Does Investor Attention Foretell Stock Trading Activities? Evidence from Twitter Attention. Chen Gu and Denghui Chen

Transcription

1 Does Investor Attention Foretell Stock Trading Activities? Evidence from Twitter Attention Chen Gu and Denghui Chen First version: December, 2017 Current version: July, 2018 Abstract This paper investigates the impact of Twitter attention, measured by abnormal number of tweets on stock trading activities. We find that Twitter attention has predictive power for future stock volatility and trading volume. A heightened number of tweets is followed by high volatility and trading volume over the next trading day. This finding is robust when focusing on international markets and controlling for other attention measures. We also find that high Twitter attention strengthens stock price adjustment to recommendation changes whereas it alleviates post-announcement price drift. These findings suggest that market underreaction to new information is related to limited investor attention from social media. JEL classification: G12; G14; G15; G17 Keywords: Twitter Attention; Social Media; Volatility Predictability; Trading Volume; Analyst Recommendation; * Corresponding author. Economics Department, the Office of the Comptroller of the Currency. Tel: , Fax: denghui.chen@occ.treas.gov. Chen Gu is an assistant professor of finance at Shanghai Business School, Research Center of Finance. Denghui Chen is a financial economist at the Office of the Comptroller of the Currency. The views expressed in this paper are those of the authors alone and do not necessarily reflect those of the Office of the Comptroller of the Currency or the U.S. Department of the Treasury. The authors take responsibility for any errors.

2 1 1. Introduction Investor attention fluctuations impact stock market trading activities. Andrei and Hasler (2015) argue that high attention accelerates the investor learning process, thereby inducing high volatility in theory. Ben-Rephael, Da and Israelsen (2017) show empirically that firm-specific news searching and reading activity, a proxy for institutional investor attention, is closely related to volatility and volume. In the past decade, the use of social networks, such as Twitter and Facebook, has grown exponentially. With this rapid growth and significant adoption of social media among investors, this paper aims to provide insights into the effect of investor attention extracted from Twitter and consequently enhance our understanding on how attention affects stock markets. One main challenge to investigate the impact of firm-level attention from Twitter is to disentangle all relevant tweets related to specific firms. In our study, we construct an investor attention proxy, namely Twitter attention which is measured by the abnormal number of tweets associated with individual stocks classified by Bloomberg, and investigate and validate whether investor attention is associated with volatility and trading volume of individual stocks. We also examine whether investor attention facilitates the incorporation of new information into stock prices. Using firm-specific Twitter publication count data from Bloomberg, we find a comovement between Twitter attention and return volatility and trading volume at a daily frequency. More notably, social media coverage has predictive power for future market activities. A heightened number of tweets is followed by high volatility and trading volume over the next day. As the attention effect is also likely to be present in foreign stock markets, we further validate whether our results are consistent using FTSE 100, CAC 40, and DAX 30 component stocks. The results suggest that the effect of Twitter attention on stock volatility and volume persists among international stock markets. We also compare Twitter attention to various investor attention proxies in terms of forecasting volatility and trading volume. We find that the predictive power of Twitter attention is

3 2 not subsumed to that of news attention, institutional investor attention and volume based attention, and remains largely unchanged after controlling them. This finding supports the view that volatility and trading volume predictivity of Twitter attention is complementary to those of other attention measures. The intuition for the positive relationship between social media coverage and return volatility and trading volume can be understood as follows. The increasing influence of social media and decreasing threshold of information access have changed how investors receive and process financial information, thus an abnormal number of tweets is likely to provide unique proxies to observe collective investor attention. When investors pay high attention to particular stocks via social media, prices immediately adjust to information because learning is fast. Price volatility and trading volume are proportional to the rate of information flow (e.g., Tauchen and Pitts, 1983; Ross, 1989). Therefore, high attention induces high return volatility and high trading volume. In contrast, low investor attention gradually incorporates new information into prices, and thus results in low return volatility and low trading volume. We next investigate whether social media exposure facilitates incorporation of information into prices. We do so by examining the incremental impact of high Twitter attention on event day return and return drift following an analyst recommendation announcement. We find that the immediate responsiveness of stocks to recommendation changes is stronger when the Twitter attention associated with the stocks is high on the day of the recommendation change. In addition, well documented post-announcement price drift disappears when the recommendation changes are announced on high Twitter attention days. These findings together show that market underreaction to news about recommendation revision, to some extent, is related to limited investor attention on social media. This paper provides a further step towards understanding whether social media activity is associated with stock market activities, such as volatility and trading volume. The level of activity on

4 3 media platforms, such as Twitter, provides a measure of investor attention or general interest about the occurring real-time events for given stocks. Existing literature of Twitter attention mainly focus on tweets that reference stock indices. For instance, Mao, Wang, Wei and Liu (2012) find that the daily number of tweets that mention S&P 500 stocks is helpful to predict the stock price and traded volume. In a more recent study, Oliveira, Cortez, and Areal (2017) assess the value of Twitter data to forecast stock returns, volatility and trading volume of diverse indices and portfolios, and they find that Twitter data is helpful to predict stock market behavior. However, we still know little about how individual stocks react to firm-specific Twitter attention. Our study sheds new light on this issue. Our analysis can be considered an empirical application of Andrei and Hasler s (2015) theoretical framework and extends extensive empirical literature on limited investor attention (e.g. Dellavigna and Pollet, 2009; Hou, Peng, and Xiong, 2009; Boehmer and Wu, 2013). We provide evidence that attention generated from social networks, as a proxy for investor attention, leads to an increase in stock volatility and trading volume. In addition, Twitter attention facilitates the incorporation of information about analyst recommendation revisions into prices and alleviates postannouncement return drift. The paper proceeds as follows. Section 2 provides a review of the related literature. Section 3 introduces the key variables. Section 4 examines the relationships between Twitter Attention and Volatility and Volume. Section 5 investigates the impact of Twitter attention on price responses to analyst recommendation revisions. Section 6 concludes the paper. 2. Related Literature 2.1 Investor Attention and Stock Markets Existing literature has studied the relationship between investor attention and stock markets from various aspects. For example, Andrei and Hasler (2015) propose a theoretical framework and investigate whether investor attention and uncertainty are key determinants of asset prices. They find

5 4 that stock return variance and risk premia increase with both attention and uncertainty. In a more empirical setting, Yuan (2015) shows that market-wide attention-grabbing events predict the trading behavior of investors and market returns. In another event study literature, DellaVigna and Pollet (2009) compare the response to earnings announcements on Friday, when investor attention is likely to be lower, to the response on other weekdays. They find that investor inattention leads to a less immediate response and more drift for Friday announcements, and explain that post-earnings announcement drift is caused by underreaction to information resulting from limited attention. Hou, Peng, and Xiong (2009) find that price underreaction to earnings news decreases with investor attention, but price continuation caused by investors overreaction increases with attention. Moreover, Chen, Liu, Lu, and Tang (2016) examine the role of investor attention in scheduled macroeconomic announcements, and observe the reactions of futures prices to CPI announcements are stronger to bad CPI news, more sensitive in high-inflation periods, and less pronounced on Fridays. The investor attention from these studies are related to news announcements (e.g., earnings and macroeconomic announcements) and typically available at low frequencies (weekly or monthly). Recently, a growing body of empirical literature has been using internet search activity as a proxy for investor attention, and it enhances researchers ability to analyze the impact of investor attention on daily stock trading activities. For example, Da, Engelberg and Gao (2011) propose search volume in Google captures retail investor attention. Vlastakis and Markellos (2012) use the Google Trends database to study information demand and supply for 30 Dow Jones Industrial Average Index component stocks. They find that demand for information at the firm level is positively related to historical and implied volatility and trading volume, even after controlling for market return and information supply. Dimpfl and Jank (2016) further report that internet research queries improve volatility forecasting at the index level. Some international evidence from Aouadi, Arouri, and Teulon (2013) shows that Google search volume is a reliable proxy of investor attention, and that investor attention is strongly correlated with trading volume and is a significant determinant

6 5 of illiquidity and volatility in the French stock market. Ben-Rephael et al. (2017) find that institutional attention responds more quickly to major news relative to retail investor attention based on Google search activity. 2.2 Social Media and Stock Markets Investors have increasingly relied on information from Twitter and other social media when making trading decisions. Existing literature on Twitter can be divided into two strands. The first strand analyzes whether qualitative information extracted from tweets influences stock returns. For example, Bollen et al. (2010) find Twitter mood is correlated to the value of the Dow Jones Industrial Average (DJIA) over time, and the accuracy of DJIA predictions can be significantly improved by the inclusion of specific public mood dimensions. Mao, Counts, and Bollen (2015) develop an indicator of online investor sentiment using Twitter updates and Google search queries, and find that Twitter and Google bullishness are positively correlated to investor sentiment and high Twitter bullishness predicts increases in stock returns. Bartov, Faurel and Mohanram (2018) find that the aggregate opinion in individual tweets can predict the forthcoming quarterly earnings of a company, and there is a positive association between the aggregate opinion and the immediate abnormal stock price reaction to the quarterly earnings announcement. Gu and Kurov (2018) propose that firm-level Twitter sentiment contains information about fundamentals, which is gradually incorporated into prices. Thus, Twitter sentiment has return predictability. The second strand of Twitter literature examines levels of social media coverage and trading activities. For example, Mao, Wang, Wei and Liu (2012) find that the daily number of tweets is correlated with certain stock market indicators at the stock market level, industry sector level, and individual company level (only Apple Inc. examined). Our study is related to this strand of literature. Instead of market and industry level, we focus on the more granular level and examine the effect of firm-level Twitter attention on stock return volatility and trading volume. In addition, we show that

7 6 Twitter attention facilitates the incorporation of information about recommendation revisions into stock prices. 3. Data and Sample Selection 3.1 Twitter Attention We download the number of tweets that reference publicly traded companies from Bloomberg. Bloomberg counts tweets in real time and reports the daily firm-level Twitter publication count every morning about 10 minutes before the U.S. stock market opens. The number reflects the level of social media exposure during the past 24 hours, from 9:20 a.m. the previous day to 9:20 a.m. on the current day, for a given firm. Thus, the Tuesday morning count indicates social media activities from 9:20 a.m. Monday to 9:20 a.m. Tuesday. Bloomberg gathers tweets from Twitter and StockTwits and applies a Ticker rule to identify whether a tweet is relevant for a given company. For example, if a tweet contains #AMZN from Twitter or contains $AMZN from StockTwits, it is classified as being related to Amazon.com, Inc. 1 The ticker rule allows Bloomberg to focus on messages related to firms financial information and to better capture investor attention. 2 We define daily Twitter attention (ATA) as the ratio between stock i s current day s number of tweets and its maximum daily number of tweets in the previous 10 trading days. 3 This is used to investigate the linear relationships between Twitter attention and volatility and abnormal trading volume. Hereafter, we use ATA i,t to represent stock i s Twitter attention between 9:20 a.m. on day t and 9:20 a.m. on day t+1. Moreover, we create a dummy version of Twitter attention, D ATA i,t, which is 1 Twitter introduces prefix # to group messages around a specific topic. StockTwits, which is a financial communications platform for the financial and investing community, use prefix $ to allow its users to identify the content of their post. 2 Da, Engelberg, and Gao (2011) argue that Google search frequencies with ticker symbols are less ambiguous than search frequencies with full company names in terms of capturing investor attention on given firms. 3 Our main results are robust when Twitter attention is defined as the ratio between stock i s current day s number of tweets and its average daily number of tweets in the previous 10 trading days. Our main results are also robust to the length of the rolling window (5 days, 15 days, 20 days).

8 7 equal to 1 when ATA i,t is greater than 1, and 0 otherwise. This variable captures the impact of high Twitter attention. Both attention variables are of our interest and serve different empirical purposes Volatility and Trading Volume Following Vipul and Jacob (2007), we use the Rogers and Satchell (1991) extreme value volatility estimator to estimate historical firm-level daily volatility (VOL). This approach is widely used in the volatility literature and doesn t rely on high-frequency data. The estimator is computed as follows: Volatility i,t = (H i,t C i,t )(H i,t O i,t )+ (L i,t C i,t )(L i,t O i,t ), (1) where H i,t, L i,t, O i,t and C i,t are the log-transformed highest, lowest, opening and closing prices of stock i on day t. We use trading volume during regular trading hours from Bloomberg to calculate daily abnormal trading volume (AV). Daily AV is computed by dividing the difference between volume for stock i on day t and the mean volume for stock i in the previous 10 trading days by the mean of volume for stock i in the previous 10 trading days. Both abnormal trading volume and return volatility are expressed in percentage points Sample Construction Our sample includes all Russell 3000 stocks and their corresponding daily number of Twitter publications from Bloomberg. Bloomberg started to release Twitter publication count for publicly traded companies at the beginning of Based on data availability, our sample period is from January 2015 to December Some stocks may have few tweets during a certain period. This generates spurious outliers of the Twitter attention measure. Thus, we drop stock-day observations if the daily average number of tweets is less than 1 in the past 10 trading days for a given firm. For instance, if Tesla s average tweet count is less than 1 between t-10 and t-1, its Twitter attention on

9 8 day t would be excluded for our analysis. These conditions lead to the full sample of our analysis, which contains 1,132,574 stock-day observations. We download analyst recommendation data from I/B/E/S. Then we construct a recommendation subsample by using filters in Jegadeesh and Kim (2013) and Ben-Rephael, Da and Israelsen (2017) The criteria are: (1) at least one analyst issues a recommendation for the stock within 180 calendar days; (2) there should be at least two active recommendations for the stock as of the day before the revision; (3) a recommendation is active for up to 180 days after issuance; (4) there should be one recommendation change for one firm during one recommendation revision day. Using timestamp from I/B/E/S, we define recommendation revision day t starts from 9:20 a.m. on day t-1 and ends at 9:20 am on day t, so Twitter attention and recommendation changes have same calendars. Our final recommendation subsample contains 4,378 recommendation changes. 3.4 Descriptive Statistics Table 1 provides summary statistics of our full sample and the subsample used for recommendation change analysis. Panel A shows that the mean ATA is 0.45 for the full sample, but it is 1.09 for the RecChng sample, which indicates that recommendation releasing days are associated with larger Twitter attention. A similar pattern can be observed for VOL and AV when comparing the average values of these variables in the full sample with the RecChng sample. Notably, the abnormal trading volume spikes about 7 times on days when recommendation changes occur. [Insert Table 1 Here] Panel B shows that the average of Twitter attention dummy is 8.77% in the full sample. We can interpret as 8.77% of all stock-trading days in our sample experiences high Twitter attention. High Twitter attention day frequency increases to 29.20% for the RecChng sample suggesting that 29.20% of the recommendation changes coincide with high Twitter attention days. Panel C shows that both VOL and AV are substantially larger on days of high Twitter attention compared to the days

10 9 when Twitter attention is low. For example, the full sample average volatility during low attention days (D ATA i,t =0) is 1.75% while the average volatility during high attention days (D ATA i,t =1) is 2.19%. 4. Empirical Results This section shows the effect of Twitter attention on stock volatility and trading volume. We also compare the Twitter attention to various attention proxies. Finally, we test the role that social media exposure plays in international stock markets. 4.1 Twitter Attention, Stock Volatility, and Abnormal Trading Volume Univariate Test Results We first use a univariate test to examine the relationship between firm-specific Twitter attention and their daily stock volatilities and abnormal trading volume. We begin by sorting stocks into quintiles based on Twitter attention every trading day. We then examine the stock s average current and next day volatility and abnormal trading volume across varying quintiles. [Insert Table 2 here] Panel A of Table 2 shows the univariate test results for volatility. We sort stocks into quintiles based on daily Twitter attention where Q1 is the quintile of stocks with the lowest Twitter attention while Q5 is the quintile of stocks with the highest Twitter attention. Column 1 shows the concurrent relationship between Twitter attention and daily volatility. As seen in the column, average volatilities increase across increasing Twitter attention quintiles in general. Average daily volatility for stocks in the quintile (Q1) with the lowest Twitter attention is 1.65%, and the average is 1.98% in Q5 with the highest Twitter attention. The volatility difference between high (Q5) and low (Q1) Twitter attention quintiles is 0.33% and statistically significant. In economic terms, the difference is approximately equal to 18.4% (0.33%/1.79%) of the sample daily average volatility reported in Panel A of Table 1. A similar pattern can be observed when investigating the relationship between Twitter

11 10 attention and volatility over the following trading day. The difference of average next day volatility between Q5 and Q1 reported in Column 2 of Panel A is 0.38%, and the difference is statistically significant. Panel B of Table 2 reports the daily abnormal trading volume sorted by quintile Twitter attention. We can observe that both current and next trading volume increase monotonically with Twitter attention. High Twitter attention quintile (Q5) is associated with abnormally high trading volume, while low Twitter attention quintile (Q1) is related to abnormally low trading volume. In addition, differences between the highest Twitter attention quintile (Q5) and lowest Twitter attention quintile (Q1) are statistically significant. This suggests that stocks are more actively traded when they are animatedly discussed on Twitter. Overall, Table 2 provides some preliminary evidence that social media exposures affect stock market trading activities Multivariate Test Results Univariate tests do not consider other factors, such as volatility clustering, which might influence the relationship between Twitter attention and return volatility and trading volume. Thus, we next explore the Twitter attention effect in a multivariate setting. We do so by applying Fama and MacBeth (1973) type of regressions. Specifically, we first carry out time-series regressions for each firm, and then report the cross-sectional averages of the coefficient estimates from the first-stage regressions and the corresponding Newey-West (1987) t-statistics. 4 The model specification is: Y i,t+1 = a + bx i,t + cvol i,t + dav i,t + ε i,t+1, (2) where Y i,t+1 represents VOL i,t+1 and AV i,t+1 in separate model specifications in Table 3, X i,t represents D ATA i,t or ATA i,t. The coefficient estimate on D ATA i,t reflects abnormal trading activities on high Twitter attention days relative to low Twitter attention days, while ATA i,t is used to capture the 4 As a robustness check, we run cross-sectional regressions for each period first, and then report the average of the timeseries coefficient estimates. The results, available upon request, are qualitatively similar to those reported in the paper.

12 11 changes in volatility and trading volume when Twitter attention changes by one unit. As discussed above, ATA i,t and D ATA i,t are based on tweets between 9:20 a.m. on day t to 9:20 a.m. on day t+1. The coefficient estimates on ATA i,t and D ATA i,t are our main parameter of interest. The model controls for lagged value of volatility (VOL i,t ) and trading volume (AV i,t ). Including these controls in the regression is important since volatility and trading volume autocorrelation in conjunction with contemporaneously correlated volatility and trading volume and attention can generate spurious evidence of lead-lag relation. [Insert Table 3 here] Panel A of Table 3 presents the multivariate regression results for the effect of Twitter attention on volatility. Column 1 shows a statistically significant coefficient of 0.54 on the high Twitter attention dummy variable. The result suggests that stocks that experience high Twitter attention on the previous day are more volatile than stocks that have low Twitter attention on the prior day. Column 2 reports the linear relationship between Twitter attention and volatility. It shows that Twitter attention can predict next trading day return volatility. A one unit increase in Twitter attention leads to a 0.26% increase in next trading day volatility. Finally, as one would expect, lagged values of volatility and volume both exhibit strong explanatory power for future volatility. Trading volume is another well-known measure of trading activity. In Panel B, we investigate the relationship between firm-specific Twitter attention and abnormal trading volume. Comparing to the results in Panel A, we observe similar results for trading volume. Both dummy and continuous measures of Twitter attention have a positive significant impact on daily abnormal trading volume. Column 1 indicates that trading volume is 55% higher than the average level in the previous two weeks when Twitter attention is high. Column 2 shows that one unit increase in Twitter attention is related to a 21% increase in abnormal trading volume over the next trading day. Overall, the results in Table 3 support our univariate tests in Table 2, and suggest that social media coverage is a driver of stock volatility and abnormal trading volume.

13 Twitter Attention vs. Other Investor Attention Measures Various existing literature has studied the role of investor attention on stock trading activities. For example, Ben-Rephael, Da and Israelsen (2017) argue that news searching and reading activity levels in Bloomberg terminals reflect institutional investor attention, since Bloomberg service subscribers are mostly institutional investors; Mitchell and Mulherin (1994) use number of news articles published in traditional media to measure investor attention; Loh (2010) proposes an attention measure based on trading volume. In this section, we empirically compare our Twitter attention with other attention measures in terms of forecasting volatility and trading volume. This comparison allows us to test whether Twitter attention provides innovative forecasting power relative to existing attention measures. In addition, given that people have spent substantial amounts of time on social media in recent years, it is of interest to explore whether the effect of Twitter attention on trading activity is stronger than traditional media based attention measures. We compare our Twitter attention dummy (D ATA i,t ) with three other dummy versions of ANA attention measures. is a news attention dummy variable which is equal to 1 when stock i s current day s number of news articles published on Bloomberg is greater than its maximum daily number of tweets in the previous 10 trading days. D AIA i,t is the dummy version of Ben-Rephael, Da and Israelsen (2017) institutional investor attention; D AVA i,t is a trading volume based attention dummy variable which is equal to 1 when the daily volume for stock i in day t is greater than the mean daily volume for stock i in the previous 10 trading days. We use dummy versions of attention instead of continuous versions due to three reasons. Firstly, the Bloomberg institutional investor attention has no continuous values. Secondly, we are most interested in the effect of high attention. Dummy variables make interpretation easier. Thirdly, the relationship between attention and trading activity may not be linear. In this aspect, attention dummies can be less criticized than continuous attention measures.

14 13 [Insert Table 4 Here] Table 4 provides the correlations across investor attention measures. We can observe that Twitter attention is positively correlated with all other three attention measures. But the correlations are not very high. The largest correlation is with institutional investor attention at 0.22, while the lowest correlation is with trading volume based attention at It suggests that Twitter attention is different from other attention measures and provides innovative volatility and trading volume predictability. We further explore the effects and statistical differences of different investor attentions in Table 5. The model specification is similar to Equation (2). Y i,t+1 = a + bd ATA i,t + cd K i,t + ε i,t+1, (3) where Y i,t+1 represents daily volatility (VOL i,t ) in panel A and abnormal trading volume (AV i,t ) in panel B. In the first 7 columns, the regressions are univariate or bivariate that contains Twitter attention measure (D ATA i,t ) and one of the three alternative investor attention measures, including new attention (D ANA i,t ), institutional investor attention (D AIA i,t ) and trading volume based attention (D AVA i,t ). The regressions in column (8) include all attention measures. [Insert Table 5 Here] Panel A of Table 5 presents the regression results of the impact of different investor attention measures on stock volatility. As a benchmark, Column 1 shows that Twitter attention predicts next day volatility. It has a coefficient estimate of 0.68 and an R 2 of 3.2%. In Column 2, news attention has a coefficient estimate of 0.48 and an R 2 of 1.9%. It suggests that new attention is a good positive volatility predictor, but its predictivity is lower than that of Twitter attention. In Column 3, the bivariate model reports the joint forecasting power of Twitter attention and news attention. The coefficient estimates on the two attention measures are still positive and statistically significant. The R 2 of the bivariate model is only slightly below the sum of univariate regression R 2 reported in Column 1 and Column 2. These findings are consistent with the low correlation between Twitter attention and news attention documented in Table 4.

15 14 Column 4 reports the power of Ben-Rephael, Da and Israelsen (2017) institutional investor attention in terms of forecasting volatility. As seen in the column, the institutional investor attention can be used to predict next day volatility. The univariate regression generates an R 2 of 1.9%, which is equal to that of news attention, but smaller than that of Twitter attention. In Column 5, we run the model including both the Twitter attention and the institutional investor attention. The bivariate regression shows that including the institutional investor attention has little impact on the statistical and economic significance of Twitter attention. It indicates that institutional investor attention doesn t dominate Twitter attention. Column 6 shows that volume based attention has significant explanatory power for volatility over the next trading day. The corresponding R 2 of 2.3% is higher than those of news attention and institutional investor attention, but lower than that of Twitter attention. Column 7 documents that the volatility predictivity of Twitter attention survives after controlling for volume based attention. In the last column, we run a kitchen-sink regression that includes all attention proxies in one model. We find that Twitter attention has the largest coefficient estimate and remains statistically significant. All attention proxies together explain 6.9% of variation in volatility. Panel B of Table 5 reports volume predictability of the four attention proxies. Univariate predictive regressions (Column 1, 2, 4, 6) show that all attention measures are good trading volume predictors. The magnitude of coefficient estimates suggests that the differential impact of high vs. low Twitter attention on trading volume is larger than that of other attention measures. Twitter attention and volume based attention both have an R 2 of 3.9%, which is larger than those of news attention and institutional investor attention. Bivariate predictive regressions (Column 3, 5, 7) show that volume predictivity of Twitter attention is not subsumed to that of news attention, institutional investor attention and volume based attention, respectively. Finally, the kitchen-sink regression (Column 8) shows that Twitter attention survives in the model containing all attention proxies.

16 15 In short, our findings suggest that Twitter attention provides incremental power beyond other attention measures in forecasting volatility and trading volume. 4.3 International Evidence Non-US listed stocks are also frequently discussed on social media, such as Twitter and Facebook. Thus, we would expect the attention effects reported in section 4.1 to be persistent globally. In other words, we expect Twitter attention to have a significant and positive impact on volatility and abnormal volume for stocks listed on non-us stock exchanges. In this subsection, we test this conjecture by examining the effect of Twitter attention on FTSE 100, CAC 40, and DAX 30 component stocks. [Insert Table 6 Here] Table 6 presents the effect of Twitter attention on stock volatility and trading volume in UK, ATA France, and Germany blue-chip stock markets. Panel A shows that the coefficient estimates of and ATA i,t are positive and statistically significant for all three foreign stock exchanges when the dependent variable is next day volatility. The international results are consistent with our findings based on the U.S. stock market and suggest that high Twitter attention would predict a volatility spike in the next day for large firms listed on stock exchanges in Britain, France, and Germany. However, compared to the U.S. based results reported in Table 3, we find the overall impact of Twitter attention is relatively weaker on foreign stock markets than on the U.S. stock market. For example, a one unit change in attention is associated with a 0.54% change in volatility for Russell 3000 stocks, while FTSE 100 stocks only change by 0.07% after a same level change in attention. This is not surprising since Twitter is more popular in the US than in European countries. Panel B shows that trading volume predictability of Twitter attention is statistically and economically significant in the international stock markets. Overall, Table 6 results suggest that the impact of Twitter attention on stock markets is not limited to the U.S.

17 Various Rolling Windows In above sections, we measure Twitter attention using the ratio between the current day number of tweets about stock i and its maximum number of tweets in the previous 10 trading days. This section provides robustness checks on the effect of Twitter attention on volatility and trading volume with 5-, 15-, and 20-day rolling windows. [Insert Table 7 Here] Table 7 presents the multivariate regression results of the effect of Twitter attention on stock volatility and trading volume with three different rolling windows. Panel A shows that the coefficient estimates on Twitter attention variables are consistently positive and statistically significant when predicting volatility. Panel B shows the effect of Twitter attention on abnormal trading volume remains largely unchanged after changes in rolling windows. Overall, Table 7 shows that the forecasting power of Twitter attention is robust to the length of the rolling window. 5. Twitter Attention and Price Response to Analyst Recommendation Revision Extensive literature finds that analyst recommendations contain valuable information that moves market prices (e.g., Womack, 1996; Jegadeesh et al., 2004). Womack (1996) provides evidence of price drift after announcement. The return drift reflects the on-going digesting of the information about the recommendation changes after the day it has been released and suggests that the market underreacts to recommendation revisions on announcement days. In this section, we examine the impact of high social media attention on the incorporation of information about analyst recommendation into prices. If high social media exposure facilitates the information incorporation process, we would expect a stronger immediate market reaction and weaker post-announcement price drift when recommendation revision days coincide with high Twitter attention days. The regression specification is:

18 17 Return h ATA i /H = a + b 1 RecChng i,t + b 2 + ATA b3 RecChng i,t + Controlt 1,t 5 + ε i,t+h, (4) where Return i h is cumulative open-open excess return for stock i over period h. H represents the length of period h. Excess returns are defined as the residuals from a regression involving an individual firm s open-to-open log return against Russell 3000 index open-to-open log return. 5 RecChng i,t is stock i s recommendation change occurring between 9:20 AM on day t-1 and 9:20 AM on day t. Thus, if period h is from market open on day t-1 to market open day t, we investigate the impact of Twitter attention on announcement day returns. We focus on three periods returns: the announcement day return (h=(open t-1, open t)) and two post-announcement returns: average daily return over a month after the announcement (h=(open t, open t+22)) and average daily return over a quarter after the announcement (h=(open t, open t+63)). Our main variable of interest in this model is the interaction term between RecChng i,t and D ATA i,t (RecChng i,t D ATA i,t ). The coefficient on ATA RecChng i,t identifies the incremental impact of having high Twitter attention. We obtain analyst recommendation data from I/B/E/S. I/B/E/S recommendation score ranges from 1 to 5, where 1 represents a strong buy and 5 represents a strong sell. For interpretational convenience, we reverse I/B/E/S numerical scores so that a positive recommendation revision (RecChng i,t > 0) corresponds to an upgrade and a negative recommendation revision (RecChng i,t < 0) corresponds to a downgrade. Figure 1 depicts the frequency of high Twitter attention days (D ATA i,t = 1) between 5 days before and after analyst recommendation revision days. The frequency is simply equal to the daily cross-section mean of the Twitter attention dummy. As seen in the figure, high Twitter attention days are more likely to occur as announcement days approach. The frequency of high attention days increases from about 10% 5 days prior to announcements to about 30 % on announcement days. 5 Regression results are robust when excess returns are defined as the residuals of the Fama-French-Carhart four-factor model. However, using risk factors based on close-to-close returns to risk-adjust open-to-open returns may lead to biased excess returns.

19 18 Moreover, on the first day after the recommendation revision, the frequency decreases to below 20%. Afterwards, the frequency is lower than the sample average level. Overall, the pattern of frequency of high Twitter attention days confirms the view that recommendation revision events capture attention from social media. [Insert Figure 1 Here] Next, we test the effect of Twitter attention on announcement returns and post-announcement drifts. Following Ben-Rephael et al (2017), we estimate the regression in equation (5) using pooled OLS with standard errors clustered by day and firm. We also control for the stock s cumulative return from day t-5 to t-1 (Return t 1,t 5 ), the log of the stock s average market capitalization from day t-5 to t-1 (Size t 1,t 5 ), the log of the stock s average trading volume from day t-5 to t-1 (Volume t 1,t 5 ), standard deviation of the stock s daily return between day t-5 and t-1 (Std t 1,t 5 ) and the stock s average bid-ask spread in percentage from day t-5 to t-1 (Spread t 1,t 5 ). Column 1 of Table 8 reports market reaction to recommendation changes while Columns 2 and 3 focus on post-announcement periods. In all three columns, RecChng i,t has a positive and significant coefficient estimate. It suggests that stock prices react positively to recommendation revisions on announcement days, but the initial market reaction is incomplete. Stock prices continue to move in the right direction in response to RecChng i,t over 21 and 63 trading days after the announcement, confirming the existence of post-recommendation announcement drift. ATA More interestingly, the coefficient estimates on RecChng i,t are significantly positive (1.59) for the announcement return and significantly negative (-0.03 and -0.02) when dependent variables are post-recommendation announcement returns. This pattern suggests that high Twitter attention makes the market response to recommendation changes stronger and alleviates initial underreaction to news. Moreover, the sum of coefficients on RecChng i,t D ATA i,t and RecChng i,t for price drift regression is insignificant. Thus, when Twitter attention is high, there is no postrecommendation announcement drift. Overall, these results together are consistent with our

20 19 conjecture that Twitter attention facilitates incorporation of information about recommendation changes into prices. 6 [Insert Table 8 Here] 6. Conclusion Social media has increasingly become a major channel to disseminate financial information. However, the challenge to disentangle social media content related to specific firms has hindered researchers ability to analyze the impact of investor attention from social media on stock trading activities. In this study, we empirically investigate and validate whether social media is associated with volatility and volume and consequently facilitates incorporation of new information into stock prices. We observe Twitter attention has predictive power for future market volatility and trading volume. A heightened number of tweets is followed by high volatility and trading volume over the next day. The effect is complementary to that of other existing attention measures. Moreover, the finding is not U.S specific. We find similar evidence based on international stock markets, such as FTSE 100, CAC 40, and DAX 30. We next investigate whether social media exposure facilitates incorporation of information into prices. We do so by examining the incremental impact of high Twitter attention on event day return and return drift for analyst recommendation announcements. We find that the immediate responsiveness of stocks to recommendation changes is stronger and well documented postrecommendation announcement return drift disappears when the recommendation changes are announced on high Twitter attention days. These findings together show that market underreaction to news on announcement days, to some extent, is related to limited investor attention on social media. 6 We admit that high announcement returns could cause high Twitter attention on announcement days. In untabulated results, we find that when investors pay high attention via social media on the day before an announcement, the stock s price reaction is stronger. In addition, we find that Twitter attention makes announcement day stronger for recommendations released during after-hours. These checks largely alleviated concerns about potential reverse causality.

21 20 This paper provides a further step towards understanding the role of social media activity in stock markets and has important implications for market participants. The predictive power of Twitter attention for volatility and trading volume suggests that investor trading behaviors are influenced by the information flows generated from social media. The role of Twitter attention in facilitating the incorporation of information about recommendation revisions also indicates that social media makes stock markets more efficient.

22 21 References Andrei, D. and M. Hasler., Investor Attention and Stock Market Volatility, Review of Financial Studies, 28, Aouadi, A., Arouri, M., and Teulon, F., Investor attention and stock market activity: Evidence from France, Economic Modelling, 35, issue C, Bartov E, Faurel L, and Mohanram P., 2018, Can Twitter Help Predict Firm-Level Earnings and Stock Returns? Accounting Review, 93, Ben-Rephael, A., Da, Z., D. Israelsen, R., It Depends on Where You Search: Institutional Investor Attention and Underreaction to News. Review of Financial Studies Boehmer, E., J.J. Wu., Short selling and the price discovery process. Review of Financial Studies 26, Bollen, J., Mao, H., Zeng, X., Twitter Mood Predicts the Stock Market. Journal of Computational Science. 2, 1-8. Chen, J., Liu, Y., Lu, L., Tang, Y., 2016, Investor Attention and Macroeconomic News Announcements: Evidence from Stock Index Futures. Journal of Futures Markets, 36, Da, Z., Engelberg, J., Gao, P., In search of attention. Journal of Finance, 66, DellaVigna, S., Pollet, J. M., Investor inattention and Friday earnings announcements. Journal of Finance 64, Dimpfl, T., Jank, S., Can Internet Search Queries Help to Predict Stock Market Volatility? European Financial Management, 22, Fama, E., MacBeth. J., Risk, Return, and Equilibrium: Empirical Tests. Journal of Political Economy, 81, Gu, C., Kurov, A., Informational Role of Social Media: Evidence from Twitter Sentiment. Working paper. Hou, K., Peng, L., Xiong, W., A tale of two anomalies: The implications of investor attention for price and earnings momentum. Working Paper. Jegadeesh, N., Kim, J., Krische, S. D., Lee, C., Analyzing the analysts: When do recommendations add value? Journal of Finance, 59, Loh, R. K., Investor Inattention and the Underreaction to Stock Recommendations. Financial Management, 39, Mao, H, Counts S., Bollen J., Quantifying the effects of online bullishness on international financial markets. ECB working paper. Mao, Y., Wang, B., Wei, W., Liu B., Correlating S&P 500 Stocks with Twitter Data. Working Paper.

23 22 Mitchell, M. L., Mulherin, J. H., The Impact of Public Information on the Stock Market. The Journal of Finance, 49, Newey, W., West, K., A Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. Econometrica, 55, Oliveira, Cortez, Areal The impact of microblogging data for stock market prediction: Using Twitter to predict returns, volatility, trading volume and survey sentiment indices. Expert Systems with Applications, 73, Ross, S. A., Information and volatility: The no-arbitrage martingale approach to timing and resolution irrelevancy. Journal of Finance, 44, Rogers, L. C., Satchell, S. E., Estimating Variance from High, Low and Closing Prices. The Annals of Applied Probability, 1, Tauchen, G. E., Pitts, M., The Price Variability-Volume Relationship on Speculative Markets, Econometrica, 51, Vlastakis, N., Markellos, R. N., Information demand and stock market volatility. Journal of Banking & Finance 36, Vipul, V., Jacob, J., Forecasting performance of extreme value volatility estimators. Journal of Futures Markets, 27, Womack, K. L., Do brokerage analysts' recommendations have investment value? Journal of Finance, 51, Yuan, Y., Market-Wide Attention, Trading, and Stock Returns. Journal of Financial Economics, 116,

24 23 Table 1 Summary Statistics The table reports the summary statistics of firm-level Twitter attention measure (ATA), daily trading volatility (VOL) and abnormal trading volume (AV) from January 2015-December ATA is defined as the ratio between stock i s current day s number of tweets and its maximum daily number of tweets in the previous 10 trading days. D ATA is a dummy variable which is equal to 1 when ATA is greater than 1. AV is obtained by dividing the difference between volume for stock i on day t and the mean volume for stock i in the previous 10 trading days by the mean of volume for stock i in the previous 10 trading days. VOL is volatility, which is estimated based on Rogers and Satchell s (1991) approach. Our initial sample includes all Russell 3000 stocks with a daily number of Twitter publications from Bloomberg. We report results for the full sample (Full Sample) and the analyst recommendation changes sample (RecChng Sample). The Full Sample includes 1,070,505 day stock observations, and the RecChng Sample includes 4,075 stock observations. Panel A. Basic Summary Statistics for key variables Full Sample RecChng Sample Mean Std Mean Std ATA VOL AV Panel B. Sample Averages of D ATA Full Sample RecChng Sample % total obs Panel C. Sample Averages conditioning on D ATA Full Sample RecChng Sample D ATA =0 D ATA =1 D ATA =0 D ATA =1 VOL AV

25 24 Table 2 Univariate Test-Relation between Twitter Attention and Volatility and Trading Volume Panel A shows the relationship between firm-specific Twitter attention (ATA) sorted into quintiles at daily basis and current day volatility (Column 1) and next day volatility (Column 2). ATA is defined as the ratio between stock i s current day s number of tweets and its maximum daily number of tweets in the previous 10 trading days. Daily Volatility (VOL) is estimated based on Rogers and Satchell s (1991) approach. Panel B shows the relation between firm-specific Twitter attention sorted into quintiles at daily basis and current day abnormal trading volume (Column 1) and next day abnormal trading volume (Column 2). Daily abnormal volume (AV) is obtained by dividing the difference between daily volume for stock i in day t and the mean daily volume for stock i in the previous 10 trading days by the mean of daily volume for stock i in the previous 10 trading days. Q1 is the quintile of stocks with the lowest Twitter attention while Q5 is the quintile of stocks with the highest Twitter attention. The sample period is from January 2015 to December We report corresponding t-statistics in parentheses. ***, **,* represent statistical significance at 1%, 5%, and 10% respectively. Current Day (1) Next Day (2) Panel A: Daily Volatility Q Q Q Q Q Q5-Q (64.35)*** 0.38 (74.55)*** Panel B: Daily Abnormal Trading Volume Q (-11.11)*** (-16.77)*** Q (-19.45)*** (-9.71)*** Q (16.36)*** 1.68 (7.75)*** Q (41.25)*** 6.45 (29.15)*** Q (43.84)*** (66.35)*** Q5-Q (45.46)*** (69.09)***