key: cord-1002684-a5tse3xj
authors: Gharib, Cheima; Mefteh-Wali, Salma; Jabeur, Sami Ben
title: The bubble contagion effect of COVID-19 outbreak: Evidence from crude oil and gold markets
date: 2020-07-28
journal: Financ Res Lett
DOI: 10.1016/j.frl.2020.101703
sha: f8911d56532e408b2cb576350d45130bfa4a7f9c
doc_id: 1002684
cord_uid: a5tse3xj

This paper examines the causal relationship between crude oil and gold spot prices to assess how the economic impact of COVID-19 has affected them. We analyze West Texas Light crude oil (WTI) and gold prices from January 4, 2010, to May 4, 2020. We detect common periods of mild explosivity in WTI and gold markets. More importantly, we find a bilateral contagion effect of bubbles in oil and gold markets during the recent COVID-19 outbreak.

The price of WTI crude oil dropped into a negative level on April 2020. This severe collapse is a direct result of a fall in demand caused by the economic slowdown generated by Coronavirus pandemic and failed negotiations between Russia and OPEC (Organization of the Petroleum Exporting Countries) to reduce daily barrel production. The financial markets have reacted to this oil price crash and the global panic with large drops. The important uncertainty in markets all over the world has created a risk-averse environment that has driven investors toward safe-haven assets, notably gold. Previous studies such as Hillier et al. (2006) and Kaul and Sapp (2006) document that the correlations between gold and other assets are weak or negative. Baur and McDermott (2010) and Baur and Lucey (2010) show that gold serves as a hedge and a safe haven for the US, the UK, and German stocks or bonds. This property can be a priori seen in the recent evolution of gold price. Indeed, despite the mid-March decrease, the gold price has been on a new upward course since March 19, 2020.

As oil and gold are the most commonly traded commodities in the world and lead economic variables, their price movements have important implications for the world economy and the financial markets. It is therefore of crucial practical significance to analyze their co-movement and causality. Although there is a vast body of literature on the prices relationship between gold and oil (Bedoui et al., 2019; Ewing and Malik, 2013; Narayan et al., 2010; Soytas et al., 2009; Zhang and Wei, 2010) , the economic impact of the COVID-19 pandemic on the dependence structure between oil and gold prices has not yet been analyzed. Our study fills this gap in the literature and provides three contributions. First, we assess how the recent health and economic crisis has affected the relationship between oil and gold. This allows us to assess to which extent gold can act as a hedge, a safe haven and/or a diversifier against oil price movements during the period of COVID-19 outbreak. Second, our long analysis period (January 4, 2010 , to May 4, 2020 enables us to examine how the relationship between oil and gold has progressed and to analyze the dependence and risk contagion between oil and gold. Third, in addition to the previous empirical contributions, our study presents a third methodological novelty. Indeed, various techniques are applied to capture the co-movement of both commodities during the period of COVID-19 outbreak. We use rolling window of the linear correlation between gold and crude oil prices and find that the dependence between WTI and gold varies across time due to the presence of bubbles and crises. Therefore, we employ the Phillips and Shi (2018) technique to identify a common bubble for both commodities during the period of COVID-19 pandemic. Then, we employ the linear Granger (1969) and the non-parametric nonlinear version of the Granger non-causality test of Diks and Panchenko (2005, 2006) for causality analysis. To identify changes in the causal relationship, we complete our causality analysis via time-varying Granger causality tests of Thoma (1994) , Swanson (1998) and Shi et al. (2018) . Identifying changes in the causal relationship allows to detect and to date the causality between gold and oil markets. To the best of our knowledge there are no studies, using time-varying Granger causality tests, to identify changes in the causal relationship between gold and WTI.

Overall, our study of how gold price co-moves with the oil price fluctuations will contribute to our understanding both of how COVID-19 is impacting the price of the most important commodities; as well as how gold might act as a hedge or safe haven during this period.

The Phillips and Shi (2018) procedure is based on the recursive rolling window of Phillips, Shi and Yu (PSY; 2015 a, b) , which is more efficient particularly if multiple bubbles are present in the sample period.

The PSY test can be conducted for each observation of interest ranging from to 1. The recommended setting of √ with T the sample length. Suppose the observation of interest is . The PSY calculates the ADF statistic respectively from a backward expanding sample sequence. Let and be the start and end points of the regression sample. The ADF statistic calculated from this sample is . We fix the end point of all samples on the observation of interest such that and allow the start point to vary within the range [ ].

Under the null hypotheses of , estimate the following equation:

The PSY statistics are the ultimate values of all ADF statistics expressed as follows:

The exuberance date is assumed to be where the PSY test statistic first exceeds its critical valuethe first time it has stopped for this episode. Similarly, the collapse date is considered to be when the supremum test statistic consequently falls below its essential values-a second stopping period for this episode. Suppose there is only one episode of the sample originating from to . According to Phillips and Shi (2018) , estimated periods and termination dates are given by the equations (3) and (4):

where the quantile of the distribution of the of equation (2).

In this paper, we use three tests of time-varying Granger causality of Thoma (1994) , Swanson (1998) and Shi et al. (2018) . Thoma (1994) and Swanson (1998) suggested using forward expanding and rolling window Wald tests, respectively, to detect changes in causal relationships. In contrast, the test of Shi et al. (2018) is based on the recursive rolling window, or evolving procedures of PSY (2015 a, b). The process of the corrective bootstrap algorithm is detailed below:

Step 1: We estimate the VAR(1) model under the null hypothesis of no Granger causality: in the first step from y 1t (WTI return) to y 2t (gold return) and in the second step from y 2t (gold return) to y 1t (WTI return).

Step 2: For data size, we run a bootstrap sample calculated as follows:

where and are the residuals of the VAR(1) model.

Step 3: Calculate the statistic sequences for the forward, rolling and recursive evolution by bootstrapped series. The statistic sequences for each test is expressed as follows:

 Thoma (1994) test based on the forward procedure: Step 4: For B = 1,...,499, repeat Steps 2 and 3.

Step 5: The forward, rolling and recursive processes are now provided by the 95% percentile of the , , and , respectively.

We analyze the daily WTI and gold prices from January 4, 2010, to May 4, 2020. The nominal WTI price is provided by the U.S. Energy Information Administration (EIA) and the gold price by www.gold.org. Table 1 shows the statistical properties of the oil and gold market used in this study. The skewness indicates that oil prices are skewed negatively while gold prices are positively skewed. The kurtosis of returns is greater than 3, indicating that oil and gold returns are leptokurtic. Both the prices and returns have a not-normal distribution. As expected, the Jarque-Bera test rejects the null hypothesis for the Gaussian distribution at a significance level of 5%. The price series are nonstationary, as confirmed by Figure 1 . In contrast, the return series are stationary, indicating the presence of volatility clustering. As anticipated, the augmented Dickey-Fuller test rejects the null hypothesis of nonstationarity.

The ADF test results in Table 1 show that the crude oil and gold prices have the same order one I(1) process. Hence there may be a cointegration relationship or a long-term relationship between the two sequences. Based on modeling needs, to eliminate seasonal factors in the crude oil and gold prices, we use the first-differencing of log time series i.e., return series as in Ding et al. (2017) and Holmes et al. (2020) .

[Insert Table 1 around here] [Insert Figure 1 around here]

We use a 30-day window to estimate the linear correlation of Pearson. Specifically, we use the overlapping rolling window to move forward one-day observations. Figure 2 reports the overlapping rolling windows of the linear correlation between gold and WTI prices. The changing correlations between crude oil and gold price vary between a positive and a negative value. As can be seen, some strong changes in the dependence between WTI and gold vary across time due to the presence of bubbles and crises in WTI and gold markets. However, linear correlation may be inadequate to fully reflect the dependence between WTI and gold.

[Insert Figure 2 around here] 4. Results and discussion 4.1. Testing for crisis identification Figure 3 and Table 2 report the results of crisis identification with 95% critical values obtained by the bootstrap procedure of Phillips and Shi (2018) . We date a negative mildly explosive episode between late November 2014 and late March 2015 in WTI nominal price. Our results are in line with Fantazzini (2016), Su et al. (2017) and Zaho et al. (2020) . This negative financial bubble is explained by the excess capacity in the oil market (Baumeister and Kilian, 2016), the increased leverage of oil firms (Domanski et al., 2015; Tokic, 2015) and the increase in non-OPEC oil exports (Sharm and Escobari, 2018). Su et al. (2017) show that crude oil prices also react to non-fundamental factors (e.g., speculation, geopolitics, USD exchange rates). We also date two short-lived bubbles at the beginning of 2016 and 2018 in WTI nominal price.

Moreover, our empirical results indicate two short, mildly explosive episodes in 2020: the first between March 6 and April 1, 2020, and the second from 14 to 29 April 2020 in WTI. These falling oil prices are explained by a combination of supply and demand issues as well as uncertainty about the future. Indeed, COVID-19 outbreak has had a negative impact on the world economy (Yilmazkuday, 2020; Maijama'a et al., 2020; Aloui et al., 2020) and specifically on oil demand because of border closures. In addition, the Russia-Saudi Arabia oil price war of 2020 and the insufficient storage capacity also contributed to the oil price crash. As a consequence, on April 21, 2020, the price of WTI dropped below zero for the first time in recorded history (-$39).

In the gold market, we detect a positive, mildly explosive episode in July-August 2019 due to the collapse of the U.S. dollar. This can be explained by the safe-haven properties of gold. 1 Beckman et al. (2015) showed that gold serves as both a hedge and a safe haven in times of market stress or turmoil. As mentioned by Su et al. (2017) , when investors anticipate the depreciation of the U.S. dollar, they tend to move away from the dollar and buy commodities (i.e., crude oil or gold). Furthermore, we also detect two short, mildly explosive episodes in gold prices in 2020: from February 20 to 27, 2020, and from March 4 to 11, 2020. This larger increase in the price of gold in late February could be explained by the oil price fall and decline in global stock markets. Due to the global spread of COVID-19 and oil price collapse, stock markets all over the world have responded in terms of growing risks and decline of stock prices U.S. 2 , European and Asian stock market indices 3 reported their largest singleweek declines since the 2008 financial crisis. Recently, Corbet et al. (2020) show a positive and significant relationship between WTI and Chinese stock markets. Sharif et al. (2020) have found that oil prices were leading the US market at both low and high frequencies in the period from January, 21st 2020 to March, 30, 2020. Since gold is considered to be a safe asset, investors shift over to this investment. Traditionally, gold (Baur and Lucey, 2010; Baur and McDermott, 2010) , is considered as a safe-haven investment during times of financial turmoil.

[Insert Table 2 

For a complete analysis of WTI and gold price interactions, we use causality tests to test the WTI versus gold returns and vice versa. We apply the linear non-Granger test (1969) and the nonparametric-nonlinear version of the Granger non-causality test of Diks and Panchenko (2005, 2006) .

As shown in Table 3 , the results reveal the rejection a bi-directional causality between gold and WTI returns. Additionally, we observe a unidirectional nonlinear causality running only from the WTI returns to the gold returns at a significance level of 10%. More importantly, this rejection of causality may possibly be explained by the time-varying behavior on the dynamic correlations and the presence of explosive processes or bubbles in these markets.

For implementing the time-varying Granger causality tests, the minimum window size is f 0 = 0.2, which contains 519 observations. The critical values are obtained from a bootstrapping procedure with 499 replications. The empirical size is 5% and is controlled over a three-year period. Following Shi et al. (2018) , the three tests of time-varying Granger causality are run with the assumption of homoscedasticity/heteroscedasticity of the residual error term for the VAR(1).

[Insert Figure 4 around here] Figure 4 indicate that the test statistics of the forward causality of Thoma (1994) are always below their critical values over the whole sample period. Consequently, the null hypothesis of no Granger causality from oil price to gold over the whole sample period cannot be rejected. Under the homoscedastic assumption, and based on the rolling test of Swanson (1998) (panel (c) of Figure 4 ), we find an episode of causality running from oil to gold from May 24 to August 2, 2013. After this date, we detect multiple causality periods from oil to gold: from March 11 to 19, 2020; from March 23 to April 13, 2020; and the last from April 17 to 20, 2020. In contrast, under the heteroscedastic assumption (panel (d) of Figure 4 ), we cannot reject the null hypothesis of no Granger causality from oil price to gold over the whole sample period. This result highlights the danger of inattention of the homoscedasticity/heteroscedasticity in financial time series analysis. Under the homoscedastic assumption and with the consistent recursive evolving algorithm of Shi et al. [Insert Figure 5 around here] Figure 5 shows the time-varying Wald test statistics for causal effects running from gold to oil returns. The first interesting feature of the results is that the Thoma test (1994) based on the entire sample would suggest evidence of causality from gold to WTI for the episodes from March 5 to March 13, 2020, and from March 17 to May 4, 2020 (panel (a) of Figure 5 ). In contrast, under the heteroscedastic assumption (panel (b) of Figure 5 ), we paint only three episodes: from March 5 to 13, 2020; from March 20, 2020 to April 1, 2020; and the last episode from April 6 to 17, 2020.

By applying the rolling test of Swanson (1998) under the heteroscedastic assumption (panel (c) of Figure 5 ), we find a longer episode of causality running from gold to oil from September 11, 2019, to March 13, 2020. We also detect two short causality episodes from March 17 to April 17, 2020, and from April 22 to May 4, 2020. In contrast, under the heteroscedastic assumption (panel (d) of Figure 5 ), we paint two short periods from December 18, 2019, to January 2, 2020. After this date, we detect multiple shorter periods: from January 6 to 21, 2020, and from February 21 to 26, 2020. We also find three short episodes in March 2020 (March 6, from March 19 to 24, and from March 26 to 27). Under the homoscedastic assumption and the consistent recursive evolving algorithm of Shi et al. (2018) (panel (e) of Figure 5 ), we plot a longer episode of causality running from gold to oil during the period from July 31, 2019, to May 4, 2020. In contrast, under the heteroscedastic assumption (panel (f) of Figure 5) , we refer to a longer episode of causality from October 9, 2019, to March 13, 2020, and two shorter episodes from March 19 to 27, 2020, and April 9 to 17, 2020. Our results are in line with the findings of Shi et al. (2018) . The recursive evolving approach of Shi et al. (2018) offers the best finite sample performance, followed by the rolling window algorithm of Swanson (1998) .

In this study, we used the bootstrap technique of Phillips and Shi (2018) to identify the bubbles in the crude oil and gold markets from January 2010 to May 2020. Our results indicate that there are common bubbles in the WTI oil and gold markets in March 2020 and April 2020. The dates of the 2020 crash correspond to the global COVID-19 outbreak. To investigate the causal relationships between WTI oil and gold markets, we used three tests of time-varying Granger causality. We detect a bilateral contagion effect of bubbles in oil and gold markets during the COVID-19 pandemic.

Our findings have practical implications. They are instructive for policy-makers who have to make decisions on financial stability measures. Indeed, the identification of explosive behavior and bubbles is of great importance in light of the links between oil prices and the overall economic activity, including other commodities prices. They are also relevant for different market participants (predictions of price changes, portfolio diversification, crosshedging and cross-speculation). 

Notes: The first, second rows show daily WTI and gold prices and returns respectively from January 4, 2010, to May 4, 2020.

Notes: The rolling windows bivarite correlation between daily WTI and gold prices from January 4, 2010, to May 4, 2020 The solid line is the test statistic sequence. The blue discontinue line is the 5% critical value sequence. The shaded areas are the causality episode i.e. the periods where the test statistic exceeds its 95% boostraped critical value. The first, second and third rows show the sequences of test statistics obtained from the forward recursive test of Thoma (1994) , rolling window test of Swanson (1998) and recursive evolving test of Shi et al. (2018) respectively. The columns of Figure 4 and 5 refer to the homoscedasticity and heteroskedasticity assumption of the residual of the VAR(1) model. The solid line is the test statistic sequence. The blue discontinue line is the 5% critical value sequence. The shaded areas are the causality episode i.e. the periods where the test statistic exceeds its 95% boostraped critical value. The first, second and third rows show the sequences of test statistics obtained from the forward recursive test of Thoma (1994) , rolling window test of Swanson (1998) and recursive evolving test of Shi et al. (2018) respectively. The columns of Figure 4 and 5 refer to the homoscedasticity and heteroskedasticity assumption of the residual of the VAR(1) model.

Test for A Note on the Hiemstra-Jones Test for

TESTING FOR MULTIPLE BUBBLES: HISTORICAL EPISODES OF EXUBERANCE AND COLLAPSE IN THE S&P 500

TESTING FOR MULTIPLE BUBBLES: LIMIT THEORY OF REAL-TIME DETECTORS

Identifying price bubble periods in the energy sector

COVID-19 pandemic, oil prices, stock market, geopolitical risk and policy uncertainty nexus in the US economy: Fresh evidence from the wavelet-based approach

Change Detection and the Causal Impact of the Yield Curve

World oil prices, precious metal prices and macroeconomy in Turkey

When Will Occur the Crude Oil Bubbles?

Money and output viewed through a rolling window

Subsample instability and asymmetries in money-income causality

Mad March: how the stock market is being hit by COVID-19

The crude oil market and the gold market: Evidence for cointegration, causality and price discovery

Identifying bubbles and the contagion effect between oil and stock markets: New evidence from China