Does oil price respond to macroeconomic uncertainty? New evidence

Abstract

Using a new measure of economic policy-related uncertainty (EPU), this study evaluates whether macroeconomic uncertainty affects oil price or vice versa. Specifically, we investigate mean and volatility spillovers between the EPU index and oil returns and the underlying drivers for the time-varying correlation. Our results illustrate the importance of policy uncertainty. Although the mean spillover of EPU on oil returns is negligible in the long run, the mean spillover in the short term and the volatility spillover of EPU are significant for oil spot and futures returns. Moreover, we provide evidence that both oil supply shocks and real economic shocks lead the correlation between the EPU and oil returns to fluctuate over time.

1 Introduction

Sharp and erratic oil price movements are not new and have been a dominant phenomenon during the past two decades. Numerous studies generally cite two reasons for this phenomenon. Some studies claim that sharp movements in fundamental oil supply and demand, particularly the rapidly growing demand in emerging economies such as China, lead to strong oil price fluctuations (Hamilton 2009; Kilian and Vigfusson 2011), whereas others argue that excessive speculation and arbitrage trading activities by index investors play an important role (Lombardi and Van Robays 2011; Tang and Xiong 2012; Singleton 2013).

However, one factor that has been overlooked in this debate is that economic uncertainty is typically very high in periods of strong oil price volatility. The dynamic pattern of these two variables is puzzling but suggestive of important information of transmission. A rich strand of literature has confirmed the significant effects of uncertainty on macroeconomic performance (Grier and Perry 2000; Bhar and Hamori 2004; Fountas et al. 2006), while little attention has been paid to oil markets. Therefore, this study aims to examine the impact of economic uncertainty on oil markets. Specifically, we analyze the information spillover (in terms of mean and volatility spillovers) and dynamic correlation between economic uncertainty and oil price and assess the factors at the origin of changes in the degree of dynamic correlation.

The contributions can be summarized as follows. First, although several studies have investigated the relationship between oil price and macroeconomic variables, such as output, inflation, unemployment and GDP growth (Hooker 1996; Ferderer 1996; Barsky and Kilian 2004; Baumeister et al. 2010; Elder and Serletis 2010; Kilian and Vigfusson 2011; Aguiar-Conraria and Soares 2011; Hamilton 2011; Kim 2012; Bhar and Mallik 2013), little research has analyzed how uncertainty in the global economy affects the oil market. A notable exception is Guo and Kliesen (2005) who find that standard macroeconomic variables do not forecast realized oil price volatility, which suggests that changes in the supply and demand for crude oil that raise the variance of future crude oil prices tend to reflect stochastic disturbances. However, Dvir and Rogoff (2009) argue that the real price of oil historically has tended to be both more persistent and more volatile whenever rapid industrialization in the world economy coincided with uncertainty regarding access to supply. This disagreement may relate to the inappropriate measure of macroeconomic situations. By using a new and appropriate measure of uncertainty, this study evaluates the information transmission and dynamic correlation between uncertainty and oil prices.

Second, unlike most studies that focus on uncertainty with respect to oil price itself or oil supply/demand (Pindyck 2004; Kilian 2009; Kellogg 2014; Bredin et al. 2011; Baumeister and Peersman 2013), we focus on macroeconomic uncertainty. To the best of our knowledge, this study is the first to examine the time-varying correlation and the information spillover between the oil market and uncertainty external to the oil market, namely economic policy-related uncertainty (EPU), a new and appropriate¹ measure proposed by Baker et al. (2013).

Third, we assess the effects of different types of oil shocks that are at the origin of changes in the degree of dynamic correlation between EPU and oil prices. Numerous studies document that the volatility puzzle in the crude oil market can be rationalized by the fact that the price elasticity of both oil supply and oil demand has decreased considerably (Regnier 2007; Kilian 2009; Baumeister et al. 2010; Baumeister and Peersman 2013). Conversely, other studies touch on the relationship between oil price and economy activity (Hooker 1996; Ferderer 1996; Barsky and Kilian 2004; Guo and Kliesen 2005; Elder and Serletis 2010; Baumeister et al. 2010). This study links these two ideas by demonstrating that the degree of dynamic correlation between EPU and oil prices can be explained by both oil price shocks and changes in the real economy.

The remainder of the study is organized as follows. Section 2 discusses several possible ways in which macroeconomic uncertainty could affect oil price dynamics. Section 3 describes the data. Section 4 presents the econometric methodology. Section 5 illustrates empirical results and analysis, and Sect. 6 concludes.

2 How can uncertainty affect the dynamics of oil price?

Some studies help clarify why oil prices fluctuate dramatically in times of greater economic uncertainty.

First and most importantly, both oil demand and supply could be less responsive due to an option value to wait created by economic uncertainty. The elasticity of oil supply could be lower because oil producers prefer to delay changing their production until more information on demand shock is available. This link has been well documented (Litzenberger and Rabinowitz 1995; Guiso and Parigi 1999; Bloom et al. 2007; Bloom 2009; Mohn and Misund 2009; Kellogg 2014).

Second, uncertainty could affect price setting in the oil spot and futures markets. Fundamentalists, who predict the oil price to converge on its fundamental value, have a stabilizing effect on the price of oil, but chartists usually have a destabilizing effect on the price of oil (Reitz and Slopek 2009). Ellen and Zwinkels (2010) and Singleton (2013) argue that heterogeneous beliefs concerning uncertainty in future economic would lead to price drifts and irrational volatility, even booms or busts.

Third, oil futures markets may respond quickly to economic uncertainty and then transfer information to oil spot markets. Alquist and Kilian (2010) demonstrate that increased uncertainty about future oil supply shortfalls under plausible assumptions causes the spread to decline. Increased uncertainty also causes precautionary demand for oil to increase, resulting in an immediate increase in the real spot price. The authors also provide an empirical analysis of this indicator, providing evidence of how shifts in uncertainty about future oil supply shortfalls affect the real spot price of crude oil. Consistent with the view that hedging against oil price movements could weaken the responsiveness of oil demand and supply (Baumeister and Peersman 2013), greater macroeconomic uncertainty would lead to an increased use of futures contracts, therefore decreasing the elasticity of oil demand and supply.

3 Data

To measure uncertainty, we use the monthly economic policy-related uncertainty index by Baker et al. (2013) that has recently gained considerable attention (Johannsen 2014; Johnson and Lee 2014; Jones and Olson 2013; Yin and Han 2014). Figure 1 plots the EPU index. The EPU index spikes near close presidential elections, Gulf Wars I and II, the 9/11 terrorist attacks and other major shocks. Recently, the index rose to historic highs after the Lehman bankruptcy and TARP legislation, the 2010 midterm elections, the Eurozone crisis and the US debt-ceiling dispute.

Fig. 1

Index of economic policy-related uncertainty

The oil spot price is the fob price of West Texas Intermediate crude oil (WTI), and the oil futures price is the closing price of WTI in the New York Mercantile Exchange (NYMEX), which is determined by rolling over the nearby futures contract on the first trading day of the contract’s expiration month. The data are provided by the EIA database. We employ monthly data over the period from January 1986 to March 2013.

The descriptive statistics for the change of uncertainty index, spot returns and futures returns are displayed in Panel A of Table 1. The Ljung–Box Q tests with 6 and 12 lags indicate a strong autocorrelation in each series. Panel B of Table 1 reports the unit root test results. Using augmented Dickey–Fuller tests, ERS tests and Lee and Strazicich (2003) test (break model with two breaks), all the first differences of logarithmic of series are stationary. Panel C of Table 1 reports the ARCH tests results. The ARCH LM test rejects the null hypothesis of homoscedasticity for all three variables, indicating that ARCH-type models are appropriate. The Tse (2000) LM test also rejects the null hypothesis of constant correlation. The results displayed in Panel D of Table 1 reveal that both the trace and maximal eigenvalue tests significantly reject the null hypothesis of none, one or two co-integrating vectors in the system.

Table 1 Descriptive statistics of change of uncertainty index, spot returns and futures returns

4 Methodology

Based on the descriptive statistics of all data series, we chose to model the series as Engle’s (2002) multivariate DCC-GARCH model allowing for time-varying correlations.

4.1 Mean equations

To examine patterns of price transmission across various markets, we use the following vector error-correction model (VECM) to specify conditional mean changes of the uncertainty index and returns of oil spot and futures prices:

$$\begin{aligned} r_{u,t}= & {} \mu _u +\sum _{i=1}^p {\alpha _{u,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{u,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{u,k} r_{f,t-k} } \nonumber \\&+\,\lambda _u \left( {P_{s,t-1} -P_{u,t-1} } \right) +\tau _u \left( {P_{f,t-1} -P_{u,t-1} } \right) +\varepsilon _{u,t} , \end{aligned}$$

(1)

$$\begin{aligned} r_{s,t}= & {} \mu _s +\sum _{i=1}^p {\alpha _{s,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{s,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{s,k} r_{f,t-k} } \nonumber \\&+\,\lambda _s \left( {P_{u,t-1} -P_{s,t-1} } \right) +\tau _s \left( {P_{f,t-1} -P_{s,t-1} } \right) +\varepsilon _{s,t} , \end{aligned}$$

(2)

$$\begin{aligned} r_{f,t}= & {} \mu _f +\sum _{i=1}^p {\alpha _{f,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{f,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{f,k} r_{f,t-k} } \nonumber \\&+\,\lambda _f \left( {P_{u,t-1} -P_{f,t-1} } \right) +\tau _f \left( {P_{s,t-1} -P_{f,t-1} } \right) +\varepsilon _{f,t} , \end{aligned}$$

(3)

where $P_{u,t}$ is the logarithmic level of uncertainty index on time t and $P_{s,t}$ and $P_{f,t}$ are the logarithmic prices of oil spot price and oil futures price, respectively. Equations $r_{u,t} =P_{u,t} -P_{u,t-1}$, $r_{s,t} =P_{s,t} -P_{s,t-1}$, and $r_{f,t} =P_{f,t} -P_{f,t-1}$ correspond to logarithmic changes of the uncertainty index and respective logarithmic returns of oil spot and futures prices. Parameters p, q and r are the lag orders to be determined using the Bayesian information criterion (BIC). The co-integrating vectors, $P_{s,t-1} -P_{u,t-1}$, $P_{f,t-1} -P_{u,t-1}$, $P_{u,t-1} -P_{s,t-1}$, $P_{f,t-1} -P_{s,t-1}$, $P_{u,t-1} -P_{f,t-1}$ and $P_{s,t-1} -P_{f,t-1}$, act as error-correction terms for the long-term relationships between the returns, determined using the Johansen methodology. The error vector $\varepsilon _t =\left( {\varepsilon _{u,t} ,\varepsilon _{s,t} ,\varepsilon _{f,t} } \right) ^{\prime }$ can be written as $\varepsilon _t =H_t^{1/2} \eta _t$, where $\eta _t =\left( {\eta _{u,t} ,\eta _{s,t} ,\eta _{f,t} } \right)$that refers to standardized error is an independently and identically distributed sequence of random vectors with zero mean and covariance matrix I. The matrix $H_t$ represents the Cholesky decomposition of:

$$\begin{aligned} H_t =Var\left( {\varepsilon _{u,t} ,\varepsilon _{s,\mathrm{t}} ,\varepsilon _{f,t} \left| {\Omega _{t-1} } \right. } \right) = \left[ {{\begin{array}{c@{\quad }c@{\quad }c} {h_{u,t} }&{} {h_{us,t} }&{} {h_{uf,t} } \\ {h_{su,t} }&{} {h_{s,t} }&{} {h_{sf,t} } \\ {h_{fu,t} }&{} {h_{fs,t} }&{} {h_{f,t} } \\ \end{array} }} \right] , \end{aligned}$$

where $\Omega _{t-1}$ is the information set at $t-1$. Therefore, $\varepsilon _t \sim N\left( {0,H_t } \right)$.

This approach is widely used in previous studies to describe price interactions among various informationally linked markets because it captures both the short- and long-term effects of information flow across markets. Specifically, short-term effects are reflected by cross-market lagged returns in these equations, and long-term effects are captured by long-run equilibrium errors, defined as the difference between market prices of any two markets in the last period.

4.2 Variance equations

Given that volatility is a source of information, an examination of volatility spillover can help us further understand the information transmission across markets. The time-varying variances of the changes of the uncertainty index and returns of oil spot and futures prices are specified as follows:

$$\begin{aligned} h_{u,t}= & {} \omega _u +\theta _{uu} h_{u,t-1} +\theta _{us} h_{s,t-1} +\theta _{uf} h_{f,t-1} +\delta _{uu} \varepsilon _{u,t-1}^2 +\delta _{us} \varepsilon _{s,t-1}^2 +\delta _{uf} \varepsilon _{f,t-1}^2 , \nonumber \\\end{aligned}$$

(4)

$$\begin{aligned} h_{s,t}= & {} \omega _s +\theta _{su} h_{u,t-1} +\theta _{ss} h_{s,t-1} +\theta _{sf} h_{f,t-1} +\delta _{su} \varepsilon _{u,t-1}^2 +\delta _{ss} \varepsilon _{s,t-1}^2 +\delta _{sf} \varepsilon _{f,t-1}^2 , \nonumber \\\end{aligned}$$

(5)

$$\begin{aligned} h_{f,t}= & {} \omega _f +\theta _{fu} h_{u,t-1} +\theta _{fs} h_{s,t-1} +\theta _{ff} h_{f,t-1} +\delta _{fu} \varepsilon _{u,t-1}^2 +\delta _{fs} \varepsilon _{s,t-1}^2 +\delta _{ff} \varepsilon _{f,t-1}^2 .\nonumber \\ \end{aligned}$$

(6)

The conditional volatility is influenced by past residual shocks from both its own markets and other markets. Consequently, this model can capture both volatility clustering in each market and volatility spillovers across markets.

Unlike the setting in CCC-GARCH, which restricts the correlation coefficients to be constant over time, DCC-GARCH allows them to vary through time:

$$\begin{aligned} \rho _{us,t}= & {} \left( {1-\kappa _1 -\kappa _2 } \right) \bar{{\rho }}_{us} +\kappa _{1,us} \rho _{us,t-1} +\kappa _{2,us} \pi _{us,t-1} , \end{aligned}$$

(7)

$$\begin{aligned} \rho _{uf,t}= & {} \left( {1-\kappa _1 -\kappa _2 } \right) \bar{{\rho }}_{uf} +\kappa _{1,uf} \rho _{uf,t-1} +\kappa _{2,uf} \pi _{uf,t-1} , \end{aligned}$$

(8)

$$\begin{aligned} \rho _{sf,t}= & {} \left( {1-\kappa _1 -\kappa _2 } \right) \bar{{\rho }}_{sf} +\kappa _{1,sf} \rho _{sf,t-1} +\kappa _{2,sf} \pi _{sf,t-1} , \end{aligned}$$

(9)

where $\rho _{us,t}$, $\rho _{uf,t}$ and $\rho _{sf,t}$ are the correlation coefficients at time t, and for $m\ge 2$

$$\begin{aligned} \pi _{us,t-1}= & {} \frac{\sum \nolimits _{l=1}^m {\eta _{u,t-1} \eta _{s,t-1} } }{\sqrt{\left( {\sum \nolimits _{l=1}^m {\eta _{u,t-1} } } \right) \left( {\sum \nolimits _{l=1}^m {\eta _{s,t-1} } } \right) }},\\ \pi _{uf,t-1}= & {} \frac{\sum \nolimits _{l=1}^m {\eta _{u,t-1} \eta _{f,t-1} } }{\sqrt{\left( {\sum \nolimits _{l=1}^m {\eta _{u,t-1} } } \right) \left( {\sum \nolimits _{l=1}^m {\eta _{f,t-1} } } \right) }},\\ \pi _{sf,t-1}= & {} \frac{\sum \nolimits _{l=1}^m {\eta _{s,t-1} \eta _{f,t-1} } }{\sqrt{\left( {\sum \nolimits _{l=1}^m {\eta _{s,t-1} } } \right) \left( {\sum \nolimits _{l=1}^m {\eta _{f,t-1} } } \right) }}. \end{aligned}$$

Moreover, to investigate whether fluctuations in oil prices breed uncertainty, we also include the conditional standard variances of each variable in the mean equations as follows:

$$\begin{aligned} r_{u,t}= & {} \mu _u +\sum _{i=1}^p {\alpha _{u,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{u,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{u,k} r_{f,t-k} } \nonumber \\&+\,\lambda _u \left( {P_{s,t-1} -P_{u,t-1} } \right) +\tau _u \left( {P_{f,t-1} -P_{u,t-1} } \right) +\phi _{u,u} \hat{{\sigma }}_{u,t}^2 \nonumber \\&+\,\phi _{u,s} \hat{{\sigma }}_{s,t}^2 +\phi _{u,f} \hat{{\sigma }}_{f,t}^2 +\varepsilon _{u,t}, \end{aligned}$$

(10)

$$\begin{aligned} r_{s,t}= & {} \mu _s +\sum _{i=1}^p {\alpha _{s,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{s,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{s,k} r_{f,t-k} } \nonumber \\&+\,\lambda _s \left( {P_{u,t-1} -P_{s,t-1} } \right) +\tau _s \left( {P_{f,t-1} -P_{s,t-1} } \right) +\phi _{s,u} \hat{{\sigma }}_{u,t}^2 \nonumber \\&+\,\phi _{s,s} \hat{{\sigma }}_{s,t}^2 +\phi _{s,f} \hat{{\sigma }}_{f,t}^2 +\varepsilon _{s,t}, \end{aligned}$$

(11)

$$\begin{aligned} r_{f,t}= & {} \mu _f +\sum _{i=1}^p {\alpha _{f,i} r_{u,t-i} } +\sum _{j=1}^q {\beta _{f,j} r_{s,t-j} } +\sum _{k=1}^r {\gamma _{f,k} r_{f,t-k} } \nonumber \\&+\,\lambda _f \left( {P_{u,t-1} -P_{f,t-1} } \right) +\tau _f \left( {P_{s,t-1} -P_{f,t-1} } \right) +\phi _{f,u} \hat{{\sigma }}_{u,t}^2 \nonumber \\&+\,\phi _{f,s} \hat{{\sigma }}_{s,t}^2 +\phi _{f,f} \hat{{\sigma }}_{f,t}^2 +\varepsilon _{f,t} . \end{aligned}$$

(12)

where $\hat{{\sigma }}_{u,t}^2$, $\hat{{\sigma }}_{s,t}^2$ and $\hat{{\sigma }}_{f,t}^2$ denote the conditional variance of each variable, respectively, which are generated by the process of GARCH(1,1).² If the hypothesis that fluctuations in oil prices breed uncertainty is correct, then the coefficients on the conditional variances of returns on oil spots and futures ($\phi _{u,s} ,\phi _{u,f} )$ would be positive and statistically significant in the conditional mean equation of uncertainty.

To enhance the accuracy of our estimate, we adopt the BFGS algorithm with no constraints on parameters to maximize the log-likelihood function based on the BIC:

$$\begin{aligned} L\left( \Theta \right) =-\sum _{t=1}^T {\left( {\ln \left| {H_t } \right| +\varepsilon _t ^{\prime }H_t^{-1} \varepsilon _t } \right) } , \end{aligned}$$

(13)

where $\Theta$ is the parameter vector.

5 Empirical results

5.1 Full-sample estimations of return and volatility spillovers

Table 2 presents the full-sample estimation results of the mean equation. First, we focus on the cross-market coefficients of lagged returns, which measure the short-term impact of one market on another. The mean spillover effect is unidirectional from uncertainty to oil returns, in both spot and futures markets. Regarding the oil spot markets, the coefficient of uncertainty at the first and second lags is significant, whereas oil spot returns have no significant effect on uncertainty. Similar results are found regarding oil futures returns. Uncertainty has a lagged and significant impact on the oil futures market, with a slightly greater effect than the effect on oil spot markets. Therefore, an increase in EPU will depress both the oil spot and futures returns, and the oil futures markets respond in a more significant way than the spot markets.

Second, we focus on the coefficients of the error-correction terms, which reflect the price adjustment toward the long-run equilibrium relationship between any two markets. The results indicate that all the error-correlation coefficients are significant, suggesting that the markets react with respect to each other to maintain equilibrium. In this sense, economic uncertainty and oil markets are informationally efficient in the long term. Moreover, the efficiency of three markets in updating market information is inconsistent. Uncertainty adjusts to correct a disparity more quickly than oil spot and futures prices do. Besides, oil spot (futures) markets respond to the oil futures (spot) market with a large size than to EPU.

Third, the coefficients on the conditional variances of oil spot and futures returns are positive and statistically significant in the conditional mean equation of uncertainty, rejecting the hypothesis that no GARCH-in-mean effects exist from oil markets to uncertainty. The results suggest that fluctuations in oil prices may breed uncertainty related to economic policy.

Table 2 Return spillovers between economic policy-related uncertainty and oil returns (full-sample estimation)

Estimation results for conditional variance Eqs. (4)–(6) are provided in Table 3. Significant and positive bidirectional volatility spillovers between uncertainty and oil markets are observed. Fluctuations in uncertainty tend to decrease fluctuations in both oil spot and oil futures markets, and the reverse is also true.

However, spot and futures markets have different capabilities of spilling over their volatilities in terms of size and the statistical significance of coefficients. Focusing on the conditional variance of oil spot returns, the coefficient for volatility spillover from uncertainty is slightly greater than that from the oil spot and futures market. Similar results can be found in conditional variance equations of uncertainty and oil futures returns. This analysis illuminates the importance of uncertainty in information transmission, particularly in terms of volatility spillover.

Table 3 Volatility spillovers between economic policy-related uncertainty and oil returns (full-sample estimation)

5.2 Subsample estimations of return and volatility spillovers

Considering that spillovers between the uncertainty and oil prices may change dramatically over the last 20 years, we divide the sample period into two subperiods—January 1986 to December 2000 and January 2001 to March 2013—and perform subperiod analysis. To save space, we only report the results of return spillovers including conditional variances in mean equations [the model given by Eqs. (10–12)] in Table 4, while results for the model without including conditional variances of each variable in mean equations are very similar.

The results indicate that the return spillovers tend to vary considerably over time. First, focusing on the spillover effect of uncertainty on oil returns, the estimates for the first lagged uncertainty are relatively similar across the two periods, but also indicate a slightly higher coefficient estimate during 2001–2013 period, signifying that uncertainty has a stronger spillover effect on oil returns. The coefficients of both the effects of the second lagged uncertainty on oil spot and futures returns are not significant during the 1986–2000 period. Conversely, the coefficients for oil returns are significant with values of $-$0.0968 and $-$0.0947, respectively. Additionally, changes in uncertainty have a greater impact on the spot market in terms of the size of the coefficients and their significance. This observation suggests that the changes in uncertainty can be impounded for oil markets and information flows rapidly between uncertainty and oil markets. Second, regarding oil returns, none of the coefficient is significant during two periods, which is consistent with full-sample estimation results. This observation indicates that a unidirectional and asymmetric lead–lag relationship exists between uncertainty and the oil market. Third, our results reveal that all the error-correction coefficients are significant during the 2001–2013 period, suggesting that all these markets react with respect to each other to maintain equilibrium in the long term. However, none of these coefficients is significant in the first period.

Focusing on coefficients concerning conditional variances, during the 2001–2013 period, the estimated parameters are correctly signed and significant. Specifically, increased oil market volatility increases the level of uncertainty. However, this phenomenon cannot be observed in the 1986–2000 period.

Table 4 Return spillovers between economic policy-related uncertainty and oil returns including conditional variances of each variable in the mean equations (subsample estimations)

The results displayed in Table 5 indicate that the volatility spillovers also vary over time. In the 1986–2000 period, we observe no significant volatility spillovers between uncertainty and the oil market because none of the cross-market coefficients are significant. However, these effects are significant in the second period. Specifically, the coefficient for volatility spillover from uncertainty to the oil spot market is significant at the 1 % level with a value of $-$0.0256, and the effect in the opposite direction is $-$0.1191. Therefore, bidirectional spillover is present between uncertainty and oil markets, whereas the spillover effect from oil markets to uncertainty is greater than in the opposite direction.

Table 5 Volatility spillovers between economic policy-related uncertainty and oil prices (subsample estimations)

5.3 Dynamic conditional correlation

We now track time variation in spillover graphically via dynamic conditional correlation. In Fig. 2, we present the dynamic conditional correlation between uncertainty and oil futures returns along with the 90 % confidence intervals. Similar results can also be found for oil spot returns. Shaded portions of the figures are NBER recession dates, and the lines are dates during which the USA experienced oil price shocks, as determined by Kilian (2009). The most striking feature of the figures is the substantial variation in series interactions over time.

As observed in Fig. 2, the time-varying correlation between uncertainty related to economic policy and oil returns is generally positive over the period from January 1986 to February 2003, except for a rapid decrease during the period of oil shocks as well as recessions in 1991 and 2001. We observe a decreasing trend in correlation since the early part of 2003. Correlation becomes less positive and turns negative, and it remains at considerable negative correlation levels before the 2008 financial crisis. This observation may contribute to the elevated economic volatility of financial situations and the financialization of oil futures.

Although the time-varying correlation between uncertainty and oil returns is consistently negative over the 2004–2007 period, during the crisis, the correlation increases so sharply that it turns from approximately $-$0.9 in the later part of 2007 to almost +0.9 in mid-2009, and it retains a high level throughout the latter period of our sample, along with an upward trend. Interestingly, the correlation behaves heterogeneously during the three recessions covered by our sample: 1991, 2001 and 2007–2009. A plausible explanation can be found in the unprecedented bailout package for the US banking sector of 2008 and the stimulus package of 2009. However, the uncertainty remains high after these two packages, the oil market responds favorably, considering that in the later part of the financial crisis, the market is experiencing positive returns.

Fig. 2

Dynamic conditional correlation between uncertainty and oil futures returns

5.4 What explains the evolution of conditional correlations?

What explains the evolution of conditional correlations between uncertainty and oil market returns? To tackle this issue, we consider two kinds of shocks: oil shock and real economic shock. Real economic shock is a vector of changes in the real economy, namely industrial production growth, capital utilization and inflation, to account for changes in the real economy not captured by oil shock variables. Oil shock is a vector of three oil price shocks (namely supply-side shocks, oil-specific demand shocks and aggregate demand shocks) that have been estimated using the framework developed by Kilian (2009). Supply-side shocks are associated with changes in world oil production such as overproduction or supply interruptions due to war-related activities or destruction of oil facilities. Oil-specific demand shocks are associated with concerns about the future availability of oil, whereas aggregate demand shocks are associated with changes in the global economic activity. We identify the oil shocks using sign restrictions, as summarized in Table 6. The sign restrictions are imposed to hold for six months after the occurrence of the shocks, which accounts for the delayed response of the oil market variables in the early period of our sample due to institutional arrangements (Barsky and Kilian 2004). This identification has been widely used in Peersman and Van Robays (2009, 2012), Kilian and Murphy (2012) and Baumeister and Peersman (2013).

Table 6 Identification of oil shocks using sign restrictions

The oil price is the nominal refiner acquisition cost for imported crude oil taken from the US Department of Energy (DoE) database. We use the nominal price of oil because this should allow for a better identification of the different types of oil shocks. The results are robust when using the real price of oil. World oil production data are provided on a monthly basis by the US Department of Energy (DoE). We proxy aggregate economic activity by the OECD measure of global industrial production, which covers the OECD countries and six major non-OECD economies.

We then apply a Fisher transform on the estimated time-varying correlations $\rho _{ij,t}$, between uncertainty and oil market returns according to $\hbox {dcorr}_{ij,t} =\log \left( {\left( {1+\rho _{ij,t} } \right) /\left( {1-\rho _{ij,t} } \right) } \right)$ ($i,j=$ uncertainty index, oil spot price or oil futures price), and estimate various specifications of the following model (taking the correlation between uncertainty and oil futures price as an example):

$$\begin{aligned} \hbox {dccor}_{uf,t}= & {} a_f +\sum _{i=1}^I {b_{f,i} \hbox {dcorr}_{uf,t-i} } +\sum _{j=1}^J {f_{f,j} \hbox {RE}\_\hbox {shock}_{t-j} }\nonumber \\&+\sum _l^L {c_{f,l} \hbox {Oil}\_\hbox {shock}_{\mathrm{supply},t-l} } +\sum _m^M {d_{f,m} \hbox {Oil}\_\hbox {shock}_{\mathrm{specific}\_\mathrm{demand},t-m} }\nonumber \\&+\sum _n^N {e_{f,n}\hbox { Oil}\_\hbox {shock}_{\mathrm{aggregate}\_\mathrm{demand},t-n} } +\xi _t. \end{aligned}$$

(14)

Table 7 presents the results of the various forms of model (14). We find that both oil supply shocks and real economic shocks exhibit significant effects on the correlation level. Concerning the supply shock, the effect of the first and three lags of oil supply shocks is positive, whereas the effect of the second lag of oil supply shocks is negative. This result is expected because negative supply shocks would increase the uncertainty of the macroeconomic policies and increase oil returns at once, whereas the effects reverse in the next period and reverse again before the effects diminish. On average, the oil supply shocks have a significantly positive effect on the dynamic correlation. It means that due to the shortfall oil supply, the correlation between oil price and uncertainty is positive, suggesting that high uncertainty leads to high oil prices. This finding coincides with the fact that during the periods from September 1998 to September 2002, from July 2005 to September 2008, from March 2009 to January 2010, from January 2011 to November 2011 and from August 2012 to February 2013, oil supply shocks lead oil prices to rise rapidly.

Table 7 Estimations of dynamic correlations and the role of shocks

Regarding real economic shock, a significant negative effect is observed for both inflation and industrial production. We maintain that the results are expected, considering that an increase in fluctuations in inflation or industrial production would impose greater uncertainty in the macroeconomic policies, whereas oil prices would experience a bearish period, and thus the correlation between two variables will become more negative. This conclusion is consistent with the fact that during the post-crisis period combined with higher economic uncertainty, oil prices experience a sharp depression. Concerning the correlation fluctuations between 2007 and 2009, as shown in Fig. 2, three special periods should be noticed: the large interest rate cuts, the global financial crisis and the Obama election. Obviously, only during the financial crisis does the correlation between uncertainty and oil spot/futures returns remain negatively high. We can then conclude that during financial crisis, high uncertainty leads to low oil prices.

Therefore, the key factor in determining whether the effect of uncertainty on oil price is positive or negative depends on the relative strengths of two types of shocks. However, the oil supply shocks tend to establish a closer relationship between the oil market and uncertainty. This result has also been verified by Tverberg (2012) and Lütkepohl and NetŠunajev (2014) who argue that oil supply limits appear to be a primary cause, whereas the real economic activity has become less important for changes in the price of oil since the mid-1980s.

6 Conclusion

Using a new measure of macroeconomic policy-related uncertainty, we investigate the information transmission in terms of return and volatility spillovers between uncertainty and oil markets and explore the drivers behind the evolution of dynamic conditional correlations. Original results obtained include the following.

First, in the short term, the mean spillover effect between uncertainty and oil returns is negative and unidirectional from uncertainty to oil returns. An increase in policy uncertainty will depress both the oil spot and futures returns, and the oil futures markets respond in a more significant way than the spot markets. In the long term, policy uncertainty and oil markets under consideration are informationally efficient.

Second, significant and positive bidirectional volatility spillovers between uncertainty and oil markets are observed. However, policy uncertainty plays a dominate role in spilling over its volatility to oil markets in terms of size and statistical significance of the coefficients. In addition, the conditional variances of oil returns have positive and statistically significant effect on the conditional mean of policy uncertainty, suggesting that fluctuations in oil prices would increase uncertainty.

Third, we provide evidence that both oil supply shock with positive effect and real economy shock with negative effect lead the correlation between economic policy-related uncertainty and oil returns to fluctuate over time. Moreover, as to whether the effect of uncertainty on oil price is positive or negative, the key point is which type of shock dominates.