The only certainty is uncertainty: An analysis of the impact of COVID-19 uncertainty on regional stock markets.

Uncertainty surrounding COVID-19 is widespread. We investigate the timing and quantify the impact of COVID-19 related uncertainty on returns and volatility for regional market aggregates using ARCH/GARCH models. Drawing upon economic psychology, COVID-19 related uncertainty is measured by searches for information as reflected by Google Trends. Asian markets are more resilient than others. Latin American markets are most impacted in terms of returns and volatility. For most regions, there is evidence of an increasing impact of COVID-19 related uncertainty which dissipates as the crisis evolves. We confirm that Google Trends capture uncertainty by comparing this measure against alternative uncertainty measures.

Introduction

While several pandemics and serious disease outbreaks have occurred in the past, such as the Spanish flu (in 1918), SARS (2003), MERS (2012)1 and Ebola (2014), the novel coronavirus (COVID-19) outbreak of 2019-2020 ranks amongst the most severe and widespread with infections recorded in more than 200 countries (World Health Organisation (WHO), 2020). The emergence of COVID-19 has resulted in a global economic crisis coupled with severe stock market declines. Prior studies show that not only are financial markets negatively impacted by diseases and crises in general, but that the intensity and timing of impact differs across countries (Nippani and Washer, 2004; McTier et al., 2013; Bekaert et al., 2014). Ichev and Marinč (2018) report that Ebola outbreaks had a more significant impact on companies that had operations in, or were geographically nearer to, Ebola origins (in West Africa). Claessens et al. (2010) document that during the 2007-2008 financial crisis, countries with closer ties to the United States’ (US) financial system or with direct exposure to asset-backed securities were the first to be affected.

Research on the differential effects of COVID-19 across countries has identified varying intensities and timings. Liu et al. (2020) observe that Asian financial markets experienced an immediate downturn when the outbreak occurred. The impact on US and most European markets was delayed, occurring several days after outbreaks of the virus in South Korea and Italy,2 and was less severe. Similarly, Gormsen and Koijen (2020) show that only once COVID-19 spread to Italy, Iran and South Korea, did US and German stock markets decline sharply. Gunay (2020) reports a structural break in volatility for Chinese stock returns earlier (30 January 2020) than in other countries.3 Ru et al. (2020) find that market reactions to early COVID-19 outbreaks were more immediate and substantial in countries that suffered from SARS in 2003. Gerding et al. (2020) document that stock markets in countries with higher debt-to-gross domestic product ratios were more impacted.

Uncertainty surrounding COVID-19 is widespread, both with respect to the evolution of the disease itself and its economic impact (McKibbin and Fernando, 2020). Moreover, COVID-19 related uncertainty has impacted both returns and volatility in the US (Baig et al., 2020; Bretscher et al., 2020; Ramelli and Wagner, 2020) and internationally (Liu, 2020; Papadamou et al., 2020). However, no study has examined the differential impact of COVID-19 related uncertainty on regional markets around the world and the timing of these effects.

We quantify the differential impact of COVID-19 related uncertainty on returns and variance for six regional market aggregates using the ARCH/GARCH framework and structural change analysis. Drawing upon economic psychology, which proposes that individuals respond to uncertainty about specific events by searching more intensively for relevant information (Liemieux and Peterson, 2011; Dzielinski, 2012; Castelnuovo and Tran, 2017; Bontempi et al., 2019), we measure uncertainty using Google search trends data for terms related to COVID-19. We contribute to the burgeoning literature on the impact of COVID-19 on financial markets. To the best of our knowledge, this is the first study that investigates the relationship between uncertainty reflected by Google search trends and COVID-19 for regional market aggregates. We find that returns for all regions are negatively impacted by global COVID-19 uncertainty and that COVID-19 uncertainty has volatility triggering effects for all regions with the exception of Arab markets. Furthermore, we find that a number of regions show a weakening of the impact of COVID-19 uncertainty as the crisis evolved. We confirm that Google search trends are a proxy for uncertainty which drives returns and triggers volatility.

Data and Methodology

Our primary data sample spans the period between 1 January 2019 and 19 June 2020.4 For regional markets, the MSCI All Country (AC) Asia, AC Europe, Emerging Frontier Markets (EMF) Africa, Emerging Markets (EM) Latin America, North America and Arabian Markets indices are used. Returns are defined as logarithmic differences in index levels. Data is of a daily frequency and is stated according to MSCI's local currency methodology, representing performance unimpacted by foreign exchange rate movements. Table 1 reports descriptive statistics for return series.

Table 1

Descriptive statistics for returns on MSCI indices

Region	Asia	Europe	Africa	Latin America	North America	Arab Markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
Mean	0.0002	0.0002	-0.0001	-7.80E-05	0.0006	-0.0003
Median	0.0005	0.0011	0.0004	0.0001	0.0009	0.000000
Maximum	0.0525	0.0761	0.0614	0.0954	0.0911	0.0529
Minimum	-0.0503	-0.1193	-0.0925	-0.1238	-0.1282	-0.1631
Std. dev.	0.0101	0.0137	0.0153	0.0184	0.0176	0.0131
Kurtosis	8.7745	22.6102	13.0918	18.1083	18.1819	67.3530
Skewness	-0.3311	-2.0972	-1.4676	-1.6940	-1.1366	-5.5057
SW	0.9107***	0.7858***	0.8440***	0.7726***	0.7524***	0.6213***

Note: This table reports descriptive statistics for returns on the regional indices in our sample. Returns are defined as logarithmic differences in index levels. *** indicates statistical significance at the 1% level of significance. SW is the Shapiro-Wilk test statistic for normality.

Following an analysis of Google search trends, we identify nine COVID-19 related terms associated with high search volumes globally. These are: “coronavirus”, “COVID19”, “COVID 19”, “COVID”, “COVID-19”, “SARS-CoV-2”, “SARS-COV”, “severe acute respiratory syndrome-related coronavirus” and “severe acute respiratory syndrome.” We construct a search term index5 by combining search trends for the terms above. Individual index values are added and the sum is divided by nine. The highest value is adjusted to 100, with remaining values adjusted accordingly relative to this base. Index values are then differenced (Fig. 1A in the Appendix).

We apply the ARCH/GARCH framework to measure the impact of changes in search volumes on both returns and conditional variance, a proxy for risk (Brzeszczyński and Kutan, 2015). We begin with an ARCH(p) model and proceed to estimate a GARCH(p,q) model if the ARCH(p) specification exhibits residual heteroscedasticity. We also consider the IGARCH(p,q) model if the sum of the ARCH and GARCH parameters is unity or close to unity (Engle and Bollerslev, 1986).

Table 2 lists all specifications, where r is the return on index i at time t, ΔCV19I are first differences in the combined COVID-19 search term index – our measure of COVID-19 related uncertainty – and h is the conditional variance. Dum 0,1 is a shift dummy denoting the pre-COVID-19 and COVID-19 periods, defined as 1 January 2019 to 30 November 2019 and 1 December 2019 to 19 June 2020, respectively. A residual market factor, Rε, derived from returns on the MSCI AC World Market index, is included to address potential underspecification (Burmeister and McElroy, 1991). Additionally, a factor analytically derived factor set, , is incorporated into equation (1) to account for influences that may not be reflected by Rε. Factors comprising the factor analytic augmentation, accounting for both contemporaneous and lagged relationships, are derived from regional return series and are then adjusted for the impact of ΔCV19I and Rε (Szczygielski et al., 2020).6 For parsimony, only significant proxy factors are retained. Finally, autoregressive terms, r , of order identified from an analysis of a residual correlogram are included to address remaining autocorrelation if required. To identify periods for which the impact of ΔCV19I differs, breakpoints are identified using the Bai-Perron test (Carlson et al., 2000). If breakpoints are detected, the ΔCV19I variable, together with associated coefficients and shift dummies in equations (1) and (2a)/(2b)/(2c), is replaced with and , respectively, with Dum 0,1, taking on a value of one or zero otherwise for segment π between breakpoints. Equations are first estimated using maximum likelihood estimation. If residuals are non-normal, they are re-estimated using quasi-maximum likelihood estimation with Bollerslev-Wooldridge standard errors and covariance (Fan et al., 2014).

Table 2

Model specifications

Model	Specification
Mean:	r_i,t=α_i+β_i,ΔCV19IΔCV19I_tDum_0,1+β_i,IMRε_IM,t+∑k≥0kβ_i,kF_k,t+γ_ir_i,t−τ+ε_i,t	(1)
ARCH/GARCH:
ARCH(p)	h_i,t=ω_i+∑p≥1pα_iεi,t−p2+φ_i,ΔCV19IΔCV19I_tDum_0,1	(2a)
GARCH(p,q)	h_i,t=ω_i+∑p≥1pα_iεi,t−12+∑q≥1qβ_ih_i,t−q+φ_i,ΔCV19IΔCV19I_tDum_0,1	(2b)
IGARCH(p,q)	h_i,t=∑p≥1pα_iεi,t−p2+∑q≥1qβ_ih_i,t−q+φ_i,ΔCV19IΔCV19I_tDum_0,1	(2c)

Note: This table lists the specifications fitted in this study. The mean equation, equation (1), is specified in the “Mean” row. ARCH(p), GARCH(p,q) and IGARCH(p,q) specifications, equations (2a)/(2b)/(2c) respectively, follow after the “ARCH/GARCH” row.

Results

The impact of COVID-19 related uncertainty on regional markets

Panel A, Table 3 reports coefficients on ΔCV19I in the conditional mean (β ) and Panel B reports the impact of ΔCV19I on the conditional variance (φ).7 The results in Panel A, Table 3 indicate that returns for all regions are negatively and significantly impacted by ΔCV19I. The results in Panel B indicate that coefficients on ΔCV19I in the respective ARCH/GARCH models, φ, are positive and statistically significant for five regions. The negative impact of ΔCV19I on returns can be attributed to a combination of lower expected cash flows and heightened risk aversion. The adverse economic effects of COVID-19 uncertainty are likely to be associated with declining expected cash flows to firms (Ramelli and Wagner, 2020). In addition, heightened risk aversion attributable to uncertainty surrounding the pandemic means that investors will require a higher risk premium which is reflected in the forward looking discount rate (Andrei and Hasler, 2014; Cochrane, 2018; Smales, 2021). Together, lower expected cash flows and a higher discount rate translate into lower stock prices.8

Table 3

Results for specifications without breaks

Region	Asia	Europe	Africa	Latin America	North America	Arab markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
Panel A: Conditional mean (eq.(1))
Intercept	0.0001	0.0003	0.0002	0.0002	0.0004***	-0.0001
βi,ΔCV19I	-0.001814***(6th)	-0.003459***(2nd)	-0.00314***(4th)	-0.003625***(1st)	-0.003417***(3rd)	-0.001882***(5th)
βi,IM	0.5622***	0.9471***	0.6537***	0.9293***	1.1430***	0.4290***
Proxy factors:
βi1	0.0049***	0.0013**		-0.0012**		0.0021***
βi2	0.0061***	0.00876***	0.0229***			0.0111***
AR terms	-0.2639rt − 1***	-0.1128rt − 1**		-0.0747rt − 1*	-0.1791rt − 1***0.1306 rt − 5***	0.0136rt − 5***
Panel B: Conditional variance (eq.(2a)/(2b)/(2c))
Model	IGARCH(1,1)	GARCH(1,1)	GARCH(1,1)	IGARCH(1,1)	GARCH(1,2)	GARCH(1,2)
ωi		4.10E-07*	1.11E-06*		3.25E-07***	6.55E-06*
αi	0.0171**	0.1426***	0.1125***	0.0238**	0.2470***	0.2842*
β1	0.9829***	0.8376***	0.8640***	0.9762***	0.4637*	0.0120
β2					0.2618	0.6548⁎⁎⁎
φi,ΔCV19I	0.1300***(5th)	0.1460***(4th)	0.2680***(2nd)	0.5480**(1st)	0.0599(6th)	0.1720(3rd)
Panel C: Diagnostics
R¯^2	0.6907	0.8491	0.7177	0.6983	0.9404	0.4169
F-statistic	144.5589***	291.9546***	404.4670***	61.7861***	1584.791***	15.5807***
Q(1)	0.0013	1.3880	2.6949	0.1089	1.4571	1.1509
Q(10)	10.615	9.2321	12.852	11.598	8.5414	11.417
ARCH(1)	1.5753	0.0085	0.7341	0.0227	1.2751	0.0497
ARCH(10)	0.4548	0.5198	0.5901	0.9616	0.6879	1.0301
Log-likelihood	1484.054	1597.812	1378.156	1276.960	5384.058	1320.595

Note: This table reports the impact of changes in COVID-19 related uncertainty on the returns (β) and variance (φ) for regional markets. Coefficients on ΔCV19I in the conditional variance equation are scaled by 100 000. Panel A reports estimation results for the conditional mean, which also includes proxy factors derived from regional returns using factor analysis and adjusted for the impact of ΔCV19I and Rε. Panel B reports results for the conditional variance. Values in brackets (…) rank the order of absolute impact according to the magnitude of the β and φ coefficients. Panel C reports model diagnostics, with Q(1) and Q(10) being Ljung-Box tests statistics for joint residual serial correlation at the 1st and 10th orders. ARCH(1) and ARCH(10) are test statistics for the ARCH LM test for heteroscedasticity. Each model is estimated over the primary data period between 1 January 2019 and 19 June 2020 unless residuals show dependence structures in which case longer estimation periods are used. Pre-COVID-19 and COVID-19 periods are defined as 1 January 2019 to 30 November 2019 and 1 December 2019 to 19 June 2020, respectively. The asterisks ***,** and * indicate statistical significance at 1%, 5% and 10% levels of significance, respectively.

Although returns in North America are negatively impacted (β of -0.003417 (3rd)), this region does not show significant volatility triggering effects. However, the results in Panel B, Table 5 present a different picture suggesting that North American markets experienced delayed volatility triggering effects. Similarly, while returns in Europe are also impacted (β of -0.003459 (2nd)), volatility triggering effects are relatively low (φ of 0.1460 (4th)). Arab markets do not appear to experience heightened volatility associated with ΔCV19I, although returns are impacted (β of -0.00188 (5th)). The lack of volatility triggering effects is surprising, given the economic dependence on oil of Arab markets and the consequent uncertainty surrounding their macroeconomic prospects (Ashraf, 2020). However, an analysis of realised variance suggests that Arab markets showed extreme, but short-lived, heightened volatility around 7 to 9 March 2020. These dates coincide with COVID-19 cases surpassing 100 000 and a call by the WHO for more stringent actions to control the spread of COVID-19 (WHO, 2020). While φ may not be significant, forecasted conditional variance captures this volatility spike (Fig. 7A in the Appendix).

Table 5

Results for specifications with breaks

Region	Asia	Europe	Africa	Latin America	North America	Arab markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
Panel A: Conditional mean (eq.(1)) with breaks
Breakpoints	18/05/2020	24/02/2020, 26/03/2020	No breaks	26/02/2020, 26/03/2020, 13/05/2020	20/01/2020, 24/02/2020, 26/03/2020	No breaks
Intercept	0.0002	0.0003*		0.0003	0.0004***
βi,1,ΔCV19I	-0.001865***	-0.003399***		-0.002338***	-0.001972***
βi,2,ΔCV19I	-0.001504***	-0.003282***		-0.005448***	-0.003341***
βi,3,ΔCV19I		-0.003448***		-0.003980***	-0.003697***
βi,4,ΔCV19I				-0.002243***	-0.003258***
βiM	0.5581***	0.9515***		0.8811***	1.1405***
Proxy factors:
βi1	0.0049***	0.0011*		-0.0014**
βi2	0.0062***	0.0088***
AR terms	-0.2653rt − 1***	-0.1107rt − 1**		-0.0970rt − 1**	-0.1778rt − 1***
					0.0157rt − 4
					0.0062rt − 4
Panel B: Conditional variance (eq.(2a)/(2b)/(2c)) with breaks
Model	IGARCH(1,1)	GARCH(1,1)		IGARCH(1,1)	GARCH(1,2)
ωi		1.11E-06***			3.50E-07***
αi	0.0322***	0.0558*		0.0410***	0.2251***
β1	0.9678***	0.8280***		0.9590***	0.4951*
β2					0.2322
φi,1,ΔCV19It	0.1740***	0.1650***		0.6000	-0.0288
φi,2,ΔCV19It	-0.1920	0.9460**		0.6190**	0.0728**
φi,3,ΔCV19It		-0.2790		0.7780**	0.2250*
φiΔ,4,CV19It				-0.3750	-0.0519
Panel C: Diagnostics
R¯^2	0.6909	0.8450		0.710	0.9400
F-statistic	108.5422***	468.0253***		44.0270***	1054.943***
Q(1)	0.0433	1.6874		0.1303	1.6350
Q(10)	9.2844	9.5584		10.927	8.7518
ARCH(1)	0.7557	0.6367		0.0232	2.3943
ARCH(10)	0.3983	0.8531		0.3965	0.6533
Log-likelihood	1484.948	1610.274		1280.500	5395.083

Note: This table reports the impact of changes in COVID-19 related uncertainty on the returns (β) and variance (φ) for regional markets, taking into account structural breaks. Segments are identified using the Bai-Perron test of L+1 vs L sequentially determined breaks with robust standard errors (HAC) and heterogenous error distributions. Coefficients on ΔCV19I in the conditional variance equation are scaled by 100 000. Panel A reports estimation results for the conditional mean, which also includes proxy factors derived from regional returns using factor analysis and adjusted for the impact of ΔCV19I and Rε. Panel B reports the results for the conditional variance. Panel C reports model diagnostics, with Q(1) and Q(10) being Ljung-Box tests statistics for joint residual serial correlation at the 1st and 10th orders. ARCH(1) and ARCH(10) are test statistics for the ARCH LM test for heteroscedasticity. Breakpoint identifies the date on which each structural change occurs during the COVID-19 period, where the beginning of the COVID-19 period is taken as 1 December 2019. Each model is estimated over the primary data period between 1 January 2019 and 19 June 2020 unless residuals show dependence structures in which case longer estimation periods are used. Pre-COVID-19 and COVID-19 periods are defined as 1 January 2019 to 30 November 2019 and 1 December 2019 to 19 June 2020, respectively. Asterisks ***,** and * indicate statistical significance at 1%, 5% and 10% levels of significance, respectively.

Asian markets are relatively resilient to COVID-19 uncertainty (β of -0.001814 (6th) and φ of 0.1300 (5th), respectively). This may be attributable to experience that Asian countries have in dealing with pandemics (SARS and MERS outbreaks) (Lu et al., 2020; Wang and Enilov, 2020). While these results differ from those of Liu et al. (2020) and Ru et al, who report that Asian markets were severely impacted by COVID-19 infection numbers, this finding demonstrates the varying effect of COVID-19 uncertainty relative to infection numbers on stock markets.

Finally, the substantial impact of ΔCV19I on returns and volatility in African and Latin American markets (β parameters of -0.00314 (4th) and -0.003625 (1st) and φ parameters of 0.2680 (2rd) and 0.5480 (1st), respectively) can be attributed to risk aversion in relation to developing markets in times of crisis and spillovers from developed markets (Frank and Hesse, 2009; Bekaert et al., 2014). Both regions comprise two of the larger and more developed stock markets in the world, the Johannesburg Stock Exchange (JSE) (19th) and the Brazilian BM&F Bovespa (20th) (Haqqi, 2020), which are highly integrated with global markets (Nashier, 2015; Babu et al., 2016) and therefore more likely to readily reflect global developments (Szczygielski and Chipeta, 2015).9 In contrast, Arab markets, while comprising developing countries, have been found to be less globally integrated (Marashdeh and Shrestha, 2010; Alotaibi and Mishra, 2017), which is consistent with our findings that they are less impacted by COVID-19 related uncertainty. Our results are generally in line with previous studies on the differential impact of pandemics and crises on different regions (Claessens et al., 2010; Bekaert et al., 2014).

As ΔCV19I is constructed from global Google search trends, we also consider value-weighted regional versions by replacing ΔCV19I with ΔCV19R 10 in the equations in Table 2 as an extension and robustness test. Results in Table 4 show a similar pattern. Returns for all regions, with the exception of Arab markets,11 are impacted negatively although to a lesser magnitude. For example, coefficients on ΔCV19R for Latin and North America decrease to -0.000876 and -0.002296, respectively. The order of the magnitude of impact is approximately the same across regions although North American and Arab markets are now most and least impacted, respectively. We attribute this effect to the dominance of US uncertainty. Specifically, uncertainty experienced by the US dominates North American markets and also impacts all other regions (Chiang et al., 2015; Dimic et al., 2016; Smales, 2019) and hence with regional measures, US uncertainty is excluded resulting in a reduced impact. Volatility triggering effects associated with ΔCV19R are also lower, with the exception of Latin America where ΔCV19R is now associated with a coefficient of 1.0700 in the conditional variance. The generally greater impact of ΔCV19I on both returns and conditional variance indicates that regional markets likely reflect not only regional uncertainty but also spillovers from the rest of the world (see discussion that follows). Importantly, it appears that global COVID-19 related uncertainty, as opposed to region or country-specific uncertainty, primarily matters most for stock markets and volatility (see Mumtaz and Mussom, 2019).12 This is broadly consistent with the findings of Costola et al. (2020) that US, German, French, Spanish and the United Kingdom stock markets responded more to Italian Google search trends than those for their own countries. Smales (2021) also finds that global search trends had a greater impact than regional search trends on the G20 stock markets. We conclude that, overall, the results of the analysis using ΔCV19R are mostly qualitatively consistent with those for ΔCV19I.13

Table 4

Abridged regional results for specifications without breaks

Region	Asia	Europe	Africa	Latin America	North America	Arab markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
βi,ΔCV19R	-0.00078***(5th)	-0.000910***(2nd)	-0.000789***(4th)	-0.000876***(3rd)	-0.002296***(1st)	0.000158**(6th)
φi,ΔCV19R	0.0947**(4th)	0.1010***(3rd)	0.2970**(2nd)	1.0700***(1st)	0.0421(5th)	-0.0076(6th)

Note: This table reports the abridged results for the impact of changes in regional COVID-19 related uncertainty as captured by Google search trends on the returns (β) and variance (φ) for regional markets. Coefficients on ΔCV19R in the conditional variance equation are scaled by 100 000. The asterisks ***,** and * indicate statistical significance at 1%, 5% and 10% levels of significance, respectively. Fig. 8A in the Appendix presents a comparison of global and regional search term indices. Unabridged results are reported in Table 1A in the Appendix.

Table 5 reports results after accounting for breakpoints. Results in Panel A suggest that the negative impact of ΔCV19I on returns first intensified and then weakened as the COVID-19 crisis evolved, although all regions continued to be significantly impacted. No structural breaks were detected for African and Arab markets. For North America, Europe, Latin America and Asia, the results in Panel B indicate that the negative impact of ΔCV19I on volatility intensified and then weakened as the crisis evolved. The dates of breakpoints across European, North American and Latin American markets are similar, with all three experiencing breaks in late February 202014 and in late March 2020 (26 March 2020 for all three). Breakpoints in late February 2020 coincide with President Trump's request for $1.25 billion from the US Congress to respond to the COVID-19 crisis (24 February 2020) and the first reported case in Latin America (Brazil) (26 February 2020) (Onali, 2020; Taylor, 2020). Gunay (2020) also identified a structural break in volatility in North America and Europe in late February 2020. The structural break on 26 March 2020 coincides with most European, North American and Latin American countries having imposed lockdowns and restrictions and the US becoming the country most impacted by the pandemic (Taylor, 2020). We also identify a breakpoint for North America in January 2020 (20 January 2020)15 and one for Latin America in mid-May 2020 (13 May 2020).

Returns in North America are most impacted (β of -0.003697) after late February 2020 whereas returns in Europe are most impacted following the end of March 2020 (β of -0.003448). For both North America and Europe, the impact of ΔCV19I on volatility is greatest following the February 2020 breakpoint (φ of 0.2250 and φ of 0.9460, respectively), but the impact of uncertainty on volatility dissipates thereafter (and is insignificant). The delay in impact mirrors the findings of Gormsen and Koijen (2020) and Liu et al. (2020) on the effect of COVID-19 infections on markets outside of Asia and is consistent with Ichev and Marinč’s (2018) assertions that geographical proximity matters. It is only when these two regions become centres of the outbreak that volatility (and to a lesser extent returns) is most impacted in these markets. For returns in Latin America, the initial impact is less severe but more than doubles (β Δ of -0.002338 to β of -0.0054) after the end of February 2020 before declining progressively (β of -0.003980 and β of -0.002243, respectively). A similar pattern emerges with ΔCV19I triggering heightened volatility after the end of February 2020 and further after late March 2020 (significant φ and φ of 0.6190 and 0.7780, respectively) before dissipating after mid-May 2020. The dissipating effect of uncertainty on volatility thus occurs later in Latin American markets than in North American or European markets. The weakening impact of ΔCV19I on volatility can potentially be attributed to the COVID-19 crisis being viewed by economic agents as a no longer novel but persistent situation. The decline in uncertainty reflected in Fig. 1 can also mean that a higher risk premium is no longer needed as risk aversion has dissipated or decreased substantially and/or that the decline in expected cash flows due to the pandemic is not as severe as initially predicted by the markets. Alternatively, this decline may be attributable to government responses to the pandemic, such as lockdowns and/or economic stimulus packages. A role for government interventions in reducing uncertainty and volatility is suggested by Kizys et al. (2020) but not by Zaremba et al. (2020). The latter - the role of government interventions - is investigated further in Section 3.2.

Fig. 1

COVID-19 related searches over time as captured by Google search trends.

Note: This figure plots levels of the combined COVID-19 search term index created from Google search trends volumes for nine COVID-19 related search terms over the period 1 December 2019 to 19 June 2020. Levels of search volumes for individual COVID-19 related terms are also plotted.

For Asia, the effects of ΔCV19I are immediate. The respective parameters (β Δ 19 of -0.001865 and φ of 0.1740) are largest and statistically significant prior to the first breakpoint on 18 May 2020. These findings are in line with those of Liu et al. (2020) and Ru et al. (2020) regarding the timing of the impact of COVID-19 infections on Asian markets. The effects on volatility in Asia dissipate similarly to North America, Europe and Latin America but, consistent with Latin America, this occurs later than in North America and Europe (the timing of the single breakpoint for Asia is similar to that of the final break for Latin America in May 2020). A finding of no structural breaks for African markets implies that the impact of COVID-19 uncertainty is still high. For African markets, this is potentially attributable to the pandemic still being far from its peak (WHO, 2020). For Arab markets, this may reflect a return to persistently lower levels of volatility following a large but short-lived volatility spike in early March 2020.

COVID-19 related uncertainty as a factor

To confirm that ΔCV19I is indeed driving returns, we factor analyse the structure of returns during the pre-COVID-19 and COVID-19 periods. For both periods, a single factor is extracted. The higher mean communality for the COVID-19 period suggests that the extracted factor explains a greater proportion of shared variance. The higher Kaiser-Meyer-Olkin (KMO) statistic also suggests that a greater proportion of shared variance is attributable to underlying factors. Both measures point towards strengthened dependence, likely attributable to the global impact of COVID-19 (Uddin et al., 2020). Spearman correlation between factor scores and ΔCV19I is highly significant with a coefficient of -0.3240 (ordinary ρ = -0.5619). This implies that ΔCV19I is indeed part of a composite factor set driving regional returns over this period. Fig. 2 shows that the rolling correlation between factor scores summarising the drivers of returns and ΔCV19I during the COVID-19 pandemic grows steadily in magnitude from early February 2020, peaking between mid-March 2020 and late April 2020, and decreases thereafter. These increases (decreases) correspond to a growing (decreasing) negative impact on returns and higher (lower) periods of volatility attributable to ΔCV19I, notably for Europe, Latin America and North America as identified using structural break analysis.

Fig. 2

Rolling correlations between ΔCV19I and factor scores.

Note: This figure plots rolling ordinary and Spearman's correlations between factors scores and ΔCV19I on an inverted vertical axis. Factor scores are estimated for the period 1 November 2019 and 19 June 2020 and are reported for the COVID-19 period 1 December 2019 and 19 June 2020 using rolling windows of 30 observations.

To confirm that ΔCV19I reflects uncertainty during the pandemic, we compare our measure against two other measures over the COVID-19 period. The first is the Chicago Board Options Exchange (CBOE) Volatility index (VIX), which we treat as a measure of stock market uncertainty (Bekaert et al., 2013). Although this is the US version of the index, Smales (2019) shows that the VIX captures global market uncertainty. Chiang et al. (2015) and Dimic et al. (2016) also utilise the US version of this index as a measure of global market uncertainty. The second is the recently developed Twitter-based Market Uncertainty (TMU) index of Renault et al. (2020).

Fig. 3 shows that COVID-19 search term index levels move closely with the two alternative measures of market uncertainty, although with somewhat of a lag especially between the end of January 2020 and the end of the sample period. Furthermore, changes in both measures become highly correlated with ΔCV19I between the end of January 2020 and the end of April 2020, implying that both reflect ΔCV19I during this period (see Fig. 9A and 10A in the Appendix).

Fig. 3

Comparison of COVID-19 search term index, VIX and TMU index levels.

Note: This figure plots levels of the combined COVID-19 search term index created from Google search trends data for nine COVID-19 related search terms over the period 1 December 2019 to 19 June 2020 against levels of the TMU index and the VIX. The TMU index has been exponentially smoothed for illustrative purposes.

Given that these two measures appear to reflect COVID-19 related uncertainty over the COVID-19 period, we re-estimate the specifications in Table 2, replacing ΔCV19I with ΔVIX and ΔTMU. Panel A and Panel B of Table 7 show that ΔVIX and ΔTMU have a similar impact on returns and volatility over the COVID-19 period to that of ΔCV19I. Both measures impact returns negatively across all regions. ΔVIX is associated with significant volatility triggering effects across half of the regions, with the exception of European, North American and Arab markets, as in Table 3 for the latter two regions. ΔTMU triggers volatility in all regions except Arab markets. Returns on Latin American markets are now second most impacted after North American markets, whereas returns on Asian and Arab markets remain least impacted. As in Table 3, North American markets experience the lowest volatility triggering effects in response to both alternative measures although they respond significantly to ΔTMU. Conversely, Latin American markets continue to be significantly and highly impacted. Overall, our results are largely consistent with those presented in Table 3, providing support for the role of ΔCV19I as a measure of uncertainty during the COVID-19 period.

Table 7

Abridged results for specifications with alternative measures

Region	Asia	Europe	Africa	Latin America	North America	Arab markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
Panel A: Abridged specifications with ΔVIXt
βi,ΔVIX	-0.000806***(6th)	-0.002420***(3rd)	-0.002083***(4th)	-0.003270***(2nd)	-0.003911***(1st)	-0.000960**(5th)
φi,ΔVIX	0.1850**(4th)	0.1560(5th)	0.3050**(3rd)	0.7670**(2nd)	0.0180(6th)	0.9810(1st)
Panel B: Abridged specifications with ΔTMUt
βi,ΔTMU	-0.000920***(6th)	-0.002454**(3rd)	-0.001760***(4th)	-0.002774***(2nd)	-0.002828***(1st)	-0.001367***(5th)
φi,ΔTMU	0.3400***(2nd)	0.1850***(5th)	0.2300**(4th)	0.8630***(1st)	0.0489***(6th)	0.3000(3rd)

Notes: This table reports the abridged results for the impact of changes in VIX and the TMU index on the returns (β, β) and variance (φ, φ) for regional markets. Coefficients on ΔVIX and ΔTMU in the conditional variance equation are scaled by 100 000. Values in brackets (…) rank the order of absolute impact according to the magnitude of coefficients on ΔVIX and ΔTMU. The asterisks ***,** and * indicate statistical significance at 1%, 5% and 10% levels of significance, respectively. Unabridged results are reported in Tables 4A and 5A in the Appendix.

Given that ΔCV19I shows a dissipating impact on returns and volatility in Table 5 and that Fig. 2 suggests that the importance of ΔCV19I diminishes, we set out to determine whether this effect can be attributed to government responses during the COVID-19 crisis. We first construct a response measure, ΔRESP, using the Oxford COVID-19 Government Response Tracker database16 and then test model specifications by incorporating ΔRESP in place of ΔCV19I in Table 2 after adjusting returns for the impact of ΔCV19I.

Results in Table 8 show that returns for most regions, with the exception of Asia, respond negatively to government responses to the pandemic. While this measure also reflects economic support measures, it may be that containment measures (lockdowns and restrictions) dominate. This would explain a negative relationship. Four regions are significantly and negatively impacted with North American and Arab markets the most and least impacted, respectively. Moreover, response measures are associated with significant volatility triggering effects in four regions, namely Asia, Latin America, Europe and Arab markets, which show the greatest response by far. A potential reason for the positive impact is that these measures were implemented around the time of, and in response to, significant COVID-19 related events which also had an adverse impact on stock markets and volatility, and therefore responses are a proxy for the immediate impact of these events.17 These findings are in line with those of Zaremba et al. (2020) who find that stringent policy responses tend to increase return volatility in international markets. We therefore propose that the lessening importance of ΔCV19I in Table 5 is attributable to a normalisation of economic agents’ expectations.

Table 8

Abridged results for the impact of government responses

Region	Asia	Europe	Africa	Latin America	North America	Arab markets
Index	MSCI AC Asia	MSCI AC Europe	MSCI EFM Africa	MSCI EM Latin America	MSCI North America	MSCI Arabian Markets
βi,ΔRESP	0.000487**(6th)	-0.001210***(4th)	-0.001282**(3rd)	-0.003618***(1st)	-0.003133***(2nd)	-0.000449(5th)
φi,ΔRESP	0.1070***(6th)	0.2360*(4th)	0.3120(3rd)	0.3630**(2nd)	0.1540(5th)	1.9600***(1st)

Notes: This table reports the abridged results for the impact of changes in government responses to the COVID-19 crisis on the returns (β) and variance (φ) for regional markets. Coefficients for ΔRESP in the conditional variance equation are scaled by 100 000. Values in brackets (…) rank the order of absolute impact according to the magnitude of coefficients on ΔRESP. The asterisks ***,** and * indicate statistical significance at 1%, 5% and 10% levels of significance, respectively. Unabridged results are reported in Table 6A in the Appendix. Regional returns are adjusted for ΔCV19I.

Finally, we present variance forecasts obtained from ARCH/GARCH specifications against realised variance for the COVID-19 period. Plots in Figs. 2A to 7A in the Appendix18 show that our forecasts approximate changing volatility dynamics and that the increases (decreases) in variance coincide with increases (decreases) in search volumes (see Fig. 1A in the Appendix).

Conclusion

Using the ARCH/GARCH framework, we demonstrate that COVID-19 uncertainty has impacted almost all global regions, resulting in lower returns and increased volatility. Asian markets appear to be more resilient to COVID-19 related uncertainty, while European, North and Latin American markets experience a weakening of the impact of COVID-19 related uncertainty over time. The evidence of a differential impact of COVID-19 across time and regions paves the way for further research into the reasons why such effects exist and why they dissipate over time. We confirm that our measure of COVID-19 related uncertainty reflects uncertainty by showing that it moves closely with alternative measures of uncertainty during the COVID-19 period. These measures, namely the VIX and TMU index, have a similar impact on returns and volatility over the COVID-19 period. Our results, together with the analysis of the structure of the return generating process, show that COVID-19 uncertainty is part of the factor set driving regional returns, although its role has diminished substantially.

CRediT authorship contribution statement

Jan Jakub Szczygielski: Methodology, Conceptualization, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Validation, Supervision, Project administration. Princess Rutendo Bwanya: Investigation, Writing – original draft, Writing – review & editing, Project administration. Ailie Charteris: Investigation, Writing – original draft, Writing – review & editing. Janusz Brzeszczyński: Conceptualization, Writing – review & editing.

Table 6

Summary of factor analysis

Period	Factors extracted	Mean communality	KMO
Pre-COVID-19	1	0.3834	0.7177
COVID-19	1	0.6505	0.8650

Notes: This table reports the results of factor analysis applied to returns over the pre-COVID-19 and COVID-19 periods. Pre-COVID-19 and COVID-19 periods are defined as 1 January 2019 to 30 November 2019 and 1 December 2019 to 19 June 2020, respectively. The number of factors extracted for each period are reported in the second column. Mean communality is the mean proportion of common variance explained by common factors across the return series extracted on the basis of the minimum average partial (MAP) test. KMO is the Kaiser-Meyer-Olkin (KMO) index which indicates suitability for factor analysis; values of over 0.8 are deemed desirable for factor analysis although values above 0.6 are acceptable.