”, “25-30”, “20-

25”}), the curvature of the futures curve in the short-term is higher in the near-investment investment
26
We have the futures curve for (Nymex) crude oil prices from Jan-90 to Aug-03.
27
We use the notation Fi for the futures price of a contract with the nearest maturity to i months.
28
The measure of curvature that we choose is the price of a portfolio of futures contracts, where we have a long position
in the one-month and one-year maturity contracts and a short position in two six-month contracts. It is easy to see that
this can be a measure of the second derivative of the curve for short maturities (ω = F1 − 2F6 + F12 ).
2.4. Extensions - Flexible production with adjustment cost 102

region.29 This occurs in our model because the convenience yield is higher in the near-investment

region, which implies higher degrees of backwardation.

2.4 Extensions - Flexible production with adjustment cost

In solving the representative agent model in equation (2.5) we made the simplifying assumption

that the production of oil in equation (2.2) was fixed at it = ī. In this section we explore the effect

of relaxing this assumption by extending our model to include an optimal demand rate of oil. We

consider a variant of the model proposed in the previous sections, where the production technology

f (Kt , it Qt ) is flexible in the sense that the sector chooses optimally the fraction of oil to use as an

input i∗t . There are adjustment costs for this type of flexibility when the optimal demand rate deviates

from some target rate ī. Changes in the stocks of capital Kt and oil wells Qt produces adjustment in

the input rate, it .

The problem of the representative agent is similar to the one before, but the dynamics of the

stocks K and Q account for the flexible demand/production of oil and the adjustment costs, ψ. The

stochastic differential equations for these dynamics are

dQt = −(it + δ)Qt dt + Xt dIt (2.82)

 ψ 
dKt = f (Kt , it Qt ) − (it − ī)2 Kt − Ct dt + σKt dwt − β(Xt )dIt (2.83)
2

There is an extra first order constraint that determines the optimal demand for oil i∗t . It is easy to

check that this FOC is

Kt
fq (Kt , i∗t Qt ) − ψ(i∗t − ī) = St (2.84)
Qt

where fq is the first derivative of the production function with respect its second argument, i.e., the

amount of oil used as an input. In equilibrium, marginal benefit of an extra unit of oil ( fq ) minus

adjustment costs equals its marginal cost (S t ). This model nests the previous one, because as the
29
The results are similar when we use contracts with other maturities for the measure of curvature.
2.4. Extensions - Flexible production with adjustment cost 103

adjustment cost ψ tends to infinity, the optimal demand rate moves to the target rate

lim i∗t = ī (2.85)

ψ→∞

Overall, the main results remain unchanged when we try different adjustment costs. The spot price

follows a mean-reverting process and the same two regimes are present in the economy.

Figure 2.17 shows the dynamics of the demand rate and the convenience yield for different

adjustment costs. The thin lines shows the values of these variables when the economy is in the near-

investment region and the thick lines for the far-from-investment region. The plot on the left shows

the logarithm of the input ratio, log(i∗t Qt /Kt ), determined by equation (2.84). In the limiting case,

when adjustment costs are very high, this relation is linear (i∗t → ī). The more oil in the economy,

the more it is used as an input for the production technology f . With reasonable adjustment costs

(ψ = 1, ψ = 0.1) this increasing relation is also true, but there is a tendency to use more oil when

oil stocks are low. The reason is that the price of oil is lower for lower stocks near the investment

region, so this benefit (from less costs of inputs) justifies adjustments in the demand rate. When the

production technology can be adjusted costlessly (ψ = 0), what really matters is the price of oil S t

instead of the stocks of oil in the economy.30 When the oil price is high, the input rate is low, and

viceversa. In this case, the (log) input ratio inherits the two regimes present in the oil price.

The plot in the right of figure 2.17 shows the convenience yield for different adjustment costs.

Interestingly, the convenience yield is increasing in the adjustment costs in the near-investment

region (where the stocks of oil are small compared to the amount of capital in the economy). With

small adjustment costs this benefit is less important because the oil can be replaced by capital at a

small cost. In the case with no adjustment costs (ψ = 0), the convenience yield is zero and the net

convenience yield equal the negative depreciation cost. Here there are no benefits from holding oil

in the no-investment region, because it can be substituted (locally) at no cost. This is clear from

equations (2.71) and (2.84) with ψ = 0. There remains a singular component to oil prices, which

could be interpreted as a ‘convenience yield’ at the investment boundary. This result shows that

a two-input production technology is necessary but not sufficient to generate positive convenience
30
Of course, the oil/capital ratio affects the price, but when ψ = 0 the input ratio is only a function of S t . In this case,
1
we can get the input ratio in closed-form from equation (2.84), i.e. i∗t Qt /Kt = (α η/S t ) 1−η .
2.5. Conclusion 104

yields. It is necessary to have some degree of rigidity in the technology or adjustment costs (this is

similar to the findings of Carlsson et al. (2002)).

2.5 Conclusion

We develop an equilibrium model for spot and futures oil prices. Our model considers the commod-

ity as an input for a production technology in an explicit way. This feature endogenizes one of the

main assumptions in standard competitive models of storage, i.e. the demand function. Our model

generates positive convenience yields and long period of backwardation in futures curves without

the necessity of running out of oil, like in the standard “stock-out” literature. Convenience yields

arise endogenously due to the productive value of the oil, which is consistent with the predictions

of the “Theory of Storage”. This convenience yield is high when the stocks of commodity are low,

and viceversa. By modeling explicitly risk-averse agents, we can investigate risk-premia associated

with holding of stocks of commodities versus futures contracts.

Equilibrium spot price behavior is endogenously determined as the shadow value of oil. Our

model makes predictions about the dynamics of oil spot prices and futures curves. The equilibrium

price follows an heteroscedastic mean-reverting process. The spot price is non-Markov, because

there are two regimes in our economy that depend on the distance to the investment region. For

reasonable parameters, the futures curves are most of the time backwardated. Also, the two regimes

imply that two futures curve with similar spot prices can have very different degrees of backwarda-

tion.

We calibrate the model using futures price and economic aggregates data. We find that the model

captures many of the stylized facts of our data set. In particular, our model can reproduce the mean

and volatilities of futures prices for maturities up to 72 months and also the average consumption

of oil-output and output-consumption of capital ratios. We estimate a linear approximation version

of our model with crude oil prices from 1982 to 2003. Our empirical specification successfully

captures spot and futures data. Finally, the specific empirical implementation we use is designed

to easily facilitate commodity derivative pricing that is common in two-factor reduced form pricing

models.
2.6. Tables 105

2.6 Tables
2.6. Tables 106

Production technologies
Productivity of capital K, α 0.23
Oil share of output, η 0.04
Demand rate for oil, ī 0.07
Volatility of return on capital, σ 0.263
Depreciation of oil, δ 0.02
Irreversible investment
Fixed cost (K component), βK 0.016
Fixed cost (Q component), βQ 0.272
Marginal cost of oil, βX 17
Agents preferences
Patience, ρ 0.05
Risk aversion, γ 1.8

Table 2.1: Parameters from the calibration exercise.

Historical data Model

Mean Vol Mean Vol
Crude oil futures prices (US$/bbl)
1-months contract 23.62 6.38 22.38 4.47
6-months contract 22.61 4.88 22.15 4.19
12-months contract 21.65 3.87 21.86 3.86
18-months contract 21.05 3.21 21.64 3.58
24-months contract 20.71 2.81 21.42 3.31
36-months contract 20.42 2.43 21.09 2.92
48-months contract 20.28 2.25 20.77 2.56
60-months contract 20.23 2.10 20.49 2.27
72-months contract 20.42 1.88 20.25 2.05
Macroeconomic ratios
consumption of oil-output ratio 2.16% 0.01 2.1% 0.01
output-consumption of capital ratio 1.8 0.08 2.2 0.01

Table 2.2: Historical and implied moments by the model using parameters in Table 2.1.
2.6. Tables 107

Parameter Estimate Std.Error t-ratio

λ1 0.984 0.312 3.2
λ2 5.059 1.589 3.2
α -0.248 0.103 -2.4
κ1 0.319 0.116 2.7
κ2 0.055 0.148 0.4
σ1 0.251 0.010 25.8
σ2 0.808 0.063 12.9
S Max 39.8 0.399 99.7

Table 2.3: Quasi-maximum likelihood estimates for the regime-switching model for weekly deflated
Brent crude oil prices between Jan-1982 and Aug-2003.

F1 oil prices far-from-investment state near-investment state

($/barrel) Nobs F1 F1 − 2F6 + F12 Nobs F1 F1 − 2F6 + F12
30- 41 32.4 114.9 32 33.1 181.8
25-30 93 27.3 3.0 35 27.9 92.7
20-25 189 21.8 31.9 17 21.7 41.5
15-20 237 18.1 -7.2 13 18.2 -34.1
10-15 54 13.5 -28.7 2 12.4 -182.0

Table 2.4: Sample mean of the shortest maturity contract (F1 ) and average short-term curvature
of the futures curve (F1 − 2F6 + F12 ) under different regimes and for different groups of crude
oil prices between Jan-1990 and Aug-2003. The active regime is inferred by the estimation of the
regime-switching model.
2.7. Figures 108

2.7 Figures
 45 150
2.7. Figures

35 125

F01 F01
25 100
F18 F18

15 75

Futures Prices (Copper)

Futures Prices (Crude Oil)

5 50
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

500 750

650

400
550
F01 F01
F18 F18
450
300

Futures Prices (Gold)

350
Futures Prices (Silver)

200 250
Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02 Jan-90 Jan-92 Jan-94 Jan-96 Jan-98 Jan-00 Jan-02

Figure 2.1: F01 and F18 futures contracts on crude oil, copper, gold and silver between 1/2/1990 to 8/25/2003. Oil prices are in dollars per barrel,
copper prices are in cents per pound, gold prices are in dollars per troy ounce and silver prices are in cents per troy ounce.
109
 45 150
2.7. Figures

35 125

25 100

15 75

Futures Curves (Copper)

Futures Curves (Crude Oil)

5 50
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2
Maturity (years) Maturity (years)

500 800

700

400
600

500
300

400

Futures Curves (Gold)

Futures Curves (Silver)

200 300
0 1 2 3 4 0 1 2 3 4
Maturity (years) Maturity (years)

Figure 2.2: Monthly term structures of futures prices on crude oil, copper, gold and silver between 1/2/1990 to 8/25/2003. Oil prices are in dollars
per barrel, copper prices are in cents per pound, gold prices are in dollars per troy ounce and silver prices are in cents per troy ounce.
110
 40
2.7. Figures

30

20

Crude Oil Futures Curves (1990-2003)

10
0 0.5 1 1.5 2 2.5 3 3.5

Maturity (years)

Figure 2.3: Sample futures curves from 1990-2003 for Brent crude (deflated by U.S. Consumer Price Index).
111
 2.7. Figures

24
6

22 Data Data
4
Model Model

20
2

Vol of Futures ($/bbl)

Mean of Futures ($/bbl)

18 0
0 5 10 0 5 10

Maturity (years) Maturity (years)

Figure 2.4: Mean and volatility of futures prices for different maturities. The markers show the moments from historical annual crude oil prices
from 1949 to 2002. The lines show the moments implied by the model using the parameters in Table 2.1.
112
 Commodity price St Commodity price St
0
15

0
200
400

-6.2
-10.266
-6
z1

-10.016
-5.9
z1=z2

-9.766
-5.7
-9.516
-5.6
-9.266
-5.5
-9.016
-5.3
-8.766
parameters are from Table 2.1.

-5.2
-8.516
-5
-8.266
-4.9
-8.016
-4.8
-7.766
-7.516 -4.6
-7.266 -4.5
-7.016 -4.3
-6.766 -4.2
-6.52E+00 -4.1
-6.266 -3.9
-6.016 -3.8
-5.766 -3.6
zt=Log[Qt/Kt]
zt=Log[Qt/Kt]

-5.516 -3.5
-5.266 -3.4
z2

-5.019 -3.2
-4.769 -3.1
-4.519 -2.9
-4.269 -2.8
-4.02E+00 -2.7
0
0

0.05
0.2
0.4

Probability Probability
PDF
PDF

Price
Price

Probability Probability
0.0
0.2
0.4
0.6

0.00
0.05
0.10
0.15

1 / 17 5.5 / 6.1
32 / 48
6.6 / 7.2
63 / 79
7.7 / 8.3
94 / 110
125 / 141 8.8 / 9.3
156 / 172
9.9 / 10.4
187 / 203
218 / 234 11 / 11.5
249 / 265
12.1 / 12.6
280 / 296
311 / 327 13.2 / 13.7

Commodity price St
342 / 358

Commodity price St
14.3 / 14.8
373 / 389
404 / 420 15.4 / 15.9
435 / 451
16.5 / 17

Figure 2.5: Effect of fixed costs component of the investment in the simulated probability density function of the state variable zt and of the
commodity price S t . The upper row is for the case without fixed costs and the lower low is when βK = 0.2 and βQ = 3.4. The rest of the
113 2.7. Figures
 30 Near-investment
Far-from-investment region
2.7. Figures

region

20

10

Commodity price St
0
z1 zSmax z2
zt=Log[Qt/Kt]

Figure 2.6: Oil price S t as a function of the logarithm of the oil wells-capital ratio, zt . The vertical dashed-line is at zS max and separates two
regions. The thin line shows the oil price in the near-investment region (z1 < zt ≤ zS max ) and the thick line is the oil price in the far-from-
investment region (zt ≥ zS max ). We use the parameters in Table 2.1. In particular, the fixed cost components of the investment are βK = 0.016 and
βQ = 0.272, and the marginal cost of oil is βX = 17. The equilibrium critical ratios are z1 = −7.19, zS max = −6.64 and z2 = −5.57.
114
 Near-investment
Far-from-investment region
2.7. Figures

region

0.8

T-t=1 yr
0.6
T-t=2 yrs
T-t=5 yrs
T-t=10 yrs
0.4

Investment Probability
0.2

0
z1 zSmax
zt=Log[Qt/Kt]

Figure 2.7: Probability of investing during an interval of time T − t as a function of the logarithm of the oil wells-capital ratio. The thin lines show
the probability in the near-investment region and the thick line in the far-from-investment region. The equilibrium critical ratios are z1 = −7.19
and zS max = −6.64.
115
 SMax 31
-3.346 29
2.7. Figures

27
25
23
βx 21
19
-4.346 17
15
13
11
9
Commodity price St

-5.346 7
5
3
1
-1
-6.346 -3
-5
-7
z
-9 2
-11
-13
-7.346 -15
-17
-19
z
-21Smax

zt=Log[Qt/Kt]
-23
-8.346 -25
z1
Chronological time t

Figure 2.8: Simulations for the logarithm of the oil wells-capital ratio zt (below) and the oil price S t (above) over time. The thin lines show
these variables in the near-investment region and the thick lines show them in the far-from-investment region. We use the parameters in Table 2.1.
For the path below, the critical ratios are z1 = −7.19, zS max = −6.64 and z2 = −5.57, while for path above the (equilibrium) maximum price is
S Max = 27.60 and the marginal cost of oil is βX = 17.
116
 2.7. Figures

40%
25% Far-from-investment region Far-from-investment region

0%
0%
Near-investment region Near-investment region
-25%

-40%

Drift µ(St)
-50%

Volatility σ(St)
-75% -80%
0 10 20 30 0 10 20 30

Commodity price St Commodity price St

Figure 2.9: Return and instantaneous volatility of returns in oil price S t . The horizontal dashed-line separates the two regimes. The thin lines
below the dashed-lines show the variables under the near-investment regime and the thick lines under the far-from-investment regime. We use the
parameters in Table 2.1. In particular, the fixed cost components of the investment are βK = 0.016 and βQ = 0.272, and the marginal cost of oil is
βX = 17. The endogenous upper bound for the price is S Max = 27.60.
117
 Commodity price St
0
20

-7.2
z1

-7.1
-6.9
-6.8
-6.7
-6.6
-6.4
-6.3
-6.2
-6.1
-5.9
-5.8
-5.7
-5.6
z2

-5.4
-5.3
-5.2
-5.1
-4.9
zt=Log[Qt/Kt]

-4.8
-4.7
-4.5
-4.4
-4.3
-4.2
-4
0.00
0.10

Probability
PDF
Price

Probability
0.0
0.1
0.2
0.3

9 / 9.9
10.7 / 11.6
12.5 / 13.4
14.3 / 15.2
16.1 / 17
17.8 / 18.7
19.6 / 20.5
21.4 / 22.3

Commodity price St
23.2 / 24.1
24.9 / 25.8
26.7 / 27.6

Figure 2.10: Simulated probability density function for the state variable zt and the commodity price S t using the parameters in Table 2.1.
118 2.7. Figures
 -6 -4 6
Xt/Qt
6
-6
2.7. Figures

Xt/Qt
4
4 -8
-7
-10
2
2

zt=Log[Qt/Kt]
zt=Log[Qt/Kt]
z1
-12 z1

Investment ratio
Investment ratio

-8 0 -14 0
0 0.01 0.02 0.03 0.04 0 500 1000 1500 2000

Fixed costs βK Variable costs βX

40 -7 6
z1
-6
z1
30 -7.2
5
-8
20 -7.4

4
-10

zt=Log[Qt/Kt]
zt=Log[Qt/Kt]
10 -7.6

Investment ratio
Investment ratio

Xt/Qt
Xt/Qt
-12 0 -7.8 3
0 0.02 0.04 0.06 0.08 0.5 1 1.5 2 2.5

Oil share of income η Risk aversion γ

Figure 2.11: Investment strategy {z1 , z2 } for different investment cost structure (βK , βX ), levels of risk aversion γ and oil share of income η. The
thick (below) line corresponds to the investment trigger z1 , while the thin (above) line is the returning point z2 . To summarize both fixed cost
components in the parameter βK , we assume that βQ = βK βX for this plots.
119
 30
2.7. Figures

25

S=20
S=25
S=Max
20
S=25
S=20

Futures price
S=15
15

10
0 2 4 6 8 10

Maturity T-t

Figure 2.12: Futures curves for contracts on oil for different spot prices. The thick curves are for spot prices in the far-from-investment region
and the thin lines when the spot price is in the near-investment region. We use the parameters in Table 2.1 and the endogenous upper bound for
the price is S Max = 27.60.
120
 40%
2.7. Figures

Near-investment
Far-from-investment region
region
35%

30%

25%

20%

15%

10%

5%

Net convenience yield (% of S)

0%

-5%
z1 zSmax
zt=Log[Qt/Kt]

Figure 2.13: Convenience yield as a function of the state variable zt when it = ī. The thick line is the convenience yield when the economy is in
the far-from-investment region and the thin for the economy in the near-investment region. We use the parameters in Table 2.1 and the critical
values for zt are z1 = −7.19 and zS max = −6.64.
121
 1.2
2.7. Figures

0.8
S=20
S=25
0.6 S=25
S=20
0.4 S=15

Relative volatility
0.2

0
0 2 4 6 8 10

Maturity T-t

Figure 2.14: Relative volatility of futures contracts on oil to spot price volatility for different spot prices and maturities. The thick curves are for
spot prices in the far-from-investment region and the thin lines when the spot price is in the near-investment region. We use the parameters in
Table 2.1 and the endogenous upper bound for the price is S Max = 27.60.
122
 2.7. Figures

0.5
0.2 Far-from-investment region

0
0
5 10 15 20 25 30 35 40 45

-0.5

Drift µ(St)
-0.2

Volatility σ(St)
Near-investment region

-1
-0.4 5 10 15 20 25 30 35 40 45

Crude oil price St Crude oil price St

Figure 2.15: Maximum likelihood estimes for the expected return and instantaneous volatility of returns in oil prices between Jan-1982 and
Aug-2003. The thin lines below the dashed-lines show the variables under the near-investment regime and the thick lines are the variables under
the far-from-investment regime. The estimates for the parameters that generate these plots are shown in Table 2.3.
123
 35
2.7. Figures

30

0.8
25

20 0.6

15
0.4

10

Deflated crude oil price St

0.2
Smoothed inference of state NI

0 0
Jan-86 Jan-90 Jan-94 Jan-98 Jan-02

Figure 2.16: Historical Brent crude oil prices between Jan-1982 and Aug-2003 deflated by the US Consumer Price Index (thick line) and inferred
probability of being in the near-investment state (thin line).
124
 2.7. Figures

-4.25 15%

ψ=0 ψ=inf
10%
ψ=1
-4.5
ψ=0.1
5%
ψ=0.1
-4.75
0%
ψ=1 ψ=0
ψ=inf

Input ratio Log[itQt/Kt]

Net convenience yield
-5 -5%
-8.75 -8.25 -7.75 -7.25 -6.75 -8.75 -8.25 -7.75 -7.25

zt=Log[Qt/Kt] zt=Log[Qt/Kt]

Figure 2.17: Logarithm of the input ratio for the numeraire production technology and net convenience yield as a function of zt when the
production is flexible and for different adjustment costs. The thin lines show these variables in the near-investment region and the thick lines
show them in the far-from-investment region. We use the parameters in Table 2.1. In particular, the target input ratio is ī = 5% and the depreciation
rate is δ = 2%.
125
Appendix 2.A. Numerical Techniques 126

Appendix 2.A Numerical Techniques

In this appendix we delineate the numerical algorithm used to solve the Hamilton-Jacobi-Bellman

equation in (2.24) with boundary conditions represented by equations (2.28) to (2.31). We present

the solution method for the case with a flexible production technology and adjustment costs (see

Section 2.4)), since this one nests the fixed production technology case.

The first step is to use the homogeneity of the solution to reduce the state space (see Subsection

2.1.3). After this is done, the solution of the problem is represented by a nonlinear second order

ODE in the state variable zt = log(Qt /Kt ). The boundary conditions are also expressed in terms

of zt . Now, we need to determine the value function j(z) in equation (2.33). The nonlinear HJB

equation for j(z) depends on (i) the optimal controls c∗t and i∗t in the case of a flexible production

technology, and on (ii) the optimal investment strategy {z1 , z2 } determined by the boundary condi-

tions. Unfortunately, the optimal controls themselves depend on the value function j(z). The same

is the situation for the optimal investment strategy. These imply that j(z), c∗t , i∗t , z1 and z2 need to be

simultaneously determined.

We use an iterative method to solve for j(z). The main idea is to build a conditionally linear

ODE for j(z) so it is possible to apply a finite-difference scheme. The selection of the initial guess

is extremely important for the convergence of the iteration. We assume that j0 (z) = 1 which cor-

responds to the solution when the oil is not relevant for the production technology (η = 0). In this

case we also know that it is never optimal to invest z01 → ∞.

For every iteration m (for m = 0 . . . ∞) we do the following steps:

• Determine the optimal controls c∗m and i∗m as a function of jm (z). For the optimal consump-

tion c∗m we use equation (2.41) and for the optimal input ratio i∗m we use equation (2.84).

Most of the times the latter equation does not imply a closed-form expression for i∗m and it

needs to be solved numerically.

• We recognize that the ODE for jm+1 (z) determines the value function when it is optimal not

to invest in new stocks of commodity. We name this function as jm+1

noinv (z). We calculate

the coefficients of the ODE for jm+1

noinv (z). It is important to notice that this ODE is linear

conditional on c∗m and i∗m .

Appendix 2.A. Numerical Techniques 127

• Determine the optimal commodity/capital ratio zm+1

2 using the super contact condition in

equation (2.46). Conditional that it is optimal to invest in new commodity stocks, the re-

turning point is always zm+1

2 independent of what was the value of zt before investment was

made. Using this argument we define the extended value matching condition as

 
z (β − β ) 1−γ
m m+1 

 1 − β + e 
jm+1 
inv (z) = j (z2 ) 
K X Q
 . (2.A1)
(zm+1 )
1 + e 2 βX

This equation represents the value function when the representative agent is forced to invest.

• Use a finite-difference scheme to solve for the value function jm+1

noinv (z). The finite difference

discretization defines a tridiagonal matrix that needs to be inverted to determine the value of

jm+1
noinv (z). Instead of doing this, we eliminate the upper diagonal of this matrix. At this point

the value of jm+1 m+1

noinv (z) depends only on the value of jnoinv (z − ∆z). We choose a zmin negative

enough to ensure that at that level it is optimal to invest, and then we solve the value function

for higher zt . At every point we choose the maximum of the value from investing ( jm+1
inv (z))

and the value of no investing which comes from the finite-difference scheme. This maximum

determines the value of jm+1 (z). The optimal trigger zm+1

1 is endogenously determined when

the representative agent is indifferent between investing and postponing the investment. The

algorithm described above is a more efficient way than solving independently for jm+1
inv (z) and

jm+1
noinv (z) and then choosing j
m+1 (z) = max( jm+1 (z), jm+1 (z)).
inv noinv

• Check for the convergence condition. If it not satisfied we start a new iteration with the

updated value of jm+1 (z).

Once j(z) has converged it is straight forward to calculate spot commodity prices from equation

(2.62). For the futures prices we use an implicit finite-difference technique. This is simpler than the

solution for j(z) since the coefficients of the PDE and boundary conditions and boundaries {z1 , z2 }

are known at the beginning of the scheme.

BIBLIOGRAPHY 128

Bibliography
Arrow, K. J. and Kurz, M. (1970). Public Investment, the Rate of Return and Optimal Fiscal Policy.
Johns Hopkins Press, Baltimore.

Bailey, R. and Chambers, M. (1996). A theory of commodity price fluctuations. Journal of Political
Economy, 104(5):924–957.

Bessembinder, H., Coughenour, J. F., Paul J, S., and Smoller, M. M. (1995). Mean-reversion in
equilibrium asset prices: Evidence from the futures term structure. Journal of Finance, VOL. L
NO.1 (March):361–375.

Bobenrieth, E. S. A., Bobenrieth, J. R. A., and Wright, B. D. (2002). A commodity price process
with a unique continuous invariant distribution having infinity mean. Econometrica, 70(3):1213–
1219.

Brennan, M. J. (1958). The supply of storage. American Economic Review, 48:50–72.

Brennan, M. J. (1991). The price of convenience and the valuation of commodity contingent claims.
in D. Lund and B. Oksendal, Eds.: Stochastic Models and Option Values, North Holland.

Brennan, M. J. and Schwartz, E. S. (1985). Evaluating natural resource investments. Journal of

Business, v58 n2:135–57.

Caballero, R. J. and Engel, E. M. R. A. (1999). Explaining investment dynamics in u.s. manufac-

turing: A generalized (s,s) approach. Econometrica, 67(4):783–826.

Carlson, M., Khokher, Z., and Titman, S. (2002). An equilibrium analysis of exhaustable resources
investments. University of Texas Working Paper.

Casassus, J. and Collin-Dufresne, P. (2002). Maximal affine model of convenience yields implied
from interest rates and commodity futures. Carnegie Mellon University Working Paper.

Cox, J. C., Ingersoll Jr., J. E., and Ross, S. A. (1985). An intertemporal general equilibrium model
of asset prices. Econometrica, 53:363–384.

Deaton, A. and Laroque, G. (1992). Competitive storage and commodity price dynamics. Journal
of Political Economy, 104:896–923.

Dixit, A. (1991). A simplified treatment of the theory of optimal regulation of brownian motion.
Journal of Economic Dynamics and Control, v15 n4:657–73.

Dixit, A. (1993). The Art of Smooth Pasting. Vol. 55 in Fundamentals of Pure and Applied Eco-
nomics, Harwood Academic Publishers.

Dixit, A. K. and Pindyck, R. S. (1994). Investment under Uncertainty. Princeton University Press,
Princeton.

Duffie, D. (1996). Dynamic Asset Pricing Theory. Princeton: University Press.

Dumas, B. (1991). Super contact and related optimality conditions. Journal of Economic Dynamics
and Control, v15 n4:675–85.
BIBLIOGRAPHY 129

Dumas, B. (1992). Dynamic equilibrium and the real exchange rate in a spatially separated world.
Review of Financial Studies, 5:153–180.

Fama, E. F. and French, K. R. (1987). Commodity futures prices: some evidence on forecast power,
premiums, and the theory of storage. Journal of Business, v60 n1:55–73.

Finn, M. G. (1995). Varience properties of solow’s productivity residual and their cyclical implica-
tions. Journal of Economic Dynamics and Control, v19:1249–1281.

Finn, M. G. (2000). Perfect competition and the effect of energy price increases on economic
activity. Journal of Money, Credit and Banking, v32n3:400–416.

Gibson, R. and Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil contin-
gent claims. Journal of Finance, v45n3 (July):959–76.

Grossman, S. J. and Laroque, G. (1990). Asset pricing and optimal portfolio choice in the presence
of illiquid durable consumption goods. Econometrica, 58:2551.

Gustafson, R. L. (1958). Carryover levels for grains: A method for determining amounts that are
optimal under specified conditions. Technical Bulletin No. 1178, U.S. Department of Agriculture,
Washington, DC.

Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and
the business cycle. Econometrica, 57:357384.

Harrison, J. M. (1990). Brownian Motion and Stochastic Flow Systems. Krieger Publishing Com-
pany, Florida.

Harrison, M., Sellke, T., and Taylor, A. (1983). Impulse control of brownian motion. Mathematics
of Operations Research, v8 n3:454–466.

Hartley, P. M. and Rogers, L. C. G. (2003). Two-sector stochastic growth models. University of

Cambridge Preprint Series.

James L. Paddock, D. R. S. and Smith, J. L. (1988). Option valuation of claims on real assets: The
case of offshore petroleum leases. Quarterly Journal of Economics, 103:479–508.

Jeanblanc-Picque, M. and Shiryaev, A. N. (1995). Optimization of the flow of dividends. Russian

Math. Surveys, v50:257–277.

Kaldor, N. (1939). Speculation and economic stability. The Review of Economic Studies, 7:1–27.

Karatzas, I. and Shreve, S. E. (1998). Methods of Mathematical Finance. Applications of Mathe-

matics 39, Springer-Verlag, New York.

Kim, C.-J. (1993). Dynamic lenear models with markov-switching. Journal of Econometrics, 60:1–
22.

Kogan, L. (2001). An equilibrium model of irreversible investment. Journal of Financial Eco-

nomics, v62 n2 November:201–45.

Litzenberger, R. H. and Rabinowitz, N. (1995). Backwardation in oil futures markets: theory and
empirical evidence. Journal of Finance, 50:1517–1545.
BIBLIOGRAPHY 130

Mamaysky, H. (2001). The term structure of interest rates and the price of durable goods. Working
Paper Yale University.

McDonald, R. and Siegel, D. (1986). The value of waiting to invest. Quarterly Journal of Eco-
nomics, v101 n4:707–27.

Merton, R. C. (1975). An asymptotic theory of growth under uncertainty. The Review of Economic
Studies, 42:375–393.

Newbery, D. M. and Stiglitz, J. E. (1981). The Theory of Commodity Price Stabilization, A Study of
the Economics of Risk. Claredon Press, Oxford.

Ross, S. A. (1997). Hedging long run commitments: Exercises in incomplete market pricing. Eco-
nomic Notes by Banca Monte, 26:99–132.

Routledge, B., Seppi, D., and Spatt, C. (2000). Equilibrium forward curves for commodities. Jour-
nal of Finance, LV n 3:1297–1338.

Scarf, H. (1960). The optimality of (s; s) policies in the dynamic inventory problem. In Mathe-
matical Methods in the Social Sciences K. Arrow, S. Karlin and P. Suppes (ed.), Stanford, CA:
Stanford University Press, pages 196–202.

Scheinkman, J. A. and Schechtman, J. (1983). A simple competitive model with production and
storage. The Review of Economic Studies, 50(3):427–441.

Schwartz, E. S. (1997). The stochastic behavior of commodity prices: Implications for valuation
and hedging. Journal of Finance, v52n3 (July):923–73.

Schwartz, E. S. and Smith, J. E. (2000). Short-term variations and long-term dynamics in commod-
ity prices. Management Science, v47n2 (July):893–911.

Sundaresan, M. (1984). Consumption and equilibrium interest rates in stochastic production

economies. Journal of Finance, 39:77–92.

Telser, L. G. (1958). Futures trading and the storage of cotton and wheat. Journal of Political
Economy, 66:233–55.

Uppal, R. (1993). A general equilibrium model of international portfolio choice. Journal of Finance,
v48 n2:529–53.

Wei, C. (2003). Energy, the stock market, and the putty-clay investment model. American Economic
Review, v93:313–323.

Williams, J. C. and Wright, B. D. (1991). Storage and Commodity Markets. Cambridge University
Press, England.

Working, H. (1948). Theory of the inverse carrying charge in futures markets. Journal of Farm
Economics, 30:128.

Working, H. (1949). The theory of the price of storage. American Economic Review, 39:1254–62.

Wright, B. D. and Williams, J. C. (1982). The economic role of commodity storage. Economic
Journal, 92(367):596–614.