Macroeconomic Uncertainty and the Impact of Oil Shocks

Transcription

1 Macroeconomic Uncertainty and the Impact of Oil Shocks Ine Van Robays February 2012 Abstract This paper evaluates whether macroeconomic uncertainty a ects the impact of oil shocks. Using a structural threshold VAR model, we endogenously identify di erent regimes of uncertainty in which we estimate the e ects of oil demand and supply shocks. The results show that higher uncertainty tends to signi cantly increase the oil price impact of oil shocks for a given change in oil production, implying a lower price elasticity of oil demand and supply in the uncertain regime. The di erence in oil demand elasticities is economically meaningful as the price impact of oil supply shocks might easily double in volatile macroeconomic times. Accordingly, heightened uncertainty about the macroeconomic outlook can explain time variation in the price elasticity of oil and hence in oil price volatility. In addition, also the impact of oil shocks on economic activity appears to be signi cantly reinforced by uncertainty. JEL classi cation: E31, E32, Q41, Q43 Keywords: Oil price, uncertainty, price elasticity, threshold VAR, sign restrictions Ghent University, Department of Financial Economics, W. Wilsonplein 5D, B-9000 Gent, Belgium, ine.vanrobays@ugent.be. I am grateful to Nathan Balke for providing his computer codes on which my estimations are based, and to Gert Peersman for useful comments and suggestions. I acknowledge nancial support from the Interuniversity Attraction Poles Programme - Belgian Science Policy [Contract No. P6/7] and Belgian National Science Foundation. 1

2 1 Introduction The remarkable increase in oil price volatility over the past decade sparked a discussion about its driving factors. Many studies argue that the increased oil price uctuations can be explained by fundamental oil supply and demand-side factors (Baumeister and Peersman 2008, 2010; Hamilton 2009, Kilian and Murphy 2010), whereas others claim that also nancial speculation could have played some role (Lombardi and Van Robays 2011, Masters 2009, Singleton 2011). However, given that the nancial crisis lead to a global downfall in economic activity and substantially increased uncertainty about the macroeconomic outlook, a possible determinant went unnoticed, namely uncertainty itself. It is well documented that increased uncertainty can in uence the decision behavior of economic agents and cause a delay in the production and consumption decision, thereby lowering the quantity adjustment and increasing the price response after shocks (Bernanke 1983, Pindyck 1991, Litzenberger and Rabinowitz 1995, Bloom et al. 2007). Analogously, uncertainty could a ect the way in which oil shocks impact on oil prices and production, and subsequently on economic activity. In this paper, we evaluate whether the impact of oil shocks di ers in times of increased uncertainty. In contrast to most of the recent literature on uncertainty, we do not regard uncertainty as a shock but as a channel of transmission through which the impact of oil supply and demand shocks might be reinforced. 1 Using a threshold vector autoregressive (TVAR) model, we endogenously identify high and low uncertainty regimes based on a measure of macroeconomic volatility crossing an estimated threshold. Conditional on being in a particular regime, we quantify the impact of di erent types of oil shocks on oil prices, oil production and economic activity. Similar to Peersman and Van Robays (2009, 2012), Baumeister and Peersman (2010) and Baumeister, Peersman and Van Robays (2010), we identify three types of oil shocks using sign restrictions; oil supply shocks, oil demand shocks driven by economic activity, and other oil demand shocks labeled oilspeci c demand shocks. Our results show that the impact of oil demand and supply shocks tends to di er substantially when macroeconomic uncertainty is high. Oil supply shocks appear to have a much stronger e ect on oil prices for a given response of oil production, implying that 1 The literature on the e ects of uncertainty shocks was pioneerd by Bloom (2009), and has expanded quickly since then. 2

3 the price elasticity of oil demand is lower in the high uncertainty regime. The di erence in estimated oil demand elasticities is economically very signi cant, as the price impact of a similar change in oil production might easily double when the oil supply shock hits the economy in uncertain times. More speci cally, we estimate the impact oil demand elasticity to decline from a range of to when uncertainty is low, to to when uncertainty is high. Not only oil prices and oil production react di erently, but also economic activity appears to react more aggressively to oil shocks when macroeconomic volatility is already high. A stronger positive e ect on inventories in the high uncertainty regime might be explained by increased precautionary demand. We nd similar results following the oil demand shock driven by economic activity, implying that also the oil supply elasticity is signi cantly lower when uncertainty is higher, although the di erence across regimes is somewhat less pronounced. Hence, we show that di erent levels of macroeconomic uncertainty can explain time variation in the price elasticity of oil, and therefore in oil price volatility. For example, Hamilton (2009) and Kahn (2009) argue that a lower price elasticity could explain why fundamental oil supply and demand shocks impacted more strongly on oil prices over the last decade, and we empirically demonstrate that this could have been the case because of higher uncertainty. In contrast to the other oil shocks, we do not nd clear-cut di erences in the estimated price elasticity following the oil-speci c demand shock. However, also after this shock, the real economic e ects appear to be more severe in uncertain times. Finally, we observe that oil demand shocks driven by economic activity lower macroeconomic uncertainty, whereas oil supply shocks and other oil demand shocks that raise oil prices increase uncertainty. This paper relates to di erent strands of the oil literature. Several studies have touched upon the relationship between uncertainty and oil prices, but mostly they only regard uncertainty with respect to the oil price itself, i.e. oil price volatility instead of macroeconomic volatility more generally (Bredin et al. 2011, Elder and Serletis 2010, Ferderer 1996, Lee, Ni and Ratti 1995, Pindyck 2004). 2 Moreover, these papers mostly document the real e ects of increased oil price volatility, whereas this paper nds an endogenous explanation for variation in oil price volatility itself. On the other hand, this paper also builds upon the studies that have documented an increase in the volatility of the oil price over time, 2 Two exceptions to this are Pindyck (1980) and Litzenberger and Rabinowitz (1995), although their focus is di erent. Pindyck (1980) concentrates on the theoretical e ect of demand and oil reserves uncertainty on expected oil price behavior, and Litzenberger and Rabinowitz (1995) focus on explaining backwardation in oil futures markets. 3

4 and those that relate this increased price volatility to varying elasticities of oil demand and supply (Lee, Ni and Ratti 1995, Ferderer 1996, Regnier 2007). In a time-varying framework, Baumeister and Peersman (2008, 2010) nd a structural break in the shortrun oil price elasticity of oil demand and supply around the mid 1980s that can explain both the observed increase in oil price volatility and decline in oil production volatility. They argue that the oil price elasticities of oil demand and supply have behaved relatively stable since then. We, however, observe that there is still considerable time variation in the price elasticities since the mid 1980s, which we endogenously explain by variation in the level of macroeconomic uncertainty. The main contribution of this paper is therefore to document the relationship between macroeconomic uncertainty and the impact of oil shocks, and empirically test whether this uncertainty e ect might explain why oil price elasticity, and hence oil price volatility, appears to change over time. The remainder of this paper is organized as follows. In the next section, we rst describe the threshold VAR model and its speci cation, test for threshold e ects and explain the identi cation strategy. The empirical results are discussed in Section 3. In Section 4, we discuss some potential channels through which macroeconomic uncertainty can a ect the price elasticity of oil demand and supply, and Section 5 concludes. 2 Model and Identi cation 2.1 Threshold VAR model To evaluate the role of macroeconomic uncertainty on the oil market, we rely on a structural threshold vector autoregressive (TVAR) model. The threshold model is attributed to Tong (1978) and has been extensively used afterwards, see Hansen (2011) for an overview. The TVAR model enables us to endogenously identify di erent regimes with respect to one endogenous transition variable, which is called the threshold variable. In our case, this is a function of macroeconomic uncertainty. The di erent regimes are determined by the value of this threshold variable with respect to a certain threshold which is estimated within the model. Once the di erent regimes are identi ed, we generate the impulse response functions conditional upon the regime to compare the estimated e ects. In Markov-Switching models, in contrast, the transition variable is typically not observed, which makes the TVAR model particularly attractive for addressing our research question. We estimate a 4

5 two-regime TVAR model of the following form: Y t = 1 + A 1 Y t + B 1 (L) Y t A 2 Y t + B 2 (L) Y t 1 It (c t d ) + u t The vector of endogenous variables Y t captures the global dynamics in the oil spot market, i.e. world oil production (Q oil ), the price of crude oil expressed in US dollars (P oil ), a measure of world economic activity (Y w ) and oil inventories (I oil ). To model di erent uncertainty regimes, we also add a measure of macroeconomic uncertainty denoted by U. The variable c t d is the threshold variable and I t (:) is an indicator function that takes value one when the d-lagged value of the threshold variable is higher or equal to the estimated threshold, and zero otherwise. This indicator function thus determines the regimes based on the value of c t d relative to. As the threshold variable c t d is a function of macroeconomic uncertainty and subsequently an endogenous variable in the TVAR model, shocks to the oil market as well as to macroeconomic uncertainty are allowed to determine whether the economy is in a high or low uncertainty regime. 3 a vector of constants, B (L) is a matrix polynomial in the lag operator L and A is the contemporaneous impact matrix of the vector of orthogonalized error terms u t. The TVAR model allows for non-linearity in the e ects across regimes as each regime has di erent autoregressive matrices. If I t = 0, the dynamics of the system are given by 1, A 1 and B 1 (L), and if I t = 1, the relevant coe cients are 1 + 2, A 1 + A 2 and B 1 (L) + B 2 (L). Note that the contemporaneous impact of the shocks is allowed to vary, which is crucial for our analysis of the price elasticities on impact for example. The oil price is the nominal re ner acquisition cost of imported crude oil, which has extensively been used in the literature as the best proxy for the free market global price of imported crude oil. We proxy global economic activity by the OECD measure of global industrial production, which covers the OECD countries and the six major non-oecd economies, including e.g. China and India. Following Kilian and Murphy (2010), we proxy global crude oil inventories as total US crude oil inventories scaled by the ratio of OECD petroleum stocks over US petroleum stocks. Similar to Baum and Wan (2010), we measure macroeconomic uncertainty by the conditional variance series of US GDP growth, or in other words its volatility. We generate a monthly GDP series by interpolating quarterly GDP using industrial production based on the Chow-Lin procedure, after which we model the conditional variance as a GARCH(1,1) model. 4 To ensure robustness 3 We discuss possible endogeneity issues later in this section. 4 The GARCH(1,1) gives the best speci cation for modelling the conditional variance according to is 5

6 of our ndings, we construct two additional measures of uncertainty; one is the conditional variance of US industrial production growth, and the other is the Chicago Board of Exchange VXO stock market volatility measure. This VXO index is based on a hypothetical at the money S&P100 option, and is the measure of uncertainty used by Bloom (2009). We constructed a monthly series of the VXO index by taking monthly averages of the daily closing price. As noted by Baum and Wan (2010), these di erent measures capture di erent types of uncertainty. The measure based on GDP growth is designed to re ect the overall uncertainty of the macroeconomic environment, whereas the measure based on industrial production disregards uncertainty about the service sector. The VXO stock market volatility measure is more closely related to nancial market uncertainty. Given the focus of this paper, we use the rst measure as our preferred indicator of macroeconomic uncertainty. The results however indicate that the main conclusions hold for the other measures of uncertainty as well. We do not use the world industrial production measure to proxy uncertainty for the reason that we restrict the response of this variable in the identi cation of the structural oil shocks, as explained later on in Section 2.3. Using this index could therefore signi cantly bias the results. 5 The TVAR model is estimated using monthly data over the period 1986: :07. We choose 1986 as our starting point for the reason that Baumeister and Peersman (2008, 2010) document a substantial decline in the price elasticity of oil demand and supply around this date. Not accounting for this structural break could seriously bias our estimations of the pice elasticities, and the identi cation of the regimes could be in uenced by the fact that macroeconomic volatility was much higher before the Great Moderation. 6 various information criteria. We estimated the conditional variance over the period to avoid a possible bias due to the Great Moderation. 5 For example, following oil demand shocks driven by economic activity, the reaction of world industrial production is positive which mechanically pushes uncertainty upwards, even if uncertainty would otherwise be una ected or fall, which is what we nd for all other measures of uncertainty. If we construct macroeconomic uncertainty using the world industrial production measure, the positive impact response of uncertainty to global oil demand shocks is up to 13 times as high compared to using US GDP growth, indicating that this using this measure could signi cantly bias our results.moreover, restricting the maximum response of uncertainty to be equal to the threshold, or restricting the response of uncertainty to be nonpositive following the oil demand shock driven by economic activity signi cantly changes the results in favour of our benchmark results, i.e. the estimated elasticity becomes smaller following the global oil demand shock. 6 An important drawback of starting our sample in 1986 is that this considerably limits the number of observations available for estimation. Doing a similar excercise for the 1970s-1980s would be an interesting We 6

7 include four lags of the endogenous variables based on the conventional lag length criteria. 7 All the variables are transformed to monthly growth rates by taking the rst di erence of the natural logarithm, except for the measure of macroeconomic uncertainty. In general, the results are robust to di erent speci cations of the variables and the structural TVAR model Test for Threshold E ects and Identi cation of Regimes Before testing whether the model is indeed non-linear, we have to decide on the exact speci cation of the threshold variable. Typically, the threshold variable is assumed to have a certain delay in determining the regimes and is modeled as a moving average process depending on the persistence of the series (Balke 2000). As it is plausible that higher macroeconomic volatility could immediately result in higher uncertainty, we ideally would set the delay parameter equal to zero. For econometrical reasons, however, we put it equal to one to exclude the possibility that the regimes are directly driven by oil shocks, which could hamper the interpretation of the results. 9 We model the threshold variables as a moving average of order three to allow for some persistence in the uncertainty regimes, which corresponds to the average volatility of the past quarter, as the measures of uncertainty that we employ are highly volatile. To test for the signi cance of threshold e ects, we use the approach described in Balke (2000). If the threshold value would be known, the test of linearity under the null hypothesis against the presence of threshold behavior would simply come down to testing whether 2 = A 2 = B 2 (L) = 0. However, as this is not the case, we have to rely on nonstandard inference. A commonly used approach is to estimate the model for each possible value of the threshold variable using least squares. The range of possible thresholds is trimmed by a certain percentage to allow for su cient observations in each regime. Similar robustness check of our results, but the degrees of freedom available for estimation are too limited for the results to be reliable. The ndings for the oil demand elasticity are however robust to estimating the TVAR model over the period Although Hamilton and Hererra (2004) argue that at least a year worth of lags should be included in the model to allow for su cient dynamics, we choose to only include four lags given the limited number of observations in the higher uncertainty regime in particular. 8 Our conclusions hold for the real oil price, reasonable variation in the number of lags given our data sample (2,3 and 5 lags) and di erent measures of uncertainty as described in the main text. These results are available upon request. 9 We discuss the possibility of endogeneity in more detail later on. 7

8 to Balke (2000), we choose a trimming parameter of 15 percent. Conditional on each threshold, we calculate the Wald statistic that evaluates the hypothesis of equality between the regimes. Three di erent summary test statistics are generated: the maximum Wald statistic of all possible threshold values (sup-wald), the average Wald statistic (avg-wald) and a statistic calculated as a function of the sum of the exponential Wald statistics for all possible thresholds (exp-wald). For the reason that the distribution of these test statistics is non-standard, we rely on the simulation technique proposed by Hansen (1996) to simulate the unknown asymptotic distributions. This enables us to derive the p-values associated with the test statistics. The estimated threshold value is the one that maximizes the log determinant of the variance-covariance matrix of residuals. Table 1 shows the threshold test results for the di erent measures of uncertainty. There is strong evidence for signi cant threshold e ects for all measures of uncertainty according to the three Wald test statistics. 10 The threshold based on the rst and preferred measure of macroeconomic uncertainty using US GDP growth is estimated to be , which splits the sample into high and low uncertainty regimes that represent respectively 22 and 78 percent of all observations. To put this into perspective, Panel A of Figure 1 illustrates the threshold variable, the estimated threshold and the identi ed regimes for this measure of uncertainty. The shaded areas correspond to the high uncertainty states, when the threshold variable surpasses the threshold value. A rst notable period of higher uncertainty is when stock markets around the world crashed on Black Monday at the end of A second prolonged period of elevated uncertainty is identi ed to be the US recession that started at the end of High macroeconomic volatility also caused increased uncertainty at the end of 1998, which could be explained by the dragging e ects of the Asian nancial crisis and the Russian nancial crisis in August 1998 which lead global markets into turmoil because of the Russian debt default and the bail-out of the Long-Term Capital Management (LTCM) hedge fund. The next period of prolonged high uncertainty can be associated with the US recession of the early 2000s and more generally, the global slowdown in economic activity in many industrialized countries. The 9/11 Terrorist Attacks elevated uncertainty for a few months at the end of Not surprisingly, a nal marked period of high economic volatility and uncertainty is the nancial crisis that hit global markets in summer 2008, and lead to a substantial fall in 10 Moreover, the results are robust for the threshold variable modelled as a moving average process of order 4, 5 and 6 and a trimming parameter of 10 percent, as suggested by Hansen (1999). These results are available upon request. 8

9 economic activity world-wide. Note that before the onset of the nancial crisis, overall macroeconomic uncertainty was already elevated, which can be ascribed to a recession in the US and a slowdown in economic growth in other major industrialized countries. More recently, concerns about the sovereign debt crisis in the euro area and its possible economic e ects might explain why uncertainty is again higher. These periods of higher uncertainty correspond well with those identi ed in Bloom (2009). 11 Panel B and Panel C of Figure 1 show the identi ed regimes for the uncertainty measure based on the conditional variance of US industrial production and the VXO stock market volatility index. Although it is clear that the VXO measure captures nancial market uncertainty more closely, the di erent regimes of high and low uncertainty correspond well with those identi ed using GDP volatility. 12 It is well known that oil shocks can lower economic activity and cause recessions (e.g. Hamilton 1983, 2009; Bjørnland 2000; Peersman and Van Robays 2009, 2012). Accordingly, as higher oil price movements might also cause higher uncertainty, the results might be subject to an endogeneity bias. Assuming that macroeconomic uncertainty is strictly exogenous with respect to oil shocks, however, might not be realistic. For that reason, the TVAR model allows macroeconomic uncertainty to endogenously respond to oil shocks. This is important as the oil demand shock driven by economic activity is by de nition closely related to macroeconomic volatility. However, the regimes are determined by the value of the threshold variable, which is a function of macroeconomic uncertainty. The threshold variable is de ned as a moving average process of macroeconomic uncertainty and is assumed to only switch regimes with a delay of one period. Hence, oil shocks will not cause a regime shift within the same month as the shock hits, and by modelling the threshold variable as a three-month moving average process, there should be some persistence in the increase of macroeconomic uncertainty before it can trigger a regime switch. 13 Moreover, the correlation between oil price changes and macroeconomic uncertainty equals 11 Bloom (2009) identi es 17 volatility shock events that substantially increased uncertainty, which he uses as arguably exogenous shocks to empirically evaluate the e ect of uncertainty shocks. Most of these shocks are caused by economic events, war or terrorism. 12 Using the VXO index, high uncertainty is concentrated around the Black Monday event, the Russian and Long-Term Capital Management (LTCM) default, 9/11 Terrorist attack, the Enron and Worldcom accounting scandals, Gulf War II and the nancial crisis. The working paper version of Bloom (2009) provides more details on these events. 13 Probably for this reason, the regimes of high uncertainty that we identify are mostly associated with recessions, when macroeconomic volatility is higher than the threshold for several months in a row. 9

10 -0.25 and is signi cant. Most of the high uncertainty events identi ed are also not directly linked to oil shocks (see also Bloom 2009), and the results are robust to using nancial uncertainty instead of macroeconomic uncertainty. Second, looking ahead at the results con rms that a possible endogeneity bias is small if not negligible. The di erent types of oil shocks do not signi cantly a ect uncertainty on impact in the high uncertainty regime, and in the lower regime, there is no signi cant e ect on uncertainty over all horizons. This is also the case when we estimate a linear VAR model containing the same set of variables over the total sample. In addition, the conditional variance decompositions show that the contribution of the oil shocks in explaining variability in macroeconomic uncertainty is very small. 14 Finally, if we restrict the impact response of macroeconomic uncertainty to the di erent oil shocks to be zero in our estimations, the results turn out to be very similar. 2.3 Identifying Oil Shocks using Sign Restrictions Independent on whether the model is non-linear or not, we face the problem that the contemporaneous errors could be correlated. In order to make the shocks orthogonal and thereby econometrically interpretable, we need to impose structure on the model to identify the di erent shocks. Given that we only want to evaluate whether uncertainty acts as a reinforcer of oil shocks, we are only interested in identifying the oil shocks. The oil literature has increasingly recognized that di erent factors can drive oil price movements, and that the economic e ects of those shocks crucially depend on the underlying source of the oil price change (Kilian 2009, Peersman and Van Robays 2009, 2012). Following Baumeister and Peersman (2010), Peersman and Van Robays (2009, 2012) and Baumeister, Peersman and Van Robays (2010), we will therefore separate three di erent types of oil shocks by using sign restrictions: oil supply shocks, oil demand shocks driven by global economic activity and oil-speci c demand shocks. We will not explicitly model oil speculation shocks for the reason that its importance as well as its de nition is still strongly debated. 15 We also exactly want to assess whether uncertainty caused shocks to 14 The contribution of the oil supply shock, oil demand shock driven by economic activity and the oil-speci c demand shock to the contemporaneous median variance decompositions of macroeconomic uncertainty are respectively 3, 4 and 6% in the upper regime, and 3, 2, 3% in the lower regime. Estimated over the full sample, all oil shocks contribute about 4% on impact. 15 Lombardi and Van Robays (2011) identify an ine cient nancial shock to oil futures markets in addition to the fundamental oil shocks. We refer to that paper for an extension of this framework to 10

11 oil supply and demand fundamentals to have a stronger price impact over the past decade, and that this increased volatility should not necessarily be linked to nancial activity. We identify the oil shocks by relying on the following set of sign restrictions: 16 STRUCTURAL SHOCKS Q oil P oil Y w I oil U Oil supply < 0 > 0 0 Oil demand driven by economic activity > 0 > 0 > 0 Oil-speci c demand > 0 > 0 0 The sign restrictions are derived from a simple supply-demand scheme of the oil market. An oil supply shock is an exogenous shift of the oil supply curve to the left and therefore moves oil prices and production in opposite directions. Production disruptions caused by military con icts in the Middle-East are natural examples. As oil prices are higher, global industrial production will not increase following this supply shock. In contrast, shocks on the demand side of the oil market will result in a shift of oil production and oil prices in the same direction. On the one hand, demand for oil can endogenously increase because of changes in macroeconomic activity. A change in the demand for commodities from emerging economies like China or India for example, will shift world economic activity, oil prices and oil production in the same direction. We de ne such a shock as an oil demand shock driven by economic activity. On the other hand, oil demand can also vary for reasons not related to economic activity. We label these shocks as oil-speci c demand shocks. Shocks to expected oil demand in the future, which increases oil inventory demand as a precaution, and oil-gas substitution shocks are two examples. In contrast to the demand shock driven by economic activity, oil-speci c demand shocks do not have a positive e ect on global economic activity as oil prices are higher. These sign conditions are imposed on the estimated impulse response functions of the variables in the TVAR model, and we follow the procedure of Peersman (2005) and Peersman and Van Robays (2009, 2012) for estimation and inference. 17 To evaluate the change in impact of oil shocks on the oil price elasticity under di erent regimes of uncertainty, we construct conditional impulse response functions, i.e. condiinclude shocks related to speculation. 16 The sign restrictions are shown for oil shocks that increase the oil price. 17 Similar to Peersman and Van Robays (2009, 2012) and Baumeister and Peersman (2008, 2010), we impose the sign conditions to hold for the rst year after the shocks which shoul improve identi cation, see Paustian (2007). The results on the elasticities are however robust to only imposing the signs on impact. 11

12 tional upon a speci c uncertainty regime. Although the impulse response functions are assumed to be linear within a regime, the size and persistence of the responses to similar oil shocks can di er across regimes. Mostly, the e ects of shocks in a TVAR model are evaluated using so-called generalized impulse response functions, which, in contrast to conditional impulse response functions, allow that shocks can cause a switch in regime over the duration of the response. 18 For the reason that we use sign restrictions instead of a recursive identi cation scheme, however, constructing generalized impulse response functions proves to be very di cult. Up to our knowledge, only Candelon and Lieb (2011) have used TVAR models in combination with sign restrictions, and they make the same assumption as we do here. A drawback of not allowing the shocks to cause switches in regimes during the response might be that the conditional impulse response functions are only informative in the short run. Concerning the estimation of the impact price elasticities of oil demand and supply, however, this assumption should not make any di erence. 3 E ects of Oil Shocks in Di erent Uncertainty Regimes Figure 2 shows the estimated e ects of the variables in the TVAR model to di erent types of oil shocks in the two regimes. In order to make the e ects comparable across regimes, we normalized the contemporaneous response of oil production to a one percent change. The conditional impulse responses are accumulated and shown in levels over the rst two years after the shock. The shaded responses in the gure represent the 68 percent posterior probability range of the estimated e ects in the high uncertainty regime and the dotted ones represent those conditional on low uncertainty. Although the estimation uncertainty surrounding the e ects in the upper regime are generally larger than those in the lower regime because of the limited number of observations, we still nd some notable di erences in the e ects due to uncertainty. The rst column in Figure 2 shows the e ects of an exogenous oil supply shock. For a similar impact change in oil production, the oil price response tends to be much stronger in the high uncertainty regime. As the response of oil production relative to that of the oil price following an oil supply shock gives an estimate of the price elasticity of oil demand, this implies that the elasticity of oil demand falls substantially when uncertainty is high. 18 See for example the working paper version of Calza and Sousa (2006) for more details, as they construct both the conditional and the generalized impulse response functions. 12

13 Speci cally, we estimate the oil demand elasticity to decrease from within a range of to in the low uncertainty regime, to a value within the range of to in the high uncertainty regime. Although the impact elasticities somewhat overlap across regimes, additional tests show that this di erence is statistically signi cant. Given that the oil price elasticity in the high uncertainty regime might easily be less than half its value of the low uncertainty regime, the e ect is also economically very signi cant. The estimated oil demand elasticities are well within the range estimated in the literature. At the lower end of the probability range, Hamilton (2009), Dahl (1993) and Cooper (2003) report oil demand elasticities between and -0.07, whereas Baumeister and Peersman (2010), Bodenstein and Guerrieri (2011) and Kilian an Murphy (2010) arrive at estimates ranging from to -0.44, which is at the higher end of our estimation range. Kilian and Murphy (2010) argue that allowing for endogeneity of oil price could be a reason for why they nd high oil demand elasticities. Our model however, despite modelling oil prices endogenously, also generates very low elasticities once we allow for endogenous nonlinearity in the price elasticity depending on the economic regime. Interestingly, using a time-varying VAR model and thereby allowing for non-linearity, Baumeister and Peersman (2010) estimate the median price elasticity of oil demand to uctuate within a range of to since 1986, with 68 percent posterior credible sets reaching up to -0.40, which almost exactly covers our estimation range over the two regimes. We show that the variation of the oil demand elasticity within their sample could be explained by varying levels of macroeconomic uncertainty. Not only oil production and the oil price, but also the real economic e ects of oil supply shocks appear to di er considerably when uncertainty is high. The fourth row of Figure 2 shows that economic activity appears to react more strongly and more quickly following oil shocks in the high regime. The impact response in the high uncertainty regime might be almost twice as large than when uncertainty is low, which could be explained by the increased sensitivity of the oil price. The di erence in responses also appears to be statistically signi cant. At rst sight there is no apparent di erence between the reaction of oil inventories across regimes, although beyond impact the response of oil inventories turns out to be signi cantly higher when macroeconomic uncertainty is elevated. This might be explained by a precautionary demand e ect as economic agents that have storage capacity change their inventory holdings conditional on varying economic uncertainty about the future, as argued by Pirrong (2009). Not surprisingly, the level of 13

14 macroeconomic uncertainty slightly increases following an oil supply disruption that pushes oil prices upwards when uncertainty is already high. In the low uncertainty regime, the level of macroeconomic volatility is not signi cantly a ected. responses across regimes appears to be statistically signi cant. Again, the di erence in The second and third column illustrate the estimated responses following both types of oil demand shocks. Analogous to the calculation of the oil demand elasticity, the ratio between the oil production and oil price reaction following the oil demand shocks estimates the oil supply elasticity. For the reason that we have two types of oil demand shocks, we generate an estimate of the curvature of the oil supply curve following oil demand shocks driven by economic activity and following oil-speci c demand shocks, similar to Baumeister and Peersman (2010) and Peersman and Van Robays (2012). Following the oil demand shock driven by economic activity, we also nd that the oil price response is likely to be much stronger when uncertainty is higher. 19 This implies that also the elasticity of oil supply tends to be lower when uncertainty is higher, although the di erence between the regimes is somewhat less pronounced than in the case of the oil demand elasticity. Nevertheless, the di erence is statistically signi cant. The estimated oil supply elasticity drops from a maximum value of 0.29 in the low uncertainty regime to a maximum of 0.16 when uncertainty is high. The minimum estimated elasticity of oil supply reduces from 0.05 to Again, these estimates correspond well with the estimates in the literature. Baumeister and Peersman (2010), for example, estimate the median oil supply elasticity to lie in between 0.02 and However, when the oil supply elasticity is generated through a shift in the oil-speci c demand curve, we do not longer nd a signi cant di erence across regimes, even though the estimated probability range is somewhat lower when uncertainty is higher. This nding is probably due to the fact that the oil-speci c demand shock captures a broad set of shocks, i.e. economic activity. all demand shocks that are not driven by global Shocks to expected net oil demand and oil-gas substitution shocks are two examples, and also speculation shocks are thought to be part of it. 20 For the reason that these shocks could trigger diverging responses in oil demand and supply, the estimation of the oil price elasticities could be subject to considerable noise. As noted 19 During some periods of high uncertainty, small changes in oil demand were associated with enormous variation in the oil price, which might explain why the estimation uncertainty surrounding the oil price response is so high. For example, in the fourth quarter of 2008, oil demand felt by 0.6 percent whereas oil prices plummeted by more than 111 percent. 20 See for example Kilian and Murphy (2010) and Lombardi and Van Robays (2011). 14

15 by Baumeister and Peersman (2010), the di erences in the estimated elasticities could also be explained by a di erent reaction of oil supply to both shocks in oil demand. However, following both types of oil demand shocks, we nd that the e ects on economic activity are more aggressive when the shock hits in already volatile macroeconomic times. In contrast, the response of oil inventories is not signi cantly di erent across regimes. Interestingly, macroeconomic uncertainty slightly lowers following the oil demand shock driven by economic activity, whereas it increases after the oil-speci c demand shock. These opposite reactions can be linked to the fact that the oil-speci c shocks that increase the oil price are unfavorable shocks to oil-importing countries, whereas the oil demand shock by de nition is associated with increased economic activity and could therefore be favorable, certainly in times of high uncertainty. In contrast, oil demand shocks generally do not signi cantly change the level of macroeconomic uncertainty when uncertainty is low. A simple back-on-the-envelope calculation illustrates the economic relevance of the di erence in estimated elasticities. In the aftermath of the nancial crisis that hit the global economy in summer 2008, oil demand dropped considerably. Global oil demand declined with about two percent between 2008Q3-2009Q2 and oil prices decreased from about USD 112 to USD 58 per barrel, a decline of 52 percent. Based on our estimates, uncertainty concerning the macroeconomy was already high before the nancial crisis hit (see Figure 1). If we assume the price elasticities of oil supply in the di erent regimes to be equal their average value, the part of the oil price decline that could be attributed to the uncertainty e ect would be about 10 percent. In levels, this would mean that the price of oil would only have decreased from USD 112 to about USD 69, instead of to USD 58 per barrel. 21 This strengthens the view that oil supply and demand-side fundamentals may have been responsible for most part of the sharp movements in oil prices, as high uncertainty about the macroeconomic outlook reinforced the price impact of these fundamental oil shocks, independent of any speculative activity in the oil futures market. The nding that the oil demand elasticity and the oil supply elasticity (following the oil demand shock driven by economic activity) tends to be smaller when uncertainty is higher is robust to using the other measures of uncertainty that we constructed, see Panel B and C of Figure 3 in comparison with Panel A. 21 For simplicity, we made the assumption that the two percent drop in global oil demand is entirely caused by an oil demand shock driven by economic activity, and that the drop in production is equal to the drop in demand. These and the other assumptions made could be restrictive, and therefore these results should be interpreted with caution. 15

16 4 How Can Uncertainty A ect the Oil Market? A lower price elasticity of oil demand and supply during uncertain economic times means that shocks hitting the oil market generate larger price responses but smaller quantity reactions compared to more certain times. In this section, we discuss several possible ways in which macroeconomic uncertainty can negatively impact on the price elasticity of oil demand and supply, which are not mutually exclusive. First, both oil demand and oil supply could be less responsive because of an option value to wait. Under the condition that the action to be decided on is irreversible, uncertainty creates an option value to wait through which investors are willing to forego current returns in order to gain from more information that will become available in the future. Bernanke (1983) shows that this concept is successful in explaining cyclical uctuations in investment, and in more recent work, Bloom et al. (2007) and Bloom (2009) con rm that rms delay the decision-making process because of higher uncertainty. 22 Accordingly, following an oil demand shock that occurs when macroeconomic volatility is already high, crude oil producers could decide to wait with changing their production until more information is available on the persistence of the oil shock as well as its impact on the already fragile economy. This option value to wait would then lower the elasticity of oil supply. Analogously, the elasticity of oil demand could be lower as oil consumers prefer to wait with reducing their demand following an oil supply shock that pushes oil prices upwards. In addition, uncertainty could reduce the tendency of oil consumers to substitute oil for other energy products, or at least delay substitution until there is more certainty about the e ect of the oil shock. Second, the existence of futures markets might also play a role in explaining a lower elasticity of both oil demand and supply. Baumeister and Peersman (2010) note that as producers and consumers are hedged against movements in the price of oil, they could lower their responsiveness of oil demand and supply. Accordingly, if higher macroeconomic uncertainty leads to an increased use of futures contracts, which is plausible given that futures markets exist to transfer risks, it could cause the oil price elasticity of demand and supply to decline. An additional channel through which the oil supply elasticity could decline during uncertain periods is discussed by Litzenberger and Rabinowitz (1995). They explain why oil 22 There is an extensive literature that deals with the e ect of uncertainty on investment dynamics, some examples are Henry (1974a,b), Pindyck (1991), Brennan and Schwartz (1985), Majd and Pindyck (1987), Elder and Serletis (2010) and Bredin et al. (2011). 16

17 producers that own reserves could prefer to lower production when uncertainty increases, thereby aggravating the price e ect. In a two period equilibrium model, Litzenberger and Rabinowitz (1995) show that uncertainty increases the value of oil reserves for any level of the extraction cost, or at least keeps it constant. Key is that having oil reserves below the ground can be viewed as owning a call option for the oil producer, in the sense he can buy oil against the extraction cost, which is the exercise price in this case. The optimal level of oil production is determined by a trade-o between the net value of oil above the ground, which is the spot price minus the extraction costs, and its value below the ground, which is the price of the call option. Oil producers will not extract oil as long as the net value of oil above the ground is lower than the price of the call option. As uncertainty about demand rises, and the outlook becomes riskier, the value of the call option will increase. Therefore, oil producers will prefer to leave oil reserves below the ground when uncertainty rises, aggravating the price e ect of oil shocks. Litzenberger and Rabinowitz (1995) also nd empirical support for their argumentation. Moreover, uncertainty could also a ect the price setting mechanism in the oil spot and futures markets, without the need for immediate oil demand and supply adjustments. Singleton (2011) touches upon di erent possible ways in which increased uncertainty can have an e ect on the price elasticity, although not explicitly. First, in a simple model he formulates the spot price as a function of a risk premium that investors demand for trading in the spot and futures market. As the willingness to bear risk changes over time, the risk premium and its e ect on the price of oil also vary. In uncertain economic times, investors are typically less inclined to take risks, or are confronted with more risk, which increases the risk premium. 23 Given a certain change in oil production, time-varying risk premia could explain higher price responsiveness following oil shocks in volatile times. A second way in which uncertainty can a ect the reaction of oil prices in futures markets is through heterogenous beliefs. Singleton (2011) describes that investors can have di erent opinions about public information concerning the future course of economic events and their impact on the price of oil. He argues that these heterogenous beliefs can induce higher price volatility, price drifts and even booms and busts in prices. Singleton (2011) adds that because of these heterogeneous beliefs, and the fact that investors learn about the economic environment, the release of new information about oil supply and demand can have a 23 In the model of Singleton (2011), a risk premium exists as in reality expectations deviate from the risk neutral setting. Therefore, the risk premium is also related to risk aversion, in addition to capturing the consequences of limits to arbitrage including nancial market frictions. 17

18 large e ect on prices. Although he uses these arguments to back up a potential role for speculation, they can also explain why in times of higher macroeconomic uncertainty, when investors beliefs typically diverge more than in normal times, shocks to oil demand and supply create a larger impact response on prices given a certain oil production change Conclusions This paper analyzes whether the impact of oil shocks di ers in times of high uncertainty compared to when macroeconomic volatility is low. As it is well documented that uncertainty can a ect the decision behavior of economic agents, it could equally impact on the strength at which shocks to oil fundamentals a ect oil prices, oil production and economic activity. Several important insight emerge from our analysis. First, a test for the signi cance of threshold e ects indicates that the model is non-linear and behaves di erently in regimes of higher uncertainty which are mostly associated with periods of slowing economic growth, recessions and nancial crises. Second, higher macroeconomic uncertainty causes oil prices to respond more strongly given a certain change in oil production following oil supply shocks, indicating that the elasticity of oil demand decreases when uncertainty is higher. This also appears to be the case for oil demand shocks driven by economic activity, although the di erence in estimated oil supply elasticities is somewhat less pronounced. The reduction in the oil price elasticity in the high uncertainty regime is statistically and economically signi cant, certainly for the elasticity of oil demand as the price e ect of an equal impact change in oil production could easily be twice as large when uncertainty is high. A third, possibly related nding is that the e ect of all types of oil shocks on economic activity is more aggressive in times when macroeconomic volatility is already high. Finally, macroeconomic uncertainty itself increases following oil shocks that are not driven by higher economic growth. These ndings are robust to variations in the speci cation of the model, the threshold test, identi cation of the shocks and the measure of uncertainty. As far as we are aware, this is the rst paper considering a role for macroeconomic uncertainty in explaining changes in the impact of oil shocks in general, and that endogenously explains variations in the elasticity of oil demand and supply 24 Singleton (2011) shows a high correlation between the dispersion of forecasts of oil prices by professionals surveyed by Consensus Economics, a proxy for heterogenous beliefs, and the oil price movements in This co-movement could also indicate that higher uncertainty leads to higher disagreement among forecasters, and in turn to stronger oil price uctuations. 18

19 over time. In the discussion on the driving factors behind the recent rollercoaster ride in oil prices, our ndings imply that the contribution of oil demand and supply shocks to the oil price could be larger than previously estimated, once the non-linearity of the price elasticity of oil demand and supply is taken into account. We leave the analysis of the channels of transmission through which higher macroeconomic uncertainty a ects the price elasticity of oil demand and supply as an interesting avenue for future research. 19