Media Attention vs. Sentiment as Drivers of Conditional Volatility Predictions: An Application to Brexit
Introduction
It is well understood that media coverage (a proxy for investors’ attention) and tone (a proxy for market sentiment) are important drivers of trading volumes, stock returns, and to some extent, also risk (see, among others, Broadstock and Zhang, 2019; Brown and Cliff, 2005; Da, Engelberg, and Gao, 2011; Tetlock, 2007). While the evidence on the predictive power of coverage and sentiment for stock returns and trading volume is pervasive, the available results concerning the connection between media attention and tone and standard measures of risk, such as conditional variance, were initially tenuous at best. For instance, Mitchell and Mulherin (1994) and Berry and Howe (1994) relate aggregate stock market volume and volatility to broad measures of news about firms and the economy. Both studies reported weak correlations of less than 0.12 between market volatility and the news items. However, Antweiler and Frank (2004) studied the frequency and tone of stock message-board posts on Yahoo! Finance and Raging Bull for 45 large US stocks in the year 2000. They find that message-board posting frequency predicts return volatility and trading volume, even controlling for the frequency of WSJ stories.
We perform analyses that to the best of our knowledge are new to test whether on-line media attention and sentiment have any predictive power for risk, measured as the conditional variance of returns of the FTSE 100 stock index (and, as a robustness check, of the FTSE 250 stock index). We base our investigation on the number of articles and the tone extracted from more than half a million on-line news items concerning the June 2016 Brexit referendum vote over a daily 2015-2018 sample, available from Global Database of Events, Language, and Tone (GDELT). Notably, we conduct our analysis within a classical Generalized Autoregressive Conditional Heteroskedasticity (GARCH) framework, i.e., we seek to measure whether and how the in- and out-sample predictive performance of off-the shelf GARCH can be improved by augmenting the model to include exogenous scores that reflect media coverage and tone.
Importantly, there is some theoretical backing for our empirical tests. For instance, Andrei and Hasler (2014) develop an asset-pricing model that derives a link between asset prices, as well as volatility, and investors' attention. In their model, the level of attention varies with economic conditions, and the model shows that higher attention and higher uncertainty are associated with higher stock return variance. This link between attention and volatility is supported by some empirical literature (see, e.g. Da, Engelberg, and Gao, 2011; Dimpfl and Jank, 2016; Guidolin et al., 2017). Paradoxically, we know less about the link between sentiment and equity risk, even though Da, Engelberg, and Gao (2015) construct a new measure of investor sentiment (the Financial and Economic Attitudes Revealed by Search, FEARS) index, by aggregating the volume of Google Trends queries related to concerns such as recession, unemployment and bankruptcy and report that FEARS predicts volatility. 1
In the perspective of showing the importance of attention and sentiment in forecasting the volatility of asset returns, the Brexit vote represents a highly relevant case for several reasons. Indeed, Brexit has undoubtfully raised concerns among investors (see Adesina, 2017), while Arshada, Rizvib, and Haroon (2020) find that the efficiency of the British stock market worsened drastically because of the uncertainty caused by the Brexit vote. Our key result is that while recent findings that media attention increases the conditional variance of asset returns are confirmed, it is mostly the (absolute value of the) tone of the on-line media flow that generates risk.2 Therefore, it is not just the presence of media chatter and coverage that generates risk, but also (or especially) the fact the news come to reflect strong views, with a specific tone that seems to affect stock prices more, thus offering precise forecasting power.
Section 2 describes our data and the alternative models that we propose to forecast conditional volatility. Section 3 discusses the empirical results. Section 4 concludes.
Section snippets
Data
Our source of data is the Global Database of Events, Language, and Tone (GDELT), which collects news from print, broadcast, and online media in more than 100 languages and from all over the world and analyze them using natural language processing algorithms.3 To construct the series employed in this paper we have relied on the Global Knowledge Graph tool (henceforth, GKG). The GKG is a large collection of press articles organized around a “nameset”, a set of
Methodology
The first specification that we propose is a GARCHX obtained by adding each of the two news-related series as exogenous variables to a standard GARCH:where Rt is the (daily) log return defined as is the value of the FTSE 100 index at time t, zt is a normally distributed random shock, and is the conditional variance at time t given the information set at time t − 1, xi = 1, t = tonet while
In-sample forecasting results
Table 2 shows the in-sample estimates and the measures of in-sample fit for the models discussed in Section 3. Notably, the models including the growth in the number of articles either in a symmetric or an asymmetric fashion fail to outperform a standard GARCH(1,1) model, in-sample.9
Conclusion
We propose alternative augmented GARCH models that include measures of the coverage and the tone of the media for a specific event, the 2016 Brexit referendum. We assess the predictions of these models vs. a standard GARCH. The models that include the media tone outperform GARCH in three specifications: when it is always considered (GARCHX), when it is considered above q threshold of previous predicted variance (GARCHND), and under an asymmetric effect related to the sign of the exogenous
Author's Statement (on Behalf of All Authors)
1. 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.
2. All authors have contributed in equal shares to all the parts of this submission.
3. The data have been duly shared and made available.
References (21)
Estimating volatility persistence under a Brexit-vote structural break
Finance Research Letters
(2017)- et al.
Social-media and intraday stock returns: The pricing power of sentiment
Finance Research Letters
(2019) - et al.
No news is good news: An asymmetric model of changing volatility in stock returns
Journal of Financial Economics
(1992) - et al.
The heterogeneous impact of Brexit: Early indications from the FTSE
European Economic Review
(2018) - et al.
Textual sentiment in finance: A survey of methods and models
International Review of Financial Analysis
(2014) - et al.
Does twitter predict Bitcoin?
Economics Letters
(2019) - et al.
Impact of Brexit vote on the London stock exchange: A sectorial analysis of its volatility and efficiency
Finance Research Letters
(2020) - et al.
Investor attention and stock market volatility
Review of Financial Studies
(2014) - et al.
Is all that talk just noise? The information content of Internet stock message boards
Journal of Finance
(2004) - et al.
Policy uncertainty and international financial markets: the case of Brexit
Applied Economics
(2018)
Cited by (7)
Google search trends and stock markets: Sentiment, attention or uncertainty?
2024, International Review of Financial AnalysisThe US banking crisis in 2023: Intraday attention and price variation of banks at risk
2023, Finance Research LettersMedia attention and corporate greenwashing behavior: Evidence from China
2023, Finance Research LettersEU Climate Change News Index: Forecasting EU ETS prices with online news
2023, Finance Research LettersIntrospecting predictability of market fear in Indian context during COVID-19 pandemic: An integrated approach of applied predictive modelling and explainable AI
2021, International Journal of Information Management Data InsightsCitation Excerpt :Liu et al. (2021) argued fear sentiment owing to COVID-19 pandemic augmented the possibility of market crash. In general media coverage, sentiments, etc. have been observed to possess some explanatory capacity in governing the volatility of financial markets Guidolin & Pedio (2021). The present work resorts to Google Search Volume Index (GSVI) (Swamy, Dharani & Takeda, 2019; Afkhami, Ghoddusi & Rafizadeh, 2021) on different keywords to extract the cognate sentiment effectively.
We thank three anonymous referee for constructive feedbacks and Alessandra Poli for excellent research assistance.
JEL codes: C53, C58, G17.