Trading on Twitter: Using Social Media Sentiment to Predict Stock Returns: Trading on Twitter

Abstract

Decision making is often based on the rational assessment of information, but recent research shows that emotional sentiment also plays an important role, especially for investment decision making. Emotional sentiment about a firm's stock that spreads rapidly through social media is more likely to be incorporated quickly into stock prices (e.g., on the same trading day it was expressed), while sentiment that spreads slowly takes longer to be incorporated into stock prices and thus is more likely to predict stock prices on future days. We analyzed the cumulative sentiment (positive and negative) in 2.5 million Twitter postings about individual S&P 500 firms and compared this to the stock returns of those firms. Our results show that the sentiment in tweets about a specific firm from users with less than 171 followers (the median in our sample) had a significant impact on the stock's returns on the next trading day, the next 10 days, and the next 20 days. Interestingly, sentiment in tweets from users with fewer than 171 followers that were not retweeted had the greatest impact on future stock returns. A trading strategy based on these findings produced meaningful economic gains on the order of an 11–15% annual return.

Full text

Decision Sciences
Volume 00 Number 0
xxxx 2016
© 2016 Decision Sciences Institute
Trading on Twitter: Using Social Media
Sentiment to Predict Stock Returns
Hong Kee Sul
Wharton Research Data Service, The Wharton School, University of Pennsylvania
Philadelphia, PA 19104, e-mail: hongkee@wharton.upenn.edu
Alan R. Dennis
Operations and Decision Technologies Department, Kelley School of Business
Indiana University, Bloomington, IN 47405, e-mail: ardennis@indiana.edu
Lingyao (Ivy) Yuan†
Department of Supply Chain and Information Systems, College of Business, Iowa State
University, Ames, IA 50011, e-mail: lyuan@iastate.edu
ABSTRACT
Decision making is often based on the rational assessment of information, but recent
research shows that emotional sentiment also plays an important role, especially for
investment decision making. Emotional sentiment about a firm’s stock that spreads
rapidly through social media is more likely to be incorporated quickly into stock prices
(e.g., on the same trading day it was expressed), while sentiment that spreads slowly
takes longer to be incorporated into stock prices and thus is more likely to predict stock
prices on future days. We analyzed the cumulative sentiment (positive and negative)
in 2.5 million Twitter postings about individual S&P 500 firms and compared this to
the stock returns of those firms. Our results show that the sentiment in tweets about a
specific firm from users with less than 171 followers (the median in our sample) had
a significant impact on the stock’s returns on the next trading day, the next 10 days,
and the next 20 days. Interestingly, sentiment in tweets from users with fewer than 171
followers that were not retweeted had the greatest impact on future stock returns. A
trading strategy based on these findings produced meaningful economic gains on the
order of an 11–15% annual return. [Submitted: December 4, 2014. Revised: April 1,
2016. Accepted: April 12, 2016.]
Subject Areas: Twitter, Emotion, Sentiment, Stock returns, and S&P500.
INTRODUCTION
Almost 75% of adult Internet users use social media, and this percentage is in-
creasing (Pew-Research, 2014). Twitter is one of the most popular social media
†Corresponding author.
1

2Trading on Twitter
platforms in the world. Not only has the number of people using social media in-
creased dramatically, so too has the amount of use. In 2015, there were about 300
million Twitter users worldwide, who sent an average of 500 million tweets per
day (“About Twitter, Inc.,” 2015). Users have integrated social media into many
aspects of their daily life (Ellison, 2007), including investment decision making
(Oh & Sheng, 2011). Numerous professional and amateur investors and analysts
use Twitter to post news articles, and opinions, often providing information and
comments more frequently than the professional news media (Sprenger, Tumasjan,
Sandner, & Welpe, 2014).
Stock returns, or the profits from trading stocks, are influenced by many
factors. Along with fundamental factors and transaction costs, investor sentiment
also plays an important role in influencing stock return (Baker & Wurgler, 2007).
Market sentiment can be expressed in many ways. The development of social
media provides a new meaningful channel for users to share information and their
personal feelings. As such, it also serves as a convenient method to capture market
sentiment.
Prior research has studied whether the emotional content of tweets can be
used to predict stock returns. Bollen, Mao, and Zeng (2011b) assessed the emo-
tional state (calm, alert, sure, vital, kind, and happy) in 10 million tweets that were
not related to the stock market. They found that the amount of one state, “calm,”
was significantly positively correlated with changes in the Dow Jones Industrial
average (DJIA) several days later; in other words, when there was a great deal of
“calm” in tweets on a given day, the DJIA tended to rise over the following days.
Oh and Sheng (2011) examined 200,000 tweets from StockTwits that focused on
specific stocks and classified each tweet as “bullish,” “bearish,” or “neutral” to
create a “bullishness” index for each stock. They found the 5-day rolling average
of the bullishness index was useful in predicting stock price movements. Sprenger
et al. (2014) also used machine learning to create a different bullishness index that
they too found to be predictive of stock returns several days later. Smailovi´
c, Grˇ
car,
Lavraˇ
c, and ˇ
Znidarˇ
siˇ
c (2014) used machine learning to examine sentiment (i.e.,
positive emotion) in tweets and found it to be predictive of stock returns several
days later. Risius, Akolk, and Beck (2015) examined emotional states (happiness,
affection, satisfaction, fear, anger, depression, contempt) and positive and negative
sentiment, and found negative sentiment and “depression” to predict stock returns
on the following day.
These findings are promising in suggesting that the emotional state and sen-
timent in tweets can be used to predict stock returns, but there are still many
unanswered questions. Although empirical research has shown that certain emo-
tional states and sentiments in tweets can predict stock price movements, there is
a lack of theory to explain why they influence stock returns days later. We argue
that the Gradual Information Diffusion model (Hong & Stein, 1999) is useful in
understanding how tweets are linked to future stock returns. Under this theoreti-
cal perspective, information (in our case, sentiment) influences stock prices as it
spreads through the investing public. Sentiment that spreads quickly has an imme-
diate influence on prices, while sentiment that spreads slowly has a slower effect.
Sentiment that spreads slowly opens the door for a trading strategy that capitalizes
on the stock returns from slowly rising or falling prices.

Sul, Dennis, and Yuan 3
We analyzed almost 2 years’ worth of data collected from Twitter and linked
it to the average daily stock returns of firms in the S&P 500. Our results show
that the sentiment in tweets about specific firms was significantly related to stock
returns on subsequent days. Tweets from individuals with fewer followers had
a stronger impact on future returns than tweets from those with many followers,
because their tweets took longer to spread. Likewise, tweets that were not retweeted
took longer to spread and were linked to greater future stock returns.
PRIOR RESEARCH AND THEORY
Information and Stock Return Prediction
Whether stock returns can be predicted has long been a debate. Based on the
Efficient Market Hypothesis (EMH), early research argued that stock returns are
random and cannot be predicted (Eppen & Fama, 1969; Dockery & Kavussanos,
1996). Research shows that new information, especially news, is a major factor in-
fluencing stock returns and quickly leads to stock price changes (Malkiel & Fama,
1970; Hong, Lim, & Stein, 2000; Qian & Rasheed, 2007). Under EMH, positive or
negative news (e.g., merger, terrorist attack) is quickly factored into a stock price
within minutes, so there is little opportunity to profit, unless of course, one has
insider knowledge of an event before it occurs. Mass media outlets play an impor-
tant role in disseminating information to a broad audience, especially individual
investors (Fang & Peress, 2009). This suggests that information contained in social
media such as Twitter, which also reaches a broad audience, may be linked stock
returns (Bollen et al., 2011a; Oh & Sheng, 2011; Smailovi´
c et al., 2014; Risius
et al., 2015; Sprenger et al., 2014).
EMH assumes that information travels quickly and that investors are rational
and capable of understanding the full implications of the information they receive
(Hong & Stein, 1999). An alternative view is the Gradual Information Flow (GIF)
model of Hong and Stein (1999), although the term GIF did not emerge until
later (Hong & Stein, 2007). GIF argues that the modern world differs in two
important ways from that assumed by EMH. First, some information is private,
known only to some investors, and this information diffuses more slowly than
public information. Second, investors have cognitive limitations and biases that
limit their ability to fully process all implications of the information they receive. In
general, investors are either news followers who use fundamental firm information
to make investment decisions or momentum traders who use past changes in
stock prices to make investment decisions. Both act under bounded rationality,
and because they focus primarily on the information relevant to their investing
style, they overlook other types of information they receive, and thus prices do not
respond to new information as quickly as EMH would predict.
GIF predicts that the speed of information diffusion through the investing
public influences how quickly stock prices change in response to new information.
Under reasonably efficient markets, information diffuses rapidly among the in-
vesting public and is quickly incorporated into stock prices (Hong & Stein, 1999).
Conversely, if information diffuses more slowly, it will take longer for that informa-
tion to be fully incorporated into their prices, and thus there may be opportunities

4Trading on Twitter
to profit from information before it is fully incorporated into prices (Hong et al.,
2000). Under GIF, information should spread rapidly for stocks covered by the
mass media but more slowly for stocks not covered by the media. Research shows
that stocks not covered by the mass media earn significantly higher future returns
than stocks that are covered, after controlling for risk characteristics (Merton,
1987; Fang & Peress, 2009), suggesting that the speed of information diffusion is
important in understanding how information may be used to predict future stock
prices and thus the returns that can be made by investing.
Information Diffusion in Social Media
Twitter is a social media platform in which users post short text messages of up to
140 characters, called tweets. Anyone can open a Twitter account and begin sending
tweets. Users can subscribe to or “follow” other users, and the followers are notified
immediately when a user tweets. Many Twitter users have few followers, while
commentators, journalists, and celebrities have thousands or more. The median
number of followers has gradually increased over time and was about 100 in 2014
(Liu, Kilman-Silver, & Mislove, 2014). Most users follow more people than they
have followers; the median number of users followed has gradually increased over
time and was around 140 in 2014 (Liu et al., 2014).
During the past several years, Twitter has drawn interests of researchers
from multiple disciplines. Current research on Twitter includes several streams.
One stream is its impact on information diffusion and supporting communica-
tion/collaboration (Honey & Herring, 2009) in many different contexts. Using
Twitter during a talk show decreased the psychological distance between the host
and his/her audience (Larsson, 2013). In the context of education, Twitter is a
potential learning tool in classrooms (Dhir, Buragga, & Boreqqah, 2013). Twitter
has become an important tool to spread information during natural disasters and
social crises (Sakaki, Okazaki, & Matsuo, 2010; Oh, Agrawal, & Rao, 2013).
Another research stream using Twitter is designing and developing network
analysis techniques and algorithms. The abundant data exchanged on Twitter every
minute provide researchers, especially those in computer science, the opportunity
to observe the social network change. Other related techniques, such as text mining
and data mining techniques, also became more refined by studying Twitter data.
A third stream is using Twitter to predict individual behavior. Using opinion
mining tools and sentiment analysis techniques, researchers are able to predict
election results (Tumasjan, Sprenger, Sandner, & Welpe, 2010), hospital-associated
mortality (Daley et al., 1988), and heart disease in middle-aged and older persons
(Gordon, Castelli, Hjortland, Kannel, & Dawber, 1977).
Due to its popularity, the investment community has adopted Twitter. This
community uses the convention of tagging stock-related tweets with a dollar sign
($) followed by the firm’s stock ticker symbol. For example, an individual tweeting
about PepsiCo would include $PEP in the tweet. A sample tweet from our data:
“$PEP has been strong all day. And who doesn’t love those Frito-Lay snacks? Be
honest."
Any Twitter user can send a tweet and include a stock ticker with a dollar
sign to indicate that he or she thinks the tweet contains financial information.

Sul, Dennis, and Yuan 5
Depending upon how many followers that user has, that information may reach a
few users, many users, or even tens of thousands of users. Other users can “retweet”
the information to their followers so that the information in the original tweet will
spread throughout a broad audience of Twitter users—and to non-Twitter users if
some users choose to spread the information using other media.
In his seminal work on networks, Barab´
asi (2002) shows just how inter-
connected we are. Twitter is a directed graph network in that connections are
directional. I receive information from users I follow, but they do not receive in-
formation from me unless they follow me back. The speed at which information
spreads through such a network depends upon how many followers a user has
(Barab´
asi, 2002). The general formula for the number of hops it takes to reach any
other node in a network is d=Log N/ log k, where Nis the total number of nodes
in the network, and kis the average number of connections per node (i.e., follow-
ers) (Barab´
asi, 2002). The number of active traders is on the order of 10 million
depending upon how one defines active (trade-IQ, 2011). It is difficult to estimate
the average number of followers in this community; links in networks typically
follow a power law distribution—not a normal distribution (Barab´
asi, 2002)—and
our data were no different, so we use the median of 171. Using these data, we
see that it takes about 3.1 hops for information from one node to reach any other
node. Of course, Twitter is not the only mechanism through which information is
spread. Individual investors can talk with or e-mail other investors. Most people
know 200–5,000 people by name (Barab´
asi, 2002), which suggests that we are
three to four hops away from anyone else on the planet (Barab´
asi, 2002).
The speed of information diffusion influences whether the information is
quickly incorporated in stock prices or takes longer—perhaps days—to be fully
disseminated and incorporated into stock prices (Hong & Stein, 1999). Empirical
studies of networks show that the number of connections is not randomly dis-
tributed (Barab´
asi, 2002). In every network, there are hubs, individuals who have
substantially more connections than the average (Barab´
asi, 2002). These hubs are
often opinion leaders who facilitate the rapid diffusion of information through the
network (Barab´
asi, 2002). Some professional analysts routinely tweet information
and thus have a large number of followers; Jim Cramer of CNBC’s Mad Money,
for example, has over 650,000 followers. These individuals are the hubs in the
diffusion of investment information, reaching a significant proportion of investors
in one hop.
If a Twitter user is a hub (i.e., has many followers), the information he
or she tweets will spread more quickly than if the same information is tweeted
by a user who has few followers (Barab´
asi, 2002). Tweets from a hub not only
reach more people in a single hop, but also tend to be more influential (Barab´
asi,
2002). If a Twitter user has many followers, any information he or she tweets will
be quickly disseminated, and stock prices should quickly change to incorporate
that information that same day, and there should be little or no effect on stock
returns on future days. For example, during the day on July 7, 2011, Twitter
account howardlindzon (which has about 200,000 followers) tweeted, “Looks like
Howard will be adding more $aapl on a good close. He’s predictable.” In this tweet
of 14 words, two were positive, making it a positive tweet on Apple. Apple had
closed the previous day at $351.76, and rose to close at $357.20 that day.

6Trading on Twitter
Conversely, if a Twitter user has few followers, information should be slower
to disseminate because it will take more hops to reach a critical mass of investors
and because it will be less influential than tweets from a hub (Barab´
asi, 2002).
Therefore, there is more likely to be a relationship between that information and
stock returns on future days because it will take longer for that information to
reach many investors and be incorporated into stock prices. For example, during
the day on August 8, 2011, several tweets came from multiple twitter accounts
with less than 100 followers. User ibshakey tweeted, “Still love $AAPL. Continue
to love gold. The US in general, not so much.” User cronked tweeted, “After that
last quarter, how can you not buy $AAPL here? There are some bright spots out
there. Not all is lost.,” and drewmethey tweeted, “$AAPL I have to buy Apple
here. It’s just too cheap!!!! Can’t resist.” Each tweet contained two positive words,
making all three a positive tweet on Apple, Inc. From the previous close price of
$373.62, the stock price fell to $353.21 at the end of the day. However, on the next
day and over the next 10 days, the stock price gradually rose, closing a month later,
on September 8, at $384.14.
Sentiment and Contagion
Much of the investment information shared using traditional media and social
media is facts and opinions, but individual behavior is not only the outcome of
rational decision making. Emotions triggered by these facts and opinions can
also influence decisions (Bechara & Damasio, 2005). Twitter provides a good
environment to foster the sharing of emotion (Bollen, Pepe, & Mao, 2011b) because
the length of each tweet is restricted to 140 characters. The limitation on length of
tweets encourages users to be brief and get to the point (Oh & Sheng, 2011). Thus,
a short message can provide a focused and more intense trigger for the receiver.
Individual moods, emotions, and other affects are influenced by both in-
ternal factors and external factors. Internal factors include personality, individual
competency, and so on. External factors include experiences, and information the
individual receives. Different affects have different impacts on individuals (Frijda,
1994). Affects can be broad and vague or acute and specific. Affects may have a
long-term influence; their effects can also be short term.
Emotion, as one type of affect, has the characteristics of having a clear trig-
ger and a short but more intense effect (Frijda, 1994). Emotion is a subjective
feeling related, triggered by a stimulus such as an event, an object, or information
in one’s environment. Once the stimulus conditions, the stimulus itself, or the sup-
porting cognition, perceptions, or other triggers are no longer active, the emotion
will disappear. Emotion can be highly contagious (Schoenewolf, 1990; Hatfield,
Cacioppo, & Rapson, 1993).
There are many ways to conceptualize the way emotion is expressed, but two
dominant approaches have emerged (Russell, 1980, 2003; Calvo & Kim, 2013).
The classic approach, used by Bollen et al. (2011a) and Risius et al. (2015) is
to consider specific emotional states, such as joy, anger, sadness, etc. The other
approach, used by Smailovi´
c et al. (2014) is the dimensional model in which
emotional affect is conceptualized as having two dimensions: valence (positive or
negative) and arousal (high or low) (Osgood, Suci, & Tannenbaum, 1957; Russell,

Sul, Dennis, and Yuan 7
1980, 2003; Cacioppo, Petty, Losch, & Kim, 1986); some authors also include
a third dimension of dominance (Bradley & Lang, 1994). Neither approach is
more or less “correct” (Calvo & Kim, 2013), and it is straightforward to map
emotional states onto the dimensional model (Russell, 1980, 2003; Bradley &
Lang, 1994). For example, the “calm” emotional state studied by Bollen et al.
would be considered neutral valence and low arousal (Russell, 2003).
Both models are commonly used, although Calvo and Kim (2013) conclude
that researchers in natural language processing are more likely to use the emotional
states model, while researchers in psychology are more likely to use the dimen-
sional model. When natural language processing researchers use the dimensional
model, they commonly focus on only the valence dimension, which they term
“sentiment” (e.g., Wiebe, Wilson, & Cardie, 2005; Abbasi & Chen, 2008; Oh &
Sheng, 2011; Smailovi´
c et al., 2014). In this study, we focus on sentiment, as has
been common in financial research (e.g., Tetlock, 2007).
Sentiment affects decision making (Bakamitsos, 2006). According to Con-
strual Level Theory (CLT), positive and negative sentiment may have different
effects (Liberman & Trope, 1998; Bar-Anan, Liberman, & Trope, 2006; Fujita,
Trope, Liberman, & Levin-Sagi, 2006). Positive sentiment increases abstract con-
strual, that is, the adoption of abstract, future goals, while negative sentiment
triggers a focus on immediate and proximal concerns and reduces the adoption
of abstract future goals (Liberman & Trope, 1998; Eyal, Liberman, Trope, &
Walther, 2004; Bar-Anan et al., 2006; Fujita et al., 2006; Labroo & Patrick, 2009).
Positive sentiment is more likely to induce individuals to make a decision than
negative sentiment, which tends to slow the decision process (Qiu & Yeung,
2008). Positive sentiment also may induce an individual to act on a decision
(Frijda, 1994). Positive sentiment can increase consumers’ impulse to buy in the
context of electronic commerce (Parboteeah, Valacich, & Wells, 2009) but increase
an individual’s resistance to temptation in other contexts (Fedorikhin & Patrick,
2010). Thus, positive sentiment and negative sentiment are more than the opposite
ends of the same dimension; they can trigger different behaviors. For this reason,
researchers often have measured them separately to capture their full effects (e.g.,
Tetlock, Saar-Tsechanksky, & Macskassy, 2008; Risius et al., 2015).
Sentiment is contagious (Schoenewolf, 1990; Hatfield et al., 1993). Social
contagion is “the tendency to automatically mimic and synchronize expressions,
vocalizations, postures, and movements with those of another person’s and to
converge emotionally” (Hatfield et al., 1993). Contagion happens implicitly and
explicitly (Singer & Lamm, 2009). Sentiment is expressed through facial im-
pressions, physical gestures, vocal tones, and written words. When individuals
exchange messages via written text, photos, audio, or even video on social media,
the message sender’s sentiment also is exchanged. Thus, tweets and the sentiment
they contain have the potential to influence the receiver’s behavior (Risius et al.,
2015).
Hypotheses
We argue that the sentiment contained in social media tweets will have a direct
effect on stock returns in a manner similar to the effects that professional news

8Trading on Twitter
media have on stock returns. Positive sentiment should be associated with positive
returns and negative sentiment should be associated with negative returns (Tetlock
et al., 2008).
The effects of sentiment spread in the same manner in which information
spreads through the network. Thus, the speed of diffusion is important. If a tweet
about a specific firm is sent by a hub (a user who has many followers), the sentiment
it contains will spread faster than the sentiment sent by a user who has few followers
because more individuals will see it immediately, and it will be more influential
(Barab´
asi, 2002). Sentiment that is spread more quickly will be incorporated into
prices faster, so that it will have an effect on returns sooner (Hong & Stein, 1999;
Tetlock et al., 2008). Its effects are more likely to be seen on the same trading day
on which it was tweeted. Thus, it will have little effect on stock returns on future
days, because its effects are immediately incorporated into prices (Hong & Stein,
1999; Tetlock et al., 2008).
In contrast, sentiment contained in tweets from those with fewer follow-
ers will take longer to disseminate because fewer people will see them on the
first hop, and it will be less influential (Barab´
asi, 2002). Thus, the number of fol-
lowers affects the speed of sentiment diffusion. This less visible sentiment from
users with few followers is spread more slowly and will take longer to affect stock
prices (Hong & Stein, 1999). Therefore, it will have a larger effect on stock returns
on future trading days (Hong & Stein, 1999; Tetlock et al., 2008). Thus:
H1: The sentiment in tweets about a specific firm sent by individuals with
few followers is directly related to stock returns on future trading days.
This sentiment diffusion process will also be affected by the extent to which
tweets are “retweeted”—that is, whether an individual who receives a tweet resends
it to his or her followers. A study of 37 billion public tweets found that the
percentage of retweets has increased over time: about 5% in 2010, 10% in 2011,
20% in 2012, and 25% in 2013 (Liu et al., 2014).
Individuals retweet for a variety of reasons. The most common reasons are
because they believe the tweet’s information would be of interest to their followers
or to express support for the original tweeter (Macskassy & Michelson, 2011; Liu
et al., 2014). In the investing context where tweets are deliberately tagged with the
$ and ticker symbol, we theorize that most tweets are retweeted because the sender
believes they have potential information for other investors.
Retweets affect the diffusion process. Retweeting someone else’s tweet is a
deliberate signal that the user believes the tweet would be of interest to his or her
followers. Retweeting spreads the sentiment in the tweet faster than if the tweet
was not retweeted and makes the tweet more influential because now two people
advocate for its content, not one. Sentiment in tweets that are retweeted will be
more quickly incorporated in stock price, so it will have less of an effect on returns
on future trading days. Therefore, it is the combination of few followers and not
being retweeted that leads to the greatest stock returns on future days because
the sentiment in these tweets will take the longest to diffuse through the network.
Therefore, we hypothesize:

Sul, Dennis, and Yuan 9
H2: The sentiment in tweets about a specific firm sent by individuals with
few followers that are not retweeted is directly related to stock returns on
future trading days.
METHODOLOGY
Financial Data
To ensure sufficient reliability of Twitter data, we focused only on firms that are
part of the S&P 500. Financial data, including the closing price of each stock in the
S&P 500, were obtained from Compustat, Center for Research in Security Prices
(CRSP), Institutional Brokers’ Estimate System (IBES), and Kenneth French’s
Web site (Rai, Patnayakuni, & Seth, 2006). The sample period is from March 2011
to January 2013.
Twitter Data
This study used data collected from Twitter. The focus of this article is on whether
the sentiment in tweets about an individual firm can predict stock returns. Thus,
it is important to match tweets to specific firms. The convention in Twitter is to
precede the stock ticker symbol with a dollar sign ($) to indicate that a tweet
contains investment information about a firm. We collected all public tweets that
contained the relevant $ symbol with an S&P 500 stock ticker from Twitter using a
developer account. We retrieved 3,475,428 tweets during the sample time period.
Of all the tweets, 16.02% were retweets. We excluded all the tweets that contained
more than one ticker symbol because we could not be sure if the information in the
tweet pertained to one firm or all firms equally. For example, a tweet like “I also
like long $AAPL @347.40 . . . and short $RIMM @62.70” would be excluded
from the analysis. This produced a final sample of 2,503,385 tweets. An inspection
of 500 randomly selected tweets found no tweets from the firm itself. Figure 1
shows the distribution of the tweets by days of the week.
The Sentiment in Tweets
There are many approaches to sentiment analysis (Feldman, 2013). We used the
word analysis strategy. Each word in a tweet was matched to a dictionary of terms
to determine its sentiment. We used the Harvard-IV dictionary (Jorgenson & Vu,
2005), which is a commonly used source for word classification in the financial
content analysis of popular press articles and Web news sites, used, for example,
by Tetlock (2007), Tetlock et al. (2008), and Da, Engelberg, and Gao (2011). There
are other dictionaries that could be used (e.g., the financial dictionary of Loughran
and McDonald (2011)), but these dictionaries are designed for the analysis of legal
and financial documents which contain formal English (e.g., 10K filings), not the
slang version of English used in Twitter.
CLT argues that positive and negative sentiment may have different effects
(Liberman & Trope, 1998; Bar-Anan et al., 2006; Fujita et al., 2006), so it is
important to track both positive and negative sentiment because they may have
different effects. Empirical research on stock returns has shown that sometimes

10 Trading on Twitter
Figure 1: Number of tweets (top) and percentage distribution (bottom) by the day
of the week.
0
50000
100000
150000
200000
250000
300000
350000
400000
450000
Mon Tue Wed Thurs Fri Sat Sun
Number of Tweets by the Day of the Week
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
Mon Tue Wed Thurs Fri Sat Sun
Percentage of Tweets by the Day of the Week
positive sentiment has an impact (Smailovi´
c et al., 2014), while other times negative
sentiment has an impact (Risius et al., 2015).
We counted all words in the tweets that had the “NEG” tag in the Harvard-IV
dictionary as words that conveyed a negative sentiment. We counted “POS” tagged
words as words conveying a positive sentiment. Although this approach has been
widely used in prior research (Tetlock et al., 2008), it is an imperfect measure of
sentiment, because it cannot detect subtle meanings in English, such as sarcasm or
semantic word groups that combine positive and negative words (e.g., “not bad”:

Sul, Dennis, and Yuan 11
Xie et al., 2015). Likewise, we did not include emoticons in our analysis, so this
is another limitation (Xie et al., 2015). An analysis of 100,000 randomly selected
tweets from our sample found 430 to contain emoticons (i.e., less than 1%).
We used three separate measures to better model sentiment, as has been done
in prior research (Tetlock et al., 2008). If our measures do not accurately capture
sentiment, then we are less likely to find a significant relationship, so this approach
is a more conservative test of the relationship between sentiment and stock returns
than human analysis of the tweets, which would be effectively impossible given
our sample size of over 2.5 million tweets. Table 1 shows descriptive statistics
about the tweets.
Because we are using daily stock returns as our dependent variable, we
combined all tweets for each firm on a given day. Daily returns are defined as
close-to-close daily returns, so we match day treturn with firm level Twitter
content on day tup to the market close time of 4 p.m. New York’s time. Any tweet
that was posted after 4 p.m. was treated as day t+1. Following Tetlock et al.
(2008), we used three variables to measure sentiment. Sentiment is measured as
following, where P,N, and Tare the daily aggregate number of positive, negative,
and total words for each day for a given firm.
Sentiment =
⎧
⎪
⎪
⎨
⎪
⎪
⎩
neg1 ≡N
T
pos1 ≡P−N
P+N
pos2 ≡log 1+P
1+N
.(1)
Conceptually, neg is the ratio of the amount of negative sentiment to the
total communication (positive, negative, and neither). Pos1 is a normalized ratio
(on a –1 to +1 scale) of the overall positive or negative sentiment expressed
(omitting words with no sentiment). Pos2 is an unstandardized ratio of positive
to negative sentiment, but log adjusted to capture the potential for diminishing
marginal effects. All three measures may produce similar results, but we included
all three for greater insight. Descriptive statistics can be found in Table 2.
Analysis
To answer the question of whether social media have sentiment information that can
predict future returns, we examine whether the speed of information dissemination
(i) reflected by the number of followers and (ii) retweet history is associated with
future returns.
To answer the first research question, we test the following equation:
CARi
t, t+n=α+β0sentimentui
t+β1sentimentoi
t+γCVi
t+i
t,(2)
where CARi
t,t+nis the cumulative abnormal return about firm i from day t+1to
day t+n; sentimentui
tis the sentiment about firm ion day texpressed in tweets
from users with a number of followers at or under a given threshold; sentimentoi
tis
the sentiment about firm ion day texpressed in tweets from users with a number
of followers over a given threshold; and CV are five control variables, as described
below.

12 Trading on Twitter
Table 1: Descriptive statistics of tweets.
March April May June July August September October November December January February
2011 2011 2011 2011 2011 2011 2011 2011 2011 2011 2012 2012
Total stock tickers 497 497 498 497 498 498 498 498 498 498 498 497
Total number of tweets 126,926 216,303 197,978 160,119 189,655 244,763 205,893 215,702 199,471 171,531 190,144 189,079
Average number of
tweets per firm
255.38 435.22 397.55 322.17 380.83 491.49 413.44 433.14 400.54 344.44 381.82 380.44
Average number of
words per tweet for all
firms
15.299 15.615 15.437 15.358 15.161 15.130 14.804 14.953 14.662 15.267 15.160 14.735
Average number of
positive words per
tweet for all firms
0.400 0.433 0.401 0.359 0.261 0.331 0.289 0.322 0.309 0.322 0.349 0.336
Average number of
negative words per
tweet for all firms
0.245 0.264 0.247 0.223 0.254 0.220 0.204 0.217 0.195 0.203 0.229 0.207
Average percentage of
positive words for all
firms
0.026 0.027 0.026 0.023 0.023 0.022 0.019 0.021 0.021 0.021 0.023 0.023
Average percentage of
negative words for all
firms
0.016 0.017 0.016 0.014 0.016 0.014 0.014 0.014 0.013 0.013 0.015 0.013
SD % positive words for
all firms
0.012 0.013 0.011 0.012 0.012 0.012 0.013 0.012 0.012 0.011 0.012 0.011
SD % negative words for
all firms
0.009 0.010 0.010 0.008 0.010 0.008 0.009 0.008 0.009 0.008 0.008 0.008

Sul, Dennis, and Yuan 13
Table 2: Descriptive statistics of the firm/trading day data.
NMean Variance SD Min 25th Percentile 50th Percentile 75th Percentile Max
Pos1 Sentiment i,t 119,727 0.254 0.474 0.688 –1.000 –0.111 0.333 1.000 1.000
Pos 2 Sentiment i,t 119,727 0.347 0.704 0.839 –4.500 –0.182 0.452 0.860 5.226
Neg Sentiment i,t 119,727 0.040 0.002 0.042 0.000 0.000 0.035 0.060 0.667
Surprise i,t 119,727 0.001 0.001 0.031 –2.080 0.000 0.000 0.000 3.950
Control 2i,t-30 ,t-2112,938 –0.005 0.008 0.092 –1.010 –0.052 0.000 0.047 0.652
Control 1 i,t-1112,858 0.000 0.000 0.019 –0.673 –0.008 0.000 0.008 0.532
Upgrade i,t 119,727 0.023 0.029 0.171 0.000 0.000 0.000 0.000 9
Downgrade i,t 119,727 0.021 0.023 0.151 0.000 0.000 0.000 0.000 5
Notes: Surprise i,t: Earnings surprise, relative to median analyst estimate.
Control 1 i,t-30,t-2: Past returns, cumulative abnormal return from the [–30, –2] trading window.
Control 2 i,t-1: The abnormal return on the prior trading day.
Upgrade i,t: The number of financial analyst upgrades for company i on day t.
Downgrade i,t: The number of financial analyst downgrades for company i on day t.

14 Trading on Twitter
Table 3: Definitions of variables.
ARi
tThe ARi
tis the abnormal return of firm ion date t, adjusted
using the size and book-to-market matched characteristic
portfolio’s return.
ARi
t=Ri
t−Pf oi
t,
where Ri
t=ln(ri
t+1), and Pf oi
t=1
nn
j=1wj(Rj
t). Note
that ri
tis the daily holding period return (ret) in the CRSP
daily stock database (CRSP.DSF) including and wjis the
value weight of the jth firm in the portfolio of firms
j=1∼n, such that n
j=1wj=1. The size and
book-to-market characteristic portfolio was formed using the
30th and 70th NYSE book-to-market percentiles and the
median NYSE market equity.
CARi
t, t+nThe CARi
t, t+nis future cumulative abnormal return of firm ion
date t, the dependent variable in our regressions. It is the
summation of the abnormal returns of the next ndays
starting day t+1.
CARi
t,t+n=
n

j=1ARi
t+j
Control1i,t The Control1i,t is the abnormal return of firm ifor the date
t−1. Control1i,t =ARi
t−1
Control2i,t The Control2i,tis the cumulative abnormal return, or the
summation of the abnormal return, of firm ifor the previous
30 days excluding the abnormal return of the previous date.
Control2i,t =
30

j=2ARi
t−j
Earnings Surprisei,t An earnings surprise is calculated for each firm ion each
earnings announcement date. The Earnings Surprisei,t is
calculated as the difference between the actual EPS (actual)
and the median EPS (medest) from the IBES summary
statistics database (IBES.STATSUM).
Upgradei,t An upgrade/downgrade is recorded as 1 if an analyst
increased/decreased the IBES recommendation code
Downgr adei,t (ireccd) from IBES recommendation detail database
(IBES.RECDDET). The Upgradei,t ,Downgradei,t is the
summation of the number of upgrade/downgrade for all
analysts for firm ion the same date.
Equation (2) examines future abnormal returns for days 1 to nafter the tweets
were made. Table 3 shows how we calculated CARi
t,t+n. We have chosen to use
three time periods: next day returns (i.e., n=1), next-day-to-10th-day returns
(i.e., n=10), and next-day-to-20th -day returns for a longer view (i.e., n=20).
These are trading days, so 10 days is approximately 2 weeks, and 20 days is
approximately one month. These time periods are consistent with prior research
(e.g., Tetlock et al., 2008; Fang & Peress, 2009; Chen, De, Hu, & Hwang, 2014).

Sul, Dennis, and Yuan 15
Table 4a: Correlations among the sentiment variables used in H1.
Pos1 Over Pos2 Over Neg Over Pos1 Under Pos2 Under
Pos2 Over .905
Neg Over –.800 –.741
Pos1 Under .253 .253 –.193
Pos2 Under .242 .276 –.196 .905
Neg Under –.191 –.193 .199 –.811 –.766
Notes: All correlations are significant at p<.001. Cells in gray are correlations among
variables used in the same regression analysis.
We used five control variables. Stock returns exhibit autocorrelation so we
included two control variables to control for autocorrelation: control1 and control2
capture the abnormal returns on the day before (i.e., t– 1) and the cumulative return
over the prior 30 days, respectively (Tetlock et al., 2008; Chen et al., 2014) (Table
3). The third control variable is earnings surprise, calculated as the actual earnings
per share for a given firm announced on a given day minus the median analyst
earnings per share prediction, where the median analyst prediction is the “Median
Estimate” from IBES Summary. The last two control variables were the upgrades
and downgrades on the company from professional stock analysts as control vari-
ables because upgrades and downgrades can influence stock returns (Chen et al.,
2014). We counted the number of upgrades and downgrades on the specific firm’s
stock on the same trading day as the tweets and included these numbers as controls.
As is commonly done in financial research, we obtained analyst recommendations
from the IBES, categorized each change in recommendation as either an upgrade
or downgrade, and counted the number of each on each trading day (Chen et al.,
2014) (Table 3).
To examine H1, the impact of the number of followers, we need to test
whether β0is positive when sentiment is pos1 or pos2 and test whether β0is
negative when sentiment is neg for the different trading periods. We split the
tweets into two groups based on the number of followers of the tweeters, those
with many followers and those with few followers. The question is, what is “many”
and “few?” The median number of followers in our sample was 171, so we selected
this as the break point for assigning tweets into groups of users with few followers
and many followers. The sample size was 48,538 because we can analyze the data
only when there are tweets from individuals both over and under the threshold on
the same day for the same firm.
To examine H2, the combined impact of number of followers and retweets,
we divided the tweets into four groups: many followers and retweeted; many
followers and not retweeted; few followers and retweeted; and few followers and
not retweeted. We used the same break point (171) as the threshold for assigning
tweets into groups with few followers and many followers. The sample size was
8,245 because we can analyze data only when there are tweets in all four groups
on the same day for the same firm.
Tables 4a and (b) shows the correlations among the sentiment variables. If the
sentiment in tweets from those with few and many followers are highly correlated,
multicollinearity could bias the results. The correlations indicate little risk due to

16 Trading on Twitter
Table 4b: Correlations among the sentiment variables used in H2.
1234567891011
1. Pos1 over not retweeted
2. Pos1 over retweeted .189
3. Pos1 under not retweeted .103 .067
4. Pos1 under retweeted .078 .209 .102
5. Pos2 over not retweeted .927 .235 .125 .117
6. Pos2 over retweeted .174 .910 .072 .265 .239
7. Pos2 under not retweeted .096 .077 .915 .115 .131 .091
8. Pos2 under retweeted .074 .232 .101 .906 .121 .337 .123
9. Neg over not retweeted –.793 –.151 –.085 –.063 –.750 –.141 –.080 –.062
10. Neg over retweeted –.145 –.436 –.007 –.074 –.142 –.405 .021 –.087 .151
11. Neg under not retweeted –.069 –.003 –.495 –.001 –.050 .002 –.436 –.002 .066 .144
12. Neg under retweeted –.067 –.073 .010 –.411 –.059 –.090 .041 –.383 .065 .304 .199
Note: All correlations shown in bold are significant at p<.001. Cells in gray are correlations among variables used in the same regression analysis.

Sul, Dennis, and Yuan 17
multicollinearity, with the highest correlation between variables in the same model
being less than .30. We included Variance Inflation Factors in all analyses and
found that most were less than 1.2, and none exceeded 2.0, indicating that it is
highly unlikely that our data suffer from multicollinearity.
One of the issues with large data sets is that the traditional approach of
using pvalues can be misleading because the large sample size means that any
relationship is likely to be significant (Lin, Lucas, & Shmueli, 2013). Lin et al.
(2013) offer several strategies for the analysis of large data sets. We adopt three
of their recommendations, plus a fourth traditionally used in the analysis of large
sample stock return data. First, we present confidence intervals for the size of
effects. Second, we conduct a series of robustness checks using alternate models
to see the extent to which our results are dependent on the specific models we use.
Third, we examine the predictive ability of the models by comparing them to a
controls-only model using symmetric mean absolute percent error (SMAPE), the
absolute value of the difference between the predicted and actual divided by the
mean of the absolute value of the predicted and the absolute value of the actual
(Armstrong, 1985; Makridakis, 1993; Tofallis, 2015). Finally, one of the strongest
tests of the practical significance of models used to predict stock returns is a trading
strategy analysis—a test of whether an investor who builds a trading strategy using
the results would experience a profit after accounting for trading costs (Tetlock
et al., 2008).
RESULTS
Impact of Followers
We begin with H1, which argues that future abnormal returns would be directly
related to the sentiment in tweets from users with few followers. Table 5 shows
that the beta coefficients are significant and in the hypothesized direction for all
three measures (pos1, pos2, and neg) for all three time periods (next day, next-to-
10th day, and next-to-20th day), except for next day returns for neg, which is in
the hypothesized direction but not significant.
The adjusted R2for these analyses are in a similar range to those in other
studies of stock returns, such as (Chordia, Roll, & Subrahmanyam, 2002; Tetlock,
2007; Tetlock et al., 2008; Bollen et al., 2011a ; Chen et al., 2014). SMAPE values
for the models with only the five control variables for next day returns, next-to-
10th -day, and next-to-20th -day are 1.9231, 1.8685, and 1.8210, respectively. The
SMAPE values for all nine models in Table 5 are below the SMAPE values for
the controls-only models, indicating they are a better fit. Based on these results
(8 of 9 hypothesis tests significant, R2equivalent to R2in prior research, and lower
SMAPE), we conclude that H1 is supported.
Combined Impact of Followers and Retweets
H2 argues that future abnormal returns would be directly related to the sentiment
in tweets from users with few followers that are not retweeted. Table 6 shows
that the beta coefficients on the sentiment in tweets from those with few followers
that were not retweeted are significant and in the hypothesized direction for all

18 Trading on Twitter
Table 5: Regression results of emotional sentiment on abnormal returns by number
of followers.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.140 1.020*1.103
Sentiment Under i,t 0.388** 1.512*** 2.582***
Control 1 i,t-10.004 0.017 –0.011
Control 2i,t-30 ,t-20.001 0.013*** 0.022***
Surprise i,t 0.010*** 0.004 0.001
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.003
Intercept 0.000*** –0.003*** –0.006***
SMAPE 1.9208 1.8662 1.8200
Adj. R20.001 0.001 0.001
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.105 0.618 0.133
Sentiment Under i,t 0.349** 1.405*** 2.674***
Control 1 i,t-10.004 0.017 –0.010
Control 2i,t-30 ,t-20.001 0.012*** 0.022***
Surprise i,t 0.010*** 0.004 0.001
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.003
Intercept 0.000*** –0.004*** –0.007***
SMAPE 1.9190 1.8648 1.8187
Adj. R20.002 0.001 0.002
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t 2.313 –13.140 –6.829
Sentiment Under i,t –3.265 –22.735*** –52.175***
Control 1 i,t-10.004 0.018 –0.010
Control 2i,t-30 ,t-20.001*0.013*** 0.022***
Surprise i,t 0.010*** 0.004 0.001
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.003
Intercept 0.000*** –0.001** –0.003***
SMAPE 1.8200 1.8187 1.8201
Adj. R20.001 0.001 0.002
Notes: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.
three measures (pos1, pos2, and neg) for all three time periods (next day, next-to-
10th day, and next-to-20th day), except for next-to-10th day returns for neg which
is in the hypothesized direction but not significant. Table 7 presents confidence
intervals for the betas.
The adjusted R2for these analyses are equivalent to or substantially higher
(by an order of magnitude—i.e., 1,000%) than adjusted R2in prior studies (Chordia

Sul, Dennis, and Yuan 19
Table 6: Regression results of emotional sentiment and retweeting on abnormal
returns by number of followers.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t –0.531 2.043 1.871
Sentiment OverRe i,t 0.162 –0.407 –1.000
Sentiment UnderNo i,t 0.917*3.598** 5.850**
Sentiment UnderRe i,t 0.468 –0.037 –0.122
Control 1 i,t-10.020*0.006*0.036
Control 2i,t-30 ,t-20.003 0.043*** 0.082***
Surprise i,t 0.008** 0.000 0.007
Upgrade i,t 0.002** –0.002 –0.005
Downgrade i,t –0.001 –0.005*0.004
Intercept –0.001*** –0.005*** –0.009***
SMAPE 0.2219 0.2148 0.2172
Adj. R20.002 0.008 0.013
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t –0.126 0.778 –0.116
Sentiment OverRe i,t –0.327 –0.412 –1.214
Sentiment UnderNo i,t 0.831** 3.353*** 4.649***
Sentiment UnderRe i,t 0.522 0.221 0.901
Control 1 i,t-10.020*0.064 0.036
Control 2i,t-30 ,t-20.004 0.043*** 0.008***
Surprise i,t 0.008*0.000 0.007
Upgrade i,t 0.002 –0.002 –0.005
Downgrade i,t –0.001 0.005 –0.004
Intercept –0.001*** –0.005*** –0.009***
SMAPE 0.2217 0.2146 0.2172
Adj. R20.003 0.009 0.013
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t 3.008 –57.433 –79.755
Sentiment OverRe i,t 8.017 10.249 25.681
Sentiment UnderNo i,t –14.274*–40.149 –67.144*
Sentiment UnderRe i,t –7.962 3.638 2.038
Control 1 i,t-10.020*0.064*0.035
Control 2i,t-30 ,t-20.003 0.004*** 0.082***
Surprise i,t 0.009*0.000 0.007
Upgrade i,t 0.002*–0.002 –0.005
Downgrade i,t –0.001 0.005*0.004
Intercept 0.000 0.000 –0.003
SMAPE 0.2219 0.2148 0.2172
Adj. R20.002 0.008 0.012
Notes: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.

20 Trading on Twitter
Table 7: Confidence intervals for beta for sentiment in non-retweeted tweets from
individuals with 171 or fewer followers.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment UnderNoi,t 0.053 to 1.781 0.946 to 6.250 2.048 to 9.652
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment UnderNoi,t 0.213 to 1.449 1.459 to 5.247 1.932 to 7.366
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment UnderNoi,t –0.005 to –28.543 3.635 to –83.933 –4.361 to –129.927
et al., 2002; Tetlock, 2007; Tetlock et al., 2008; Bollen et al., 2011a; Chen et al.,
2014). SMAPE values for the models with only the five control variables for
next day returns, next-to-10th -day, and next-to-20th -day are 1.8476, 1.7763, and
1.7042, respectively. The SMAPE values for all nine models in Table 6 are below
the SMAPE values for their matching controls-only models, indicating they are
a better fit. Based on these results (eight of nine hypothesis tests significant, R2
equivalent to or higher than R2in prior research, and lower SMAPE), we conclude
that H2 is supported.
These results support our arguments that the speed of sentiment diffusion
affects future returns. When sentiment spreads the slowest (i.e., tweets sent by those
with fewer than the median number of followers [171] that are not retweeted), it
affects stock returns on future trading days.
Robustness Checks
We conducted several robustness checks. We ran a separate analyses for H1 and H2
treating missing values as zero emotion (which produced sample sizes of 83,891)
and found the same pattern of results.
We conducted a separate analysis for H1 using a different split in the number
of followers. We used 1,000 followers as the threshold between many and few. A
user with 1,000 followers falls in the top 4% of all Twitter users (“About Twitter,
Inc.,” 2014), so they are what Barab´
asi (2002) would call hubs. The Twitter users
with over 1,000 followers were typically well-known media or analysts, such as
CNN, ABC, WSJ, CNBC, Fox News, Fortune Magazine, and Jim Cramer, who
is a writer, TV show host, and co-founder of TheStreet.com. This produced the
analysis of “few” and “many” followers, with a 96–4% split. The split using 1,000
followers followed the same pattern as with the median split (Table A1 in the
Appendix). We conclude that this hypothesis is robust to the choice of threshold
for the number of followers. For this hypothesis, as long as one is not a hub (i.e., the
top 4% of all Twitter users), it takes days before sentiment in your tweets spreads.
We conducted a similar analysis for H2 using the 1,000 follower threshold
(96–4% split). The sample size here was 9,014. The pattern here was different;
only three of the nine hypothesis tests were significant (Table A2 in the Appendix).
We conclude that this hypothesis is not robust to the choice of threshold for what
“few” followers means. If we consider retweeting behavior, then the threshold

Sul, Dennis, and Yuan 21
Table 8: Annualized returns from a trading strategy using sentiment in non-
retweeted tweets from individuals with 171 or fewer followers.
Holding Period
1 Day 10 Days 20 Days
Without trading costs 11.44% 17.91% 12.59%
With trading costs –28.57% 15.65% 11.41%
number to identify “few” followers must be such that we do not consider only
hubs and nonhubs in the network.
We examined the effects of the sentiment contained in tweets only from
those with many followers, omitting the sentiment in tweets from those with
few followers. Four of the nine hypothesis tests were significant (Table A3 in
the Appendix), but the SMAPE values are not improved. We conclude that the
sentiment in tweets from those with many followers is not consistently related to
future returns.
If the sentiment in tweets does affect the market, we are likely to see an
effect on the same day by users with many followers. For example, if Jim Cramer
tweets positively or negatively about a specific stock then its price should move
quickly. Table A4 in the Appendix shows the same day effects (i.e., t=0);
there is a significant same day effect for the sentiment in tweets from users with
many followers both with and without considering the sentiment in tweets from
those with few followers. All three betas on the sentiment from those with many
followers are greater than the corresponding betas for those with few followers,
with significance of p<.001. The SMAPE values are lower than those of the
corresponding controls-only models.
Effectiveness of a Trading Strategy
One important question is whether these results can be used to build a profitable
trading strategy. We theorized, and the empirical results show that the tweets from
users with few followers that are not retweeted lead to the greatest abnormal
returns. We followed the approach of Tetlock et al. (2008) and constructed two
equally weighted portfolios, one long, one short. At the close of each trading day,
we analyzed the sentiment in that day’s tweets about specific firms using pos2 (we
choose pos2 because it takes into account both positive and negative sentiment).
We purchase firms in the top 10% and short sell the firms in the bottom 10%. Not
all firms receive tweets each day, so the number of firms varies from day to day.
We used three different holding periods (1 day, 10 days, and 20 days), and at the
end of the holding period, we close out our long and short positions. Because we
are simultaneously taking long and short positions, there is no need to consider
market return as a control; any rise or fall in the market as a whole is controlled
for by the simultaneous long and short positions.
Table 8 presents the annualized returns of the trading strategy with and
without trading costs for the three different holding periods. Following Tetlock

22 Trading on Twitter
et al. (2008), we assume round-trip trading costs of 10 basis points (i.e., the
total cost to buy and sell). The 1-day holding period produces positive returns,
but because the strategy executes trades every day, the returns become negative
after including trading costs. The trading strategies using 10- and 20-day holding
periods produce significant positive returns, both before and after trading costs.
These returns compare favorably with those in Tetlock et al. (2008), who found
trading strategies using sentiment in news stories produced returns of 23.17%
before trading costs and –2.71% after trading costs (i.e., a loss). In other words, the
results in Table 8 show that a trading strategy with a 10- or 20-day holding period
that balances long and short positions results in meaningful positive returns.
DISCUSSION
We theorized that the sentiment in tweets is related to stock returns of individual
stocks based on how fast that sentiment spreads through the market. The sentiment
in tweets sent by users with few followers, which diffuses more slowly than
sentiment in tweets sent by those with many followers, is significant in predicting
the firm’s stock returns one trading day, 10 trading days, and 20 trading days after
the tweets were posted. The sentiment in tweets from those with few followers
that were not retweeted had the strongest effect on returns on future days, as this
sentiment takes the longest to spread through the market. A trading strategy with
a 10- or 20-day holding period built on these factors shows meaningful annual
returns.
We argue that a social contagion process is at work. Tweets spread positive
or negative sentiment about a stock through the market and can influence prices,
and thus the returns from trading those stocks. Sentiment can spread quickly;
for example, network hubs like Jim Cramer can send a tweet that has positive
or negative sentiment about a stock and his 650,000 followers immediately see
its sentiment. If these followers act, the stock price can respond very quickly.
Thus, we conclude that Twitter users with many followers have a market impact
similar to traditional news media; the impact of the sentiment in their tweets
disseminates rapidly and is quickly incorporated into stock price. However, there
are no significant stock returns on future trading days and thus it is difficult to profit
from a trading strategy based on them. In contrast, if a user with few followers
sends a tweet, few people will see it, and even if they quickly act on its sentiment,
the small number of trades will have little immediate impact on the stock price.
Over time, however, as the sentiment diffuses through the market, the sentiment
will gradually affect the stock price. The sentiment in tweets from users with fewer
followers had a stronger impact on stock returns the next trading day and over
the next 10–20 days compared to tweets from users with many followers. Because
the change is gradual, there is an opportunity to profit from this as a trading
strategy. The diffusion is slower for tweets that are not retweeted, and thus they
have the greatest impact on returns on future trading days and offer the greatest
opportunities to profit from a trading strategy based on this.
Our results offer similar conclusions to other research based on the GIF
model of information diffusion. GIF argues that markets are generally efficient at
the macro level, but if we examine them at a micro level, we see that it is possible

Sul, Dennis, and Yuan 23
to uncover situations in which markets are not perfectly efficient, because human
behavior is not perfectly efficient at spreading information. Studies in line with
GIF have suggested that it takes longer for information about stocks that are not
routinely covered by the mass media to be absorbed into their price (Merton, 1987;
Fang & Peress, 2009). Thus, it is possible to use information to predict future
stock returns for firms with systematically slower information diffusion, even after
controlling for risk characteristics (Merton, 1987; Fang & Peress, 2009). Thus, we
believe the GIF model provides a good theoretical foundation for understanding
why Twitter sentiment can be used to predict future stock returns in some cases
but not in others.
There exist at least two possible theoretical explanations for how the speed
of information diffusion influences stock returns. The first is that the sentiment of
tweets “causes” changes in stock prices. Individuals post tweets when they believe
they have useful comments about an individual stock. These comments may have
facts as well as an underlying sentiment. Sentiment is highly contagious (Schoe-
newolf, 1990; Hatfield et al., 1993), and it influences how investors make buy/sell
decisions as the sentiment spreads through the public. A cumulative positive senti-
ment triggers positive thoughts about the company and leads to a purchase decision,
raising the stock price. A cumulative negative sentiment induces negative thoughts
and thus leads to sell decisions, decreasing the stock price. The rate at which this
sentiment spreads through the market is influenced by the number of followers of
the sender and whether the tweet is retweeted, so that sentiment in tweets from
those with few followers that are not retweeted takes a longer time to spread and
thus takes longer to influence stock prices; this leads to significant stock returns
on future trading days.
A second possible explanation is that tweets “reflect” the underlying infor-
mation that influences individual stock returns. In this case, it is not the sentiment
of the tweets themselves that influence stock returns, but rather the tweets reflect
how investors feel about the stock and are a leading indicator of their buy/sell de-
cisions. Investors planning to buy a stock have positive sentiment about the stock
and communicate this sentiment in their tweets. Likewise, investors planning to
sell a stock communicate negative sentiment in their tweets.
We believe that the first explanation, that the sentiment of tweets causes
stock price changes, best explains our findings because the number of followers
and whether the tweet was retweeted were significantly related to stock price
changes. Two assumptions would have to hold for the explanation that tweets reflect
information to be viable. First, investors would have to tweet their information days
before acting on it, which is illogical; no rational investor would share information
likely to affect prices before acting on it. Second, the underlying information would
need to spread through the social network in the same manner as the tweets and
retweets but via a different mechanism in order for us to find the relationships
we did. This is a less likely explanation than the simpler explanation that it is the
tweets themselves that influence behavior.
CLT argues that positive and negative sentiment may have different effects
(Liberman & Trope, 1998; Bar-Anan et al., 2006; Fujita et al., 2006). Previous
research has shown that both positive and negative sentiment in tweets can affect
stock returns, but no study has found both to have effects (Smailovi´
c et al., 2014;

24 Trading on Twitter
Risius et al., 2015). Interestingly, we found both positive and negative sentiment
to directly affect returns.
Investors make investment decisions using a variety of information sources,
with Twitter being just one of many possible sources. The economic magnitude of
the relationships in our study is moderate to high (Tetlock et al., 2008). The trading
analysis showed positive annual returns for 10- and 20-day holding periods after
considering trading costs. The economic significance of these effects is meaningful.
Limitations
This study also has several limitations. We only studied firms in the S&P 500. We
have no empirical data to argue that our results apply or do not apply to smaller
firms or firms traded in other markets that are not covered by the S&P 500. We
studied one specific time period in the life of the market, so it could be that the
market conditions that led to our findings no longer apply. Likewise, we studied
the same time period in the life of Twitter, and because Twitter behavior changes
over time (Liu et al., 2014), it may be that Twitter users behave differently now,
and the behaviors we observed no longer occur.
The fundamental theory underlying our research is the GIF model. Our results
are driven by the speed of diffusion of sentiment, so one important theoretical
limitation is if this sentiment is based on already widely diffused fundamental
information (e.g., a rise in oil prices that could negatively influence transportation
stocks) then this sentiment is likely to have little effect on stock returns, because
investors have already acted and prices have already changed. We did not examine
the extent to which the sentiment in the tweets we analyzed was based on already
disseminated fundamental information, so this is an avenue for future research.
Another potential limitation is homophily, the possibility that individuals
similar to each other tend to post similar Tweets (Aral, Muchnik, & Sundararajan,
2009; Shalizi & Thomas, 2011). Under this argument, the changes in stock prices
are not due to social contagion but are because people similar to each other in the
number of followers use similar trading strategies. This is possible, but we view
this as less likely than social contagion because it could only be true if trading
behaviors were related to the number of followers and the retweet history of tweets.
This is possible, but requires additional, somewhat convoluted theorizing to link
the number of Twitter followers and retweeting history to trading behaviors. While
homophily is useful in understanding some tweeting behaviors, it is often not as
powerful as other theoretical models (Macskassy & Michelson, 2011). So, using
Occam’s razor, we conclude that the social contagion of sentiment is a better
explanation for our results.
Implications for Research
Despite these limitations, we believe that these results have implications for future
research. Our work builds on recent research showing that the “calmness” or
“depression” in tweets (Bollen et al., 2011a; Risius et al., 2015), their “bullishness”
or “beariness” (Oh & Sheng, 2011; Sprenger et al., 2014), and their sentitment
(Smailovi´
c et al., 2014; Risius et al., 2015) can be useful in predicting stock returns.
We use social contagion based on the GIF model as underlying theory and show
that factors which influence the speed of sentiment diffusion (number of followers,

Sul, Dennis, and Yuan 25
retweeting) significantly affect the stock returns on future trading days. We offered
two possible explanations for the theoretical mechanism that links the sentiment
in tweets to future stock returns. We need more research to better understand the
underlying theoretical mechanism that links sentiment to stock returns.
Our research shows that who sends the tweets is an important factor in
explaining how the sentiment in tweets affects stock returns. Ironically, users with
many followers (i.e., those with more than the median number of 171) have no
significant influence on stock returns on the next trading day or subsequent days.
The sentiment in their tweets is quickly incorporated into stock prices leading
to no future returns. In contrast, sentiment expressed by Twitter users who have
few followers—and thus diffuses slowly—has significant and meaningful impacts
on stock returns on future trading days. We used a simple analysis that divided
users into two groups, over and under the median number of followers in our data
set. We believe that this calls for more research into who expresses the sentiment
in tweets and how this can be used to explain stock price movements and better
predict stock returns. The number of connections in a social network typically
follows a power law distribution (Barab´
asi, 2002), so an analysis that uses more
than two categories to better capture this distribution may better model the speed
of sentiment diffusion and provide additional insight.
Our research also shows that what happens to the tweet after it is sent has a
significant impact on stock returns. Tweets that are retweeted have a faster impact
on stock prices and thus do not predict stock returns on future trading days, whereas
tweets that are not retweeted can predict future returns. We believe this calls for
more research into retweeting behavior. For example, how do investors react to
the tweets they receive that are and are not retweeted? The most common reason
for retweeting is because the sender believes the tweet’s information would be of
interest to their followers (Macskassy & Michelson, 2011). Do retweeted tweets
appear more important and thus get more attention, so they are more likely to
influence behavior?
Prior studies examining how emotion is linked to stock returns have used
different approaches to measuring it, including emotional states (e.g., calm, hap-
piness, depression) and sentiment (e.g., positive emotional valence, “bullishness”)
(Calvo & Kim, 2013). We used the Harvard IV Psychological Dictionary to as-
sess the positive and negative sentiment expressed in the tweets. There are many
dictionaries designed to categorize words based on sentiment, such as Loughran
and McDonald Financial Sentiment Dictionaries (Loughran & McDonald, 2011).
We did not include emoticons in our analysis, which could be examined in future
research.
We used the formulas of Tetlock et al. (2008) to build three different measures
of sentiment (pos1, pos2, neg) that provided essentially the same conclusions (with
some minor differences among them). There are many other formulas and machine
learning techniques that can be used to develop sentiment metrics that are more
sophisticated (e.g., Oh & Sheng, 2011; Sprenger et al., 2014; Smailovi´
c et al.,
2014). One key challenge in sentiment analysis is understanding semantics in
groups of words. A Twitter post may have both positive and negative terms, so if
one considers the semantic rules of groups of words, the meaning may become
clearer (e.g., “this is not bad.”) (Xie et al., 2015). Additional research is needed

26 Trading on Twitter
that uses different, more sophisticated, analysis strategies to better understand if
different approaches are better at predicting stock returns on future trading days.
In this study, we use Twitter as the social media platform. There are many
other social media platforms that may also provide insights into future stock
returns. We hope our work can spawn future research on this topic. What are the
impacts of Facebook, LinkedIn, or other Web media?
Finally, we examined the impact of sentiment on stock returns at a daily level.
Future research could use market microstructural data to examine how emotional
state and sentiment impact markets in real time.
Implications for Practice
We believe that this study has two implications for practice. The first is providing
guidance to investment decision making. Our results show that a trading strategy
built on the analysis of the sentiment in tweets from users with few followers that are
not retweeted produces significant positive returns after considering trading costs.
Tweets are available publically and can be retrieved using Twitter development
accounts, so this may be an investable trading strategy. Combining this with a
focus on firms that have little coverage from the traditional media may also increase
returns (Merton, 1987; Fang & Peress, 2009).
A second implication is that firms should carefully monitor how they use
Twitter. Most firms manage formal financial information that could impact stock
prices because there are numerous financial regulations in place. Because the
sentiment of tweets is linked to future stock prices, firms need to monitor the
sentiment in their tweets in addition to the “rational” information they contain.
CONCLUSION
We found that the sentiment in social media postings can predict stock returns
on future trading days. Tweets from users with few followers (i.e., less than the
median of 171 followers) that were not retweeted had an impact on future returns
10 and 20 days later, while those from users with many followers and those that
were retweeted had no impact on future returns. The findings are consistent with
our hypothesis that sentiment that is diffused slowly takes longer to be incorporated
into prices, while sentiment that is diffused faster will be quickly incorporated into
prices and thus will have little association with returns on future days.
REFERENCES
Abbasi, A., & Chen, H. (2008). CyberGate: A design framework and system for
text analysis of computer-mediated communication. MIS Quarterly,32(4),
811–837.
About Twitter, Inc. (2015). Accessed March 14, 2016. Available at:
https://about.twitter.com/company
Aral, S., Muchnik, L., & Sundararajan, A., (2009). Distinguishing influence-based
contagion from homophily-driven diffusion in dynamic networks. Proceed-
ings of the National Academy of Sciences,106(51), 21544–21549.

Sul, Dennis, and Yuan 27
Armstrong, J. S. (1985). Long range forecasting: From crystal ball to computer
(2nd ed.). New York, NY: Wiley.
Bakamitsos, G. A. (2006). A cue alone or a probe to think? The dual role of affect
in product evaluations. Journal of Consumer Research,33(3), 403–412.
Baker, M., & Wurgler, J. (2007). Investor sentiment in the stock market. Journal
of Economic Perspectives,21(2), 129–151.
Bar-Anan, Y., Liberman, N., & Trope, Y. (2006). The association between psycho-
logical distance and construal level: Evidence from an implicit association
test. Journal of Experimental Psychology-General,135(4), 609–622.
Barab´
asi, A. L. (2002). Linked. Cambridge, MA: Persues Publishing.
Bechara, A., & Damasio, A. R. (2005). The somatic marker hypothesis: A neural
theory of economic decision. Games and Economic Behavior,52(2), 336–
372.
Bollen, J., Mao, H., & Zeng, X. (2011a). Twitter mood predicts the stock market.
Journal of Computational Science,2(1), 1–8.
Bollen, J., Pepe, A., & Mao, H. (2011b). Modeling public mood and emotion:
Twitter sentiment and socio-economic phenomena. Proceedings of the Fifth
International AAAI Conference on Weblogs and Social Media, Barcelona.
Bradley, M. M., & Lang, P. J. (1994). Measuring emotion: The self-assessment
manikin and the semantic differential. Journal of Behavior Therapy and
Experimental Psychiatry,25(1), 49–59.
Cacioppo, J. T., Petty, R. E., Losch, M. E., & Kim, H. S. (1986). Electromyographic
activity over facial muscle regions can differentiate the valence and intensity
of affective reactions. Journal of Personality and Social Psychology,50(2),
260–268.
Calvo, R. A., & Kim, S. M. (2013). Emotions in text: Dimensional and categorical
models. Computational Intelligence,29(3), 527–543.
Chen, H., De, P., Hu, Y. J., & Hwang, B.-H. (2014). Wisdom of crowds: The value
of stock opinions transmitted through social media. Review of Financial
Studies,27(5), 1367–1403.
Chordia, T., Roll, R., & Subrahmanyam, A. (2002). Order imbalance, liquidity,
and market returns. Journal of Financial economics,65(1), 111–130.
Da, Z., Engelberg, J., & Gao, P. (2011). In search of attention. The Journal of
Finance,66(5), 1461–1499.
Daley, J., Jencks, S., Draper, D., Lenhart, G., Thomas, N., & Walker, J. (1988).
Predicting hospital-associated mortality for Medicare patients. JAMA: The
Journal of the American Medical Association,260(24), 3617–3624.
Dhir, A., Buragga, K., & Boreqqah, A. A. (2013). Tweeters on campus: Twitter a
learning tool in classroom? Journal of Universal Computer Science,19(5),
672–691.
Dockery, E., & Kavussanos, M. G. (1996). Testing the efficient market hypothe-
sis using panel data, with application to the Athens stock market. Applied
Economics Letters,3(2), 121–123.

28 Trading on Twitter
Ellison, N. B. (2007). Social network sites: Definition, history, and scholarship.
Journal of Computer-Mediated Communication,13(1), 210–230.
Eppen, G. D., & Fama, E. F. (1969). Cash balance and simple dynamic portfolio
problems with proportional costs. International Economic Review,10(2),
119–133.
Eyal, T., Liberman, N., Trope, Y., & Walther, E. (2004). The pros and cons of tem-
porally near and distant action. Journal Personality and Social Psychology,
86(6), 781–795.
Fang, L., & Peress, J. (2009). Media coverage and the cross-section of stock returns.
The Journal of Finance,64(5), 2023–2052.
Fedorikhin, A., & Patrick, V. M. (2010). Positive mood and resistance to temp-
tation: The interfering influence of elevated arousal. Journal of Consumer
Research,37(4), 698–711.
Feldman, R. (2013). Techniques and applications for sentiment analysis. Commu-
nications of the ACM,56(4), 82–89.
Frijda, N. H. (1994). Varieties of affect: Emotions and episodes, moods, and
sentiments. In P. Ekman & R. J. Davidson (Eds.), The nature of emotion:
Fundamental questions. Oxford, UK: Oxford University, 197–202.
Fujita, K., Trope, Y., Liberman, N., & Levin-Sagi, M. (2006). Construal levels and
self-control. Journal of Personality and Social Psychology,90(3), 351–367.
Gordon, T., Castelli, W. P., Hjortland, M. C., Kannel, W. B., & Dawber, T. R.
(1977). Predicting coronary heart disease in middle-aged and older persons.
JAMA: The Journal of the American Medical Association,238(6), 497–499.
Hatfield, E., Cacioppo, J. T., & Rapson, R. L. (1993). Emotional contagion. Current
Directions in Psychological Science,2(3), 96–99.
Honey, C., & Herring, S. C. (2009). Beyond microblogging: Conversation and
collaboration via Twitter. Proceedings of the 42nd Hawaii International
Conference on System Science, Waikoloa.
Hong, H., Lim, T., & Stein, J. C. (2000). Bad news travels slowly: Size, ana-
lyst coverage, and the profitability of momentum strategies. The Journal of
Finance,55(1), 265–295.
Hong, H., & Stein, J. C. (1999). A unified theory of underreaction, momentum
trading, and overreaction in asset markets. The Journal of Finance,54(6),
2143–2184.
Hong, H., & Stein, J. C. (2007). Disagreement and the stock market. Journal of
Economic Perspectives,21(2), 109–128.
Jorgenson, D. W., & Vu, K. (2005). Information technology and the World econ-
omy. The Scandinavian Journal of Economics,107(4), 631–650.
Labroo, A. A., & Patrick, V. M. (2009). Psychological distancing: Why happiness
helps you see the big picture. Journal of Consumer Research,35(5), 800–809.
Larsson, A. O. (2013). Tweeting the viewer—Use of Twitter in a talk show context.
Journal of Broadcasting & Electronic Media,57(2), 135–152.

Sul, Dennis, and Yuan 29
Liberman, N., & Trope, Y. (1998). The role of feasibility and desirability consid-
erations in near and distant future decisions: A test of temporal construal
theory. Journal of Personality and Social Psychology,75(1), 5–18.
Lin, M., Lucas, H. C. Jr., & Shmueli, G. (2013). Too big to fail: Large samples
and the p-value problem. Information Systems Research,24(4), 906–917.
Liu, Y., Kilman-Silver, C., & Mislove, A. (2014). The tweets they are a-changin’:
Evolution of Twitter users and behavior. Proceedings of the Eighth Interna-
tional AAAI Conference on Weblogs and Social Media, Ann Arbor, MI.
Loughran, T., & McDonald, B. (2011). When is a liability not a liability? Textual
analysis, dictionaries, and 10-Ks. The Journal of Finance,66(1), 35–65.
Macskassy, S. A., & Michelson, M. (2011). “Why do people retweet? Anti-
homophily wins the day!” Proceedings of the Fifth International AAAI Con-
ference on Weblogs and Social Media, Barcelona.
Makridakis, S. (1993). Accuracy measures: Theoertical and practical concerns.
International Journal of Forecasting,9(4), 527–529.
Malkiel, B. G., & Fama, E. F. (1970). Efficient capital markets: A review of theory
and empirical work. The Journal of Finance,25(2), 383–417.
Merton, R. C. (1987). A simple model of capital market equilibrium with incom-
plete information. The Journal of Finance,42(3), 483–510.
Oh, O., Agrawal, M., & Rao, H. R. (2013). Community intelligence and social
media services: A rumor theoretic analysis of tweets during social crises.
MIS Quarterly,37(2), 407–426.
Oh, O., & Sheng, O. (2011). Investigating predictive power of stock micro blog sen-
timent in forecasting future stock price directional movement. Proceedings
of the International Conference on Information Systems, Shanghai.
Osgood, C., Suci, G., & Tannenbaum, P. (1957). The measurement of meaning.
Urbana, IL: University of Illinois.
Parboteeah, D. V., Valacich, J. S., & Wells, J. D. (2009). The influence of website
characteristics on a consumer’s urge to buy impulsively. Information Systems
Research,20(1), 60–78.
Pew-Research. (2014). Social media use over time, accessed Novem-
ber 11, 2014, available at http://www.pewinternet.org/data-trend/social-
media/social-media-use-all-users/.
Qian, B., & Rasheed, K., (2007). Stock market prediction with multiple classifiers.
Applied Intelligence,26(1), 25–33.
Qiu, C., & Yeung, C. W. (2008). Mood and comparative judgment: Does mood
influence everything and finally nothing? Journal of Consumer Research,
34(5), 657–669.
Rai, A., Patnayakuni, R., & Seth, N. (2006). Firm performance impacts of digitally
enabled supply chain integration capabilities. MIS Quarterly,30(2), 225–
246.
Risius, M., Akolk, F., & Beck, R. (2015). Differential emotions and the stock
market—The case of company-specific trading. European Conference on

30 Trading on Twitter
Information Systems, Completed Research Papers, Munster, Germany, Paper
147.
Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and
Social Psychology,39(6), 1161–1178.
Russell, J. A. (2003). Core affect and the psychological construction of emotion.
Psychological Review,110(1), 145–172.
Sakaki, T., Okazaki, M., & Matsuo, Y. (2010). Earthquake shakes Twitter users:
Real-time event detection by social sensors. Proceedings of the 19th Inter-
national Conference on the World Wide Web. Raleigh, NC.
Schoenewolf, G. (1990). Emotional contagion: Behavioral induction in individuals
and groups. Modern Psychoanalysis,15(1), 49–61.
Shalizi, C. R., & Thomas, A. C. (2011). Homophily and contagion are generically
confounded in observational social network studies. Sociological Methods
& Research,40(2), 211–239.
Singer, T., & Lamm, C. (2009). The social neuroscience of empathy. Annals of the
New York Academy of Sciences,1156(1), 81–96.
Smailovi´
c, J., Grˇ
car, M., Lavraˇ
c, N., & ˇ
Znidarˇ
siˇ
c, M. (2014). Stream-based active
learning for sentiment analysis in the financial domain. Information Sciences,
285(32), 181–203.
Sprenger, T. O., Tumasjan, A., Sandner, P. G., & Welpe, I. M. (2014). Tweets and
trades: The information content of stock microblogs. European Financial
Management,20(5), 926–957.
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in
the stock market. The Journal of Finance,62(3), 1139–1168.
Tetlock, P. C., Saar-Tsechanksky, M., & Macskassy, S. (2008). More than words:
Quantifying language to measure firms’ fundamentals. The Journal of Fi-
nance,63(3), 1437–1467.
Tofallis, C. (2015). A better measure of relative prediction accuracy for model
selection and model estimation, Journal of the Operational Research Society,
66(8), 1352–1362.
trade-IQ. (2011). The U.S. active trader market; Report Preview, accessed De-
cember 8, 2015, available at http://tradeiq.blogspot.com/2012/02/us-active-
trader-market-report-preview.html.
Tumasjan, A., Sprenger, T. O., Sandner, P. G., & Welpe, I. M., (2010). Predicting
elections with twitter: What 140 characters reveal about political sentiment.
Proceedings of the Fourth International AAAI Conference on Weblogs and
Social Media, Washington, DC.
Wiebe, J., Wilson, T., & Cardie, C. (2005). Annotating expressions of opinions
and emotions in language. Language Resources and Evaluation,39(2-3),
165–210.
Xie, Y., Chen, Z., Zhang, L. K., Cheng, Y., Honbo, D., & Agrawal, A. (2015).et al.
MuSES: Multilingual sentiment elictitation system for social media data,
IEEE Intelligent Systems,29(4), 34–42.

Sul, Dennis, and Yuan 31
Appendix
Table A1: Regression results of emotional sentiment on abnormal returns by
number of followers using a breakpoint of 1,000 followers (a 96–4% split). The
SMAPE values for the controls-only models are 1.9182, 1.8694, and 1.8116.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.193 0.428 0.088
Sentiment Under i,t 0.384** 1.248** 1.871**
Control 1 i,t-1 0.006 0.022 0.007
Control 2i,t-30, t-20.002*0.016*** 0.027***
Surprise i,t 0.010*** 0.006 0.005
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.002
Intercept 0.000*** –0.003*** –0.006***
SMAPE 1.9146 1.8594 1.8113
Adj. R20.002 0.001 0.001
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.164 0.114 –0.625
Sentiment Under i,t 0.301** 1.145*** 1.988***
Control 1 i,t-10.006 0.022 0.007
Control 2i,t-30, t-20.002*0.016*** 0.027***
Surprise i,t 0.010*** 0.006 0.005
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002** 0.000 –0.002
Intercept 0.000*** –0.003*** –0.006***
SMAPE 1.9142 1.8589 1.8107
Adj. R20.002 0.001 0.001
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t 1.885 –4.643 –1.104
Sentiment Under i,t –2.922 –20.715** –14.126***
Control 1 i,t-1 0.006 0.023*0.026***
Control 2i,t-30, t-20.002*0.016*** 0.007
Surprise i,t 0.010*** 0.006 0.005
Upgrade i,t 0.002*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.002
Intercept 0.000*–0.002*** –0.004***
SMAPE 1.9170 1.8597 1.8112
Adj. R20.002 0.001 0.001
Notes: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.

32 Trading on Twitter
Table A2: Regression results of emotional sentiment and retweeting on abnormal
returns by number of followers using a breakpoint of 1,000 followers (a 96–4%
split). The SMAPE values for the controls-only models are 1.8577, 1.7660, and
1.7168, respectively.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t –1.021 2.814 0.672
Sentiment OverRe i,t 0.146 –0.481 0.309
Sentiment UnderNo i,t 1.185*1.240 3.737
Sentiment UnderRe i,t 0.526 0.998 0.922
Control 1 i,t-1–0.003 0.058*0.037
Control 2i,t-30 ,t-20.008*** 0.047*** 0.082***
Surprise i,t 0.007*0.001 –0.005
Upgrade i,t 0.002 –0.002 –0.005
Downgrade i,t –0.002*** 0.003 0.002
Intercept –0.001** –0.005*** –0.008***
SMAPE 0.3253 .3145 0.3158
Adj. R20.004 0.009 0.012
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t –0.655 1.079 –0.745
Sentiment OverRe i,t –0.165 –1.127 –0.604
Sentiment UnderNo i,t 0.896*1.444 3.643*
Sentiment UnderRe i,t 0.384 1.186 0.856
Control 1 i,t-1–0.003 0.059*0.038
Control 2i,t-30 ,t-20.008*** 0.047*** 0.083***
Surprise i,t 0.008*0.001 –0.005
Upgrade i,t 0.002 –0.002 –0.005
Downgrade i,t –0.002*** 0.003 0.002
Intercept –0.001** –0.005*** –0.008
SMAPE 0.3252 0.3144 0.3156
Adj. R20.004 0.009 0.012
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment OverNo i,t 6.373 –59.201*–39.345
Sentiment OverRe i,t 4.147 32.940 29.771
Sentiment UnderNo i,t –13.255 –27.298 –65.048
Sentiment UnderRe i,t –8.617 –21.707 –22.808
Control 1 i,t-1–0.003 0.059*0.037
Control 2i,t-30 ,t-20.008*** 0.046*** 0.082***
Surprise i,t 0.008*0.001 -0.005
Upgrade i,t 0.002 –0.002 –0.005
Downgrade i,t –0.002*** 0.003 0.002
Intercept 0.000 0.000 –0.002
SMAPE 0.3257 0.3145 0.3157
Adj. R20.004 0.009 0.012
Notes: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.

Sul, Dennis, and Yuan 33
Table A3: Regression results of emotional sentiment on abnormal returns using
only those with more than 171 followers. The SMAPE values for the controls-only
models are .9678, .9399, and .9216.
(a) Pos1 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.056 0.832** 0.952*
Control 1 i,t-1 0.002 0.011 -0.014
Control 2i,t-30 ,t-20.001 0.007*** 0.013***
Surprise i,t 0.012*** 0.012*0.014
Upgrade i,t 0.002*** 0.001 –0.001
Downgrade i,t –0.002*** 0.001 –0.004**
Intercept 0.000*** –0.003*** –0.006***
SMAPE 0.9677 0.9399 0.9216
Adj. R20.001 0.000 0.001
(b) Pos2 Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t –0.041 0.587*0.395
Control 1 i,t-10.002 0.011 –0.010
Control 2i,t-30 ,t-20.001 0.012*** 0.013***
Surprise i,t 0.012*** 0.004 0.014
Upgrade i,t 0.002*** 0.000 -0.001
Downgrade i,t –0.002*** 0.000 –0.004***
Intercept 0.000*** –0.003*** –0.004***
SMAPE 0.9677 0.9399 0.9216
Adj. R20.001 0.000 0.001
(c) Neg Next Day Next-to-10th-Day Next-to-20th-Day
Sentiment Over i,t 0.149 –12.808** –9.950
Control 1 i,t-10.001 0.018 –0.013
Control 2i,t-30 ,t-20.001 0.013*** 0.013***
Surprise i,t 0.012*** 0.004 0.014
Upgrade i,t 0.001*** 0.000 –0.001
Downgrade i,t –0.002*** 0.000 –0.003
Intercept 0.000*** –0.001** –0.003**
SMAPE 0.9678 0.9399 0.9216
Adj. R20.001 0.001 0.001
Note: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.

34 Trading on Twitter
Table A4: Regression results of emotional sentiment on abnormal returns by
number of followers for the same trading day. The SMAPE values for the controls-
only model are 1.7935 and .9286, respectively.
(a) Pos1 Both Over and Under Over Only
Sentiment Overi,t 1.635*** 1.135***
Sentiment Under i,t 0.508**
Control 1 i,t-1–0.003 –0.003
Control 2i,t-30 ,t-20.001 0.001
Surprise i,t 0.027*** 0.022***
Upgrade i,t 0.011*** 0.010***
Downgrade i,t –0.017*** –0.015***
Intercept 0.000*** 0.000***
SMAPE 1.7409 0.9018
Adj. R20.036 0.027
(b) Pos2 Both Over and Under Over Only
Sentiment Over i,t 1.627*** 1.318***
Sentiment Under i,t 0.506***
Control 1 i,t-1–0.004 –0.004
Control 2i,t-30 ,t-20.001 0.001
Surprise i,t 0.027*** 0.022***
Upgrade i,t 0.011*** 0.009***
Downgrade i,t –0.017*** –0.015***
Intercept 0.000*** 0.000***
SMAPE 1.7206 0.8910
Adj. R20.038 0.028
(c) Neg Both Over and Under Over Only
Sentiment Over i,t –23.415*** –16.503***
Sentiment Under i,t –7.596**
Control 1 i,t-1–0.002 –0.003
Control 2i,t-30 ,t-20.001 0.001
Surprise i,t 0.027*** 0.023***
Upgrade i,t 0.011*** 0.009***
Downgrade i,t –0.017*** –0.015***
Intercept 0.000*** 0.000***
SMAPE 1.7495 0.9086
Adj. R20.035 0.027
Notes: The coefficients are multiplied by 1,000.
*pࣘ.05, ** pࣘ.01, *** pࣘ.001.

Sul, Dennis, and Yuan 35
Hong Kee Sul is a research support director at Wharton Research Data Services,
The Wharton School, University of Pennsylvania. In 2015, he graduated with a PhD
in Finance from Indiana University, Kelley School of Business. Dr. Sul received
a Master of Science degree from Seoul National University and a Bachelor of
Science degree from Korean Advanced Institute of Science and Technology. His
work has been presented in several conferences including the 2016 American
Economic Association Annual Meetings, 2014 Financial Management Association
Annual Meeting Doctoral Consortium, and the 47th Annual Hawaii International
Conference on System Sciences.
Alan Dennis is a professor of information systems and holds the John T. Chambers
Chair of Internet Systems in the Kelley School of Business at Indiana University. He
is a Fellow of the AIS and has written more than 150 research papers. His research
focuses on three main themes: the use of computer technologies to support team
creativity and decision making; IS for the subconscious; and the use of the Internet
to improve education. He is the AIS Vice President of Conferences, the editor in
Chief of Foundations and Trends in Information Systems, and co-editor in Chief
of AIS Transactions on Replication Research. Prof. Dennis also has written four
books, two on data communications and networking, and two on systems analysis
and design. His most recent start-up company is NameInsights.com, which uses
big data and analytics to help parents pick baby names.
Lingyao (Ivy) Yuan is an assistant professor of information systems of College
of Business at Iowa State University. Her research interests include on the impact
of noncognition behavior and decision making, especially the impact of emotion,
on computer-mediated communication, decision making, and collaboration. She
has conducted research in the fields of electronic commerce and social media.
She has been published in the Group Decision and Negotiation as well as several
conferences including 47th Annual Hawaii International Conference on System
Sciences,and 2013 INFOMRS Annual Meeting. She received a Master of Science in
Information Technology from University of North Carolina Charlotte in 2011 and
Bachelors of Management Information Systems from University of International
Business and Economics in 2009.