text
Executive Tweets Richard M Crowley Assistant Professor of Accounting Singapore Management University rcrowleysmuedusg Wenli Huang Assistant Professor of Accounting School of Accounting and Finance Hong Kong Polytechnic University wlhuangpolyueduhk Hai Lu Professor of Accounting Rotman School of Management University of Toronto HaiLuRotmanUtorontoCa This draft May This research was supported by the Hong Kong General Research Fund Executive Tweets Abstract We examine the behavior of executives on Twitter focusing on the impact of the US Securities and Exchange Commissions SEC Report on social media usage by corporations and executives as well as on firmspecific information disclosure We document a limited impact of the SEC report however we document substantial evidence that managers do release information related to their firms and potentially of interest to investors particularly around important events such as earnings announcement and conference calls SEC filings press releases and news articles We further document that investors react to executives tweets that are financial in nature reacting more strongly to such tweets when posted by executives than when such tweets are posted by their firms Lastly we document a method for proxying for executive effort using the content of executives tweets by examining the similarity between executive and firm tweets We find that our similaritybased effort measure captures onequarterahead growth opportunities for firms Keywords Social media executives dissemination Twitter executive effort JEL Codes Page Executive Tweets Introduction Over the past decade social media has become more and more central to the way in which individuals and corporations exchange information with Twitter alone having million active users per Furthermore actions of executives on Social media have caught the attention not just of individuals but also the media and the US Securities and Exchange Commission SEC On July the CEO of Netflix Reed Hastings posted a message on Facebook that contained information useful for backing out Netflix financial performance The post in question led to an increase in Netflix stock price of on the same day and would spawn an SEC investigation into how executives and firms use social media to disclose or disseminate information At the end of the investigation the SEC determined that social media accounts both of firms and executives were public enough to disclose information through while maintaining compliance with Regulation Fair disclosure According to the SEC an increasing number of public companies are using social media to communicate with their shareholders and the investing public We appreciate the value and prevalence of social media channels in contemporary market communications and the Commission supports companies seeking new ways to communicate and engage with shareholders and the market SEC As Miller and Skinner point out the SECs guidance along with the tremendous growth and penetration of social media makes it likely that social media will become an increasingly important component of firms disclosure strategies One immediate implication of the SECs guidance is that executives can also use social media to As of per Twitters Shareholder Letter Page broadcast company information to their investors without legal consequence as far as the information is not fabricated In recent years the focus on executives on Twitter has increased with wellknown executives like Richard Branson Virgin Tim Cook Apple Aaron Levie Box Elon Musk Tesla SpaceX and Satya Nadella Microsoft having millions of followers each Furthermore media attention of executives on Twitter has grown rapidly in recent years with press articles on CEO tweets in the first four months of Likewise the US Presidential election brought additional focus to the use of Twitter by wellknown individuals In this study we examine four research questions First did the SECs conclusion in April affect the use of Twitter by executives Second we examine if executives appear to post information that would be useful to investors Third we examine the impact that executives tweets have on the stock market Lastly we examine whether it is possible to use executives Twitter data to capture the level of effort executives put into their firms Our findings document that the SECs report following the investigation of Reed Hastings post did not have any appreciable effect on executives Instead a mixture of demographic and firm characteristics appears to drive adoption from through We then turn our attention to what executives are posting on Twitter examining the composition of financial information businessrelated nonfinancial information and other information they post We again find little impact of the SECs report however we find strong and consistent evidence of that executives do tweet about firmrelevant information and that they are particularly likely to do so around many salient firm events including earnings announcements and earnings conference calls and filings filings press releases and news Based on a Factiva search for CEO and tweet in the headline of an article excluding web content Page articles As such we find strong evidence that executives do tweet about information that is relevant to investors We then shift our focus to stock price reaction to executives tweets We document a strong reaction to any tweets posted by executives that are financial in nature and that the reaction to financial tweets is stronger to tweets posted by managers than to financial tweets posted by firms Lastly we construct a new measure of executive effort based on the similarity between executives tweets and their firms tweets We find that this measure appears to be capable of picking up at least some part of effort and can be used to predict onequarterahead Tobins Q Our paper makes important contributions to three literatures First we contribute to the burgeoning research in finance and accounting on social media Our paper is the first in the literature to examine both the drivers of executives joining Twitter as well as the first to examine the determinants of executives tweeting about their company in terms of financial and nonfinancial information We document that executives appear to be quite intune with dissemination and disclosure by and about their firm and that they increase tweeting of financial information around earnings announcements earnings conference calls and filings Furthermore for broader disclosures including filings press releases and news articles we find that executives increase tweeting of both financial as well as nonfinancial business information on Twitter Second we contribute the regulation literature While Crowley Huang and Lu documented a limited impact of the SEC Report on the format of tweets posted by firms this study examines the impact of this SEC report on executives Twitter behavior We document limited impacts of this report on the rate of executives joining Twitter as well as on executives tweet content Page Lastly we contribute to the managerial compensation literature by demonstrating a Twitterbased measure of executive effort We construct a measure of executive effort based on the content executives disclose on Twitter along with a firmspecific baseline the firms own Twitter accounts We show that the similarity of tweets between executives and their firm can be used to predict the growth opportunities one quarter ahead consistent with capturing executive effort We also conduct a test to rule out an alternative explanation that the marketing or PR function is the driver behind both the similarity and improved growth Literature and Hypothesis Development Literature review As manager and firm disclosure are endogenously linked with managers influencing both firms outcomes and disclosures this study relates to the emerging literature on firms use of social media Over the past decade social media has transformed the way firms engage with their customers investors and the market Considering the proliferation of social media in marketing campaigns it is not surprising that the impact of social media on corporate marketing and sales has been substantially studied in recent marketing literature For instance Kumar Bhaskaran Mirchandani and Shah show that social media can be used to generate growth in sales return on investment and positive word of mouth Of all social media outlets Twitter is considered by many firms as their primary choice of the social medium due to its simple social short and tangible features The extent of use of Twitter by firms as compared to other social media channels is discussed in Jung Naughton Tahoun and Wang which finds that there are more SP firms on Twitter than on all other examined platforms Facebook YouTube LinkedIn Google and Pinterest Furthermore it is commonly viewed that firms Twitter Page followers are more likely to be present or potential investors while other outlets such as Facebook and LinkedIn are mainly used for social interaction and professional networking A separate strand of literature has focused on predicting firmlevel stock performance andor stock market performance by leveraging content on Twitter Bollen Mao and Zeng and Mao Wei Wang and Liu examine how aggregate investor mood on Twitter can be used to help predict changes in the stock indices Sprenger Tumasjan Sandner and Welpe looks more closely at individual firms analyzing a set of almost tweets related to SP firms over the span of six months They find that tweet sentiment volume disagreement appears to be associated with stock returns volume volatility Curtis Richardson and Schmardebeck examine how investor response to earnings news is related to investors activity on Twitter finding a positive associate between the two More recently Bartov Faurel and Mohanram examine both firmlevel stock returns and earnings and find that information contained on Twitter helps predict both More specific to firm dissemination and disclosure resent academic research has examined the extent to which firms disseminate information on Twitter as well as the capital market consequences of this dissemination Blankespoor Miller and White examined the impact of technology firms disseminating hyperlinks to earnings announcement press releases via Twitter finding that this dissemination facilitates a decrease in information asymmetry Lee Hutton and Shu examines the context of consumer product recalls They show that during a recall firms can limit the negative price reaction to the announcement of the recall by using social media In a discussion paper Miller and Skinner provide a framework identifying several important themes in the disclosure literature and encourage future research to continue exploring emerging forces in disclosure such as the role of social media Jung et al provide large Page sample evidence on the determinants and market consequences of firms decision to disseminate financial news through social media They find that firms are less likely to disseminate when the news is bad and when the magnitude of the bad news is worse consistent with strategic behavior Crowley Huang and Lu provide largesample evidence on firms dissemination strategies on Twitter using a sample of all tweets by SP firms from through They find that firms disseminate more financial information on Twitter around earnings announcements and filings and filings if the information contained in the firms other disclosures is sufficiently positive or negative also consistent with strategic behavior While this result appears to suggest that firms may be providing a more complete set of disclosure on Twitter this does not necessarily apply to all types of disclosure As shown in Crowley Huang Lu and Luo the equilibrium for CSRrelated disclosure on Twitter appears to involve companies with lower CSR ratings tweeting more about CSR than higher CSR rated firms One last strand of literature examines why individuals use Twitter While Twitter first went online in as early as in it was documented that individuals were using Twitter was to share and seek out information Java Song Finin and Tseng Toubia and Stephen provide an indepth investigation of the motivations of users to contribute content to Twitter They focus exclusively on noncommercial accounts on Twitter and through a field study they document that users derive utility both intrinsically directly through posting tweets as well as through imagerelated effects indirectly through perception of themselves by others Lin and Lu document an additional incentive for users to join Twitter their peers Using a questionnaire they find that intrinsic utility is a driver of Twitter usage but that the presence of individuals peers on Twitter drives further intrinsic utility On the other hand determinants for a company having a Twitter account are largely financial in nature While the motivations for having Page a Twitter account may be different across individuals and companies accounts it is likely that both firm and individual incentives affect executives decision to have a Twitter account and to post on Twitter tweet That is an executive may post on Twitter for intrinsic or imagerelated benefits or they may post on Twitter for companyrelated reasons Our study is also related to the literature on managerial incentives for disclosure Prior literature on dissemination and disclosure by managers is largely focused on management earnings forecasts earnings announcements and conference calls It is commonly argued that firms and managers have incentives to withhold negative information Delaying disclosure of negative news is documented by Kothari Shu and Wysocki which finds that investors react more strongly to negative news than to positive news about cutting dividends and management earnings forecasts consistent with asymmetric disclosure of positive and negative news More recently Bertomeu Ma and Marinovic use a structural model to estimate the rate at which managers withhold earnings forecasts finding that nondisclosure is strategic at a rate of about one of every four withheld management earnings forecasts Withholding negative information is in the best interest of the manager in these cases but this leads to an increase investors uncertainty over the firms performance In the compensation theory literature significant focus has been given to moral hazard related to managers effect on firm disclosure A common feature of managerial compensation models is that a managers compensation is contingent on firm or management disclosure Thus withholding or delaying negative information can be interpreted as a manifestation of this moral hazard As managers control the information environment within their firm any disclosures representing the firm or the manager are interpreted rationally and compensation contracts attempt to back out any influence of the manager on the information environment In a perfect information Page environment investors would directly know a mangers effort and could contract on this level of effort to in expectation induce a firstbest outcome Margiotta and Miller However the opacity of the information process leads to a reliance on expost disclosure that is correlated with or noisily captures managers actual effort leading to a secondbest contracting equilibrium In the archival compensation literature the construct of effort is generally passed over in favor of examining the relationship between compensation and firm performance directly Murphy et al However as shown by Margiotta and Miller investors would stand to gain by knowing the effort exerted by managers As executives Twitter accounts are not directly representative of the firm but instead are representative of the manager it is possible that information disclosed on Twitter is useful in understanding effort put forth by executives for their firms Hypothesis development There is relatively little research on the behavior of executives or managers on Twitter Ex ante executives may behave like any other individual on Twitter using Twitter for their own personal enjoyment It is also possible that executives will instead choose to represent their firms releasing content that is in line with their firms disclosures to aid in strategic dissemination of corporate information A third possibility is that managers will release content that fits with both previously outlined behaviors As such it is unclear if managers will tweet information that is potentially useful to other parties If managers do have a desire to strategically disseminate information a natural question is if executives disseminate or disclose any information that is useful from the perspective of investors If executives do disseminate useful information one potential issue they could have encountered was the unclear legal status of disclosing corporate information on Twitter prior to Page April On April the SEC released a report titled Report of Investigation Pursuant to Section of the Securities Exchange Act of Netflix Inc and Reed Hastings Reed Hastings on July had posted investorrelevant information in a public post on The SEC decided to investigate this post to examine if posting investorrelevant information via an executives social media account was a violation of Regulation Fair Disclosure Reg FD and if the SECs August Guidance on the Use of Company Web Sites was applicable to social media platforms The SECs primary finding was that the SEC guidance was applicable and that executives posting investorrelevant information on social media is not a violation of Reg If managers did have a desire to disseminate strategically on Twitter it is likely that the SECs report would drive the adoption of Twitter by executives and that the report would also bring about a change in the content of tweets posted by executives SEC Such a change should be particularly salient with respect to financial content as financial content is perhaps most clearly investor relevant A finding of the SECs investigation bolsters this expectation as the investigation found that there is uncertainty concerning how Regulation FD and the Commissions Guidance apply to disclosures made through social media channels On the other hand the SEC did have preexisting guidance on the use of company web sites since SEC Furthermore given that the SEC found that social media use was Congrats to Ted Sarandos and his amazing content licensing team Netflix monthly viewing exceeded billion hours for the first time ever in June When House of Cards and Arrested Development debut well blow these records away Keep going Ted we need even more PM UTC July The SEC does however suggest informing investors that a certain social media account is used for disclosure or dissemination of important information for investors An example of such as disclosure for Twitters press release is as follows Twitter has used and intends to continue to use its Investor Relations website and the Twitter accounts of jack nedsegal twitter and TwitterIR as means of disclosing material nonpublic information and for complying with its disclosure obligations under Regulation FD Available at Page consistent with the guidance it is possible that some firms and executives would have no issue with disclosing investorrelevant information on social media even before the SEC report Thus it is possible that there was no impact of the SEC report on executives behavior on social media Alternatively if most executives have no interest in posting companyrelated information on Twitter then there should likewise be no effect of the SEC report on executives adoption of Twitter or on their tweet content For firm Twitter accounts there is some limited evidence from Crowley et al that the content of tweets by firms were unaffected by the SEC report but that firms usage of media and links in financial tweets may have increased after the report Given the differences in motivations for firm and individual Twitter account usage however it is unclear if this result would follow through for executives Twitter accounts Given the contrasting expectations on the effect of the SEC report on executive behavior on Twitter this leads to our first of hypothesis in alternative form Hypothesis The rate of executives joining Twitter and tweeting investorrelevant of information increased after the SEC report was released An increase in executive activity on Twitter would directly show a desire by executives to post investorrelevant information on social media However a lack of an effect of the SEC report would not directly show a lack of interest by executives in posting information about their companies on social media due to the possibility that they expected to be covered by the SEC guidance If this is the case then we should expect that executives react to various events their company is experiencing For instance if managers desire to post investorrelevant information on Twitter then the release of an earnings announcement holding an earnings conference call or release of a or filing should drive executives to post about financial information on Twitter Likewise the release of important documents with a broader focus such Page as filings and press releases should drive financial as well as broader businessrelated dissemination or disclosure on Twitter by executives as should news about the firm covered in the press If managers do not want to post about their firms however we would expect their behavior on Twitter to not respond to such important events This leads to our second hypothesis Hypothesis Executives tweet more about financial information on days with financialrelated information disclosure earnings announcements or earnings conference calls and filings Hypothesis Executives tweet more about financial and other businessrelated information on days with businessrelated information disclosure or dissemination filings press releases news press If results consistent with either Hypothesis or Hypothesis are found then a natural question to ask is if investors find the information useful While there are examples of social media posts that have moved the market such as Reed Hastings Facebook post on the number of hours watched by customers or Elon Musks about taking Tesla public it is possible that the majority of information posted by executives is largely the same as what is already available elsewhere As such we might expect the information to have no effect on the market On the other hand there is evidence that even tweets by investors are able to predict market movements Bollen et al Mao et al Sprenger et al and it is natural to expect an executive to be able to post more useful information than most investors Furthermore even repeated dissemination of the same information has been shown to be able to affect stock prices Tetlock As such even if executives play it safe and avoid posting new disclosures on Am considering taking Tesla private at Funding secured PM UTC Aug Page Twitter the information may still be useful to investors and may still move stock prices This leads to our third hypothesis Hypothesis Tweets by executives particularly tweets about financial information lead to changes in their firms stock price We will examine this hypothesis by examining absolute returns and their relationship with executives tweets Extension Tweet similarity and executive effort Given the incentives for tweeting already discussed previously intrinsic imagerelated peer effects tweets by executives are likely to have a more human element than most firm disclosure or dissemination To some extent tweets provide a window into the minds of the individuals tweeting and do so in a way that is unintentional from the posters perspective Examples of other studies leveraging Twitter to tap into the minds of users include Eichstaedt et al which used Twitter data to predict county level mortality rates from atherosclerotic heart disease and De Choudhury Counts and Horvitz which used tweets to detect indications of depression among individuals on Twitter Notably Eichstaedt et al find that their model using only Twitter data performs just as well as traditional models leveraging socioeconomic health and demographic data for atherosclerotic heart disease mortality prediction The previously discussed studies peer into the minds of Twitter users to capture a construct that while not causally related to the act of tweeting o the content of the tweets manifests itself in the behaviors of users on Twitters In a similar vein we propose leveraging Twitter data in a novel way to capture archival evidence linking managerial effort and firm performance As mentioned previously the construct of effort is generally not examined in the archival compensation Page literature due to the complexity and opacity of measuring the construct This would likewise be the case for using executive tweets alone as we would have no baseline for what we might expect a high effort manager to tweet However due to firms participation in Twitter we propose leveraging firm tweets as a natural baseline for what is important for the firm We then construct a measure of similarity based on executive tweets and firm tweets by assessing how similar the underlying meanings of the tweets are If firms tweets serve as a reasonable baseline for what the firm is focused on at a given point in time then the similarity measure should capture the extent to which the executive and firm are in line with one another which in turn reflects the effort that the manager is exerting for the company Following traditional managerial compensation models one would thus expect a firm to be better positioned if its executive is exerting more effort Our fourth and final hypothesis is thus Hypothesis The similarity of tweets between executives and firms positively predicts future firm growth opportunities Data and methodology Data and sample selection Our sample spans the years through and covers all SP firms that were contained in the SP index between January and September Twitter handles for companies and CEOs were identified between September and October while Twitter handles for CFOs were identified in April Our sample is based on a mix of data from GNIP and the Twitter API as described in Crowley Huang and Lu This paper uses an extended version of the corpus used for Crowley et al including all tweets from the prior corpus and adding an additional one year of data derived from the Twitter API In total we have identified firm accounts and executive accounts executive Page accounts and are excluded from our sample as the executives did not release any tweets while they were CEOs or CFOs of SP firms or had their accounts set to private and firm accounts were excluded due to having no tweets or being set to private Our financial data executive data and stock return data are from Compustat Fundamentals Quarterly Execucomp and CRSP respectively For identifying information events we use the following sources IBES for earnings announcement times Capital IQ for earnings call times WRDS SEC Analytics Suite for and times Ravenpack PR Edition for Press release times and Ravenpack Dow Jones Edition for News article times and content Our full sample consists of all firmexecutive pairs that were in the SP any time between January and September that has complete control variable information in Compustat that is in CRSP and that has a CEO a CFO or both in Execucomp This sample is comprised of approximately million firmexecutivetrading day fiscal quarter observations Measure construction A key difficulty and feature of our data is that nearly all the data Tweets and all information events are tracked to the second of announcement As such we standardize all data by assigning each tweet or event to an NYSE trading day If a tweet or event occurs on a trading day and is released prior to PM in New York City we treat that day as the trading day If the tweet or event occurs after PM and seconds in New York City or if the tweet or event occurs on a day where US stock exchanges are closed we code the trading day to be the next day that US stock exchanges are open We take care to factor in issues such as the timezones data are derived from generally either New York time or GMT as well as to account for daylight savings time Page Twitter measures Our primary measures derived from our Twitter data are counts of the tweets posted by executives and firms by trading day Tweets are aggregated to the trading day level as described in section To measure the content of tweets we use the TwitterLDA algorithm by Zhao et al to machine learn the contents of tweets from to TwitterLDA has been used for tweet classification in Crowley et al TwitterLDA itself is a modified version of the LDA algorithm to adjust for the short length of tweets as short documents are a noted problem for LDA LDA Latent Dirichlet Allocation is a machine learning algorithm by Blei Ng and Jordan that classifies the thematic content ie topics of text in a Bayesian manner without any oversight from the researcher ie LDA is an unsupervised algorithm LDA has grown in popularity in the accounting literature and has been used in numerous studies see eg Dyer Lang and SticeLawrence Huang Lehavy Zang and Zheng Brown Crowley and Elliott For consistency with prior literature we use the same TwitterLDA model as used by and described in detail in Crowley et al This model classifies tweets into different machine learned topics We then cluster these topics into three overarching categories of information financial nonfinancial business henceforth nonfinancial and other Financial tweets are likely to be the most informative as financial information is crucial for investors Nonfinancial business tweets are companyrelevant tweets covering topics such as business events marketing conference participation and customer support and thus may be of interest to investors Lastly other tweets are likely unrelated to the firm and may be about daytoday life sports travel or other interests To categorize a tweet we determine which of the topics of the TwitterLDA model the tweet most relates to by applying the weighted dictionaries generated by TwitterLDA Page to each tweet and picking the topic with the highest weight for each tweet We map each tweet to a category based on its For daily analyses we aggregate tweets by counting the number of tweets in each category on the trading day In some tests we also use whether the executive or firm has a Twitter account with at least tweet as of a given day or fiscal quarter end For control variables we include the log of one plus the number of followers of each account the log of one plus the number of accounts the executive or firm Twitter account is following the number of tweets posted by the account over the previous five trading days and the log of one plus the total number of tweets posted by the account Of the control variables derived from Twitter data followers and following are all left censored measures as Twitter provides these measures at the time the information is accessed not historically For testing Hypothesis we introduce a finegrained measure of content or meaning of text to the accounting literature called Universal Sentence Encoder USE The USE algorithm by Cer et al leverages neural networks to process text on the order of sentences or short paragraphs factoring in word order As such this model breaks away from the typically bagofwords based approaches used in accounting such as dictionaries or LDA allowing it to ascribe a more precise meaning to a sentence Furthermore the short nature of tweets means that we can encode whole tweets easily with USE For our analysis we use a model pretrained on a variety of online information sources including Wikipedia web news web questionanswer pages and discussion forums Cer et al Given that Twitter is likewise a source of general webcontent we expect this model to transfer well to our context The output of the algorithm is a unit vector for each tweet mapping the tweet to an area of a vector The topics are hand classified in Crowley et al The financial category contains topic the nonfinancial category contains topics and the other category contains topics Appendix B contains examples of tweets in each category Page space that abstractly captures the meaning of the tweet While the vector space itself is not comprehendible in a human sense the vector space encodes similar meanings all in the same local area We leverage this feature as demonstrated in Cer et al to precisely measure the similarity of executives tweets with their firms tweets Examples of sentences encoded with this algorithm and their respective similarities are provided in Appendix C along with a more detailed description of our methodology To construct our measure for each executive tweet we identify all tweets by the executives firm within a trading day window dropping any executive tweets that do not have a corresponding firm tweet in the trading day window We then search for the closest firm tweet to the executive After distances are calculated for each executive tweet we take the mean of all of an executives tweets each fiscal quarter as an overall measure of the difference between the executives tweets and the executives firms tweets Lastly we convert the average distance measure to a similarity measure by normalizing the distance to be between and and subtracting this normalized distance from Our final measure labeled as Tweet Similarity is equal to if the executive and firm tweets in a fiscal quarter are identical and approaches as firm and executive tweets become less related Event measures The first event we examine was the SEC report which occurred on Tuesday April To differentiate between observations before and after the SEC report we construct a variable Post SEC equal to for April or later In quarterly tests we code Post SEC as if the fiscal quarter started after April We measuring distance using Euclidean distance and implement the Approximate Nearest Neighbor matching algorithm of Arya and Mount to efficiently compute exact matches between executive and firm tweets Our results are unchanged if we define Post SEC as the first fiscal quarter ending after April thus including the fiscal quarter during which the SEC report was released Page The other events we examine are all firmlevel events for which we have intraday data that we map to trading days as we did with tweets The first event is earnings announcements Using IBES data we create a variable Earnings announcement equal to if there is an annual or quarterly earnings announcement on the trading day otherwise The second measure we create is Earnings call which captures if there is an earnings conference call on a given day based on Capital IQs conference call schedule As earnings announcements and earnings conference calls overlap significantly are on the same trading days we present results using an aggregate measure of the two which we call Earnings event Our third type of measure captures releases of SEC filings Using WRDS we construct and filing which captures if there was a or filing released on the given trading day We also construct filings which captures the number of filings released on a given Our remaining events are all derived from Ravenpack For all Ravenpackbased measures we include only the first instance of an article by keeping only the earliest item matching an RPSTORYID and we require an item to have a relevance score of at least as recommended by Ravenpack Press releases is the count of unique press releases by a firm on a given trading day using Ravenpack PR Edition News articles is similarly constructed being the number of unique news articles about a firm on a given trading day using Ravenpack Dow Jones We include this measure as a count as opposed to a binary measure as we find that a nontrivial amount of days with filings contain multiple filings In our sample the most filed on the same trading day is however some SP firms whose managers were not on Twitter have filed as many as filings on the same day We include our measures of press releases and news articles as counts as these can be significantly clustered together For press releases we find that of days with a press release contain multiple unique press releases while of days with a news article contain multiple unique news articles In Section as an additional analysis we disaggregate our measure of news articles by subject to analyze the impact of specific subjects of news on executive tweeting Page Return measures Our primary return measures are based on market model return MMR We calculate betas using months of lagged daily returns trading days using SP returns For tests involving stock returns we present results using a precise window restricted to only day This eliminates any concern of the reverse causality of executives tweeting due to stock price movements Studies in accounting often use windows of or to account for information leakage or expectations of investors about scheduled events Tweets by executives however are not a type of disclosure that should be expected or that is mandatory in general and thus we expect less leakage of their effect to prior daystimes For robustness we also test our model using windows of and day t we also confirm our results using SP adjusted returns equivalent to assuming a beta of for all firmdays Empirical methodology and results Methodology Executive adoption of Twitter To examine Hypothesis we use our quarterly sample as described in Section We examine the impact of the SEC report on executives joining Twitter using a Cox proportional hazards survival model log Exec on Twitter is our event of interest and it is until an executive has posted hisher first tweet after which it becomes Our main variable of interest is Post SEC which captures the impact of the SEC report and is for fiscal quarters that are entirely after April the date of Page the For results consistent with Hypothesis we would expect to be positive and significant This model also serves as a determinants model to explain some of the variation in executives that did and did not join Twitter As such we also include the age of the executive Executive Age as the more frequent presence of younger individuals on social media is well We also include various financial variables used in the prior literature including firm size Size return on assets ROA market to book ratio MTB and debt to assets ratio Debt To control for and examine any potential links between firms Twitter activity and executives joining Twitter we also include measures of if the firm is on Twitter Firm on Twitter the number of followers the firm has logFollowersFirm the number of accounts the firm is following logFollowingFirm the number of tweets the firm posted in a given quarter TweetsFirm and the total number of tweets the firm has posted over time Total tweetsFirm Lastly we include industry fixed effects GICS sector as executives at more hightech industries are likely to be more aware of Twitter All variables in the regression are defined in Appendix A Executives tweet determinants Next we address Hypotheses and by examining determinants of different categories of tweets financial nonfinancial other For these tests we restrict to only executivefirmday observations where Exec on Twitter is ie days where the executive has a Twitter account and has already tweeted at least once since opening the account Our main focus will be on the quantity of tweets posted by executives around different events To implement this we introduce a new regression structure to the accounting literature using Poisson pseudo maximum likelihood PPML regression with robust standard errors and highdimensional fixed effects HDFE as Our results are robust to including the fiscal quarter overlapping the release of the report For instance Pew Research Center shows that in the US of year old individuals used Twitter as of January dropping monotonically with age until the age group at usage Page implemented in Correia Guimares and Zylkin PPML regression is interpretable like with the added benefits of being able to reliably use large amounts of fixed effects and being robust to sparse dependent variables ie dependent variables that are mostly Our main regression specification for Hypothesis is E Our dependent variable in these regressions is the count of all tweets in a certain category such as financial tweets Event is an indicator or count variable and it is either Post SEC or one of the event measures discussed in Section The variable Firm topic tweets controls directly for tweets on the same trading by an executives firm This serves to control for the possibility that the manager is simply responding to firm dissemination or disclosure as opposed to the events themselves For executives we retain executive age and control for other factors using an executive fixed effect For other controls we include the same financial controls as in the tests of Hypothesis We also include the Twitter controls for a firm including if the firm is on Twitter Firm on Twitter measures of the logged number of followers the firm account has logFollowersFirm and the number of accounts the firm is following logFollowingFirm the number of tweets the firm posted over the past trading week Recent tweetsFirm and the number of tweets the firm has posted to date LogTotal tweetsFirm We augment these Twitter controls by including the same measures for the executives Twitter accounts as well except for the on twitter measure which is always for executives in our The authors of this paper have made their work publicly available for Stata on SSC via the ppmlhdfe package Coefficients of the PPML regressions are logscale like with Poisson regression As such an easy interpretation of a coefficient is that is the incidence rate ratio IRR The IRR is multiplicative for instance an IRR of indicates that a change in the variable for the coefficient of leads to a increase in the dependent variable all else held constant Page sample Lastly we include a comprehensive collection of fixed effects firm executive which differentiates between CEO and CFO within firm as well as changes in management as well as year and month to capture any linear time trends in tweeting behavior We use the same controls and fixed effects throughout all regressions testing Hypothesis Market reaction to executives tweets To address Hypothesis we directly examine stock returns around executive tweets For these tests we use the same sample as we used for Hypothesis We examine absolute market model return We focus on absolute abnormal returns as this has been used as a reliably capture stock market reaction to disclosures with no ex ante known directional impact see eg Hope Hu and Lu To Hypothesis we use a linear regression with HDFE and robust standard As our events are precisely tracked intraday our primary tests use a window of just day though we also test windows of and day t for robustness Our independent variable of interest is Exec topic tweets where a positive and significant coefficient would be consistent with Hypothesis As with our regression tests for Hypothesis we control for the executives firms tweets the executives age financial controls twitter account controls for both the executive and firm and a set of fixed effects including firm executive year and month As it is potentially interesting to include executives not on Twitter as a control sample we also run all analyses related to stock return on the full sample used in the tests of Hypothesis Our results are robust to this specification We implement this regression model using reghdfe available in the Stata SSC and described in Correia Page Tweet similarity For Hypothesis we return to our firmexecutivequarter sample specification from our tests of Hypothesis We include all firmexecutivequarters where the firm tweeted within a trading day window of at least one executive tweet This ensures that we can construct our similarity measure described in Section To test Hypothesis we use a linear regression with HDFE and robust standard errors to examine the relationship between future Tobins Q and tweet similarity using the following specification If tweet similarity is reflective of the effort managers exert for their firm then we would expect to see a positive and significant coefficient on We include the same control variables as in regression equation and also include current quarter Tobins Q We also include a large collection of fixed effects including firm executive year and month fixed effects One potential concern with the methodology used for this regression is that there is a correlated omitted variable driving both the firms and the executives tweets particularly a firms marketing or PR strategy If some or all of the managers tweets are driven by marketing or PR then our results may simply reflect the outcome of a strategy of pushing marketing or PR information through both the firms and managers Twitter account To identify any potential effects of this alternative explanation we hand classified all clients used for posting tweets on Twitter in our sample over unique Twitter clients In our classification we identified if a client is typically used for marketing activities or not of which we found such clients A few examples of such clients include products by Cision and Hootsuit as well as adcampaign specific platforms such as CocaCola The Big Game We find that executives tweeted using such Page clients in firmexecutivequarter observations We also identified firmspecific Twitter clients such as Intuit Social Media Team and FEI Company but we find that not a single executive tweet was sent from a firm specific client over our full sample To rule out this potential alternative explanation we rerun regression equation excluding the observations where the executive used a marketing Twitter client at least once Results Univariate statistics Figure presents statistics on our sample of Twitter accounts The Panel A present the percentage of executives in our sample that have Twitter accounts with at least tweet in each year Predictably we find that the number of executives with Twitter accounts increases over time starting at just executives in and peaking at executives almost in In terms of industries not tabulated for brevity we find that communication services and information technology have the highest proportion of executives on Twitter with and of unique executives in those industries having a Twitter account while in an executive position at the firm In total our sample of executives with Twitter accounts consists of executives across executivefirmyears The second part of Panel A presents the collective number of tweets by executives per year We see a sustained increase in the number of tweets each year including in the most recent year of our sample Table Panel B presents the sample of firm Twitter accounts As with the manager accounts we find that communications and information technology have the highest rate of adoption of Twitter at and respectively Likewise usage of Twitter by firms does increase over time starting at of firms in and peaking at of firms in Overall our sample contains firms that had a Twitter account One noticeable difference Page between executive and firm tweeting is the change in the number of tweets in While tweets by executives continued to rise in tweeting by firms decreased in Table Panel C presents the distribution of different types of tweet content across the sample period We find that executives have a higher proportion of tweets relating to business matters than firms including both financial and nonfinancial tweets and that this difference is consistent throughout the sample In the final year of our sample executives post financial content nonfinancial content and other content Executive adoption of Twitter and the SEC report Our test of Hypothesis is presented in Table contained within a model showing determinants of executives joining Twitter following regression equation For Hypothesis we expect to observe a positive and significant coefficient on Post SEC Based on the full sample model presented in Table it appears that the SEC report did not lead to an increase in the number of executives on Twitter The model presented in Table also provides a determinants model of some of the characteristics of executives on Twitter First as one might expect younger CEOs are more likely to be on Twitter With respect to firm characteristics we find that executives at large firms and those with higher growth high MTB are more likely to be on Twitter Both types of firms are likely to be more visible which may explain the influence of these factors We also find that the number of tweets the firm posted in a quarter is a strong positive predictor of an executive joining Twitter indicating that a firms disclosure or dissemination strategy on Twitter may impact its investors approach to social media Lastly executives in the information technology and communications services industries are more likely to join Twitter likely due to the hightech nature of the industries Page Executives tweet determinants Univariate statistics for our daily sample restricted to executives on Twitter are presented in Table Panel A This sample consists of firm days with managers posting an average of tweets per day and their firms post around tweets per day Among managers on Twitter around are CEOs and are CFOs and approximately of their firms are also on Twitter In terms of what managers tweet about the most common category is nonfinancial followed by other Other nonfirm related tweets account for of tweets by executives Firms follow a similar pattern where nonfinancial tweets are still the most common followed by other tweets One other interesting phenomenon is the relationship between executives and their firms having Twitter accounts While rare there are some instances where an executive has a Twitter account but their firm does not To explore the variation in executives and firms having Twitter accounts we split our sample along these two dimensions and examine the four determinants that drove executive adoption of Twitter Panel B of Table presents these results First we document that executives on Twitter are indeed younger on average Furthermore among firms whose executive is not on Twitter the executive tends to also be younger Second we document that size appears to only be a factor in an executive joining Twitter when the firm itself is also on Twitter differenceindifference t statistic pvalue Also consistent with our findings from Table we find that firms with executives on Twitter tweet significantly more per day Lastly We find that tech firms are more likely to be on Twitter and have their executive on Twitter Furthermore of firms with both the firm and executive on Twitter are in the tech industry differenceindifference t statistic pvalue Page A univariate analysis for Hypothesis is present in Panel B of Table The first three columns of the panel present the composition of tweets by category across the full sample while the latter columns present the composition of tweets in a narrow window of four months around the SEC report For the full sample we see a large increase in all types of tweets except financial tweets which has a marginally significant increase As the SEC report should be most salient for posting financial information on Twitter this result does not strongly support Hypothesis Likewise for the window around the SEC report we observe no univariate change in tweeting across any category of information As such we do not observe much evidence of differences in what executives are posting that we can attribute to the SEC report itself To further examine the effect of the SEC report on tweet content we run regression equation using Post SEC as the event In a multivariate setting presented in Panel A of Table we observe that the SEC report did not drive any change in the overall level of tweeting nor any change in tweeting of any specific category of tweets As such these results indicate that against Hypothesis the SEC report did not appear to have any effect on the content of firm tweets Before moving on to Hypothesis we take note of some interesting determinants of executive tweets that are presented in Table Panel B First we note that firm tweets about the same topic are a positive driver of every topic We also note that for financial tweets older executives are more likely to post them and in fact older executives do not post a significantly different level of other categories of tweets from younger executives Also we note that an executive whose firm is not on Twitter is significantly more likely to post both financial and nonfinancial tweets as evidenced by the negative coefficient on Firm on Twitter Page Our tests of Hypothesis are presented in Table In Panel A we examine the impact of earnings announcements and earnings conference calls on executive tweets We observe a large increase in the number of financial tweets posted around these events with no increase in any other category Given the financial nature of earning announcements and earnings conference calls this presents strong evidence for Hypothesis Panel B presents results for Hypothesis while examining SEC filings Consistent with the findings from Crowley et al that firms post more financial tweets around and filings we find that and filings lead to an increase in executives posting financial information on Twitter Furthermore filings lead to an increase in posting of every category of tweet likely due to the varied nature of what can be disclosed in an filing Echoing the results for filings Panel C of Table presents the results of our press release test Like filings press releases can contain many different types of disclosure or dissemination and consequently appear to increase all categories of tweeting by executives Similarly Panel D shows that executives also generally react to the amount of news coverage they receive This test shows that executives also respond to information events that are external to the firm Overall the findings in Table Panels A through D present strong results in favor of Hypothesis While it appears that content posted by executives did not change in response to the SEC report Hypothesis we find ample evidence that executives tweeting behavior is responding to a wide variety of information events including bother voluntary and mandatory events as well as both internal and external events Our results are consistent and robust when separately examining the impact of earnings announcements or earnings conference calls Page Market reaction to executives tweets Given the results for Hypotheses it appears credible that executive tweets could contain useful information for investors To test we run regression equation and present the results in Table Panel A When we examine financial tweets in the first column we see a strong positive coefficient for both executive tweets as well as firm tweets For executive tweets on day we observe that an increase of financial tweet by an executive leads to an increase in absolute abnormal return of Furthermore this result is robust to the other three windows used though most of the effect comes from time For firm tweets we observe a similar pattern but with less economic significance only a increase in absolute abnormal return per tweet and there is seemingly no effect on contemporaneous returns Thus it appears that financial tweets by executives provide a significant amount of useful information to market participants However we find relatively little reaction to nonfinancial and other tweets indicating that most useful information to investors is contained only in executives financial Tweet similarity To test Hypothesis we use a fiscal quarterbased sample Table presents univariate statistics for the quarterly sample including the measures of tweet similarity between executives and firms We find a reasonable amount of variation in the measures with Tweet similarity varying from for the percentile up to at the percentile Furthermore we note that the mean median of Tweet similarity is Table presents our test of Hypothesis following regression equation The first two columns present results using the full sample of executives with tweets overlapping their firms As these three models together could be construed as testing a joint hypothesis of statistical tests we note that our presented result for financial tweets is still statistically significant at the level after applying a Bonferroni correction Page tweets In the coefficient on Tweet similarity is positive and statistically significant this indicates that when a firm has an executive that tweets more similarly to what the firm is tweeting the firms Tobin Q rises on This result supports Hypothesis and is consistent with tweet similarity capturing the concept of executive effort As a firms tweets can form a baseline expectation for what the firm is currently focused on and based on prior literature demonstrating that social media posts provide some insight into the mind or thoughts of users by comparing an executives own tweets to their firm we can capture whether the executive appears to be focusing on the same concepts as the firm thus capturing at least some part of executive effort To rule out the alternative story of high similarity quarters being a product of a firms marketing or PR functions we present results removing the executivefirmquarters in which an executive sent a tweet from a Twitter client that is marketingrelated While this restriction further decreases the size of our sample for this test our results continue to hold for this sample This provides strong evidence that the marketing or PR sending tweets is not a correlated omitted variable driving our results from the full sample Robustness checks Specific types of news articles In our tests of Hypothesis we examined the number of news articles released by Dow Jones owned sources This test included news on any topic It is also interesting to consider specific news subjects such as news about financials or MA and their differential impact on executives tweeting behavior Following Crowley et al we look at news articles on specific subjects constructing counts of articles for each subject for each firm on each trading day These news Our results are robust to an alternative specification of our similarity measure using Manhattan distance norm as the distance measure underlying our similarity computations Using this alternative metric all results are inferentially identical Page subject count measures include financial news acquisition news management forecast news executive news and analyst forecast news In untabulated tests we examine the impact of each of the five news subjects on executive tweets We expect the first three subjects financial acquisition and management forecast news to impact executives financial tweeting Executive news is not typically financial in nature and any impact should instead be concentrated in nonfinancial tweets Lastly based on the lack of firm on Twitter reaction to analyst forecasts documented in Crowley et al we would similarly expect managers to not react as well Our results for executives financial tweets are in line with expectations executives increase tweeting of financial information around financial news acquisition news and management forecast news Likewise we observe a positive impact of acquisition news on nonfinancial tweeting as well though acquisition news also appears to positively impact other nonbusinessrelated tweeting We also do not find any reaction to analyst forecast news as expected Interestingly however we also do not observe any reaction to executive news In a supplementary analysis we examine executives reaction to management forecast news based on the sentiment of the news positive negative or neutral following Crowley et al and find that executives only increase financial tweeting are positive and neutral news showing that there appears to be some bias in what managers are disclosing or disseminating on Twitter This result further differentiates the disclosure behavior of executives from their firms as firms are more likely to release financial tweets are positive and negative news and less so around neutral news Crowley et al Page Conclusion This paper examines the tweeting behavior of executives CEOs and CFOs We find that there is a limited impact of the SEC Report on executives rate of joining Twitter as well as on the content of their tweets In contrast we document that managers post financialrelated tweets around important financial events such as earnings announcements earnings conference calls and and filings We also document executives posting both financial and nonfinancial businessrelated information on Twitter around more general information events for their firms such as filings press releases and news media We then find that tweets by executives that are financial in nature are reacted to by the market and are reacted to more strongly than financial tweets by firms We then extend our analysis to create a novel measure of executive effort based on the similarity between executives and firms tweets We document that our tweet similarity based measure of executive effort positively predicts future growth opportunities for firms consistent with higher tweet similarity executives exerting more effort for their firm Page References Arya S and D M Mount Approximate Nearest Neighbor Queries in Fixed Dimensions In Proceedings of the Fourth Annual ACMSIGACTSIAM Symposium on Discrete Algorithms Bartov E L Faurel and P S Mohanram Can Twitter Help Predict FirmLevel Earnings and Stock Returns Accounting Review Bertomeu J P Ma and I Marinovic How Often Do Managers Withhold Information Working paper University of California San Diego University of Minnesota and Stanford University Available at Blankespoor E G S Miller and H D White The Role of Dissemination in Market Liquidity Evidence from Firms Use of TwitterTM Accounting Review Blei D M A Y Ng and M I Jordan Latent Dirichlet Allocation J Mach Learn Res March Bollen J H Mao and X Zeng Twitter Mood Predicts the Stock Market Journal of Computational Science Brown N C R M Crowley and W B Elliott What Are You Saying Using Topic to Detect Financial Misreporting Working paper University of Illinois UrbanaChampaign and Singapore Management University Available at Cer D Y Yang SY Kong N Hua N Limtiaco R St John N Constant M GuajardoCespedes S Yuan C Tar YH Sung B Strope and R Kurzweil Universal Sentence Encoder ArXiv preprint Available at Correia S A Feasible Estimator for Linear Models with MultiWay Fixed Effects Working paper Duke University Available at httpscorreiacomresearchhdfepdf Correia S P Guimares and T Zylkin Ppmlhdfe Fast Poisson Estimation with HighDimensional Fixed Effects ArXiv preprint Available at Crowley R M W Huang and H Lu Discretionary Dissemination on Twitter Working paper Singapore Management University Hong Kong Polytechnic University and University of Toronto Available at Crowley R M W Huang H Lu and W Luo Do Firms Tweet Social Responsibility Evidence from Machine Learning Analysis Working paper Singapore Management University Hong Kong Polytechnic University University of Toronto and Peking University Curtis A V J Richardson and R Schmardebeck Investor Attention and the Pricing of Earnings News Working paper University of Washington and University of Arkansas Available at De Choudhury M S Counts and E Horvitz Social Media As a Measurement Tool of Depression in Populations In Proceedings of the Annual ACM Web Science Conference Dyer T M Lang and L SticeLawrence The Evolution of Textual Disclosure Evidence from Latent Dirichlet Allocation Journal of Accounting and Economics Page Eichstaedt J C H A Schwartz M L Kern G Park D R Labarthe R M Merchant S Jha M Agrawal L A Dziurzynski M Sap C Weeg E E Larson L H Ungar and M E P Seligman Psychological Language on Twitter Predicts CountyLevel Heart Disease Mortality Psychological Science Hope OK D Hu and H Lu The Benefits of Specific RiskFactor Disclosures Review of Accounting Studies Huang A H R Lehavy A Y Zang and R Zheng Analyst Information Discovery and Interpretation Roles A Topic Modeling Approach Management Science Java A X Song T Finin and B Tseng Why We Twitter Understanding Microblogging Usage and Communities In Proceedings of the WebKDD and SNAKDD Workshop on Web Mining and Social Network Analysis Jung M J J P Naughton A Tahoun and C Wang Do Firms Strategically Disseminate Evidence from Corporate Use of Social Media Accounting Review Kothari S P S Shu and P D Wysocki Do Managers Withhold Bad News Journal of Accounting Research Kumar V V Bhaskaran R Mirchandani and M Shah Practice Prize WinnerCreating a Measurable Social Media Marketing Strategy Increasing the Value and ROI of Intangibles and Tangibles for Hokey Pokey Marketing Science Lee L F A P Hutton and S Shu The Role of Social Media in the Capital Market Evidence from Consumer Product Recalls Journal of Accounting Research Lin KY and HP Lu Why People Use Social Networking Sites An Empirical Study Integrating Network Externalities and Motivation Theory Computers in Human Behavior Mao Y W Wei B Wang and B Liu Correlating SP Stocks with Twitter Data In Proceedings of the First ACM International Workshop on Hot Topics on Interdisciplinary Social Networks Research Margiotta M M and R A Miller Managerial Compensation and the Cost of Moral Hazard International Economic Review Mikolov T K Chen G Corrado and J Dean Efficient Estimation of Word Representations in Vector Space ArXiv preprint Available at Miller G S and D J Skinner The Evolving Disclosure Landscape How Changes in Technology the Media and Capital Markets Are Affecting Disclosure Journal of Accounting Research Murphy K J Executive Compensation Where We Are and How We Got There In Handbook of the Economics of Finance Pennington J R Socher and C Manning Glove Global Vectors for Word Representation In Proceedings of the Conference on Empirical Methods in Natural Language Processing EMNLP Pew Research Center Social Media Use Demographics and Statistics March Available at Page Sprenger T O A Tumasjan P G Sandner and I M Welpe Tweets and Trades The Information Content of Stock Microblogs European Financial Management Tetlock P C All the News Thats Fit to Reprint Do Investors React to Stale Information The Review of Financial Studies Toubia O and A T Stephen Intrinsic vs ImageRelated Utility in Social Media Why Do People Contribute Content to Twitter Marketing Science Securities and Exchange Commission SEC Commission Guidance on the Use of Company Web Sites Release Nos Washington DC SEC Report of Investigation Pursuant to Section of the Securities Exchange Act of Netflix Inc and Reed Hastings Release No Washington DC SEC Zhao W X J Jiang J Weng J He E P Lim H Yan and X Li Comparing Twitter and Traditional Media Using Topic Models In Advances in Information Retrieval Page Appendix A Variable Definitions Variable Dependent variables Exec on Twitter Exec topic tweets MMRetwindow Tobins Definition An indicator for if the executive has both opened a Twitter account and posted hisher first tweet by the given date The number of tweets posted on the given day daily tests or quarter quarterly tests by the executive CEO or CFO Market model return calculated using a daily frequency using beta with respect to the SP index calculated over the prior quarter trading days Tobins Q for the next fiscal quarter calculated using Compustat Quarterly items as winsorized at and Independent variables and filing An indicator for if a or filing was released on the given trading day based on FINDEXDATE from WRDS SEC Analytics filings The number of filings released on a given trading day based on FINDEXDATE from WRDS SEC Analytics Earnings event An indicator for if the firm released an earnings announcement annual or quarterly on the given trading day IBES or conducted an earnings conference call on the given trading day Capital IQ Exec topic tweets The number of tweets posted on the given day daily tests or quarter quarterly tests by an executive CEO or CFO News articles The number of unique news articles in Dow Jones owned publications released about the firm on a given trading day using intraday data from Ravenpack Dow Jones Edition filtering for at least relevance and a unique RPSTORYID Post SEC if the date is at or later than April else Press Releases The number of unique press releases released by the firm on a given trading day using intraday data from Ravenpack PR Edition filtering for at least relevance and a unique RPSTORYID Tweet Similarity The similarity of tweets by an executive to those within a day window by the executives firm averaged across the fiscal quarter Distance is measured as the minimum Euclidean distance norm between the USE vector representing the executives tweet and the USE vectors representing the executives firms tweets Similarity is given by Distance scaling to the range of where is most similar Page Control variables executive Executive age CEO CFO Control variables financial Debt MTB ROA Size Tobins Q Firm tweet content variables Firm topic tweets TweetsFirm Control variables Twitter Firm on Twitter logFollowersEntity logFollowingEntity Recent tweetsEntity Age of the CEO in years Execucomp Indicator for if the executive is the CEO of the company Execucomp Indicator for if the executive is the CFO of the company Execucomp Debt as a portion of assets calculated as total liabilities ltq divided by total assets atq winsorized at and Market to book value calculated as market value mkvaltq divided by total assets atq winsorized at and Return on assets calculated as Net income niq divided by total assets atq winsorized at and Log of assets atq winsorized at and Tobins Q for the next fiscal quarter calculated using Compustat Quarterly items as mkvaltq ltqatq ltq winsorized at and The number of tweets posted on the given day daily tests or quarter quarterly tests by the firm The number of tweets posted in a given quarter by the firm across all categories Note Excluding Firm on Twitter and Recent tweets these control variables are generally backfilled due to the pointintime nature of the data Firm data is first available as of January CEO data is first available as of January and CFO data is first available as of June For days missing these measures after the first date of availability the most recent previous nonmissing observation is used An indicator for if the firm associated with the executive is both on Twitter and has tweeted at least once by the given date The log of one plus the number of Twitter followers that the entity firm or executive The log of one plus the number of twitter accounts that the entity firm or executive is following on Twitter The number of tweets posted by the entity firm or executive over the window covering the prior five trading days Page logTotal tweetsEntity Fixed Effects Executive Firm Industry Month Year The log of one plus the number of tweets that the entity has posted up to the given date The coperrol ID of the executive as provided by Execucomp The gvkey of the firm as provided by Compustat Quarterly The GIC Sector of the firm as provided by Compustat Quarterly The month of the trading date January February The year of the trading date Page Appendix B Tweet examples by category Financial Omar Ishrak MedtronicCEO CEO of Medtronic Tweet ID Continuing to execute in both our product SGA cost reduction initiatives will provide consistent EPS leverage MDTEarnings Mike Jackson CEOMikeJackson CEO of AutoNation Tweet ID With ample credit great products strong Toyota Honda inventorywe raised our sales forecast to mid million vehicles Marcelo Claure marceloclaure CEO of Sprint Tweet ID was a transformational year Positive operating income for the first time in years httpstcohxEkNDlpWO Nonfinancial Mark T Bertolini mtbert CEO of Aetna Tweet ID Arriving in Atlanta A day meeting with customers is better than any day in the office But I do love all the folks back in Hartford too o Jim Whitehurst Jwhitehurst CEO of Redhat Tweet ID Great time chatting with our Customer Platform team Keep up the great work LifeAtRedHat httpstcoOTfvfqhmfa Carl Bass carlbass CEO of Autodesk Tweet ID Giving keynote tomorrow at Talking about the good bad of and the future of software Other Bob Carrigan BobCarrigan CEO of Dun Bradstreet Tweet This wont play well in the home office but the Flyers are making an amazing comeback against the Bruins Series now tied Go Philly Carl Bass carlbass CEO of Autodesk Tweet ID Another great day of spring skiing in the Alps Tony Thomas TonyThomasWIN CEO of Windstream Tweet ID Hail uncool Mother Nature showing her fury Page Appendix C USE Method Universal Sentence Encoder USE is an algorithm developed by Cer et al for generating embeddings of sentences An embedding is vector that can represent a meaning within an abstract highdimensional vectors space Other examples of embeddings include word embedding algorithms like Mikolov Chen Corrado and Dean and GloVe Pennington Socher and Manning which are both used in accounting in Brown et al While a word embedding algorithm maps words to their meanings a sentence embedding algorithm like USE takes this a step farther mapping whole sentences to the meaning of the sentences themselves In the case of USE this can be accomplished in two different ways using a Deep Averaging Network DAN or a transformer architecture For our implementation we leverage the pretrained DANbased model provided on TensorFlow The USE methodology converts each tweet in our data into unit vectors that map somewhere into a vector space Within this space the closer two vectors are the more similar the meaning of the tweets the vectors represent To calculate the distance between vectors our primary measure uses Euclidean distance as this is the default distance metric used by the USE model in TensorFlow For robustness we also calculate distances To convert to similarity scores we normalize the distances such that the theoretical maximum distance becomes For distance we normalize by dividing by as the farthest distance under an norm for any anydimensional unit vectors is For distance we normalize by dividing by as the maximum distance between ndimensional unit vectors can be calculated as Then we subtract the normalized distance from in order to convert to similarity The DAN based pretrained USE algorithm we use is available at Page Example Twitterlike text similarities Note This figure shows some Twitterlike text a mix of tweets shortened tweets and contrived text for illustrative purposes The first second third three messages represent financial nonfinancial other content For financial note how the algorithm can pick up the similarity between earnings losses in the context of yearoveryear and operating income as well as how it applies a slightly higher similarity to the tweets that are both positive as compared to mixes of positive and negative For nonfinancial note how it understands that the third message is relatively abstract and could sensibly link to the other two examples yet the other two themselves by being more specific receive a relatively lower similarity Lastly note how for the first and third messages it can tell that both are about hockey The first message only mentions a couple of team names Flyers Bruins as hints that the message is hockeyrelated yet it strongly matches this message with the more generic hockeyrelated third message Page Figure Twitter accounts and Tweets by year Page Table Executives joining Twitter VARIABLES Post SEC Fiscal quarter sample Firm on Twitter Size ROA MTB Debt Executive age logFollowersFirm logFollowingFirm TweetsFirm logTotal tweetsFirm Industry FE Subjects Failures Joins Observations Yes Note This table presents the results of regression equation on the fiscal quarter sample of executives use a Cox proportional hazards model Failure under the Cox model is taken to be the fiscal quarter in which an executive first joins Twitter has posted at least one tweet Z statistics are presented in parentheses and significance is denoted as follows denotes p denotes p and denotes p Page Table Univariate Statistics daily sample of executives on Twitter Panel A Univariate statistics Variable N Mean SD Manager on Twitter Post SEC Firm on Twitter Size ROA MTB Debt CFO CEO Executive age logFollowersFirm logFollowingFirm Recent tweetsFirm logTotal tweetsFirm logFollowersExec logFollowingExec Recent tweetsExec logTotal tweetsExec Page Panel B Differences across significant determinants of executives joining Twitter Variable Executive Age Exec not on Twitter Exec on Twitter Difference Firm not on Twitter Firm on Twitter DID Other Variable Size Exec not on Twitter Exec on Twitter Firm not on Twitter Firm on Twitter DID Other Variable TweetsFirm Firm not on Twitter Exec not on Twitter Exec on Twitter Difference Difference Exec not on Twitter Exec on Twitter Firm not on Twitter Firm on Twitter DID Firm on Twitter Variable Tech Industry Other Difference Note Panel A presents univariate statistics of the sample of trading dayexecutivefirm observations restricted to trading days where the executive had an active Twitter account with at least tweet posted that day or prior Panel B presents twobytwo splits on Executives and firms having Twitter accounts The presented variables are those that are determinants of executives having a Twitter account as shown in Table Page Table Executives tweeting behavior and the SEC Report Panel A Determinants of executive tweeting VARIABLES Post SEC Firm topic tweets Size ROA MTB Debt Executive age Financial Nonfinancial Other Firm on Twitter logFollowersFirm logFollowingFirm Recent tweetsFirm logTotal tweetsFirm logFollowersExec logFollowingExec Recent tweetsExec logTotal tweetsExec Constant Firm FE Yes Yes Yes Exec FE Yes Yes Yes Year FE Yes Yes Yes Month FE Yes Yes Yes Page Pseudo Rsq Sample size Dropped observations Observations Panel B Univariate differences in tweeting before and after the SEC Report Tweet type Before After Diff t Before mo After mo Diff t Financial Nonfinancial Other Observations Note Panel A presents the results of regression equation on the daily sample of executives using a Poisson pseudo maximum likelihood PPML regression The dependent variables are counts of the number of tweets posted by a manager on a given trading day Z statistics are presented in parentheses and significance is denoted as follows denotes p denotes p and denotes p Panel B presents univariate differences in the number of tweets posted for each of the three tweet categories financial nonfinancial other before and after the SEC Report The first three columns report the differences across the tradingday sample restricted to observations where managers had an active account with at least tweet on the trading day or prior The latter three columns further restrict the sample to the two months before and after the SEC Report ie from February to June Page Table Executive response on Twitter to information events Panel A Executive tweets and earnings announcementsconference calls VARIABLES Financial Nonfinancial Other Earnings Yes Yes Yes Event All Controls Firm Exec year month FE Pseudo Rsq Yes Yes Yes Panel B Executive tweets and SEC filings VARIABLES Financial and All Controls Yes Yes Yes Firm Exec year month FE Yes Yes Yes filing filings Pseudo Rsq Nonfinancial Other Panel C Executive tweets and press releases VARIABLES Financial Nonfinancial Other Press Releases All Controls Yes Yes Yes Firm Exec year month FE Yes Yes Yes Pseudo Rsq Panel D Executive tweets and news articles VARIABLES Financial Nonfinancial Other News articles All Controls Yes Yes Yes Firm Exec year month FE Yes Yes Yes Pseudo Rsq Note This table presents the results of regression equation on the daily sample of executives using a Poisson pseudo maximum likelihood PPML regression The dependent variables are counts of the number of tweets posted by a manager on a given trading day Z statistics are presented in parentheses and significance is denoted as follows denotes p denotes p and denotes p Page Table Market response to executive tweets VARIABLES Exec topic tweets Financial Nonfinancial Other Firm topic tweets Size ROA MTB Debt Executive age Firm on Twitter logFollowersFirm logFollowingFirm Recent tweetsFirm logTotal tweetsFirm logFollowersExec logFollowingExec Recent tweetsExec logTotal tweetsExec Constant Firm FE Exec FE Year FE Month FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Adjusted Rsq Observations Page Note This table presents the results of regression equation on the daily sample of executives using a linear regression with high dimensional fixed effects HDFE The dependent variable in all panels is absolute market model return on day t Exec topic tweets and Firm topic tweets capture counts of different tweet types for each column financial tweets nonfinancial tweets and other tweets in the first second and third column respectively t statistics are presented in parentheses and significance is denoted as follows denotes p denotes p and denotes p Page Table Univariate Statistics quarterly sample of executives and firm both on Twitter Tweet similarity TweetsExec TweetsFirm Tobins Tobins Q Size ROA MTB Debt CEO CFO Executive age logFollowersFirm logFollowingFirm logTotal tweetsFirm logFollowersExec logFollowingExec logTotal tweetsExec N Mean SD Note This table presents univariate statistics of the sample of fiscal quarterexecutivefirm observations restricted to fiscal quarters with at least of both executive and firm tweets within a window of each other Page Table Tweet similarity and future firm growth opportunities Tobins VARIABLES Tweet similarity TweetsExec TweetsFirm Tobins Q Size Full sample No marketing client usage ROA MTB Debt Executive age logFollowersFirm logFollowingFirm logTotal tweetsFirm logFollowersExec logFollowingExec logTotal tweetsExec Constant Firm FE Yes Yes Exec FE Yes Yes Year FE Yes Yes Month FE Yes Yes Adjusted Rsq Sample size Dropped observations Observations Page Note This table presents the results of regression equation on the quarterly sample of executives requiring at least firm and executive tweet within a day window of each other using a linear regression with high dimensional fixed effects HDFE The second column further restricts the sample to exclude any executivefirmquarter in which the executive posted a tweet using a marketingoriented Twitter client The dependent variable in all panels is quarter ahead Tobins Q t statistics are presented in parentheses and significance is denoted as follows denotes p denotes p and denotes p Page