You are on page 1of 21

JUMPING HEDGES: AN EXAMINATION OF MOVEMENTS IN COPPER SPOT AND FUTURES MARKETS

WING H. CHAN* DENISE YOUNG

Price risk is an important factor for both copper purchasers, who use the commodity as a major input in their production process, and copper reners, who must deal with cash-flow volatility. Information from NYMEX cash and futures prices is used to examine optimal hedging behavior for agents in copper markets. A bivariate GARCH-jump model with autoregressive jump intensity is proposed to capture the features of the joint distribution of cash and futures returns over two subperiods with different dominant pricing regimes. It is found that during the earlier producerpricing regime this specification is not needed, whereas for the later exchange pricing era jump dynamics stemming from a common jump across cash and futures series are signicant in explaining the dynamics in

The authors thank the Editor and an anonymous referee for helpful comments and suggestions. The rst author thanks the Social Science and Humanities Research Council of Canada for nancial support. *Correspondence author, School of Business and Economics, Wilfrid Laurier University, Waterloo, Ontario N2L 3C5, Canada; e-mail: wchan@wlu.ca Received September 2004; Accepted April 2005

Wing H. Chan is an Assistant Professor at the School of Business and Economics at Wilfrid Laurier University in Waterloo, Ontario, Canada. Denise Young is an Associate Professor with the Department of Economics at the University of Alberta in Edmonton, Alberta, Canada.

The Journal of Futures Markets, Vol. 26, No. 2, 169188 (2006) 2006 Wiley Periodicals, Inc. Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/fut.20190

170

Chan and Young

both daily and weekly data sets. Results from the model are used to undertake both within-sample and out-of-sample hedging exercises. These results indicate that there are important gains to be made from a timevarying optimal hedging strategy that incorporates the information from the common jump dynamics. 2006 Wiley Periodicals, Inc. Jrl Fut Mark 26:169188, 2006

INTRODUCTION The hedging of risk is an important factor in the day-to-day operations of copper producers and consumers. This article extends the applied literature on optimal hedging decisions, within the context of copper markets, by examining the joint behavior of cash and futures returns in a bivariate GARCH model that incorporates jump dynamics. In the process, the general literature on copper price behavior is also extended. The context of this application is a market in which sellers of rened copper generally contract with mills (brass and wire rod) and other industrial customers (Edelstein, 2001). Consumers of refined copper generally operate on narrow margins, therefore are sensitive to price risk (Figuerola-Ferretti & Gilbert, 2001). Producers, concerned with cashflow variability, also have an incentive to consider risk-management strategies (Rockman & Kennedy, 2003). The price of rened copper has historically uctuated dramatically over time in response to a variety of supply and demand factors. Price uctuations have tended to be greater in recent decades, where most transactions take place on exchange markets, compared to earlier periods, where a producer-pricing system dominated North American markets. In order to shield themselves from these uctuations, fabricators, manufacturers, and reners have the option of hedging by contracting in futures markets. Within the literature that looks at the modeling of cash and futures prices in commodity markets, the use of GARCH models to capture volatility clustering in metal prices has become common in recent years. One strand of the literature, such as Agirkay, Booth, Hatem and Mustafa (1991), Bracker and Smith (1999), Smith and Bracker (2003), focuses on univariate modeling of spot or futures metal prices. Other studies, including those of Baillie and Myers (1991), Brunetti and Gilbert (2000), and Moschini and Myers (2002), look at the joint behavior of daily cash and futures prices. The latter strand of the literature has highlighted the time variation in both the behavior of prices over time and the optimal strategies that can be employed to hedge against price risk. Although these studies have often captured much of the observed behavior in the prices of various metals, models allowing for jumps,
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

171

which have been successfully applied to the analysis of foreign-exchange and stock-market returns (Chan, 2003; Chan, 2004; Chan & Maheu, 2002; Maheu & McCurdy, 2004), may improve on the performance of current models of resource price behavior. Jumps, specified to parsimoniously model unusual news events as part of the latent news process,1 have the potential to add a layer of richness through the ability of these models to capture both smooth and sudden movements in price volatility. The jump intensity is allowed to change over time following an approximate autoregressive moving average (ARMA) process. The autoregressive jump intensity (ARJI) parametrization provides a channel for the likelihood of future jumps to depend on the history of the jump dynamics. In the present application, it is found that there is a significant common jump component in copper cash and futures returns and these dynamics play an important role in the determination of optimal hedging behavior for agents in copper markets.2 The structure of the article is as follows. First some background information regarding copper markets and the characteristics of copper prices is given. A GARCH-jump model of the joint distribution of copper spot and futures returns and a model of optimal hedging behavior are presented. The data and results showing the important role of jump dynamics in copper markets are discussed. Within-sample and out-of-sample hedging exercises are described, with a comparison of time-varying hedging strategies implied by the GARCH-jump model to constant-hedging and no-hedging strategies. A conclusion ends the article. COPPER MARKETS AND PRICES For many firms (brass and wire-rod mills, foundries, chemical plants, etc.), refined copper is a primary input into the production process. These firms face several uncertainties, including those related to the price of their primary input, rened copper, which can uctuate signicantly over time. Similarly, producers of rened copper face volatility in their cash ow as prices uctuate. Previous to the 1980s, copper markets were characterized by the coexistence of producer and exchange prices for copper. By the end of the late 1970s, producer pricing came to an end in copper markets (Slade, 1988).
1

Andersen (1996) develops a microstructure model to link the information flow to large price movements. 2 Other studies have also considered jumps in the context of optimal hedges (Chang & Chang, 2003; Chang, Chang, & Fang, 1996).
Journal of Futures Markets DOI: 10.1002/fut

172

Chan and Young

It has been argued that one of the potential benets of the producer pricing system, in effect in North America until the late 1970s, was the price smoothing undertaken by producers with market power (FiguerolaFerretti & Gilbert, 2001). In the producer-pricing era, customers generally entered into long-term contracts with North American producers. The producer prices, although not guaranteed, did tend to remain stable for long periods of time. And in times of tight supply, producer prices tended to be much more stable than exchange prices, with producers often rationing supplies instead of raising prices (Edelstein, 2001; Slade, 1991b). As a result, there was comparatively little incentive to use future contracts to hedge price risk. The move away from producer pricing on the part of the major North American producers changed these incentives. At the very least, the move to exchange pricing on the part of North American producers changed the environment in which agents operate and led them to re-evaluate their strategies regarding price risk. Slade (1991a) and Figuerola and Gilbert (2001) look at spot metal prices before and after the producer-pricing era. Although, in general, there are signicant differences between the two eras, the results are sensitive to which metals are included and the time spans considered in the postproducer-pricing era. Another study by Chang, Chen, and Chen (1990) looks at the properties of monthly returns associated with futures contracts. Dividing their sample into pre- and post-producer pricing eras for copper, they nd that the riskiness of futures contracts differs across the two eras, with evidence of statistically signicant systematic risk appearing only in the latter period. Many recent studies examine daily copper prices. Bracker and Smith (1999) and Smith and Bracker (2003) focus on daily futures price returns and note that GARCH or EGARCH models perform best in terms of capturing the volatility dynamics. Furthermore, the variance of the returns is not constant, differing across subsamples. In terms of optimal hedging decisions, it is the joint behavior of the spot and futures prices that matters. Although there have been studies of the joint time-series behavior of several commodities (see, for example, Baillie & Myers, 1991; Brunetti & Gilbert, 2000; Moschini & Myers, 2002), little work has been done in terms of copper markets. Varela (1999) looks at cross-sectional relationships between realized cash prices and futures prices for copper, but does not consider their joint time series properties. Because copper futures markets have often been characterized by backwardation, as opposed to contango (see NYMEX,

Journal of Futures Markets

DOI: 10.1002/fut

Jumping Hedges

173

2001), it may well be the case that modeling joint futures and spot price behavior and the corresponding appropriate hedging strategies for copper may lead to different results from those obtained for other commodities. In the following section, a bivariate GARCH-jump model for copper markets is presented. THE ARJI-GARCH MODEL FOR COPPER GARCH models of joint distributions for spot and futures prices in resource markets have been examined, for example, by Baillie and Myers (1991). Brunetti and Gilbert (2000) allow for volatilities with common origins and volatility spillovers from one market to another within a (fractionally integrated) FIGARCH specication. Moschini and Myers (2002) add further dynamics to the GARCH specication through the introduction of time-varying parameters as a modication of the commonly used BEKK specication of Engle and Kroner (1995). Although none of these studies look at copper per se, they provide frameworks for capturing the joint behavior of related prices and their volatilities. An alternative approach, which allows the researcher additional flexibility in capturing changing volatilities over time for interrelated assets is the class of GARCH models incorporating jump dynamics, as in Chan (2003, 2004). In these models, interrelationships between series are driven by two distinct sources: normal random shocks and systematic jumps. These GARCH-jump models allow for a mixture of smooth volatility movement with abrupt changes or jumps, where these jumps can arrive at varying frequencies and can either be specic to one asset or common across both. In some applications, such as an examination of joint distributions of stock returns, it is expected that some rm-specic news might affect one firm and not others. In this particular application, looking at spot and futures price returns for the same physical commodity, it is assumed that jumps are correlated and impact both series simultaneously. The parametrization of the model allows the frequency of events that create these jumps to vary across time, resulting in a nonconstant arrival rate. A further advantage of GARCH-jump models, in comparison to standard multivariate GARCH setups, is their ability to account for unconditional leptokurtosis in the underlying data in their parametrization. The joint distribution of copper cash and futures price returns (Rc and Rf) is modeled as having an autoregressive conditional jump

Journal of Futures Markets

DOI: 10.1002/fut

174

Chan and Young

intensity GARCH or ARJI-GARCH structure Rct mc fc Rct1 Pct Jt Rft mf ff Rft1 Pft Jt (1) (2)

Each return fluctuates around an asset-specific mean (m). The assetspecic random error terms Pct and Pft are assumed to be bivariate normal random disturbances, distributed independently of the common jump component, Jt. These asset-specic random errors are assumed to have a zero mean and variancecovariance matrix Ht allowing for GARCH effects. For Ht, the bivariate BEKK specification of Engle and Kroner (1995) is used: ~ P ~ Ht C C A P t1 t1 A B Ht1B (3)

where C is an upper triangular parameter matrix, A and B are parameter ~ is dened as the sum of the normal matrices with no restrictions. P t1 disturbance and jump component. The jump component Jt, which provides a parametrized structure for capturing excess kurtosis in the returns, is formulated as Jt a Ytj E a a Ytj b
j0 j0 nt nt

(4)

The expected value is subtracted from the jump component, entering into the mean equation with zero mean. The jump counter nt follows a Poisson distribution with parameter lt as the average number of jumps on day t: exp(lt ) ljt P (nt j 0 t1 ) , j! j 0, 1, 2, p (5)

and t1 is the information set up to t 1. The time-varying jump sizes are normally distributed as Yt N (ut, d2 t ). When specied in this manner, nt jumps arrive between times t 1 and t. This arrival rate is governed by the jump-intensity parameter lt, which is allowed to vary across time according to an approximate ARMA process: lt g0 g1lt1 g2jt1 (6)

To ensure that the jump intensities lt are positive at all times, a sufcient condition is g0 0, g1 g2, and g2 0. The ex-post error, jt1, captures
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

175

the unexpected number of jumps in the previous period. That is, jt1 is the innovation or unpredictable component of the counting process, measured ex-post as jt 1 E (nt1 | t1 ) lt1 a jP (nt i j 0 t i ) lti
j0

(7)

Because of the observed returns Rt and Bayes rule, the probability of the occurrence of j jumps at time t can be inferred ex post, with the lter dened as P (nt j 0 t ) f (Rt 0 nt j, t1 ) P (nt j 0 t1 ) , P (Rt 0 t1 ) j 0, 1, 2, p (8)

where f (Rt 0 nt j, t1 ) is the conditional density function assuming normality. The normal density can then be used to construct P (Rt 0 t1 ) by iterating over all possible numbers of jumps. Further details can be found in Chan and Maheu (2002). The jump size is captured through Ytj, whose parameters dene the jump-size distribution. The mean of the distribution is allowed to vary asymmetrically over time as a function of the size and sign of recent returns in both spot and futures markets: ut u0 u1Rct1(1 I(Rct1)) u2Rct1I(Rct1) u3Rft1(1 I(Rft1)) u4Rft1I(Rft1) (9)

where I(x) 1 if x 0 and 0 otherwise. The variance of the jump size distribution is allowed to be a function of the GARCH variances for cash and futures returns: d2 t d0 d1hct d2hft (10)

where all parameters are restricted to be positive so that the variance covariance matrix of the joint distribution will be well dened. Overall, with this ARJI specication, several aspects of jump behavior are parametrized. A discrete counting variable governs the arrival rate of jumps, or the jump intensity. As a result of the ARMA structure on lt, this discrete random counting variable has a time-varying conditional mean and variance of lt, and unconditional mean g0(1 g1). The model also incorporates dynamics that govern the conditional mean and variance of the size of individual jumps. The conditional mean of the jump
Journal of Futures Markets DOI: 10.1002/fut

176

Chan and Young

size is allowed to react asymmetrically to the size and sign of past returns for both assets. The variance, modeled as a function of GARCH volatility, is allowed to be sensitive to contemporaneous market volatility. Given the jump component, the joint density of the returns is well defined and can be used to estimate the parameters with the use of maximum-likelihood methods. Inferences on jumps can be easily made with the filter defined in Equation (8). A truncation point has to be selected as the maximum number of jumps for the density function. For this application, it is chosen to be sufciently large so that the likelihood function and parameter estimates stabilize to a set of converged values. Once the joint distribution of cash and futures returns has been adequately specied, this information can be used to examine optimal hedge ratios over time. Here, the hedging ratio is dened as the value of futures divided by the value of the cash position. For an agent who hedges a ratio rt1 in the futures market, the expected return and variance on this portfolio position (Rpt) will be E (Rpt ) E (Rct ) rt1E (Rft ) Var (Rpt ) Var (Rct ) r2 t1Var (Rft ) 2rt1Cov (Rct, Rft ) The variance-minimizing hedge is calculated as r* t1 Cov (Rct, Rft ) Var (Rft ) (13) (11) (12)

where both terms can be retrieved from the joint distribution of (Pct Jt) and (Pft Jt), which will be a function of the parameters in Ht, lt, ut, and d2 t . The covariance between spot and futures prices is driven by not only the comovement between the normal disturbances represented by the off-diagonal element in Ht, but also the variance of jump size d2 t . For example, given k common jumps in both returns and hcft as the covariance between the normal disturbances, the covariance between spot and futures returns is dened as: Cov (Rct, Rft ) hcft kd2 t (14)

In the case of zero expected returns on holding futures, this will also generally be the expected utility-maximizing hedge (Baillie & Myers, 1989). The effect of jumps on the optimal hedging ratio depends on the relative size of the covariance versus the variance of futures returns. For most cases, as the absolute size of covariance is smaller than the futures variance, it will be the case that |r*| 1, and common jumps will increase the size of the optimal hedge ratio in absolute terms. In other
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

177

words, it is more costly to hedge with the presence of common jumps in this scenario because the spot and futures prices are moving more closely together. Jumps are directly linked to the cost of risk management through the optimal hedging ratio. The results of the ARJI-GARCH model estimation from the copper data set and the corresponding hedging exercises are discussed in the following sections. DATA The data consist of daily observations on NYMEX cash and futures prices for copper from January 2, 1975 to July 18, 2003, inclusively. As in Moschini and Myers (2002), the futures price is defined by rolling over expiration. These data have also been used to construct weekly observations by aggregation. Returns are calculated as 100 times the rst difference of the log prices. The data from 19751999, inclusive, are used for model estimation. The remaining observations are reserved for out-of-sample exercises. Table I presents summary statistics for the observations used for estimation purposes. There are 1253 daily and 261 weekly observations for the pre-1980 sample (corresponding to an era where many agents operated within a producer-pricing system), and 5917 daily and 1228
TABLE I

Summary Statistics
Pre-1980 Cash Daily
Mean 0.0584 SD 1.5935 Min 8.6074 Max 10.8802 Skewness 0.1717 Kurtosis 5.1868 Q(5) Q 2(5) N 5.9931 [0.3068] 15.2626 [0.0092] 1253

Post-1980 Future Cash Daily


0.0054 1.5351 12.8712 10.5758 0.7065 7.9443 8.1364 [0.1488] 64.1367 [0.0000] 5917

Future Daily
0.0049 1.6187 13.0923 8.4022 0.3821 4.4907 3.5196 [0.6204] 109.9124 [0.0000] 5917

Weekly
0.2808 3.4029 10.0643 17.0345 0.0093 1.9896 5.3208 [0.3779] 3.7859 [0.5806] 261

Daily
0.0545 1.5446 5.6068 6.4219 0.0838 1.1246 4.2154 [0.5188] 53.4057 [0.0000] 1253

Weekly
0.2544 3.2565 8.0300 16.7479 0.2628 2.1892 3.2230 [0.6656] 3.4975 [0.6237] 261

Weekly
0.0264 3.3452 15.6094 18.1657 0.0307 2.0934 3.4143 [0.6365] 38.0936 [0.0000] 1228

Weekly
0.0230 3.4647 20.5622 17.0445 0.1021 2.7438 5.3898 [0.3701] 41.3182 [0.0000] 1228

Note. P values are in squared brackets. Q (i ) is the Ljung box statistics for serial correlations in lag i robust to heteroskedasticity applied to the return series. Q 2 is the same test applied to the squared returns. N is the number of observations.

Journal of Futures Markets

DOI: 10.1002/fut

178

Chan and Young

weekly observations for the post-1980 sample (where agents are operating exclusively under an exchange pricing system). The pre- and post-1980 samples yield quite different summary statistics, with the exception of similar standard deviations. The strong negative skewness in both spot and futures returns lends support to the consideration of a common jump structure under the exchange pricing system. Note that there is evidence of excess kurtosis in all series; however, values are much larger for the exchange pricing period with the values of 7.9443 and 2.0934 for the daily and weekly cash returns compared to the values of 5.1868 and 1.9896 from the producer pricing period. Both squared return series in the post-1980 period are autocorrelated, whereas there is no evidence of autocorrelation in the return series in either period. These general features of the data suggest that a GARCH model with added jump features to capture the leptokurtosis provides an appropriate modeling framework for the return series for the later exchange pricing period. ARJI-GARCH MODEL RESULTS To confirm that dismantling the producer pricing system has led to riskier copper prices and therefore the common jump component is benecial in copper hedging, the models for both the producer pricing system (pre-1980) and the exchange pricing system (post-1980) have been estimated for a comparative analysis. The main features of the estimated ARJI-GARCH models are presented in Table II. During the producer pricing period, the jump structures are not supported by either the daily or weekly data. For the daily data, the average number of jumps are essentially zero with the constant parameter g0 in the jump intensity structure being statistically insignicant. For the weekly data, even though the jump intensity parameters are signicant, they give an average number of jumps g0(1 g1) with a very small value of 0.0067 representing almost zero jumps a day. In addition, all jump sizes parameters ui and di are not signicantly different from zero. These results suggest that the time-varying common jump structure is not a suitable characterization of spot and futures copper returns from the producer pricing period.3
3

Testing for jumps is complicated by the lack of identication of the nuisance parameters ui and di under the null hypothesis. Drost, Nijman, and Werker (1998) propose a kurtosis-based test and Saphores, Khalaf, and Pelletier (2002) develop a Monte Carlo simulation-based test for the presence of jumps. However, the power of these tests are not clear in a bivariate setting. Therefore, the significance of jump structure parameters as an alternative to testing the presence of jumps is examined.
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

179

TABLE II

Estimates of Bivariate Jump Models


Pre-1980 Daily
Mean mc mf fc ff Jump intensities g0 g1 g2 Jump sizes u0 u1 u2 u3 u4 d0 d1 d2
2 QC 2 QF QC,F WC (4) WF (4) WC,F (4) ln L

Post-1980 Weekly Daily


0.0058 (0.0121) 0.0092 (0.0181) 0.0615 (0.0120) 0.1941 (0.0150)

Weekly
0.0893 (0.1232) 0.1563 (0.1274) 0.1883 (0.0260) 0.2045 (0.0265)

0.0846 (0.1136) 0.0762 (0.1188) 0.0952 (0.0260) 0.3599 (0.0306)

1.1712 1.1565 0.0166 0.1683

(0.6953) (0.6904) (0.0452) (0.0478)

0.7855 (1.3016) 0.4480 (0.7511) 0.1893 (0.0926)

0.0045 (0.0007) 0.3344 (0.0786) 0.0580 (0.0482) 57.3570 3.5425 0.0000 0.9998 0.8227 0.0000 0.0000 1.4456 1.2087 0.8718 2.1446 11.1867 16.6328 5.3754 1180.12

0.1379 (0.0292) 0.3112 (0.0266) 0.2813 (0.0433) 0.0296 (0.0373) 0.0094 (0.0218) 0.4930 (0.4216) 0.1336 (0.0129) 0.2776 (0.0269) 0.8249 (0.0749) 0.1968 (0.3422) 0.6707 (0.1704) 2.2809 2.3245 2.4709 32.4763 12.5222 54.1886 16587.13 [0.1309] [0.1273] [0.1159] [0.0000] [0.0138] [0.0000]

0.2798 (0.1179) 0.3278 (0.0395) 0.9457 (0.1975)

0.0355 (0.0725) 0.0494 (0.0427) 0.5173 (0.2496) 0.2323 (0.0393) 0.3486 (0.0633) 0.6365 (0.0824) 0.3699 (0.1509) 0.0000 (0.0024) 2.3511 [0.1251] 8.5710 [0.0034] 7.3155 [0.0068] 4.0620 [0.3977] 8.3861 [0.0784] 47.9221 [0.0000] 4133.56

(38.4527) (3.1935) (0.0001) (0.7805) (2.6830) (10.3668) (0.0226) (1.0176) [0.2715] [0.3504] [0.1430] [0.0254] [0.0022] [0.2508]

0.1912 (0.1670) 0.1177 (0.0471) 0.5289 (0.2542) 0.0473 (0.0235) 0.2094 (0.0673) 0.0000 (0.5052) 0.0000 (0.0131) 0.1712 (0.0361) 1.8038 0.8870 1.3685 1.9879 5.4287 0.3582 4702.51 [0.1792] [0.3462] [0.2420] [0.7379] [0.2451] [0.0792]

Note. Standard errors are in parentheses and p values are in squared brackets. Q are Ljung box statistics for serial correlations of the rst lag robust to heteroskedasticity applied to the cross product of the standardized residuals. Q 2 is the same test for squared residuals. W is the Wooldridges (1991) regression-based tests for model specications. Subscripts C and F refer to cash and futures returns, respectively. The BEKK-GARCH parameters are not reported.

The post-1980 period, which will provide the focus for the remainder of the discussion is now discussed. It is found that the common jump structure is appropriate for modeling joint behavior of cash and futures returns from the exchange market era. The GARCH parameters in the BEKK specication are not reported for the sake of conserving space. In general, there are no signicant changes in the estimated GARCH structure once jumps are incorporated. Volatility clustering with high persistence is revealed in both data series with or without the jump component. In other words, the common jump structure extracts additional information from the spot and futures series that is not captured by the GARCH model.
Journal of Futures Markets DOI: 10.1002/fut

180

Chan and Young

There is strong evidence of signicant correlated jump behavior in both the daily and weekly data from the exchange pricing period. The estimated unconditional mean of the counter for the arrival rate of jumps is 0.2 for the daily data, and a little over double that at 0.42 for the weekly data. The conditional jump intensity (lt) is persistent, with the estimates of g1 being quite similar at approximately 0.3 for the two data sets. The results in terms of the effects of the past shock, as measured via g2, is different across the two data sets, with a much smaller coefcient for the daily data. Although g2 is greater than g1 for the weekly data, violating one of the sufficient conditions, the conditional jump intensity is positive at all points. Overall, the conditional jump intensity reveals that the likelihood of future common jumps depends on the past history of jump dynamics. These results have important implications for the optimal hedging strategy. A rm should take the jump risk into account when constructing its optimal futures hedging ratio. The role of jumps in the model is to capture the effects of unanticipated news announcements on the copper market parsimoniously. Therefore, a higher likelihood of jumps around the time of news announcements is one of the main elements to be considered in the firms hedging decision. The autoregressive jump intensity structure provides an intuitive way of incorporating news effects through the measurable shocks in Equation (7) and supplies the predicted probability of jumps in the future. The effect of these shocks will propagate into the future likelihood of jumps and becomes important for constructing the optimal hedging ratio, as is shown explicitly in Equation (14). The time-varying jump intensities implied by the empirical model are depicted in the top panels of Figures 1 and 2. A constant jump intensity model is obviously not a realistic depiction of copper markets. It can be seen that although the conditional jump intensity is generally less than 1 for the daily data and less than 2 for the weekly data, it can reach occasional highs at approximately 2 and 4 for the respective data sets. It is not unusual to observe on average one common jump a day or two common jumps a week in copper spot and futures returns. As far as mean jump sizes are concerned, the estimates of the u parameters indicate that only futures returns behavior has a signicant impact on mean jump size for both daily and weekly data. These impacts are asymmetric, with changes in the size of positive returns having the bigger impact on mean jump size, as shown by the estimated u3 and u4 having values of 0.1336 and 0.2776 for the daily data and 0.0473 and 0.2094 for the weekly data. From the middle panels of Figures 1 and 2,
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

181

FIGURE 1

Common jumps in daily copper cash and futures returns.

Journal of Futures Markets

DOI: 10.1002/fut

182

Chan and Young

FIGURE 2

Common jumps in weekly copper cash and futures returns.

Journal of Futures Markets

DOI: 10.1002/fut

Jumping Hedges

183

it can be seen that there have been occasional periods of unusually large jumps, with some of the major ones occurring in the late 1980s and late 1990s. The sharp economic recession in the early 1980s depressed world mine production, creating low copper prices and excess supply in the market. As mines scaled down production, this set the stage for a tight supply in the late 1980s (Edelstein, 2001). Together with the added speculative influence from the exchange-based pricing system, this translated to high volatilities in prices. The large jumps in 1980s are mostly likely the consequence of these market conditions. The late 1990s jump behavior would appear to correspond to the copper trading scandal of June 1996 reported in Edelstein (2001) and Smith and Bracker (2003). The industry was shocked by the revelation by Sumitomo Corporationof a $2.6 billion loss on unauthorized copper trades by its head trader. The industry speculated that Sumitomo held a large unreported amount of copper, which in turn led to a sharp price drop, creating jump-like features in the returns data. For the daily and the weekly data, the parameters d2 on the futures GARCH effect hft, in the time-varying jump size variance specication are also signicant, with values of 0.6707 for the daily data and 0.1712 for the weekly data. This suggests the magnitude of jumps depends on the volatility of the futures prices. The bottom panels in Figures 1 and 2 show that it is also during these same periods from the 1980s and 1990s had the largest variances in the jump sizes. These results show that not only is the jump intensity changing over time, but the jump size also varies substantially over the sample period. Autoregressive jump intensity and asymmetric time-varying jump sizes are both essential components in modeling the joint behavior of copper spot and futures returns. The diagnostics for the ARJI-GARCH model are also presented at the bottom of Table II. The Ljung-Box statistics for the squared residuals Q2 i and their cross-products Q C , F , all with p values well above 0.05, indicate no remaining autocorrelation for either the daily or weekly models. Furthermore, the Wooldridge model specication tests Wi, robust to distributional assumptions, do not indicate any specication problems in the variance and covariance processes for the weekly data. Although the ARJI and GARCH components are picking up much of what seems to be occurring with the daily data, the Wooldridge statistics for conditional variances and covariance reject the null of no misspecification errors. These results indicate that there is room for further improvement in terms of modeling daily returns behavior.
Journal of Futures Markets DOI: 10.1002/fut

184

Chan and Young

To summarize a signicant common jump component in daily and weekly cash and futures returns can be found. This indicates that copper cash and futures returns after the producer-pricing era should not be modeled solely as a bivariate GARCH process. For these copper markets, jump dynamics should also be considered. The information from the implied joint distribution is now used to undertake hedging exercises with both within-sample and out-of-sample data. HEDGING EXERCISES The cases of no hedging (rt1 0), a constant hedge obtained as the slope from a linear regression of cash returns on futures returns, and a time-varying risk-minimizing hedge from Equation (13) with the use of the full estimated variancecovariance matrix of the composite (Pct Jt) and (Pft Jt) error terms are examined. Given the insignicance of the mean parameter of the futures returns mf in both the daily and weekly models, the risk-minimizing optimal hedging ratios can also be considered as the expected utility-maximizing hedges. The optimal hedging ratios for the daily and weekly data are displayed in Figure 3. For the daily data, there are periods in the 1990s where the optimal hedge implies the purchase of futures contracts instead of shorting futures. There are no such periods evident in the weekly data. The optimal hedging ratios vary from 0.259 to 0.987 for the daily data and 0.544 to 1.169 for the weekly data. The range of optimal daily hedging ratios is relatively larger, which reflects the higher volatility of daily data. For both data sets, substantial variation in the optimal hedge is evident over time. Under normal market conditions, the spot and futures prices are expected to be positively correlated, and the futures price should reect the cash price, interest rates, insurance, and the cost of storage for copper. To hedge the price risk of holding copper to be sold sometime in the future, a rm should sell some particular number of futures contracts according to the optimal hedging ratio to offset the potential losses from the spot market. The positive relationship between the spot and futures prices may be true for the low frequency (weekly or monthly) data, but occasionally the relationship may turn negative, as is reflected in the covariance in the highly volatile daily data. This may easily be caused by tight supply in the copper market, which leads to futures prices progressively lower than the cash prices the further out is the delivery date of the contract. This so-called backwardation happens frequently in the copper market (NYMEX, 2001). With the presence of backwardation
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

185

FIGURE 3

Optimal hedging ratios for copper.

and negative correlations between spot and futures prices, the firms should purchase futures contracts to hedge the price risk instead of shorting futures as suggested in the standard setup. The negative optimal hedging ratios found in the daily data match with a period of backwardation in copper markets. Table III presents summary statistics regarding the effects of constant and optimal time-varying hedges on portfolio returns and variances. In all cases but one, both for in-sample and out-of-sample data, the mean returns are the highest for the case where the optimal timevarying jump hedge is implemented. In fact, it is only by following a time-varying hedging strategy that overall returns are positive. The daily in-sample returns for the portfolios with no hedging strategy is 0.0055% and with jump hedges it is 0.0002%. Similarly, the weekly
Journal of Futures Markets DOI: 10.1002/fut

186

Chan and Young

TABLE III

Effects of Hedging on Portfolio Returns and Variances


Daily
In-sample returns No hedge Constant hedge Jump hedge In-sample variance reductions Constant hedge Jump hedge Out-of-sample returns No hedge Constant hedge Jump hedge Out-of-sample variance reductions Constant hedge Jump hedge 0.0481 2.1618 1.0003 10.9960 0.0088 0.0044 0.0022 0.0437 0.0158 0.0133 0.1973 1.8732 7.1612 9.7648 0.0055 0.0041 0.0002 0.0303 0.0120 0.0072

Weekly

Note. Variance reductions are defined as the average percentage changes in variance compared to no hedging.

in-sample returns have been improved from 0.0303% to 0.0072% with a common jump component. The same story can be told for the out-ofsample statistics, with weekly returns switching from a negative value of 0.0437% to positive returns of 0.0133%. The benet of adopting the jump hedging strategy is clearly shown by the average percentage reduction in the variance of the portfolio. Within the estimation data sample, jump hedges lower the portfolio variance by 1.87% daily on average. However, for the weekly data, much of the reduction can be attributed to having a constant hedge mechanism, with a percentage reduction in variance of 7.16%, which is not much smaller than the 9.76% obtained from jump hedges. The superiority of the jump hedge over constant hedge is strongly supported by the out-ofsample results. The daily data show a minimal impact due to a constant hedge (0.0481%), whereas jump hedges provide a reduction of over 2% in the variance. Furthermore, the variance reduction from the jump hedge is over 10 times that from the constant hedge with the weekly data. Overall, there are substantial gains from undertaking a timevarying hedging strategy based on the information obtained regarding the structure of the joint ARJI-GARCH distribution of cash and futures returns for copper.
Journal of Futures Markets DOI: 10.1002/fut

Jumping Hedges

187

CONCLUSIONS The focus of this article has been the incorporation of jump dynamics into GARCH models in order to improve understanding of optimal hedge ratios, especially in the more volatile exchange price era. This examination has been undertaken in the context of copper markets, where riskmanagement via hedging strategies can be an important consideration in day-to-day operations. The cash and futures returns data have autocorrelated squared standardized residuals and are leptokurtic, indicating a GARCH model augmented with jump dynamics provides a good candidate for modeling their behavior over time. Given that both returns accrue to the same physical commodity, a jump parametrization that considers common jumps for the two series is chosen. The bivariate ARJI-GARCH models t the data well, leading to an enriched understanding of the dynamics of returns in copper markets. Specication tests indicate that the weekly model is well specied, but there remain uncaptured characteristics in the daily data. The results of the within-sample and out-of-sample hedging exercises point toward an advantage to considering a time-varying hedge with common jumps for agents in todays copper markets. BIBLIOGRAPHY
Agirkay, V., Booth, G., Hatem, J., & Mustafa, C. (1991). Conditional dependence in precious metal prices. Financial Review, 26, 367386. Andersen, T. G. (1996). Return volatility and trading volume: An information ow interpretation of stochastic volatility. Journal of Finance, 51, 169204. Baillie, R. T., & Myers, R. J. (1989). Modeling commodity price distributions and estimating the optimal futures hedge (Working Paper No. 201). New York: Columbia University, Center for the Study of Futures Markets. Baillie, R. T., & Myers, R. J. (1991). Bivariate garch estimation of the optimal commodity futures hedge. Journal of Applied Econometrics, 6, 109124. Bracker, K., & Smith, K. (1999). Detecting and modeling changing volatility in the copper futures market. Journal of Futures Markets, 19, 79100. Brunetti, C., & Gilbert, C. (2000). Bivariate garch and fractional cointegration. Journal of Empirical Finance, 7, 509530. Chan, W. H. (2003). A correlated bivariate poisson jump model for foreign exchange. Empirical Economics, 28, 669689. Chan, W. H. (2004). Conditional correlated jump dynamics in foreign exchange. Economics Letters, 83, 2328. Chan, W. H., & Maheu, J. M. (2002). Conditional jump dynamics in stock market returns. Journal of Business & Economic Statistics, 79, 377389. Chang, C. W., & Chang, J. S. K. (2003). Optimum futures hedge in the presence of clustered supply and demand shocks, stochastic basis, and rms costs of hedging. Journal of Futures Markets, 23, 12091237.
Journal of Futures Markets DOI: 10.1002/fut

188

Chan and Young

Chang, C. W., Chang, J. S. K., & Fang, H. (1996). Optimum futures hedges with jump risk and stochastic basis. Journal of Futures Markets, 16, 441458. Chang, E., Chen, C., & Chen, S. N. (1990). Risk and return in copper, platinum, and silver futures. Journal of Futures Markets, 10, 2939. Drost, F. C. Nijman, T. E., & Werker, B. J. M. (1998). Estimation and testing in models containing both jumps and conditional heteroskedasticity. Journal of Business & Economic Statistics, 16, 237243. Edelstein, D. (2001). Copper. Minerals Yearbook, U.S. Geological Survey. Engle, R. F., & Kroner, K. F. (1995). Multivariate simultaneous generalized arch. Econometric Theory, 11, 122150. Figuerola-Ferretti, I., & Gilbert, C. (2001). Price variability and marketing method in non-ferrous metals: Slades analysis revisited. Resources Policy, 27, 169177. Maheu, J. M., & McCurdy, T. H. (2004). News arrival jump dynamics and volatility components for individual stock returns. Journal of Finance, 59, 755793. Moschini, G. C., & Myers, R. J. (2002). Testing for constant hedge ratios in commodity markets: A multivariate garch approach. Journal of Empirical Finance, 9, 589603. NYMEX. (2001). A guide to metals hedging. New York: New York Mercantile Exchange. Rockman, A., & Kennedy, M. (2003). To hedge or not to hedge: Considering a strategy. In H. Cameron (Ed.), Hedging in the mining industry: Strategy, control, and governance (pp. 512). PriceWaterhouseCoopers. Saphores, J. D., Khalaf, L., & Pelletier, D. (2002). On jumps and arch effects in natural resource prices: An application to stumpage prices from Pacific Northwest national forests. American Journal of Agriculture Economics, 84, 387400. Slade, M. (1988). Pricing of metals. Kingston: Centre for Resource Studies. Slade, M. (1991a). Market structure, marketing method, and price instability. Quarterly Journal of Economics, 106, 13091339. Slade, M. (1991b). Strategic pricing with customer rationing: The case of primary metals. Canadian Journal of Economics, 42, 70100. Smith, K., & Bracker, K. (2003). Forecasting changes in copper futures volatility with garch models using an iterated algorithm. Review of Quantitative Finance and Accounting, 20, 245265. Varela, O. (1999). Futures and realized cash or settle prices for gold, silver, and copper. Review of Financial Economics, 8, 121138. Wooldridge, J. M. (1991). On the application of robust, regression-based diagnostics to models of conditional means and conditional variances. Journal of Econometrics, 47, 546.

Journal of Futures Markets

DOI: 10.1002/fut

You might also like