Effects of the first wave of COVID-19 pandemic on implied stock market volatility: International evidence using a google trend measure

Abstract

This paper investigates the relationship between investors' attention, as measured by Google search queries, and equity implied volatility during the COVID-19 outbreak. Recent studies show that search investors' behavior data is an extremely abundant repository of predictive data, and investor-limited attention increases when the uncertainty level is high. Our study using data from thirteen countries across the globe during the first wave of the COVID-19 pandemic (January–April 2020) examines whether the search “topic and terms” for the pandemic affect market participants’ expectations about future realized volatility. With the panic and uncertainty about COVID-19, our empirical findings show that increased internet searches during the pandemic caused the information to flow into the financial markets at a faster rate and thus resulting in higher implied volatility directly and via the stock return-risk relation. More specifically for the latter, the leverage effect in the VIX becomes stronger as Google search queries intensify. Both the direct and indirect effects on implied volatility, highlight a risk-aversion channel that operates during the pandemic. We also find that these effects are stronger in Europe than in the rest of the world. Moreover, in a panel vector autoregression framework, we show that a positive shock on stock returns may soothe COVID-related Google searches in Europe. Our findings suggest that Google-based attention to COVID-19 leads to elevated risk aversion in stock markets.

1. Introduction

The coronavirus disease 2019 (COVID-19) is the first global deadly pandemic after more than a century (the last one was the 1918 flu pandemic, also known as the Spanish flu). The first known outbreak of COVID-19 was around the end of 2019 in China and in less than three months, it spread across the globe, causing a huge number of infections and deaths in more than 200 countries.

This paper investigates the link between the dynamics of implied volatility indices in thirteen countries across the globe and investor attention as measured by Google search queries during the COVID-19 outbreak. To assess the adverse impact of COVID-19 on financial markets, academics examine various channels. Given the anxiety and uncertainty brought on by COVID-19, it is not surprising that search terms for the pandemic may have predictive power over market participants' expectations about future realized volatility. Information has become one of the most valuable assets in the financial market. One of the most important assets on the financial market today is information and scholars are coming to recognize the predictive value of data collected across various digital platforms (Da et al., 2011; Dai, Xiong, Liu, Huynh, & Sun, 2021; Jiang, Zhu, Zhang, Yan, & Shen, 2021), with search behavior data being an extremely abundant repository of predictive data. Peng and Xiong (2006) modeled investors’ endogenous attention allocation and predicted that investor-limited attention increases when the uncertainty level is high.

Google search query data are increasingly being used in the literature across several disciplines for measuring variables and phenomena ranging from the spread of flu (Ginsberg et al., 2009; Polgreen, Chen, Pennock, Nelson, & Weinstein, 2008) and election outcomes (Metaxas & Mustafaraj, 2012), through tourist numbers (Choi & Varian, 2012) to consumer behavior (Carrière-Swallow & Labbé, 2013) and economic statistics and key economic figures (Choi & Varian, 2012). In the macro-finance field, two main strands of research examine search queries provided by Google: the first group focuses on how Google search volumes relate to financial markets regarding returns, volatility, and liquidity (indicatively, Mondria, Wu, & Zhang, 2010; Drake, Roulstone, & Thornock, 2012; Kristoufek, 2013; Preis, Reith, & Stanley, 2010; 2013; Poutachidou & Papadamou, 2021), while the second body of papers attempts to create sentiment/uncertainty indicators based on Google searches that can explain several macroeconomic variables (see Costola, Iacopini, & Santagiustina, 2021 and the references therein).

The current study adopts a rather different viewpoint and relates to the literature that examines how markets and market agents react to exogenous events, such as natural disasters, social unrest, political upheavals, and violent events (see Kollias, Papadamou, & Arvanitis, 2013 and Liu, Peng, Hu, Dong, & Zhang, 2019 for a relevant discussion); search queries data extends existing attempts, as it measures the public's attention to unexpected catastrophic events and gives timely feedback on investment dynamics (Liu et al., 2019). Furthermore, no previous disease outbreak has affected economic activity as the COVID-19 pandemic did. As a result, a vast amount of literature that studies the negative impact of the COVID-19 pandemic on financial markets has emerged in a short period (see among others (a) effects on returns: Alfaro, Chari, Greenland, & Schott, 2020; Gormsen & Koijen, 2020; Ramelli & Wagner, 2020; Zhang, Hu, & Ji, 2020; (b) financial contagion: Corbet, Hou, Hu, Lucey, & Oxley, 2020; Corbet, Larkin, & Lucey, 2020; Akhtaruzzaman, Benkraiem, Boubaker, & Zopounidis, 2022; Akhtaruzzaman et al., 2021, Akhtaruzzaman et al., 2021, Akhtaruzzaman et al., 2021; Zehri, 2021; Zorgati & Garfatta, 2021; Alshater, Alqaralleh, & El Khoury, 2023; (c) risk aversion: Fassas, 2020; (d) oil and/or gold markets: Akhtaruzzaman, Boubaker, Chiah, & Zhong, 2021; Akhtaruzzaman, Boubaker, Lucey, & Sensoy, 2021; (e) bond market: Papadamou, Fassas, Kenourgios, & Dimitriou, 2021; Cicchiello, Cotugno, Monferrà, & Perdichizzi, 2022; Akhtaruzzaman et al., 2022, Akhtaruzzaman et al., 2022; (f) effects on volatility: Bai, Wei, Wei, Li, & Zhang, 2021; Maghyereh, Abdoh, & Awartani, 2022; Syed, 2022; Vuong, Nguyen, & Huynh, 2022; Wang & Liu, 2022; Fassas, 2023).

Chundakkadan and Nedumparambil (2022) show that significant declines in stock markets during the COVID-19 pandemic cannot be solely attributed to economic activity restrictions caused by lockdowns. Changes in investor sentiment also play a significant role in these market downturns. As a result, there is an increasing amount of research that investigates the connection between investor sentiment and stock market dynamics in the context of the COVID-19 pandemic. This paper falls into the emerging strand of COVID-related empirical work (Cevik, Kirci Altinkeski, Cevik, & Dibooglu, 2022; Lyócsa, Baumöhl, Výrost, & Molnár, 2020; Simionescu & Raišienė, 2021; Smales, 2021; Szczygielski, Bwanya, Charteris, & Brzeszczyński, 2021; Vasileiou, 2021) in which investor sentiment is captured by Google search volume. However, our study is the first to address the predictive power of Google searches on implied equity volatility during the COVID-19 pandemic for 13 countries across the globe. Our results show that COVID-19 sentiment not only generated excess volatility in financial markets, as existing empirical evidence suggests, but also resulted in market participants’ expecting higher volatility in the future as well. Our findings remain robust with alternative specifications.

Based on the assumption that information is instantaneously incorporated into prices when it arrives, Da et al. (2011) are the first to propose a new and direct measure of investor attention using search frequency in Google. Since then, an emerging literature documented the impact of investors' attention on asset market microstructure and asset prices and volatility (Alqahtani, Wither, Dong, & Goodwin, 2020; Chronopoulos et al., 2018; Dimpfl & Jank, 2016; Ding & Hou, 2015; Goddard, Kita, & Wang, 2015; Joseph, Wintoki, & Zhang, 2011; Schneller, Heiden, Heiden, & Hamid, 2018; Smith, 2012; Vlastakis & Markellos, 2012; Vozlyublennaia, 2014). Our paper is motivated by the papers of Andrei and Hasler (2015) – who develop a theoretical model that shows how investors’ attention affects the dynamics of asset returns – and Da, Engelberg, and Gao (2015) – who find that internet search volume for relevant terms such as “recession” and “unemployment” is contemporaneously related to the S&P500 VIX. Cevik et al. (2022) proxy for negative and positive investor sentiments using the Google Search Volume Index and show that there is a significant relationship between investor sentiment and stock market returns and volatility. Against this background, the current study extends this literature, by measuring investor attention to the coronavirus pandemic, following a similar approach to Da et al. (2011), and analyzing the relation between Internet search activity and implied volatility in a period of extraordinary stress and uncertainty.

In this regard, we formulate four main research questions: (1) Does the level of investors' attention, as indicated by Google's search volume index, provide valuable insights into stock market implied volatility? (2) Does the level of investors' attention, as indicated by Google's search volume index, have a significant impact on returns of equity indices? (3) Do the dynamics of the relationship between stock index returns and changes in implied volatility differ during times of crisis? (4) Do investors exhibit a higher demand for information during periods of pandemic shocks? To investigate the aforementioned questions, we construct a composite index using the terms "coronavirus" and the thematic relevance of "coronavirus" as a direct measure of investors' attention. Our analysis encompasses a diverse set of thirteen countries worldwide, utilizing a sample of equity and implied volatility indices.

The rationale behind our research hypotheses is that the time variation of individuals’ internet search activity reflects noise trading, which should increase stock market volatility (Graham, Nikkinen, & Peltomäki, 2020). The progress in information technology provides a great opportunity to develop an ex-ante proxy for quantifying investors' attention as empirical evidence indicates that both economic agents and market participants rely on online information search as an integral component of their investment decision-making process. However, we should consider that users who search for these terms are not necessarily investors because there are many other types of users. Therefore, although internet search queries are a likely proxy for retail investors' attention, we cannot observe whether the individuals searching are also the same that trade and cause the higher volatility (Dimpfl & Jank, 2016). Still, regardless of the connection between Google search queries and noise traders, it has been documented in the empirical literature (Da et al., 2015; Dimpfl & Jank, 2016; Vlastakis & Markellos, 2012) that retail investor attention contains information about future volatility and enhances price efficiency (Chen, 2020).

This paper investigates the link between implied volatility and investor attention during the first wave of the COVID-19 pandemic (January–April 2020) in thirteen countries across the globe. We measure implied volatility using the 30-day constant maturity equity volatility indices (calculated based on the model-free variance swap methodology) for Germany, France, Italy, Netherlands, United Kingdom, Switzerland, Russia, China, India, Japan, Korea, United States of America and Australia, while investor attention is proxied by Google search queries. Our study involved creating an equally weighted synthetic index using the search term "coronavirus" and its corresponding topic. Google Trends played a crucial role in distinguishing between search "topics" and "terms." Regardless of language or country, Google Trends aggregates data from diverse search queries related to a specific topic. We specifically focused on the volume of Google search queries categorized under the topic "Coronavirus" for our analysis. This approach included incorporating synonyms and phrases associated with the coronavirus, even if they were not explicitly searched for. Furthermore, Google Trends provides commonly used search queries or "terms" that are relevant to each topic.

The novelty of our study is that it not only analyzes the influence of COVID-19-related sentiment on stock returns but also demonstrates how investors' attention influences their expectations of future equity volatility. Additionally, we reveal an indirect mechanism wherein a positive surge in Google searches related to COVID-19 has an immediate and statistically significant negative impact on returns, leading to subsequent rises in implied volatility. Our main findings are summarized as follows: first, we identify short-run causality from Google search queries data to implied volatility dynamics. Second, there is a negative short-run effect of Google trend measure for COVID-19 on equity indices returns. Third, there is evidence that the persistence (leverage effect) in the VIX becomes stronger as Google search queries intensify. Fourth, Google carries different short-run predictive information in Europe relative to the rest of the world. Finally, we show that a positive shock on stock returns may calm down Google searching about COVID-19 in Europe.

The rest of the paper is organized as follows. Section 2 presents the literature review and the hypotheses to be tested. Section 3 presents the data sample and a preliminary analysis, while the description of the methodology is presented in Section 4. The empirical results and a robustness analysis are presented in Section 5 and discussed in detail in Section 6. Section 7 reports the summary and concluding remarks.

2. Background and hypotheses

Our paper attempts to contribute to the recent path of literature that studies the effects of investors’ attention, as proxied by Google search volume, on financial assets. A literature studying the impact of investor attention on the dynamics of asset prices has emerged during the last two decades (see Baker & Wurgler, 2007 and Bekaert, Engstrom, & Xu, 2022 for a review). One important channel through which investors express their demand for information is through internet searches (Drake et al., 2012). The appeal of search-based sentiment measures is more transparent when compared with alternatives (see Da et al., 2015 for a discussion).

Da et al. (2011) were the first to treat Google Search as a direct measure of investor attention; their empirical findings show that an increase in search volume for Russell 3000 stocks predicts higher stock prices in the next two weeks. Subsequently, several papers studying the connection between investor attention, as measured by search queries, market returns, and volatility have emerged over recent years. Indicatively, studies documenting this link include (Bijl, Kringhaug, Molnár, & Sandvik, 2016; Chen & Lo, 2019; Dimpfl & Kleiman, 2019; Drake et al., 2012; Joseph et al., 2011; Kim, Lučivjanská, Molnár, & Villa, 2019; Vlastakis & Markellos, 2012) for individual stocks; Dzielinski (2012), Vozlyublennaia (2014), Hamid and Heiden (2015), Da et al. (2015), Chronopoulos et al., (2018), Dimpfl and Jank (2016) and Graham et al. (2020) for stock indices; Goddard et al. (2015) and Smith (2012) for exchange rates; Vozlyublennaia (2014) and Dergiades, Milas, and Panagiotidis (2015) for bonds (Afkhami, Cormack, & Ghoddusi, 2017; Li, Tang, & Li, 2020; Vozlyublennaia, 2014); for commodities; Da et al. (2015) for ETFs; Yung and Nafar (2017) for REITs and Philippas, Rjiba, Guesmi, and Goutte (2019) for Bitcoin.

Extensive research has documented the effect of investors' attention on asset prices and volatility, but only a very limited number of studies have investigated the link of investors' attention to implied volatility, one of the most popular market-based measures of investor sentiment (Whaley, 2000). In particular, Vlastakis and Markellos (2012), Da et al. (2015) and Ruan and Zhang (2016) show that investors’ attention is significantly positively related to implied volatility, while Nikkinen and Peltomäki (2020) show that the effects of information demand on realized stock returns and the VIX index are instantaneous.

Based on this empirical literature, we test four hypotheses in order to analyze the relationship between investors' attention, as proxied by Google search queries, and market aggregate risk-return dynamics. Empirical research in finance has long been investigating the link between volatility and the rate at which information flows into financial markets (see Kalev, Liu, Pham, & Jarnecic, 2004 and Da et al., 2011 for a review of the relevant literature); one of the most intuitive explanations for commonly observed volatility patterns is that volatility is proportional to the rate of information arrival (Smith, 2012). Similarly, Smales (2021) suggests that the Google's search volume index (GSVI) provides a direct and timely measurement of the retrieval of available information, while Costola et al. (2021) demonstrates that the Google's search data contribute to explaining the dynamic of stock market returns for Italy, Spain, and Germany. Implied volatility incorporates equity investors' estimate of the realized volatility of the underlying index over the next thirty calendar days. In this regard, we want to investigate whether Google search volumes correlate with implied volatility. Thus, the first hypothesis is as follows:

H1

Investors' attention as proxied by Google's search volume index adds information to a market-based measure of volatility.

Our second hypothesis stems from the attention-induced price pressure hypothesis of Barber and Odean (2013), which suggests that increased attention leads to increased buying, and thus pushes prices and returns higher. Theoretical and empirical studies indicate that attention plays a crucial role in shaping asset pricing dynamics. Nevertheless, it is important to acknowledge that our empirical analysis is based on specific samples, such as individual investors or a limited period, which may not fully represent the broader market. Additionally, by treating all investors' attention as homogeneous, we may overlook the potential variations in attention patterns and behaviors among different types of investors. Taking these considerations into account, we are going to examine the relationship between Google search volumes and equity returns, as measured by the thirteen benchmark indices under review. As a result, the second hypothesis is as follows:

H2

Investors' attention as proxied by Google's search volume index significantly affects daily equity indices returns.

The third hypothesis relates to one of the most noticeable stylized facts in finance; the negative correlation of stock index returns with changes in volatility. Two main hypotheses have been developed to explain this relationship: the leverage effect proposed by Black, 1976 and the volatility feedback effect (Campbell & Hentschel, 1992; French, Schwert, & Stambaugh, 1987; Fleming, Ostdiek, & Whaley, 1995). Although both hypotheses expect a negative relationship between volatility and returns, they explain the causality in opposite directions; the former suggests that causality runs from returns to volatility, whereas the latter runs the other way around. Empirical evidence (Bekaert & Wu, 2000) generally supports both hypotheses, as it documents that causality runs in both directions. According to Carr and Wu (2017), crises have displayed a concerning pattern that causes unease among policymakers and financial managers: a significant negative financial event frequently enhances the likelihood of subsequent similar events. In this regard, out third hypothesis tests whether:

H3

The negative relationship between implied volatility and return is affected by the pandemic outbreak.

Finally, investor attention, as measured by Google search volumes, is found to be strongly time-varying and higher in periods of high volatility. In particular, relatively scant empirical evidence (see Andrei & Hasler, 2015 and the references therein) shows that investors demand more information as a shock to index returns occurs and as the level of risk aversion increases, while Dimpfl and Jank (2016) find a strong co-movement and bi-directional relationship between stock market indices' realized volatility and the search queries for their names. On the other hand, implied volatility indices represent market participants’ best collective forecast of the forward thirty-day realized volatility of the underlying stock index. Given the above, the fourth hypothesis is as follows:

H4

Increased implied volatility, as proxied by the model-free volatility indices, leads to increased investors' attention.

Thus, we are going to test whether the documented realized volatility and investors’ attention relationship holds for implied volatility as well.

3. Data and preliminary analysis

The development of COVID-19 in late 2019 in China and its contagion on other countries around the world over the beginning of 2020 leads us to focus on a sample period from 02 January 2020 to 09 April 2020. Google trends metrics provide useful information concerning the attention of the crowd on the pandemic of COVID-2019 over this period. The metrics achieve the highest level (one hundred) on the day of this period where attention is highest, and the rest days are presented in reference to that.

Adopting this approach, we opted to create a synthetic index that is equally weighted. This index is based on the search term "coronavirus" and the corresponding topic of "coronavirus." Google Trends provides support for distinguishing between search "topics" and "terms".1 Irrespective of the country or language in which the search queries are expressed, Google Trends consolidates data for various search queries associated with a specific topic (Anastasiadis & Papadamou, 2022). In our study, the data series is derived from the volume of Google search queries categorized under the topic "Coronavirus". This approach includes synonyms and phrases linked to the coronavirus, even if they have not been specifically searched for. For instance, it encompasses individuals who may have clicked on an image displaying Coronavirus statistics or watched videos related to the symptoms or potential development of the disease. For each topic, Google Trends also reports search queries or “terms” related to the topic, which are frequently performed by users. The search “term” is the word or phrase that a specific user types into a search engine to discover results based on it. We checked the relevance of search queries reported by Google Trends concerning the “coronavirus” search topic used for this study. For example, Google Trends considers the phrases “coronavirus consequences”, “coronavirus disease” and “coronavirus news” as relevant to the “Coronavirus” topic.

By creating this synthetic indicator, we can capture any temporal changes starting from the early months of 2020 and spanning over the first four months of our analysis period (the end of the sample period coexists with the closing of the stock markets for the Easter holidays in the USA and most European countries). This allows us to identify and analyze any variations in the data during this timeframe.

The thirteen sample countries’ selection is based on the availability of an implied volatility index and the existence of coronavirus victims. Therefore, we use data on general stock market indices and implied volatility indices (VIX) for Germany (DEU), France (FR), Italy (IT), Netherland (NL), United Kingdom (UK), Switzerland (CH), Russia (RUS), China (CHN), India (IND), Japan (JPN), Korea (KOR), United States of America (USA) and Australia (AUS). All the included equity volatility indices have a constant 30-day maturity and are based on the model-free variance swap methodology introduced by CBOE, which is based on the fair value calculation of a variance swap (Demeterfi, Derman, Kamal, & Zou, 1999).2

Fig. 1a, Fig. 1b, Fig. 2, Fig. 3 include the graphical depiction of the three variables of interest, Google Trend Metric (GTR), Implied Volatility (VIX), and Stock Prices, for each country in a vertical manner. By looking at European countries in Fig. 1a, Fig. 1b a and b, the beginning of the search concerning COVID-19 can be placed in the middle of January (specifically the indicator takes off on day 20/01/2020). However, this is also true for all countries in the sample (see Fig. 2, Fig. 3 ). By looking at these figures, the first wave of increase in Google trend indicator concerning COVID-19 coexists with a small increase in implied volatility and a short decline in stock prices (this effect seems to be clearer in China and Korea relevant to the rest countries). Over the second wave of increase in Google trend indicator beginning on 19/2/2020, the drop in stock prices and the increase of implied volatility are common and unified across all countries and regions.

Fig. 1a.

Google Trend on Coronavirus, Stock & VIX prices in European Countries. Notes: The horizontal lines of the figure present a) the GTR metric (Google Trend on Coronavirus) in each country, b) the Stock prices (stock prices of general index) in each country, c) the VIX prices (Implied Volatility index prices) in each country. The tickers for each country are presented as follows CH: Switzerland; DEU: Germany; FR: France; IT: Italy. M1, M2, M3, M4 in the x-axis of the sub-figures represent months of year 2020.

Fig. 1b.

Google Trend on Coronavirus, Stock & VIX prices in European Countries. Notes: The horizontal lines of the figure present a) the GTR metric (Google Trend on Coronavirus) in each country, b) the Stock prices (stock prices of general index) in each country, c) the VIX prices (Implied Volatility index prices) in each country. The tickers for each country are presented as follows NL: Netherland; RUS: Russia; UK: United Kingdom. M1, M2, M3, M4 in the x-axis of the sub-figures represent months of year 2020.

Fig. 2.

Google Trend on Coronavirus, Stock & VIX prices in Asian Countries. Notes: The horizontal lines of the figure present a) the GTR metric (Google Trend on Coronavirus) in each country, b) the Stock prices (stock prices of general index) in each country, c) the VIX prices (Implied Volatility index prices) in each country. The tickers for each country are presented as follows CHN: China; IND: India; JPN: Japan; KOR: Korea. M1, M2, M3, M4 in the x-axis of the sub-figures represent months of year 2020.

Fig. 3.

Google Trend on Coronavirus, Stock & VIX prices in USA and Australia. Notes: The horizontal lines of the figure present a) the GTR metric (Google Trend on Coronavirus) in each country, b) the Stock prices (stock prices of general index) in each country, c) the VIX prices (Implied Volatility index prices) in each country. The tickers for each country are presented as follows USA: United States; AUS: Australia. M1, M2, M3, M4 in the x-axis of the sub-figures represent months of year 2020.

Panel A of Table 1 presents the descriptives statistics for rate of changes of stock, VIX and Google Trend indices in a panel data framework. As can be easily seen, GTR and VIX changes present the higher means and the higher standard deviation when compared with the relevant stock market indices measures. Worth mentioning the fact that Google trend changes present the highest values in terms of skewness and kurtosis.

Table 1.

Summary of Descriptive Statistics and panel unit root tests.

PANEL A	VIX changes	Stock changes	GTR changes
Mean	0.0206	−0.0041	0.0154
Maximum	0.6195	0.1041	1.0147
Minimum	−0.5680	−0.1854	−0.3403
Std. Dev.	0.1189	0.0316	0.1264
Skewness	0.7166	−0.8193	3.4279
Kurtosis	5.7993	7.2707	23.4897
Observations	767	767	767
Cross sections	13	13	13

PANEL B	VIX changes				Stock changes				GTR changes
	Cross-				Cross-				Cross-
Method	Statistic	Prob.	sections	Obs.	Statistic	Prob.	sections	Obs.	Statistic	Prob	sections	Obs.
Null: Unit root (assumes common unit root process)
Levin, Lin & Chu t-stat*	−23.431	0.00***	13	767	−26.159	0.00***	13	767	−19.979	0.00***	13	767
Null: Unit root (assumes individual unit root process)
Im, Pesaran and Shin W-stat	−23.593	0.00***	13	767	−23.625	0.00***	13	767	−21.332	0.00***	13	767
ADF - Fisher Chi-square	426.631	0.00***	13	767	424.321	0.00***	13	767	344.034	0.00***	13	767
PP - Fisher Chi-square	460.068	0.00***	13	767	525.935	0.00***	13	767	424.341	0.00***	13	767

Notes: Panel A of the table reports the descriptives statistics for VIX changes, stock changes and google trend indicator (GTR) changes. Panel B of the table includes the panel unit root tests results for VIX changes, stock market changes and google trend indicator (GTR) changes. The main research hypothesis tested is the following: H₀: unit root is present. The upper part of the table presents the values of the Levin, Lin & Chu t-statistic and the relevant P-values (see column entitled Prob.). This test assumes common unit root processes across the 13 markets. The lower part of the table presents the relevant values for tests assuming individual unit root process across markets. *** indicates statistical significance at the 1% level.

Given that Fig. 1a, Fig. 1b, Fig. 2, Fig. 3 leave no doubts of non-stationary series, we proceed on first logarithmic changes of level data for GTR, VIX and Stock Indices3 to investigate stationarity via panel unit root tests. Therefore, the first step of our empirical investigation involves a few panel unit root tests applied to these changes (VIX changes, GTR changes, Stock changes). The results in Panel B of Table 1 indicate strong evidence against non-stationarity for the changes of all series under review. Here on, our focus will be on changes more than level data.

Assumed the well-known negative relationship between stock market returns and stock market implied volatility in finance (Whaley, 2000), we come to assess firstly any direct impact of Google Trend metric concerning COVID-19 on stock market implied volatility, and secondly any indirect effect working via this well-known relationship. More specifically, we expect that uncertainty concerning a spread of a pandemic with dangerous health results adds to future stock market volatility measured by the implied volatility measure, but also may strengthen in absolute terms the relationship between stock market returns and implied volatility changes. A significant drop in the stock market may increase implied volatility more in a contagion environment of a pandemic.

In order to have a preliminary picture of the relationship pattern between VIX and stock market indices, we apply the dynamic conditional correlation approach (DCC) of Engle (2002). DCC is a parsimonious estimation technique of a dynamic correlation between two series and provides useful information on their correlation fluctuations over time. The level of DCC correlations between VIX and stock market returns over the period 02/01/2020 through 09/04/2020 is shown in Fig. 4 .

Fig. 4.

Temporal Dynamic correlations. Notes: Dynamic correlations between VIX index and stock market returns for each country. The vertical lines represent the following dates: 20/1/2020 and 19/2/2020. The abbreviations are analyzed as follows. VAEX, AEX indicate the volatility and stock market indices of the Netherland, respectively. VCAC, CAC40 indicate the French volatility and stock market indices, respectively. VDAX, DAX indicate the German volatility and stock market indices, respectively. VIX, SPX indicate the US volatility and stock market indices, respectively. IVUKX30, FTSE indicate the UK volatility and stock market indices, respectively. AXVI, AXJO indicate the Australian volatility and stock market indices, respectively. RVI, RTS indicate the Russian volatility and stock market indices, respectively. VHSI, Hang Seng indicate the Hong Kong volatility and stock market indices, respectively. INDIA VIX, NIFTY50 indicate the Indian volatility and stock market indices, respectively. VSMI, SMI indicate the Swiss volatility and stock market indices, respectively. JNIV, NIKKEI indicate the Japanese volatility and stock market indices, respectively. VKOSPI, KOSPI indicate the Korean volatility and stock market indices, respectively. IVMIB30, FTSEMIB indicate the Italian volatility and stock market indices, respectively.

At all points, the degree of correlation between the series is found to be low, with a magnitude less than −0.2 in most cases. However, it is worth highlighting the change of correlation near and after the milestone dates of January 20, 2020, and February 19, 2020. The level of correlation fluctuates considerably during these periods, supporting changes in investors’ appetite for risk. This change is asymmetric and depends heavily on the market under examination. However, while DCC provides information on the correlation characteristics of our series over time, it sheds no light on the causal effects among series. This supports the use of panel data techniques for examining the series behavior in more depth, by including more explanatory variables.

4. Methodology

In the empirical analysis we study thirteen countries using daily data for a period of three and half months in a panel data framework. Panel data estimation allows us to control for individual heterogeneity and reduces estimation bias (Wooldridge Jeffrey, 2002).

Firstly, following previous literature indicating the strong relationship between volatility and return (Bekaert & Wu, 2000; Bollerslev, Tauchen, & Zhou, 2009; Fleming et al., 1995; Hibbert, Daigler, & Dupoyet, 2008), we formulate a model that incorporates any possible GTR effects on volatility changes as shown below:

$${{VIX}_{changes}}_{i,t}=a_0+a_1{GT{R}_{changes}}_{i,t}+a_2{GT{R}_{changes}}_{i,t-1}+a_3{Stoc{k}_{changes}}_{i,t}+a_4\left({Stoc{k}_{changes}}_{i,t}\times {GT{R}_{changes}}_{i,t}\right)+a_5\left({Stoc{k}_{changes}}_{i,t}\times {GT{R}_{changes}}_{i,t-1}\right)+{\mu }_i+{\phi }_t+e_{i,t}$$ (1)

where, ${{VIX}_{changes}}_{i,t}$ measures the rate of change on stock market implied volatility index (i corresponds to each country of the sample and t on each day of the sample) and ${Stoc{k}_{changes}}_{i,t}$ refers to stock market price daily changes for each cross section i, while the rate of change on Google trend measurement about COVID-19 is measured for each country i over day t by the ${GT{R}_{changes}}_{i,t}$ and can also be treated as an independent variable. As can be seen in our model, given that we don't know whether the Google trend variable affects contemporaneous or in a time lag the changes on volatility index (VIX), we estimate a wider model allowing for time dynamic direct and indirect effects. Furthermore, since underlying equity index returns essentially incorporate all relevant information regarding the equity market under review, we can capture the response of investor sentiment (as proxied by implied volatility) to GTR shocks more accurately.

Under this specification, the $a_1\text{,}$ $a_2$ coefficients capture any direct effects of the uncertainty concerning the COVID-19 epidemic on risk taking at stock markets, as proxied by implied volatility. If anxiety about the future possible negative social and/or economic effects of the pandemic does directly discourage risk-taking, we expect the $a_{1}and/or$ $a_2$ coefficient to be positive and statistically significant. However, in our model we also quantify any indirect effect working through the well-established negative relationship between stock returns and implied volatility changes (we expect $a_3\le 0$, and statistically significant). In addition, we expect that a cross-term of stock returns with the Google Trend metric may increase in absolute terms this negative relationship (we expect $a_4\le 0$ or $a_5\le 0$, and statistically significant).

The parameter $a_0$ represents the overall constant in the model, while the μ_i and ${\phi }_t$ represent respectively the cross section and time specific effects (random or fixed). The $e_{i,t}$ are the error terms for i = 1, 2, …,13 cross-sectional units, observed for t = 1, 2, …,T daily periods. This model can help us investigate any significant contemporaneous and/or dynamic relationship between the variables of interest. Furthermore, since the futures market data essentially captures all relevant information regarding the stock market under review, we can quantity the response of investor sentiment to epidemic anxiety shocks more effectively.

Secondly, in order to investigate any causality between implied stock market volatility changes, stock market returns and Google Trends changes, we proceed to the estimation of a panel VAR model including ${VIX}_{changes},{Stock}_{changes}andGT{R}_{changes}$ as endogenous variables, using the GMM estimator of Abrigo and Love (2016). Impulse response analysis and variance decomposition may highlight any significant causal effects and any time delayed response of variables of interest. Therefore, our general first PVAR model is defined as follows:

$$Z_{i,t}={\Gamma }_0+{\Gamma }_1{{\rm Z}}_{it-1}+f_i+e_{i,t}$$ (2)

in which, $Z_{i,t}$ is a three-variable vector {${VIX}_{changes},{Stock}_{changes},GT{R}_{changes}$}. This panel VAR model allows for “individual heterogeneity” in the levels of the variables by introducing fixed effects, denoted by $f_i$ in Equation (2).

5. Empirical results

Table 2 presents the estimation results for the pooled, fixed and random effects models regarding Equation (1). In the upper part of column (1) the independent variables of the models are presented, while in the lower part some diagnostics and specification tests are also provided. In column (2) the expected signs of the coefficients of interest are given. We follow a top-down econometric approach, beginning with a wide model including all variables of the model presented in equation (1), and finally keeping only the statistically significant ones (the results of this reduced model are presented in Table 2).

Table 2.

Panel data estimation results for changes on stock market VIX.

Dependent Variable		VIX changes
Independent Variables	Exp. Sign	Pooled with PCSE		Fixed Effects with PCSE		Random Effects with PCSE		Pooled OLS with PCSE and AR(1)	Fixed Effects with PCSE and AR(1)
Model		(1)	(2)	(1)	(2)	(1)	(2)	(2)	(2)
C	+/−	0.0067	0.0083	0.0108	0.0112	0.0096	0.0103	0.0078	0.0077
		(0.04)**	(0.01)**	(0.00)***	(0.00)***	(0.21)	(0,00)***	(0.00)***	(0.00)***
GTR changes(t)	+	0.1206	0.1352	0.0926	0.0993	0.1028	0.1105	0.1398	0.1405
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
GTR changes(t-1)	+	0.0338		−0.0093		0.0034
		(0.12)		(0.71)		(0.91)
Stock changes(t)	–	−1.9618	−2.0118	−1.5027	−1.5058	−1.6523	−1.6656	−2.1183	−2.1220
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
Stock changes(t) x GTR changes(t)	–	−3.4598		−0.8889		−1.5054
		(0.11)		(0.66)		(0.33)
Stock changes(t) x GTR changes(t-1)	–	−3.2075	−4.3030	−3.7950	−3.6266	−3.6301	−3.7408	−4.1490	−4.2142
		(0.04)**	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
ρ [AR(1) coeff.]	+/−							−0.1654	−0.1675
								(0.00)***	(0.00)***
Country Effects		no	no	yes	no	yes	no	no	yes
Time Effects		no	no	yes	yes	yes	yes	no	no
R2		35.71%	35.33%	55.77%	56.53%	34.90%	34.55%	34.27%	36.09%
F Test		86.09***	140.53***	13.87***	17.33***	38.52***	63.72***	113.21***	28.04***
Durbin-Watson stat.		2.35	2.33	2.76	2.75	2.65	2.28	2.03	2.03
N =(ixT)		767	767	767	767	767	767	754	754
Specification tests
Cross section F-test (pooled OLS vs. FEM) p-value		(0.99)	(0.99)
Period F-test (pooled OLS vs. FEM) p-value		(0.00)***	(0.00)***
Hausman test (FEM vs REM) - Period Random					(0.08)*
Test of cross-sectional independence by Frees				(0.00)***
Modified Wald test for group wise heteroskedasticity				(0.00)***

Notes: This table presents the estimation results for VIX changes regressed on stock market changes, google trends indicator (GTR) contemporaneously (t) and with a lag (t-1), and on cross terms of stock changes with GTR measurement each time. The pooled, fixed and random effects models with panel corrected standard errors (PCSE) are estimated regarding Equation (1) for the period from 02 January 2020 to 09 April 2020. P-values are in parentheses. *,** and ***indicate statistical significance at the 10%, 5%, and 1% level respectively. The column Exp. Sign refers to expected signs of the coefficient of the relevant explanatory variable. Coefficient ρ is estimated assuming a common across countries autoregressive model [(AR(1)] for the residuals. Next panel of the table reports the adjusted R² coefficient, the F-test, the Durbin–Watson value and the number of observations [market units (i) x time observations (t)]. The lower part of the table presents specification tests for the models estimated. Initially pooled versus fixed effects are tested, and then a Hausman test is applied between fixed versus random effects models for cases where previously fixed effects models are selected versus pooled models. Finally, p-values of tests for cross sectional independence and group wise heteroscedasticity are reported.

In terms of estimation methods, we begin with the pooled estimation and by conducting a fixed versus pooled estimation F-test, we proceed with the one that is suggested by the test. We conduct a cross section and a period F-test in order to select the most suitable model. Moreover, the Hausman test helps us in order to select between Fixed and Random Effects models, by giving a small lead to the latter. However, for the reason of consistency and robustness we present the results of all types of estimation methods. We must mention at this point that the results of the F-test for cross-sectional correlation in fixed effect and modified Wald test for group-wise heteroskedasticity leads us to use panel corrected standard errors (PCSE) in all of our estimations.4

By looking at the columns of Pooled, Fixed and Random Effects models, we can conclude that: (a) stock return increases coexist with reductions in implied volatility changes; (b) the direct effect of Google trend on implied volatility is positive, contemporaneous only, and statistically significant at the 1% and 5% level of significance; and (c) the indirect effect is also statistically significant. More specifically, based on our findings, an increase in Google Trend about the COVID-19 pandemic over the previous day can strengthen the negative relationship of stock returns on the returns of the implied volatility of the stock index. However, this is not happening contemporaneously. Our three main findings are robust across the estimation methods applied to the reduced model that keeps only the statistically significant variables presented in Table 2. The constant term in all cases is positive and statistically significant, reflecting an average daily VIX change. The R-squared across all estimations varies between 34.5% and 56.3%, with the highest value on the model without cross section (or country) effects, but with time effects.

In the next step, given the values of the Durbin-Watson tests and allowing the residuals to follow an autoregressive one process, we proceed to the estimation of the reduced model with the Pooled OLS5 /Fixed Effects with PCSE and an AR(1) term common across all cross sections. The results of these estimations are presented in the right last two columns of Table 2. In terms of the magnitude of the three coefficients of interest, all are higher in absolute terms in both cases of estimation, without reducing statistical significance. The negative statistically significant value of the coefficient of the autoregressive term implies a mean reversion process on VIX changes across all stock markets in those thirteen countries studied.

We continue our empirical analysis by investigating whether there is any feedback effect across the three variables of interest in a PVAR framework. The impulse response analysis can highlight the magnitude of the effect during the time and, most importantly, the direction of the effect across the three variables. The GMM estimation in a PVAR framework allows us to treat any bi-directional relationship between return and volatility in the stock market across all 13 countries, but also any bi-directional or uni-directional relationship between Google Trend metrics about the COVID-19 epidemic and the other two stock market variables.

Impulse response analysis in Fig. 5 highlights clearly the direction of these interdependencies. As expected, a positive shock on stock market returns reduces significantly implied volatility in the relevant market by reaching its bottom immediately and dies off after almost two days. However, a bi-directional relationship between these two variables seems to be present to some degree, since the response of stock returns to a positive implied volatility shock is negative, reaching its bottom after one day and disappearing after two days. By focusing on the last row of diagrams in Fig. 5, we observe that a positive shock on Google search for COVID-19 pandemic has an immediate statistically significant positive effect on stock market implied volatility and a clear negative effect on stock returns. In both cases, this effect diminishes after two days. By looking at diagrams in the diagonal, VIX changes and GTR changes present a quite similar pattern as long as their autoregressive part is considered in contrast to stock changes.

Fig. 5.

Impulse response analysis for base PVAR model. Notes: This figure depicts the impulse response analysis between stock returns (stock_changes), volatility indices changes (VIX_changes) and Google trend metric (gtr_changes). The variable after symbol “:” is the variable that responds to a positive shock in variable before symbol “:”.

By looking at the response of Google trend changes about the COVID-19 pandemic to a positive shock on stock changes and VIX changes respectively, we can highlight the following two points. On the one hand, a positive shock of stock returns significantly reduces the rate of searching about the COVID-19 pandemic, while on the other hand, a positive shock on implied stock market volatility does not seem to increase searching for the COVID-19 pandemic in a statistically significant way.

5.1. Robustness check

By experimenting with different groups of stock markets, we investigate the robustness of our central findings. More specifically, we re-estimate Equation (1) in its reduced form as appeared in Table 2 by decomposing the effect of each coefficient (concerning direct and indirect effects of Google Trend metric about COVID-19 on VIX changes) on two groups: the European versus non-European markets and the Asian versus non-Asian markets.6

A particular dummy $\left(D_j\right)$ is constructed, taking the value of one if market k belongs to j group (j takes value European or Asian markets) and zero otherwise. Similarly, the dummy $\left({1-D}_j\right)$ is constructed, taking the value of one if market k does not belong to the j group (i.e., non-European or non-Asian markets) and zero otherwise. Therefore, model 2 (the reduced model) is re-estimated as follows:

$${{VIX}_{changes}}_{i,t}=b_0+b_1^a\left[\left(D_j\right)\times {GT{R}_{changes}}_{i,t}\right]+b_1^b\left[\left({1-D}_j\right)\times {GT{R}_{changes}}_{i,t}\right]+b_2{Stoc{k}_{changes}}_{i,t}+b_3^a\left[\left(D_j\right)\times \left({Stoc{k}_{changes}}_{i,t}\times {GT{R}_{changes}}_{i,t-1}\right)\right]+b_3^b\left[\left(1-D_j\right)\times \left({Stoc{k}_{changes}}_{i,t}\times {GT{R}_{changes}}_{i,t-1}\right)\right]+{\mu }_i+{\phi }_t+e_{i,t}$$ (3)

Our attention is given to the comparison between $b_1^a$ and $b_1^b$ referring to “direct effects of GTR changes on Vix” and $b_2^a$ and $b_2^b$ referring to “indirect effects of GTR changes on VIX”, between j group versus rest markets each time.

Table 3 presents the results in line with Table 2. The only difference is that for each estimation method we present two different columns: one for the estimation of European markets versus non-European markets and another for the estimation of Asian versus non-Asian markets. Therefore, index j takes values EUR for European markets or Asia for Asian markets each time. Let us, firstly, focus on European versus non-European markets results. The direct effect of GTR changes on VIX changes is higher in magnitude for European markets versus non-European markets ($b_1^a$ >$b_1^b$). In Asian markets, the direct effect is smaller in magnitude compared to the rest markets. Additionally, when comparing the direct effect on Asian markets versus the direct effect on non-European markets, it can be implied that USA and Australian markets contribute positively relative to Asian markets. These main findings remain the same in the case that an AR(1) term is estimated, but the magnitude of the direct effect is in general increased.

Table 3.

Panel data estimation results for changes on stock market VIX in Europe vs. Asia.

Dependent Variable		VIX changes
Independent Variables	Exp. Sign	Pooled with PCSE		Fixed Effects with PCSE		Random Effects with PCSE		Pooled OLS with PCSE and AR(1)	Pooled OLS with PCSE and AR(1)	Fixed Effects with PCSE and AR(1)	Fixed Effects with PCSE and AR(1)
Model		j = EUR	j = Asia	j = EUR	j = Asia	j = EUR	j = Asia	j = EUR	j = Asia	j = EUR	j = Asia
C	+/−	0.0082	0.0082	0.0112	0.0111	0.0103	0.0103	0.0085	0.0084	0.0084	0.0084
		(0.01)**	(0.01)**	(0.00)***	(0.00)***	(0.17)	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
Dj x GTR changes(t)	+	0.1640	0.0715	0.1115	0.0710	0.1244	0.0772	0.2525	0.1535	0.2519	0.1538
		(0.00)***	(0.12)	(0.00)***	(0.07)*	(0.00)***	(0.04)**	(0.00)***	(0.04)**	(0.00)***	(0.04)**
(1-Dj) x GTR changes(t)	+	0.0931	0.1592	0.0744	0.1099	0.0846	0.1226	0.1921	0.2552	0.1913	0.2546
		(0.01)**	(0.00)***	(0.04)**	(0.00)***	(0.01)**	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
Dj x Stock changes(t)	–	−2.1195	−2.2504	−1.5771	−1.7971	−1.7503	−1.9450	−2.1343	−2.2972	−2.1402	−2.3095
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
(1-Dj) x Stock changes(t)	–	−1.8588	−1.9269	−1.4287	−1.4051	−1.5704	−1.5670	−1.9396	−1.9720	−1.9431	−1.9745
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
Dj x Stock changes(t) x GTR changes(t-1)	–	−5.4529	−3.5387	−3.9877	−3.7593	−4.2297	−3.7197	−4.7851	−2.6949	−4.9186	−2.8032
		(0.00)***	(0.07)*	(0.02)**	(0.02)**	(0.01)**	(0.02)**	(0.00)***	(0.25)	(0.00)***	(0.23)
(1-Dj) x Stock changes(t) x GTR changes(t-1)	–	−1.9794	−4.8256	−2.9839	−3.7633	−2.8230	−3.9512	−1.5600	−4.3682	−1.5504	−4.3790
		(0.39)	(0.01)**	(0.14)	(0.03)**	(0.16)	(0.02)**	(0.55)	(0.01)**	(0.52)	(0.01)**
ρ [AR(1) coeff.]	+/−							−0.1620	−0.1655	−0.1644	−0.1678
								(0.00)***	(0.00)***	(0.00)***	(0.00)***
Country Effects		no	no	no	no	no	no	no	no	yes	yes
Time Effects		no	no	yes	yes	yes	yes	no	no	no	no
R2		35.60%	35.37%	56.45%	56.55%	34.99%	34.74%	38.19%	38.11%	37.32%	37.23%
F Test		71.58***	70.88***	16.51***	16.58***	32.34***	32.30***	67.47***	67.24***	24.59***	24.51***
Durbin-Watson stat.		2.32	2.33	2.74	2.75	2.28	2.28	2.06	2.06	2.06	2.06
N =(ixT)		767	767	767	767	767	767	754	754	754	754
Specification tests
Cross section F-test (pooled OLS vs. FEM) p-value		(0.99)	(0.99)
Period F-test (pooled OLS vs. FEM) p-value		(0.00)***	(0.00)***
Hausman test (FEM vs REM) Period Random				(0.13)	(0.16)
Test of cross-sectional independence by Frees				(0.00)***	(0.00)***
Modified Wald test for group wise heteroskedasticity				(0.00)***	(0.00)***

Notes: This table presents the estimation results for changes on stock market VIX in Europe vs. Asia, regressed on cross terms of the dummy variables Dj and (1-Dj) with i) google trends indicator (GTR) contemporaneously (t), ii) stock changes contemporaneously (t) and finally iii) stock changes and google trends indicator (GTR) with a lag (t-1). By using dummy variables Dj, (1-Dj) we decompose the sample between j geographic area and the rest. The j takes the value of either EUR for European markets studied or Asia for Asian markets. The pooled, fixed and random effects models with panel corrected standard errors (PCSE) regarding Equation (3) for the period from 02 January 2020 to 09 April 2020. P-values are in parentheses. *,** and ***indicate statistical significance at the 10%, 5%, and 1% level respectively. The column Exp. Sign refers to expected signs of the coefficient of the relevant explanatory variable. Coefficient ρ is estimated assuming a common across countries autoregressive model [(AR(1)] for the residuals. Next panel of the table reports the adjusted R² coefficient, the F-test, the Durbin–Watson value and the number of observations [market units (i) x time observations (t)]. The lower part of the table presents specification tests for the models estimated. Initially pooled versus fixed effects are tested, and then a Hausman test is applied between fixed versus random effects models for cases where previously fixed effects models are selected versus pooled models. Finally, p-values of tests for cross sectional independence and group wise heteroscedasticity are reported.

Stock return is statistically significant and negatively related to VIX changes in all cases, but the magnitude of the coefficient is higher in absolute terms in Asian markets and then follows European markets. When looking at the indirect effect of Google trend searching for COVID-19 on this stock return VIX relationship, we can argue that this COVID-19 search when increasing strengthens this relationship in absolute terms. More specifically, this indirect effect of GTR changes on VIX via the stock return channel is present and statistically significant in European and Asian markets only. Again, when looking at the magnitude of the coefficients, the European case presents higher values.

In order to investigate further the significance of our findings in a framework that incorporates as explanatory variables in the models both lagged dependent variable and the rest independent variables of interest, we proceed with the generalized method of moments (GMM) estimator.7 In both the fixed and random effects settings, the lagged dependent variable is correlated with the error term, even if the disturbances are not autocorrelated. An instrumental variable procedure corrects for the endogeneity as well as the correlation between the lagged difference of the dependent variable and ε_i,t−1. Blundell and Bond (1998) show that the system GMM estimator overcomes these problems preferable to that of Arellano and Bond (1991) when the dependent variable and/or the independent variables are persistent.

Therefore, the left-hand side of Table 4 (Model A) presents the Panel GMM estimation results for VIX changes regressed on stock market changes and google trends indicator (GTR) contemporaneously (t), and on cross terms of stock changes with GTR_t-1 measurement for the period from 02 January 2020 to 09 April 2020 across all countries studied. As can be seen our results concerning signs and significance of coefficients do not change compared to previous results from FE and RE models.

Table 4.

Panel GMM estimation results for changes on stock market VIX in case of Europe vs. Asia and whole sample.

Dependent Variable	VIX changes
Independent Variables	Exp. Sign	System GMM	Independent Variables	Exp. Sign	System GMM
Model A		All countries	Model B		j = EUR	j = Asia
VIX changes(t-1)	+/−	−0.2279	VIX changes(t-1)	+/−	−0.4002	−0.1250
		(0.00)***			(0.00)***	(0.32)
C	+/−	0.0056	C	+/−	0.0024	0.0060
		(0.00)***			(0.58)	(0.04)**
GTR changes(t)	+	0.0531	Dj x GTR changes(t)	+	0.1355	0.1648
		(0.01)**			(0.04)**	(0.30)
			(1-Dj) x GTR changes(t)	+	0.1693	0.0605
					(0.17)	(0.04)**
Stock changes(t)	–	−1.6939	Dj x Stock changes(t)	–	−0.988	−2.5684
		(0.00)***			(0.09)*	(0.00)***
			(1-Dj) x Stock changes(t)	–	−1.5651	−1.6834
					(0.04)**	(0.00)***
Stock changes(t) x GTR changes(t-1)	–	−5.8253	Dj x Stock changes(t) x GTR changes(t-1)	–	−3.1651	−2.4563
		(0.09)*			(0.04)**	(0.47)
			(1-Dj) x Stock changes(t) x GTR changes(t-1)	–	−2.48617	−6.6170
					(0.32)	(0.04)**
Number of Observations N =(ixT)		767	Number of Observations N =(ixT)		767	767
Arellano-Bond test AR(1) (p-value)		(0.04)**	Arellano-Bond test AR(1) (p-value)		(0.01)**	(0.03)**
Arellano-Bond test AR(2) (p-value)		(0.55)	Arellano-Bond test AR(2) (p-value)		(0.93)	(0.24)
Sargan test (p-value)		(0.99)	Sargan test (p-value)		(0.99)	(0.99)

Notes: The left hand side of the table (Model A) presents the Panel GMM estimation results for VIX changes regressed on stock market changes, google trends indicator (GTR) contemporaneously (t), and on cross terms of stock changes with GTR_t-1 measurement for the period from 02 January 2020 to 09 April 2020 across all countries studied. The column Exp. Sign refers to expected signs of the coefficient of the relevant explanatory variable. The right hand side of the table (Model B) reports the Panel GMM estimation results for changes on stock market VIX in Europe vs. Asia, regressed on cross terms of the dummy variables Dj and (1-Dj) with i) google trends indicator (GTR) contemporaneously (t), ii) stock changes contemporaneously (t) and finally iii) stock changes and google trends indicator (GTR) with a lag (t-1). By using dummy variables Dj, (1-Dj) we decompose the sample between j geographic area and the rest. The j takes the value of either EUR for European markets studied or Asia for Asian markets. The lower part of the table presents results for Arrelano-Bond tests concerning the hypothesis that the error term is not serially correlated and the results for Sargan test examining the overall validity of the instruments. Finally, p-values are in parentheses. *,** and ***indicate statistical significance at the 10%, 5%, and 1% level respectively.

Similarly, the right-hand side of Table 4 (Model B) reports the Panel GMM estimation results for changes on stock market VIX in Europe vs. Asia, regressed on cross terms of the dummy variables Dj and (1-Dj) with i) google trends indicator (GTR) contemporaneously (t), ii) stock changes contemporaneously (t) and finally iii) stock changes and google trends indicator (GTR) with a lag (t-1). By using dummy variables Dj, (1-Dj) we decompose the sample between j geographic area and the rest, as we did in the previous section. The j takes the value of either EUR for European markets studied or Asia for Asian markets. Again our findings underline the importance of GTR changes on VIX changes that is particularly present in European markets.

The lower part of the Table 4 presents results for Arrelano-Bond tests concerning the hypothesis that the error term is not serially correlated and the results for Sargan test examining the overall validity of the instruments. As can be seen, the Sargan test provides no evidence of misspecification, while the serial correlation tests point to first- but no second-order autocorrelation of the residuals, which is in accordance with the assumptions underlying the selection of the instruments.

Furthermore, in order to examine the results across different groups of stock markets using the PVAR model, we follow the same methodology of decomposing the effect of Google trend searching on the COVID-19 pandemic in European versus non-European markets and Asian versus non-Asian markets. Impulse response analysis (see Fig. 6, Fig. 7 ) confirm our previous findings that the effects are stronger for European markets versus the rest of the markets. Stock market returns respond negatively to an increase in searches about COVID-19 with higher statistical significance in the case of European versus the rest markets. Worth also mentioning, that the negative response of Google trend searching for COVID-19 due to a positive stock market return shock is mainly attributed to European markets. This implies a bi-directional relationship between stock market investing and searching about the consequences of a pandemic.

Fig. 6.

Impulse response analysis for PVAR model on European vs. non-European markets. Notes: This figure depicts the impulse response analysis between stock returns (stock_changes), volatility indices changes (VIX_changes) and Google trend metric (gtr_changes). gtr _eur refers to the Google trends metric on European markets and gtr_exeur on non-European markets (the rest markets in our sample). The variable after symbol “:” is the variable that responds to a positive shock in a variable before symbol “:”.

Fig. 7.

Impulse response analysis for PVAR model on Asian vs. non-Asian markets. Notes: This figure depicts the impulse response analysis between stock returns (stock_changes), volatility indices changes (VIX_changes), and Google trend metric (gtr_changes). gtr_asia refers to the Google trends metric on Asian markets and gtr _exasia on non-Asian markets (the rest markets in our sample). The variable after symbol “:” is the variable that responds to a positive shock in a variable before symbol “:”

Finally, in order to investigate further the robustness of our main findings concerning the indirect effect of the investors’ attention to COVID-19 on stock returns and implied volatility changes, we proceed to estimate the dynamic correlations in each country from a GARCH(1,1)-DCC model. Then, we estimate any positive relationship between absolute values of these correlations with our metric about COVID-19 attention. The estimates presented in Table 5 show that an increase in COVID-19 attention increases, in absolute terms, the correlation between stock returns and implied volatility changes for all groups of countries, confirming our previous findings. The fixed effect model is preferable to the pooled model. Moreover, the Hausman test supports the fixed effect versus the Random effects model.

Table 5.

Panel data estimation results for Google Trend metric on dynamic conditional correlation between VIX & Stock changes.

Dependent Variable		abs (correlation between VIX & Stock changes)
Independent Variables	Exp. Sign	Pooled with PCSE and AR(1)			Fixed Effects with PCSE and AR(1)
Model		j = All	j = EUR	j = Asia	j = All	j = EUR	j = Asia
C	+/−	0.03349	0.0345	0.0348	0.097461	0.1008	0.0978
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.01)**	(0.01)**
Dj x GTR changes(t-1)	+	0.026026	0.0296	0.0462	0.019484	0.0229	0.0210
		(0.02)**	(0.00)***	(0.00)***	(0.02)**	(0.01)**	(0.20)
(1-Dj) x GTR changes(t-1)	+		0.0204	0.0222		0.0011	0.0191
			(0.23)	(0.04)**		(0.94)	(0.04)**
ρj [AR(1) coeff.]	+/−	0.955131	0.9550	0.9488	0.861368	0.9139	0.7969
		(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***	(0.00)***
Country Effects		no	no	no	yes	yes	yes
R2		92.45%	92.59%	92.58%	93.80%	94.80%	93.90%
F Test		4681.83***	2394.23***	2392.78***	818.78***	881.25***	739.27***
Durbin-Watson stat.		2.13	2.14	2.14	2.03	2.04	2.04
N =(ixT)		767	767	767	767	767	767
Specification tests
Cross section F-test (pooled OLS vs. FEM) p-value		(0.00)***	(0.00)***	(0.00)***
Hausman test (FEM vs REM)					(0.00)***	(0.00)***	(0.00)***
Test of cross-sectional independence by Frees					(0.00)***	(0.00)***	(0.00)***
Modified Wald test for group wise heteroskedasticity					(0.00)***	(0.00)***	(0.00)***

Notes: This table presents the results for the relationship between Google Trend metric and the dynamic conditional correlation between VIX and stock changes for three country groups (All, Europe, Asia). The “abs” refers to the absolute values of the dynamic conditional correlation estimates from the estimated DCC-GARCH models. The independent variable regressed on cross terms of the dummy variables Dj and (1-Dj) with lagged google trends indicator (GTR_t-1). By using dummy variables Dj, (1-Dj) we decompose the sample between j geographic area and the rest. The j takes the value of either EUR for European markets studied, Asia for Asian markets or All for the markets of the whole sample. The pooled and fixed effects models with panel corrected standard errors (PCSE) for the period from 02 January 2020 to 09 April 2020. P-values are in parentheses. *,** and ***indicate statistical significance at the 10%, 5%, and 1% level respectively. The column Exp. Sign refers to expected signs of the coefficient of the relevant explanatory variable. Coefficient ρ is estimated assuming a common across countries autoregressive model [(AR(1)] for the residuals. Next panel of the table reports the adjusted R² coefficient, the F-test, the Durbin–Watson value and the number of observations [market units (i) x time observations (t)]. The lower part of the table presents specification tests for the models estimated. Initially pooled versus fixed effects are tested, and then a Hausman test is applied between fixed versus random effects models for cases where previously fixed effects models are selected versus pooled models. Finally, p-values of tests for cross sectional independence and group wise heteroscedasticity are reported.

6. Discussion

This study contributes to the growing empirical literature exploring the relationship between investors' attention, as proxied by Google search volume, and market volatility. Our empirical results show that investors' attention as proxied by Google's search volume adds information to a market-based measure of volatility – thus confirming the first research hypothesis (H1) – as the respective coefficients of Google searches are statistically significant under all specifications. This fact can be explained by the notion that as investors become more anxious regarding the pandemic, they search more and expect higher volatility of stock returns in the future. Existing empirical evidence (Liu, Dai, Huynh, Zhang, & Zhang, 2022; Sun, Wu, Zeng, & Peng, 2021) confirms that investors sentiment regarding COVID-19 includes predictability regarding stock returns, but we also show that investors' attention, as proxied by Google searches, also relates with equity market implied (or expected) volatility. The conclusion that high Google search volumes for COVID–19 increase stock market volatility aligns with Lyócsa et al. (2020) and Chundakkadan and Nedumparambil (2022) and complements their work as we employ implied (and not realized) volatility. Our findings are also consistent with Bai et al. (2021) who show that COVID-19 pandemic upwardly influenced risk perceptions.

Our empirical investigation also confirms the second hypothesis (H2), as we show that a positive shock on Google search for the COVID-19 pandemic has an immediate statistically significant negative effect on stock returns (which diminishes after two days). Since investors’ attention has a negative connotation during the outbreak of the pandemic, we expect that increased attention can be perceived as negative, thus leading to lower prices and negative returns.

Our study also contributes to the long-standing academic literature that examines the relationship between market volatility and equity returns. This relation can be studied either by using realized (historic)/ex-post volatility or implied/ex-ante volatility. Investors are mostly concerned about ex-ante risk, thus, using implied volatility (instead of historic volatility) provides a better gauge for determining stock returns. Examining this relationship during the outbreak of the pandemic is an appealing and unprecedented case to investigate. Our empirical findings show that the correlation between stock returns and implied volatility changes for all countries under review increased during the pandemic, thus confirming the third research hypothesis (H3), which suggests that the negative relationship between implied volatility and return is affected by the pandemic outbreak. This result is important since in periods of high expected volatility, stock returns become strongly mean reverting, and thus, perhaps attractive buying opportunities arise for those with long investment horizons (Moreira & Muir, 2019).

Regarding the fourth hypothesis (H4) concerning the effect of implied volatility on investors’ attention, our results show that although a positive shock on implied volatility increases Google search volume, the respective relationship is not statistically significant. On the contrary, a positive shock of stock returns significantly reduces the search rate for the COVID-19 pandemic. This means that market participants knowing that the stock market is a leading indicator of future economic activity and thus, of the possible economic consequences of the pandemic, interpret positive returns as an indication of milder economic consequences from this pandemic and reduce internet searching regarding COVID-19.

7. Concluding remarks

While the relationship between stock returns and implied stock market volatility is widely acknowledged and well-studied in existing academic literature, the outbreak of the COVID-19 pandemic represents a quite attractive and unprecedented case to re-investigate this relation. In this context, we study the effects of investors' attention on implied stock market volatility in thirteen countries across the globe during the first wave of the COVID-19 pandemic (January–April 2020). We show that direct effects of increased uncertainty due to COVID-19 contribute to increased stock market implied volatility, but we also find indirect effects, as we find that a positive shock on Google search for the COVID-19 has an immediate statistically significant negative effect on returns which induces further increases in implied volatility.

The level of a volatility index is derived directly from option prices and therefore reflects all positive or negative market-moving events that affect the underlying equity index. It is a volatility measure that is independent of any option pricing model, therefore representing the risk-neutral measure of the expected volatility, and thus incorporating both stock market uncertainty and investors’ risk aversion. As investors become more anxious and increase their Google searches, they also expect higher expected volatility in the underlying equity market. To secure protection against future potential losses, they are willing to pay higher premiums for out-of-the-money options, while they simultaneously buy out-of-the-money call options to participate in a potential stock price reversal. This leads to more expensive options and higher implied volatility.

Our results also support the attention-induced price pressure hypothesis, as we find that a positive shock of stock returns significantly reduces investors’ attention, as proxied by search volume for the COVID-19 pandemic. This means that investors consider equity index returns as an indication of milder economic consequences from the COVID-19 pandemic and thus reduce the respective internet searching. This finding is mainly dominant in the case of European markets. Therefore, our results complement previous studies which show that implied volatilities in the stock markets are affected by investor attention in a “Google” or “internet” based economy (Da et al., 2015).

Our findings also highlight an investor sentiment channel that provides useful insights to investors and policymakers. The pandemic led to negative equity returns not only because of the expected negative effects on economic activity and corporate profits, but also by adding to investors' uncertainty. Understanding the links between investors’ decisions during a pandemic crisis and asset price variability is critical for designing and implementing the policy measures needed in markets and economies, as the presence of fear among investors can lead to a direct decline in stock prices, and when panic arises, it intensifies the adverse effects of the pandemic on equity markets. Our findings show that policymakers should attempt to reduce the uncertainty associated with a global pandemic, aiming to positively influence the expectations of investors and businesses. This, in turn, could potentially alleviate the market declines caused by the repricing of variance risk.

However, it is important to recognize certain limitations inherent in the study. Firstly, the analysis is restricted to the first wave of the pandemic, which may limit the applicability of the findings to other phases or periods of the pandemic. Secondly, the study primarily focuses on the effects of attention on implied volatility and does not explicitly consider other potential drivers of market volatility, such as news events or macroeconomic indicators. Nevertheless, it is assumed that these factors are encompassed within the underlying equity indices returns. Moreover, the study relies exclusively on Google searches for COVID-19 as a proxy for capturing investors' attention, neglecting the availability of alternative sources, particularly in the realm of social media. This limitation suggests the necessity for further research to address these gaps and encompass a broader range of factors that can influence investor sentiment and market volatility throughout different stages of the pandemic. By incorporating additional data sources and exploring various forms of media, a more comprehensive analysis can be conducted to enhance our understanding of the relationship between attention, investor sentiment, and market volatility. Finally, further research might help shed light on the risk-taking monetary policy transmission channel (Delis, Hasan, & Mylonidis, 2017; Fassas & Papadamou, 2018) and how this may be affected by the pandemic which affects investors’ risk tolerance and attention, as measured by Google trend metrics. As there is an anticipation of a rise in the occurrence of pandemics, which includes the spread of contagious diseases, in the upcoming years, this would be beneficial for global financial stability and economic growth.

Declarations

No Conflict of Interest.

Availability of data and materials

Available upon request by the authors.

Funding

No Funding.

CRediT authorship contribution statement

Stephanos Papadamou: Conceptualization, Methodology, Software, Data curation, Writing – original draft, Writing – review & editing. Athanasios P. Fassas: Conceptualization, Methodology, Software, Data curation, Writing – original draft, Writing – review & editing. Dimitris Kenourgios: Data curation, Writing – original draft, Supervision. Dimitrios Dimitriou: Visualization, Investigation, Software, Validation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

The authors do not have permission to share data.