Content uploaded by Charles Martineau
Author content
All content in this area was uploaded by Charles Martineau on Aug 31, 2024
Content may be subject to copyright.
Social Media-Driven Noise Trading: Liquidity Provision
and Price Revelation Ahead of Earnings Announcements
Edna Lopez Avila, Charles Martineau, and Jordi Mondria∗
August 29, 2024
ABSTRACT
Social media attention before earnings announcements is overly optimistic, fails to pre-
dict fundamentals, and generates buying pressure, leading to a 58 bps stock return
as intermediaries seek higher returns for providing liquidity. Such price pressure dis-
torts the price informativeness of fundamentals. A return reversal occurs immediately
following announcements as markets correct mispricing. How stock prices respond to
earning news is endogenous to the effect of social media in the pre-announcement price
formation. A pre-announcement trading strategy based on expected social media at-
tention yields 40 bps monthly alphas. When noise trading is systematically driven, it
can deter liquidity provision and price revelation.
JEL Classification: G12, G14, G40.
Keywords: earnings announcements, inventory risks, investor attention, liquidity pro-
vision, price efficiency, return reversal, social media
∗A previous version of this paper was circulated under the title “When Crowds Aren’t Wise: Biased
Social Networks and its Price Impact.” We thank Pat Akey, Ing-Haw Cheng, Alex Chinco, Roberto Cram,
Alex Corhay, Valeria Fedyk, Goutham Galapakrishna, Vincent Gregoire, Ryan Israelsen, Runjing Lu, Marina
Niessner, Yoshio Nozawa, Terry Odean, Chay Ornthanalai, Mike Simutin, Christopher Shiller, and Xiaofei
Zhao for valuable feedback. We thank the seminar participants at the 2023 NFA, 2023 Future of Financial
Information, 2023 Trans-Atlantic Doctoral, 2023 Waseda Summer Workshop in Finance, 2024 MFA, 2024
SOFiE conferences. We acknowledge the financial support from the Canadian Securities Institute, the Social
Sciences and Humanities Research Council (SSHRC), and TD MDAL. Jay Cao at TD MDAL and Kevin Li
provided excellent research assistance. Lopez Avila: Rotman School of Management, University of Toronto,
105 St-George, Toronto ON, Canada, M5S 3E6 (edna.lopez.avila@rotman.utoronto.ca), Martineau: Rotman
School of Management and UTSC Management, University of Toronto, 105 St-George, Toronto ON, Canada,
M5S 3E6 (charles.martineau@rotman.utoronto.ca), Mondria: University of Toronto, 150 St. George Street,
Toronto ON, Canada, M5S 3G7 (jordi.mondria@utoronto.ca).
1. Introduction
Classical frameworks of market efficiency (Friedman et al.,1953,Fama,1965) consider only
the presence of investors who possess valuable skills in evaluating a firm’s prospects for future
fundamentals as the main source of information affecting price efficiency. Investors that trade
on irrelevant information (“noise traders”) are assumed to be normally distributed, have no
permanent price impact, and are met in the market by arbitrageurs who trade against them
and quickly eliminate mispricing (Grossman and Stiglitz,1980,Kyle,1985). Noise trading
has no role in price formation.
Social investment networks have become the primary source of investment advice for
many retail investors (Cookson, Mullins, and Niessner,2024), a driver of persistent flow of
non-fundamental trading motives (Pedersen,2022,Waters, Zhang, Zhou, and Santosh,2024),
and are an attraction device generating systematic noise trading (Barber, Huang, Odean,
and Schwarz,2022).1While Barber, Odean, and Zhu (2008,2009) show that systematic
noise trading can move prices, how it affects price efficiency and the informational content
of stock prices remains unclear.2
This paper focuses on the friction that arises from social media-driven systematic noise
trading to price revelation.3Models of price formation suggest that trades originating from
social media should not affect price revelation if they reflect “noise.” But when such noise is
systematically driven, it can worsen liquidity provision as intermediaries face an imbalance
between the quantity sought by buyers and sellers at a given price. Intermediaries may
1Heimer (2016) find that peer interactions on social networks influence investors’ trading decisions.
De Long, Shleifer, Summers, and Waldmann (1990) propose a model of correlated noise trading impact-
ing informational efficiency.
2Han and Yang (2013) analyze a rational expectations equilibrium model with endogenous information
acquisition and argue that social communication networks can worsen price revelation.
3We follow the terminology of Brunnermeier (2005) to distinguish between two components of price
revelation: “price efficiency” and “price informativeness”. Price efficiency relates to the price revelation of
public information and how fast information is impounded into prices and price informativeness reflects the
absolute level of information to future fundamentals (see also Biais, Hillion, and Spatt,1999,Weller,2018,
Boguth, Fisher, Gregoire, and Martineau,2024).
1
absorb the order imbalance into their own account but through price concession to minimize
inventory risks by setting prices above (below) fundamental value in response to buy (sell)
order imbalances (So and Wang,2014). Such price concession will worsen the informational
content of stock prices.
We exploit the cross-sectional variation in the coverage of stocks on social media in
the days leading to earnings announcements. Such days are periods of high information
uncertainty (Akey, Gr´egoire, and Martineau,2022) and low liquidity where traders abstain
from trading (Lee, Mucklow, and Ready,1993). Liquidity providers’ demand curves are
steeper during that time due to increases in inventory risks associated with unexpected
price changes following announcements (So and Wang,2014). Consequently, stock prices
are expected to be more sensitive to systematic noise trading driven by social media chatter
ahead of earnings announcements.
The main punchline of our paper is that stocks widely discussed on social media are
associated with retail buying pressure, generating a significant positive price drift ahead
of earnings announcements as intermediaries require a higher rate of return to mitigate
inventory risks. Such systematic noise trading worsens price revelation ahead of earnings
announcements as the price concession required by intermediaries adds noise to prices, and
a price reversal follows after the earnings news release. Conditioning on the fundamental
news, i.e., earnings surprises, stock prices can diverge away from fundamentals ahead of
announcements, but markets quickly correct the mispricing following the announcement. A
key implication to future research is that theories should consider the role of systematic
noise trading in price formation. Moreover, understanding how social media influences pre-
announcement price dynamics is critical. Peng, Wang, and Zhou (2022), Campbell, Drake,
Thornock, and Twedt (2023), Hirshleifer, Peng, and Wang (2024) have examined the role of
social media in price formation following earnings announcements. We show that examining
only the effect of social media on price formation following announcements can lead to biased
2
inferences, as post-announcement price responses are endogenous to the effect of social media
to pre-announcement price formation.
We focus on the social network StockTwits, the largest investor-focused social media
platform, which averages more than one million monthly posts covering most stocks. Our
sample covers the period of 2013 to 2022 and provides the widest coverage of stocks, more
than any other social media platform. Cookson, Lu, Mullins, and Niessner (2022) find that
StockTwits’ coverage correlates more strongly with returns than other popular social media
platforms (i.e., Twitter or SeekingAlpha) and that StockTwits fairly represents the aggregate
information shared on social media. We find that StockTwits activity increases five days
before earnings announcements. This increase in coverage is not observed for other news
sources such as newswires. StockTwits activity is not only concentrated in large stocks (top
NYSE breakpoint quintile). Small market capitalization (bottom quintile) stocks receive as
much coverage as large firms.
Aggregating StockTwits users’ self-labeled “bullish” and “bearish” sentiment tags reveal
a positive bias among users. Over 60% of the stock-earnings announcement observations
exhibit a bullish outcome, defined as having more than 80% of the sentiment-tagged posts
about that specific stock-earnings announcement observation tagged bullish. We show that
users’ sentiments do not predict earnings announcement fundamentals, i.e., earnings sur-
prises.4We further show that such excess positivism displayed on StockTwits transmits into
correlated retail buying pressure ahead of announcements.5
4Previous studies find that the content of social media posts can predict earnings surprises, (e.g., Chen,
De, Hu, and Hwang,2014,Dim,2020,Bartov, Faurel, and Mohanram,2018). The first two papers use
Seeking Alpha, a platform where non-anonymous users can write lengthy articles. Seeking Alpha provides
limited coverage before announcements, and only a handful of stocks receive coverage. Bartov, Faurel, and
Mohanram (2018) employ Twitter data for an earlier period, specifically 2009-2012.
5Bradley, Hanousek Jr, Jame, and Xiao (2023) and Hu, Jones, Zhang, and Zhang (2021) examines the
social media platform Reddit/Wall Street Bets and find that greater Reddit attention relates to more retail
trading. Kakhbod, Kazempour, Livdan, and Schuerhoff (2023) classify StockTwits users into “skilled”,
“unskilled”, and “antiskilled” and find that 56% of users are “antiskilled” and create optimistic beliefs and
are the most influential in changing followers’ beliefs.
3
When intermediaries (market makers) face an uninformed and balanced buying and sell-
ing order flow, prices remain efficient (Grossman and Stiglitz,1980). But when faced with
systematic noise trading, the findings of So and Wang (2014) suggest that intermediaries will
require a higher expected return to meet the demand as they seek liquidity in the opposite
direction to hedge their inventory risk. We find that stocks with high StockTwits attention
(i.e., top attention quintile) generate a 58 bps price pressure ahead of announcements and
a 45 bps price reversal on announcement dates. By comparison, the average return to a
comparable matched sample of stocks with low attention shows no evidence of price reversal.
Additional tests show that the effect of StockTwits’ attention to price pressure and reversals
is robust to earnings surprises, firm size, the number of analyst following, and news cover-
age before announcements. We further show that buying pressure induced by social media
activity ahead of announcements predicts the reversal.6
We then examine the cross-sectional variation in the reversal magnitudes. If the buying
pressure is a result of inventory risk, the reversal magnitude will be greater for stocks with
ex-ante higher anticipated return volatility on announcement dates. As argued by So and
Wang (2014), intermediaries can take several days to unwind their net positions. Conse-
quently, providing liquidity ahead of scheduled news events, characterized by high volatility,
increases their exposure to inventory risks (Madhavan and Smidt,1993) and the risk of mar-
gin calls (Comerton-Forde, Hendershott, Jones, Moulton, and Seasholes,2010). We use firm
size7, option implied volatility prior to announcements, and the average of previous earn-
ings announcement dates’ absolute returns as proxies of inventory risks (anticipated return
volatility). For stocks with high StockTwits attention, the reversal magnitudes are higher for
small firms and stocks with high option implied volatility and prior earnings announcement
date absolute returns. Low-StockTwits attention stocks show little cross-sectional varia-
6Price pressure ahead of earnings announcements as a result of social media lends support to the findings
of Laarits and Sammon (2023) that retail-heavy stocks are more expensive to trade ahead of announcements.
7As shown in Martineau (2022), smaller firms have larger absolute returns on announcement dates.
4
tion in reversals. The increased inventory risks arising from social media-driven systematic
noise trading have a striking influence on the autocorrelation of returns and the information
content of stock prices.
We next show that social media stock activity ahead of earnings announcements can
distort the price informativeness of stocks’ fundamentals. Conditioning on announcements’
earnings surprises, the average price run-up before announcements pushes prices closer to
fundamentals for positive earnings surprise announcements, albeit not a reflection of infor-
mative trading. However, for negative earnings surprises, the buying price pressure ahead
of announcements deviates prices away from fundamentals. On the announcement date,
markets quickly adjust prices to public information and correct any mispricing generated
by social media information production. In other words, the release of public information
quickly eliminates mispricing caused by the social-media-driven price pressure. We find no
evidence that social media attention results in slower price formation following announce-
ments.8This last result contrasts with the findings of Campbell, Drake, Thornock, and
Twedt (2023). They find that social media activity on Twitter slows price formation fol-
lowing earnings announcements.9We show that post-announcement price formation are
endogenous to pre-announcement price responses caused by social media attention.10 Our
paper illustrate why one would find social media to be associated with slower price discovery
post-announcement if one disregards pre-announcement price dynamics.
Our results exhibit a contemporaneous relationship between attention and price pressure
ahead of announcements. This raises the question of whether intermediaries can foresee
which stocks will gather the most social media attention ahead of earnings announcements
and if there is a trading strategy to exploit the price pressure.11 We developed a trading
8See Martineau (2022) for a review on the evolution of stock return post-earnings announcement drifts.
9Other papers that examine the relationship between social media and post-earnings announcement price
dynamics include Ding, Shi, and Zhou (2023) and Hirshleifer, Peng, and Wang (2024).
10See Boguth, Gr´egoire, and Martineau (2019) and Fisher, Martineau, and Sheng (2022) for evidence of
investor attention increasing before scheduled macroeconomic announcements.
11We do not argue that this strategy is expected to be profitable when taking into account transaction
5
strategy that buys stocks expected to receive high attention, sells those expected to receive
low attention, and closes the position the day before the announcement. We used posts from
Earnings Whispers, a firm known for its earnings forecasts and strong social media following,
to indicate which stocks might receive more social media attention. At the beginning of each
week, Earnings Whispers posts on StockTwits and other social media platforms the week’s
most anticipated earnings releases. We find that stocks mentioned in these posts attract
more attention from retail investors, generate more retail buying pressure, and earn higher
returns in the days leading up to their earnings announcements compared to stocks not
mentioned. A pre-announcement long-short strategy earns monthly abnormal returns of 45
basis points.
We present a simple theoretical framework explaining how social media may influence
optimistic trading behaviors that can result in systematic noise trading. The model is based
on the concept of “wishful thinking” from Caplin and Leahy (2019), which posits that indi-
viduals derive utility from their beliefs and thus tend to interpret information optimistically.
The model predicts that investors will display positive (negative) optimism when seeking
to buy (sell) stocks. It is well-known that retail investors are more inclined to buy than
sell (Barber and Odean,2008) and, consistent with our findings, we expect investors to dis-
play more positive optimism. Furthermore, our model indicates that this optimistic bias is
stronger when information is easily manipulated, as evidenced by the significant engagement
with Earnings Whispers’ posts on StockTwits. This simple model provides interesting av-
enues for future research to explore the role of social media in shaping investors’ beliefs and
trading behaviors ahead of scheduled announcements.
costs.
6
1.1. Related literature and contributions
This paper contributes to the growing literature examining the role of social media in finan-
cial markets. Antweiler and Frank (2004) find a link between stock market volatility and
messages on online investment message boards. Cookson and Niessner (2020) is the first
paper to study the content of StockTwits messages. The authors find a relationship between
users’ disagreement and trade volume. Cookson, Fos, and Niessner (2021) find that greater
investor disagreement measured from StockTwits facilitates informed trading and short sell-
ers. Cookson, Engelberg, and Mullins (2020) study StocksTwits users’ political partisanship
and how it influences investors’ expectations in the wake of COVID-19. More closely related
to our paper, Cookson, Engelberg, and Mullins (2023) find that StockTwits users choose se-
lective exposure to confirmatory information, i.e., echo chambers.12 With retail traders more
likely to buy than to sell a stock (Barber and Odean,2008), such users are more likely to be
influenced by positive information, which can explain the positive-biased sentiment of Stock-
Twits posts ahead of earnings announcements. Our paper further shows that prices reflect
the positive-biased sentiment, but markets correct mispricing following public information.
Cookson, Niessner, and Schiller (2022) find that corporate managers are also influenced by
the content of StockTwits when evaluating the prospects of a merger and acquisition. The
recent episodes of GameStop and other “meme-stocks” provided compelling evidence of cor-
related noise trading resulting from social media (Bradley, Hanousek Jr, Jame, and Xiao,
2023). Our paper shows that episodes of correlated noise trading resulting from social media
are much more pervasive than documented in the media and such noise trading impacts
market markers inventory risks and price revelation. Jia, Redigolo, Shu, and Zhao (2020)
find that social media posts from Twitter does not predict rumor realization of mergers.
Similarly, we find that social media posts do not predict earnings fundamentals.
12Jiao, Veiga, and Walther (2020) provide evidence consistent with echo chambers from an aggregate
source of social media platforms collected by MarketPsych Data.
7
Other research examines the impact of alternative social media outlets on asset prices and
retail trading, such as Reddit/Wall Street Bets (Kang, Lou, Ozik, Sadka, and Shen,2024),
Seeking Alpha (Chen, De, Hu, and Hwang,2014,Dim,2020,Farrell, Green, Jame, and
Markov,2022), and Twitter (Jiao, Veiga, and Walther,2020,Campbell, Drake, Thornock,
and Twedt,2023,Bianchi, G´omez-Cram, and Kung,2024). In contrast to these existing
papers, we document how social media influences retail trading ahead of earnings announce-
ments, how it can impact inventory risks and price revelation, and how quickly markets
correct social media-driven mispricing following the release of public information.
Our paper further relates to the growing literature on retail trading. Barber and Odean
(2000), Barber and Odean (2008), Barber, Lee, Liu, and Odean (2009), and Barber, Huang,
Odean, and Schwarz (2022) find that retail investors are generally uninformed and make
systematic mistakes when selecting stocks. Another strand of the literature finds otherwise.
Hvidkjaer (2008), Kaniel, Saar, and Titman (2008), Kaniel, Liu, Saar, and Titman (2012),
Kelley and Tetlock (2013), Barrot, Kaniel, and Sraer (2016), Boehmer, Jones, Zhang, and
Zhang (2021) show that retail order imbalance positively predicts future returns at short
horizons. We also find that retail order imbalances are positively associated with returns
but that such trades do not reflect fundamentals and lead to a price reversal on announcement
dates. Peress and Schmidt (2021) find that noise trading follows the common assumption
of i.i.d.-normal only at monthly and lower frequencies but at daily frequencies to be more
correlated. We show that social media can be responsible for such correlated noise trading.
Finally, Peress and Schmidt (2020) show that market makers are more concerned with trading
against insiders than noise traders. When faced with systematic noise trading ahead of
scheduled announcements, we show that market makers are concerned and require a higher
expected return to meet the demand from noise traders.
8
2. Data and Methodology
2.1. StockTwits
We retrieve data on StockTwits posts from January 2013 to December 2022 through Rapi-
dAPI, starting our analysis from 2013 to ensure quality stock coverage as per Cookson and
Niessner (2020). By identifying posts with $-tagged tickers (e.g., $AAPL), we matched these
to tickers in the CRSP database, totaling 150,262,272 stock-specific posts. Figure 1 presents
the total monthly number of stock-specific posts on StockTwits and reveals a significant
increase in posting activity during the COVID-19 pandemic, with monthly posts exceeding
four million, compared to approximately one million posts per month before the pandemic.
This activity returned to pre-pandemic levels in 2022.
On StockTwits, users can label their posts as “bullish” or “bearish” to express their
sentiment toward a stock.13 They can also indicate their trading experience level as novice,
intermediate, or professional.14 We follow Cookson and Niessner (2020) in assigning posts
made after 4 p.m. to the following trading day. This approach allows us to align our analysis
with daily stock returns, calculated from 4 p.m. to 4 p.m. the next trading day. Weekend
posts are similarly assigned to the next trading day.
A stock’s “attention” (or coverage) and user sentiments on StockTwits are the main two
variables that we construct. When measuring how much attention a stock receives on social
media, the standard approach is to divide the total number of posts for a stock on a given
day by the total number of posts on the platform on that day. Comparing a particular
stock post activity to other stocks’ posts controls for the non-stationary time-series in posts
activity on StockTwits. We compute StockTwits attention as follows:
Atti,t =posti,t
PN
iposti,t
,(1)
1356% of our stock-day observations have sentiment-tagged StockTwits posts.
14According to Cookson and Niessner (2020), 20% of users identify as professionals, 52% as intermediates,
and 28% as novices.
9
where posti,t is the number of posts for stock ion date t. The denominator is the sum of
posts for all stocks on StockTwits on date t.
To measure investors’ sentiment associated with StockTwits messages on a given day,
we calculate the daily proportion of bullish posts to the total number of bullish and bearish
posts for stock ion the day t, as follows:
Senti,t =bulli,t
bulli,t +beari,t
,(2)
where “bull” and “bear” correspond to the number of posts with bullish and bearish tags,
respectively. When calculating attention (sentiment) over a longer window, e.g., 5 days, we
sum the number of posts (bullish posts) over the 5-day period and divide the total number
of StockTwits posts over the 5-day period (sum of bearish and bullish posts).
2.2. Earnings, news, and stock-level data
We supplement our StockTwits data with analyst forecasts and earnings announcement dates
from Thomson Reuters I/B/E/S. We include earnings announcements in IBES that meet
the following criteria: the earnings date is reported in Compustat, the stock price five days
before the announcement is available in CRSP, and the stock price is available in Compustat
as of the end of the quarter. We calculate the Surprise in earnings announcements as the
difference between the firm’s earnings per share for the quarterly earnings announcement and
the consensus analysts’ forecast, divided by the stock price five days prior to the earnings
announcement day. We compute analysts’ forecasts by taking the median of all analysts’
estimates issued within the 90 days preceding the earnings announcement date. Lastly, we
winsorize the earnings surprise at the 1st and 99th percentiles.
Following Gregoire and Martineau (2022), we also gather analyst recommendation news
events and other firm-level news for our sample of stocks from Ravenpack. All news events
occurring after 4 p.m. or on weekends are attributed to the next trading day.
10
Additionally, we retrieve daily stock returns from CRSP, the five factors from Kenneth
French’s website, and intraday trading data from TAQ. With the trading data, we follow
the methodology outlined in Boehmer, Jones, Zhang, and Zhang (2021) and the suggested
adjustments in Barber, Huang, Jorion, Odean, and Schwarz (2024) to identify trades by
retail investors and construct various retail trading measures such as order imbalances.
Table 1 reports summary statistics for high and low-attention stocks. We define high-
attention stocks if Att belongs to the top quintile five days before earnings announcements
for stocks with the same earnings announcement dates. High-attention stocks have a higher
average market capitalization, stock price volatility, absolute abnormal returns on announce-
ments, and analyst following. A key takeaway from this table is that when comparing high-
attention stocks to low-attention stocks, it is important to use a set of matched low-attention
stocks as pre-earnings liquidity and asset price dynamics vary across stocks (Liu, Wang, Yu,
and Zhao,2020). Also, stocks that receive the highest StockTwits activity ahead of an-
nouncements are more volatile, with higher absolute returns on the announcement date.
These stock characteristics attract the most investor attention (Barber and Odean,2008).
2.3. StockTwits activity around earnings announcements
Table 2 reports a breakdown of the coverage across NYSE market capitalization breakpoint
quintile. It shows the count of stock-earnings observations with at least one StockTwits
message, one analyst recommendation, or one newswire mention from five to one day be-
fore the announcements. Additionally, the table reports the number of observations without
StockTwits posts, analyst recommendations, or newswire coverage. A key insight from this
table is the broader scope of StockTwits in covering stocks before earnings announcements
compared to analysts and newswires. Approximately 24% of pre-announcement StockTwits
posts pertain to the smallest firms, and 37% to the largest. In contrast, analyst recommen-
dations and newswire reports before earnings announcements predominantly focus on the
11
largest stocks, accounting for 66% of recommendations and 73% of newswire, respectively.
Only 7% of the earnings announcements in the sample lacked StockTwits posts in the five
days leading up to the announcements. For the smallest stocks, only 13% lacked StockTwits
messages. However, the absence of analyst recommendations and newswire coverage for the
smallest stocks significantly jumps to 98% and 81%, respectively, highlighting a disparity in
coverage based on firm size.
We plot in Figure 2 the abnormal activity in StockTwits posts and newswires articles
using boxplots five days before to five days after earnings announcements. Abnormal post
(newswire coverage) activity is computed as the daily log number of posts (newswire) minus
the log of the average daily number of posts (newswire) from 20 to 6 days before the earnings
announcements. The figure shows an increase in abnormal StockTwits post activity in
the days approaching earnings announcements, with a notable 50% increase on the day
before the announcement. In contrast, newswire activity shows a modest rise on the day
before the earnings announcements, yet below the benchmark period (t= [−20,−6]) and
increases following announcements.15 This figure highlights the significant role of social
media networks in disseminating information about stocks before earnings announcements,
bridging a gap not covered by traditional news sources. Unlike newswires, which often report
earnings results post-release, social media platforms enable investors to monitor real-time
discussions and sentiments regarding a stock leading up to its earnings announcement.
We then examine users’ sentiment on StockTwits in the 60 days leading up to earnings
announcements and compare it to analysts’ recommendations.16 Figure 3 shows the fraction
of stock-earnings observations according to StockTwits tagged-sentiment and analyst recom-
mendation sentiment. We define sentiment as in equation (2) and split sentiment ratio into
15Gregoire and Martineau (2022) and Li, Ramesh, Shen, and Wu (2015) show that analyst recommenda-
tions typically follow earnings announcements. We find no increase in abnormal analyst recommendations
ahead of earnings announcements.
16It has been shown that analysts’ forecasts exhibit predictable biases (Kothari, So, and Verdi,2016,
Van Binsbergen, Han, and Lopez-Lira,2023) and over-optimism (Cowen, Groysberg, and Healy,2006).
12
five buckets: [0-20%], (20-40%], (40-60%], (60-80%], and (80-100%], i.e., from very bearish to
very bullish. Sentiment on StockTwits regarding upcoming earnings is predominantly posi-
tive. Over 60% of the stock-earnings announcement observations exhibit a bullish outcome,
defined as having more than 80% of the sentiment-tagged posts about that specific stock-
earnings announcement observation tagged bullish, while fewer than 5% of observations show
a similar dominance of bearish posts (bucket [0-20%]). In contrast, analyst recommendations
display a less pronounced positive bias and exhibit a more balanced distribution. Approxi-
mately 55% of stock-earnings observations has over 80% bullish recommendations, and 30%
of stock-earnings observations have a majority of bearish recommendations (bucket [0-20%]).
This figure reveals a significant inclination among StockTwits users towards sharing and en-
gaging with positively biased posts on StockTwits. Selecting posts five days before earnings
announcements shows similar positive-biased sentiment on StockTwits.
3. How Informative Is Social Media Ahead of Earnings
Announcements?
We first examine the informativeness of StockTwits’ sentiment about earnings fundamentals
ahead of earnings announcements. We then investigate the relationship between StockTwits’
attention and retail trading.
3.1. StockTwits sentiment does not predict earnings fundamentals
Having determined that StockTwits users display a predominantly positive sentiment, ques-
tions arise regarding the informativeness of their posts about earnings fundamentals. We
investigate this question by estimating the following regression:
Surpi,t =β1Senti,t +β2
1
Att
i,t ×Senti,t +β3
1
Att
i,t + Γ′Controlsi,t +αi+αt+εi,t ,(3)
where Surpi,t is the earnings surprise for stock-earnings iannounced on date t,Senti,t
represents the sentiment as defined in equation (2), and
1
Att
i,t is a dummy variable set to one
13
if the stock’s StockTwits attention, as defined in equation (1), falls within the top quintile,
otherwise it is set to zero. Both sentiment and attention are computed from posts made
from five days to one day prior to earnings announcements. The regression also includes an
interaction term between sentiment and attention to examine whether stocks with a higher
volume of posts yield a more accurate sentiment prediction of earnings surprises. The control
variables are the buy-and-hold abnormal returns, sentiment from analyst recommendations
computed as in equation (2) based on recommendation outlook being bullish or bearish,
and RavenPack newswire sentiment, all measured in the five days leading up to the earnings
announcements. αiand αtcorrespond to firm- and year-fixed effects through the parameters.
Table 3 presents the results for the full sample, large caps (the top three NYSE market
capitalization quintiles), and small caps (bottom two quintiles). The model specifications
defined in columns (1)–(3) exclude stock-earnings observations with no tagged sentiment.
In columns (4)–(6), we treat stock-earnings observations without sentiment data as having
a neutral sentiment (i.e., Sent = 0.5). Across all model specifications, we find no statisti-
cally significant evidence that the sentiment (Sent) expressed in StockTwits posts predicts
earnings surprises and similarly when interacting sentiment with attention (
1
Att ×Sent).
In a robustness check, we estimate equation (3) using the change rather than the level
of sentiment. Table IA1 of the Internet Appendix reports the results and finds no statisti-
cally significant evidence that the change in sentiment predicts earnings surprises. Cookson
and Niessner (2020) find that StockTwits’ users that are self-labelled as professionals are
indeed more sophisticated than novice users. They find that professional posts’ sentiments
are positively related to future returns. We examine whether sentiment posts for novice,
intermediate, and professional self-labeled users predict earnings surprises and report the
findings in Table IA2 of the Internet Appendix. Consistent with our previous findings, we
find no statistically significant evidence that sentiment predicts earnings surprises across all
user types.
14
3.2. StockTwits activity induces buying pressure
Barber and Odean (2008) find that retail investors are net buyers of attention-grabbing
stocks, e.g., stocks in the news. We confirm this finding by examining the relationship be-
tween StockTwits attention and retail trading. We calculate retail trading orders following
the methodology of Boehmer, Jones, Zhang, and Zhang (2021), incorporating the adjust-
ments suggested by Barber, Huang, Jorion, Odean, and Schwarz (2024). We retrieve the
number of retail trades, trading volume, and dollar volume and compute retail order im-
balance measures. Barber, Lin, and Odean (2023) show that focusing on the number of
trades rather than the volume provides a more accurate reflection of attention-induced retail
trading.17 We proceed to estimate the following regression model:
Retail OIi,t =β
1
Att
i,t +αi+αt+ϵi,t,
where Retail OIi,t in the regression specification represents retail order imbalance computed
five to one day before the stock’s earnings announcement ireleased on date tusing the
number of trades, volume, dollar volume, and their corresponding detrending transformation.
We choose 50 to six before announcements as the detrending window.
We present the findings in Table 4. In all model specifications, higher attention is associ-
ated with positive retail order imbalances and is statistically significant at the 1% level. The
βestimate varies from 0.007 to 0.027, and computing retail order imbalance using trades
indicates stronger buying pressure, consistent with the arguments of Barber, Lin, and Odean
(2023) that trades best capture retail investor attention. These increases in retail order im-
balances are economically significant. For example, the point estimates in Panels A and B of
0.027 and 0.013 correspond to a four and 1.3 times increase in retail order imbalance relative
to the unconditional mean of 0.0068 and 0.017, respectively.
We also examine whether the component of the order imbalance driven by social media
17Specifically, Barber, Lin, and Odean (2023) find that smaller retail trades tend to focus on stocks that
capture significant attention and are inversely related to future returns.
15
attention relates to earnings surprises. Table IA3 of the Internet Appendix reports that the
fitted component of retail order imbalance, from regressing retail order imbalance on
1
Att,
does not predict earnings surprises, but the orthogonal component does.
4. Social Media-Driven Systematic Noise Trading, Price
Pressure, and Liquidity Provision
The preceding sections show that social media activity ahead of earnings announcements is
associated with more retail buying pressure. We next examine how the price pressure relates
to inventory risks for intermediaries and how it impacts the price revelation of earnings news.
4.1. Price pressure and return reversals
The evidence reported in Section 3 suggests that the content on StockTwits does not relate
to earnings fundamentals, and retail investors are net buyers of stocks with high StockTwits
attention. When facing buying pressure (a positive net order imbalance) ahead of announce-
ments, intermediaries should be compensated via a price concession by setting prices above
fundamental value, resulting in a positive expected return. As market markers unwind their
net position following announcements and adjust the excess of price concession will result in
a negative expected return. We follow So and Wang (2014) and use market-adjusted returns
around announcements to proxy for intermediaries’ inventory balances. We use the extent of
negative autocorrelation (i.e., return reversal) from the pre-to-post announcement as a proxy
for the expected returns that market makers demand to provide liquidity to net buyers of
high-attention social media stocks.
Figure 4 plots the difference in the buy-and-hold abnormal returns for high-attention and
matched-low-attention stocks. Matched stocks are assigned based on the firm size, industry
(GIC), buy-and-hold abnormal returns from 30 to 6 days prior to the announcement, and
earnings surprises for the same year-quarter. The figure shows a significant return divergence
16
between high- and low-attention stocks of more than 1%, with the most significant increase
occurring five days before the announcement. Following the announcement, we observed a
reversal of more than 50 bps, and the difference between high- and low-attention stocks is
not statistically different from zero one day after the announcement.
We next conduct a “diff-in-diff ” analysis to validate the robustness of our findings in Table
5. The first difference compares the effect of price pressure before earnings announcements
to using a randomly selected ‘pseudo-earnings-announcement’ date in place of the actual
announcement date. Following So and Wang (2014), we select pseudo-announcement dates
from randomly selecting a pseudo-date 50 to 20 days window prior to actual announcement
dates. Columns (1) and (2) report the average BHAR, in percent, around earnings announce-
ments (EA) and pseudo-earnings announcements for high-attention stocks, respectively, and
the difference is reported in columns (3). Columns (4) and (5) report the difference between
the low-attention-match stocks and for the full sample of low-attention stocks. Columns (6)
and (7) report the “Diff-in-Diff”.
For high-attention stocks, column (1) reports a 58 bps and -45 bps in BHAR[-5,-1] and
BHAR[0,1], respectively. Column (3) reports a 49 bps increase (t-statistic of 4.68) in BHAR
relative to the pseudo earnings dates in the five days leading earnings announcements. The
“Diff-in-Diff” columns (6) and (7) report a 65 bps (t-statistic of 5.27) and 62 bps (t-statistic of
4.10) increase in BHAR relative to low attention stocks, respectively. On the announcement
date ([0,1]), the pre-announcement increase in BHAR is reversed. Columns (3) report a
decrease of 47 bps (t-statistic=5.42). Columns (4) and (5) report no significant decrease in
abnormal returns for the low-attention stocks, and the diff-in-diff columns report a decrease
of 58 bps (t-statistic=-5.04) and 45 bps (t-statistic=-3.72), respectively.
We examine the robustness of the return reversal to alternative explanations that can
result in higher returns leading to earnings announcements. For example, it has been well-
documented that around earnings announcements, stocks earn a risk premium (Barth and
17
So,2014). Moreover, information leakage can result in pre-earnings announcement drifts
(Akey, Gr´egoire, and Martineau,2022). To further examine the robustness of price pressure
to alternative explanations, we run the following regression
BH AR[−5,−1]i,t =β
1
Att
i,t + Γ′Controlsi,t +αt+ϵi,t ,(4)
where the control variables are earnings surprises, firm size (log market capitalization), ana-
lyst following, news sentiment, and abnormal newswire coverage. We report the findings in
Table 6 for the sample comprised of high-attention StockTwits stocks and their correspond-
ing matched low-attention stocks. We further report the results for the full sample in Table
IA4 of the Internet Appendix. In the univariate analysis for the matched sample, column
(1) reports an increase in BHAR of 79 bps for high-attention stocks. With the additional
control variables and fixed effects, column (2) reports an increase in BHAR of 108 bps for
high-attention stocks. We then repeat the same analysis with this time the announcement
date return BH AR[0,1] as the dependent variable and including the BHAR[−5,−1] as an
additional control variable. Column (3) reports a 69 bps decrease in BHAR for high-attention
stocks and a 53 bps decrease in column (4) when including all the control variables and the
fixed effects.
In column (5), we replace the attention dummy in equation (4) for the fitted and residual
components from regressing retail order imbalance using trades five days before announce-
ments onto attention as defined in equation (1). Column (5) reports that the fitted compo-
nent (Retail OIfit ) predicts the reversal and not the orthogonal component. We conclude
that social-media-induced buying pressure ahead of announcements results in a price reversal
on announcement dates.
4.1.1. Inventory risks
If the reversal documented in Section 4.1 results from intermediaries requiring a higher
compensation to manage inventories when facing social-media-driven buying retail pressure,
18
the magnitude of the reversal is expected to be larger when there is greater uncertainty
regarding the market’s reaction to earnings news. We select a firm’s market capitalization,
historical absolute announcement returns, and the 3-day average implied volatility before
announcements as proxies for inventory risks. With respect to firm size, Martineau (2022)
shows that absolute returns on announcement date are higher for small firms.
Table 7 presents time-series average announcement-window returns after independently
double-sorting observations into high and matched low-attention StockTwits stocks (rows)
and a quintile sort for a corresponding inventory risk proxy (columns). Panels A to C present
the results sorting on firm size, historical announcement returns, and implied volatility,
respectively.
For high-attention stocks, the average return reversal for small stocks is -2.49 bps (0.26
bps for large stocks), -1.06 bps for stocks with high absolute announcement returns (0.08
bps for low absolute returns), and -0.72 bps for high implied volatility stocks (0.05 bps
for low implied volatility). Column “High-Low” confirms statistical significance at the 1%
between the high and low sorted groups. We do not find such relationship for low-attention
stocks, except when sorting on size. However, column “Low” (bottom size quintile) of Panel
A reports a reversal of -2.49 bps for high-attention stocks and -0.56 bps for low-attention
stocks. The difference is statistically significant with a t-statistics of -6.91. Overall, the effect
of social media-induced price pressure ahead of earnings announcements is more pronounced
for stocks where intermediaries encounter limited risk-bearing capacity.
4.2. Social Media and Price Revelation
A return reversal on earnings announcement is evidence consistent with market efficiency.
After announcements, markets correct for the “inefficiency” caused by temporary price de-
viation ahead of announcements as compensation for intermediaries to accommodate the
buying pressure. This section examines how such buying pressure impacts price informa-
19
tiveness with respect to fundamentals ahead of announcements and price efficiency following
announcements. We follow the terminology of Brunnermeier (2005) to distinguish between
“price informativeness” and “price efficiency,” the main two components of the price dis-
covery process. Price informativeness reflects the absolute level of information to future
fundamentals (see also Biais, Hillion, and Spatt,1999,Weller,2018,Boguth, Fisher, Gre-
goire, and Martineau,2024) and price efficiency relates to the price revelation of public
information and how fast information is impounded into prices.
To examine the impact of price pressure on price informativeness, we graphically depict
in Figure 5 the buy-and-hold abnormal returns (BHAR) around earnings announcements
for high and matched low-StockTwits attention stocks. Panels A and B show the BHAR
for positive earnings surprise (top two quintiles) and negative earnings surprise (bottom two
quintiles), respectively. We rescale the figure such that BHAR is equal to zero at t=−6.
Both panels show positive upward price drifts leading to earnings announcements for stocks
with high StockTwits attention. In contrast, stocks with low coverage show no price drifts
before positive earnings surprises and downward price drifts for negative earnings surprises.
Five days before the announcement, the difference between BHAR for high vs low attention
is approximately 60 bps in both panels. This figure conveys that in days leading to earnings
announcements, social media can diminish price informativeness, i.e., in the case of low
earnings surprises, prices deviate from future fundamentals to be revealed on announcement
dates. In the case of positive earnings surprises, prices converge to fundamentals, not because
social-media-induced trading reflects fundamentals but because social-media induced trading
results in buying pressure, which pushes prices toward fundamentals.
The second main insight from this figure is how markets correct mispricing upon the
release of earnings announcements. For positive earnings surprises (Panel A), markets take
into account the heightened pre-announcement level and adjust prices less than those with
low attention such that there is no significant difference in total returns. In other words,
20
markets are efficient at adjusting prices to fundamental news post-announcement, indepen-
dently of the stock’s social media popularity. In Panel B, the BHAR of stocks with high
StockTwits attention deviate from fundamentals ahead of earnings announcements with neg-
ative earnings surprises, but at the time of the announcement, the BHAR quickly converge
to those with low StockTwits attention. In the days that follow the earnings announcement,
we do not observe significant price drifts, consistent with the findings of Martineau (2022)
that markets quickly process earnings news.
These news findings are important in light of the results reported in Campbell, Drake,
Thornock, and Twedt (2023) and Ding, Shi, and Zhou (2023). These authors conclude that
stocks with more Twitter and Seeking Alpha coverage before earnings announcements lead
to smaller price reactions to earnings surprises, i.e., a lower earnings response coefficient
(ERC), and conclude that social media “slows down” the price discovery process. Their
conclusion is much different from ours. We reexamine this premise that social media atten-
tion following earnings announcements dampens the price discovery process by running the
following regression
ARi,t =β1Surpi,t +β2
1
Att
i,t +β3Surpi,t ×
1
Att
i,t + Γ′Controlsi,t +αi+αt+εi,t ,
where AR corresponds to the abnormal return on the announcement date in columns (1)–(3),
buy-and-hold abnormal return (BHAR) from the announcement date to the next trading day
in column (4), and two to five days after the announcement in column (5). Table 8 reports
the results. Columns (1) to (4) report a negative earnings response coefficient (Surpi,t ×
1
Att
i,t )
of -0.18 to -0.23 for stocks with high StockTwits attention, corresponding to approximately
a decline of 25% to the total earning response. These results lend support to the findings of
Campbell, Drake, Thornock, and Twedt (2023). However, simply examining the regression
coefficient is misleading. We show in Figure 5 that stocks with more social media activity
ahead of earnings announcement have the same efficient price level as stocks with low social
media activity following announcements. The reason we obtain a negative earnings response
21
coefficient is simple. More than 67% of earnings announcements are associated with positive
earnings surprises. Therefore, the negative relationship between social media activity and
earnings surprises results from the buying price pressure leading to earnings announcements,
which diminishes the price response to positive earnings surprises as shown in Figure 5.
Column (5) of Table 8 provides additional evidence that high StockTwits attention leading
to earnings announcements does not result in a continuation of price drifts two to five days
following announcements as the loading on Surpi,t ×
1
Att
i,t is not statistically different from
zero.
5. Is Social Media-Driven Noise Trading Predictable?
Our results up to this point find a contemporaneous relationship between social media at-
tention and price pressure leading to earnings announcements. We next show that social-
media-induced systematic retail trading is ex-ante forecastable and highlight the economic
significance of the price reversal. We propose a long-short strategy centered on investor
attention using an ex-ante measure of expected StockTwits attention to upcoming earnings
announcements.
The ex-ante measure we utilize is based on social media posts about the “most anticipated
earnings announcements” by Earnings Whispers (EW). EW aggregates analysts’ earnings
forecasts and provides insights into which stocks are likely to experience significant move-
ments after earnings announcements and are more likely to have post-earnings announce-
ment drifts. Starting in January 2016, during the weekend, Earnings Whispers releases a
post highlighting the most “anticipated” earnings announcements. Figure 6 presents two
examples of such posts. Typically, these posts are shared across all social media platforms
over the weekend. As of February 2024, EW has 150,000 followers on StockTwits, 450,000
on X(formerly known as Twitter), and 116,000 on Instagram. We leverage these posts as
an ex-ante indicator of the stocks most likely to attract retail investors.
22
Before proceeding to our long-short strategy, we first document how stock returns, at-
tention, sentiment, and retail order imbalance change conditioning on appearing on a post
made by EW. Table 9 presents the results from the following regression:
ypost
i,t =β1
1
Ewhispers
i,t + Γ′Controlsprior
i,t +αi+αt+εi,t,(5)
where ypost
i,t corresponds to the stock ibuy-and-hold abnormal return (in percent), Stock-
Twits attention (in percent), StockTwits sentiment, and retail trade order imbalance from
the beginning of the earnings week (Monday) to t−1, i.e., the day before the earnings are an-
nounced on date t.
1
Ewhispers
i,t is a dummy variable equal to one if stock iappears on the EW
“most anticipated earnings” and zero otherwise.18 The control variables (Controlsprior) are
the buy-and-hold abnormal return, StockTwits attention and sentiment, and retail order im-
balance from the prior week (Monday to Friday). We also control for the upcoming earnings
surprise (Surp) and the absolute surprise (|Surp|), as well as the abnormal news coverage
and average newswire sentiment spanning the last ten days prior to the announcement. We
assign a neutral score of 0.5 for stocks with no sentiment-tagged posts.
Table 9 reports a statistically significant (at the 1% and 5% level) positive impact of a
stock appearing in an EW post on its return, attention, sentiment, and retail order imbalance.
We find an increase in BHAR by 51 basis points, a 2.7 basis points increase in attention
(a 75% increase relative to the unconditional mean of 3.6 basis points), 2.3% increase in
sentiment (a 4% increase relative to the unconditional mean), a 6.8% increase in retail
trading, and a 1.1% increase in net retail order imbalance (a 55% increase relative to the
unconditional mean) for stocks appearing in an EW post.
Having demonstrated the impact of the EW posts on investor attention, we next form
long portfolios of stocks that appear on the EW posts and short the other stocks with
earnings for that particular week but that do not appear on the EW posts. We value-weight
18We make sure to select only Earnings Whispers posts occurring on the weekend and on Monday before
9:30 am. The number of stocks-earnings observations appearing on a EW post that have an earnings
announcement on Monday is 6.8% (431 observations) of the total sample (6,301 observations).
23
the stocks to create the portfolios.19 The portfolios are rebalanced weekly. Once we obtain
the daily portfolio returns for the long and short sides, we accumulate the daily returns to
the monthly frequency. The “Long-Short” portfolio buys the long portfolio and sells the
short portfolio. We rebalance the portfolios weekly. We execute the buy (sell) order starting
Monday at 4 p.m. and hold the stock until t-1, 4 p.m., i.e., the last session of regular trading
hours before the earnings announcement.
Table 10 presents the average monthly value-weighted long-short portfolio returns in
percentage points for the pre- and post-announcement period, where the post-announcement
period consists of the announcement date and the following trading day. The long-short
column reveals that based on this strategy, the long-short portfolio generates abnormal
returns. The long-short value-weighted portfolio consistently earns significant abnormal
returns, whether using CAPM alphas, three-factors alphas, five-factors alphas, or five-factors
with momentum alphas, which abnormal returns ranging from 0.43% per month (t-statistics
= 2.81) to 0.45% (t-statistics = 2.37). It is worth noting that the majority of this spread is
attributable to the long side, where the abnormal returns for the long portfolio range from
0.51% to 0.55% per month (t-statistics=2.29 and t-statistics=3.21, respectively) whereas the
abnormal returns for the short side range from 0.06% to 0.11%. If one forms a long-short
portfolio using the same strategy and hold the portfolio over t= [0,1], columns “Post-
announcement” report that it results in a negative alpha ranging from -0.22% to -0.34%.
The negative alphas are due to the return reversal.
In the Internet Appendix, Figure IA1 shows the time series of the long, short, and long-
short cumulative returns since the initial portfolio inception. The largest increase in the
performance occurs at the outset of the COVID-19 pandemic, a period of growing retail
trading (Ozik, Sadka, and Shen,2021,Martineau and Zoican,2023) and investor atten-
tion to social media (see Figure 1). This figure provides evidence that the dynamics in
19From 2016 to 2022, we were not able to retrieve the EW “most anticipated earnings” posts for a total
of 15 weeks in our sample. When a post is missing, we replace the missing week with the risk-free rate.
24
social-media-induced retail trading impact the magnitude of price pressure ahead of earn-
ings announcements.
6. Implications to Future Research
We next discuss alternative social media platforms, precisely, the Reddit forum WallStreet-
Bets and Seeking Alpha, and demonstrate why StockTwits is the most appropriate platform
to examine the impact of social media on stock prices ahead of earnings announcements. We
then present a simple theoretical framework to rationalize why investors consume optimisti-
cally biased information and trade on such information. This simple model provides fruitful
avenues for future theoretical work to explore the role of social media in shaping investors’
beliefs and in generating systematic noise trading.
6.1. Alternative social media platforms
StockTwits is not the only social media platform examined in the literature. Seeking Alpha
and the Reddit forum WallStreetBets are two other platforms that have been examined.
Cookson, Lu, Mullins, and Niessner (2022) highlights the importance of distinguishing sen-
timent and attention across different investor social media platforms.20 Chen, De, Hu, and
Hwang (2014) and Dim (2020) find that post sentiment on the social media platform Seek-
ing Alpha predicts earnings surprises. An important distinction between StockTwits and
Seeking Alpha is that Seeking Alpha contributors are not anonymous and get compensated
for their posts.21 A more closely related paper is Kang, Lou, Ozik, Sadka, and Shen (2024),
which examines social media content from WallStreetBets around earnings announcements
from 2020 to 2021 and finds that increased social discussion reduced pre-earnings turnover,
20Pyun (2024) further demonstrate the importance of examining real-time (synchronous) group chats such
as Discord and compare them with forum-style (asynchronous) postings on Reddit’s WallStreetBets.
21Farrell, Green, Jame, and Markov (2022) find that the ability of retail order imbalances to predict stock
returns increases in the intraday following a Seeking Alpha publication. Ding, Zhou, and Li (2020) find that
Seeking Alpha coverage reduces individual stock return comovement with the market.
25
return drift, and higher earnings response coefficients.22 These findings depart from ours.
StockTwits activity ahead of earnings announcements is excessively optimistic, fails to pre-
dict earnings surprises, and induces attention-based buying pressure that increases stock
returns followed by a reversal on announcement dates.
The difference between our findings and other social media papers can result from the
differences in coverage across platforms. Table IA5 highlights the significant discrepancy
in coverage ahead of earnings announcements for StockTwits, WallStreetBets, and Seeking
Alpha from 2018 to 2021. The table reports the number of stock-earnings observations with
at least one post and the number of posts five days ahead of earnings announcements by
NYSE quintile breakpoints. StockTwits has the broadest coverage, with more than 38,000
stock-earnings announcement observations compared to 5,543 and 2,515 observations for
WallStreetBets and Seeking Alpha, respectively. The number of posts exceeds 5.5 million
for StockTwits, close to 33,000 for WallStreetBets, and 3,500 for Seeking Alpha. Small and
large firms are widely discussed on StockTwits, whereas more than 50% of the posts are
about the largest firms (top quintile) on the other platforms. Stock prices of small firms are
more sensitive to price impact. Consequently, a social media platform that widely covers
small stocks will play a more determinant role in understanding the impact of social media
on aggregate price dynamics of small firms.
6.2. Theoretical implications
Our results raise the question: Why would investors trade on optimistically-biased informa-
tion? We present a rational-based model based on a special case of Caplin and Leahy (2019)
to demonstrate why investors might consume optimistically biased information and trade on
such information in a systematic way.23 The model is motivated by our empirical findings
22Bradley, Hanousek Jr, Jame, and Xiao (2023) find that recommendations shared on WallStreetBets are
significant predictors of returns and cash-flow news.
23The importance of wishful thinking in financial markets is further highlighted in Cassella, Dim, and
Karimli (2023). The authors find that investors who are optimistic about a stock’s prospect react to negative
26
and has implications to future theoretical work to better understand social media-driven
systematic noise trading.
Consider a wishful-thinking investor who is considering buying q > 0 shares of an asset
with price pbefore the release of the company’s earnings announcement. For simplicity, we
will abstract how qand pare determined and take them as given. After the release of the
earnings announcement, the asset payoff ˜vcan take two values: a high value vH=p+v
after a positive surprise or a low value vL=p−vafter a negative surprise, where v > 0
and vH> vL. There is an objective probability for each value. With probability ¯πHthere
is a positive surprise and a high realization of the asset vH, and with probability ¯πLthere
is a negative surprise and a low realization of the asset vL. An alternative interpretation of
the objective probabilities is that these probabilities represent the consensus or mainstream
opinion in case there are agents with heterogeneous information.
The model assumes that wishful-thinking investors have subjective beliefs about the
probability realization of ˜v. We denote πHas the subjective probability of a positive surprise
vHand πLas the subjective probability of a negative surprise vL. These subjective beliefs
may differ from objective beliefs, but deviating from objective beliefs is costly. We represent
the cost of deviating from objective beliefs by the Kullback-Leibler distance:
1
θπHln πH
¯πH
+1
θπLln πL
¯πL
.
The parameter θrepresents the ease with which the agent can manipulate their beliefs.
The larger is θ, the greater the amount of evidence the agent would need before they reject
their chosen beliefs in favor of the objective ones. In other words, the larger θ, the more
likely the investor is to opt for subjective beliefs. The lower the θ, the more costly it is to
deviate from the objective beliefs.
The investor’s expected utility of holding the asset and manipulating beliefs is then given
news by shifting their optimistic expectations to a longer forecast horizon.
27
by:
EU (πH, πL) = q(πHvH+πLvL−p)−1
θπHln πH
¯πH−1
θπLln πL
¯πL
.(6)
The investor understands the preferences and that the beliefs differ from the objective be-
liefs. The wishful thinking investor will choose subjective beliefs πHand πLby maximizing
expected utility in (6), taking into account that πH+πL= 1. The optimization problem
leads the investor to choose the following subjective beliefs:24
πH=¯πHexp (θqvH)
¯πHexp (θqvH) + ¯πLexp (θqvL).(7)
The investor chooses to distort beliefs towards states with positive surprises vHso that
πH>¯πHfor ¯πH∈(0,1). The investor exhibits wishful thinking behavior by being over-
optimistic about the high utility states. In other words, the wishful-thinking investor obtains
utility from anticipating future events. At the extremes, when the objective probability
is either zero or one, subjective probabilities are equal to objective probabilities, and the
investor is rational. A wishful-thinking investor will not get any utility for dreaming about
impossible events. As the cost of manipulating beliefs decreases (θincreases), the beliefs
become even more distorted towards positive surprises. The same effect appears the more
shares qthe investor is considering to buy; as qincreases the subjective probability πH
deviates more from the objective probability ¯πHand thus more positive optimistic biased
are investors.
We can observe how a wishful-thinking investor distorts beliefs in a numerical example
in Figure 7. In this figure, we set the following parameters: vH= 3, vL= 1, θ=.5 and
q= 1,3,5. The solid line represents the beliefs of a wishful-thinking investor given by
(7). The dashed line represents the beliefs of a rational investor that uses the objective
beliefs πRational
H= ¯πH. The figure shows that the wishful-thinking investor distorts beliefs
towards positive surprises. Even when the probability of a positive surprise is less likely
24See Section IA for derivations.
28
than a negative surprise ¯πH<0.5, the wishful thinking investor may distort beliefs so that
πH>0.5. In words, even when the consensus is that there will be a negative surprise,
the wishful-thinking investor may think that a positive surprise is more likely (for example,
when ¯πH= 0.4, then πH>0.5). As the consensus probabilities get closer to the extremes,
when events are almost certain, then wishful-thinking investors resemble rational investors.
Figure 7 shows how beliefs get distorted as the number of shares qincreases. As the stakes
increase, there is an increase in the distortion of beliefs.
The wishful thinking investor will choose to purchase qunits of the asset at price p
when the expected utility in equation (6) with subjective beliefs given by (7) is positive
EU (πH, πL)≥0, which happens when:
¯πH≥exp (θqp)−exp (θqvL)
exp (θqvH)−exp (θqvL)=1
1 + exp (θqv)= ¯πcutof f
H.
Thus, a wishful thinking investor will choose to purchase the qshares of an asset at price p
when ¯πH≥¯πcutoff
H. Instead a rational investor with πRational
H= ¯πHwould choose to purchase
the qshares of an asset at price pwhen ¯πH≥0.5. We can see that a wishful-thinking
investor would make the same choices as a rational investor only when it is infinitely costly
to distort beliefs (θ= 0). For any θ > 0, the wishful thinking investor will have a lower
cutoff to purchase the asset than a rational investor such that ¯πcutoff
H<0.5.
We believe that more theoretical work on wishful thinking could shed some light on the
role of social media in financial markets. In Banerjee, Davis, and Gondhi (2024), wishful
thinking leads to endogenous disagreement. Their findings show a connection between wish-
ful thinking and market outcomes, including return volatility, price informativeness, trading
volume, and return predictability, which match empirical evidence presented in this paper.
29
7. Conclusion
Social media activity during the week leading up to announcements is overly optimistic
and does not forecast the earnings fundamentals. This optimism leads to systematic noise
trading, resulting in a buying pressure that increases stock returns by more than 50 bps
ahead of earnings announcements as market makers seek higher returns to provide liquidity.
Noise trading, when correlated, can increase inventory risk for market makers when providing
liquidity before anticipated information events.
Our findings show that social media activity ahead of earnings announcements can distort
price revelation because of the higher required rate of return by market makers adds noise to
prices. Conditioning on announcements’ earnings surprises, the average price run-up pushes
prices closer to fundamentals for positive earnings surprise announcements, albeit not a
reflection of informative trading, and away from fundamentals ahead of announcements with
negative earnings surprises. On the announcement date, markets efficiently adjust prices
quickly to new information and correct any mispricing generated by social media information
production. We further show that social media-driven noise trading is predictable.
A key implication of our paper to future research is that theories should consider the
role of systematic noise trading in price formation. Moreover, analyzing the effect of social
media on post-announcement price formation without considering pre-announcement price
dynamics can lead to biased inferences on the role of social media in price discovery. Fi-
nally, our paper does not claim that there is no useful information about upcoming earnings
on StockTwits, but that, on average, the information content is uninformative about earn-
ings fundamentals. Differentiating between “skilled” and “unskilled” social media content
creators, or “finfluencer,” is a fruitful avenue for future research. Kakhbod, Kazempour,
Livdan, and Schuerhoff (2023) have already made admirable progress in that regard.
30
References
Akey, Pat, Vincent Gr´egoire, and Charles Martineau, 2022, Price revelation from insider trading: Evidence
from hacked earnings news, Journal of Financial Economics 143, 1162–1184.
Antweiler, Werner, and Murray Z. Frank, 2004, Is all that talk just noise? The information content of
internet stock message boards, Journal of Finance 59, 1259–1294.
Banerjee, Snehal, Jesse Davis, and Naveen Gondhi, 2024, Choosing to disagree: Endogenous dismissiveness
and overconfidence in financial markets, Journal of Finance 79, 1635–1695.
Barber, Brad M., Xing Huang, Philippe Jorion, Terrance Odean, and Christopher Schwarz, 2024, A (sub)
penny for your thoughts: Tracking retail investor activity in TAQ, Journal of Finance 79, 2403–2427.
Barber, Brad M., Xing Huang, Terrance Odean, and Christopher Schwarz, 2022, Attention-induced trading
and returns: Evidence from robinhood users, Journal of Finance 77, 3141–3190.
Barber, Brad M., Yi-Tsung Lee, Yu-Jane Liu, and Terrance Odean, 2009, Just how much do individual
investors lose by trading?, Review of Financial Studies 22, 609–632.
Barber, Brad M., Shengle Lin, and Terrance Odean, 2023, Resolving a paradox: Retail trades positively
predict returns but are not profitable, Journal of Financial and Quantitative Analysis pp. 1–35.
Barber, Brad M., and Terrance Odean, 2000, Trading is hazardous to your wealth: The common stock
investment performance of individual investors, Journal of Finance 55, 773–806.
, 2008, All that glitters: The effect of attention and news on the buying behavior of individual and
institutional investors, Review of Financial Studies 21, 785–818.
, and Ning Zhu, 2008, Do retail trades move markets?, Review of Financial Studies 22, 151–186.
, 2009, Systematic noise, Journal of Financial Markets 12, 547–569.
Barrot, Jean-Noel, Ron Kaniel, and David Sraer, 2016, Are retail traders compensated for providing liquid-
ity?, Journal of Financial Economics 120, 146–168.
Barth, Mary E., and Eric C. So, 2014, Non-diversifiable volatility risk and risk premiums at earnings an-
nouncements, The Accounting Review 89, 1579–1607.
Bartov, Eli, Lucile Faurel, and Partha S. Mohanram, 2018, Can twitter help predict firm-level earnings and
stock returns?, The Accounting Review 93, 25–57.
Biais, Bruno, Pierre Hillion, and Chester Spatt, 1999, Price discovery and learning during the preopening
period in the paris bourse, Journal of Political Economy 107, 1218–1248.
Bianchi, Francesco, Roberto G´omez-Cram, and Howard Kung, 2024, Using social media to identify the effects
of congressional viewpoints on asset prices, Review of Financial Studies 37, 2244–2272.
Boehmer, Ekkehart, Charles M Jones, Xiaoyan Zhang, and Xinran Zhang, 2021, Tracking retail investor
activity, Journal of Finance 76, 2249–2305.
31
Boguth, Oliver, Adlai J. Fisher, Vincent Gregoire, and Charles Martineau, 2024, Noisy FOMC returns?
Information, price pressure, and post-announcement reversals, Working paper.
Boguth, Oliver, Vincent Gr´egoire, and Charles Martineau, 2019, Shaping expectations and coordinating
attention: The unintended consequences of fomc press conferences, Journal of Financial and Quantitative
Analysis 54, 2327–2353.
Bradley, Daniel, Jan Hanousek Jr, Russell Jame, and Zicheng Xiao, 2023, Place your bets? The market
consequences of investment research on reddit’s wallstreetbets, Working paper.
Brunnermeier, Markus K., 2005, Information leakage and market efficiency, Review of Financial Studies 18,
417–457.
Campbell, Brett, Michael Drake, Jacob Thornock, and Brady Twedt, 2023, Earnings virality, Journal of
Accounting and Economics 75, 101517.
Caplin, Andrew, and John V. Leahy, 2019, Wishful thinking, NBER Working Paper.
Cassella, Stefano, Chukwuma Dim, and Tural Karimli, 2023, Optimism shifting, Working paper.
Chen, Hailiang, Prabuddha De, Yu Jeffrey Hu, and Byoung-Hyoun Hwang, 2014, Wisdom of crowds: The
value of stock opinions transmitted through social media, Review of Financial Studies 27, 1367–1403.
Comerton-Forde, Carole, Terrence Hendershott, Charles M. Jones, Pamela C. Moulton, and Mark S.
Seasholes, 2010, Time variation in liquidity: The role of market-maker inventories and revenues, Journal
of Finance 65, 295–331.
Cookson, J. Anthony, Joseph E. Engelberg, and William Mullins, 2020, Does partisanship shape investor
beliefs? Evidence from the COVID-19 pandemic, Review of Asset Pricing Studies 10, 863–893.
, 2023, Echo chambers, Review of Financial Studies 36, 450–500.
Cookson, J. Anthony, Vyacheslav Fos, and Marina Niessner, 2021, Does disagreement facilitate informed
trading? evidence from activist investors, Working paper p. 45.
Cookson, J. Anthony, Runjing Lu, William Mullins, and Marina Niessner, 2022, The social signal, Journal
of Financial Economics, forthcoming.
Cookson, J. Anthony, William Mullins, and Marina Niessner, 2024, Social media and finance, Working paper.
Cookson, J. Anthony, and Marina Niessner, 2020, Why don’t we agree? Evidence from a social network of
investors, Journal of Finance 75, 173–228.
, and Christoph Schiller, 2022, Can social media inform corporate decisions? Evidence from merger
withdrawals, Working paper.
Cowen, Amanda, Boris Groysberg, and Paul Healy, 2006, Which types of analyst firms are more optimistic?,
Journal of Accounting and Economics 41, 119–146.
De Long, J. Bradford, Andrei Shleifer, Lawrence H Summers, and Robert J. Waldmann, 1990, Noise trader
risk in financial markets, Journal of Political Economy 98, 703–738.
32
Dim, Chukwuma, 2020, Should retail investors listen to social media analysts? Evidence from text-implied
beliefs, Working paper.
Ding, Rong, Yukun Shi, and Hang Zhou, 2023, Social media coverage and post-earnings announcement drift:
Evidence from Seeking Alpha, The European Journal of Finance 29, 207–227.
Ding, Rong, Hang Zhou, and Yifan Li, 2020, Social media, financial reporting opacity, and return comove-
ment: Evidence from Seeking Alpha, Journal of Financial Markets 50, 100511.
Fama, Eugene F., 1965, The behavior of stock-market prices, Journal of Business 38, 34–105.
Farrell, Michael, T. Clifton Green, Russell Jame, and Stanimir Markov, 2022, The democratization of in-
vestment research and the informativeness of retail investor trading, Journal of Financial Economics 145,
616–641.
Fisher, Adlai, Charles Martineau, and Jinfei Sheng, 2022, Macroeconomic attention and announcement risk
premia, The Review of Financial Studies 35, 5057–5093.
Friedman, Milton, et al., 1953, The case for flexible exchange rates, Essays in positive economics 157, 33.
Gregoire, Vincent, and Charles Martineau, 2022, How is earnings news transmitted to stock prices?, Journal
of Accounting Research 60, 261–297.
Grossman, Sanford J., and Joseph E. Stiglitz, 1980, On the impossibility of informationally efficient markets,
American Economic Review 70, 393–408.
Han, Bing, and Liyan Yang, 2013, Social networks, information acquisition, and asset prices, Management
Science 59, 1444–1457.
Heimer, Rawley Z, 2016, Peer pressure: Social interaction and the disposition effect, Review of Financial
Studies 29, 3177–3209.
Hirshleifer, David, Lin Peng, and Qiguang Wang, 2024, News diffusion in social networks and stock market
reactions, Review of Financial Studies, forthcoming.
Hu, Danqi, Charles M. Jones, Valerie Zhang, and Xiaoyan Zhang, 2021, The rise of Reddit: How social
media affects retail investors and short-sellers’ roles in price discovery, Working paper.
Hvidkjaer, Soeren, 2008, Small trades and the cross-section of stock returns, Review of Financial Studies 21,
1123–1151.
Jia, Weishi, Giulia Redigolo, Susan Shu, and Jingran Zhao, 2020, Can social media distort price discovery?
Evidence from merger rumors, Journal of Accounting and Economics 70, 101334.
Jiao, Peiran, Andre Veiga, and Ansgar Walther, 2020, Social media, news media and the stock market,
Journal of Economic Behavior & Organization 176, 63–90.
Kakhbod, Ali, Seyed Mohammad Kazempour, Dmitry Livdan, and Norman Schuerhoff, 2023, Finfluencers,
Working paper.
33
Kang, Namho, Xiaoxia Lou, Gideon Ozik, Ronnie Sadka, and Siyi Shen, 2024, Innocuous noise? Social
media and asset prices, Working paper.
Kaniel, Ron, Shuming Liu, Gideon Saar, and Sheridan Titman, 2012, Individual investor trading and return
patterns around earnings announcements, Journal of Finance 67, 639–680.
Kaniel, Ron, Gideon Saar, and Sheridan Titman, 2008, Individual investor trading and stock returns, Journal
of Finance 63, 273–310.
Kelley, Eric K., and Paul C. Tetlock, 2013, How wise are crowds? Insights from retail orders and stock
returns, Journal of Finance 68, 1229–1265.
Kothari, Sagar P., Eric So, and Rodrigo Verdi, 2016, Analysts’ forecasts and asset pricing: A survey, Annual
Review of Financial Economics 8, 197–219.
Kyle, Albert S., 1985, Continuous auctions and insider trading, Econometrica pp. 1315–1335.
Laarits, Toomas, and Marco Sammon, 2023, The retail habitat, Working paper.
Lee, Charles MC, Belinda Mucklow, and Mark J. Ready, 1993, Spreads, depths, and the impact of earnings
information: An intraday analysis, Review of Financial Studies 6, 345–374.
Li, Edward Xuejun, K. Ramesh, Min Shen, and Joanna Shuang Wu, 2015, Do analyst stock recommendations
piggyback on recent corporate news? An analysis of regular-hour and after-hours revisions, Journal of
Accounting Research 53, 821–861.
Liu, Bibo, Huijun Wang, Jianfeng Yu, and Shen Zhao, 2020, Time-varying demand for lottery: Speculation
ahead of earnings announcements, Journal of Financial Economics 138, 789–817.
Madhavan, Ananth, and Seymour Smidt, 1993, An analysis of changes in specialist inventories and quota-
tions, Journal of Finance 48, 1595–1628.
Martineau, Charles, 2022, Rest in peace post-earnings announcement drift, Critical Finance Review 11,
613–646.
, and Marius Zoican, 2023, Retail trading and analyst coverage, Journal of Financial Markets 66,
100849.
Ozik, Gideon, Ronnie Sadka, and Siyi Shen, 2021, Flattening the illiquidity curve: Retail trading during the
covid-19 lockdown, Journal of Financial and Quantitative Analysis 56, 2356–2388.
Pedersen, Lasse Heje, 2022, Game on: Social networks and markets, Journal of Financial Economics 146,
1097–1119.
Peng, Lin, Qiguang Wang, and Dexin Zhou, 2022, Social networks, trading, and liquidity, Journal of Portfolio
Management 48, 196–215.
Peress, Joel, and Daniel Schmidt, 2020, Glued to the tv: Distracted noise traders and stock market liquidity,
Journal of Finance 75, 1083–1133.
34
, 2021, Noise traders incarnate: Describing a realistic noise trading process, Journal of Financial
Markets 54, 100618.
Pyun, Chaehyun, 2024, Synchronous social media and the stock market, Journal of Financial Markets p.
100915.
So, Eric C., and Sean Wang, 2014, News-driven return reversals: Liquidity provision ahead of earnings
announcements, Journal of Financial Economics 114, 20–35.
Van Binsbergen, Jules H., Xiao Han, and Alejandro Lopez-Lira, 2023, Man versus machine learning: The
term structure of earnings expectations and conditional biases, Review of Financial Studies 36, 2361–2396.
Waters, Brian, Xingtan Zhang, Wei Zhou, and Shrihari Santosh, 2024, Exploiting non-fundamental flows,
Working paper.
Weller, Brian M., 2018, Does algorithmic trading reduce information acquisition?, Review of Financial
Studies 31, 2184–2226.
Figure 1. StockTwits Activity Over Time
This figure shows the monthly number of stock-specific posts on StockTwits. The sample
period is from January 1, 2013, to December 31, 2022.
2013-Mar
2013-Jul
2013-Nov
2014-Mar
2014-Jul
2014-Nov
2015-Mar
2015-Jul
2015-Nov
2016-Mar
2016-Jul
2016-Nov
2017-Mar
2017-Jul
2017-Nov
2018-Mar
2018-Jul
2018-Nov
2019-Mar
2019-Jul
2019-Nov
2020-Mar
2020-Jul
2020-Nov
2021-Mar
2021-Jul
2021-Nov
2022-Mar
2022-Jul
2022-Nov
0.0M
1.0M
2.0M
3.0M
4.0M
5.0M
Number of posts per month
Figure 2. Abnormal Attention Around Earnings Announcements
This figure shows a boxplot representing the distribution of the number of abnormal Stock-
Twits posts and newswires coverage five days around earnings announcements. The whiskers
correspond to the 5th and 95th percentiles. Abnormal attention (newswire coverage) is com-
puted as the daily log number of StockTwits posts (newswire articles) minus the log of the
average daily number of posts (newswire articles) from 20 to 6 days before the earnings
announcements. The sample period is from January 1, 2013, to December 31, 2022.
-5 -4 -3 -2 -1 0 1 2 3 4 5
Days since earnings announcement
−1
0
1
2
3
Abnormal attention
StockTwits
Newswire
Figure 3. StockTwits Sentiment is Optimistic
This figure shows the fraction of stock-earnings observations by (1) the fraction of bullish
StockTwits posts and (2) the fraction of bullish (positive) analyst recommendations sixty
days before earnings announcements. The sample period is from January 1, 2013, to De-
cember 31, 2022.
[0 −20] (20 −40] (40 −60] (60 −80] (80 −100]
% of bullish views
0
10
20
30
40
50
60
Fraction of observations (%)
StockTwits posts
Analyst recommendations
Figure 4. Return Reversals After Earnings Announcements
This figure plots the average difference in buy-and-hold abnormal return (BHAR, in %)
between high-StockTwits attention stocks and matched low-StockTwits attention stocks
around earnings announcements (t= 0). High-attention stocks correspond to the top quin-
tile stock-earnings announcements with the highest coverage on StockTwits five days before
announcements. Matched stocks are assigned based on the same industry, the BHAR 30
to six days before announcements, firm size, and earnings surprises. The 95% confidence
intervals are represented by the error bars. The sample period is from January 1, 2013, to
December 31, 2022.
−20 −15 −10 −5 0 5 10
Days since announcement
−0.25
0.00
0.25
0.50
0.75
1.00
1.25
1.50
BHAR difference between high and low attention-matched stocks (%)
Figure 5. Cumulative Returns Around Earnings Announcements Conditioning on Stock-
Twits Attention and Earnings Surprises
This figure shows the buy-and-hold abnormal returns (BHAR, in %) around earnings an-
nouncements for stocks with high-StockTwits attention and matched-low StockTwits at-
tention 20 days before to 10 days after earnings announcements. High attention stocks
correspond to the top quintile stock-earnings announcements with the highest coverage on
StockTwits five days before announcements. Matched stocks are assigned based on the same
industry, the BHAR 20 to six days before announcement, firm size, and on earnings surprises.
Panels A and B show the cumulative returns around earnings announcements with positive
(top two quintiles) and negative earnings surprises (bottom two quintiles), respectively. The
shaded area corresponds to the 95% confidence intervals. The plots are rescaled to zero at
t=−6. The sample period is from January 1, 2013, to December 31, 2022.
Panel A. Positive earnings surprises
−20 −18 −16 −14 −12 −10 −8−6−4−2 0 2 4 6 8 10
Days since announcement
−1
0
1
2
3
Cumulative abnormal return
High attention
Low attention
Panel B. Negative earnings surprises
−20 −18 −16 −14 −12 −10 −8−6−4−2 0 2 4 6 8 10
Days since announcement
−3
−2
−1
0
1
Cumulative abnormal return
High attention
Low attention
Figure 6. Earnings Whispers Social Media Posts
This figure shows two examples of Earnings Whispers StockTwits posts about upcoming
earnings announcements for the week of August 29, 2022 and September 26, 2022.
Panel A. August 29, 2022
Panel B. September 26, 2022
Figure 7. Plots of πHfor Wishful Thinking Versus Rational Investors
This figure presents the relationship between πHand ¯πHfor a rational and wishful thinking
agent. Dashed black line represents a rational agent. Solid lines represent wishful-thinking
investors for different quantities q. We set vH= 3, vL= 1, θ = 0.5.
0.0 0.2 0.4 0.6 0.8 1.0
¯
πH
0.0
0.2
0.4
0.6
0.8
1.0
πH
Rational Agent
Wishful, q = 1
Wishful, q = 3
Wishful, q = 5
Table 1
Summary Statistics on High and Low StockTwits Coverage
This table reports summary statistics of the stock-earnings announcement sample for high
and low StockTwits attention stocks. High-attention stocks correspond to the top quintile
stock-earnings announcements with the highest coverage on StockTwits five days before the
announcement. Volatility is the standard deviation of 30 daily returns ending ten days before
announcements. Abn ret. and Abn |ret.|corresponds to the abnormal and absolute abnormal
returns on the earnings announcement date, respectively. The sample period is from January
2013 to December 2022.
High attention Low attention
Mean 25th Median 75th Mean 25th Median 75th
Market cap (mill.$) 35,179 849 5,482 28,335 4,720 351 1,204 3,758
Volatility (%) 3.23 1.50 2.36 3.87 2.57 1.41 1.99 3.05
Abn ret. (%) -0.43 -4.66 -0.35 3.77 -0.03 -3.59 0.04 3.61
Abn |ret.|(%) 6.41 1.84 4.23 8.47 5.57 1.49 3.60 7.35
Surprise (%) -1.90 -0.05 0.06 0.26 -0.23 -0.09 0.06 0.29
N. analysts 10.16 4.00 9.00 15.00 5.33 2.00 4.00 7.00
Table 2
Summary Statistics on Coverage
This table reports summary statistics on StockTwits coverage, analyst recommendations,
and newswire coverage in Panels A to C, respectively. N. posts, N. rec., and N. news
correspond to the total number of StockTwits posts, analyst recommendations, and Dow
Jones newswires, respectively, five to one day before earnings announcements. The sample
period is from January 2013 to December 2022.
A. StockTwits coverage
StockTwits coverage No StockTwits coverage
NYSE qnt Stock-EA obs. N. posts % of posts Stock-EA obs. % with no coverage
1 (small) 31,145 2,213,564 24 4,820 13
2 19,662 1,204,046 13 1,218 6
3 15,406 1,132,371 12 678 4
4 13,953 1,200,506 13 479 3
5 (large) 13,780 3,320,317 37 315 2
Total 93,946 9,070,804 100 7,510
B. Analyst recommendation
Analyst recommendation No analyst recommendation
NYSE qnt Stock-EA obs. N. rec. % of rec. Stock-EA obs. % with no rec.
1 (small) 779 1,350 5 35,186 98
2 999 2,326 8 19,881 95
3 1,231 2,031 7 14,853 92
4 1,592 3,497 13 12,840 89
5 (large) 3,129 18,185 66 10,966 78
Total 7,730 27,389 100 93,726
C. Newswire
News No news
NYSE qnt Stock-EA obs. N. news % of news Stock-EA obs. % with no news
1 (small) 6,837 24,520 4 29,128 81
2 5,372 28,763 5 15,508 74
3 5,527 37,890 6 10,557 66
4 6,890 75,524 12 7,542 52
5 (large) 11,120 461,581 73 2,975 21
Total 35,746 628,278 100 65,710
Table 3
StockTwits’ Sentiment Does Not Predict Fundamentals
This table reports estimates for the full sample, large caps (the top three NYSE market
capitalization quintiles), and small caps (bottom two quintiles) of the following regression:
Surpi,t =β1Senti,t +β2
1
Att
i,t ×Senti,t +β3
1
Att
i,t + Γ′Controlsi,t +αi+αt+εi,t ,
where Surpi,t is the earnings surprise (%) for stock ion earnings announcement t.Sent
is the sentiment on StockTwits as defined in equation (2) over five days before earnings
announcements, and
1
Att
i,t is a dummy variable equal to one if the stock attention on Stock-
Twits belongs to the top quintile. Controlsi,t is a vector of control variables corresponding
to the buy-and-hold abnormal returns (BHAR[-5,-1]), sentiment from analyst recommen-
dations (Analysts sent), and RavenPack newswire sentiment (News sent), all measured in
the five days leading up to the earnings announcements. αiand αtare stock and time-fixed
effects, respectively. Columns(1)–(3) exclude observations for which posts have no sentiment
tags. Columns (4)–(6) include non-tagged sentiment posts