Chris Brooks and Marcel Prokopczuk ICMA Centre – University of Reading

May 2011

ICMA Centre Discussion Papers in Finance DP2011-09 Copyright 2011 Brooks and Prokopczuk. All rights reserved. ICMA Centre University of Reading Whiteknights PO Box 242 Reading RG6 6BA UK Tel: +44 (0)1183 787402 Fax: +44 (0)1189 314741 Web: www.icmacentre.ac.uk Director: Professor John Board, Chair in Finance The ICMA Centre is supported by the International Capital Market Association

The Dynamics of Commodity Prices

Chris Brooks and Marcel Prokopczuk∗

May 19, 2011 Abstract

In this paper we study the stochastic behavior of the prices and volatilities of a sample of six of the most important commodity markets and we compare these properties to those of the equity market. We observe a substantial degree of heterogeneity in the behavior of the series. Our ﬁndings show that it is inappropriate to treat diﬀerent kinds of commodities as a single asset class as is frequently the case in the academic literature and in the industry. We demonstrate that commodities can be a useful diversiﬁer of equity volatility as well as equity returns. Risk measurement and options pricing and hedging applications exemplify the economic impacts of the diﬀerences across commodities and between model speciﬁcations.

JEL classification: G10, C32 Keywords: Commodity prices, stochastic volatility, jumps, Markov Chain Monte Carlo ∗

Both authors are members of the ICMA Centre, University of Reading. Contact: Marcel Prokopczuk, ICMA Centre, Henley Business School, University of Reading, Whiteknights, Reading, RG6 6BA, United Kingdom. e-mail: [email protected] Telephone: +44-118-378-4389. Fax: +44-118-931-4741.

I.

Introduction

Following the seminal work of Samuelson (1965), it is now widely accepted that commodity prices ﬂuctuate randomly. Understanding the nature of this stochastic behavior is of crucial importance for decision makers engaging in commodity markets. Yet traditionally, with the probable exception of gold, commodities have not been in the focus of investors. However, interest has grown enormously over the last decade for a variety of reasons. First, following the relatively poor performances of stocks and Treasuries, investors have sought previously unexplored asset classes as potential new sources of returns. Second, the low correlation of commodity returns with equities and their ability to provide a hedge against inﬂation make them useful additions to portfolios. The liberalization of numerous markets has also increased corporates’ requirements for hedging. This increased investment and hedging interest has led to a fast growth of the commodity derivatives market. The aim of this paper is to study the stochastic behavior of commodity prices, both from an individual perspective but also concerning the cross-market linkages between commodities, and between commodities and the equity market. As such, this is the ﬁrst piece of research to comprehensively apply a range of stochastic volatility models to commodities from several market segments. A good understanding of commodity prices’ behavior and their interdependences as well as their relation to the equity market is important for investors, producers, consumers, and also policymakers. We study the following issues for six major commodity markets. First, we investigate whether volatility does indeed behave stochastically and we estimate several models for the returns and volatility processes. Second, we examine the volatility of volatility, its persistence, and whether changes in prices and volatility are correlated. We additionally allow for jumps in both prices and volatility. We also investigate the linkages between the commodity markets considered. We determine whether volatilities are correlated across markets and whether the prices or the volatilities of diﬀerent commodities jump at the same time. Moreover, we analyze the same questions for the linkages between each commodity and the equity market. Our main ﬁndings are as follows. First, we ﬁnd that, within the stochastic volatility framework, the models that allow for jumps provide a considerably better ﬁt to the data 1

than those which do not, although there is little to choose between the models allowing for jumps in returns only and those allowing for jumps in both returns and volatility. Second, we observe alternate signs in the relationships between returns and volatilities for diﬀerent commodities – negative for crude oil and equities, close to zero for gasoline and wheat, and positive for gold, silver, and soybeans. We attribute these diﬀerences to variations in the relative balances of speculators and hedgers across the markets. We also ﬁnd evidence of considerable diﬀerences in both the intensity and frequency of jumps, although all commodities are found to exhibit more frequent jumps than the S&P 500. We conclude that commodities have very diﬀerent stochastic properties, and therefore that it is suboptimal to consider them as a single, uniﬁed asset class. To analyze the economic implications of the diﬀerences across commodities and between model speciﬁcations, as exemplars, we employ four important applications focused on risk measurement, and options pricing and hedging. Regarding the stochastic behavior of prices, equity markets, and especially equity index markets, have received a great deal of attention. The non-normality of equity returns has been documented extensively.

Motivated by the poor performance of

the Black-Scholes-Merton options pricing formula that produced the well known smile phenomenon, researchers have extended the simple Brownian motion in various directions. Merton (1976) is probably the ﬁrst to suggest adding a discontinuous jump component to the continuous Brownian price process, and Ball and Torous (1985) empirically conﬁrm the merits of this approach.

Scott (1987) and Heston (1993)

suggest modelling volatility stochastically, and the latter study is able to derive a semi-closed form solution for European option prices in the environment where volatility is stochastic. The two ideas are combined by Bates (1996) and Bakshi et al. (1997), who propose employing stochastic volatility and price jumps to improve the description of the asset’s price process with the aim of enhancing the accuracy of options pricing. Finally, Duﬃe et al. (2000) develop a general aﬃne framework for asset prices, and suggest a model with stochastic volatility, price jumps, and, additionally, jumps in volatility. Andersen et al. (2002), Chernov et al. (2003), Eraker et al. (2003), and Eraker (2004) study various versions of stochastic volatility models, with the latter two concluding that jumps in prices and volatility improve the description of S&P 500

2

price dynamics as well as the pricing of options written on the index. Asgharian and Bengtsson (2006) use this class of models to study the cross-market dependencies of various equity index markets. Compared to this array of research on equity markets, state-of-the-art empirical studies on commodity markets are sparse. Of the few examples there are of such studies, Brennan and Schwartz (1985), Gibson and Schwartz (1990), Schwartz (1997), and Schwartz and Smith (2000) study the ability of continuous-time Gaussian factor models with constant volatility to describe the stochastic behavior of the futures curve, mostly considering the crude oil market. Sorensen (2002) and Manoliu and Tompaidis (2002) apply the model of Schwartz and Smith (2000) to the agricultural and natural gas futures markets, respectively. Paschke and Prokopczuk (2009) study the cross-market dependence structure of prices in the energy segment only. Casassus and Collin-Dufresne (2005) mainly study the nature of risk premia in four diﬀerent commodity markets; they also allow for the possibility of discontinuous price jumps, which Aiube et al. (2008) conclude are important in the crude oil market. The possibility of stochastic volatility in commodities is considered by Geman and Nguyen (2005) for the soybean market, and by Trolle and Schwartz (2009), for crude oil. Finally, a three-factor incorporating prices, interest rates and the convenience yield is constructed by Liu and Tang (2011) to capture the stochastic relationship between the convenience yield level and its volatility for industrial commodities.1 Papers considering multiple commodity markets and their dependencies rely mainly on simple return analyses. For example, Erb and Harvey (2006) and Gorton and Rouwenhorst (2006) study the beneﬁts of investing in commodity markets; Kat and Oomen (2007a) and Kat and Oomen (2007b) conduct statistical analyses of commodity returns. By contrast, we evaluate the cross-linkages between commodities and between commodities and equities in a more formal, rigorous framework. The remainder of this paper is structured as follows. Section II describes the models estimated and the procedures employed to obtain the parameters, while Section III 1

There is an enormous number of papers that model time-varying volatilities and correlations within a discrete time GARCH-type framework. In the commodities area, many of these are on the oil market and focus on the objective of determining eﬀective hedge ratios. A full survey of such work is beyond the scope of this paper, but relevant studies include Serletis (1994), Ng and Pirrong (1996), Haigh and Holt (2002), Pindyck (2004), Sadorsky (2006), Alizadeh et al. (2008), and Wang et al. (2008).

3

presents and discusses the data and the empirical results. Section IV analyzes the economic implications of employing diﬀerent models in terms of Value-at-Risk and options values and hedging errors. Finally, Section V concludes, while further details of the Markov Chain Monte Carlo procedure are presented in an appendix.

II. A.

Models and Estimation

Spot Price Models

The ﬁrst model speciﬁcation employed includes stochastic volatility only, and is therefore denoted SV. The log spot price Yt = log St is assumed to follow the dynamics dYt = µ(τ )dt +

√

Vt dWtY

(1)

where WtY is a standard Brownian motion.2 To capture potential seasonal eﬀects in the price dynamics, we assume a trigonometric function for the drift component in (1) µ(τ ) = µ + η sin(2π(τ + ζ))

(2)

where τ ∈ [0, 1] denotes the time fraction of the year elapsed. The average drift rate is denoted by µ. The parameter η > 0 controls the amplitude of the function and therefore captures the strength of the seasonal eﬀect, whereas ζ governs the periodicity of the process’ drift capturing the form of the seasonality. For markets that do not show any seasonal behavior we set η = 0, yielding a constant drift. For the volatility √ Vt , we consider the square-root process dVt = κ(θ − Vt )dt + σ

√ Vt dWtV

(3)

where WtV is a second Brownian motion with dWtY dWtV = ρdt. The variance process is mean-reverting towards the long-run mean θ with speed κ. The parameter σ captures the volatility of volatility (vol-of-vol) and the kurtosis of returns increases for higher 2

Alternatively, one could specify the spot price dynamics as mean-reverting process. We have done this, however, the empirical results were inferior to the non-stationary Brownian motion.

4

values of σ. The correlation between returns and volatility captures the skewness of the returns distribution. Positive values impliy a skew to the right, and negative values a skew to the left. This speciﬁcation guarantees the positiveness of price and volatility at all times. Except for the seasonal adjustment, it is identical to the model proposed by Heston (1993). Empirical evidence suggests that a pure continuous speciﬁcation of the spot price cannot capture all salient features observed in ﬁnancial data, and in particular the possibility of rapid price movements, i.e. jumps. To allow for jumps in the price dynamics, we follow Bates (1996) and add a Poisson process NtY to speciﬁcation (1), yielding the SVJ model dYt = µ(τ )dt +

√

Vt dWtY + Zt dNtY .

(4)

The intensity λY of NtY is assumed to be constant and the jump sizes are assumed to be generated by a normal distribution, i.e. Zt ∼ N (µY , σY2 ). Assuming that jumps occur infrequently but are relatively large (which is the natural perception of jumps as opposed to the diﬀusion components in prices), the jump component will mainly aﬀect the tails of the return distribution. The speciﬁcations for the variance process and the seasonality adjustment remain unchanged. Bakshi et al. (1997) and Bates (2000) ﬁnd that although the inclusion of jumps in returns helps to describe the behavior of equity prices and the pricing of options, the model is still severely misspeciﬁed. Jumps in returns are transient, however, and hence a more persistent component is needed. Therefore, Duﬃe et al. (2000) introduce a model speciﬁcation allowing for jumps in both returns and volatility. The return process remains identical to (4), but the variance process is amended by the jump process NtV , i.e. dVt = κ(θ − Vt )dt + σ

√ Vt dWtV + Ct dNtV .

(5)

We assume that prices and volatility jump simultaneously, i.e., NtV = NtY , which is commonly denoted as a SVCJ model. This assumption is motivated by the idea that periods of stress when prices jump are often accompanied by high levels of uncertainty,

5

resulting in a jump of volatility. However, the jump sizes are not equal. Following Duﬃe et al. (2000), we assume the variance jump size to be exponentially distributed, i.e. Ct ∼ exp(µV ). To allow for dependence between the jump sizes in returns and volatility, the jumps in returns are conditionally normally distributed with Zt |Ct ∼ N (µY + ξCt , σY2 ). Due to the inclusion of the jump component in the variance process, the long-run mean changes to θ + µV λ/κ.

B.

Estimation Approach

In this section, we brieﬂy outline the Markov Chain Monte Carlo (MCMC) estimation approach we use to estimate all models. MCMC belongs to the class of Bayesian simulation-based estimation techniques. The main advantage of the MCMC methodology is the fact that it allows us to estimate the unknown model parameters and the unobservable state variables, i.e. the volatility, the jump times and the jump sizes, simultaneously in an eﬃcient way. Jacquier et al. (1994) show how MCMC methods can be used to estimate the parameters and latent volatility process of a stochastic volatility model. Johannes et al. (1999) extend this approach for jumps in returns and ﬁnally, Eraker et al. (2003) estimate the model including jumps in returns and volatility via MCMC methods.3 In order to be able to estimate the models, it is necessary to express them in discretized form. Using a simple Euler discretization, the SVCJ model is given as4 Yt = Yt−∆t + µ(τ )∆t +

√ Vt−∆t εYt + Zt Jt

and Vt = Vt−∆t + κ(θ − Vt−∆t )∆t + σ

√ Vt−∆t εVt + Ct Jt .

(6)

(7)

The innovations εYt and εVt are normal random variables, i.e. εYt ∼ N (0, ∆t), and εVt ∼ N (0, ∆t) with correlation ρ. The jump times Jt take the value one if a jump occurs and zero if not, i.e. Jt ∼ Ber(λ∆t). The discretized versions of the SVJ and SV models are obtained by dropping the respective jump components. In the following, we 3 4

For an excellent introduction to MCMC estimation techniques, see Johannes and Polson (2006). As we work with daily data, the discretization bias is negligible.

6

set ∆t = 1, i.e. one day. The general idea of the MCMC methodology is to break down the high dimensional posterior distribution into its low dimensional complete conditionals of parameters and latent factors which can be eﬃciently sampled from. The posterior distribution p(Θ, V, J, Z, C|Y ) provides sample information regarding the unknown quantities given the observed quantities (prices). By Bayes rule we have p(Θ, V, J, Z, C|Y ) ∝ p(Y |V, J, Z, C, Θ)p(V, J, Z, C|Θ)p(Θ)

(8)

where Y is the vector of observed log prices; V , J, Z, and C contain the time series of volatility, jump times and jump sizes, respectively; Θ is the vector of model parameters; p(Y |V, J, Z, C, Θ) is usually called the likelihood, p(V, J, Z, C|Θ) provides the distribution of the latent state variables, and p(Θ) the prior, reﬂecting the researcher’s beliefs regarding the unknown parameters. To keep the inﬂuence of the priors small, we specify extremely uninformative priors. As p(Θ, V, J, Z, C|Y ) is high dimensional, it is not possible to directly sample from it. Therefore, it is necessary to simplify the problem by breaking down the posterior distribution into its complete conditional distributions which fully characterize the joint posterior. Whenever possible, we use conjugate priors which allow us to directly sample from the conditional. If this is not possible, we rely on a Metropolis algorithm. For more details on the precise speciﬁcations, see the Appendix. The

output

of

the

simulation

procedure

is

a

set

of

G

draws

{Θ(g) , V (g) , J (g) , Z (g) , C (g) }g=1:G , that forms a Markov Chain and converges to p(Θ, V, J, Z, C|Y ). Estimates of the parameters, the volatility paths, and the jump sizes are obtained by simply taking the mean of the posterior distribution. For the jump times, one additional step is required. As each draw of J is a set of Bernoulli random variables, i.e. taking on the value one or zero, the mean over all draws will provide a time series of jump probabilities. To obtain estimates for the jump times, we follow Johannes et al. (1999) and identify the jump times by choosing a threshold

7

probability, i.e. we estimate the jump times Jˆ as 1 if p(J = 1) > α t ˆ Jt = 0 if p(J = 0) ≤ α t

(9)

The threshold α is chosen such that the number of jumps identiﬁed corresponds to the estimate of the jump intensity λ.

III. A.

Data and Empirical Results

Data

In this paper, we employ daily spot price data for a variety of commodities traded in the US. The series are chosen to reﬂect the relative importance of those markets, and also to ensure the availability of as long a span of high quality data as possible. We consider two energy commodities, namely crude oil (CL) and gasoline (HU). From the metal markets, we include gold (GC) and silver (SI) in the study. Finally, we consider soybeans (S) and wheat (W) as representatives of the agricultural commodities market. All commodity data are obtained from the Commodity Research Bureau.5 Additionally, we employ S&P 500 index data (obtained from Bloomberg) to enable us to put the results into the perspective of the existing literature on equity dynamics and to investigate the relationship of commodity and equity markets in Subsection D. The data period covered is more than 25 years, spanning January 1985 to March 2010, and yielding 6290 observations per commodity and for the S&P 500. Table 1 provides descriptive summary statistics and Figure 1 shows time series of the six commodity markets considered. Several points are worth noting. Compared to the S&P 500, the mean return is smaller for all commodities. The lowest average return is observed for the wheat market, which is barely positive, and the highest for the gold market. The standard deviation is higher than the S&P 500 for all but the gold market. The highest levels of volatility are observed for the energy commodities, with annualized values of 42.82 % and 44.13 %, which is more than twice the volatility of 16.00 % and 18.92 % observed in the gold and the equity markets. The kurtosis 5

See www.crbtrader.com.

8

is, however, the highest for the S&P 500, while the lowest values are observed in the agricultural markets. However, these values are still higher than can be explained by a simple normal distribution. The smallest and largest returns are observed for the energy markets, both twice as large as for the S&P 500, which is to some extent surprising as the kurtosis levels are substantially lower. The gold and soybeans markets exhibit the smallest range between the minimum and maximum values.

TABLE 1 AND FIGURE 1 ABOUT HERE

B.

Univariate Analysis

Table 2 reports the estimation results for the SV model, i.e. the stochastic volatility model without jumps. All but the parameters aﬀecting the mean equation and some of the correlations are statistically signiﬁcant. The estimates for the speed of mean reversion, κ, are of similar size in all commodity markets and are also comparable to those of the S&P 500. Only the estimate for the silver market is somewhat higher, indicating a slightly less persistent volatility process. The long term variance levels, θ, of all but the gold market are signiﬁcantly higher than the corresponding equity estimates. √ The S&P 500 annualized long term volatility is 17.58 % (= θ · 252), which is very close to the unconditional volatility, and also of similar size as in other studies.6 The highest levels of annualized long term volatility are observed for the crude oil and gasoline markets, at 36.00 % and 37.61 % respectively. The diﬀerence across the commodity markets is considerable. The vol-of-vol parameters, σ, are highly signiﬁcant, which conﬁrms the stochastic nature of volatility in the commodity markets considered. Again, with the exception of the gold market, higher values than for the equity market are observed, implying heavier tails and greater deviance from normality. Moreover, there are substantial diﬀerences across markets – for example, the crude oil series exhibits a vol-of-vol which is twice as big as for soybeans. Figures 2 plots the ﬁtted volatility process in annualized percentages obtained from the SV model for the six commodities. It is evident that the crude oil and gasoline 6

For example, Eraker et al. (2003) report 15.10 % for the period 1985 to 1999.

9

series are the most variable overall, while gold in particular is the least. The oil-based products showed two particular spikes in volatility during 1986 and 1990, both of which can be tied to economic events that occurred at those times. In March 1986, Saudi Arabia changed its policy and signiﬁcantly expanded production to increase market share; other OPEC members followed, leading to price drops in both crude oil and gasoline. August 1990 marked the start of the ﬁrst gulf war when Iraq invaded Kuwait. Furthermore, we can observe the eﬀect of Hurricane Katrina hitting the US gulf coast in August 2005, leading to a short-term supply bottleneck in gasoline (although it had little noticeable eﬀect on the price of crude oil). Wheat was the only commodity in the sample that showed an upward trend in volatility from the late 1990s onwards. Moreover, it is also of interest to note that the Lehman default of September 2008 increased levels of volatility and uncertainty across all markets.

TABLE 2 AND FIGURE 2 ABOUT HERE

The estimates for the correlation of the underlying and the variance process, ρ, are most interesting, as we can observe diﬀerent signs for diﬀerent markets. The crude oil market shows a negative correlation; the gasoline and wheat markets (almost) zero correlations; and the gold, silver, and soybean markets signiﬁcant positive correlations. For equity markets, ρ is usually found to be negative, which is conﬁrmed by the estimate of -0.57. The negativity of ρ in equity markets is usually explained by the leverage eﬀect, as discussed by Black (1976). However, this argument does not apply to commodity markets. A possible explanation for the diﬀerent correlations might be the fraction of hedgers and speculators in the markets. Assuming that hedging activities reduce volatility, whereas speculation activity increases volatility, this line of argument would indicate that in the crude oil market, the fraction of hedgers over speculators increases with rising prices. In the metal and the soybean markets, the opposite would hold true – that is, more hedging takes place for lower prices, while speculation dominates in times of higher than average prices. Table 3 provides the parameter estimates for the SVJ model, i.e. the model including Poisson price jumps. The rates of mean reversion, κ, substantially decrease in all cases 10

and are now lower for every commodity compared to the S&P 500, implying a higher persistence of the volatility process. The long term volatility levels (of the diﬀusion component) decrease, which is a technical eﬀect, as part of the price variation is now captured by the jump component. Interestingly, however, the change for the S&P 500 is minimal compared to the commodity markets. The vol-of-vol parameters, σ, also decrease for the same reason in that part of the excess kurtosis is now captured by the jump component. Again, the changes in the commodity markets are much bigger than for the equity case, indicating that the jump component plays an even bigger role in these markets. The correlation estimates become more pronounced compared to the SV model, i.e. the negative values decrease, whereas the positive values increase.

TABLE 3 ABOUT HERE

The jump intensity estimates, λ, lie between 0.022 for wheat and 0.088 for silver, providing evidence of signiﬁcant diﬀerences between markets. In the oil, gasoline, and wheat markets, jumps occur about six times per year. In the soybeans and gold market, we annually observe 12 to 13 jumps on average, whereas the silver price jumps about 22 times per year. These numbers are signiﬁcantly larger than those observed for equity indices. In our sample period, we estimate the S&P 500 to jump only 1.8 times per year (Eraker et al. (2003) estimated 1.5 times per year for their sample period). The mean jump sizes, µJ , are negative for all markets, although only signiﬁcant in some instances. Taking into account the standard deviation of jump sizes, σJ , one can observe a huge variation. A two-sigma interval covers almost as much probability mass on the positive line as on the negative. The most notable cases are the two energy markets. A plus two-sigma event in these two markets corresponds to a +14 % and +12 % (daily!) price jump, respectively. Analogously, a minus two-sigma event corresponds to price drops of -16 % and -14 %. The results for the other commodity markets are qualitatively similar, but less extreme. For the equity market, on the other hand, the mean jump size is signiﬁcantly negative, and 75 % of the probability mass lies on the negative line. Lastly, Table 4 reports the estimates for the SVCJ models, i.e. the models including contemporaneous jumps in prices and volatility. The speed of mean-reversion, κ, reverts 11

back to values comparable to or even higher than in the SV model, indicating a lower persistence when including jumps. This makes intuitive sense as the inclusion of jumps induces the need for higher mean-reversion speeds directly after jumps (as only positive jumps are allowed). The long-term volatility levels further decrease, as part of the variation is now captured by the volatility jump component. We still observe signiﬁcant diﬀerences across markets, ranging from 10 % for the gold market to 27.5 % for the gasoline market. The equity market lies, at 14 %, closer to the lower border of this interval. The vol-of-vol parameter, σ, slightly decreases for most markets; the silver market is a notable exception, where σ increases from 0.16 to 0.21. The correlation of the diﬀusion components mostly decrease in absolute terms; as prices and volatility are always jumping contemporaneously, part of the correlation is now captured by the jump components. Interestingly, the jump intensities decrease signiﬁcantly in some instances, such as the gold, silver, and soybean markets, whereas it remains almost unchanged in the crude oil market. Changes in the average price jump and its volatility are mostly relatively small. This is opposed to the corresponding result for the S&P 500, where the mean jump size almost doubles. The jump dependence parameter ξ is not estimated very precisely, a phenomenon which has already been observed by Eraker et al. (2003). As the values are close to zero, only modest dependence between the jump sizes is observed. The average volatility jump size lies between 0.51 for the gold, and 4.04 for the gasoline market.

TABLE 4 ABOUT HERE

Although model choice is not the main focus of this paper, in order to compare the ﬁt of the three diﬀerent models to the data, we make use of the Deviance Information Criterion (DIC) proposed by Spiegelhalter et al. (2002) and in particular applied to stochastic volatility models by Berg et al. (2004). The DIC can be regarded as a generalization of the Akaike Information Criterion (AIC), and trades oﬀ adequacy and complexity in a similar fashion.7 As for the AIC, smaller DIC values indicate a better model ﬁt. 7

AIC equals DIC in the special case of ﬂat priors.

12

The DIC scores are reported in Table 5. First, we can observe that the ﬁt of the SV model is by far the worst for every market, providing evidence of the beneﬁts of including a price jump component. Comparing the DIC scores of the SVJ and the SVCJ models is less conclusive as most of the scores are close to each other for a given commodity. However, the SVJ values are always lower for all commodity markets whereas the SVCJ obtains the lowest score for the S&P 500. As the values for the SVJ and SVCJ models are always close together and it is therefore diﬃcult to make a ﬁnal decision on which model to prefer, we will, in the spirit of robustness, use both in the subsequent analysis.

TABLE 5 ABOUT HERE

C.

Cross-Commodity Market Analysis

The MCMC estimation approach employed not only provides us with parameter estimates but also with estimates for the latent state variables, i.e. volatility, jump times, and jump sizes.

We can use this information to analyze the cross-market

dependence structure of these state variables.8 correlations of the diﬀerences in volatility, i.e.

√

In a ﬁrst step, we calculate the √ Vt − Vt−1 , to analyze the extent

to which the volatilities in the various markets move together. Table 6 reports these correlations, as well as return correlations to enable us to put the volatility results into perspective.

TABLE 6 ABOUT HERE

As expected, return correlations are highest between related commodities belonging to the same market segment. However, this correlation is still far from perfect. In particular, the moderate degree of correlation between crude oil and gasoline of 0.49 8

Alternatively, one could try to estimate a multivariate version of the model employed in order to analyze these dependencies. This approach is, however, computationally infeasible as it would involve a 14x14 covariance matrix when considering all seven markets together.

13

is interesting. The strongest correlation is observed for gold and silver with a value of 0.68. Return correlations between commodities of diﬀerent segments are weak or close to zero, never exceeding 0.11. Looking at the volatility correlations, one can identify a similar pattern. There are, however, some diﬀerences. For example, the volatility correlation between soybeans and wheat is substantially smaller than the corresponding return correlation, indicating that the prices move more closely together than the volatilities in these markets. Another interesting point is the correlation of crude oil with the agricultural commodities. The returns are mildly correlated, indicating some dependence across markets; the volatilities’ correlations are, on the other hand, close to zero. Comparing the results for the SVJ and the SVCJ model, it is interesting to observe that for some instances, the estimated correlation is almost identical (e.g. CL-HU or S-W), whereas in other instances, the results changes substantially (e.g. GC-SI). This might be a consequence of the fact that the number of jumps identiﬁed in the gold and silver markets changes substantially between the two model variants (this can be seen from the changes in the estimated value of λ). Next, we analyze the simultaneous jump probabilities for each pair of commodities. To do this, we simply count the numbers of simultaneous jumps and divide it by the ∑T

sample size T , i.e., we calculate the following quantity:

t=1

T

Jˆti Jˆtj

. To put these numbers

into perspective, we also compute the probability of a simultaneous jump assuming independence which is given by the product of the two jump intensities. Table 7 provides the results.

TABLE 7 ABOUT HERE

We can observe that the simultaneous jump probabilities of commodities belonging to the same segment (i.e. the pairs CL-HU, GC-SI, and S-W) are about 4-10 times higher than one would expect under independence. For all other pairs, the probabilities are very similar, indicating that jumps of non-related commodities are independent of each other. Compared to the results for the dependence in volatility, there are some interesting diﬀerences to be observed. 14

The simultaneous jump probability

between soybeans and wheat is halved from 0.46 % to 0.21 % when including jumps in volatility. This is in contrast to the volatility correlation, which remains almost equal. Furthermore, one can observe a substantial decrease of simultaneous jump probabilities for most cases, whereas the correlations of volatilities remain rather constant, or even increase, when including volatility jumps.

D.

Commodity and Equity Markets

We now investigate the dependencies between the individual commodity markets and the equity market, i.e.

the S&P 500.

Table 8 reports the return and volatility

correlations. It is well known that return correlations between commodities and equities are quite low, which is the reason why commodities are often considered to be a diversiﬁer for a traditional portfolio of stocks and bonds. From Table 8, it can be seen that the same holds true for volatility, i.e. equity and commodity volatilities are almost uncorrelated. Consequently, commodities not only serve as a return diversiﬁer, but as a volatility diversiﬁer at the same time.

TABLE 8 ABOUT HERE

Table 9 displays the daily simultaneous jump probabilities in equity and commodity markets as well as the corresponding probabilities assuming independence. For the SVJ model, the greatest diﬀerence is observed for the gold (0.13 % vs 0.04 %) and oil (0.10 % vs 0.02 %) markets, indicating that jumps in these two commodies are related to jumps in the equity market. When considering the results for the SVCJ model, one can see that the joint jump probabilities are all very small. However, compared to the previous results for the SVJ model, the diﬀerences from the case assuming independence become more pronounced, indicating that the likelihood of a simultaneous jump in volatility is higher than a simultaneous jump in prices. This is especially true for the soybeans market.

TABLE 9 ABOUT HERE

15

When a simultaneous jump occurs, it might be that the prices in both markets jump in the same, or alternatively, in opposite directions. Moreover, the jumps might be more or less pronounced than average jumps. To investigate this issue, we compute the average jump sizes for each commodity market conditional on the occurrence of a jump in the equity market and vice versa. Table 10 reports the results. Interestingly, the average jump size in the equity market is much more negative than the unconditional average jump size of −2.59 (SVJ model). Surprisingly, this eﬀect is strongest for the soybean and wheat markets, i.e. a jump in the agricultural markets induces an average jump size which is 2-3 times bigger than normal. A consideration of the average jump sizes in the commodity markets given that the equity market jumps reveals that the energy commodities tend to jump upward if the equity market jumps downwards. This is diﬀerent to their unconditional average jump size which is negative. The other commodity jumps react less to jumps in the equity market; the results for the SVCJ model are qualitatively similar but with some diﬀerences in terms of strength of the eﬀects.

TABLE 10 ABOUT HERE

IV. A.

Economic Implications

Risk Management: Value-at-Risk

To analyze how the diﬀerences of the commodity dynamics impact upon economic quantities of interest, we ﬁrst consider a simple risk management application and compute the Value-at-Risk (V aR) for a position in each commodity.9

The V aR

estimates are obtained via Monte Carlo simulation. For each commodity, we simulate 100,000 price (and volatility) paths using the parameter estimates obtained in the previous section. The time horizon for the analysis is one year; to avoid negative volatilities due to the discrete implementations of the models, we choose a time step of 1/10th of a day (and rescale the daily parameters accordingly). 9

For a recent consideration of the impact of the distribution of commodity returns and model choice on Value-at-Risk, see Cheng and Hung (2011).

16

TABLE 11 ABOUT HERE

Table 11 reports the V aRs for various conﬁdence levels. The V aRα is calculated as the diﬀerence of the initial value of the position, which is normalized to 100, and the α%-quantile of the proﬁt and loss distribution. First, one can observe that the V aR increases when jumps are included in the model, which is to be expected since jumps increase the kurtosis of the return distribution. For all commodities, we can observe an increase when including contemporaneous jumps in volatility. Interestingly, this is not the case for the S&P 500. Here, slightly decreased V aR values are obtained. Most likely, this result is a consequence of the fact that, in this case, the jump intensity in the SVJ model is almost twice as high as for the SVCJ model. This demonstrates that in terms of price risk management, the inclusion of jumps in volatility is crucial for most commodities considered but less important for the S&P 500. When comparing the V aR values across commodities, one can immediately notice substantial diﬀerences. As expected, gold exhibits by far the lowest values. The energy commodities exhibit the highest values, a result which is mainly driven by their overall volatility levels, the higher standard deviations of jumps and the negative correlations of returns and volatility compared to the other assets. It is interesting to see that the two agricultural commodities exhibit quite diﬀerent behavior, however, with silver lying somewhere in the middle. On one hand, silver’s strong positive correlation of returns and volatility acts as a risk reducer; on the other hand, it exhibits a high jump intensity and a high volatility of jump sizes which tend to increase the V aR.

B.

Options Valuation

One of the advantages of the continuous time models employed in this study is that semi closed-form options pricing formulas exist. These formulas are based on Fourier analysis, which was initially proposed by Heston (1993). Duﬃe et al. (2000) provide general solutions for all aﬃne models and for the SVCJ model in particular. In the following, we use the model parameters estimated in the previous section to analyze the implications for the value of European call options. As we have estimated all parameters from historical price data under the physical measure, we need to make 17

an assumption regarding the market price of volatility and jump risk. Broadie et al. (2007) review the most recent literature on the sign and size of these risk premia for the S&P 500, concluding that the evidence is inconclusive; most studies report insigniﬁcant risk premia from a statistical and/or economic point of view. Due to these ﬁndings, and for simplicity, in the following we assume zero volatility and jump risk premia.10 To make the results comparable, we assume that each underlying currently trades at S = 100. The current variance level is assumed to be equal to its long run average, and the risk-free rate is assumed to be zero. Using the estimated parameter values, we calculate the values of at-the-money (ATM; X = S), in-the-money (ITM; X = 0.95S), and out-of-the-money (OTM; X = 1.05S) call options where X denotes the strike price. Table 12 reports the results of this computation. The left columns report the value of the option, the right columns express the option’s value relative to the value of the corresponding ATM option.

TABLE 12 ABOUT HERE

One can observe large diﬀerences across assets. Naturally, options written on the underlyings with the highest volatility levels are most expensive. The value of the gold and S&P 500 OTM options is already close to zero (although they are only 5% out-of-the-money), whereas the values for crude oil and gasoline remain at higher levels. The diﬀerences in jump intensities, jump amplitudes and the correlations of returns with volatilities lead to some interesting diﬀerences in the relative values of ITM and OTM options across commodities.11 For example, comparing the soybeans and wheat markets, one can observe that the relative value of ITM options is higher for soybeans, whereas the relative value of OTM options is higher for wheat. Comparing the prices for the SV and SVJ with the SVCJ model, one can see that prices are lower for the latter, i.e. neglecting the possibility of volatility jumps leads to increased option prices. To further compare the consequences of the diﬀerent behavior of the commodities 10 The results would, of course, change if this assumption were not valid, especially if the nature of the risk premia diﬀers across commodities. This is a non-trivial question which we leave for future research. 11 When considering implied volatilities instead of relative values, these diﬀerences would manifest themselves in the shape of the volatility smile/skew.

18

considered for options valuation, we analyze the values of the most popular exotic contracts, i.e. barrier options. To save space, we focus our attention on at-the-money down-and-out call and put options and vary the value of the barrier between 90 % and 100 % of the current spot level. To take the diﬀerent levels of volatility in the markets into account, we normalize the value of the barrier option by the value of a corresponding plain-vanilla option. Consequently, all resulting values must be between zero and one.

FIGURE 3 ABOUT HERE

Figure 3 displays the results of these computations. One can observe that, especially for the down-and-out put options, the value is quite distinct across commodities, even though we have already adjusted for the absolute volatility level by normalizing with the values of plain-vanilla options. This is due to the fact that the probability of hitting the barrier is more sensitive to the volatility and jump level than the simple option value. Inspecting the ﬁgures more closely, one can observe diﬀerences between model speciﬁcations. For example, the relative value of the down-and-out put option on wheat for a barrier of 90 % is around 0.4 for the SV and SVJ models but increases to 0.5 in the SVCJ model, whereas the value of the corresponding silver option remains at 0.6 for all three models. From this result, one can deduce that the inclusion of jumps in the volatility dynamics is more relevant for wheat options. A similar observation can be made for soybeans and the energy commodities.

C.

Options Hedging

Lastly, we investigate the consequences of the diﬀerent stochastic behavior of commodities and diﬀerent models in terms of the delta hedging of options. Consider, for example, the inﬂuence of the correlation parameter. Depending on the options position, one is long or short the underlying and long or short volatility. For example, a short put position translates into a long position in the underlying and a short position in volatility. Thus, depending on the sign of the correlation, one is either naturally diversiﬁed or not. By implementing a standard delta hedge, one neglects this 19

relationship, which will have consequences on the hedging outcome. Similarly, the jump frequency as well as the mean and standard deviation of the jump sizes will impact upon the hedging success. We set up our simulation analysis as follows. We consider a short position in an ATM put option with a maturity of one month. The hedging horizon τ is one week, the current variance level is assumed to be equal to the long run mean, and interest rates are assumed to be zero. To investigate the potential hedging error, we assume that the true price dynamics are given by either the SV, the SVJ, or the SVCJ model and we calculate the corresponding options value.12 We then consider a standard delta hedging strategy, i.e. we calculate the option’s delta using Black’s formula to set up the hedging portfolio. Next, we simulate the price and volatility dynamics until τ and calculate the new option value to obtain the payoﬀ of the hedging portfolio. This procedure is repeated 10,000 times to obtain a distribution of hedging errors.

TABLE 13 ABOUT HERE

Table 13 reports the mean, standard deviation, skewness, and kurtosis of the hedging errors. The mean hedging error is negative in all cases and larger (in absolute terms) for the energy commodities, which is as expected due to the higher volatility levels. The same observation applies to the standard deviation of the errors. When considering the skewness and kurtosis, one can observe some interesting diﬀerences across commodities and models. Considering the SV model ﬁrst, we see that all distributions are skewed to the left, i.e. big negative outcomes are more likely than big positive outcomes. Interestingly, the skews for the agricultural commodities and the S&P 500 are bigger than for crude oil. Moreover, looking at the kurtosis of the hedging errors, it becomes clear that crude oil exhibits the thinnest tails, i.e. extreme negative (and positive) outcomes are less likely than for the other markets. The picture completely changes when introducing jumps in the price and variance processes. Whereas the mean and volatility of the hedging errors remain at similar levels, or even decrease, the skewness, and, most importantly, the kurtosis, drastically 12

As in the previous section, we assume a risk premium of zero.

20

increase. This holds true for all markets, but with diﬀerent rates. Under the SVCJ model, the kurtosis of crude oil rockets from 2.6 to more than 38, which is now more than twice the value observed for the S&P 500. By contrast, the rise in kurtosis for the soybeans case is least dramatic with an increase from 4.4 to 8.6. Overall, one can make out substantial diﬀerences across commodities and models when considering the third and fourth moments of the hedging errors, which translate into a signiﬁcant degree of model risk when delta hedging in these markets.

V.

Conclusions

This paper has examined the stochastic behavior of the prices and volatilities of a sample of six of the most important commodity markets. Using a Bayesian Markov Chain Monte Carlo estimation approach, three separate stochastic volatility-type models are estimated and compared for each commodity price series, and for the S&P 500 by way of comparison. Fairly intuitively, correlations between the returns are high fore pairs of commodities from the same sub-class but almost zero across sub-classes. The same pattern holds for the relationships between the simultaneous jump probabilities: within market segments, jumps often occur in tandem whereas they are essentially independent across segments. We are able to demonstrate that not only are return correlations between commodities and the stock index low, as is well documented, but the correlations between commodity and stock volatilities are also low. This is an important result since it shows that commodities may be an even more useful portfolio constituent than previously thought as they can act as a volatility diversiﬁer as well, which is important for options portfolios or any portfolio of securities with embedded contingent claims. The paper examines whether jumps occur simultaneously across commodities and equities; there is some evidence that jumps in the two asset classes do occur together, notably for gold and oil, although the results vary somewhat between models. Finally, we test the economic impact of employing one stochastic volatility model rather than another by estimating value at risk, by considering the prices of European calls and barrier options written on each commodity, and by evaluating the eﬀectiveness

21

of delta hedging.

Again, we ﬁnd substantial diﬀerences both across commodities

and between models, indicating the heterogeneous nature of this asset class and the importance of judiciously choosing the most appropriate speciﬁcation for each individual series.

22

Appendix A: MCMC Estimation Details This Appendix provides more detailed information on the MCMC estimation procedure. For a general introduction to MCMC techniques see, e.g., Koop (2003). The Gibbs sampling technique allows one to draw each parameter and state variable of the joint posterior sequentially. For many parameters, conjugate priors can be used to derive the conditional posterior distribution, which is a standard distribution and therefore easily sampled from. Details are given below: µ: we use a standard normal distribution as the prior. κ and θ:

we use a truncated normal distribution bounded at zero and

hyperparameters of 0 and 1 as priors for each parameter. σ and ρ:

as the posteriors for σ and ρ are not known, we follow the

reparametrization suggested by Jacquier et al. (1994) and use the inverse Gamma distribution with hyperparameters 1 and 200 as priors. η: we use an exponential distribution with a hyperparameter of 0.05 as a prior, yielding a truncated normal distribution as a posterior. ζ: the posterior distribution for ζ is non-standard, and we apply an independence Metropolis algorithm, drawing from the uniform distribution on the unit interval. λ: we use a Beta distribution with hyperparameters 2 and 40. µV : we use a Gamma distribution with hyperparameters 1 and 1. µJ : we use a Normal distribution with hyperparameters 0 and 100 as priors. σJ2 : we use an Inverse Gamma distribution with hyperparameters 5 and 20 as priors. ξ: we use a standard normal distribution as a prior. The posteriors of Jt , Zt and Ct are non-standard but provided in the appendix of Eraker et al. (2003). The conditional posterior distribution of the variance path V is also non-standard and given by p(V |Y, J, Z, C, Θ) ∝

T ∏

p(Vt |Vt−1 , Vt+1 , Y, J, Z, C, Θ)

(10)

t=1

where the posterior function for each Vt is given as p(Vt |Vt−1 , Vt+1 , Y, J, Z, C, Θ) ∝ 23

1 − 1 (ω1 +ω2 +ω3 ) e 2 Vt

(11)

with (Yt+1 − µ(τ ) − Jt+1 Zt+1 )2 Vt (Vt − θκ − (1 − κ)Vt−1 − σρ(Yt − µ(τ ) − Jt Zt ) − Jt Ct )2 = σ 2 Vt−1 (1 − ρ2 ) (Vt+1 − θκ − (1 − κ)Vt − σρ(Yt+1 − µ(τ ) − Jt+1 Zt+1 ) − Jt+1 Ct+1 )2 = σ 2 Vt (1 − ρ2 )

ω1 =

(12)

ω2

(13)

ω3

(14)

We use a random walk Metropolis algorithm to sample from. The volatility of the error term in this procedure is calibrated such that the acceptance probability is within the range 30-50%. See Koop (2003) on details of calibrating the Random Walk Metropolis algorithm. Each parameter (and state variable) is sampled 100,000 times (i.e., G = 100, 000) and we discard the ﬁrst 30,000 ’burn-in’ draws as is standard practice with MCMC estimations.

24

References Aiube, F., Baidya, T., and Tito, E. (2008). Analysis of commodity prices with the particle ﬁlter. Energy Economics, 30:597–605. Alizadeh, A., Nomikos, N., and Pouliasis, P. (2008). A markov regime switiching approach for hedging energy commodities.

Journal of Banking and Finance,

32:1970–1983. Andersen, T., Benzoni, L., and Lund, J. (2002).

An empirical investigation of

continuous-time equity return models. Journal of Finance, 62:1239–1284. Asgharian, H. and Bengtsson, C. (2006). Jump spillover in international equity markets. Journal of Financial Econometrics, 4:167–203. Bakshi, G., Cao, G., and Chen, C. (1997). Empirical performance of alternative option pricing models. Journal of Finance, 52:2003–2049. Ball, C. and Torous, W. (1985). On jumps in common stock prices and their impact on call option pricing. Journal of Finance, 40:155–173. Bates, D. (1996). Jumps and stochastic volatility: Exchange rate processes implicit in Deutschemark options. Review of Financial Studies, 9:69–108. Bates, D. (2000).

Post-’87 crash fears in S&P 500 futures options.

Journal of

Econometrics, 94:181–238. Berg, A., Meyer, R., and Yu, J. (2004). Deviance information criterion for comparing stochastic volatility models. Journal of Business & Economic Statistics, 22:107–120. Black, F. (1976). Studies of stock price volatility changes. In Proceedings from the American Statistical Association, Business and Economic Statistics Section, pages 177–181. American Statistical Association. Brennan, M. and Schwartz, E. (1985). Evaluating natural resource investments. Journal of Business, 58:135–157.

25

Broadie, M., Chernov, M., and Johannes, M. (2007). Model speciﬁcation and risk premia: Evidence from futures options. Journal of Finance, 62:1453–1490. Casassus, J. and Collin-Dufresne, P. (2005). Stochastic convenience yield implied from commodity futures and interest rates. Journal of Finance, 60:2283–2331. Cheng, W.-H. and Hung, J.-C. (2011). Skewness and leptokurtosis in GARCH-typed VaR estimation of petroleum and metal returns. Journal of Empirical Finance, 18:160–173. Chernov, M., Ghysels, E., Gallant, A., and Tauchen, G. (2003). Alternative models for stock price dynamics. Journal of Econometrics, 116:225–257. Duﬃe, D., Pan, J., and Singleton, K. (2000). Transform analysis and asset pricing for aﬃne jump-diﬀusions. Econometrica, 68(6):1343–1376. Eraker, B. (2004). Do stock prices and volatility jump? reconciling evidence from spot and option prices. Journal of Finance, 59:1367–1403. Eraker, B., Johannes, M., and Polson, N. (2003). The impact of jumps in volatility and returns. Journal of Finance, 58:1269–1300. Erb, C. and Harvey, C. (2006). The strategic and tactical value of commodity futures. Financial Analysts Journal, 62:69–97. Geman, H. and Nguyen, V.-N. (2005). Soybean inventory and forward curve dynamics. Management Science, 51:1076–1091. Gibson, R. and Schwartz, E. (1990). Stochastic convenience yield and the pricing of oil contingent claims. Journal of Finance, 45:959–976. Gorton, G. and Rouwenhorst, K. (2006). Facts and fantasies about commodity futures. Financial Analysts Journal, 62(2):47–68. Haigh, M. and Holt, M. (2002). Crack spread hedging: accounting for time-varying volatility spillovers in the energy futures markets. Journal of Applied Econometrics, 17(3):269–289.

26

Heston, S. (1993). A closed–form solution for options with stochastic volatility with applications to bond and currency options. Review of Financial Studies, 6:327–343. Jacquier, E., Polson, N., and Rossi, P. (1994). Bayesian analysis of stochastic volatility models. Journal of Business and Economic statistics, 12:371–389. Johannes, M., Kumar, R., and Polson, N. (1999). State dependent jump models: How do us equity indices jump. Working Paper. Johannes, M. and Polson, N. (2006). MCMC methods for ﬁnancial econometrics. In AitSahalia, Y. and Hansen, L., editors, Handbook of Financial Econometrics. Elsevier. Kat, H. and Oomen, R. (2007a). What every investor should know about commodities part i: univariate return analysis. Journal of Investment Management, 5:4–28. Kat, H. and Oomen, R. (2007b). What every investor should know about commodities part ii: multivariate return analysis. Journal of Investment Management, 5:40–64. Koop, G. (2003). Bayesian Econometrics. Wiley. Liu, P. and Tang, K. (2011).

The stochastic behavior of commodity prices

with heteroskedasticity in the convenience yield. Journal of Empirical Finance, 18:211–224. Manoliu, M. and Tompaidis, S. (2002). Energy futures prices: term structure models with Kalman ﬁlter estimation. Applied Mathematical Finance, 9:21–43. Merton, R. (1976). Option pricing when underlying stock returns are discontinuous. Journal of Financial Economics, 3:125–144. Ng, V. and Pirrong, S. (1996). Price dynamics in reﬁned petroleum spot and futures markets. Journal of Empirical Finance, 2:359–388. Paschke, R. and Prokopczuk, M. (2009). Integrating multiple commodities in a model of stochastic price dynamics. Journal of Energy Markets, 2(3):47–82. Pindyck, R. (2004). Volatility and commodity price dynamics. Journal of Futures Markets, 24:1029–1047. 27

Sadorsky, P. (2006). Modeling and forecasting petrolum futures volatility. Energy Economics, 28:467–488. Samuelson, P. (1965).

Proof that properly anticipated prices ﬂuctuate randomly.

Industrial Management Review, 6:41–49. Schwartz, E. (1997). The stochastic behavior of commodity prices: Implications for valuation and hedging. Journal of Finance, 52:923–973. Schwartz, E. and Smith, J. (2000). Short-term variations and long-term dynamics in commodity prices. Management Science, 46:893–911. Scott, L. (1987).

Option pricing when the variance changes randomly: Theory,

estimation, and an application. Journal of Financial and Quantitative Analysis, 22:419–438. Serletis, A. (1994).

A cointegration analysis of petroleum futures prices. Energy

Economics, 16:93–97. Sorensen, C. (2002). Modeling seasonality in agricultural commodity futures. Journal of Futures Markets, 22(5):393–426. Spiegelhalter, D., Best, N., Carlin, B., and van der Linder, A. (2002). Bayesian measures of model complexity and ﬁtting. Journal of the Royal Statistical Society B, 64:583– 639. Trolle, A. and Schwartz, E. (2009). Unspanned stochastic volatility and the pricing of commodity derivatives. Review of Financial Studies, 22:4423–4461. Wang, T., Wu, J., and Yang, J. (2008). Realized volatility and correlation in energy futures markets. Journal of Futures Markets, 28:993–1011.

28

Crude oil

Gasoline

150 3 100 2 50

0 1985

1

1995

0 1985

2005

Soybeans

1995

2005

Wheat 1200

1500 1000 800

1000

600 400

500

200 0 1985

1995

0 1985

2005 Gold

1995

2005

Silver

1200

2000

1000 1500

800 600

1000

400 500 200 0 1985

1995

0 1985

2005

1995

2005

Figure 1: Price Series This ﬁgure shows the historical price series for the six commodities considered. All ﬁgures are in US dollars.

29

Crude oil

Gasoline

100

100

80

80

60

60

40

40

20

20

0 1985

1995

0 1985

2005

Soybeans 100

80

80

60

60

40

40

20

20 1995

0 1985

2005

Gold 100

80

80

60

60

40

40

20

20 1995

1995

2005

Silver

100

0 1985

2005

Wheat

100

0 1985

1995

0 1985

2005

1995

2005

Figure 2: Estimated Volatility This ﬁgure shows the estimated volatility processes for the six considered commodities under the SV model (annualized in percent).

30

SV: Down−and−out−Call

SV: Down−and−out−Put

1

1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0 90

92

94

96

98

100

0 90

CL HU GC SI S W SP

92

SVJ: Down−and−out−Call 1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

92

94

96

98

100

0 90

SVCJ: Down−and−out−Call 1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

92

94

96

98

98

100

92

94

96

98

100

SVCJ: Down−and−out−Put

1

0 90

96

SVJ: Down−and−out−Put

1

0 90

94

100

0 90

92

94

96

98

100

Figure 3: Down-and-out Options This ﬁgure displays the value of down-and-out barrier options as a fraction of the corresponding plain vanilla options value. The values on the x-axis correspond to the knock-out barrier. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500.

31

32

HU

GC

SI

S

W

SP

Mean 0.0186 0.0189 0.0204 0.0165 0.0076 0.0008 0.0311 Std. Dev. 2.6975 2.7803 1.0076 1.7621 1.5226 2.1192 1.1920 Skewness -0.6996 -0.3748 0.1052 -1.1182 -0.6291 -0.3507 -1.3842 Kurtosis 18.0155 10.0331 11.5751 17.7078 8.1172 9.1778 32.8066 Min -40.0011 -31.4158 -7.2327 -23.6716 -13.1820 -20.4226 -22.8997 Max 21.2765 23.0015 10.2073 13.4045 7.8667 13.1596 10.9572

CL

This table reports descriptive statistics for the continuous percentage returns used in the study. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Table 1: Descriptive Statistics

33

κ 0.0204 (0.0033) 0.0182 (0.0031) 0.0140 (0.0025) 0.0291 (0.0041) 0.0157 (0.0029) 0.0197 (0.0033) 0.0177 (0.0025)

θ 5.1432 (0.4032) 5.6137 (0.4452) 0.9697 (0.1243) 2.7776 (0.2155) 2.2171 (0.2275) 3.8276 (0.3040) 1.2261 (0.1206)

ρ η ζ CL -0.1266 (0.0512) HU -0.0039 (0.0533) 0.0702 (0.0218) 0.0528 (0.0793) GC 0.3826 (0.0481) SI 0.3138 (0.0459) S 0.3206 (0.0512) 0.0460 (0.0118) 0.1363 (0.0692) W 0.0223 (0.0542) 0.0946 (0.0136) 0.4186 (0.0425) SP -0.5678 (0.0387) -

CL HU GC SI S W SP

µ 0.0575 (0.0230) 0.0333 (0.0265) 0.0161 (0.0082) 0.0330 (0.0155) 0.0335 (0.0149) 0.0060 (0.0206) 0.0334 (0.0101)

σ 0.3676 (0.0247) 0.3356 (0.0238) 0.1358 (0.0085) 0.3082 (0.0185) 0.1917 (0.0121) 0.2740 (0.0202) 0.1597 (0.0089)

This table reports the means and standard deviations (in parentheses) of the posterior distributions for each parameter of the SV model. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Table 2: Parameter Estimates: SV

34

κ 0.0096 (0.0019) 0.0100 (0.0027) 0.0096 (0.0019) 0.0125 (0.0025) 0.0105 (0.0021) 0.0092 (0.0021) 0.0132 (0.0024) µJ -1.4218 (0.7941) -1.1414 (0.7866) -0.0492 (0.1471) -0.4627 (0.1901) -0.8150 (0.2658) -0.7974 (0.6487) -2.5866 (0.8356)

µ CL 0.0654 (0.0221) HU 0.0413 (0.0274) GC 0.0159 (0.0081) SI 0.0512 (0.0163) S 0.0555 (0.0160) W 0.0091 (0.0204) SP 0.0380 (0.0100)

λ CL 0.0249 (0.0047) HU 0.0241 (0.0071) GC 0.0531 (0.0117) SI 0.0879 (0.0163) S 0.0471 (0.0107) W 0.0225 (0.0064) SP 0.0071 (0.0023)

σJ 7.7162 (0.7108) 6.5266 (0.8605) 1.8276 (0.1794) 2.7129 (0.2476) 2.2938 (0.2473) 4.7285 (0.5420) 2.8435 (0.5423)

θ 3.9773 (0.4431) 4.5790 (0.6264) 0.7863 (0.1201) 2.0745 (0.2420) 1.9251 (0.2390) 3.2832 (0.3668) 1.2217 (0.1401) ζ

0.0369 (0.0568) 0.1464 (0.1169) 0.4215 (0.0363) -

0.0909 (0.0197) 0.0360 (0.0177) 0.0944 (0.0116) -

ρ -0.2894 (0.0648) -0.0202 (0.0744) 0.4660 (0.0591) 0.5507 (0.0653) 0.3477 (0.0620) -0.0443 (0.0727) -0.6417 (0.0369) η

σ 0.2106 (0.0159) 0.2182 (0.0222) 0.1001 (0.0059) 0.1649 (0.0175) 0.1439 (0.0105) 0.1682 (0.0165) 0.1381 (0.0094)

This table reports the means and standard deviations (in parentheses) of the posterior distributions for each parameter of the SVJ model. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Table 3: Parameter Estimates: SVJ

35

µJ CL -1.5291 (1.1932) HU -2.0728 (1.7065) GC -0.3294 (0.5197) SI -0.6968 (0.6375) S -0.7981 (0.8480) W -1.6252 (0.9708) SP -4.4478 (0.9411)

CL HU GC SI S W SP

µ 0.0622 (0.0227) 0.0441 (0.0275) 0.0124 (0.0082) 0.0283 (0.0159) 0.0338 (0.0157) 0.0164 (0.0207) 0.0421 (0.0100)

θ 2.4719 (0.4271) 3.0628 (0.6079) 0.4098 (0.0700) 1.2995 (0.2321) 1.0821 (0.1656) 1.7065 (0.2534) 0.7874 (0.0787)

σ 0.2085 (0.0219) 0.2151 (0.0343) 0.0961 (0.0080) 0.2110 (0.0213) 0.1447 (0.0158) 0.1373 (0.0249) 0.1264 (0.0098)

σJ ξ µV 7.9167 (0.7397) -0.0170 (0.4167) 2.3176 (0.9789) 6.7416 (0.9186) 0.1425 (0.2745) 4.0436 (1.3707) 2.7819 (0.3301) 0.5546 (0.6923) 0.5108 (0.1470) 3.9967 (0.4951) -0.0099 (0.3746) 1.5656 (0.6734) 2.8019 (0.3667) -0.0612 (0.2793) 1.7129 (0.6099) 4.9274 (0.5244) 0.0478 (0.2229) 3.0470 (0.7534) 2.2486 (0.5251) 0.0519 (0.2012) 2.7594 (0.7914)

κ 0.0173 (0.0041) 0.0200 (0.0041) 0.0185 (0.0032) 0.0374 (0.0076) 0.0262 (0.0046) 0.0240 (0.0045) 0.0225 (0.0041)

ζ

λ 0.0221 (0.0043) 0.0182 (0.0063) 0.0172 (0.0047) 0.0305 (0.0098) 0.0174 (0.0056) 0.0177 (0.0038) 0.0041 (0.0012)

0.0851 (0.0204) 0.0277 (0.0684) 0.0456 (0.0124) 0.1455 (0.0715) 0.0902 (0.0109) 0.4144 (0.0393) -

η

ρ -0.1623 (0.0623) -0.0348 (0.0721) 0.3391 (0.0553) 0.2583 (0.0594) 0.2960 (0.0657) -0.0146 (0.0791) -0.5831 (0.0395)

This table reports the means and standard deviations (in parentheses) of the posterior distributions for each parameter of the SVCJ model. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Table 4: Parameter Estimates: SVCJ

Table 5: Model Comparison This table reports the DIC scores for the model speciﬁcations examined: SV, SVJ, and SVCJ. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

SV CL 27,934 HU 29,169 GC 16,096 SI 23,683 S 21,569 W 25,395 SP 17,437

SVJ SVCJ 26,756 26,779 28,444 28,534 14,634 14,871 21,458 21,776 20,657 20,992 24,609 24,725 16,969 16,819

36

Table 6: Correlations of Returns and Volatilities This table reports return and volatility correlations. Panel A reports the correlations of returns. Panel B displays the correlations of changes in volatility calculated from the estimated volatility process of the SVJ model. Panel C displays the correlations of changes in volatility calculated from the estimated volatility process of the SVCJ model. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Panel A: Returns CL HU GC SI S W

CL 1.0000 0.4894 0.0539 0.0404 0.1094 0.0969

HU

GC

1.0000 0.0104 0.0060 0.1098 0.0755

1.0000 0.6849 0.0006 0.0030

SI

S

W

1.0000 -0.0106 1.0000 -0.0157 0.3941

1.0000

Panel B: SVJ Volatilities CL HU GC CL 1.0000 HU 0.3084 1.0000 GC -0.0084 0.0410 1.0000 SI -0.0257 0.0090 0.6375 S -0.0574 -0.0157 -0.0028 W 0.0368 0.1163 0.0340

SI

S

W

1.0000 0.0124 1.0000 0.0440 0.1146

1.00001

Panel C: SVCJ Volatilities CL CL 1.0000 HU 0.3051 GC 0.0392 SI 0.0118 S -0.0058 W -0.0003

HU

GC

1.0000 0.0244 0.0274 0.0041 0.0265

1.0000 0.4175 0.0067 0.0081

37

SI

S

W

1.0000 0.0282 1.0000 0.0314 0.1045

1.0000

Table 7: Simultaneous Jump Probabilities This table reports the simultaneous jump probabilities of returns in the lower triangular. The upper triangular reports probabilities assuming independence. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Panel A: SVJ CL CL HU 0.40% GC 0.17% SI 0.17% S 0.08% W 0.08%

HU 0.06% 0.14% 0.25% 0.13% 0.05%

GC 0.12% 0.12% 2.23% 0.32% 0.22%

SI 0.21% 0.21% 0.45% 0.62% 0.29%

S 0.12% 0.11% 0.24% 0.43% 0.46%

W 0.06% 0.05% 0.11% 0.20% 0.11% -

S 0.03% 0.03% 0.03% 0.06% 0.21%

W 0.04% 0.03% 0.03% 0.06% 0.03% -

Panel B: SVCJ CL CL HU 0.35% GC 0.08% SI 0.10% S 0.03% W 0.03%

HU 0.04% 0.05% 0.06% 0.06% 0.05%

GC 0.04% 0.03% 0.65% 0.06% 0.02%

38

SI 0.08% 0.06% 0.06% 0.08% 0.06%

Table 8: Correlation of Equity and Commodities This table reports the correlation of equity (S&P 500) and commodity market returns and the correlations of volatility. The ﬁrst row presents the correlations between the returns of the S&P and the commodity; the second and third rows present the correlations of the ﬁrst diﬀferences in volatility between the S&P and the commodity for the SVJ and SVCJ models respectively. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

Returns Volatility: SVJ Volatility: SVCJ

CL 0.0298 0.0398 0.0289

HU 0.0459 0.0433 0.0592

GC -0.0418 0.0470 0.0403

39

SI S -0.0630 0.0792 0.0429 -0.0151 0.0369 0.0299

W 0.0700 0.0418 0.0358

Table 9: Simultaneous Jump Probabilities for Equity and Commodities In parentheses we report the simultaneous jump probabilities assuming independence. Upper part: SVJ, lower part SVCJ. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

SVJ SP

CL 0.10% (0.02%)

HU 0.02% (0.02%)

GC 0.13% (0.04%)

SI 0.10% (0.07%)

S 0.05% (0.04%)

W 0.03% (0.02%)

SVCJ SP

CL 0.03% (0.01%)

HU 0.03% (0.01%)

GC 0.05% (0.01%)

SI 0.06% (0.02%)

S 0.06% (0.01%)

W 0.03% (0.01%)

40

Table 10: Conditional Jump Sizes in Equity and Commodity Markets ’xx’ stands for the respective commodity. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500. The sample period is January 2, 1985 to March 31, 2010.

SVJ CL SP/xx -3.60 xx/SP 3.87

HU GC SI S W -4.43 -4.95 -5.70 -6.51 -8.32 14.84 0.34 -1.11 -2.03 -2.82

SVCJ CL SP/xx -4.39 xx/SP 9.11

HU -5.41 7.01

GC SI S W -5.45 -4.87 -5.05 -5.06 0.02 -0.54 -2.47 -5.07

41

Table 11: Value-at-Risk This table reports the Value-at-Risk (V aRα ) estimates for a horizon of one year and a conﬁdence level of α based on a Monte Carlo simulation with 100,000 replications. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500.

Panel A: SV V aR5% V aR1% 36.96 53.53 41.27 56.60 17.66 25.28 27.89 39.02 24.41 35.03 38.72 51.61 20.71 34.19

V aR0.1% 69.31 70.09 33.51 49.91 46.37 64.33 48.80

Panel B: SVJ V aR10% V aR5% V aR1% CL 32.79 42.77 58.42 HU 35.23 43.75 58.26 GC 13.98 18.00 25.12 SI 25.50 31.32 41.16 S 21.68 27.45 37.41 W 33.01 40.27 53.11 SP 17.03 24.27 37.83

V aR0.1% 73.08 71.01 32.75 50.64 47.59 64.59 51.66

Panel C: SVCJ V aR10% V aR5% V aR1% CL 35.03 44.50 60.22 HU 39.52 48.84 63.71 GC 15.44 19.98 28.54 SI 28.12 34.97 46.26 S 23.42 30.07 42.02 W 35.62 43.37 56.73 SP 15.28 22.43 36.10

V aR0.1% 73.94 76.18 37.98 57.14 54.34 69.09 50.85

V aR10% CL 26.78 HU 32.03 GC 13.53 SI 21.45 S 18.44 W 30.99 SP 13.60

42

Table 12: Option Prices This table reports the option prices for a horizon of one month. The price of the underlying is normalized to 100. The strike price is set to 95 (ITM), 100 (ATM), and 105 (OTM). The left part (Absolute) reports the options price, the right part (Relative) reports the options price relative to the ATM option price. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500.

ITM CL 7.01 HU 7.14 GC 5.23 SI 6.01 S 5.80 W 6.51 SP 5.49

ITM CL 7.15 HU 7.17 GC 5.24 SI 6.02 S 5.81 W 6.51 SP 5.54

ITM CL 6.65 HU 6.79 GC 5.12 SI 5.82 S 5.53 W 6.07 SP 5.38

Panel A: SV Absolute Relative ATM OTM ITM ATM OTM 4.07 2.11 172% 100% 52% 4.26 2.32 168% 100% 55% 1.77 0.38 295% 100% 22% 3.00 1.31 200% 100% 44% 2.69 1.03 215% 100% 38% 3.53 1.67 185% 100% 47% 1.99 0.35 276% 100% 17% Panel B: SVJ Absolute Relative ATM OTM ITM ATM OTM 4.20 2.21 170% 100% 53% 4.28 2.33 167% 100% 54% 1.77 0.37 295% 100% 21% 2.98 1.25 202% 100% 42% 2.68 0.99 217% 100% 37% 3.51 1.64 185% 100% 47% 2.06 0.38 269% 100% 18% Panel C: SVCJ Absolute Relative ATM OTM ITM ATM OTM 3.56 1.66 187% 100% 47% 3.80 1.89 179% 100% 50% 1.38 0.20 372% 100% 14% 2.60 0.94 223% 100% 36% 2.19 0.63 253% 100% 29% 2.87 1.09 211% 100% 38% 1.71 0.19 314% 100% 11%

43

44

HU -0.4864 0.8618 -1.3783 4.1403

CL HU Mean -0.4037 -0.4062 Std 0.8425 0.7275 Skew -4.7612 -3.4210 Kurt 38.8046 24.8022

CL HU Mean -0.4896 -0.4745 Std 0.8772 0.7944 Skew -3.4585 -2.5196 Kurt 22.9281 12.2694

Mean Std Skew Kurt

CL -0.4441 0.8305 -1.0725 2.5727

Panel A: SV GC SI -0.1964 -0.3395 0.3510 0.6464 -1.4260 -1.3480 4.2224 4.1102 Panel B: SVJ GC SI -0.1967 -0.3341 0.3237 0.5446 -2.1526 -2.4347 8.5793 10.5993 Panel C: SVCJ GC SI -0.1444 -0.2634 0.2879 0.4967 -3.5175 -2.9353 23.9426 16.6159 S -0.2192 0.3785 -2.1156 8.5904

S -0.3028 0.4823 -1.8658 5.7959

S -0.3018 0.5272 -1.4964 4.4456 SP -0.2309 0.3659 -2.0724 8.3335

SP -0.2319 0.3781 -1.4691 4.1992

W SP -0.2997 -0.1873 0.5319 0.3034 -3.7515 -2.4437 25.8874 15.2599

W -0.3990 0.6497 -2.5869 13.1202

W -0.3955 0.7072 -1.4902 4.6943

This table reports the mean, standard deviation (std), skewness (skew), and kurtosis (kurt) of the hedging error when delta-hedging an ATM option using Black’s formula. The option has a maturity of one month, and the hedging horizon is one week. CL stands for crude oil, HU for gasoline, GC for gold, SI for silver, S for soybeans, W for Wheat, and SP for the S&P 500.

Table 13: Hedging Error