Understanding Dynamic Conditional Correlations between Oil, Natural Gas and Non-Energy Commodity Futures Markets.

• Author: Behmiri, Niaz Bashiri

1. INTRODUCTION

   The links between energy and non-energy commodity futures markets have deepened in recent times: the liberalization of capital flows, the development in market trading technologies and in new financial instruments, and the improvement in information transmission have all contributed to an increased integration between commodity markets (Ji and Fan, 2012). After the equity market collapse in 2000, commodities became an increasingly popular asset class, eligible for portfolio diversification, thanks to the negative correlation between commodity futures returns and stock returns (Gorton and Rouwenhorst, 2006). (1) Furthermore, the boom of biofuels in the 2000s increased the links between energy and some agricultural markets (FAO, 2008).

   Basak and Pavlova (2016) show theoretically that the correlations between commodity futures, as well as the equity-commodity correlations, increase with the financialization of these commodity markets. In recent years, empirical analysis reports an increase in the correlations between commodities, which might therefore have limited the benefits of a diversification strategy from equity to commodity futures markets (Cheung and Miu, 2010; Daskalaki and Skiadopoulos, 2011). As volatilities spillover across markets, knowledge of these dynamics is crucial for investors and financial institutions in terms of portfolio construction and risk management. Investigating the dynamics of the correlations between these markets is essential to develop accurate asset pricing models and hedging strategies, as well as to minimize the contagion risk in the occurrence of a crash in one of these markets (Barunik et al., 2015).

   Several authors have recently focused on the linkages between commodities futures markets, looking at the time-varying correlations produced through different multivariate GARCH specifications. With respect to the volatilities and correlations between energy and agricultural futures markets, researchers find statistically significant volatility spillovers from oil to agricultural markets, with a change in the dynamics of volatility transmission after the second half of the 2000s. These results are obtained using different methodologies, such as bivariate EGARCH (Ji and Fan, 2012), bivariate stochastic volatility models (Du et al., 2011), and the VAR-BEKK-GARCH and VAR-DCC-GARCH models (Mensi et al., 2014). Some authors focus in particular on the agriculture-ethanol-fossil fuels link, which has had a great deal of attention from the early 2000s. Adopting multivariate GARCH models, they find strong volatility linkages, both in the U.S. and in the emerging markets (Chang and Su, 2010; Serra, 2011; Wu et al., 2011; Du and McPhail, 2012; Trujillo-Barrera et al., 2012; Gardebroek and Hernandez, 2013; Wu and Li, 2013).

   Another area that has attracted much interest in the last decade is the relationship between energy commodities and metals. This relationship is far from simple, with metals being traded for both industrial use and hedging strategies. As for the spillovers between metal and energy markets, significant transmission of volatility between metals and oil prices is found using different methodologies such as the Markov-switching space state model (Choi and Hammoudeh, 2010), univariate and bivariate GARCH (Ewing and Malik, 2013), and the Hidden Markov Decision Tree (Charlot and Marimoutou, 2014).

   These papers do not, however, investigate under which circumstances these correlations change. Indeed, a number of studies investigate the effects of macroeconomic and financial factors on the volatility of commodity futures returns (Hammoudeh and Yuan, 2008; Batten et al., 2010; Sanders and Irwin, 2011; Irwin and Sanders, 2012; Manera et al., 2016). However, with respect to time-varying correlations, the literature so far has investigated the factors affecting the correlations between commodities and stock markets (Silvennoinen and Thorp, 2013; Buyuksahin and Robe, 2014), or those between energy futures returns (Karali and Ramirez, 2014; Bunn et al., 2017).

   Silvennoinen and Thorp (2013) find that the correlations between stocks, bonds and commodity futures returns have increased for most commodities, often when the VIX volatility index was high, thus pointing to strong financial influences. Their results are consistent with the analysis of Cheung and Miu (2010) and Daskalaki and Skiadopoulos (2011). Buyuksahin and Robe (2014) concentrate on the role of financialization in commodity markets on stock-commodity co-movement, and show that the speculative activity of hedge funds that trade actively in both equity and commodity future markets has explanatory power for the correlation between stocks and commodities. The predictive power of the speculative activity is, however, weaker in periods of higher stress in the financial markets. Karali and Ramirez (2014) analyze the time-varying volatility and spillover effects in energy futures markets, finding that macroeconomic variables and political and weather-related events have an effect on the volatilities and their correlations. More recently, Bunn et al. (2017) have provided evidence that the strength of the speculative activity is an important determinant of co-movements between oil and gas returns, even once fundamentals have been taken into account.

   Overall, the empirical evidence on the factors which influence the time-varying correlations between energy and non-energy commodities futures markets is lagging behind. The drivers of these correlation patterns over time are a field still not explored, but important to understanding whether the diversification benefits of commodities to equity market investors have weakened or not. From a public policy perspective, it is relevant to understand whether dynamic conditional correlations between commodities respond to monetary policies.

   We provide fresh evidence to fill this gap in the literature. To the best of our knowledge, this is the first attempt to investigate these correlations within a unique framework: i.e. with a common methodology to obtain the time-varying correlations, looking at the same period of analysis, and considering common explanatory variables, thus allowing a direct comparison of the results found across different sets of commodities.

   First, we estimate a dynamic conditional correlation (DCC) multivariate GARCH (Engle, 2002), which allows for covariance and correlation spillovers. The analysis considers ten commodities at weekly frequency: West Texas Intermediate (WTI) crude oil and natural gas; five agricultural commodities (corn, oats, rice, soybeans, and wheat); and three metals (copper, gold, and silver) over the period spanning from 1998:w1 to 2014:w22. We observe that the correlations between energy and metals futures markets are larger than those between energy and agricultural commodities. The DCCs peaked around the 2008 financial crisis and subsequently decreased.

   We then investigate under which circumstances energy and non-energy commodities futures returns display larger dynamic conditional correlations. We consider macroeconomic fundamentals, financial market characteristics and speculative activity, following the prior literature that suggests these factors might matter for commodity futures returns correlations. We estimate an Autoregressive Distributed Lag (ARDL) model by means of the Pooled Mean Group (PMG) estimator (Pesaran et al., 1999). Our analysis suggests that macroeconomic and financial factors influence the agriculture-energy and metals-energy correlations. Speculative activity in the energy markets is significant in explaining correlations with agricultural commodities, but not those with metals.

   The paper is structured as follows: Section 2 focuses on the first step of our analysis: namely, retrieving the DCCs, by discussing the data, the methodology, and the characteristics of the DCCs we obtain. Section 3 concerns the second step of the econometric analysis: understanding these correlations. It describes the explanatory variables, the econometric specification and presents the results, as well as some robustness checks. Finally, Section 4 concludes.

2. OBTAINING THE DCCS BETWEEN COMMODITIES FUTURES MARKETS

   2.1. Data description: Commodity futures returns

   We develop our analysis on a sample of ten commodities futures belonging to three different industries: energy (West Texas Intermediate (WTI) crude oil and natural gas); agriculture (corn, oats, rice, soybeans, and wheat), and metals (copper, gold, and silver). We collect data at weekly frequency for the period ranging from 1998:w1 to 2014:w22. (2)

   The weekly price series for these commodities are obtained by rolling over their first nearby contracts on the second Thursday of the maturity month. The returns are computed as [r.sub.it] = log([P.sub.it]/[P.sub.it-1]), where [r.sub.it] is the corresponding return, [P.sub.it] is the weekly future real price, obtained by subtracting realized inflation, calculated from the U.S. consumer price index (CPI), base year 2010, from the nominal price; i=1.. .10 defines the commodity futures market; and t is the week. Table 1 reports the descriptive statistics for the commodities futures returns. (3)

   As a first check, we test the stationarity of the commodity futures returns. The augmented Dickey Fuller (1979) unit root test confirms the stationarity of all returns at the 1% significance level. Then, we inspect the residuals obtained from the OLS regression of each series of returns on a constant term. The Lagrange multiplier test suggests the existence of the ARCH effects for all returns at the 1% and 5% levels of significance. The Breusch-Godfrey test for higher-order serial correlation in the disturbance provides evidence of serial correlation for oats and WTI, while no serial correlation is detected for the other commodities.

   2.2 Methods: The multivariate GARCH

   Commodity futures volatilities are known to move together...