2.1. The effect of geopolitical tensions on the oil market

The analysis of geopolitical risks on the oil market, however, is not new in the literature. Recently, ^{Antonanakis et al. (2017}) analysed the impact of uncertainty shock (as measured with the ^{Caldara and Iacoviello (2018})’s, geopolitical risk index) on the stock-oil returns covariance. The results reveal that geopolitical risks-broadly defined, and not specifically referred to the oil market-triggers a negative effect on oil returns and volatility, and to a lesser extent to the stock (S&P500)-oil returns covariance. It is commonplace in the literature to proxy geopolitical tensions with a wide range of uncertainty indexes and, more scarcely, with oil-specific uncertainty measures. One exception of the latter kind of research is ^{Joo and Park (2017}). By making use of GARCH-in-mean specification, the authors find that the uncertainty of stocks (in the United States, Japan, South Korea, and Hong Kong) and oil returns carry significant negative time-varying effects of uncertainty over returns in sub-periods comprehended between 1995 and 2015.

^{Kang and Ratti (2013a}) claim that oil price shocks and EPU are interrelated and influence stock market returns in the United States. The authors argue that a positive oil-market-specific demand shock significantly raises EPU and reduces real stock returns. Also, ^{Bekiros et al. (2015}) find that EPU information does matter in predicting changes in oil prices. Moreover, ^{Kang and Ratti (2013b}) find that oil-specific demand shocks account for 31% of conditional variation in the EPU. Similarly, ^{Maghyereh et al. (2016}) make use of a series of implied volatility indexes in 11 major stock exchanges to investigate the directional connectedness between oil and equities between 2008 and 2015. The results support episodic bi-directional information spillovers largely dominated by the transmission from the oil to equity markets, and not the other way around. ^{Antonanakis et al. (2014}) also examine the dynamic relationship between changes in oil prices and the EPU over 1997 to 2013, finding a negative feedback relationship between oil price shocks and EPU shocks.

More related to geopolitical tensions and violent conflict, particularly terrorism and wars, ^{Kollias et al. (2013}) finds that wars have a significant negative effect on the covariance between oil price and returns of four big stock markets (S&P500, the European DAX, CAC40, and FTSE100). Interestingly, terrorist incidents have an impact in just two indexes; CAC40 and DAX. ^{Guidolin and La Ferrara (2010}) find that, especially in the Middle East, oil futures systematically exhibit a downturn in response to conflicts in this region, analysing 101 events with the case study methodology. Some other articles highlight how terrorism deteriorates economic sentiment (^{Drakos and Kallandranis, 2015}) and lower income per capita growth (^{Gaibulloev and Sandler, 2009}).

Thus, the literature associates oil price shocks as a trigger to general uncertainty, but this relationship is evolving across time and is dependent on one-off events such as terrorist attacks and violent conflicts. In this article we partially support this view by finding that oil prices cause tensions and news in OPEC-countries only. Moreover, it is a wider spectrum of geopolitical tensions and news that cause oil prices, it forecasts and the dispersion around those forecasts. So, it is likely that non-OPEC tensions and news increase global uncertainty (as well as stocks, economic sentiment, and income, among others) through oil prices rather than tensions and news coming from OPEC countries. This important distinction is possible to make thanks to the construction of our newly proposed, oil-specific measure of geopolitical tensions and news.

2.2. The economics of OPEC countries

The OPEC was established in Baghdad, Iraq, and effective as from January 1961. Since then, a lot of attention has been attracted to a particular OPEC conference scheduled twice a year whose outcome consists of a market quota setting for participant countries. There is much speculation in the days surrounding these conferences as it supposedly is the price setting mechanism managed by OPEC. A long-standing research in this matter possibly begins with ^{Griffin and Teece (1982}), ^{MacAvoy (1982}), and ^{Draper (1984}), when analysing the effect of this meeting outcome-decoded as an increase, no change, or decrease in quota-on oil-market-based securities. A similar purpose is followed in ^{Deaves and Krinsky (1992}), ^{Wirl and Kujundzic (2004}), ^{Guidi et al. (2006}), and ^{Hyndman (2008}) among others, as well as studies including other OPEC issues such as reserves (^{Taylor and van Doren, 2005}, and ^{Considine, 2006}). The results achieve certain consensus when quotas are reduced and oil prices are then increased, but this influence has declined since mid-1980s. This finding is in line with the evidence suggesting OPEC as a strong price setter during the 1970s.

OPEC’s effective power has been analysed thoroughly from an economic point of view by researchers and policy makers (^{Pindyck, 1978}; ^{Salant, 1976}; ^{Teece, 1982}; ^{Moran, 1982}; ^{Hochman and Zilberman, 2015}). Many diverse events have occurred since OPEC’s establishment-mainly wars and other political instability realisations-and there is no consensus about the role of OPEC as price setter after the 1980s (^{Loderer, 1985}; ^{Smith, 2005}; ^{Fattouh, 2005}). Most remarkably, ^{Almoguera et al. (2011}) suggest that the ability of OPEC to set prices since its creation is rather episodic. The authors find that during the period from 1974 until 2004, OPEC acted similar to a Cournot competition when sharing the global market with non-OPEC oil producers. Their empirical results, as the authors argue, are in favour of specific but non-time-robust price rises due to OPEC’s comparison to the price level under competition⁵.

The extent to which OPEC sets prices and the effects of non-market externalities in oil spot prices are questionable. It is also questionable if oil price forecasters are aware or affected by these externalities when making their predictions. This is important because major oil supply disruptions bring attached detrimental effects of large unexpected shocks affecting stock indices (^{Hammoudeh and Eleisa, 2004}; ^{Hammoudeh and Li, 2004}; ^{Pollet, 2005}; ^{Malik and Hammoudeh, 2007}; ^{Driesprong et al., 2008}; ^{Balcilar et al., 2015}) and even leading to recessions (^{Hamilton, 2003}, ²⁰⁰⁹). Oil prices also carry a substantial amount of information to other prices affecting global inflation (see ^{De Gregorio et al., 2007}, ^{Neely and Rapach, 2011}, and ^{Medel, 2015}, ²⁰¹⁶ for details).

Besides the impact on the level, comprehensive literature also analyses the impact of OPEC news on oil price volatility. Some examples are ^{Deaves and Krinsky (1992}), ^{Horan et al. (2004}), ^{Fattouh (2005}), ^{Lin and Tamvakis (2010}), ^{Aguiar-Conraria and Wen (2012)}, ^{Cairns and Calfucura (2012}), ^{Brémond et al. (2012}), ^{López and Muñoz (2012}), ^{Schmidbauer and Rösch (2012}), and ^{Mensi et al. (2014}) among others.

It is a less clear-cut if just OPEC-related news is the only driver of oil price shocks, or if it is necessary to include a wider spectrum of supply disruptions such as political instability, wars, or any news due to non-market externalities as well as news on alleviating oil supply. This is important because certain OPEC countries have been subject to substantial geopolitical risks and tensions not necessarily affecting the organisation’s members countries only. For this reason, a key feature of this article is considering OPEC as one of many other oil-market-based news-generator devices for both oil supply contractions and expansions.

3.1. Granger causality

The notion of Granger causality is as simple as it is useful, and different from ordinary causality. It states that if lagged values of a variable x_t predict current values of another variable y_t, and that forecast of y_t includes lags of x_t as well as y_t , then x_t Granger cause y_t (short notation: x_t → y_t). In this article, we make use of the ^{Hsiao (1981}) version of Granger causality, extending it to a joint significance F-test of a whole set of parameters associated with the independent variable (xt) that cause the dependent variable (y_t). Formally, this corresponds to testing if all the lags of x_t are jointly statistically different from zero in the following regression:

where lags of y_t, control for autocorrelation,

are parameters to be estimated with, say, ordinary least squares, assuming . The autoregressive orders (p_y p_x) in equation (1) vary from one to six lags to control for autocorrelation (p_y) and to extend the Granger causality hypothesis testing (p_x).

Statistical inference is carried out by testing the joint null hypothesis NH: Θ1 = ... = Θ_px = 0 (x_t do not Granger cause

). The vector that contains the restrictions is F-distributed with (p_x,T- (p_y + p_x + 1)) degrees of freedom (where T is the sample size). A formal treatment can be found in ^{Harvey (§8.7, 1990}), ^{Hamilton (§11.2, 1994}), and ^{Patterson (§8.5, 2000}).

Notice that this inference is possible to make only if the coefficients are unbiased. In order to check for this statistical requirement, we provide the results of the Breusch- Godfrey test for residuals’ autocorrelation (^{Breusch, 1978}; ^{Godfrey, 1978}). The suitability of this test relies on its ability to deal with nonstochastic regressors, such as the lagged values of the dependent variable, and testing higher-order autoregressive schemes, such as AR(p), with p > 1. The test is built-into a Lagrange multiplier test and proceeds as follows. The regression test assumes that residuals in equation (1) Ε_t follow a p-th autoregressive process and including the information of the independent variables of equation (1) (labelled as X _t, and Β is the vector of parameters of equation (1)):

where v_t is a white noise residual. The Breusch-Godfrey null hypothesis to be tested is that NH: Ϋ₁ = ... = Ρ_p, that is, there is no serial correlation up to order p. We perform the test by setting this /vorder up to six lags. Breusch (1978) and Godfrey (1978) have shown that (T - v)R² _Ε is chi-squared distributed with p degrees of freedom, where R² _Ε is the goodness-of-fit coefficient of equation (2). Thus, if the p-value exceeds a chosen level of significance, we do not reject the null hypothesis, meaning that all Ρ1 = ... = Ρ_p coefficients are zero, and no evidence of serial correlation is found.

3.2. Short-sample bias

Our analysis relies essentially on econometric estimations considering our geopolitical tensions and news variable which is available from 2001.1 to 2012.3 (135 observations) in monthly frequency. This fact could imply a short-sample bias and could invalidate the statistical inference obtained from coefficient estimates, i.e. the F-tests. Moreover, despite that all series used in estimations are stationary (as we test in subsection 3.4), they show a high level of persistence. Thus, this poses the risk of autocorrelated residuals; invalidating the inference based on F-tests.

To alleviate these problems, we make use of the ^{Newey and West (1987}) standard deviation estimator, which accounts for both heteroskedasticity and serial correlation. By setting key parameters -that will be shown below-, we will be able to use the Newey-West estimator to alleviate short-sample bias too. This is because the estimator corrects the off-diagonal elements of coefficients’ variance(-covariance) matrix as well as heteroskedasticity. Thus, this correction goes beyond the case where the variance matrix Ω is different from σ₂ I. The estimator consists in an extension to ^{White (1980})’s variance estimator when the problem is heteroskedasticity of a general unknown form.

The baseline ordinary least squares variance matrix corresponds to:

with Ω being of unknown form. To materialise how the Newey-West estimator operates, consider the case of T = 4 and k = 3 (a constant plus two variables). In this case, we have that:

where X _s = (1 X _s2 X _s3). The shape of equation (4) implies that σ_ts provide weights associated with observations that differs in t - s periods. If σ_ts = 0 for t ≠ s, there is no serial correlation of an order greater than “s”.

Define h ≡ t - s, so, s ≡ t - h and σ_t,h-t = σ_h-t,t, meaning that what matters for the correlation control is the time difference h. For example, if Ε_t are generated by a MA(2) process, all terms for which |h| > 2 must be zero. The Newey-West estimator operates here in three ways. First, σ_ts is replaced by

, where s = t-h and are the residuals obtained with ordinary least squares. Second, the issue of how many autocovariances to include is latent. To determine this bandwidth, assume that e_t follows an MA() process, and so, the autocovariances to include should not exceed . Considering the frequency of our series, we set our baseline estimates with a bandwidth of six in oil price series (to control for possible seasonality), and one for the geopolitical risk measures. Finally, the Newey-West estimator introduces the weights w_h on the products , with s = t-h, ensuring that the variance matrix is positive definite. These weights are calculated using the Bartlett window, and are of the shape w_h = 1 - [h/( + 1)] for h =,…,. Thus, the weights decline from

/( + 1) to 1/( + 1). With these three considerations, the Newey-West estimator of is:

and, thus, the Newey-West estimator of the variance matrix

of is:

and the estimator is said heteroskedasticity and autocorrelation consistent.

For robustness purposes, we also conduct the full exercise making use of the jackknife estimator of coefficients standard deviation. The jackknife first-order unbiased estimator serves primarily in cases where some observations could be influencing the overall statistic. The estimator repeatedly calculates the standard deviation each time omitting just one of the dataset’s observations. If y_¡ is the i-th observation of the data with i = 1, ..., T, the jackknife estimator of the standard deviation makes use of the mean:

where is the mean using the entire sample excluding the i-th observation. Thus, solving for y_i we have:

and more generally:

These are the pseudovalues that configure the jackknife estimator, corresponding to the mean of those pseudovalues with a standard deviation equivalent to the standard deviation calculated for

(^{Tukey, 1958}):

As is possible noticing, the jackknife estimator is a valid alternative dealing with heteroskedasticity and the over-representation that few observations could have in a short-sample environment; but not necessarily dealing better than the Newey-West estimator under serial correlation. For this reason, our baseline estimates are based on the Newey-West estimator, whereas jackknife-based results are available for robustness only.

3.3. An application to the oil market

We label our measure of geopolitical tensions and news as “GT&N’ which is constructed, as mentioned above, as the sum of 10 dummy variables related to the oil market. Adding specific 9 out of 10 non-OPEC related variables we generate the “GT&N-NO” variable, while the remaining dummy concerning purely OPEC is labelled as “GT&N-O” (thus, GT&N = GT&N-O + GT&N-NO); see (Table 1).

By means of Granger causality we provide evidence on the following hypotheses:

• H1: Do GT&N cause the Brent oil price (P^Oil): GT&N → P^Oil?

• H2: Do GT&N cause oil price forecasts

and

• H3: Do Gr&ZValfect the dispersion of oil price forecasts

In order to conclude about the reliability of the GT&N variable, it is expected that all these hypotheses must have statistical significance in the shown direction. At the same time, a unidirectional relationship is expected in H1, with GT&N causing P^Oil but not the other way around. This is merely to ensure that GT&N is exogenous and is actually measuring unexpected news.

If oil price expectations are orthogonal to oil producers' information set, it should follow that

. Also, greater tensions are associated with uncertainty about future oil prices. For that reason, it is expected that , and the inverse should not hold; again, if the GT&N is exogenous and measuring unexpected news⁶.

Our analysis involves oil price forecasts for two reasons. The first one is the true interest in investigating to which extent both forecast level and dispersion are affected by the GT&N variable. The second reason is to stress out the reliability of the newly proposed GT&N measure and its components.

The analysis continues by testing the same set of three hypotheses making use of the GT&N-O and GT&N-NO variables. Notice that given the geographical proximity of the majority of considered oil-producer countries, it is difficult to fully isolate both measures and some intertemporal interaction may exist in specific events. However, we do not impose an orthogonality condition between them, opting for preserving the benefit of simplicity and easy-to-read results.

3.4. Dataset

The analysis is made considering a time window spanning from 2001.1 until 2012.3 in monthly frequency, comprising 135 observations. Notice that the GT&N variable is available as from 1999. So, the limiting part of the analysis are the oil price forecasts, starting in 2001. The GT&N variable is constructed by adding 10 categorical dummy variables, in which the value of one is assigned to an unexpected event (geopolitical tension or news) associated to an oil supply expansion, minus one to an oil supply contraction, and zero otherwise. There is one category fully deserved for OPEC events while the remaining belong to non-OPEC countries.

A total of 204 events are identified across the 10 categories listed in (Table 1). More detailed descriptions on the type of events are included in each category that can be found in Appendix A as well as the time-series graph of the GT&N variable in Figure A1. For a daily individual-level identification, see Appendix A of López and Muñoz (2012). The sources of geopolitical tensions and news are Bloomberg, The Wall Street Journal, Financial Times, and the United States Energy Information Administration and are manually coded comprehensively according to informational content. The GT&N variable is not recoded to, for instance (-1,0,1) after adding its components, to preserve its intensity.

Actual oil price means to the annual percentage change of the Brent oil price measured in USD per barrel (source: Bloomberg;

100 x ((Oil price_t /Oil price_t-12) 1). Oil price forecasts corresponds to the annual percentage change of the 12 months ahead forecast contained in the monthly Consensus Forecast (CF) report, but using the actual value as denominator (e [p^oil] = 100 x (Oil price forecast_t/Oil price_t-12) - 1)). The point estimator displayed in CF report corresponds to the mean of the answers at the same horizon, ranging 65-70 respondents. Each report also shows the maximum and the minimum point value answered by respondents; and, respectively and is the forecast 12-months- ahead. Hence, the difference ,measures the dispersion or, in other words, the degree to which the consensus is achieved in forming oil price forecasts; the greater the uncertainty, the smaller the consensus achieved.

(Figure 1) exhibits the time series of actual oil prices, and CF expectations and dispersion. Notice that exogenous to all of these variables, including GT&N, there is a noticeable impact of the 2008-09 Global Financial Crisis initiated after the bankruptcy of Lehmann Brothers investment bank in the United States. As shown in (Figure A1), we notice a number of disturbances during 2001 (due to the 9/11 terrorist attacks), 2003 (Iraq War), mid-2005 (Lebanon War), and the 2011-12 period (Arab Spring). (Table 2) presents the descriptive statistics of all involved series using the transformation that achieves stationarity according to the Augmented Dickey-Fuller (ADF), Kwiatkowski, Phillips, Schmidt, and Shin (KPSS), and Phillips-Perron (PP) tests.

GT&N variable description

In this appendix, we provide extended descriptions of the 10 dummy variables used in the construction of the GT&N variable. A time series plot of the 10 variables is presented in (Figure A1).

1. United Nations Oil for Food Program (1995-2003) [+]. Programme developed by the United Nations established in 1995 as a response to Iraqi citizen’s claims affected by economic sanctions imposed in the aftermath of Gulf War of 1991. The programme allows Iraq to sell petroleum in world markets in exchange for food, medicines, and other humanitarian help, aiming to bind Iraqi military capacity. The programme finishes in 2003. The events referred to this programme are United Nations’ resolutions on Iraqi global oil market quotas, similar to the impact of new discoveries.

2. United States relations with Libya and Iran (1996-2004) [-]. Events considered in this category are related to the sanctions imposed on Iran and Libya promulgated in 1996. This act imposes economic sanctions on entrepreneurial-kind relations with Iran and Libya. The programme is a response to the nuclear agenda and support provided by Iran to certain terrorist associations (Hezbolla, Hammas, and Jihad). On 19 December, 2003, Libya announced its intention to leave the nuclear programme as well as the development of massive destruction weapons and the beginning of a new era of cooperation with the United States.

3. Iraq War and post-war period (2003-2011) [-]. News related to the United States’ invasion to Iraq in March 2003, and Saddam Hussein’s capture in December 2003. It also includes events related to the installation of the provisional government in Iraq and reestablishment of Iraq’s international affairs.

4. Iran post Iraq War (start in 2005) [-]. Accounts for events related to justified hearsays of the re-establishment of a nuclear programme during the administration of president Mahmoud Ahmadinejad starting in August 2005.

5. Terrorist attacks [-]. Constitutes events referred to terrorist attacks to productive installations in the Middle East, or terrorist targets. 9/11 attacks are included within this category.

6. Lebanon War (2006) [-]. Also referred as Israel-Hezbolla War o July War, is a 34-day-long conflict occurred in Lebanon spanning from 12 July to 14 August, 2006, after a ceasefire, statement of the United Nations. The conflict had a de facto end on 8 September, 2006 when Israel unblocked maritime restrictions over Lebanon.

7. Arab Spring (2011) [-]. Constitute waves of anti-government demonstrations and strikes in Arab countries starting on 18 December, 2010 in Tunisia. Governments of Tunisia, Egypt, Libya, and Yemen were overthrown. Civilian demonstrations took place in Bahrain and Syria; massive movement strikes in Algeria, Iraq, Jordan, Kuwait, Morocco, and Oman; minor events were noticed also in Lebanon, Mauritania, Saudi Arabia, Sudan, and Western Sahara.

8. Use of the United States Strategic Petroleum Reserve [+]. The Strategic Petroleum Reserve (SPR) is the world’s greatest for-emergency reserve of oil, whose capacity achieves more than 700 million of barrels. This variable accounts for the United States announcements on sales with stabilisation purposes or domestic emergencies. An in-depth and up-to-date analysis of the use of the SPR can be found in ^{Demirer and Kutan (2010}).

9. New announcements on discoveries, and site exploration [+]. News related to oilfield discoveries, explorations, drills, and strategic alliances between firms in order to exploit Middle East oilfields.

10. Purely OPEC announcements [+/-]. Announcements on OPEC’s quotas reassignment or major maintenance works. This variable by itself constitutes the GT&N-O measure. In contrast, the sum of the previous nine make up GT&N-NO.

Robustness results using the jackknife estimator

Baseline results of (Tables 3-5) suggest that all considered geopolitical tensions and news do affect the oil price, its forecasts and the dispersion around those forecasts. Moreover, non-OPEC news and tensions also cause the current oil prices as well as forecasts and its dispersion in immediate time. Finally, OPEC news and tensions are caused by oil prices and exhibit a feedback relationship with oil price forecasts’ dispersion. This relegates OPEC as a source of information for oil price expectations formation and are ultimately a wider range of geopolitical tensions and news affecting actual oil price.

In this appendix, we perform the same analysis making use of the jackknife coefficients’ standard deviation instead of the Newey-West estimator. The results using the GT&N variable are reported in (Table B1). The first panel indicates that the fifth lag of geopolitical tensions and news cause the oil price, supported by the rejection of serial correlation hypothesis (baseline results are significant from the third lag onwards). Similar to the baseline results, H1 Inverse indicates that the relationship is unidirectional from GT&N to the oil price. The second panel shows that geopolitical tensions and news cause oil price forecasts from the fifth lag to sixth, and there is not a feedback relationship between variables. The results are supported by non-autocorrelated residuals and are qualitatively similar to baseline results. In the same line, the third panel also reveals a feedback relationship between GT&N and oil price forecasts’ dispersion, meaning that tensions and news affect current oil prices and, at the same time, the dispersion triggers news and tensions on oil supplier countries.

The results using purely OPEC news are presented in (Table B2). The first panel provides similar results to baseline estimations, rejecting the causality of OPEC’s geopolitical tensions and news to oil prices, but supporting the causality the other way around-from actual oil price to OPEC tensions and news. A small twist compared to baseline results is found in the second panel in which, besides OPEC causality of oil price forecasts, the latter cause the former with the first lag, transforming the link between both variables into a feedback relationship. Notice that the first lag of the forecast series causing OPEC tensions and news is not necessarily invalidating when considering that the current price actually causes OPEC news and tensions. It is likely that, in a persistent series such as oil price, lags of actual variable determine its one-step-ahead forecast. Interestingly, the third panel suggests that OPEC tensions and news cause forecast dispersion with the third lag, and also the causality goes in the opposite direction. This result supports the core claim of this article, giving a secondary role to OPEC as price setter, but still being relevant for forecasters and expectations formation.

Finally, the results using non-OPEC geopolitical tensions and news are presented in Table B3. First and second panels are virtually the same to those of Table 5, with GT&N-NO causing both the oil price and its forecasts. Both relationships are unidirectional and supported by residuals’ non-autocorrelation, confirming that there are a broad range of events that affect oil prices, and not necessarily those OPEC-based only. The third panel establishes an independent, no-relationship between non-OPEC geopolitical tensions and news and oil price forecasts’ dispersion. The baseline results find that the first lag of GT&N-NO causes forecasts’ dispersion, which is now eroded, and the GT&N-NO role is relegated to affect level forecasts only. This discrepancy reflects the methodological difference between the estimators; thus, suggesting that a few observations (likely coinciding with those with more intensity) command the causality of GT&N-NO over the forecasts’ dispersion.

In sum, qualitative robustness results remain, but in some cases, the statistical inference comes out “weaker” than the baseline estimations. By “weaker” we mean finding fewer statistically significant cases when testing any proposed hypothesis, but still supporting the baseline conclusions.