Additional Dissemination of Corporate
Earnings Information via Twitter and
Investor Trading Behavior -
Evidence from Dutch listed companies
University of Groningen
MSc. Accountancy
Master Thesis
Abstract: This research focuses on the additional dissemination of earnings information via Twitter, which is a new dimension of corporate information dissemination. It investigates whether additional dissemination of earnings information via Twitter influences investors’ attention and if so, how stock prices respond to this additional information dissemination. Furthermore, this research incorporates the possibility of noise making by a firm on its Twitter timeline around the earnings announcement. The focus of this research is on tweets referring to press releases announcing both quarterly and yearly earnings, in order to examine the effect of additional dissemination via Twitter exclusively. Using regression analysis on Dutch listed companies in the period from March 2007 till August 2015, it is found that there is no statistically significant differential reaction in immediate stock price response between Twitter-disseminated earnings announcements and non-Twitter-disseminated earnings announcements. Moreover, it is also found that the act of noise making of a firm on its Twitter timeline does not influence investors’ attention. Therefore, it can be concluded that the additional dissemination of earnings information via Twitter cannot be considered to influence investors’ attention. The results may indicate that the Twitter audience do not include sufficient (potential) investors, or that the investors within the sample of this study are mostly institutional investors, instead of individual investors.
Student: Eline Brinkhuis Student Number: 1977601
Supervisor: Yasemin Karaibrahimoglu Second assessor: Reggy Hooghiemstra Date: January 25, 2016
2 TABLE OF CONTENTS
I. Introduction ... 5
II. Background and Related Literature ... 8
A. Inefficient Markets ... 8
B. Dissemination Literature ... 8
C. Limited Attention and Information Processing ... 10
D. Direct Access Information Technologies ... 11
E. Social Media and Financial Markets... 13
III. Hypothesis Development ... 14
A. Investor Attention and Twitter Dissemination ... 14
B. Noise Making ... 15
IV. Research Design ... 16
A. Data and Sample ... 16
B. Summary Statistics ... 19
C. Methodology ... 23
Immediate Stock Price Response ... 24
Noise Making ... 25
V. Results ... 26
A. Graphical Evidence ... 26
Immediate Stock Price Response ... 26
3
B. Regressions ... 28
Immediate Stock Price Response ... 28
Noise Making ... 30
VI. Additional Analyses ... 33
A. Delayed Stock Price Response ... 33
B. Volume Response ... 34
VII. Robustness Analyses ... 35
A. Beginning Year 2010 ... 35
B. Handling Outliers Differently ... 36
C. Noise Making as Continuous Variable ... 36
D. Methodology Sensitivity ... 37
VIII. Conclusion and Discussion ... 37
References ... 42
Appendices ... 47
Appendix A - Pre-Announcement Stock Return ... 47
Appendix B - Correlation Matrix Firm Characteristics ... 48
Appendix C – Immediate Stock Price Responsiveness ... 49
Appendix D - Graphical Evidence Noise Making ... 50
Appendix E – Additional Analysis Delayed Stock Price Response (Graphical Evidence)51 Appendix F – Additional Analysis Volume Response (Graphical Evidence) ... 52
Appendix G – Additional Analysis Delayed Stock Price Response (Regressions) ... 53
Appendix H – Additional Analysis Volume Response (Regressions) ... 54
4
Appendix J – Robustness Analysis, winsorized at 10% tails ... 56
Appendix K – Robustness Analysis, winsorized at 1% tails ... 57
Appendix L – Robustness Analysis, outliers dropped... 58
5 I. INTRODUCTION
Do the way and scope of corporate information dissemination matter? The answer to this question is still unclear for capital markets. In accordance with the efficient market hypothesis, it has long been assumed that investors’ access to publicly disclosed corporate information is unlimited and that the way this information is distributed by firms does not matter. An efficient capital market is defined as one in which security prices fully reflect all available information (Fama, 1970). In other words, once information is publicly disclosed by a firm, it is assumed to be available to all investors in the market. Under efficient market conditions, there is therefore no reason for firms to spend resources to promote a news story once it has been disclosed (Solomon, 2012). Nevertheless, in practice prices do not respond instantaneously to news, since acquiring and processing information is costly and time-consuming to investors (Solomon, 2012; DellaVigna and Pollet, 2009). Hirshleifer and Teoh (2003) characterize investors as imperfect processors of publicly available information. Even in today’s digital world, where all information is a ‘click of the mouse’ away, finding information is time-consuming (Dyck and Zingales, 2003). Recent research shows that, since investors’ attention and resources are limited, variation in the dissemination of firm-initiated disclosures may have different economic consequences (Blankespoor et al., 2014; Soltes, 2010; DellaVigna and Pollet, 2009; Hirshleifer and Teoh, 2003). Particularly, it is found that security prices do not reflect new earnings information immediately once the information is disseminated (Bhagwat and Burch, 2014; Curtis et al., 2014). Additionally, improvements in technology and the rise of new information technologies facilitating digital communication, mostly on the Internet, create new channels for firms to disseminate information. One of these new channels for information dissemination, the social media platform Twitter, together with the notion of investors’ inattention and their limited time and resources form the basis of this study.
Particularly, this research focuses on Twitter and its ability to improve information dissemination. The way Twitter is organized, allows firms to bypass information intermediaries and contact investors directly on a frequent and real-time basis, which ultimately reduces investors’ information acquisition costs (Blankespoor et al., 2014). Specifically, this research investigates whether earnings information which is additionally disseminated via Twitter influences investors’ attention and if so, how stock prices respond to this additional information dissemination. This study focuses on the use of Twitter, i.e. tweets posted by the firm itself, surrounding earnings announcements for Dutch listed companies.
In the Netherlands, the Dutch Authority for Financial Markets (AFM) is responsible for supervising the operation of the financial markets. The AFM requires all listed companies to make sensitive information (e.g. an earnings announcement) publicly available by issuing a press release. In order to reach a broader set of stakeholders, and thus improve information dissemination, firms may decide to use additional
6 communication channels. For example, in the case of an earnings announcement, firms may complement their traditional dissemination channels by pushing their earnings information to investors via Twitter postings (so-called tweets). This research focuses on tweets referring to the official earnings announcement posted on the earnings announcement date by the official corporate Twitter account. Focusing on tweets referring to already made earnings announcements allows to examine the impact of broader dissemination exclusively, because tweets associated to earnings announcements do not provide new information to the market (Blankespoor et al., 2014; Dyck and Zingales, 2003).
Earlier research shows that improved dissemination of information and public disclosures impact both market liquidity and information asymmetry, as well as stock prices (Blankespoor et al., 2014; Soltes, 2010; Bushee et al., 2010; Easley and O’Hara, 2004; Dyck and Zingales; 2003; Diamond and Verrecchia, 1991). The results of earlier performed research thus indicate a link between a firm’s information structure and both its market liquidity and stock price, indicating that firms can (partly) influence these factors by affecting the way they disseminate information to investors. This research further investigates this link by examining the relationship between the additional dissemination of earnings information of Dutch listed firms ‘tweeting’ their earnings announcements and investors’ trading behavior.
The additional dissemination of earnings information via Twitter is expected to increase investors’ attention to earnings news and ultimately influence stock price returns. The relationship between investors’ attention and stock price responses is established by DellaVigna and Pollet (2009) and forms the starting point of this research. In line with DellaVigna and Pollet (2009) this study examines investors’ attention to earnings surprises. Firms’ earnings announcements disseminated additionally via Twitter are compared to those not disseminated via Twitter. If additional Twitter dissemination indeed increases investors’ attention to the earnings announcement, the immediate response to the non-Twitter-disseminated earnings announcements should be less pronounced than to the Twitter-non-Twitter-disseminated announcements. The effect of the additional dissemination of earnings information via Twitter on the immediate stock price response forms the main analysis of this study.
Since investors will revisit their decisions in the following period, the earlier neglected information should be eventually incorporated into the stock prices and thus the delayed response to the earnings announcements should be larger for non-Twitter-disseminated announcements. This delayed stock price response is analyzed additionally to the immediate stock price response. Moreover, if these stock price responses are indeed a consequence of varying levels of investors’ attention, these stock price responses must be accompanied by a higher immediate trading volume response to Twitter-disseminated announcements compared to non-Twitter disseminated announcements. The effect of additional Twitter dissemination on the immediate volume response is analyzed as a second additional analysis.
7 This research incorporates the possibility of ‘noise making’ on the Twitter timeline by the firm itself around earnings announcements. As DeHaan et al. (2015) indicate, firms’ managers may have incentives to vary the level of attention of investors to either good or bad news in order to influence the market price response to the news. While focusing on the corporate use of Twitter around earnings announcements, it may be expected that managers try to hide or highlight particular earnings information by playing with the number of tweets posted within a time period around the announcement, in order to attract or distract investors’ attention. As this noise making indeed appears to influence investors’ attention, the immediate stock price responses is expected to be larger for noise making firms than for non-noise making firms.
While there is a vast amount of literature regarding firm disclosures, media and information dissemination in general (Engelberg and Parsons, 2011; Beyer et al., 2010; Soltes, 2010; Fang and Peress, 2009; Hirshleifer and Teoh, 2003; Chan, 2003; Diamond and Verrecchia, 1991 amongst others), studies examining the effects of the corporate use of social media are scarce. Although some papers link Twitter to the financial market by analyzing the use of Twitter in predicting stock market prices and other stock market indicators (Chen et al., 2014; Mao et al., 2012; Bollen et al., 2011; Zhang et al., 2011), analyzing Twitter volume spikes (Wei et al., 2015; Mao et al., 2013), or more generally by investigating its potentials for corporate use (Gruber et al., 2015; Rybalko and Seltzer, 2010; Krishnamurthy et al., 2008), the additional dissemination power of Twitter remains more or less neglected. By investigating the use of Twitter in disseminating information, and by principally analyzing its effect on investor behavior, this research aims to contribute to the existing literature. In particular, it will be the first analyzing the use of social media by Dutch listed firms. This is remarkable, since Dutch firms are leading in the use of social media; in 2014, 53% of the Dutch firms used social media and Dutch firms are most active on Twitter in the European Union (CBS, 2015).
Mainly, this research contributes to the dissemination literature by analyzing the dissemination power of Twitter around earnings information, and the dissemination power of newly developed (more technology driven) information channels in general. By linking additional Twitter dissemination to investors’ attention, this research also adds to the investor attention literature by analyzing investors’ inattention from an information dissemination perspective. Finally, considering that hundreds of investment strategies are based on investor attention theories, by highlighting the possibilities of noise making and analyzing its differences in effects on the immediate stock price reaction, this research will provide insights in the strategies of managers’ earnings information provision. Thus, if a link between the dissemination of earnings information via Twitter and investors’ inattention can be established, this research is of practical importance as well.
8 The remainder of this research is structured as follows: section II provides background information on the main issues of this research, including information technologies and Twitter, as well as a review of the related literature. The hypothesis development follows in section III. Section IV introduces the data and explains the used methodology. Section V gives an overview of the results of this study and the additional analyses and several performed robustness checks are presented in section VI and VII respectively. Section VIII provides the conclusion and discussion of this study, followed by the references and appendices.
II. BACKGROUND AND RELATED RESEARCH
A. Inefficient Markets
Traditional financial theory is based on the efficient market hypothesis, which suggests that rational utility-maximizing investors consider all publicly available information about a firm when buying or selling a firm’s stock, while market forces eliminate the less rational investors (Fama, 1965; Friedman, 1953). Since under the efficient market hypothesis publicly disclosed information is assumed to be available to all investors in the market, there is no reason for firms to spend resources to promote their news stories once they have been disclosed (Solomon, 2012). In other words, efficiency implies that not the form of disclosure itself, but the information content of disclosure is valued by the market (Scott, 2012). In current academic literature, the efficient market hypothesis is challenged and called into question more and more. Important here are the notions of under reaction and overreaction of stock prices to news indicating that news is incorporated only slowly into stock prices but meanwhile stock prices tend be overpriced after good news and have low average returns afterwards (Barberis et al., 1998). An often mentioned anomaly to efficient markets, which is closely related to the notion of under reaction and therefore related to this research, is the post-earnings announcement drift. This post-earnings announcement drift indicates the tendency of stock returns to continue to drift up for “good news” firms and down for “bad news” firms even after earnings are announced (Ball and Brown, 1968). Thus, this research is based on the notion of inefficient markets.
B. Dissemination Literature
Considering the market inefficiency, it is expected that the way and the scope of corporate information dissemination matter. Easley and O’Hara (2004) acknowledge this, by arguing that the quality of information affects asset pricing, and thus the way information is provided to the markets is clearly important. Several empirical papers investigated the effects of improved or broadened information dissemination.
9 First, Singhvi and Desai (1971) argue that inadequate corporate disclosure in annual reports will widen fluctuations in the market price of a security, since in the absence of adequate information investment decisions are based on less objective measures. Diamond and Verrecchia (1991) show that revealing public information to reduce information asymmetry can reduce the cost of capital of a firm, since increased demand from large investors will be attracted due to the increased liquidity of the firm’s securities. Klibanoff et al. (1998) find that news event lead some investors to react more quickly.
In the existing literature on improved dissemination, the role of the media in financial markets is analyzed regularly. Chan (2003) investigated the relationship between news coverage and momentum, supporting the notion of under- and overreaction to news. Similarly, Antweiler and Frank (2006) report a short-term momentum for many days after the publication of a news story in the Wall Street Journal as well as a longer term reversal for many events. In contradiction, Busse and Green (2002) find that stock prices respond to segments reports within seconds of the initial mention and that positive reports are even fully incorporated within one minute.
Fang and Peress (2009) find that firms with high media coverage earn lower stock returns than firms with no media coverage. They conclude that their found media effect can be attributed to the ‘investor recognition hypothesis’ of Merton (1987). This hypothesis reasons that stocks with lower investor recognition need to offer higher returns to compensate investors for being imperfectly diversified. In other words, investors’ attention lowers the cost of capital. Fang and Peress (2009) suggest that mass media outlets such as newspapers play an important role in the dissemination of information to a broad audience, especially to individual investors and that the breadth of information dissemination affects stock returns. Moreover, Bushee et al. (2010) find that additional press coverage around a firm’s earnings announcement is associated with smaller bid-ask spreads and improvements in depth. Their finding let them argue that dissemination of earnings news is more important than the production of information by journalists. Similarly, by using an instrumental variables approach based on the amount of competing news, Soltes (2010) finds that improved dissemination of information causally lowers bid-ask spreads, increases trading volume and lowers idiosyncratic volatility. Campbell et al. (2012) examine the role of news media during financial bubbles and conclude that media reporting of recent events may have influenced asset prices, and the main contribution of the media was to provide factual information to investors which could be used in making investment decisions. Finally, Engelberg and Parsons (2011) find that local trading is strongly related to whether the local newspaper covers the earnings announcement made by the firm. By comparing the behaviors of investors with access to different media coverage of the same information event, their study represents the first systematic identification of the causal effect of the media on investor behavior.
10 C. Limited Attention and Information Processing
As mentioned before, investors can be characterized as imperfect processors of publicly available information since acquiring and processing information is costly and time-consuming. As a consequence of their limited time and resources, investors tend to rely only on a few sources to acquire information (Blankespoor et al., 2014; Hirshleifer and Teoh, 2003; Hong and Stein, 1999; Merton, 1987). In line with Hirshleifer and Teoh (2003) this research starts from the assumption that investors have limited attention and information processing power. Initially, inattention among investors seems foolish, as inattentive investors lose money by ignoring aspects of the economic environment. However, Hirshleifer and Teoh (2003) argue that, in case time and attention are costly to investors, inattentive investor behavior may be reasonable.
The (in)attention of investors has been the subject of study of several existing empirical research. This stream of empirical research can be grouped into the field of behavioral finance (Orlitzky, 2013). These behavioral theories departure from the traditional assumptions of market efficiency, investor rationality and unlimited computational capacity of investors (Orlitzky, 2013; Hong and Stein, 1999). Orlitzky (2013) points out that behavioral finance starts from the assumption that investors do not only trade on accurate information about fundamentals, but also on their unsubstantiated beliefs.
DeHaan et al. (2015) refer to market inattention or distraction as the limited capabilities of humans to acquire and process information, which prevent them from absorbing the complete set of public information. Hirshleifer and Teoh (2003) examine the issue of investors’ inattention extensively. They argue that attention is required to both encode environmental stimuli and process ideas in conscious thought. They also propose that attention must be selective and requires effort, i.e. the substitution of cognitive resources from other tasks as in Kahneman (1973). Moreover, Hirshleifer and Teoh (2003) argue that limited investors’ attention is a necessary consequence of the vast amount of available information and that it limits the information processing power of investors. This problem of limited investors’ attention as a consequence of information overload and limited resources is acknowledged by Lambert (2003) and DeHaan (2015) amongst others. DeHaan (2015) goes a step further by putting forward that investors’ inattention could even be a rational choice of investors when they are subject to extremely high information search costs.
Finally, an important consequence of limited attention is highlighted by Hirshleifer and Teoh (2003), namely that in case of limited investors’ attention disclosures with the same information content can have different effects on investors’ perceptions, dependent upon the form of the presentation. This provides an indication that limited investors’ attention conflicts with the efficient market hypothesis and thus stock
11 prices may not fully reflect new earnings information immediately after the information is disseminated, which is also supported by Bhagwat and Burch (2014). Moreover, the observation made by Hirshleifer and Teoh (2003) that observed limited investors’ attention may have an influence on investors’ behavior and asset prices in actual capital markets is reported by other researchers as well (e.g. Andrei and Hasler, 2015; Drake et al., 2012; Da et al., 2011). Huberman and Regev (2001) also conclude that inattention is associated with prices that do not fully reflect all available public information, after they report that enthusiastic public attention related to their event of study caused a permanent rise in stock prices, even though no genuinely new information was revealed to the market. In other words, the result of Huberman and Regev (2001) shows that only when investors pay attention to the new published information, prices react to it.
In the same way, Barber and Odean (2008) call investors’ attention a scarce resource. When an investor is confronted with numerous available alternatives and does not have the capacity to pay attention to all information in the environment, the option that attracts the most attention, i.e. has the strongest stimuli, is more likely to be chosen, which in turn will drive up the return of this attention-grabbing stock (Guo et al., 2015; Lou, 2014; Barber and Odean, 2008; Grullon et al., 2004; Odean, 1998). This reasoning is in line with the recognition hypothesis made by Merton (1987).
D. Direct Access Information Technologies
As Busse and Green (2002) already pointed out, there have been some fundamental changes in securities markets since the early studies of market efficiency mentioned before. Financial information is more readily available to all participants in the market, trading costs decreased significantly and technology improvements have accelerated the pace at which markets operate. In the light of this research, the technological developments in information technologies play an important role.
When a firm discloses a piece of information, a mechanism must exist to transmit or disseminate this news to investors. Without such a transmission mechanism, investors would be unable to access a firm’s disclosures (Soltes, 2010). Traditionally, firms had to develop these transmission mechanisms internally (e.g. investor-relations website), or externally in the form of third-party intermediaries (Soltes, 2010). Currently, new information technologies offer firms a third possibility by disseminating news directly to investors. Blankespoor et al. (2014) refer to these technologies as direct-access information technologies (DAITs), as they allow firms to directly access investors.
An example of a DAIT is the social network Twitter. With over 300 million monthly active users and 500 million tweets sent every day, Twitter is a real-time information network where people can discover what
12 is happening in the world, share information instantly and connect with people and business around the globe (Twitter, 2015). The use of Twitter for business purposes, e.g. announcements of earnings information, is significant (Jung et al., 2015; Bhagwat and Burch, 2014; Rybalko and Seltzer, 2010). Alexander and Gentry (2014) characterize Twitter as one of the most powerful social media platforms through which organizations communicate with their stakeholders.
An important feature of Twitter, as DAITs in general, is their use of a “push” technology instead of a “pull” technology. This push technology allows the sender (here: the firm) to transmit information to the user (here: the investor) rather than requiring the user to request the information from the sender, which consequently reduces the cost of acquiring information for investors (Blankespoor et al., 2014). The ability of smaller social networks like Twitter to deliver data to interested users over multiple possible delivery channels (smartphones, emails etc.), as pointed out by Kirshnamurthy et al. (2008), further implies that Twitter may improve information dissemination and decrease information acquisition costs.
Before the rise of the direct-access information technologies like Twitter, firms had to rely on information intermediaries, such as the press, to disseminate firm-initiated information (Bushee and Miller, 2012; Soltes, 2010). However, earlier research shows that the press is biased in covering certain types of firms (Miller, 2006; Mullainathan and Shleifer, 2005; Dyck and Zingales, 2003). Thus, relying on the press to disseminate firm-initiated information, information may not always reach investors efficiently. DAITs, like Twitter, offer the possibility to serve as a complementary information dissemination channel which bypasses these traditional information intermediaries and their related constraints and allows firms to directly reach investors on a frequent and real-time basis (Blankespoor et al., 2014).
Particularly, DAITs can be used by managers to further disseminate firm information, e.g. by tweeting links to press releases, to reach investors who would not have received the information via the traditional news sources. Ultimately, an investors’ information acquisition costs, which Blankespoor et al. (2014) describe as the investors’ time and energy spent on sifting through various sources to search and gather (i.e. “pull”) the desired information as well as any price paid to actually access this information, will decrease. Equivalently, Sprenger et al. (2014) conclude that in creating new information, compiling existing information and disseminating this information, Twitter can be considered an information intermediary. Curtis et al. (2014) acknowledge that social media allows investors to become aware of new information, thereby providing widespread dissemination of new information in extending the reach and spread of information through word of mouth. Gruber et al. (2015) denote that social media in general has increased both the speed at which information is shared and the reach of messages. Particularly in the case of Twitter, they argue that its flat hierarchy and ease of use allows information to disseminate faster via Twitter than via traditional media channels since there is no real filter. Thus, regarding investors’
13 inattention and corporate information dissemination, earlier research indicates that Twitter may offer the possibility for business to reach a greater amount of investors and that it may improve the dissemination of corporate information.
E. Social Media and Financial Markets
Since the rise of social media has occurred recently, the literature on the actual role of social media within financial markets is scarce. Only few studies relate Twitter to financial markets and information dissemination and are therefore closely related to this research.
Sprenger et al. (2014) relate Twitter to two theories on the effect of media coverage on investors’ behavior in financial markets: the information view and the salience view (Solomon et al., 2014). The information view predicts that media coverage reduces the acquisition costs of information, thereby allowing investors to improve their decision making. On the other hand, according to the salience view media coverage directs investors’ attention to that particular financial asset, ultimately increasing the asset’s demand. Sprenger et al. (2014) argue that the role Twitter plays in financial markets can be according to both the information and the salience view, i.e. Twitter could both inform investors and raise awareness depending on factors like market coverage and trading volume. Particularly, after applying computational linguistics to their comprehensive dataset of S&P 500 stock-related tweets, the event study of Sprenger et al. (2014) shows that returns prior to good news events are more pronounced than for bad news events.
Focusing on the information view, Blankespoor et al. (2014) investigate whether firms can reduce information asymmetry by more broadly disseminating their news via Twitter. For a sample of technology firms they analyze the impact of the use of Twitter in sending market participants hyperlinks to traditional press releases. They find that the dissemination of news via Twitter is associated with lower abnormal bid-ask spreads and greater abnormal depths, which is consistent with decreased information asymmetry and higher market liquidity. They conclude that managers are thus able to increase market liquidity of their firm by disseminating firm-initiated information via Twitter. By focusing on hyperlinked tweets, they constructed a unique way to isolate the impact of broader information dissemination, which made them able to conclude that in addition to the impact of the information itself, broader dissemination of that information can have real market consequences. The generalization of their research is however limited since their sample consists of technology firms only.
Conversely, Bhagwat and Burch (2014) focus on the salience view and examine how a firm’s behavior on Twitter affects earnings-news returns. They show that in case a firm tweets financial information more frequently around the announcement of earnings results, the magnitude of earnings announcement returns
14 increases. This is particularly the case when earnings surprises are small and positive, and when the firm is less visible, as measured by firm size or analyst coverage. In line with the salience view on media coverage, Bhagwat and Burch (2014) conclude that Twitter provides firms with an effective and strategic way to mitigate the limited attention of investors to news, especially when the news would otherwise have been less likely to attract attention.
More specifically, Curtis et al. (2014) use social media activity as a proxy to measure investors’ attention. They find that abnormal high levels of investors’ attention are associated with significantly higher sensitivity of market returns to earnings news. Though the effect is significant for both positive and negative earnings news, the effect for firms beating the forecast is much stronger. Finally, they conclude that investors’ attention is associated with an increase in the market responsiveness to earnings news, but also that a lack of investors’ attention is associated with an under reaction to earnings news (i.e. post-earnings announcement drift). Chawla et al. (2015) track the diffusion of retail attention to specific financial news in real time by monitoring how this news is retweeted on Twitter. A strong link between retweet speed and trading intensity is found, showing that quickly retweeted tweets are correlated with immediate trading. Moreover, a positive relationship between attention diffusion and stock returns is found. Specifically, the more users a tweet reaches after three hours, the higher the stock return on that day. However, these higher returns are completely reverted on the next day.
III. HYPOTHESIS DEVELOPMENT
A. Investor Attention and Twitter Dissemination
This research is based on the salience view as discussed by both Solomon et al. (2014) and Sprenger et al. (2014), predicting that more (social) media coverage will directly attract investors’ attention. Implicitly, this research thus assumes that, even in today’s digital world, investors still have limited attention and limited information processing power. This assumption is in line with Bhagwat and Burch (2014) and Blankespoor et al. (2014) amongst others. Moreover, as Bhagwat and Burch (2014) assumed, this research relies on the assumption that active firms’ Twitter accounts are tracked by a reasonable number of followers, which implies that earnings tweets are seen by (potential) investors. To justify, the mean number of followers per firm in the sample of this study is 16,259.
In summary, based on the discussed literature it is expected that the corporate use of Twitter around firms’ earnings announcements influences investors’ attention. Subsequently, the level of investors’ attention is expected to influence a firm’s stock returns and trading volumes (Hirshleifer and Teoh, 2003). DellaVigna and Pollet (2009) show that investors’ inattention influences stock prices, by showing that inattention is
15 associated with a less pronounced immediate response to earnings information, followed by a larger delayed response. Moreover, they find that this effect in stock returns is accompanied by lower trading volume around the announcement. To establish a link between the additional dissemination of earnings information via Twitter and investors’ attention, this research focuses on the immediate stock price response specifically.
Thus, if the additional dissemination of earnings information via Twitter indeed influences investors’ attention, similar effects as found in DellaVigna and Pollet (2009) are expected within this study. When earnings information is more broadly disseminated additionally via Twitter, the time and resources investors need to spend to acquire and process the information are expected to decrease, resulting in a higher immediate response in stock returns to the earnings announcement. Vice versa, if earnings information is less disseminated a less pronounced immediate response is expected. This expectation is captured in hypothesis 1 as follows:
Hypothesis 1: The immediate stock price response to positive (negative) earnings surprises is more positive (more negative) for earnings announcements which are additionally disseminated via Twitter than for earnings announcements which are not additionally disseminated via Twitter.
B. Noise Making
Concerning the announcements of firms’ earnings information under the assumption of investors’ limited attention, an often proposed question is whether managers try to hide bad earnings information by announcing those numbers during periods of low investors’ attention and whether managers try to highlight good earnings information by announcing it during periods of high investors’ attention (DeHaan et al., 2015). The opportunity for managers to hide bad news is a consequence of the overload of information available to investors in the information environment (Lim and Teoh, 2010; Hirshleifer et al., 2009). DeHaan et al. (2015) argue that managers do have the incentive to hide bad news in order to reduce investors’ attention, because this might influence the market price response to the announcement made, at least in the short term.
In the light of information dissemination via Twitter, it may be expected that the incentives for managers to hide or highlight information may be even greater, since Twitter creates an additional information source and manipulating the amount of information send to investors via Twitter is easier for managers. Bhagwat and Burch (2014) indeed show that the number of tweets posted by firms vary around positive earnings announcements versus negative earnings announcement. Typically, they find that financial tweeting intensity surrounding earnings news is higher for positive news than for negative news.
16 This study will refer to these abnormal levels of tweeting intensity around earnings announcements as “noise making”. Based on the expectation that managers do have incentives to hide or highlight specific earnings announcements (DeHaan et al, 2015), there may be noise on a firm’s Twitter timeline, created by the firm itself, to attract or distract investors’ attention. Distinguishing between noise making and non-noise making firms, this study investigates whether the effects found in the main analysis differ between both groups of firms. A difference in the immediate response is expected, since noise making is expected to influence investors’ attention. Because abnormal high levels of tweeting intensity around firms’ earnings announcement are expected to attract investors’ attention to the earnings announcement, the immediate stock price response is expected to be larger for noise-making firms than for non-noise making firms. This expectation leads to the following hypothesis:
Hypothesis 2: The immediate stock price response to earnings surprises for Twitter-disseminated earnings information is larger for noise making firms than for non-noise making firms.
IV. RESEARCH DESIGN
A. Data and Sample
This study focuses on quarterly and annual earnings announcements made by Dutch listed firms in the period from March 1st, 2007 until August 4th, 2015. Dutch firms showed to be most active firms on Twitter within the European Union (CBS, 2015) and are therefore chosen as the subject of study. The first date within the dataset, March 1st, 2007, is chosen since starting from March 2007 the first firms within the sample joined Twitter (TomTom and Philips).
First, a list of Dutch listed companies is extracted from Datastream containing 126 unique firms. Additionally, data on the companies’ fundamentals, industry characteristics and historical currency rates is obtained from Datastream as well. Data regarding the earnings announcements and the earnings estimates are retrieved from the I/B/E/S database, using the company’s ticker as matching variable. More specifically, data on companies’ actual earnings (i.e. announcement date, announcement time and actual EPS) are retrieved from the Detail History File of I/B/E/S while data on companies’ earnings estimates (i.e. number of estimates and median estimate) are retrieved from the Unadjusted Summary Statistics File. Specifically, the consensus analyst forecast is used as a proxy for investors’ expectations. The consensus forecast is defined as the median forecast among all the analysts that make a forecast in the last 30 calendar days before the earnings announcement (DellaVigna and Pollet, 2009). When there are multiple consensus forecasts for a firm within a quarter, only the most recent one is kept. Using the historical daily currency rates, all prices are converted to euros. Earnings announcements which occurred during the
17 weekend or on holidays are removed from the dataset, since no immediate response can be identified because exchanges are closed. Moreover, observations with missing actual values or missing median estimates are dropped, as well as observations in which the actual value or the earnings estimate is larger in absolute value than the price of the share. Penny stocks, i.e. stock which are traded at less than $5 per share (SEC, 2013), are removed from the dataset as well.
Subsequently, Twitter data (i.e. an indicator variable for a tweet mentioning the earnings announcement and the number of tweets during specific windows) is added manually using the Twitter Advanced Search. When the same earnings information was tweeted both in English and in Dutch, the Dutch tweet is considered as main announcement tweet, since most of the time the Dutch tweets had the most retweets and interaction, indicating that those tweets are more disseminated. When no tweet was posted on an active Twitter timeline at the announcement date, the correctness of the announcement date is double-checked on the firm’s corporate website. This resulted in the removal of two announcements because of incorrectness of the I/B/E/S announcement date. When counting the number of tweets within a particular time period, tweets which were posted as a reaction to a Twitter account (characterized by starting with the @-symbol) are disregarded since they do not appear on the timeline of the followers of the particular firm’s Twitter account.
Additionally, the number of followers of a particular Twitter account, measured in November 2015, is subtracted from tweetchup.com and manually added to the dataset. Moreover, data regarding media coverage in general is subtracted from the Factiva Database. The Factiva Database gives access to all media sources, e.g. newspapers, journals and magazines, both hard-copy and digitally distributed. From the Factiva database relevant media outlets are obtained from all sources and in all languages for a particular company within a particular time period (announcement date or year of announcement). When exactly the same text appeared to be published in multiple different media outlets, the observation is counted repeatedly. Since the yearly measure of media coverage in 2015 is measured during the end of November 2015, the count of media outlets for 2015-observations is replaced by the number of media outlets up to November 2015, divided by 11 and multiplied by 12. Although this method is not very precise, it is considered the most accurate measure possible given the available data.
The final sample includes 677 unique observations for 70 firms. To be able to make a fair comparison, the complete analysis is performed twice, both for announcements announcing quarterly earnings results and announcements announcing yearly earnings results separately. The quarterly sample consists of 305 observations, while the annual sample includes 372 observations.
18 In line with DellaVigna and Pollet (2009) the earnings surprise is defined as the difference between the actual earnings per share (EPS) as announced in the earnings announcement and the consensus earnings forecast, normalized by the share price five trading days before the announcement. Thus, the earnings surprise st,i is:
st,i = et,i−êt,i
Pt,i (1)
Where et,i is the EPS announced in quarter t for company i and êt,i the corresponding consensus analyst
forecast. Pt,i is the share price of company i five trading days before the announcement in quarter t. When
the fifth day prior to the announcement date appeared to be during the weekend, the stock price on the Friday before is taken into account.
To be able to measure announcement quality, the announcements are sorted into bins. Specifically, the earnings surprises are grouped into eleven quantiles. The number eleven is chosen since this separation results in an odd number of quantiles which enables to create a ‘zero surprise quantile’ and at the same time the number eleven ensures that the very extreme surprises (positive versus negative) are pulled apart. The negative earnings surprises are in quantile 1 to 5, quantile 6 contains all observations with an (exactly) zero surprise, and the positive earnings surprises are grouped in quantiles 7 to 11. The thresholds for the bins are set separately, to guarantee an equal number of non-Tweeted earnings announcements in the positive quantiles and in the negative quantiles.
To measure the immediate stock price response, cumulative abnormal returns are constructed. Ru,i denotes
the stock return of company i on day u and Ru,m denotes the return on the market on day u. The abnormal
return (ARu,i) for company i on day u is calculated according to the market model of Brown and Warner
(1980) and is estimated using the following regression equation:
ARu,i = Ru,i− αi− βi∗ Ru,m (2)
Where αi and βi are parameters in the market model. In line with DellaVigna and Pollet (2009) the
estimation window to calculate normal returns is set at [-300;-46].
Following DellaVigna and Pollet (2009), cumulative abnormal returns (CAR) are calculated over the [0;+1] window. CARs are calculated as follows:
19 All CARs are winsorized at the 5 percent tails in the main analysis. Within the robustness analyses, different ways of handling the outliers in returns are applied. The results of the robustness checks show that handling the outliers differently do not have an effect on the main conclusions.
To analyze the difference in the immediate stock price response between noise making firms and non-noise making firms, a ‘non-noise making variable’ is constructed. Within this research, a firm which is posting a high amount of tweets during the day before the announcement, the day of the announcement and the day after the announcement (i.e. the [-1;+1] window), relative to the week before and the week after the announcement (i.e. the [-7;-2] and [+2;+7] windows) is defined as a noise making firm:
Noiset,i =
# tweets [−1;+1] window
(# tweets in [−7;−2]window+ # tweets in [+2;+7]window)/4 (4)
Thus, the average number of tweets over three days within the week before and the week after the announcement are considered as ‘normal’ tweeting intensity. The higher the tweeting intensity the days around the announcement relative to the normal tweeting intensity, the more noise making. When the sum of the tweets in the week before and the week after the announcement equals zero, this amount is replaced by 1 to be able to calculate the noise variable. Based on this noise variable, a dummy noise variable is created (Noise Dummyt,i ) which equals one when noiset,i is greater than or equal to the median noise in
the sample, and equals zero when noiset,i is smaller than the median noise in the sample. Specifically, the
median noise is used as the threshold to distinguish noise making from non-noise making firms, i.e. within the remainder of this study firms who make an above-median level of noise are considered noise making firms and firms who make an below-median level of noise are considered non-noise making firms.
B. Summary Statistics
From table 1 panel A it can be observed that only one-third of all announcements is additionally disseminated via Twitter (34%). Moreover, table 1 shows that the total sample is almost equally divided into the quarterly and the yearly announcements samples, though the percentage of tweeted announcements in the quarterly sample is higher than in the annual sample. As expected, panel B shows an increase in the percentage of Twitter-disseminated earnings announcements over the years from 2007 till 2015. Although the first companies joined Twitter in March 2007, a substantive increase in tweeted earnings announcements can be identified from 2010 onwards. When the observations before 2010 are neglected, as is done in one of the robustness analyses in section VII, 46.5 percent of the 490 earnings announcements are additionally disseminated via Twitter.
20 Table 1 – Tweeted versus Non-Tweeted Earnings Announcements
Panel A: Distribution of Earnings Announcements, Tweeted versus Non-Tweeted
Total Sample QTR Sample ANN Sample
All Twee ted
Non-Tweeted
All Tweeted
Non-Tweeted
All Tweeted
Non-Tweeted
Number 677 232 445 305 138 167 372 94 278
Fraction 100% 34% 66% 100% 55% 45% 100% 25% 75%
Panel B: Tweeted versus Non-Tweeted by Year
QTR Sample ANN Sample
Year All Tweeted Non-Tweeted All Tweeted Non-Tweeted
2015 20 16 (80%) 4 (20%) 49 24 (49%) 25 (51%) 2014 34 27 (79%) 7 (21%) 49 19 (39%) 30 (61%) 2013 34 26 (76%) 8 (24%) 42 15 (36%) 27 (64%) 2012 42 26 (62%) 16 (38%) 41 16 (39%) 25 (61%) 2011 48 22 (46%) 26 (54%) 42 12 (29%) 39 (71%) 2010 43 17 (40%) 26 (60%) 46 8 (17%) 38 (83%) 2009 30 3 (10%) 27 (90%) 39 0 39 (100%) 2008 31 0 31 (100%) 45 0 45 (100%) 2007 23 1 (4%) 22 (96%) 19 0 19 (100%)
Table 2 - Differences between Tweeted and Non-Tweeted Announcements
QTR Sample ANN Sample
Tweeted Non-Tweeted Difference P-value T-test Tweeted Non-Tweeted Difference P-value T-test Earnings Surprise -0.0002 (0.0006) -0.0025 (0.0020) -0.0022 (0.0021) 0.2841 -0.0014 (0.0032) -0.0029 (0.0036) -0.0015 (0.0048) 0.7558 Market Cap (in millions) 16.1374 (0.9290) 11.0074 (1.9717) -5.1300** (2.1796) 0.0194 9.9280 (1.3452) 3.3480 (0.4235) -6.5800*** (1.4104) 0.0000 ROA 5.5014 (0.6391) 6.6245 (1.0078) 1.1231 (1.1933) 0.3475 2.8959 (0.6738) 5.0048 (0.6091) 2.1089** (0.9083) 0.0214 Total Assets (in millions) 126.78 (22.78) 24.47 (7.72) -102.31*** (24.05) 0.0000 94.21 (28.28) 12.49 (2.72) -81.72*** (28.41) 0.0053 Media Coverage (yearly) 12040 (917.16) 3712 (403.60) -8328*** (1002.03) 0.0000 5749 (873.07) 1544 (191.31) -4206*** (893.78) 0.0000 Media Coverage (anndats) 198.87 (10.82) 65.93 (5.62) -132.94*** (12.19) 0.0000 113.51 (13.32) 32.48 (3.05) -81.03*** (13.67) 0.0000
Market Capitalization, ROA and Total Assets are winsorized at the 1 percent tails. Standard errors in parentheses. ** significant at 5% level, ***significant at 1% level
21 Table 2 provides summary statistics for the tweeted versus the non-tweeted earnings announcements for both the quarterly and the yearly sample separately. Within both samples, no significant difference can be identified for the earnings surprise between tweeted and non-tweeted earnings announcements. In the yearly sample, a weak significant (5% level) difference in ROA is identified between tweeted and non-tweeted announcements, though this difference is no longer significant within the quarterly sample. Market capitalization shows to differ significantly between tweeted and non-tweeted announcements, although the level of significance is higher for the annual sample (1% level) than for the quarterly sample (5% level). Still, firms supporting their earnings announcement with a tweet have higher market capitalization than firms not tweeting their announcement. A strong significant difference (1% level) is observed in total assets and media coverage between Twitter-disseminated and non-Twitter-disseminated earnings announcements in both samples. Firms tweeting their earnings have significantly higher total assets than firms not tweeting their earnings. Within the yearly sample, firms tweeting their earnings have even more than eight times higher total asset value than non-tweeting firms. Additionally, table 2 shows that firms disseminating their earnings via Twitter are covered more in other media outlets as well. This observation holds for both measures of media coverage (yearly measure, and number of media outlets on announcement date), and both samples. Consequently, a difference in response to tweeted announcements could be due to a difference in the characteristics of companies announcing their earnings additionally via Twitter. Therefore, control variables for certain firm characteristics are incorporated later in the regression analyses. Moreover, it could be that a difference in response to tweeted versus non-tweeted announcements is caused by the pre-announcement leakage of information. However, an analysis of the stock returns during the ten days before the earnings announcement shows that pre-announcement stock returns do not differ systematically between tweeted and non-tweeted earnings announcement, conditioned on earnings news (Appendix A). Additional to table 2, Appendix B provides correlation matrices of the variables.
Table 3 reports the average earnings surprise within each of the eleven defined quantiles for the tweeted announcements versus the non-tweeted announcements. In both samples, the within-quantile average earnings surprise is practically similar for tweeted versus non-tweeted earnings announcements, except for the lowest (most negative) quantile where the average earnings surprise for non-tweeted earnings is almost twice the average surprise for tweeted earnings.
22 Table 3 – Earnings Surprises by Quantile, Tweeted versus Non-Tweeted
Panel A: Average Surprise by Earnings Surprise Quantile (QTR)
Quan tile Low 1 2 3 4 5 es = 0 6 7 8 9 10 High 11 Tweet-ed Avg. N -0.0182 12 -0.0061 11 -0.0025 11 -0.0013 11 -0.0004 11 0.0000 4 0.0004 16 0.0016 16 0.0028 15 0.0044 16 0.0106 15 Non- Tweet-ed Avg. N -0.0336 17 -0.0042 16 -0.0024 16 -0.0013 16 -0.0004 16 0.0000 8 0.0004 16 0.0009 16 0.0019 15 0.0040 16 0.0120 15
Panel B: Average Surprise by Earnings Surprise Quantile (ANN)
Quan tile Low 1 2 3 4 5 es = 0 6 7 8 9 10 High 11 Tweet-ed Avg. N -0.0573 10 -0.0179 10 -0.0074 10 -0.0040 10 -0.0017 9 0.0000 4 0.0007 9 0.0025 8 0.0069 8 0.0210 8 0.0629 8 Non- Tweet-ed Avg. N -0.1040 23 -0.0162 23 -0.0073 22 -0.0025 23 -0.0006 22 0.0000 12 0.0004 31 0.0016 31 0.0040 30 0.0094 31 0.0576 30
While focusing on noise making, it is striking to observe that within the quarterly sample the number of tweeted observations is higher in the positive surprise quantiles than in the negative surprise quantile, while the number of non-tweeted observations is more or less similar for positive and negative surprises. This may be an indication that, within the quarterly sample, managers indeed more frequently announce more positive earnings news via Twitter than negative earnings news. Within the yearly sample this is the opposite. While the number of non-tweeted observations is higher for positive announcements, the number of tweeted-observations is higher for negative announcements.
Table 4 - Summary Statistics Noise Making
QTR Sample ANN Sample
Observations 137 94 Mean 5.14 7.30 Standard deviation 7.01 13.36 Median 3.38 4 Minimum 0.48 0.67 Maximum 64 92
Note: Both samples are limited to the tweeted earnings announcements
Table 4 provides the summary statistics for the noise variable, where both samples are limited to the Twitter-disseminated earnings announcements only. The mean noise of 5.14 for the quarterly sample means that on average firms announcing their quarterly results post 5.14 times more tweets during three days around the announcement ([-1;+1] window) than during three average days in the week before and after the announcement. Equivalently for the annual sample, on average firms announcing their yearly results post 7.3 times more tweets during the three days around the announcement than during three
23 average days in the week before and after the announcement. The median noise is comparable between the quarterly and the yearly sample, i.e. 3.38 versus 4.
In table 5 the average earnings surprises by earnings surprise quantile are reported for firms making above-median noise on their Twitter timeline versus firms making below-median noise on their Twitter timeline. Although table 5 only contains tweeted earnings announcements, the thresholds for the bins are set the same as in table 3, i.e. for the total sample. Thus, the total number of observations for the quantiles 1 to 5 and for the quantiles 7 to 11 are equal, and the same as in table 3.
No consistent pattern can be identified for the distribution of noise making and non-noise making firms for good versus bad news. In other words, from table 5 it cannot be observed whether firms make more noise on their Twitter timeline in the case of bad news or in the case of good news.
Table 5 – Earnings Surprises by Quantile, High Noise versus Low Noise Panel A: Average Surprise by Earnings Surprise Quantile (QTR)
Quan tile Low 1 2 3 4 5 es = 0 6 7 8 9 10 High 11 High Noise Avg. N -0.0143 4 -0.0064 6 -0.0024 6 -0.0012 4 -0.0006 1 0.0000 3 0.0005 9 0.0016 12 0.0027 8 0.0046 9 0.0090 7 Low Noise Avg. N -0.0210 7 -0.0057 5 -0.0027 5 -0.0014 7 -0.0004 10 0.0000 1 0.0003 7 0.0015 4 0.0029 7 0.0043 7 0.0119 8 Total N 11 11 11 11 11 4 16 16 15 16 15
Panel B: Average Surprise by Earnings Surprise Quantile (ANN)
Quan tile Low 1 2 3 4 5 es = 0 6 7 8 9 10 High 11 High Noise Avg. N -0.0488 5 -0.0192 4 -0.0078 6 -0.0041 7 -0.0017 7 0.0000 4 0.0005 5 0.0020 3 0.0059 4 0.0193 4 0.0547 4 Low Noise Avg. N -0.0659 5 -00170 6 -0.0068 4 -0.0036 3 -0.0017 2 - 0 0.0008 4 0.0028 5 0.0078 4 0.0227 4 0.0710 4 Total N 10 10 10 10 9 4 9 8 8 8 8
Note: Both samples are limited to the tweeted earnings announcements. The thresholds for the quantiles are set as in table 3.
C. Methodology
This study analyzes the difference in immediate stock price response between Twitter-disseminated earnings information versus non-Twitter-disseminated earnings information. Additionally, the possibility of noise making by a firm on its Twitter timeline is incorporated. This research continues to build on the model applied by DellaVigna and Pollet (2009). In line with DellaVigna and Pollet (2009) pooled OLS regressions are performed. However, where DellaVigna and Pollet (2009) clustered standard errors on only one dimension, the standard errors within the analyses performed in this study are clustered on two
24 dimensions. Two-way clustering of standard errors, i.e. on firm and time, is recommended by Petersen (2009).
Immediate Stock Price Response
The immediate stock price responses are analyzed on two levels. First, all earnings surprise quantiles are incorporated within the model. Secondly, only the most negative surprises (1st quantile) and the most positive surprises (11th quantile) are incorporated within the model, to examine the stock price response to very positive earnings news versus very negative earnings news. To test the immediate stock price response the regression equations are the following:
CAR[0/+1]t,i = β0 + β1Tweett,i + β2 q_est,i + β3Tweett,i * q_est,i + controls + εt,i (5)
CAR[0/+1]t,i = β0 + β1Tweetit,i + β2 q_es11t,i + β3Tweett,i * q_es11t,i + controls + εt,i (6)
with CAR[0/+1]t,i being the immediate abnormal stock return for company i in quarter t. Tweett,i is a
dummy variable indicating an earnings announcement which is additionally disseminated via Twitter (tweeted=1, non-tweeted=0). q_est,i is the quantile number of the earnings surprise (1 to 11). Within the
first model, all eleven quantiles are incorporated. Within the second model, only the first and eleventh quantile are incorporated. Here, q_es11t,i is a dummy for the eleventh quantile of earnings surprises
(q_es11t,i = 1 for quantile 11 and q_es11t,i = 0 for quantile 1).
In both models, β0 represents the constant and β1 measures the difference in immediate response between tweeted and non-tweeted announcements. Within the first model, β2 measures the (linear) relationship between the quantile number (from most negative to most positive earnings surprises) and the reaction after the announcement for non-tweeted announcements. In other words, β2 captures the average response of stock return for an increase of one earnings surprise quantile for non-tweeted announcements. As the second model includes a dummy for the top quantile, here β2 measures the return to very good news (top quantile) relative to very bad news (bottom quantile) for non-tweeted announcements. Specifically, this coefficient tests whether good news is associated with higher immediate response. Although this coefficient is not the focus of this research, the term must be included in the regression equation since the interaction term is included.
The coefficient of main interest in both models is β3, representing the differential reaction for tweeted versus non-tweeted announcements. Specifically, β3 measures the top-to-bottom earnings surprise quantiles difference in immediate reaction for tweeted-announcements relative to non-tweeted announcements. Under the null-hypothesis, where tweeting earnings announcements does not influence
25 investors’ attention, β3 should equal zero. Under the alternative hypothesis, where Twitter-disseminated earnings announcements are expected to increase investors’ attention, β3 should be positive for the immediate response.
Within both models, a set of control variables are incorporated. It may be expected that the responsiveness of stocks to earnings news is related to different factors, e.g. company size or coverage in other media outlets. Table 2 showed that total assets and market capitalization, as measures of size, and two measures of media coverage indeed significantly differ for firms tweeting their announcements and firms non-tweeting. Therefore, these four factors are included as control variables within the regressions. All financial control variables are winsorized at the 1 percent tails, as is common when working with accounting data (Matsa and Miller, 2013). Moreover, indicators for the year, quarter and month of the announcement are included. In line with Petersen (2009) standard errors are clustered on two dimensions. Within the quarterly sample, standard errors are clustered on firm and quarter, while within the yearly sample, standard errors are clustered on firm and year.
Noise Making
To analyze the effect of noise making on the stock price response, both the quarterly and the annual sample are limited to the tweeted earnings announcements, since possible noise making on a firm’s Twitter timeline only matters when indeed earnings announcements tweets are posted. To test whether a difference in reaction occurs when firms make noise on their Twitter timeline around the earnings announcements, the following regressions are performed:
CAR[0;+1]t,i = β0 + β1 NoiseDummyt,i + β2 q_est,i + β3 NoiseDummyt,i * q_est,i + controls + εt,i (7)
CAR[0/+1]t,i = β0 + β1 NoiseDummyt,i + β2 q_es11t,i + β3 NoiseDummyt,i * q_es11t,i + controls + εt,i (8)
Where NoiseDummy is a dummy variable based on Noise, which is the tweeting intensity during the days around the earnings announcement in quarter t relative to the normal tweeting intensity, as stated in equation 4, for company i. All other specifications are the same as in regression equations 5 and 6, and the same control variables are added.
The coefficient of main interest in regression model 7 and 8 is β1. This coefficient tests the difference in immediate stock price reaction for noise making firms versus non-noise making firms. Under hypothesis 2, where the act of noise making on firms’ Twitter timelines around Twitter-disseminated earnings announcements is expected to influence investors’ attention and thus to raise the immediate response, β1 should be positive.
26 V. RESULTS
A. Graphical Evidence
Immediate Stock Price Response
Figure 1 – Immediate Response to Earnings Surprises (a: QTR, b: ANN)
(b) (a)
27 Figure 1 graphically shows the immediate response of stock returns to earnings surprises for Twitter-disseminated versus non-Twitter Twitter-disseminated earnings information. Figure 1a demonstrates the immediate response for the quarterly sample, while figure 1b demonstrates the immediate response for the yearly sample. No consistent pattern in the difference in immediate response between tweeted and non-tweeted earnings announcements can be observed from figure 1. The responsiveness of stock prices to earnings surprises was expected to be flatter for non-Twitter-disseminated earnings. This is not confirmed in both figure 1a and figure 1b, which may be a consequence of not finding a clear upward-sloping line for the mean immediate response in general (Appendix C). Not finding an clear upward-sloping line for the mean immediate response is striking, since stock prices are generally expected to rise in response to positive earnings surprises, while they tend to drop in response to negative earnings surprises (DellaVigna and Pollet, 2009).
From both figure 1a and 1b it is striking to see that the immediate responsiveness of stock prices is, on average, more volatile to tweeted earnings announcements than for non-tweeted earnings announcements. Thus, it may be that investor indeed react more to Twitter-active firms’ announcements, although not in the same direction.
Noise Making
Graphical evidence for the immediate stock price response for ‘high noise’ versus ‘low noise’ announcements is reported in Appendix D. For the immediate response in the quarterly sample, a more pronounced immediate reaction can be observed for the announcements surrounded by higher noise on the Twitter timeline for the most bad (quantile 1) and the most good news (quantile 11). This observation is in line with expectations, as higher noise is expected to attract attention. It is striking to observe that investors appear to react more positive to bad news and more negative to good news which is surrounded by noise. For the other quantiles, no consistent pattern can be identified. The annual sample shows a comparable immediate reaction to earnings surprises for high noise and low noise, except for the quantiles 2 and 3. For these two bad news quantiles, investors react positively to announcements surrounded by noise, while they react negative to announcements which are less surrounded by noise. This observation may give an indication that bad news (but not the most extreme bad news) is somewhat obscured by the noise. Taking the graphs for the immediate response together, no consistent pattern which supports expectations can be identified.
28 B. Regressions
Immediate Stock Price Response
Table 6 – Immediate Stock Price Response to Earnings Surprises (QTR) Dependent variable:
CAR[0/+1]t,i
All Quantiles Top and Bottom Quantile
(1) (2) (3) (4) Tweet -0.0040 0.0092 0.0010 -0.0063 (0.3477) (0.2375) (0.4347) (0.4556) Quantile 0.0014* 0.0016 (0.0650) (0.1140) Tweet * Quantile 0.0001 -0.0010 (0.4737) (0.2886) Top Quantile 0.0258*** 0.0207 (0.0003) (0.2688)
Tweet * Top Quantile -0.0380*** -0.0477
(0.0008) (0.2104) Constant -0.0060 -0.0136 0.0005 0.2511 (0.1753) (0.3759) (.) (0.1409) Observations 305 261 59 53 R-squared 0.0128 0.1541 0.0868 0.5550 Controls X X 2way Clustered SE X X X X
One-sided P-values in parentheses, * significant at 10% level, ***significant at 1% level
Table 7 – Immediate Stock Price Response to Earnings Surprises (ANN) Dependent variable:
CAR[0/+1]t,i
All Quantiles Top and Bottom Quantile
(1) (2) (3) (4) Tweet -0.0088 -0.0063 0.0206*** 0.0324*** (0.1506) (0.2746) (0.0019) (0.0078) Quantile 0.0011* 0.0008 (0.0580) (0.1698) Tweet * Quantile 0.0011 0.0006 (0.1433) (0.3175) Top Quantile 0.0211*** 0.0235** (0.0052) (0.0102)
Tweet * Top Quantile 0.0027 0.0052
(0.4266) (0.4049) Constant -0.0027 -0.0145 -0.0133** -0.1599** (0.3250) (0.3058) (0.0266) (0.0372) Observations 372 304 71 58 R-squared 0.0130 0.0458 0.0860 0.3687 Controls X X 2way Clustered SE X X X X
