ArticlePDF Available
When does attention matter? The effect of investor
attention on stock market volatility around news
releases
Daniele Ballinari1, Francesco Audrino1, and Fabio Sigrist2
1University of St.Gallen
2Lucerne University of Applied Sciences and Arts
This version: April 30, 2020
First version: December 19, 2019
Abstract
We empirically investigate how retail and institutional investor attention is re-
lated to the way stock markets process information. With a focus on 360 US
stocks in the S&P 500 universe, our results show that higher investors’ attention
around news releases is related to higher contemporaneous volatility. Further, re-
tail investor attention increases the post-announcement volatility, whereas insti-
tutional investor attention has a small but negative impact on volatility on days
following news releases on average over the cross-section of companies. These
findings are in line with the hypotheses that attention of retail investors slows
price-adjustments to new information and attention of institutional investors re-
sults in the opposite reaction. We show that these effects are heterogeneous in
the type of news and the topic of the information being released. A portfolio
allocation application highlights that these results are not only statistically sig-
nificant but also sizeable in economic terms and can lead to an overperformance
as large as dozens of basis points.
Keywords: limited attention, realized volatility, investor attention, retail in-
vestors, institutional investors, StockTwits, Twitter, news releases
JEL: G14, G17, G40
Acknowledgements
The research project is funded by the Swiss National Science Foundation (Grant
100018-172685).
Corresponding author: University of St. Gallen, Faculty of Mathematics and Statistics, Bo-
danstrasse 18, 9000 St. Gallen, Switzerland; daniele.ballinari@unisg.ch
Electronic copy available at: https://ssrn.com/abstract=3506720
1 Introduction
In today’s world, we are flooded with information of all kinds but have only limited
time and resources to properly digest it. In particular, investors have to decide where
to allocate their limited attention and, thus, not all information receives the same
amount of dedication (Kahneman, 1973). Empirical results presented in the prior
literature show that high levels of attention towards a company increase its stock price
volatility and trading volume (Antweiler & Frank, 2004; Audrino, Sigrist, & Ballinari,
2020; Dimpfl & Jank, 2016; Goddard, Kita, & Wang, 2015; Hamid & Heiden, 2015;
Rakowski, Shirley, & Stark, 2018; Vlastakis & Markellos, 2012, among others).
A possible explanation for these empirical findings is the limited attention hy-
pothesis. Theoretical models along the lines of Peng (2005), Peng and Xiong (2006),
DellaVigna and Pollet (2009), D. Hirshleifer, Lim, and Teoh (2009), Andrei and Hasler
(2014) postulate that limited attention towards specific events reduces the speed at
which the new information is properly interpreted and incorporated in stock prices.
While standard asset-pricing models assume that new information is instantaneously
reflected in stock prices, these theoretical models assume that attention is needed
for the information to be distilled and incorporated in the market value of a com-
pany. In the literature, this mechanism is often referred to as the limited attention
hypothesis. Based on the implication of these theoretical models, we first investigate
whether higher levels of attention around news releases are associated with stronger
price reactions:
Hypothesis 1 Higher attention around firm-specific news releases is related to higher
contemporaneous volatility.
Along the lines of the model introduced by Andrei and Hasler (2014), we investigate
this hypothesis by studying the contemporaneous relation between daily volatility and
investors’ attention around news releases.
Theoretical behavioral market models broadly distinguish between rational well in-
1
Electronic copy available at: https://ssrn.com/abstract=3506720
formed institutional investors and retail investors (Barberis, Shleifer, & Vishny, 1998;
Black, 1986; De Long, Shleifer, Summers, & Waldmann, 1990; Shleifer & Vishny, 1997).
Further, prior empirical research suggests that stock returns react differently to insti-
tutional and retail investor attention (Ben-Rephael, Carlin, Da, & Israelsen, 2019;
Ben-Rephael, Da, & Israelsen, 2017; Da, Engelberg, & Gao, 2011) and also that in-
stitutional and retail investors allocate attention around earnings and macroeconomic
news announcements differently (D. A. Hirshleifer & Sheng, 2019; Liu, Peng, & Tang,
2019). Concerning the impact of attention on volatility, it is thus likely that it makes a
difference whether retail or institutional investors are paying attention to information
releases. In particular, we expect that higher retail investor attention around news
releases also increases the volatility on days following the publication of news. The
motivation for this hypothesis is twofold: First, retail investors paying attention to a
news release are more likely to wrongly interpret the new information and, therefore,
create disagreement in the market and reduce the speed of price-adjustment. Second,
their noise-based trading activity adds additional risk to the market.1Further, we
expect institutional investors’ attention to have the opposite effect, i.e. increase the
speed at which stock prices adjust to the new information and therefore reduce (or at
least not increase) the return volatility on days following news releases. In summary,
the second hypothesis that we analyze is given by:
Hypothesis 2 Retail investors’ attention around firm-specific news releases increases
the return volatility on days following the publication of news. Institutional investors’
attention has the opposite effect.
Figure 1 summarizes our main findings concerning the first two hypotheses. The
figure depicts the change in volatility on the day of a news event (t= 0) and on the
following four days (t= 1,...,4) associated with a one standard deviation increase
1Models along the lines of De Long et al. (1990) postulate that noise traders add additional risk to
the market since sophisticated investors cannot anticipate how the irrational beliefs of these traders
change in the future.
2
Electronic copy available at: https://ssrn.com/abstract=3506720
Retail inv. attention
Institutional inv. attention
News, no earnings call
Earnings call
t=0 t=1 t=2 t=3 t=4 t=0 t=1 t=2 t=3 t=4
−0.025
0.000
0.025
0.050
−0.025
0.000
0.025
0.050
Figure 1: Change in volatility associated with a one standard deviation
increase in retail and institutional investors’ attention around news releases
Note: The figure depicts the fixed effects panel regression coefficients of attention
measures where the dependent variable is the realized volatility on the news event day
and the four post-release days. The gray-shaded areas represent 95% confidence intervals
computed using clustered standard errors obtained with the estimator introduced by
Arellano (1987).
in retail and institutional attention around the firm-specific news release.2Further,
we distinguish between earnings announcements and other types of firm-specific news.
Both retail and institutional attention are significantly positively related to the stock
return volatility observed on days with news releases. In other words, when investors
are paying attention to a news event, immediate price reactions are stronger as postu-
lated by the limited attention hypothesis. Higher levels of retail attention around news
events are also associated with increased volatility on days after news releases. In con-
trast, we find no evidence of a positive relation between the attention of institutional
investors to news releases and the return volatility in the post-announcement period.
In fact, our results show that higher attention levels of institutional investors around
2We use realized volatility measures based on five minute intraday returns and measure retail
investors’ attention through traders’ social media activity on StockTwits and institutional attention
with Bloomberg’s News Heat measure, an indicator of firm-specific news articles’ retrievals from
Bloomberg-terminals. More details about the attention measures and the regression framework are
given in Section 4.
3
Electronic copy available at: https://ssrn.com/abstract=3506720
earnings calls are associated with lower volatility levels two days after the event. These
results confirm our second hypothesis and show that when retail investors pay atten-
tion to news releases, price-adjustments are slower and the risk in the stock market
is higher. The attention of institutional investors, on the other hand, increases the
speed at which prices adjust to the new information and dampens the volatility on
days following news releases.
The results depicted in Figure 1 roughly distinguish between earnings announce-
ments and all other news releases. However, it is intuitively plausible that the reaction
of investors to different news types and topics is not homogeneous. As an illustra-
tive example, investors’ attention to a news article about corporate responsibility and
to one about a potential merger is most likely not associated with the same post-
announcement price reactions. For instance, when McDonald’s Corp. announced on
July 19, 2016 that they were going to sponsor the launch of the popular mobile game
Pokémon Go in Japan, the post-event volatility was low even though retail investors’
attention to this news release was very high (social media activity was above its his-
torical 97% quantile) and the contemporaneous volatility spiked (volatility level was
above its historical 86% quantile). On the other hand, when the retail company Lowe’s
Cos. Inc. announced on February 3, 2016 the acquisition of the Canadian home im-
provement chain Rona Inc. for 2.3 billion US dollars, retail investors’ attention was
high (the volume of messages shared about the company was above its historical 99%
quantile), and the stock price adjusted not only on the announcement day (volatility
above its historical 99% quantile), but also over the post-release period (on the follow-
ing two trading days the volatility was above its historical 96% quantile). Our third
hypothesis of interest is therefore:
Hypothesis 3 The relation between retail investors’ attention to news releases and
the post-release volatility is heterogeneous in the type and topic of the firm-specific news
release.
Our findings show that higher levels of retail investors’ attention are more strongly
4
Electronic copy available at: https://ssrn.com/abstract=3506720
associated with higher post-release volatility when the news release is unexpected (e.g.
"flash news") or when the new information potentially affects a company’s market
value (e.g. news about mergers and acquisitions, revenues or earnings). The novelty
and complexity of new information increase the difficulty for retail investors to cor-
rectly interpret the implications for a company’s market value and hence reduce the
speed at which prices adjust. In contrast, we do not find such a heterogeneous effect for
institutional investors’ attention, i.e. higher institutional investors’ attention to news
releases leads to lower volatility on days following news releases in a homogeneous
way for different news types and topics. Figure 2 depicts in exemplary fashion our
findings concerning Hypothesis 3. The figure shows the volatility change associated
with a one standard deviation increase in retail and institutional investors’ attention
around news releases about mergers and acquisitions, corporate responsibility and
news flashes. When retail investors are paying attention to news flashes or informa-
tion releases about mergers and acquisitions, volatility increases in the post-release
period. Instead, institutional investors’ attention to these news releases is associated
with lower volatility on the days after the announcement. For news releases about cor-
porate responsibility, neither retail nor institutional investors’ attention is significantly
associated with changes in volatility.
Finally, we apply our findings to generate out-of-sample predictions and measure
the economic value gained from this. More precisely, we take advantage of the relation
between retail investors’ attention and the increased post-announcement volatility to
improve one-day ahead volatility forecasts. The improvements thus obtained are not
only statistically significant but also economically meaningful. For example, in order
to obtain volatility predictions which take into account retail investors’ attention to
news flashes and news about acquisitions and mergers, an investor with mean-variance
preferences is willing to give up 44 and 34 basis points per year of her wealth, respec-
tively.
We contribute to the existing literature in several ways. First, we contribute to
5
Electronic copy available at: https://ssrn.com/abstract=3506720
Retail inv. attention
Institutional inv. attention
Acquisitions and mergers
Corporate responsibility
News flash
t=0 t=1 t=2 t=3 t=4 t=0 t=1 t=2 t=3 t=4
−0.05
0.00
0.05
0.10
−0.05
0.00
0.05
0.10
−0.05
0.00
0.05
0.10
Figure 2: Change in volatility associated with a one standard deviation
increase in retail and institutional investors’ attention around news releases
about mergers and acquisitions, corporate responsibility and news flashes
Note: The figure depicts the fixed effects panel regression coefficients of attention
measures where the dependent variable is the realized volatility on the news event day
and the four post-release days. The gray-shaded areas represent 95% confidence intervals
computed using clustered standard errors obtained with the estimator introduced by
Arellano (1987).
the literature concerned with the limited attention hypothesis by studying its implica-
tions for a wider range of news releases and using a novel direct measure of investors’
attention constructed from social media data. Moreover, we investigate the limited
attention hypothesis from a new perspective by analyzing the relation between atten-
tion to news and stock market volatility rather than returns. Furthermore, we extend
the analysis of the limited attention hypothesis by taking into account the type of
investors. In particular, to the best of our knowledge, this is the first study that
6
Electronic copy available at: https://ssrn.com/abstract=3506720
empirically untangles the role played by retail and institutional investors’ attention
in digesting a wide and heterogeneous range of information releases. In this respect,
our results contribute to the existing empirical and theoretical behavioral literature by
providing evidence that while the attention level of sophisticated investors fosters the
price-adjustment, noise traders’ attention slows this process down and increases post-
announcement market risk. Moreover, we contribute to the literature concerned with
predicting volatility by providing evidence that volatility forecasts can be significantly
improved based on our findings. Finally, we extend the approach to economically
evaluating volatility predictions introduced by Bollerslev, Hood, Huss, and Pedersen
(2018) to a multivariate setting and by optimally including transaction costs.
The remainder of this paper is organized as follows. In Section 2, we review the
related literature. Section 3 describes our methodology. In Section 4 we introduce
the dataset and in particular our measures of attention and the identification of news
articles. In Section 5 we present the empirical analysis of our hypotheses, and in
Section 6 we use our findings to predict volatility and improve the asset allocation of
a representative investor. Lastly, Section 7 offers some concluding remarks.
2 Related literature
Building upon the implications of the limited attention hypothesis developed in theo-
retical models (e.g. Andrei & Hasler, 2014; DellaVigna & Pollet, 2009; D. Hirshleifer
et al., 2009; Peng, 2005; Peng & Xiong, 2006), several empirical studies have exam-
ined the impact of attention on stock price reactions to news releases. D. Hirshleifer
et al. (2009) show that price and volume reactions to earnings announcements are
weaker when investors are distracted, i.e. attention is low, where distraction is mea-
sured as the number of concurrent earnings releases. Similarly, DellaVigna and Pollet
(2009) compare the price reactions observed during and after earnings announcements
made on Fridays with those observed when the announcement is made on other week
days. Their results show that the immediate price reaction to the announcement is
7
Electronic copy available at: https://ssrn.com/abstract=3506720
weaker when the earnings call happens on a Friday, since investors are generally more
distracted on the last working day of the week. Following the implications of their the-
oretical model, Andrei and Hasler (2014) provide empirical evidence that stock market
volatility and a measure of investors’ attention constructed from Google web search
queries are contemporaneously positively related. Ben-Rephael et al. (2017) analyze
the effect of institutional and retail investor attention on price adjustments observed
around earnings announcements and analyst recommendation changes. The study
of Ben-Rephael et al. (2017) is closely related to our paper, in that they investigate
the implications of attention on stock price reactions to earnings announcements and
analyst recommendation changes, distinguishing between retail and institutional in-
vestors. They find that institutional attention increases the immediate reaction to the
news and reduces the post-announcement drift in the stock price. While their results
provide limited evidence that retail investors’ attention reduces the price-adjustment
speed in the post-announcement period after earnings calls, they find no such evidence
for analyst recommendation changes. Our study differs in several aspects. First, we
analyze reactions in volatility rather then returns. This approach has several advan-
tages; e.g. our measure of volatility better reflects the trading activity of both retail
and institutional investors (see Section 4.3). Moreover, by analyzing the reaction in
volatility we are able to investigate the implications of limited attention on a wider
range of news, since we do not need to know whether the news is positive or negative.
Finally, we consider a different measure of retail investors’ attention (see Section 4.1).
For a recent survey of behavioral inattention in the field of economics and finance, we
refer to Gabaix (2019).
Besides the measures of attention used in the above-mentioned studies, several
other proxies have been proposed in the empirical literature. In their pioneering work,
Antweiler and Frank (2004) proxy attention with the posting volume on Yahoo! Fi-
nance message boards and report the predictive power of their measure for the volatility
of 45 US stocks. Barber and Odean (2008) use abnormal trading volumes, extreme
8
Electronic copy available at: https://ssrn.com/abstract=3506720
returns and the release of news as proxies for investors’ awareness of a company. Study-
ing the effect of attention on a company’s stock price around earnings calls, Aboody,
Lehavy, and Trueman (2010) use the 12-month return preceding the announcement.
Rubin and Rubin (2010) use the frequency with which a company’s Wikipedia article
is edited as a proxy of attention towards changes in firm-specific information. More
recently, a series of studies used the search-traffic on the Electronic Data Gathering,
Analysis, and Retrieval (EDGAR) system to measure investors’ attention (see, among
others, Drake, Roulstone, & Thornock, 2015; Lee, Ma, & Wang, 2015).
Several past studies have used direct attention measures for volatility prediction.
In particular, following Da et al. (2011), several studies have used Google web search
queries to proxy for investors’ attention. Using web search queries for company names
and the S&P 500 Index, Vlastakis and Markellos (2012) analyze the effect of infor-
mation demand on stock market volatility. Their results show that Google searches
at the market level have predictive power for future volatility and trading volume.
Hamid and Heiden (2015) report predictive power for attention measures constructed
from Google search queries for the keyword “dow” for the weekly volatility of the Dow
Jones Industrial Average Index. Using a similar proxy, Dimpfl and Jank (2016) aug-
ment the classical HAR model for realized volatility with an attention measure and
also report the predictive power for the Dow Jones Industrial Average Index at the
daily frequency. Furthermore, several studies have proposed using Google searches for
stock tickers as a proxy of retail investors’ attention (e.g. Ben-Rephael et al., 2019,
2017; Da et al., 2011; Drake, Roulstone, & Thornock, 2012). In our opinion, it is
questionable whether retail investors gather information about stocks by searching for
their ticker. For stock market volatility, most empirical studies document an effect of
general market-related web search queries (see, among others, Dimpfl & Jank, 2016),
whereas the impact of company-specific Google searches is limited (see e.g. Vlastakis
& Markellos, 2012). Comparing the forecasting power of different sentiment and atten-
tion measures for return volatility, the findings of Audrino et al. (2020) show that the
9
Electronic copy available at: https://ssrn.com/abstract=3506720
most relevant predictors are indeed company-specific attention measures constructed
from social media data and general market attention proxies obtained from web search
query data. Moreover, Google does not provide the absolute number of web searches,
but only a relative measure ranging from 0 to 100 based on a random sub-sample of
all searches. This might be especially problematic for stock tickers with a low number
of daily web search queries. Additionally, the daily trend measure provided by Google
is aggregated in terms of UTC time, and not Eastern Time. As such, we regard a
measure based on social media data as being more transparent and representative for
retail investors’ attention.
Our analysis is also related to the accounting literature concerned with firm disclo-
sures. Since the pioneering work of Morse (1982), the relation between news releases
and stock prices has been widely studied over the past decades. Among others, the
empirical results of Ryan and Taffler (2004) show that major price movements and
trading volumes are driven by firm-specific news releases: Roughly 65% of significant
price changes and trading volumes are explained by publicly available information.
Analyzing the role of press coverage around earnings announcements, Bushee, Core,
Guay, and Hamm (2010) report a reduced information asymmetry for companies with a
broader news coverage. Focusing on a wide range of news announcements, Boudoukh,
Feldman, Kogan, and Richardson (2019) show that news releases provide a significant
economic contribution to stock market volatility. In a recent study, Lerman (2020)
analyzes the discussion of accounting information on the social media platform Stock-
Twits and Yahoo! message boards. Her results show that online financial communities
pay attention to the release of accounting information, in particular earnings. How-
ever, she finds no evidence that higher attention is associated with an improvement in
the processing of information.
10
Electronic copy available at: https://ssrn.com/abstract=3506720
3 Methodology
In the following, we introduce the econometric framework used to test the hypotheses
introduced in the introduction. In our first hypothesis, we investigate the contem-
poraneous relation between volatility and investors’ attention to a news article. We
test Hypothesis 1 within a fixed effects panel data framework. More precisely, using a
within-transformation we estimate the following regression model:
log RVi,t =ci+X0
i,tβ
+ [(1 Ni,t)(1 Ei,t )γ1+Ni,t(1 Ei,t )γ2+Ei,t γ3]AR
i,t
+ [(1 Ni,t)(1 Ei,t )φ1+Ni,t(1 Ei,t )φ2+Ei,t φ3]AI
i,t +i,t (1)
where log(·)denotes the natural logarithm, RVi,t is the whole day realized volatility
of company ion day t,ciis a company-specific fixed effect, Xi,t is a vector of control
variables, AR
i,t and AI
i,t are the measures of retail and institutional investors’ attention,
respectively, Ni,t and Ei,t are dummy variables being equal to one if news and earnings
are released on day tabout company iand zero otherwise, and {i,t}is a zero-mean
innovation process. Note that the attention measures AR
i,t and AI
i,t are standardized
and, therefore, we can interpret the estimated coefficients as the percentual change in
volatility associated with a one standard deviation increase in attention.
In the vector of control variables, we include dummy variables for news and earnings
releases, lagged values of (log) realized volatility, the turnover-ratio defined as the
daily traded (dollar) volume divided by the market capitalization, and the log CBOE
Market Volatility Index (VIX). The inclusion of lagged volatilities is motivated by
the high persistence observed in the stock return’s second moment. More precisely,
building upon the model introduced by Corsi (2009) we include the lagged log realized
volatility and the lagged average log realized volatility observed over the previous week
and month. The inclusion of the turnover-ratio and the VIX are motivated by their
11
Electronic copy available at: https://ssrn.com/abstract=3506720
predictive power for volatility documented in several empirical studies (e.g. Audrino et
al., 2020; Buncic & Gisler, 2016; Mittnik, Robinzonov, & Spindler, 2015). The results
presented in the remainder of this paper are robust to the inclusion of other control
variables, e.g. weekday dummies and measures of earnings surprises.3
The coefficients of interest from the perspective of our first hypothesis, are γ2,γ3,
φ2and φ3. More precisely, under Hypothesis 1 we expect them to be positive, i.e.
higher levels of retail and institutional investors’ attention on the day a news article is
released (γ2and φ2) or an earnings announcement is made (γ3and φ3) are associated
with higher volatility.
Concerning Hypothesis 2, we investigate if retail investors’ attention decreases and
institutional investors’ attention increases the speed of price-adjustment to new in-
formation. In particular, we conjecture that retail investors’ attention to news is
associated with an increase in post-release volatility, i.e. the volatility observed the
days after the news release.4On the other hand, we conjecture that institutional in-
vestors’ attention to news is instead associated with a decrease in post-release volatility.
We investigate our second hypothesis by estimating the following fixed effects panel
predictive regression model:
log RVi,t+h=ci+X0
i,tβ
+ [(1 Ni,t)(1 Ei,t )γ1+Ni,t(1 Ei,t )γ2+Ei,t γ3]AR
i,t
+ [(1 Ni,t)(1 Ei,t )φ1+Ni,t(1 Ei,t )φ2+Ei,t φ3]AI
i,t +i,t+1 (2)
where we consider different post-event horizons, i.e. h= 1,2,3,4. The main differences
to the regression model defined in Equation (1) are that the dependent variable is
the realized volatility on subsequent days after news has been released, and that the
3Following Livnat and Mendenhall (2006), we define earnings surprises as the difference between
the current earnings per share (EPS) and the previous year’s EPS, divided by the stock price (i.e.
EPS are assumed to follow a seasonal random walk). The results are available from the authors upon
request.
4Throughout the paper, we will refer to the trading days after the day news has been released as
post-event, post-release or post-announcement period.
12
Electronic copy available at: https://ssrn.com/abstract=3506720
vector of control variables contains the realized volatility and average weekly and
monthly realized volatilities observed on day t, rather than t1. Under our second
hypothesis, we expect the coefficients γ2and γ3to be positive, and φ2and φ3to be
negative. In other words, higher attention of retail (institutional) investors to a news
release is expected to be associated with higher (lower) volatility in the days after the
publication.
In Hypothesis 3, we hypothesize that the relation between retail investors’ attention
to news releases and the post-release volatility is heterogeneous in the type and topic
of the firm-specific news. From an econometric perspective, we add to the regression
defined in Equation (2) additional news dummy interactions which distinguish be-
tween different types and topics of news articles. More precisely, we use the following
predictive regression model:
log RVi,t+h=ci+X0
i,tβ
+"(1 Ni,t)(1 Ei,t )γ1+ (1 Ei,t)X
j
Di,t(j)γ(j)
2+Ei,t γ3#AR
i,t
+ [(1 Ni,t)(1 Ei,t )φ1+Ni,t(1 Ei,t )φ2+Ei,t φ3]AI
i,t +i,t+1 (3)
where Di,t(j)is a dummy variable being equal to one when a news article of type or
topic jis released. For example, when comparing different types of news articles, we
include a dummy variable for “tabular material,” “full article,” “press release,” “news
flash” and “hot news flash.” Under Hypothesis 3 we expect γ(j)
26=γ(s)
2for at least one
j6=s. Note that while formally our third assumption solely focuses on retail investors’
attention, we also test for potential heterogeneities in institutional investors’ attention.
4 Data
The time period for our analysis ranges from January 2011 to December 2017. We
consider 360 companies which were part of the S&P 500 throughout the time period
13
Electronic copy available at: https://ssrn.com/abstract=3506720
of our analysis.5Our basic frequency for the analysis is daily data. We define a
day to start at 9:30 AM and end at 9:30 AM Eastern Time of the next trading day.
Accordingly, a day consists of a trading time period followed by a period where the
stock market is closed. For example, the last day of each week starts on Friday at 9:30
AM and ends on Monday morning at 9:30 AM Eastern Time when markets open for
Monday.
4.1 Attention measures
In this study, we consider two different measures of attention: One for institutional
investors’ and one for retail investors’ attention. Following the approach of Audrino
et al. (2020), we measure retail investors’ attention by counting the daily number of
messages shared on social media platforms about a company. The choice of using
social media activity to proxy attention is mainly motivated by four aspects. First,
in contrast to many measures provided by commercial vendors, our attention proxy is
transparent in terms of its construction since we have full control over the collection
and filtering approach (more details are provided in the next paragraph). Second,
financial measures which are typically used as proxies for investors’ attention, such
as extreme returns and abnormal trading volumes, not only reflect the attention and
awareness of investors towards a company, but are the outcome of different economic
forces (Da, Engelberg, & Gao, 2014). Third, the use of social media data allows us to
capture different awareness dimensions, such as attention towards the stock price, the
firm’s fundamentals, its products or the company’s social value. Finally, as discussed in
Section 2, we regard our measure based on social media data as being more transparent
and representative for retail investors’ attention compared to measures constructed
from web search queries. However, when measuring retail investors’ attention through
social media activity, we need to make the assumption that traders are in fact sharing
5Note that from 2011 to 2017, 368 companies were constantly part of the S&P 500. We excluded
eight companies due to data issues (missing intraday trade information).
14
Electronic copy available at: https://ssrn.com/abstract=3506720
their thoughts, ideas and fears about a company on the online platforms considered.6
The primary social media platform considered in this study is StockTwits.7Sim-
ilarly to Twitter, StockTwits is a social media platform where users can share short
messages with the online community, the difference being that it is specifically de-
signed for sharing thoughts, information and ideas between investors. We collect all
messages shared on these platform, i.e. tweets, which either mention a company’s name
or its cashtag, i.e. the company’s ticker symbol preceded by the dollar sign which is a
common convention adopted by stock market participants on social media platforms8
(for instance, the cashtag of Apple Inc. is $AAPL). When collecting data from the
social media platform, we account for changes of a company’s name or ticker.9As a
further robustness check, we also consider an alternative social media platform, namely
Twitter.10 We report in the main part of this paper only the results obtained with
StockTwits; the results obtained with Twitter are summarized in Appendix B. The
data for Twitter is collected using a scrapper, and for StockTwits we have obtained
the full historical data containing all short messages for the period considered in this
study. For the 360 analyzed companies, we identify and collect a total of 9,890,132
and 30,520,617 messages shared on StockTwits and Twitter, respectively. Summary
statistics of the raw total volume of social media messages collected for the analyzed
companies can be found in Panel A of Table 1. Summary statistics of the correlation
between the attention measures constructed from the two social media platforms are
reported in Panel A of Table 2. The average correlation amounts to 0.44 on days with-
out any information release, 0.57 on days with news and 0.56 on earnings call days
suggesting that the two data sources provide to some extent different information.
6See Cookson and Niessner (2019) for a discussion about the plausibility of the assumption that
investors share their “true” thoughts about companies and investment opportunities on social media.
7See https://stocktwits.com
8We also investigated the effect on our results when using a more restrictive filtering approach,
e.g. considering only messages mentioning a unique cashtag. Our results are, however, not affected
by the filtering approach.
9Between 2011 and 2017 we identified 41 out of the 360 analyzed companies that changed their
name or ticker at least once.
10See https://twitter.com
15
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 1: Summary statistics of news and social media coverage across 360 large cap
US stocks
Panel A: Number of news articles and social media posts from 2011 to 2017
Q25% Average Median Q75%
StockTwits messages 4,709.00 27,471.82 7,900.50 16,609.00
Twitter messages 20,000.00 84,772.67 32,815.00 66,512.75
News articles 1,280.75 1,689.53 1,480.00 1,849.00
Panel B: Number of days without news, news and earnings calls from 2011 to 2017
Q25% Average Median Q75%
No news 861.75 933.51 972.00 1052.00
News, no earnings call 680.00 798.49 760.00 870.25
Earnings calls 28.00 28.00 28.00 28.00
Note: In Panel A the table depicts the average, median, 25%- and 75%-quantiles across the 360
analyzed companies of the total number of news articles and social media messages published between
2011 and 2017. Panel B summarizes the average, median, 25%- and 75%-quantiles of the total number
of days a company has had no news releases, at least one news article (other than earnings calls),
and an earnings call. Not surprisingly, since each company has four earnings calls per year, all firms
in our sample have 28 earnings announcements over the eight analyzed years.
Raw daily retail investors’ attention variables are defined as the logarithm of one
plus the number of messages posted on a social media platform during a day. We
follow the approach adopted by Garcia (2013) and remove trends and seasonal com-
ponents from the raw daily attention measure by running the following regression at
the company level:
˜
AR
i,t =βL5(˜
AR
i,t) + γZt+i,t
where ˜
AR
i,t is the raw (log) posting volume about company i,Lpis a p-lag operator
which creates a vector of the past five lags, and Ztis a vector including a constant,
dummy variables for weekdays and months. The attention measure is then defined
as AR
i,t = ˆi,t/si, where ˆi,t are the estimated residuals and si, are their standard
deviations. It follows that the attention measures have zero mean and unit standard
deviation. Note that the results presented in the following are qualitatively unchanged
when defining investors’ attention as the raw (log) daily posting volume.11
11The results obtained with the raw (log) attention measures are available from the authors upon
request.
16
Electronic copy available at: https://ssrn.com/abstract=3506720
In order to measure institutional investors’ attention, we follow the approach of
Ben-Rephael et al. (2017) and use the news searching and reading activity for specific
stocks on Bloomberg terminals. More precisely, Bloomberg records the number of times
firm-specific news articles are read and searched for by terminal users. Comparing these
counts with those observed in the past 30 days, Bloomberg constructs a daily index of
attention which attains the values 0 (lowest attention), 1, 2, 3 or 4 (highest attention).
Following the approach proposed by Ben-Rephael et al. (2017), we transform these
discrete scores into continuous values and standardize the variable at the company
level to have zero mean and unit standard deviation.12 Historical data provided by
Bloomberg are missing for the first week of 2011 and from August 17 to November
2, 2011. We refer to Ben-Rephael et al. (2017) for a more detailed description of
Bloomberg’s news reading and searching index, and for a discussion of why this index
captures the attention of institutional investors.
Panel B and Panel C of Table 2 summarize the correlations between Bloomberg’s
news attention measure and the volume of social media posts shared on StockTwits
and Twitter, respectively. The correlation coefficients reported in Panel B and Panel
C are small, in particular on days with earnings calls, suggesting that the informa-
tion provided by social media platforms differs from the one that is provided by the
Bloomberg news retrieval measure. As expected, the social media platforms reflect a
different type of investors’ attention compared to the measure of institutional investors’
attention introduced by Ben-Rephael et al. (2017).
4.2 News releases
As we are interested in investors’ attention around firm-specific news releases, we need
to identify all such events. The majority of empirical studies analyzing the impact of
investors’ attention on stock markets have focused on price reactions around earnings
12Based on Bloomberg’s definition of the index, to construct a continuous attention variable, Ben-
Rephael et al. (2017) assign to the discrete scores 0, 1, 2, 3 and 4 the values -0.350, 1.045, 1.409,
1.647 and 2.154.
17
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 2: Summary statistics of the correlation between the attention measures
Panel A: Correlation between StockTwits and Twitter attention
Q25% Average Median Q75%
No news 0.32 0.44 0.44 0.55
News, no earnings call 0.45 0.57 0.58 0.70
Earnings calls 0.42 0.56 0.58 0.75
Panel B: Correlation between StockTwits and Bloomberg attention
Q25% Average Median Q75%
No news 0.08 0.13 0.14 0.18
News, no earnings call 0.15 0.21 0.21 0.27
Earnings calls -0.15 0.00 0.03 0.19
Panel C: Correlation between Twitter and Bloomberg attention
Q25% Average Median Q75%
No news -0.01 0.05 0.05 0.11
News, no earnings call 0.11 0.17 0.17 0.23
Earnings calls -0.18 0.01 0.03 0.19
Note: In Panel A the table depicts the average, median, 25%- and 75%-quantiles across the 360
analyzed companies of the correlation between the daily volume of posts shared on StockTwits and
Twitter. Similarly, Panel B and Panel C summarize the correlations between Bloomberg’s news at-
tention measure and the volume of social media posts shared on StockTwits and Twitter, respectively.
releases (e.g. Ben-Rephael et al., 2017; DellaVigna & Pollet, 2009; D. Hirshleifer et
al., 2009). In the first part of this paper, we therefore distinguish between earnings
announcements and other news releases.
For the identification of scheduled earnings releases, we use the Thomson Reuters’
I/B/E/S database, which provides dates and time-stamps of quarterly earnings calls.
The identification of news releases is done using the news article database obtained
from RavenPack News Analytics. The database covers information from Dow Jones
Newswire, regional editions of the Wall Street Journal, Barron’s and MarketWatch as
well as press releases from leading global media organizations. Using textual analysis
techniques, RavenPack News Analytics detects the entities in each news release and
determines how novel and relevant the story is for each mentioned company. Moreover,
each news release is automatically categorized based on its topic (e.g. earnings, labor
issues, legal). Finally, the database also provides information about the type of the
news release (obtained directly from the news providers), i.e. tabular material, full
18
Electronic copy available at: https://ssrn.com/abstract=3506720
article, press release, news flash and hot news flash.13 We collect all news releases with
the highest novelty and relevance for the analyzed companies.14 This is motivated by
the findings of Groß-Klußmann and Hautsch (2011), which show that the omission of
news articles containing irrelevant and old information is crucial for filtering out noise.
A similar filtering approach is also adopted by Boudoukh et al. (2019). For the 360
analyzed companies, we identify a total of 608,298 news releases from 2011 to 2017.
As for the social media data, the news coverage varies greatly among the analyzed
companies. Summary statistics of the total number of news articles published between
2011 and 2017 for the 360 analyzed US companies can be found in Panel A of Table
1. Summary statistics of the total number of days a company has had no news, at
least one news release (other than an earnings announcement), and earnings calls are
depicted in Panel B of Table 1. While the news coverage (number of articles) varies
greatly across the companies, the number of events (days with at least one news article)
is more homogeneous.
4.2.1 Relevance of news releases to retail investors
Being particularly interested in the relation between retail investors’ attention and
price reactions to news releases, we introduce the concept of relevance of a news event.
More precisely, we compare the textual data of social media posts and news article
headlines to identify news that is noticed and discussed by retail investors. The fact
that RavenPack News Analytics classifies an article as being highly relevant for a
specific company does not necessarily imply that the new information is noticed by
investors, in particular by retail investors. To this end, we use a standard approach
taken from the natural language and information science literature and compute for
each news article the cosine-similarity between its headline and social media messages
13The database provided by RavenPack News Analytics offers a much wider range of indicators than
the ones mentioned in this study; for more information we refer directly to https://www.ravenpack
.com.
14RavenPack News Analytics assigns to each news release and for each mentioned entity a novelty
and relevance score ranging from 0 to 100. In this study only news articles with a relevance and
novelty score of 100 are considered.
19
Electronic copy available at: https://ssrn.com/abstract=3506720
published on the same day.15 The main idea of cosine-similarity is to first cast doc-
uments in a (high-dimensional) vector space, and then measure the similarity among
textual data by computing the angle between the corresponding vectors:
sim(hi, mj) = < hi, mj>
||hi|| ||mj|| (4)
where hiand mjare the vectorial representation of a news headline and a social media
message, respectively, <·,·>stands for the scalar product and || · || for the Euclidean
norm. To obtain a vectorial representation of the textual data, we first clean and
tokenize all social media messages and news headlines, i.e. (i) we remove stop-words,
links, emoticons and references to other users, (ii) convert all letters to lower case and
reduce inflected words to their word stem by applying a Porter stemmer, and (iii)
break up all headlines and messages into single words (uni-grams). A term-document
matrix is then created, where each column represents a word and each row corresponds
to a document, i.e. a news headline or a message shared on Twitter or StockTwits.
In other words, each row of the matrix corresponds to the vectorial representation of
a document. In its most simple form, the matrix is filled with the relative number of
times a term occurs in a given document, often referred to as term-frequency (TF). In
this study we follow, among others, Brown and Tucker (2011) and apply a common
weighting refinement which multiplies the term-frequency with the inverse document-
frequency (IDF), resulting in the so-called term-frequency inverse document-frequency
(TF-IDF).16 This approach has the advantage that common words, such as “company,”
are down-weighted and contribute only marginally to the cosine-similarity between
two documents. Note that TF-IDF is non-negative and as such the cosine-similarity
between a news headline and a social media message is bounded between 0and +1.
Given the similarity between a news headline and Twitter or StockTwits messages, we
15As described previously, we consider a day to start at 9:30 AM and end at 9:30 AM Eastern Time
of the next trading day.
16More precisely, the inverse document-frequency is defined as the logarithm of the total number
of documents divided by the number of documents that contain a particular term, i.e. IDF(w) =
log N
nwwhere Nis the number of documents and nwthe number of documents that contain word w.
20
Electronic copy available at: https://ssrn.com/abstract=3506720
define the relevance score of a news article as:
rel(hi,t) = X
tjt
sim(hi,t, mtj)(5)
where sim(hi,t, mtj)is the cosine-similarity between headline ipublished on day tand
a social media message published at time tj, and the sum is taken over all messages
published on day t. In this study we define a news article to be relevant for retail
investors if rel(hi,t)>0.
Figure 3 shows the distribution of social media messages and news articles over one-
hour intervals of a day. More precisely, the plots in Figure 3 depict the hourly average
number (per company) of StockTwits messages (a), Twitter microblogs (b) and news
articles (c, d) published within a one-hour interval. For the volume of news releases, we
distinguish between relevant (dark-grey) and irrelevant (light-grey) articles as defined
by the cosine-similarity between the headlines and messages posted on StockTwits (c)
and Twitter (d). The social media activity is particularly concentrated over the trading
hours, with the highest number of messages shared between 9:00 AM and 10:00 AM
(Eastern Time), i.e. during market opening. Most news articles are instead published
between 7:00 AM and 8:00 AM and between 3:00 PM and 4:00 PM (Eastern Time),
i.e. before the market closes. Interestingly, when StockTwits data are used, fewer news
articles have a positive relevance score, compared to when Twitter messages are used.
In particular, a considerable share of articles published between 3:00 PM and 4:00 PM
are not discussed on the social media platform StockTwits.
4.2.2 Categorization of days based on type and topic of news
The computation of relevance scores for each news article also allows us to identify,
from the perspective of retail investors, the most relevant articles published on a
given day per company. For every company, we create two categorizations of days by
identifying the category with the highest relevance scores using the following grouping
factors: (i) news type, and (ii) topic of the articles. In other words, we first sum the
21
Electronic copy available at: https://ssrn.com/abstract=3506720
0.0
0.3
0.6
0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hours
Hourly average number of messages (per company)
(a) StockTwits messages
0.0
0.5
1.0
1.5
2.0
2.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hours
Hourly average number of messages (per company)
(b) Twitter messages
0.00
0.02
0.04
0.06
0.08
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hours
Hourly average number of news articles (per company)
(c) News articles using StockTwits to determine
relevance
0.00
0.02
0.04
0.06
0.08
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hours
Hopurly average number of news articles (per company)
(d) News articles using Twitter to determine relevance
Figure 3: Number of social media messages and news articles published in one-hour intervals over a day
Note: The figure reports the average number (across companies) of StockTwits messages (a), Twitter messages (b) and news articles, with the
relevance determined by StockTwits (c) or Twitter messages (d), published in one-hour intervals over a day. More precisely, each plot shows
the average number of social media messages published within a certain hour of the day across all companies over the entire analysis period,
i.e. from 2011 to 2017. The average number of relevant news articles published within a one-hour interval is represented in dark-grey, and
the irrelevant articles in light-grey. A news article is classified as being relevant if the cosine-similarity between its headline and social media
messages is positive.
22
Electronic copy available at: https://ssrn.com/abstract=3506720
relevance scores of all articles published on a given day per category of the chosen
grouping. We then categorize the day based on the category with the highest total
relevance score. This approach allows us to study potential heterogeneities in the effect
of investors’ attention on stock market volatility. In particular, we are not only able
to study the effect of investors’ attention around news releases in general, but we can
also shed light on potential heterogeneities due to different news types and topics.
4.3 Volatility measure
We measure daily volatility using the concept of realized volatility. I.e., instead of
treating the returns’ conditional second moment as latent, we directly estimate it using
high-frequency intraday data. In detail, we construct realized volatility measures for
the time period between 2011 and 2017 using intraday transaction prices obtained
from the NYSE Trade and Quote (TAQ) database.17 The high-frequency data are
cleaned following the procedure suggested by Barndorff-Nielsen, Hansen, Lunde, and
Shephard (2009). Realized volatility is constructed as the square root of the sum of
squared intraday log returns. Following the existing realized volatility literature, we
use a five-minute sampling frequency for calculating intraday returns.
The volatility measure thus constructed uses only data covering part of the day,
i.e. from 9:30 AM to 4:00 PM. Since we define a day to start at 9:30 AM and end
at 9:30 AM of the following trading day, we adopt an estimator suggested by Hansen
and Lunde (2005) and use overnight returns to construct a volatility measure for the
whole day. More precisely, a variance measure for the whole day is obtained by a
linear combination of the realized variance estimated from high-frequency transaction
data and the squared overnight return. The optimal combination is determined by
minimizing a mean-squared error (MSE) criterion.18 In order to obtain more robust
17For August 1, 2012, trade data of 35 companies in our sample are missing. For this day, realized
volatility measures are constructed using quote data.
18Since the true daily integrated variance is not observed, Hansen and Lunde (2005) obtain the
optimal combination of realized variance and squared overnight returns that minimizes the MSE
criterion by restricting their attention to the class of conditionally unbiased estimators. For a more
detailed description of the estimation procedure, we refer to Hansen and Lunde (2005).
23
Electronic copy available at: https://ssrn.com/abstract=3506720
optimal weights, we compute the optimal linear combination cross-sectionally.19 Our
main findings remain qualitatively the same when the classical realized volatility es-
timator (i.e no use of overnight returns) is used, when the optimal combination is
computed at the company-level, or when one of the other two estimators suggested by
Hansen and Lunde (2005) is used.
Analyzing the volatility reaction to news rather than the stock return reaction
has two main advantages. First, we do not need to know whether a news release is
positive or negative for the considered company, i.e. regardless of whether the return
is positive or negative, if prices react to news, volatility will increase. Secondly, daily
returns reflect the aggregate opinion of market participants and are highly influenced
by the most dominant market participants. If, for example, the opinions between
institutional and retail investors about a news release diverge, and the former push
stock prices up while the latter start selling their shares, the daily return will most likely
be positive since institutional investors’ trading volume is usually considerable larger.
In other words, when news is relevant to both retail and institutional investors, daily
stock returns do not always reflect the reaction of all market participants. Realized
volatility estimated from intraday trade data accounts by construction for the trades
of investors; i.e. the trades of retail investors also increase the daily volatility.20
Table 3 reports summary statistics for the three main variables of interest, i.e.
volatility, retail and institutional investors’ attention. More precisely, Panel A reports
the average volatility and attention on days with news releases relative to days without
news releases. A value above one indicates that volatility or attention is on average
higher compared to a day without any new information release. Not surprisingly,
volatility is on average higher when news is released, in particular when earnings are
announced. Attention of both retail and institutional investors follows a similar pat-
19In order to obtain true ex-ante predictions, in the out-of-sample analysis the optimal weights are
computed using only data from the rolling estimation window.
20On days with news releases, the mean squared difference between the open-to-open daily absolute
return and the whole day realized volatility is four times larger when both retail and institutional
attention are above their historical 90% quantile, compared to when only institutional attention is
above its 90% quantile.
24
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 3: Summary statistics of volatility and attention measures
Panel A: Volatility and attention on days with news releases relative to days without news
RVtRVt+1 AR
tAI
t
No news in t1.00 1.00 1.00 1.00
News, no earnings call in t1.07 1.05 1.32 1.44
Earnings calls in t1.31 1.92 6.41 3.73
Panel B: Correlation between attention and volatility
Contemporaneous corr. Lagged corr.
AR
t, RVtAI
t, RVtAR
t, RVt+1 AI
t, RVt+1 RVt, RVt+1
No news in t0.16 0.25 0.05 0.13 0.67
News, no earnings call in t0.30 0.30 0.12 0.16 0.68
Earnings calls in t0.13 0.22 0.13 -0.02 0.49
Note: Panel A reports the average volatility and attention on days with news releases relative to
days without news releases. A value above one indicates that volatility or attention is on average
higher compared to a day without any new information release. Panel B reports average (across all
companies) contemporaneous and predictive correlation coefficients between volatility and attention.
Retail investors’ attention is constructed from StockTwits data.
tern. Interestingly, volatility is also on average higher the day after the news has been
released. Panel B of Table 3 reports average (across all companies) contemporane-
ous and predictive correlation coefficients between volatility and attention. Consistent
with the widely reported high persistence in return volatility, the auto-correlation of
realized volatility is considerably high (e.g. Andersen, Bollerslev, Diebold, & Ebens,
2001). The first two columns in Panel B show that the contemporaneous correlation
between volatility and attention increases when news articles are released (other than
earnings calls), in particular for retail investors. A similar pattern is also observed for
the predictive correlation between attention and volatility.
25
Electronic copy available at: https://ssrn.com/abstract=3506720
5 The relation between investor attention, news, and
volatility
5.1 News releases and earnings announcements
In this section, we analyze our first two hypotheses. Concerning Hypothesis 1, Panel A
of Table 4 reports the estimates of the attention coefficients for the fixed effect panel
regression defined in Equation (1).21 More precisely, Panel A reports the increases
in volatility associated with a one standard deviation increase of retail investors’ and
institutional investors’ attention on days with earnings calls and days with other news
releases. For completeness, we also report the coefficients for days without any news
releases. For both retail investors and institutional investors, increases in attention on
days with news releases are associated with contemporaneous increases in volatility.
The estimated coefficients are positive and highly significant. On days with an earnings
call, a one standard deviation increase in retail and institutional attention is associated
with a 3.4% and 3.3% increase in volatility, respectively. This finding confirms the
results reported in other empirical studies (e.g. Aboody et al., 2010; Ben-Rephael et al.,
2017; DellaVigna & Pollet, 2009; D. Hirshleifer et al., 2009; D. A. Hirshleifer & Sheng,
2019). Note that the earnings day dummy variable coefficients are large and highly
significant (results not tabulated). I.e., a large part of the volatility spike observed
around earnings announcements is explained by this dummy variable. The estimated
attention coefficients are even larger on days with other news releases: the increase
in contemporaneous volatility associated with a one standard deviation increase in
attention amounts to 4.2% for retail investors and 4.6% for institutional investors.
Our findings therefore extend the results reported in the literature about earnings
announcements to a wider range of news releases. In summary, the results reported
in Panel A of Table 4 clearly confirm our first hypothesis. When investors are paying
21The results obtained when measuring retail investors’ attention with Twitter data are reported
in Table 11 in Appendix B.
26
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 4: Investors’ attention and volatility reaction
Panel A: Contemporaneous relation between attention and volatility
Estimate t-statistic
Retail investor attention
No news 0.028 17.99
News, no earnings call 0.042 17.10
Earnings calls 0.034 9.91
Institutional investor attention
No news 0.037 19.48
News, no earnings call 0.046 17.48
Earnings calls 0.033 11.05
Adjusted R258.48%
Number observations 587,762
Panel B: Relation between attention and next day’s volatility
Estimate t-statistic
Retail investor attention
No news 0.012 21.36
News, no earnings call 0.023 28.34
Earnings calls 0.030 4.93
Institutional investor attention
No news 0.002 3.31
News, no earnings call 0.002 3.99
Earnings calls 0.005 0.88
Adjusted R253.78%
Number observations 588,104
Note: The table reports the estimated volatility increase on the news release day (Panel A) and
the day after (Panel B) associated with a one standard deviation increase in retail and institutional
investors’ attention. More precisely, Panel A summarizes the attention estimates of the regression
defined in Equation (1), and Panel B those of the regression defined in Equation (2) with h= 1. For
completeness, the table also reports the estimated increase in contemporaneous and future volatility
associated with a one standard deviation increase in attention when no news article is released. The
t-statistics are computed using clustered standard errors obtained with the estimator introduced by
Arellano (1987).
attention to news releases, price reactions are stronger and volatility increases.
Panel B of Table 4 reports the estimated attention coefficients of the predictive
regression model defined in Equation (2). In detail, the panel reports the increase in
volatility observed the day after a news event for a one standard deviation increase in
investors’ attention. Further, changes in volatility observed up to four days after the
news announcement associated with a one standard deviation increase in investors’ at-
tention are summarized in Figure 1. Note that we measure investors’ attention only on
27
Electronic copy available at: https://ssrn.com/abstract=3506720
the day the news is released. Retail investors’ attention to news releases is associated
with an increase of next day’s volatility of 3% for earnings announcements and 2.3%
for other news. These estimates are highly significant. Considering up to four days
after the news release, we observe that this relation is still significantly positive and
decreasing over time for retail investors. In contrast, we find only limited evidence that
institutional investors’ attention to news releases is associated with higher volatility in
the post-event trading days. The estimated coefficient for earnings announcements is
0.005 and not significant, and for other news releases 0.002 and marginally significant.
In fact, Figure 1 shows that the relation between institutional investors’ attention to
earnings announcements and volatility becomes significantly negative two days after
the event. Further, our findings show that for other news releases, the attention of
institutional investors is not associated with changes in volatility in the post-event
period. The results presented in Panel B of Table 4 and in Figure 1 support Hy-
pothesis 2. In summary, institutional investors’ attention to earnings calls reduces the
volatility in the days after the announcement, an indication that the new information
is incorporated more quickly into prices. This finding confirms the results reported
by Ben-Rephael et al. (2017). There is instead only limited evidence of a relation be-
tween institutional investors’ attention and stock market volatility after the release of
other information. In addition, we find clear evidence that a higher attention of retail
investors to news releases is associated with a higher volatility over the next trading
day. This result is consistent with the idea that retail investors are reducing the speed
at which prices adjust to new information and increasing the risk in the stock market.
5.1.1 The relation between attention and trading volume
The results presented thus far show that retail investors’ attention to news releases is
associated with increases in volatility up to four days after the event. If the higher
volatility is in fact related to increased trading activities of investors, we would expect
attention also to be positively related to the trading volume around the news release.
28
Electronic copy available at: https://ssrn.com/abstract=3506720
Ideally, we would analyze the relation between retail investors’ attention and trading
activity. However, the trading volume of retail investors is not directly observable and
we therefore follow Dimpfl and Jank (2016) and use the total trading volume. We
adopt the approach by Garcia (2013) and use the same procedure applied previously
to the attention measures to remove trend and seasonal components from the dollar
trading volume. We regress the log trading volume at the company level on five lagged
values and dummy variables for weekdays and months (for more details see Section 4).
We then analyze the effect of investors’ attention on the normalized trading volume
ˆ
Vi,t = ˆi,t/si, , where ˆi,t are the estimated residuals from the aforementioned regression
and si, its standard deviation, within a fixed effects panel regression framework. More
precisely, we estimate the same regressions defined in Equation (1) and Equation (2)
using as a dependent variable the normalized trading volume and additionally control
for daily returns and squared returns. The results are depicted in Figure 4.
For retail investors, we observe a similar pattern as for the realized volatility. In
particular, when retail investors are paying attention to a news event, the trading
volume not only increases on the day of the news release, but also in the post-release
period. This finding indicates that higher attention of retail investors increases the
market risk and reduces the price-adjustment speed through excessive trading. In
contrast, when institutional investors pay attention to a news release we even observe
a reduction in the trading activity, i.e. a one standard deviation increase in institu-
tional investors’ attention on the news release day is associated with a reduction in
trading volume over the post-event period. When paying attention to a news release,
sophisticated investors quickly interpret the new information and rapidly react to it,
leading to a higher trading activity and stronger price adjustments on the event-day,
and to calmer markets in the following trading days. Around earnings calls, however,
institutional investors’ attention is associated with higher trading activities even on
the first day after the announcement. This finding can be explained by our definition
of a trading day: We define a day to start at 9:30 AM and end at 9:30 AM of the
29
Electronic copy available at: https://ssrn.com/abstract=3506720
Retail inv. attention
Institutional inv. attention
News, no earnings call
Earnings call
t=0 t=1 t=2 t=3 t=4 t=0 t=1 t=2 t=3 t=4
−0.1
0.0
0.1
0.2
0.3
0.4
−0.1
0.0
0.1
0.2
0.3
0.4
Figure 4: Trading volume increase associated with a one standard devi-
ation increase in retail and institutional investors’ attention around news
releases
Note: The figure depicts the fixed effects panel regression coefficients of attention mea-
sures where the dependent variable is the normalized trading volume on the news event
day and the four post-release days. The gray-shaded areas represent 95% confidence in-
tervals computed using clustered standard errors obtained with the estimator introduced
by Arellano (1987).
next trading day. As such, when earnings are released in the morning, the trades
of institutional investors are reflected in the trading volume of the day after the an-
nouncement. This effect is not observed for the volatility, since we take into account
overnight returns when measuring the second moment of stock returns. Given that
the focus of this study is on volatility, a more detailed analysis of the relation between
investors’ attention and their trading activity is left for future research.
5.1.2 Weekday effects and the reaction to news
Several studies have analyzed a similar hypothesis focusing primarily on earnings
announcements and using indirect measures of investors’ attention (or distraction).
DellaVigna and Pollet (2009), for example, analyze price-adjustments to earnings calls
in dependence on the release day, arguing that on Fridays investors are more distracted
30
Electronic copy available at: https://ssrn.com/abstract=3506720
and as such, attention to the newly released information is lower. Their results show
that the post-announcement price-drift is stronger when the earnings call happens on a
Friday. In order to ensure that our results are not driven by potential weekday effects,
we include weekday-dummy variables and interact them with the dummies for news
releases and earnings calls as a robustness check. In this way, we control for potential
spikes in volatility associated with the publication-day of the news article and earnings.
The coefficients of interest for our first two hypotheses estimated when controlling for
weekday-effects are summarized in Table 8 in Appendix A. The contemporaneous rela-
tion between attention and volatility is not affected by the additional control variables,
confirming our first hypothesis. The results for retail investors’ attention reported in
Panel B of Table 8 are qualitatively unaffected when controlling for weekday-effects.
The coefficients of institutional investors’ attention are instead now even smaller and
none of them are significant at any conventional level. Even when accounting for the
Friday-induced distraction, retail investors’ attention is still associated with increases
in volatility over the post-event period.
5.2 Heterogeneities in retail investors’ attention
In the following, we focus on our third hypothesis and analyze whether there are
heterogeneities in retail investors’ reaction to information depending on the type and
topic of the news release. Note that in Hypothesis 3 we focus on the relation between
retail investors’ attention to news and post-announcement volatility. As such we do
not report detailed results for the contemporaneous regression.22
Table 5 reports the estimated coefficients of attention variables for the regression
defined in Equation (3) with h= 1 distinguishing between the most relevant news
release types. In other words, the results summarized in Table 5 can be interpreted
as the change in volatility observed the day after a news release associated with a
22The contemporaneous regression results which account for potential heterogeneities are available
upon request from the authors. In general the contemporaneous results are in line with those ob-
tained from the regression defined in Equation (3), in the sense that we find in both cases similar
heterogeneities.
31
Electronic copy available at: https://ssrn.com/abstract=3506720
one standard deviation increase in attention on the news-event day. Instead of simply
distinguishing between earnings calls and other types of news releases, we investigate
whether the relation between volatility and attention is heterogeneous in the type of
news. Recall that we identify the most relevant news article type from the perspec-
tive of retail investors by identifying the articles most discussed on social media (see
Section 4.2). The largest increase in volatility associated with an increase in retail
attention is observed when the most relevant released information is unexpected: The
news formats “news flash” and “hot news flash” are associated with a 4.8% and 7.5% in-
crease in volatility, respectively. The estimated coefficients for institutional investors’
attention are smaller and often not significant. Only when the most relevant released
news articles are in the format of a “full article” or a “news flash” are the estimates sig-
nificant, but the volatility increases associated with a one standard deviation increase
in institutional attention amount to only 0.9% and -0.9%, respectively.
Figure 5 depicts the changes in volatility associated with a one standard deviation
increase in retail and institutional investors’ attention to press releases, news flashes
and hot news flashes, over the four days after the news release. In particular, for “press
release” and “news flash,” retail investors’ attention is associated with higher volatility
still two days after the release of the information. When institutional investors are
paying attention to these releases, volatility decreases the days after the announce-
ment. For news articles in the format of “hot news flash,” retail investors’ attention
is associated with significant increases in volatility only up to one day after the news
release.
Table 6 reports similar results but distinguishes between the most relevant topic
for retail investors. Given the large number of different news topics in our dataset,
the table reports only the most important results, i.e. those news topics for which we
find the smallest and largest estimates for changes in volatility associated with a one
standard deviation increase in retail investors’ attention. For retail investors’ atten-
tion, all estimates are positive and the majority of them also significant. On the day
32
Electronic copy available at: https://ssrn.com/abstract=3506720
Retail inv. attention
Institutional inv. attention
Press release
News flash
Hot news flash
t=0 t=1 t=2 t=3 t=4 t=0 t=1 t=2 t=3 t=4
−0.05
0.00
0.05
0.10
−0.05
0.00
0.05
0.10
−0.05
0.00
0.05
0.10
Figure 5: Change in volatility associated with a one standard deviation
increase in retail and institutional investors’ attention around different news
release types
Note: The figure depicts the fixed effects panel regression coefficients of attention
measures where the dependent variable is the realized volatility on the news event day
and the four post-release days for different news types. The gray-shaded areas represent
95% confidence intervals computed using clustered standard errors obtained with the
estimator introduced by Arellano (1987).
after a news release, the largest percentual changes in volatility associated with a one
standard deviation increase in retail investors’ attention are observed when the most
relevant published articles are about “stock prices” (5.5%), “analyst ratings” (6.6%),
“acquisition and merger” (8.5%) or “earnings” (8.8%) and “revenues” (10.4%).23 Con-
cerning institutional investors’ attention, most estimates are small in absolute terms
23Note that articles categorized as being about “earnings” or “revenues” refer to news articles about
a company’s earnings or revenues which have not been published on an earnings call days
33
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 5: Investors’ attention to news types and next day’s volatility reaction
Estimate t-statistic
Retail investor attention
No news 0.012 21.61
Not relevant news 0.010 13.12
Tabular material 0.027 17.32
Full article 0.017 9.33
Press release 0.036 18.71
News flash 0.048 25.39
Hot news flash 0.075 4.54
Earnings calls 0.030 4.90
Institutional investor attention
No news 0.002 3.62
Not relevant news 0.002 3.22
Tabular material 0.001 1.03
Full article 0.009 7.52
Press release 0.003 1.70
News flash -0.009 -5.98
Hot news flash 0.012 0.66
Earnings calls 0.005 0.90
Adjusted R253.86%
Number observations 588,104
Note: The table reports the estimated volatility increase the day after a news release associated
with a one standard deviation increase in retail and institutional investors’ attention. More precisely,
the table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news types defined
as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987).
and often not significant at any conventional level. However, the results for the news
topics “stock prices,” “analyst ratings” and “acquisition and merger” are of particular
interest: A one standard deviation increase in institutional investors’ attention is as-
sociated with a reduction in next day’s volatility of 2.2%, 1.6% and 1.7%, respectively.
Figure 6 summarizes the results for the most interesting news topics for different
time horizons. In general, the results show that when retail investors are paying
attention to these news events, volatility is not only higher on the day the news is
released but also during the subsequent four trading days. Institutional investors’
attention to a news release is only associated with a contemporaneous increase in
volatility. Further, when the most relevant articles are about stock prices, mergers
and acquisitions, or analyst ratings, institutional investors’ attention is associated
34
Electronic copy available at: https://ssrn.com/abstract=3506720
with a decrease in volatility over the four days after the release of the news. In other
words, when the most relevant published articles are about topics directly related to a
company’s market value, high attention of retail investors is associated with an increase
in volatility over the following trading days, whereas institutional investors’ attention
appears to increase the speed at which prices adjust to the new information reducing
the post-event volatility. Indeed, Ben-Rephael et al. (2017) find a similar result for
news about changes in analyst recommendations, showing that higher institutional
attention is associated with a faster price-adjustment to the new information. Our
results extend this finding by showing that this relation is also valid for other news
topics, and that retail investors’ attention has the opposite relation, i.e. it reduces the
price-adjustment speed.
In summary, we find clear evidence that the volatility increase associated with
investors’ attention is not homogeneous depending on the type or the topic of the most
relevant articles for retail investors. In particular, we find that when retail investors are
paying attention to an unexpected news event or to a news article about topics directly
linked to a company’s market value (e.g. mergers and acquisitions, analyst ratings),
the relation is stronger. Due to their complexity, the implications of these topics
for a company’s market value are generally difficult to understand, in particular for
unsophisticated investors. This finding confirms our third hypothesis. Moreover, while
for most news topics institutional investors’ attention has no significant relation to the
post-event volatility, we find that when sophisticated investors are paying attention
to news releases about analyst ratings, mergers and acquisitions, or stock prices, the
market risk decreases in the four days after the publication of the new information.
35
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 6: Investors’ attention to news topics and next day’s volatility reaction
Estimate t-statistic
Retail investor attention
No news 0.012 21.57
Not relevant news 0.010 13.05
Insider trading 0.010 5.95
Order imbalances 0.013 5.80
Credit 0.016 1.88
Credit ratings 0.020 4.29
Corporate responsibility 0.024 1.24
Technical analysis 0.026 16.57
.
.
..
.
.
Stock prices 0.055 9.81
Analyst ratings 0.066 23.65
Acquisitions and mergers 0.085 18.19
Regulatory 0.088 6.13
Earnings 0.088 12.75
Revenues 0.104 7.35
Earnings call 0.031 4.93
Institutional investor attention
No news 0.002 3.68
Not relevant news 0.003 3.26
Insider trading 0.011 8.21
Order imbalances -0.007 -3.02
Credit -0.009 -1.45
Credit ratings -0.004 -1.10
Corporate responsibility -0.002 -0.09
Technical analysis 0.002 1.20
.
.
..
.
.
Stock prices -0.022 -4.24
Analyst ratings -0.016 -6.11
Acquisitions and mergers -0.017 -3.93
Regulatory 0.005 0.33
Earnings 0.000 0.02
Revenues 0.002 0.21
Earnings call 0.005 0.90
Adjusted R253.94%
Number observations 588,104
Note: The table reports the estimated volatility increase the day after a news release associated with
a one standard deviation increase in retail and institutional investors’ attention. More precisely, the
table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news topics
defined as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987). Note that the table only
reports partial results (the news articles with the largest absolute estimates for retail investors);
results for the omitted events are reported in Table 9 in Appendix A.
36
Electronic copy available at: https://ssrn.com/abstract=3506720
Retail inv. attention
Institutional inv. attention
Stock prices
Analyst ratings
Acquisitions and mergers
Regulatory
Earnings
Revenues
t=0 t=1 t=2 t=3 t=4 t=0 t=1 t=2 t=3 t=4
0.00
0.05
0.10
0.00
0.05
0.10
0.00
0.05
0.10
0.00
0.05
0.10
0.00
0.05
0.10
0.00
0.05
0.10
Figure 6: Change in volatility associated with a one standard deviation
increase in retail and institutional investors’ attention around different news
release topics
Note: The figure depicts the fixed effects panel regression coefficients of attention
measures where the dependent variable is the realized volatility on the news-event day
and the four post-release days for different news topics. The gray-shaded areas represent
95% confidence intervals computed using clustered standard errors obtained with the
estimator introduced by Arellano (1987).
37
Electronic copy available at: https://ssrn.com/abstract=3506720
6 The economic value of accounting for attention around
news releases
In this section, we show that the empirical findings reported in the previous section
can be used to generate an economic value for investors. More precisely, we propose
two prediction models and use out-of-sample volatility forecasts in a portfolio selection
problem for a representative investor. This approach allows us to compare how much
of her wealth the investor is willing to give up in order to have access to the volatility
predictions of the more sophisticated model.
We follow the volatility prediction literature and define the benchmark model as
the heterogeneous autoregressive realized volatility (HAR) model introduced by Corsi
(2009) augmented by the (log) VIX level, the turnover-ratio and dummy variables for
news releases and earnings calls. Specifically, building upon the regression framework
used in this study, we obtain predictions from the following fixed effects panel HAR
model:
log RVi,t+1 =ci+θ(d)log RVi,t +θ(w)log RV (w)
i,t +θ(m)log RV (m)
i,t +X0
i,tθ0+i,t+1 (6)
where Xi,t contains the (log) VIX, the turnover-ratio and dummy variables for news
releases. More precisely, we include dummies for earnings calls, days without any news
articles, days with relevant news releases and not relevant ones.24 The second model
additionally includes the measure of retail investors’ attention interacted with the news
dummies. We estimate the fixed effects panel models using the within-transformation
over a rolling window of four years and produce one-day ahead forecasts of (log) realized
volatility over the following year. The starting estimation window ranges from January
2011 to December 2014. In order to obtain true ex-ante predictions, all the involved
24We categorized a day to only have not relevant news releases when the cumulative relevance score
across all articles published for a specific company is zero. Recall that the relevance score is defined as
the total cosine-similarity between news headlines and messages shared on social media (see Section
4.2).
38
Electronic copy available at: https://ssrn.com/abstract=3506720
variables are constructed using only data from the rolling window. For example, when
removing trend and seasonal components from the (log) posting volumes, we estimate
the regression coefficients using only data from the estimation window.
Results regarding the improvements in prediction accuracy achieved by accounting
for retail investors’ attention when predicting one-day ahead (log) realized volatility
are relegated to Appendix C. On average, retail investors’ attention reduces the mean-
squared prediction error by 1.17%. Moreover, the inclusion of the attention measure
based on StockTwits data improves the predictive accuracy for 351 stocks, and for 242
these improvements are statistically significant at the 5% level.
6.1 Evaluating the economic value of volatility predictions
The evaluation of the economic value of the volatility predictions closely follows the
approach introduced by Bollerslev et al. (2018). In particular, we extend their approach
to a multivariate setting, i.e. for multiple risky assets. We assume that a representative
investor has mean-variance preferences:
Et[u(Wt+1)] = Et[Wt+1 ]k
2Vart[Wt+1]
where Et[·]and Vart[·]denote the conditional expectation and variance, Wt+1 is the
investor’s wealth and kdenotes the absolute risk aversion. In each trading period, the
investor can allocate her wealth in a risk-free asset and nrisky stocks. Her wealth there-
fore evolves as Wt+1 =Wt(1 + (1 ω|
t1)Rf,t+1 +ω|
tRt+1)where ωtis a n-dimensional
vector of the shares of wealth invested in the risky assets, Rtis a random vector of
the risky assets’ returns, and Rf,t is the risk-free rate. Assuming that the relative risk
aversion γ=kWtis constant, the expected utility the investor wishes to maximize is:
U(ωt) = Wt1 + Rf,t+1 +ω|
tEt[Re
t+1]γ
2ω|
tCovt[Re
t+1]ωt
39
Electronic copy available at: https://ssrn.com/abstract=3506720
where Re
t+1 =Rt+1 Rf,t+1 is the vector of excess returns of the risky assets and
Covt[Re
t+1]is their conditional covariance matrix. As in Bollerslev et al. (2018), we
assume that the Sharpe ratio of each asset is constant over time, but we allow for differ-
ent sharp ratios across different sectors.25 The conditional expected excess returns are
Et[Re
t+1] = SREt[RVt+1 ]where RVt+1 is a n-dimensional vector of realized volatilities,
SR an-dimensional vector of Sharpe ratios, and denotes the element-wise product
of the two vectors. In this way, we avoid influencing the economic evaluation of volatil-
ity predictions by the modeling of conditional expected excess returns. Moreover, we
assume that the risky assets have a constant correlation matrix Γ, and that the con-
ditional covariance matrix is therefore Covt[Re
t+1] = diag(Et[RVt+1 ]) Γ diag(Et[RVt+1])
where diag(x)stands for the diagonal matrix with diagonal entries given by the vector
x.
The shares of wealth invested in the risky assets that maximize the expected utility
are:
ω
t=1
γdiag(Et[RVt+1]1) Γ1SR
and the optimal allocation implied by the predictions of a model is then:
b
ωt=1
γdiag(b
Et[RVt+1]1) Γ1SR
where b
Et[·]denotes the conditional expectation from a model, i.e. the one-day ahead
prediction.26 The conditional expected utility obtained with the asset allocation b
ωtis:
U(b
ωt) =Wt 1 + Rf,t+1 +1
γSR|Γ1diagEt[RVt+1]
b
Et[RVt+1]SR
1
2γSR|Γ1diagEt[RVt+1]
b
Et[RVt+1]ΓdiagEt[RVt+1 ]
b
Et[RVt+1]Γ1SR!
25In the empirical application, we estimate sector-specific Sharpe ratios. Sectors are determined
based on Standard & Poor’s Global Industry Classification Standard (GICS).
26Note that we predict the log realized volatility, i.e. b
Et[log RVt+1]. We approximate the prediction
of the realized volatility by b
Et[RVt+1]exp b
Et[log RVt+1].
40
Electronic copy available at: https://ssrn.com/abstract=3506720
and the expected utility achieved with the asset allocation ω
tis:
U(ω
t) = Wt 1 + Rf,t+1 +1
2γSR|Γ1SR!
It is easily seen that when only one risky asset is available, we obtain the same results
as in Bollerslev et al. (2018). In order to economically evaluate the utility achieved
with the predictions made by a model, Bollerslev et al. (2018) propose computing the
empirical expected utility per unit of wealth by its realized average:
UoW {b
ωt}T1
t=0 =T1
T1
X
t=0 1 + Rf,t+1 +1
γSR|Γ1diagRVt+1
b
Et[RVt+1]SR
1
2γSR|Γ1diagRVt+1
b
Et[RVt+1]ΓdiagRVt+1
b
Et[RVt+1]Γ1SR!
The economic value of the predictions b
Et[RVt+1]can then be quantified by comparing
the above realized average utility with that obtained when using the volatility pre-
dictions of a competing model. The value thus obtained can be interpreted as the
fraction of wealth an investor with mean-variance preference is willing to give up for
the volatility predictions produced by a more sophisticated model.
This approach can be further extended by taking into account transaction costs.
Specifically, we assume that the costs for trading in the assets are linear in the absolute
magnitude of the change in the respective position, i.e. |bωt,i bωt1,i|. The expected
utility a mean-variance investor is maximizing becomes:
U(ωt) = Wt1 + Rf,t+1 +ω|
tEt[Re
t+1]γ
2ω|
tCovt[Re
t+1]ωt− |ωtωt1||T Ct
where T Ctis a n-dimensional vector of transaction costs. Since the objective function
is no longer differentiable, we obtain the optimal allocation by maximizing the expected
utility with a cyclic coordinate descent algorithm (see Appendix D).
41
Electronic copy available at: https://ssrn.com/abstract=3506720
6.2 The value of attention for news-event portfolios
In the following, we analyze the value of retail investors’ attention for volatility predic-
tions in a multivariate setting, i.e. a mean-variance investor can allocate her wealth in
several risky assets and a risk-free asset. We build specific news-event portfolios that
invest exclusively in stocks for which a particular news topic or type has been published
on the day the asset allocation is determined. More precisely, on each trading day, the
investor only considers those companies for which a news article about a specific topic
(e.g. “analyst ratings”) or of a particular type (e.g. “press release”) has been released
on that day. When no stock is subject to a specific news event, the investor stays out
of the markets and invests her entire wealth in the risk-free asset. We determine the
economic value of accounting for retail investors’ attention by comparing the average
realized utility achieved by the two volatility prediction models.
The risk-free rate is the effective federal fund rate. The Sharpe-ratio is estimated
over the rolling estimation window at the industry level and kept constant over the
following year. Similarly, the correlation matrix is estimated using daily returns in the
rolling estimation window and kept fixed for the following year. Following Bollerslev
et al. (2018), we estimate transaction costs with the median bid-ask spread over the
previous nine months. In line with the literature, we set the relative risk aversion to
be 2.
Table 7 summarizes the annual performance fee (in basis points) the investor is
willing to pay for attention-based volatility predictions for different news-event portfo-
lios. The first column reports for how many days at least one company is subject to a
specific news event. The second column indicates whether transaction costs have been
taken into account, the third reports the performance fee and the last one t-statistics
of the difference in average realized utility using Newey-West standard errors.
When trading only in companies with irrelevant news releases to which retail in-
vestors do not pay attention, the performance fee the investor is willing to pay is rather
small and amounts to 10.08 basis points. In fact, when also considering transaction
42
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 7: The economic value of volatility predictions for event-based portfolios
Num.
events
Transaction
costs
Performance
fee in bps
t-statistic
Not relevant news 528 No 10.08 2.81
Yes 2.51 1.62
News flash 754 No 131.68 7.90
Yes 44.11 8.46
Press release 753 No 23.47 1.90
Yes 12.04 5.31
Acquisitions and mergers 681 No 67.22 5.84
Yes 34.67 5.79
Analyst ratings 703 No 62.13 5.92
Yes 32.67 5.26
Revenues 335 No 25.01 4.59
Yes 11.52 4.66
Earnings call 528 No 82.70 7.28
Yes 58.78 8.09
Note: The table summarizes the fraction of wealth an investor with mean-variance preferences is
willing to give up for the volatility predictions obtained from the attention-based model compared to
those obtained from the benchmark model (annualized in basis points). The investor only trades in
stocks subject to specific news types and topics. The first column reports for how many days at least
one company is subject to that event. The second column indicates whether transaction costs have
been taken into account or not. The last column reports t-statistics of the performance fees, using
Newey-West standard errors.
costs, the fee becomes even smaller and is no longer significant at any conventional
level. For the other news-events, the fees are large and highly significant. When invest-
ing in stocks for which a news article about acquisition and mergers has been released,
the investor is willing to give up 67.22 basis points per year of her wealth, and 34.67
when considering transaction costs for predictions based on attention measures.
These findings are confirmed by cross-sectional results obtained at the company
level. More precisely, we also analyze the economic value of retail investors’ atten-
tion for a mean-variance investor who can allocate her wealth in either the risk-free
or a single risky asset, the latter being one of the 360 companies considered in our
study. The results are reported in Table 10 in Appendix A. In particular, we find that
attention-based volatility forecasts lead to a higher average realized utility for 348
companies, and for 244 the improvement is significant at the 5% significance level of
a two-sided test. Moreover, the cross-sectional results reflect the same heterogeneities
presented previously. When no news or only irrelevant news is released, the economic
43
Electronic copy available at: https://ssrn.com/abstract=3506720
value of attention is low. Similarly, when articles about “credit” or “corporate respon-
sibility” are released, the value of taking into account retail investors’ attention is
low. In contrast, the mean-variance investor is willing to give up more of her wealth
for attention-based forecasts when, for example, a “news flash” or an article about
“acquisition and mergers” is released.
In summary, the news-event portfolios clearly show that taking into account retail
investors’ attention around news releases for predicting volatility is valuable in eco-
nomic terms. These results strengthen the findings presented in this paper showing
that our hypotheses hold not only in a classical fixed effect panel regression framework
but also translate into economically meaningful volatility predictions.
7 Conclusion
In this study, we introduce three hypotheses related to the behavioral finding of peo-
ple’s limited cognitive ability. More precisely, we investigated the hypotheses that
(i) when investors pay attention to news releases, the contemporaneous volatility in-
creases, (ii) the volatility observed in the subsequent trading days is positively associ-
ated with higher attention of retail investors and negatively with that of institutional
investors, and finally (iii) that this relation is heterogeneous in the type of the infor-
mation release and the topic covered by the news. Using a novel measure of retail
investors’ attention obtained from the daily posting volume on social media platforms,
we find significant evidence in favor of all three hypotheses.
Our findings show that, consistent with the noise trader models, when retail in-
vestors are paying attention to news releases, the price-adjustment to the new informa-
tion is slowed down and the market risk increases. In contrast, institutional investors’
attention to the release of news fosters price-adjustments and reduces the volatility in
the post-announcement period.
Building upon these findings, we propose a volatility prediction model that accounts
for retail investors’ attention. To quantify the economic value of these predictions, we
44
Electronic copy available at: https://ssrn.com/abstract=3506720
extend the approach introduced by Bollerslev et al. (2018) to a multivariate setting. In
particular, our findings show that around the release of unexpected news or news about
specific topics (e.g. analyst ratings, acquisition and mergers), volatility predictions
that take into account retail investors’ attention have significant economic value for
an investor with mean-variance preferences.
In summary, our findings have several implications for the academic community
and practitioners. First, we show that the attention of sophisticated institutional in-
vestors is of central importance for the price-adjustment speed to new information.
This finding is also of relevance for the asset pricing literature in that prices do not
instantaneously reflect all available information. Moreover, the type of news release
and the topic of the new information have an influence on how prices react when retail
investors are paying attention. These findings might be incorporated in models of in-
vestors’ inattention by allowing for a heterogeneous reaction to different news releases.
Finally, the findings presented in this paper contribute to the volatility prediction lit-
erature by showing that measures of retail investors’ attention around the release of
news significantly improve forecasts.
45
Electronic copy available at: https://ssrn.com/abstract=3506720
References
Aboody, D., Lehavy, R., & Trueman, B. (2010). Limited attention and the earn-
ings announcement returns of past stock market winners. Review of Accounting
Studies,15 (2), 317–344.
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Ebens, H. (2001). The distribution
of realized stock return volatility. Journal of Financial Economics,61 (1), 43 -
76.
Andrei, D., & Hasler, M. (2014). Investor Attention and Stock Market Volatility. The
Review of Financial Studies,28 (1), 33-72.
Antweiler, W., & Frank, M. Z. (2004). Is All That Talk Just Noise? The Information
Content of Internet Stock Message Boards. The Journal of Finance,59 (3),
1259–1294.
Arellano, M. (1987). Computing Robust Standard Errors for Within-groups Estima-
tors. Oxford Bulletin of Economics and Statistics,49 (4), 431-434.
Audrino, F., Sigrist, F., & Ballinari, D. (2020). The impact of sentiment and attention
measures on stock market volatility. International Journal of Forecasting,36 (2),
334-357.
Barber, B. M., & Odean, T. (2008). All That Glitters: The Effect of Attention and
News on the Buying Behavior of Individual and Institutional Investors. The
Review of Financial Studies,21 (2), 785-818.
Barberis, N., Shleifer, A., & Vishny, R. (1998). A model of investor sentiment. Journal
of Financial Economics,49 (3), 307 - 343.
Barndorff-Nielsen, O. E., Hansen, P. R., Lunde, A., & Shephard, N. (2009). Realized
kernels in practice: trades and quotes. The Econometrics Journal,12 (3), C1–
C32.
Ben-Rephael, A., Carlin, B. I., Da, Z., & Israelsen, R. D. (2019). Information Con-
sumption and Asset Pricing. Retrieved from https://ssrn.com/abstract=
3002339 (Working paper)
46
Electronic copy available at: https://ssrn.com/abstract=3506720
Ben-Rephael, A., Da, Z., & Israelsen, R. D. (2017). It Depends on Where You Search:
Institutional Investor Attention and Underreaction to News. The Review of
Financial Studies,30 (9), 3009-3047.
Black, F. (1986). Noise. The Journal of Finance,41 (3), 528-543.
Bollerslev, T., Hood, B., Huss, J., & Pedersen, L. H. (2018). Risk Everywhere:
Modeling and Managing Volatility. The Review of Financial Studies,31 (7),
2729-2773.
Boudoukh, J., Feldman, R., Kogan, S., & Richardson, M. (2019). Information, Trading,
and Volatility: Evidence from Firm-Specific News. The Review of Financial
Studies,32 (3), 992-1033.
Brown, S. V., & Tucker, J. W. (2011). Large-Sample Evidence on Firms’ Year-over-
Year MD&A Modifications. Journal of Accounting Research,49 (2), 309-346.
Buncic, D., & Gisler, K. I. (2016). Global equity market volatility spillovers: A broader
role for the United States. International Journal of Forecasting,32 (4), 1317 -
1339.
Bushee, B. J., Core, J. E., Guay, W., & Hamm, S. J. (2010). The Role of the Business
Press as an Information Intermediary. Journal of Accounting Research ,48 (1),
1-19.
Cookson, J. A., & Niessner, M. (2019). Why Don’t We Agree? Evidence from a Social
Network of Investors. Journal of Finance,Forthcoming.
Corsi, F. (2009). A Simple Approximate Long-Memory Model of Realized Volatility.
Journal of Financial Econometrics,7(2), 174-196.
Da, Z., Engelberg, J., & Gao, P. (2011). In Search of Attention. The Journal of
Finance,66 (5), 1461-1499.
Da, Z., Engelberg, J., & Gao, P. (2014). The Sum of All FEARS Investor Sentiment
and Asset Prices. The Review of Financial Studies,28 (1), 1-32.
DellaVigna, S., & Pollet, J. M. (2009). Investor inattention and friday earnings
announcements. The Journal of Finance,64 (2), 709-749.
47
Electronic copy available at: https://ssrn.com/abstract=3506720
De Long, J. B., Shleifer, A., Summers, L. H., & Waldmann, R. J. (1990). Noise trader
risk in financial markets. Journal of political Economy,98 (4), 703–738.
Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of
Business & Economic Statistics,13 (3), 253-263.
Dimpfl, T., & Jank, S. (2016). Can Internet Search Queries Help to Predict Stock
Market Volatility? European Financial Management ,22 (2), 171-192.
Drake, M. S., Roulstone, D. T., & Thornock, J. R. (2012). Investor Information De-
mand: Evidence from Google Searches Around Earnings Announcements. Jour-
nal of Accounting Research,50 (4), 1001-1040.
Drake, M. S., Roulstone, D. T., & Thornock, J. R. (2015). The Determinants and
Consequences of Information Acquisition via EDGAR. Contemporary Accounting
Research,32 (3), 1128-1161.
Gabaix, X. (2019). Chapter 4 - Behavioral inattention. In Handbook of behavioral
economics - foundations and applications 2 (Vol. 2, p. 261 - 343). North-Holland.
Garcia, D. (2013). Sentiment during Recessions. The Journal of Finance,68 (3),
1267-1300.
Goddard, J., Kita, A., & Wang, Q. (2015). Investor attention and fx market volatility.
Journal of International Financial Markets, Institutions and Money,38 , 79 -
96.
Groß-Klußmann, A., & Hautsch, N. (2011). When machines read the news: Using
automated text analytics to quantify high frequency news-implied market reac-
tions. Journal of Empirical Finance ,18 (2), 321 - 340.
Hamid, A., & Heiden, M. (2015). Forecasting volatility with empirical similarity and
google trends. Journal of Economic Behavior & Organization ,117 , 62 - 81.
Hansen, P. R., & Lunde, A. (2005). A Realized Variance for the Whole Day Based
on Intermittent High-Frequency Data. Journal of Financial Econometrics,3(4),
525-554.
Hirshleifer, D., Lim, S. S., & Teoh, S. H. (2009). Driven to Distraction: Extraneous
48
Electronic copy available at: https://ssrn.com/abstract=3506720
Events and Underreaction to Earnings News. The Journal of Finance,64 (5),
2289-2325.
Hirshleifer, D. A., & Sheng, J. (2019). The Attention Trigger Effect: Macro News
and Efficient Processing of Micro News. Retrieved from https://ssrn.com/
abstract=2890864 (Working paper)
Kahneman, D. (1973). Attention and effort (Vol. 1063). Englewood Cliffs, NJ:
Prentice-Hall.
Lee, C. M., Ma, P., & Wang, C. C. (2015). Search-based peer firms: Aggregating in-
vestor perceptions through internet co-searches. Journal of Financial Economics,
116 (2), 410 - 431.
Lerman, A. (2020). Individual investors’ attention to accounting information: Ev-
idence from online financial communities. Contemporary Accounting Research ,
Forthcoming.
Liu, H., Peng, L., & Tang, Y. (2019). Investor Attention: Endogenous Allocations,
Clientele Effects, and Asset Pricing Implications. (Working paper)
Livnat, J., & Mendenhall, R. R. (2006). Comparing the Post–Earnings Announcement
Drift for Surprises Calculated from Analyst and Time Series Forecasts. Journal
of Accounting Research,44 (1), 177-205.
Mittnik, S., Robinzonov, N., & Spindler, M. (2015). Stock market volatility: Identify-
ing major drivers and the nature of their impact. Journal of Banking & Finance,
58 (Supplement C), 1 - 14.
Morse, D. (1982). Wall Street Journal Announcements and the Securities Markets.
Financial Analysts Journal,38 (2), 69-76.
Peng, L. (2005). Learning with Information Capacity Constraints. The Journal of
Financial and Quantitative Analysis,40 (2), 307–329.
Peng, L., & Xiong, W. (2006). Investor attention, overconfidence and category learn-
ing. Journal of Financial Economics,80 (3), 563 - 602.
Rakowski, D., Shirley, S. E., & Stark, J. R. (2018). Twitter activity, investor attention,
49
Electronic copy available at: https://ssrn.com/abstract=3506720
and the diffusion of information. Financial Management .
Rubin, A., & Rubin, E. (2010). Informed investors and the internet. Journal of
Business Finance & Accounting,37 (7-8), 841-865.
Ryan, P., & Taffler, R. J. (2004). Are Economically Significant Stock Returns and
Trading Volumes Driven by Firm-specific News Releases? Journal of Business
Finance & Accounting,31 (1-2), 49-82.
Shleifer, A., & Vishny, R. W. (1997). The Limits of Arbitrage. The Journal of Finance,
52 (1), 35-55.
Vlastakis, N., & Markellos, R. N. (2012). Information demand and stock market
volatility. Journal of Banking & Finance,36 (6), 1808 - 1821.
50
Electronic copy available at: https://ssrn.com/abstract=3506720
Appendices
A Tables
Table 8: Investors’ attention and volatility reaction to news controlling for weekday-
effects
Panel A: Contemporaneous relation between attention and volatility
Estimate t-statistic
Retail investor attention
No news 0.028 18.58
News, no earnings call 0.042 17.36
Earnings calls 0.036 10.49
Institutional investor attention
No news 0.038 19.95
News, no earnings call 0.048 17.59
Earnings calls 0.033 11.11
Adjusted R259.07%
Number observations 587,762
Panel B: Relation between attention and next day’s volatility
Estimate t-statistic
Retail investor attention
No news 0.013 22.62
News, no earnings call 0.022 27.84
Earnings calls 0.033 5.46
Institutional investor attention
No news 0.000 -0.40
News, no earnings call 0.000 0.11
Earnings calls 0.005 0.96
Adjusted R254.23%
Number observations 588,104
Note: The table reports the estimated volatility increase on the news release day (Panel A) and
the day after (Panel B) associated with a one standard deviation increase in retail and institutional
investors’ attention, controlling for potential weekday-effects. More precisely, Panel A summarizes
the attention estimates of the regression defined in Equation (1), and Panel B those of the regression
defined in Equation (2) with h= 1. For completeness, the table also reports the estimated increase in
contemporaneous and future volatility associated with a one standard deviation increase in attention
when no news article is released. The t-statistics are computed using clustered standard errors
obtained with the estimator introduced by Arellano (1987).
51
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 9: Investors’ attention to news topics and next day’s volatility reaction
Estimate t-statistic
Retail investor attention
No news 0.012 21.57
Not relevant news 0.010 13.05
Insider trading 0.010 5.95
Order imbalances 0.013 5.80
Credit 0.016 1.88
Credit ratings 0.020 4.29
Corporate responsibility 0.024 1.24
Technical analysis 0.026 16.57
Investor relations 0.027 8.02
Partnerships 0.028 5.67
Dividends 0.029 7.01
Equity actions 0.029 5.48
Marketing 0.030 5.29
Products and services 0.030 9.87
Legal 0.033 3.83
Assets 0.035 4.62
Price targets 0.041 6.05
Labor issues 0.041 10.36
Stock prices 0.055 9.81
Analyst ratings 0.066 23.65
Acquisitions and mergers 0.085 18.19
Regulatory 0.088 6.13
Earnings 0.088 12.75
Revenues 0.104 7.35
Earnings call 0.031 4.93
Note: The table reports the estimated volatility increase the day after a news release associated with
a one standard deviation increase in retail and institutional investors’ attention. More precisely, the
table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news topics
defined as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987).
52
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 9: (Continued) Investors’ attention to news topics and next day’s volatility
reaction
Estimate t-statistic
Institutional investor attention
No news 0.002 3.68
Not relevant news 0.003 3.26
Insider trading 0.011 8.21
Order imbalances -0.007 -3.02
Credit -0.009 -1.45
Credit ratings -0.004 -1.10
Corporate responsibility -0.002 -0.09
Technical analysis 0.002 1.20
Investor relations -0.004 -1.62
Partnerships -0.007 -1.44
Dividends 0.005 1.42
Equity actions 0.002 0.32
Marketing 0.006 1.44
Products and services 0.002 0.85
Legal 0.001 0.20
Assets 0.002 0.29
Price targets -0.004 -0.70
Labor issues 0.005 1.58
Stock prices -0.022 -4.24
Analyst ratings -0.016 -6.11
Acquisitions and mergers -0.017 -3.93
Regulatory 0.005 0.33
Earnings 0.000 0.02
Revenues 0.002 0.21
Earnings call 0.005 0.90
Adjusted R253.94%
Number observations 588,104
Note: The table reports the estimated volatility increase the day after a news release associated with
a one standard deviation increase in retail and institutional investors’ attention. More precisely, the
table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news topics
defined as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987).
53
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 10: The economic value of retail investors’ attention in predicting volatility
(effective annual performance fee in basis points)
N TC Q25% Avg. Med. Q75%
Overall
754 No 1.07 4.21 2.99 5.61
Yes -0.23 1.50 0.31 1.77
No news and earnings calls
Not relevant news 86 No 0.00 0.11 0.04 0.13
Yes -0.08 0.01 0.00 0.06
Earnings call 12 No 0.00 0.51 0.17 0.74
Yes -0.01 0.64 0.15 0.93
News types
Tabular material 76 No 0.00 0.34 0.18 0.53
Yes -0.09 0.16 0.00 0.08
Full article 71 No 0.03 0.56 0.25 0.76
Yes -0.06 0.11 0.00 0.11
Press release 36 No 0.00 0.28 0.10 0.39
Yes -0.04 0.05 0.00 0.04
News flash 41 No 0.17 1.40 0.67 1.51
Yes -0.08 0.29 0.00 0.12
Hot news flash 6 No 0.13 0.41 0.41 0.55
Yes -0.02 0.06 0.01 0.03
News topics
Insider trading 60 No 0.00 0.32 0.11 0.46
Yes -0.05 0.06 0.00 0.09
Credit 8 No -0.04 0.02 -0.01 0.02
Yes 0.00 0.00 0.00 0.01
Credit ratings 9 No -0.03 0.02 0.00 0.04
Yes -0.02 0.01 0.00 0.02
Corporate responsibility 7 No -0.02 -0.01 -0.01 0.00
Yes -0.02 -0.01 -0.01 0.00
Technical analysis 72 No 0.00 0.32 0.16 0.49
Yes -0.08 0.15 0.00 0.06
.
.
..
.
..
.
..
.
.
Stock prices 14 No 0.01 0.74 0.13 0.95
Yes -0.02 0.21 0.00 0.14
Analyst ratings 12 No 0.05 0.41 0.15 0.52
Yes -0.02 0.06 0.00 0.04
Acquisitions and mergers 11 No 0.02 0.68 0.20 0.71
Yes -0.02 0.22 0.00 0.05
Regulatory 6 No 0.33 0.42 0.48 0.58
Yes 0.01 0.02 0.02 0.03
Earnings 8 No 0.00 0.08 0.04 0.15
Yes -0.01 0.05 0.00 0.02
Revenues 18 No 0.02 0.86 0.12 0.75
Yes -0.01 0.39 0.02 0.22
Note: The table shows summary statistics of the fraction of wealth an investor is willing to give up
in order to obtain volatility predictions from the attention-based model compared to the benchmark
model. The second column summarizes the average number of observations, and the column TC
indicates whether transaction costs have been taken into account. For each grouping criterion, we
only consider companies which have at least five observations in that category.
54
Electronic copy available at: https://ssrn.com/abstract=3506720
B Results using Twitter attention
Table 11: Investors’ attention and volatility reaction using Twitter data
Panel A: Contemporaneous relation between attention and volatility
Estimate t-statistic
Retail investor attention
No news 0.020 17.12
News, no earnings call 0.030 15.45
Earnings calls 0.015 4.15
Institutional investor attention
No news 0.038 18.78
News, no earnings call 0.048 16.92
Earnings calls 0.032 10.49
Adjusted R258.03%
Number observations 587,762
Panel B: Relation between attention and next day’s volatility
Estimate t-statistic
Retail investor attention
No news 0.009 14.31
News, no earnings call 0.018 23.48
Earnings calls 0.003 0.48
Institutional investor attention
No news 0.002 3.60
News, no earnings call 0.003 4.19
Earnings calls 0.005 0.99
Adjusted R253.69%
Number observations 588,104
Note: The table reports the estimated volatility increase on the news release day (Panel A) and
the day after (Panel B) associated with a one standard deviation increase in retail and institutional
investors’ attention. More precisely, Panel A summarizes the attention estimates of the regression
defined in Equation (1), and Panel B those of the regression defined in Equation (2) with h= 1. For
completeness, the table also reports the estimated increase in contemporaneous and future volatility
associated with a one standard deviation increase in attention when no news article is released. The
t-statistics are computed using clustered standard errors obtained with the estimator introduced by
Arellano (1987). Retail investors’ attention is measured with Twitter data.
55
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 12: Investors’ attention to news types and next day’s volatility reaction using
Twitter data
Estimate t-statistic
Retail investor attention
No news 0.009 14.35
Not relevant news 0.007 5.16
Tabular material 0.013 12.70
Full article 0.012 8.49
Press release 0.022 15.42
News flash 0.030 20.05
Hot news flash 0.060 4.57
Earnings calls 0.003 0.46
Institutional investor attention
No news 0.002 3.68
Not relevant news 0.000 0.10
Tabular material 0.004 4.82
Full article 0.007 6.08
Press release 0.003 2.57
News flash -0.004 -3.85
Hot news flash 0.009 0.53
Earnings calls 0.005 0.99
Adjusted R253.72%
Number observations 588,104
Note: The table reports the estimated volatility increase the day after a news release associated
with a one standard deviation increase in retail and institutional investors’ attention. More precisely,
the table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news types defined
as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987).
56
Electronic copy available at: https://ssrn.com/abstract=3506720
Table 13: Investors’ attention to news topics and next day’s volatility reaction using
Twitter data
Estimate t-statistic
Retail investor attention
No news 0.009 14.36
Not relevant news 0.006 4.72
Credit ratings 0.003 0.78
Insider trading 0.007 5.18
Order imbalances 0.009 6.27
Credit 0.010 1.47
Legal 0.012 2.02
Investor relations 0.012 4.59
.
.
..
.
.
Stock prices 0.040 6.93
Regulatory 0.056 3.83
Analyst ratings 0.056 18.29
Earnings 0.057 10.22
Acquisitions and mergers 0.065 17.30
Revenues 0.084 7.25
Earnings call 0.003 0.49
Institutional investor attention
No news 0.002 3.82
Not relevant news 0.000 0.11
Credit ratings -0.003 -1.12
Insider trading 0.008 6.72
Order imbalances -0.004 -2.55
Credit -0.005 -0.81
Legal 0.007 1.18
Investor relations 0.003 1.63
.
.
..
.
.
Stock prices -0.018 -3.15
Regulatory -0.005 -0.34
Analyst ratings -0.013 -5.84
Earnings 0.005 1.23
Acquisitions and mergers -0.019 -5.01
Revenues 0.011 1.30
Earnings call 0.005 1.00
Adjusted R253.80%
Number observations 588,104
Note: The table reports the estimated volatility increase the day after a news release associated with
a one standard deviation increase in retail and institutional investors’ attention. More precisely, the
table reports the coefficients for attention measures obtained from the estimation of the regression
defined in Equation (3) with h= 1. In particular, we distinguish between different news topics
defined as the most discussed articles by retail investors. The t-statistics are computed using clustered
standard errors obtained with the estimator introduced by Arellano (1987). Note that the table only
reports partial results (the news articles with the largest absolute estimates for retail investors).
57
Electronic copy available at: https://ssrn.com/abstract=3506720
C Predictive accuracy of attention-based volatility
forecasts
In the following, we briefly present the results when analyzing the predictive power of
retail investors’ attention in forecasting volatility. The models and estimation tech-
niques used are described at the beginning of Section 6. The accuracy of the forecasts
is measured in terms of mean-squared prediction error (MSE). We assess the signifi-
cance of the differences in prediction accuracy with the Diebold-Mariano test statistic
(Diebold & Mariano, 1995). Table 14 reports summary statistics (quartiles, median
and average) of reductions in MSE as well as corresponding Diebold-Mariano test
statistics when comparing predictions from a model including attention variables with
those obtained without attention variables for the cross-section of 360 stocks. Figure
7 depicts these results graphically.
Both Table 14 and Figure 7 clearly show that for the great majority of companies,
we improve the one-day ahead volatility forecasts by accounting for retail investors’
Reduction in MSE (in %)
DM−statistic
−1 0 1 2 3 4 5 0 2 4 6 8
0
10
20
30
40
Figure 7: Improvements in volatility predictions achieved by accounting
for retail investors’ attention
Note: The figure depicts the percentual reduction in MSE (left) and the respective
Diebold-Mariano statistics (right) achieved by the addition of retail investors’ attention
in a predictive volatility model. The vertical line indicates the threshold for the 5%
significance level of a two-sided Diebold-Mariano test.
58
Electronic copy available at: https://ssrn.com/abstract=3506720
attention. In fact, for 351 of the stocks, the MSE is reduced when including the
attention measure based on StockTwits data, and for 242 the improvement is significant
at the 5% level.
Table 14: Predictive accuracy of retail investors’ attention for one-day ahead realized
volatility
Q25% Avg. Med. Q75%
MSE reduction (in %) 0.70 1.17 1.17 1.62
DM statistic 1.37 2.11 2.19 2.83
Note: The first summarizes summary statistics of the reduction in MSE (in %) of one-day ahead
prediction of (log) realized volatility achieved by augmenting a benchmark model with retail investors’
attention. The second row reports instead summary statistics of Diebold-Mariano test statistics for
the difference in predictive accuracy.
59
Electronic copy available at: https://ssrn.com/abstract=3506720
D Maximization of expected mean-variance utility
with transaction costs
The objective function to be maximized is the expected mean-variance utility:
U(ωt) = Wt1 + Rf,t+1 +ω|
tEt[Re
t+1]γ
2ω|
tCovt[Re
t+1]ωt− |ωtωt1||T Ct
Dropping terms which do not affect the optimization and rewriting the objective func-
tion in terms of ωt=ωtωt1:
U(∆ωt)(ωt1+∆ωt)|Et[Re
t+1]γ
2(ωt1+∆ωt)|Covt[Re
t+1](ωt1+ωt)−|ωt||T Ct
where ωt1is the asset allocation of the previous trading period and therefore can be
treated as being known and fixed. We maximize U(∆ωt)with respect to ωtusing
a cyclical coordinate descent algorithm. Even though U(∆ωt)is non-differentiable, it
has directional derivatives along each forward and backward coordinate:
vkU= lim
τ0
U(∆ωt+τvk)U(∆ωt)
τ
=Et[Re
t+1]γ
n
X
j=1
(wt1,j + ∆ωt,j )Covt[Re
t+1,k, Re
t+1,j ] +
T Ct,k ωt0
T Ct,k ωt<0
vkU= lim
τ0
U(∆ωtτvk)U(∆ωt)
τ
=Et[Re
t+1] + γ
n
X
j=1
(wt1,j + ∆ωt,j )Covt[Re
t+1,k, Re
t+1,j ] +
T Ct,k ωt>0
T Ct,k ωt0
60
Electronic copy available at: https://ssrn.com/abstract=3506720
where vkis the coordinate direction along which ωt,k varies. For k= 1, . . . , n, we
then iteratively update ωt,k with:
ωt,k = ∆˜ω+
t,k + ∆˜ω
t,k
where ˜ω+
t,k and ˜ω
t,k are the candidate solutions to the right and left of zero (only
one of them is non-zero). They are defined as:
˜ω+
t,k = max ωt,k +Et[Re
t+1]γPn
j=1(wt1,j + ∆ωt,j )Covt[Re
t+1,k, Re
t+1,j ]T Ct,k
γVart[Re
t+1,k],0!
˜ω
t,k = min ωt,k +Et[Re
t+1]γPn
j=1(wt1,j + ∆ωt,j )Covt[Re
t+1,k, Re
t+1,j ] + T Ct,k
γVart[Re
t+1,k],0!
We repeat the iterative update of ωtuntil the improvements in utility no longer
exceed a tolerance level η.
61
Electronic copy available at: https://ssrn.com/abstract=3506720
Preprint
Full-text available
The present study demonstrates modelling of conditional volatility of NIFTY 50 using GARCH (1,1) model. The daily returns data of the Indian stock market index NIFTY 50 is used for the period ranging from April 2010- March 2020. The data is analysed using R software. The study estimates and interprets the results arrived in the summary output of the R environment and demonstrates how to forecast the volatility of the returns based on the estimated parameters. Extracting the time series of conditional volatilities is also demonstrated in the study.
Article
Full-text available
We examine the impact of Twitter attention on stock prices by examining over 21 million company specific tweets over a five‐year period. Through a quasi‐natural experiment identifying official Twitter outages, we first show that Twitter influences stock trading, especially among small, less visible securities primarily traded by retail investors. Next, we find that Twitter activity is associated with positive abnormal returns, and when tweets occur in conjunction with traditional news events, more information is spread to investors. Finally, we show that retail investor activity drives the Twitter effect, as institutional investors less actively trade the affected stocks. This article is protected by copyright. All rights reserved
Article
Full-text available
Based on high-frequency data for more than fifty commodities, currencies, equity indices, and fixed-income instruments spanning more than two decades, we document strong similarities in realized volatility patterns within and across asset classes. Exploiting these similarities through panel-based estimation of new realized volatility models results in superior out-of-sample risk forecasts, compared to forecasts from existing models and conventional procedures that do not incorporate the similarities in volatilities. We develop a utility-based framework for evaluating risk models that shows significant economic gains from our new risk model. Lastly, we evaluate the effects of transaction costs and trading speed in implementing different risk models. Received March 7, 2016; editorial decision February 3, 2018 by Editor Andrew Karolyi. Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.
Article
What moves stock prices? Prior literature concludes that the revelation of private information through trading, and not public news, is the primary driver. We revisit the question by using textual analysis to identify fundamental information in news. We find that this information accounts for 49.6% of overnight idiosyncratic volatility (vs. 12.4% during trading hours), with a considerable fraction due to days with multiple news types. We use our measure of public information arrival to reinvestigate two important contributions in the literature related to individual |$R^{2}$|s of stock returns on aggregate factors. Received February 24, 2016; editorial decision May 19, 2018 by Editor Andrew Karolyi.
Article
Online financial communities provide a unique opportunity to directly examine individual investors’ attention to accounting information on a large scale and in great detail. I analyze accounting‐related content in large samples of Yahoo! message board posts and StockTwits and find investors pay attention to a range of accounting information, fixating particularly on earnings, cash, and revenues. Consistent with the expectation that investors react to relevant information events, I find accounting‐related discussion elevated around the filings of earnings releases and 8‐K reports, but the reaction to periodic reports is confined to small firms. I also find investors expand their acquisition of accounting information and processing efforts in poor information environments. Greater attention to accounting information at earnings releases does not appear to be meaningfully associated with better information processing. This article is protected by copyright. All rights reserved.
Article
We study sources of investor disagreement using sentiment of investors from a social media investing platform, combined with information on the users' investment approaches (e.g., technical, fundamental). We examine how much of overall disagreement is driven by different information sets versus differential interpretation of information by studying disagreement within and across investment approaches. Overall disagreement is evenly split between both sources of disagreement, but within‐group disagreement is more tightly related to trading volume than cross‐group disagreement. Although both sources of disagreement are important, our findings suggest that information differences are more important for trading than differences across market approaches. This article is protected by copyright. All rights reserved
Article
We analyze the impact of sentiment and attention variables on stock market volatility by using a novel and extensive dataset that combines social media, news articles, information consumption, and search engine data. Applying a state-of-the-art sentiment classication technique, we investigate the question of whether sentiment and attention measures contain additional predictive power for realized volatility when controlling for a wide range of economic and nancial predictors. Using a penalized regression framework, we identify investors' attention, as measured by the number of Google searches on nancial keywords (e.g. "nancial market" and "stock market"), and the daily volume of company-specic short messages posted on StockTwits to be the most relevant variables. In addition, our study shows that attention and sentiment variables are able to signicantly improve volatility forecasts, although the improvements are of relatively small magnitude from an economic point of view.
Article
We study the relationship between investors? active attention, measured by a Google search volume index (SVI), and the dynamics of currency prices. Investor attention is correlated with the trading activities of large FX market participants. Investor attention comoves with comtemporaneous FX market volatility and predicts subsequent FX market volatility, after controlling for macroeconomic fundamentals. In addition, investor attention is related to the currency risk premium. Our results suggest that investor attention is a priced source of risk in FX markets.