SOCIAL MEDIA DISCLOSURE AND ANALYSTS AS INFORMATION INTERMEDIARIES By Kwangjin Lee A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Business Administration Doctor o f Philosophy 2018 ABSTRACT SOCIAL MEDIA DISCLOSURE AND ANALYSTS AS INFORMATION INTERMEDIARIES By Kwangjin Lee Using a sample of S&P 500 firms over the period 2012 2014 and Twitter data, I investigate the effect of social media disclosure on fin ancial analysts as information intermediaries. On one hand, social media is a low - cost mechanism for direct communications from the firm to its investors, so may substitute for information intermediation by analysts. On the other hand, following Mosaic the ory (Pozen , 2005), analysts (i.e., the crowd of the experts) have a comparative advantage at placing relevant pieces of information into the broader mosaic, implying that the importance of analysts as information intermediaries may increase with the volume forecast error. This finding is consistent with analysts us ing social media information as a com plement to other information sources, providing richer analyses to investors. I also find that the market reaction to analysts' forecast revisions varies positively with the level of social media activity. Together, these findings suggest that social media disclosure serves as a complement to information processing by analysts, as opposed to a substitute. This paper contributes to the literature on financial analysts by providing evidence that even in the era of social media disclosure, the role of analysts as information intermediaries remains important for the efficient functioning of capital markets. It also contributes to the literature on the impact of social media on capital markets by providing a deeper understanding of the impact of unregulated and u nstructured disclosure on the general information environment of financial markets. iii This dissertation is dedicated to my beloved family. iv ACKNOWLEDGEMENT S I am deeply indebted to my dissertation guidance committee members who provided continuous s upport and guidance throughout the Ph.D. program : Chris Hogan (Co - Chair), Marilyn Johnson (Co - Chair), Ranjani Krishnan, and Hang Nguyen. I am grateful to Chris Hogan and Marilyn Johnson for their advice, insight, and consideration throughout not only the d issertation process but also the entire doctoral program. I would also like to express my sincere gratitude to Ranjani Krishnan and Hang Nguyen for their feedback and comments. In addition, I would like to thank Bok Baik, Sung Chung, Kyonghee Kim, and Kyu ngran Lee for their helpful suggestions and feedback. I am also grateful to James Anderson, Maurice Atkinson, and Luke Weiler for their time and help in developing the financial key words dictionary. Furthermore, I want to thank Joohyung Lee, Sangmok Lee, and Jason Shin for being collegial colleagues and true friends during my doctoral study. I am also thankful to my parents, Ho - In Lee and Hae Jeung Lim, for their unwavering love and unconditional support. They have always encouraged me with love throughou t my life. My sisters Hannah, Yoonnah, and Ginnah have also encouraged me with their own experiences as they have gone through this process themselves. Last, but certainly not least, I want to thank my marvelous wife, Soo Jeong Hong. Without her endless lo ve, support, and encouragement, I would not have reached this point. She has always been there for me and shared my joy and tears from the beginning of my doctoral study. Above all, I thank God for blessing me with His endless love and grace. v TABLE OF CONTENTS LIST OF TABLES v ii LIST OF FIGURES v i ii 1. INTRODUCTION 1 2. BACKGROUND AND HYPOTHESES 6 2.1. Role of Information Technology in Corporate Disclosure 6 2.2. Changes in SEC Disclosure Regulation 7 2.3. Research R elated to Business Communication 9 2.4. Prior Literature on the Role of Twitter in Business Communication 12 2.5. The Role of Financial Analysts as Information Intermediaries 16 2.6. Development of Hypotheses 19 3. D ATA AND RESEARCH DESIGN 23 3.1. Data 23 3.2. Measures of Social Media Disclosure 24 3.3. Measures of A nalyst C overage and P roperties of F orecasts 25 3.4. Research Design 25 3.4.1. Analyst Following Model 25 3.4.1.1. Control V ariables for Analys t Following Model 27 3.4.2. Forecast Properties Model 29 3.4.2.1. Control V ariables for Forecast Properties Model 29 3.4.3. Market Reaction to Analyst Forecast Revisions Model 3 1 3.4.3.1. Control V ari ables for Market Reaction to Analyst Forecast Revisions Model 33 4. EMPIRICAL RESULTS 3 4 4.1. Descriptive Statistics 3 4 4.2. Profile A nalysis 3 5 4.3. Corr elation A nalysis 3 6 4.4. Regression R esults 3 7 4.4.1. Analyst Following 3 7 4.4.2. Forecast Error 3 8 4.4.3. Forecast Dispersion 39 4.4.4. Market Reaction to Forecast Revisions 4 0 5. SUPPLEMENTAL ANALYSIS 4 3 vi 5 .1. Analysis of Analyst Following, Analyst Forecast Errors, and Analyst Forecast Dispersion for Subsamples of Compa nies in Consumer - Oriented and Non - Consumer - Oriented Industries 4 3 5 .2. Analysis of the Market Response to Analyst Forecast Revision s using 3 - Day CAR s 4 6 5 .3. Analysis of Market Response to Analyst Forecast Revision s for Subsamples with and w ithout Pr ior Management Forecasts 4 7 6. CONCLUSION 49 APPENDICES 5 1 APPENDIX A Variable Definitions and Data Source s 5 2 APPENDIX B APPENDIX C In dustry Classification s Tables 5 6 5 8 REFER E NCES 8 1 vii LIST OF TABLES Table 1 . Sample Selection 59 Table 2 . Descriptive Statistics 6 0 Table 3 . Profile Analysis 6 1 Table 4 . Pairwise Correlation s among Variables Used in the Analysis 6 3 Table 5 . Analyst Following Regression Analysis 6 5 Table 6 . Analyst Forecast Error Regression Analysis 6 7 Table 7 . Analyst Forecast Dispersion Regression Analysis 69 Table 8 . Descriptive S tatistics for Variables Used in the Market Resp onse Analysis 7 1 Table 9 . Regression Analysis of the Market Response to Analyst Forecast Revisions Using 2 - Day CAR s 7 2 Table 10 . Analyst Following Regression Analysis for Subsamples of Consumer - Oriented Industries and Non - Consumer - Oriented Ind ustries 7 3 Table 11 . Analyst Forecast Error Regression Analysis for Subsamples of Consumer - Oriented Industries and Non - Consumer - Oriented Industries 7 5 Table 12 . Analyst Forecast Dispersion Regression Analysis for Subsamples of Consumer - Oriented Indus tries and Non - Consumer - Oriented Industries 7 7 Table 1 3 . Regression Analysis of the Market Response to Analyst Forecast Revisions Using 3 - day CAR s 79 Table 14 . Regression Analysis of the Market Response to Analyst Forecast Revisions for Subsamples wit h and without Prior Management Forecasts Using 2 - Day CARs 8 0 viii LIST OF FIGURES Figure 1 . Number of Tweets by Quarter, 2012 - 2014 3 6 1 1. INTRODUCTION Due to innovation s in information technology, there have been enormous changes in business commun ication practices over the past decade. The Internet lowered the cost of information dissemination and increased the velocity at which information travels. Similar to the introduction of the Internet, the emergence and widespread adoption of social media i ncreased information flow by facilitating interaction between Websites and information users. The use of social media not only facilitate s the dissemination of news but also encourages participation, collaboration, and information sharing (Culnan, McHuch, and Zubillaga, 2010; Chen, De, Hu, and Hwang , 2014 ; Kane , Alavi, Labianca, and Borgatti , 2014) . Prior research show s that social media is an efficient conduit for disseminating information to financial market s and affects investor behavior (Antweiler and F rank, 2004; Barnes, Lescault, and Wright, 2013; Lee, Hutton, and Shu, 2015; Chen, Hwang, and Liu, 2016; Bratov, Faurel, and Mohanram, 2017). Howev er, there is limited evidence on how social media affects sophisticated information intermediaries such as fin ancial analysts. But , the same communication may also complement the information processing activities of analysts, who hav e a comparative advantage in positioning bits of information in the broader information mosaic (Poze n, 2005 ; Yeldar, 2012 ). Lehavy, Li, and Merkley ( 2011 ) , Cao, Keskek, Myers, and Tsang ( 2014 ) , and Lev and Gu ( 2016 ) , for example, argue that the increased v olume and complexity implies an expansion in the role of financial analysts as intermediaries . Thus, the net effect of social media disclosure on the importance of analysts as information intermediaries is a n open question that I address in this paper. 2 I use a sample of S&P 500 firms over the period 2012 - 2014 and Twitter data to examine whether tweets by and about a firm are associated with analyst following , properties of analyst forecasts, and the magnitud e of the market reaction to analyst forecast revisions . I measure the amount of social media disclosure by the firm as the number of tweets from the Twitter account W ebsite and the amount of social media disclosure about the fi rm by 1 I further split tweets by the firm into those that discuss financial topics and those that do not. My results indicate that analyst following is larger and forecast errors are sm aller, the larger the number of financial tweets by the firm . A nalyst following is smaller, the larger the number of nonfinancial tweets by the firm. Forecast errors are larger, the larger the number of tweets by the public , while t he volume of tweets by the public is not significantly associated with analyst following. Forecast dispersion is not significantly associated with social media disclosure. Collectively, the se findings suggest that only financial social media disclosure provides timely, value - rel evant information to analysts. One interpretation of these results is that financial tweets by the firm reflect supply - side factors of information . When more information is supplied by the firm, analyst following is larger and analyst forecasts are more ac curate. In contrast, tweets by the public reflect the demand for information. But, demand for information, per se, does not imply more accurate forecasts. In fact, forecasts are less accurate when the volume of public tweets is large, implying that analyst s may 1 Cashtags are stock ticker symbols that are prefixed with a dollar sign. For example, tweet s about Microsoft would use $MSFT . 3 rely on misleading info rmation released by the public beliefs are heterogeneous . I also report evidence on the relation between social media and the market response to analyst forecast revisions. In this anal ysis, I regress cumulative abnormal returns ( CARs ) over the two - day window beginning with the revision release date on the average analyst forecast revision, the number of financial tweets by the firm over the revision period and the interaction between th e revision and the number of financial tweets by the firm. The coefficients on the revision variable and the interaction term are both positive and significant. In contrast, the coefficient on financial tweets is insignificant. Consistent with prior resear ch ( Brown, Foster, and Noreen, 1985; Klein, 1990; Lys and Sohn, 1990; Beyer, Cohen, Lys, and Walther, 2010 ) , the positive coefficient on the revision variable indicates that revisions contain value relevant information cash flows . T he positive coefficient on the interaction term indicates that revisions are more informative, the larger the number of financial tweets released by the firm. The coefficient on the number of tweets is insignificant, consistent with the previous impounding of the information in those tweets. T o validate the main model results and to provide enhanced perspectives about the main findings , I re - analyze the impact of social media on financial analysts as information intermediaries for subsample s of firms that are members of consumer - oriented industries and non - consumer - oriented industries . 2 In the se supplemental subsample analyses, I consistently find that financial tweets by the firm provide additional value relevant information for financial analysts. 2 Social media disclosure focus es jointly on investors and consumers, so often includes both financial information and advertising. While the difference in audience can increase the risk of misinterpretatio n, nonfinancial information such as advertising has the potential to engage investors as well (Madsen and Niessner, 2016). 4 Furthe r, for firms in consumer - oriented industries, I also find that nonfinancial tweets by the firm provide value relevant information for financial analysts. I also re - run the market response test s . T o examine the sensitivity of my main results to the length of my event window, I regress CARs over the three - day window centered on the revision release date on the average analyst forecast revision, the number of financial tweets by the firm over the revision period, and the interaction between the revision and t he number of financial tweets by the firm. To control for the effect of management forecasts issued between forecast revisions , I also re - estimate the return analysis for subsample s of observations with and without a prior management forecast between the previous and the current forecast revisions. 3 In the two additional sensitivity analyses, I consistently find that the interaction between the forecast revision and the number of financial tweets has a positive and significant coefficient. These results i ndicate that revisions are more informative, the larger the number of financial tweets released by the firm and impl y that social media disclosure supplem ent s the information used by financial analysts. From this analysis, I also find that financial tweets have a larger impact on forecast revisions in the absence of concurrent management forecast s . This finding may indicate that investors rely more on financial information shared through social media by firms when there is less information from management i n other format s such as management forecast s . My findings make at least two contributions to the literature. First, I address the question : Does the rise of social media imply that traditional intermediaries are less relevant? E vidence that the market resp onse to analyst forecast revisions is increasing in the number of financial tweets 3 Baginski and Hassell (1990) find that prior earnings forecasts by management influence subsequent financial analyst forecast revisi ons. 5 released by the firm suggest s that social media disclosure serves as a complement to information processing by analysts, as opposed to a substitute. Investors respond to the information in social media disclosures when those disclosures are released ( Zhang, Fuehres, and Gloor , 2011; Bollen, Mao, and Zeng , 2011; Ruiz, Hristidis, Castillo, Gionis, 2012; Mao, Wei, Wang, and Liu, 2012; Sprenger, Tumasjan, Sandener, and Welpe, 201 4), but also benefit from the subsequent interpretation of those disclosures by analysts. One explanation is that when the volume of social media disclosure by a firm is large, the ability of analysts to fit pieces of information into the overall mosaic is particularly valuable. Second, I provide evidence consistent with the argument that social media disclosure provides timely, value - relevant information to analysts. Prior studies show that social media is an efficient conduit for disseminating information to financial markets and affects investor behavior. However, there is limited evidence o n how information from social media affects the behavior and beliefs of sophisticated information intermediaries. To my knowledge, this is also the first study to exam ine concurrently the influence of social media disclosure s by firms and the public on financial analyst following and the properties of analyst earnings forecasts. The rest of the paper is organized as follows. Section 2 presents institutional background, summarizes related studies on social media disclosure and financial analysts, and develops hypotheses. Section 3 describes the data collection method and provides an outline of the research design. Section 4 presents results from empirical analyses of the effect of social media disclosure on analyst following and properties of analyst forecasts. Th is Section also includes a discussion of analysis of the relation between social media disclosures and the information content of analyst forecast revisions. Sec tion 5 present s results of sensitivity analyses. Section 6 summarizes the main findings and provides a conclusion. 6 2. BACKGROUND AND HYPOTHESES I begin this chapter by discussing the role of i nformation technology in corporate disclosure . I then discuss cha nge s in Regulation Fair Disclosure (Reg. FD) in response to the increased importance of social media business communication. Then, I discuss prior literature on the effect of general social media and ss prior literature on the role of financial analysts as information intermediaries. Finally, I develop my hypotheses. 2.1. Role of Information Technology in Corporate Disclosure Due to innovation s in information technology, there have been enormous changes in business communication pr actices over the past decade. The Internet lowered the cost of information dissemination and increased the velocity at which information travels. Ashbaugh, Johnstone, and Warfield (1999) and Ettredge, Richardson and Scholz (200 2 ) document that most firms use the Internet for voluntary financial information disclosure . Similar to the introduction of the Internet, the emergence and widespread adoption of social media increased information flow by facilitating interaction betwe en Websites and information users. The use of social media not only facilitate s the dissemination of news but also encourages participation, collaboration, and information sharing (Culnan et al. , 2010; Chen et al., 2014 ; Kane et al. , 2014) . Social media ha s also reduced use social media to reduce information asymmetry by disseminating news directly to investors rather than relying solely on third party intermediaries (Blankespoor, Miller, and Whit e , 201 3 ). Blankespoor et al. (201 3 technolog y feature of social media, where push technology refers to electronic communication in which the sender transmits information to the user instead of waiting until the user specifically reque sts the information. They show that by sending investors a hyperlink to a press release concurrent with the issuance of the press release , 7 acquisition costs. This br oad dissemination also increases the likelihood that all users have access to the information at the same time. 2.2. Changes in SEC Disclosure Regulation The proliferation of company W The Use of Company Web s 2008, which addressed how W ebsite disclosures could . FD . 4 Although social media can be used to disseminate information to a large number of users at a low cost, prior to April 2, 2013, the SEC con cerns about selective disclosure prohibited companies from using social media such as Twitter to initially disclose material and nonpublic information under Reg . FD . Therefore, rational information users could ignore social media platforms as an outlet for new information. On April 2, 2013, the SEC responded to public companies use of social media by issuing a report stating that initial dissemination of mandatory filings by SEC registrants through social media outlets such as Facebook and Twitter does not violate Reg . FD, so long as investors have been alerted in advance to the social media outlets that will be used. 5 In addition, the Guidance on the U se of C ompany Web s ites for D isclosure can be applied to social media p latforms. I f the information is 4 Thus, research prior to the 2008 expansion of Reg. FD to W ebsites studied I nternet financial disclosures that were already available from other sources. Ashbaugh et al. (1999) study corporate W ebsite disclosure of comprehensive financial statements and links to SEC filings. Ettredge et al. (200 2 ) examine information already filed with the SEC and other voluntary information available from other sources such as stock price, calendar events, and a list of analysts who cover the firm. 5 Report of Invest igation Pursuant to Section 21(a) of the Securities Exchange Act of 1934: Netflix, Inc., and Reed Hastings , Release No. 34 - https://www.sec.gov/litigation/investreport/34 - 69279.pdf) 8 designed to provide broad, non - n issuers would be allowed to elect not to file a Form 8 - K. Shortly after the SEC report wa s issued , many companies, including Netflix, Nielson, Dell, and AutoNation, filed a Form 8 - K detailing their intent to disseminate investor information on their social media feeds. As of 2013, Twitter, Facebook, and YouTube accounts were used to release co rporate information by 77%, 70%, and 69% of Fortune 500 companies, respectively (Barnes et al. , 2013). Therefore, rational investors are expected to pay greater attention to social media platforms as a source of new, relevant information. As social medi a bec a me more prevalent, many firms created written social media policies. Barnes and Daubitz (2017) document that 50% of Inc. 500 companies have a written social media policy incorporated into their business plan , and 21% have a stand - alone social media p olicy . 6 In total, 77% of Inc. 500 companies adopted social media policies to guide the online communications of the firm and its employees. Socialmediagovernance.com provides a social media policy database that includes uideline s. 7 For example, Apple provides retail blogging and online social media guidelines for its employees , and Cisco offers an I nternet postings policy. Accounting and consulting firms also provide services to guide (Ernst & Young, 2014; Deloitte Touche Tohmatsu, 2015 ; KPMG, 2015; PwC, 2017). The policies cover both firms and their executives , including legal and regulatory compliance risk, security risk, and reputational risk (Elliot, Grant, and Hodge, 2018). 6 Inc. is an Am erican weekly magazine that publishes news about small businesses and startups. Beginning in 1982, t he magazine publishes annual lists of the 500 and 5000 fastest - growing privately held small companies in the U.S., called the "Inc. 500" and "Inc. 5000". 7 http://socialmediagovernance.com/policies/ 9 2.3. Resea rch R elated One distinctive f eature of social media is that new platforms allow users to create and disseminate their own content about firms (Miller and Skinner, 2015). U sers formerly known a s the audience, i.e., consumers of information, are now producers of information (Rosen, 2006; K aplan and Haenlein, 2010). For example, using a measure of the bullishness of messages posted find that stock messages help predict market volatility and that disagreement among the posted messages is associated with increased trading volume. Das and Chen (2007) find that investor sentiment extracted from Internet stock message boards is significan tly related to stock indices, trading volume, and volatility. Rickett (2016) find that the financial blog, SeekingAlpha.com, serves an infomediary role for retail investors especially when information asymmetry is high, earnings quality is low, and during economic uncertainty. Prior literature also examines the impact of social media disclosures provided by a broad set of stakeholders on the investment decisions of investors. Gomez - Carrasco and Michelon (2017) investigate the influence of social media acti vism on the stock market performance of targeted firms. They focus on information published on Twitter by consumer associations and trade unions . They provide evidence that tweeting by key stakeholders has a significant impact on investors' decisions. Tang (2018) finds that third - party - generated comments about products and brands on Twitter, aggregated at the firm level, provide information that is useful in forecasting firm - level fundamentals. She finds that Twitter comments not only reflect upcoming sales , but also capture an unexpected component of sales growth . The f indings of this study suggest that user generated nonfinancial information on social media is also predictive of future firm performance. 10 The advent of social media also provide s opportunit ies for individual public opinions about firms to be more easily accessed and aggregated (Hales, Moon, and Swenson, 2018). R ecent research suggest s that various platforms provide channels for communicating information that is relevant to forecasting future performance and disclosure. Using crowdsourced forecast data from Estimize in 2012 and 2013 , Jame, Johnston, Markov, and Wolfe (2016) find that crowdsourced forecasts are incrementally useful in forecasting earnings and measuring the market's expec tations of earnings. Hales et al. (2018) examine whether the opinions employees share on social media relate to future corporate disclosures. Using a sample of approximately 150,000 employee reviews from Glassdoor.com, where employees voluntarily share the ir opinions - term business outlook, they find that employee opinions posted on social media platform s voluntary disclosure s . Together, these studies imply that users formerly known as the audience use social media disclosures to create and disseminate their own content about firms (Rosen, 2006; Miller and Skinner, 2015). Another distinctive technological feature of social media is that it is a two - way communication ch annel that allow s stakeholders to interact with manage rs and with each other (Cade, 2018; Elliott, Grant, and Hobson, 201 7 ). Thus, s ocial media implies a fundamental change in the information environment. Trinkle, Crossler, and Bélanger (2015) examine the impact of stakeholder s on social media. They find that the opinions of others, as expressed in attached comments via social media, have valuation judgments and influence investors' pe rception s . The f indings of this study also imply that social media not only provide s two - way interaction between management and non - management stakeholders , but also results in more active interaction among non - management 11 stakeholders. Using Twitter commen ts that contain product information , T ang (2018) finds that third - party - generated comments about products and brands on social media, when aggregated at the firm level, provide information that is useful in forecast ing future firm sales . S ocial media also provides firms with the opportunity to respond to comments and questions posted by stakeholders. This feature provides firms with the opportunity to mitigate reputational damage by engaging in conversations on Twitter. Accounting and consulting firms who that firms mitigat e social media risk by monitoring social media conversations and responding quickly when issues emerge . R ecent research supports this guidance (Elliot et al., 2018). Le e et al. (2015) examine how corporate social media affects the capital market consequences of consumer product recall disclosure s . They find that during a crisis triggered by a recall, quickly inform ing customers and the public of the recall on social medi a help s to minimize the spread of rumors and misinformation. They document that corporate social media, on average, attenuates the negative price reaction to recall announcements, and that the attenuation benefits vary with the level of firm involvement an d with the level of control the firm has over its social media content. 8 Gans, Goldfarb, and Lederman (2017) find that customer complain ts on Twitter increase when the on - time performance of airlines deteriorates, and that airline companies are more likely to respond to the complaints if the complaints are from airports or hubs out of which they operate a greater share of flights. This paper suggests that two - way communication using social media also plays a disciplining role by product quality. Cade (2018) examine s 8 Hsu and Lawrence (2016) also investigate how company involvement in social media affects the capital market of company involvement in mitigating the potential negative effects of social media during a product recall. Hsu and Lawrence (2016) sample covers not only consumer product recalls , but also food, drug, and automotive recalls which have greater social i mpacts. 12 disclosure strategy affect s negative influence of criticism on Twitter by directly addressing the criticism and redir ecting attention to positive information in the firm disclosure s . Together, these studies imply that an increase in two - way interactions on social media results in the provision of more comprehensive and complete information . As Miller and Skinner (2015) point out, the emergence of social media provides firms a new way of disseminating information, but the interactive features of social media bring new challenges for firms as they seek to manage the information environment. 2.4. Prior Literature on the Role of Twitter in Business Communication As social media evolved, three platforms - Facebook, Twitter, and YouTube - either absorbed or replaced other platforms to become social media market leaders. Each platform has distinctive features that facilitate differe nt types of communication among different groups of users. Twitter is ranked as the top platform by investor relations professionals (Jones, 2013). Twitter restricts tweets to 140 - characters. 9 pr oviding an ideal medium for sharing relevant information in a timely fashion, in contrast to the longer format and potentially reduced timeliness of research reports or articles (Bartov et al., 2017). Prior research consistently finds that social media has a significant influence on financial markets. For example, the mood of Twitter feeds can predict the movement of stock market indices (Zhang 9 In my sample period, Twitter restricted tweets to 140 - characters. Subsequent to my sample period, there have been several changes in the 140 - character rule. On May 24, 2016, Twitter announced that m edia such as phot os and videos, would not count against the 140 - character limit. Previously, a photo was considered to be approximately 24 characters. In addition, a ttachments and links are no longer part of the character limit. On September 26, 2017, Twitter announced it was testing 280 - character limit tweets. The 280 - character limit went live for all users on November 7. Certain characters, including CJK, emoji and most Unicode symbols, count as two characters under the new limit s . 13 et al., 2011; Bollen et al., 2011), and tweets are correlated with trading volume (Ruiz et al., 2012; Mao et al., 2012; Sprenger et al., 2014). links to press releases is associated with reduced information asymmetry, as measured by lower abnormal bid - ask spreads and greater abn ormal depth. Push dissemination is also positively associated with liquidity. Prokofieva (2014) also investigates the effect of dissemination of corporate disclosure via Twitter and find s that tweets posted by a firm decrease the information asymmetry prox ied by the abnormal spread. She also finds that this negative association is stronger for firm s with less business press or financial analyst coverage. Bhagwat and Burch (2016) provide attention to firm disclosure s . They find that tweets about earnings news increase the magnitude of announcement returns and that this effect is more significant for small, positive earning s surprise s and when the firm is less visible as measured by firm s ize or analyst coverage. Together, these three studies provide evidence that Twitter allows companies to disseminate corporate announcements more effectively , attract investors' attention , and contribute to a decrease in information asymmetry. Lee et al. (2015) document that by quickly informing customers and the public of consumer product recalls, social media disclosures help to minimize the spread of rumors and misinformation. However, they also find that social media can be a double - edged sword. Socia l media can exacerbate a crisis by spreading news to a wider audience, thereby helping the news to go viral. Their findings indicate that the benefits and costs of corporate social media usage vary with the level of control the firm has over its social med ia content. As social media disclosure bec o me s more prevalent, some top executives connect with investors directly, personally, and in real time through social media. Chen et al. (2018) find that personal tweets by CEOs and CFOs 14 contain information that bo th improves stock market liquidity and exacerbates stock return volatility. They document that executive participation on social media grabs investor attention and enable s retail investors to obtain value - relevant information to which they previously had n o access. Analyzing S&P 1500 firms' use of Twitter to disseminate quarterly earnings announcements, Jung, Naughton, Tahoun, and Wang (201 8 ) find that social media outlets are more likely to be used to disseminate quarterly earnings news when the news is p ositive, suggesting that some firms are opportunistic in their use of social media. They also find that the market reaction is stronger for firms that follow a consistent social media disclosure policy. Crowley, Huang, and Lu (2018) also investigate discretionary disclosure on Twitter. They find that firms social media disclosure activities are more active around earnings announcements, accounting filings, and firm - specific news events. Unlike Jung et al. (201 8 ), they find that firms are more likely to disseminate news on Twitter when it is significantly good or bad . 10 This finding suggests that firms are not opportunistic in their usage of social media. Although these two studies provide some contradictory findings, together they imply that managers exercise discretion regarding the level, timing, and format of disclosure on social media. There are also a few papers that provide evidence on managers use of discretion in the onmental performance, Huang, Lu, and Su (2016) find that green firms are more likely to be early adopter s of Twitter and tweet more frequently about their prosocial behavior. Yang and Liu (2017) find that 10 There are several significant differen ces between Jung et al. (201 8 ) and Crowley et al. (2018). First, the sample periods and sizes of the two studies are quite different. Second, Jung et al. (201 8 ) adopt a dictionary approach , while Crowley et al. (2018) employ a machine learning approach to identify earnings announcement related tweets. Third, Jung et al. (201 8 ) use earnings surprise s to classify good or bad news as they focus solely on earnings announcements, while Crowley et al. (2018) use both RavenPack and CAR( - 1,1) to classify good and b ad news. 15 firms with earnings increases are more willing to u se Twitter to disseminate earnings - related disclosure than are firms with earnings decreases . Yang, Liu, and Zhou (2016) investigate the effect of corporate governance on the decision to disseminate earnings related disclosure s on Twitter. They find that the dissemination of earnings news on Twitter is significantly associated with larger board size, greater gender diversity, and higher board effectiveness. Their findings provide evidence that corporate governance plays a significant role in decision s abou t social media disclosure. B aik, Cao, Choi, and Kim (2016) use geographic proximity as a measure of private information and find local Twitter users are more likely to tweet about firms with high information asymmetry, and their Twitter activity , in turn , increases the trading volume of local stocks. Together, these studies imply that as firms consistently disclose through the use of social media, investors become more informed and information asymmetry is reduced. While prior studies contribute to our und erstanding of the effects of social media on financial markets, they generally focus on the response of investors. Although investors may be the primary audience for fi nancial communication, fi rm disclosure through social media may also change the informat ion environment for other stakeholders. S ocial media disclosure is often context specific, implying that information extraction and interpretation may require the acqui sition and interpre ta ti on of objective, quantitative information or require context spec ific abilities such as industry or institutional knowledge (Blankespoor , 2018). Given the importance of financial analysts as information intermediaries, investigating the impact of social media disclosure on analyst behavior is important . 16 2.5. The Role of Fi nancial Analysts as Information Intermediaries Financial analysts are a primary information intermediary in capital market s ( Womack , 1996; Jegadeesh , Kim, and Krische, 2004; Ivkovic and Jegadeesh , 2004; Asquith, Mikhail and Au , 2005) . In response to the in of financial analysts as information intermediaries is expanding (Lehavy et al., 2011; Lev and Gu, 2016). Analysts collect information from public and private sources and interpret comp lex communication using their expertise and industry knowledge (Jacob, Lys, and Neale, 1999; Ramnath, Rock, and Shane, 2008). Prior research consistently finds that firm disclosure is an important determinant of analyst following and the properties of ana lyst forecasts. For example, Lang and Lundholm (1996) find that firms with more forthcoming direct investor relations communications have greater analyst following and more accurate analyst earnings forecasts, less dispersion among individual analyst forec asts , and less volatility in forecast revisions. Healy, Hutton, and Palepu (1999) show that firms whose disclosures provide greater information content have more accurate analyst earnings forecasts and less dispersion among individual analyst forecasts. Kr oss, Ro and Schroeder (1990) disclosures are more informative. Hope (2003) find s that a cross countries, the level of disclosure about accounting policies is inv ersely related to forecast errors and dispersion. Lehavy et al. (2011) find that less readable annual reports are associated with lower accuracy and greater dispersion of analyst forecasts. Researchers also have investigated whether and how significant changes in the information environment resulting from Reg . FD impacted the behavior of analysts and the properties of their forecasts (Irani and Karamanou, 2003; Heflin, Subramanyam, and Zhang , 2003; Baily, Li, Mao, 17 and Zhong, 2003; Agrawal, Chadha, and Ch en , 2006; Mohanram and Sunder, 2006). Collectively, subsequent to Reg . FD . 11 These findings suggest that Reg . FD decreased the quantity and quality of publicly ava ilable information and also imply that the amount of information available to Previous studies also find that analysts use soft information as well as hard information ( B radshaw, Wang , and Zhou, 2016; Huang and Mamo, 2016 ) . Using firm specific print news coverage data, Bradshaw et al. (201 6 ) find that the quantity of news coverage about a firm is positively associated with subsequent recommendation revisions, and that the tone of the n ews predicts the direction of the revisions. Huang and Mamo (2016) find that analysts' earnings forecast revisions are significantly influenced by the tone of news and that the relation between news and earnings forecast revisions is stronger when the news contains information regarding firm fundamentals. Together, these studies provide evidence that analysts are influenced by information provided by another information intermediar y , the media. formation extracti on costs (Bloomfield, 2002; Grossman and Stiglitz, 1980; Libby and Emett, 2014) , and distinctive characteristics of social media can create fundamentally different information environment s for information users. S ocial media disclosure pr ovides more context specific information (Blankespoor , 2018). Considering that the intermediary role increases in importance with the 11 Heflin et al. (2003) find neither forecast accuracy nor dispersion appear to change following Reg . FD, suggesting that Reg . FD did not restrict the information available to investors prior to earnings announcements. 18 difficulty of interpretation, context specific abilities such as the industry or institutional knowledge of financial anal ysts would be more valuable. Analyst forecast revisions have information content (Griffin, 1976; Givoly and Lakonishok , 1979; 1980; Abdel - khalik and Ajinkya, 1982; Fried and Givoly, 1982; Imhoff and Lobo, 1984; Gleason and Lee, 2003; Ivkovic and Jegadeesh , 2004 ). They occur throughout the quarter and are the result of analysis and interpretation of new information. Forecast revisions are positively associated with the sign and magnitude of stock returns (Brown et al., 1985; Klein, 1990; Lys and Sohn, 1990 ; Beyer et al., 2010). Collectively, the studies on analyst forecasts and revisions suggest analysts play a valuable role as intermediaries (Guan, Lu, and Wong, 2012) and are sophisticated users of financial information (Chava, Kumar, and Warga, 20 0 9 ). Bro wn, Call, Clement, and Sharp (2015) find that sell - side analysts are also valuable intermediaries, even for institutional investors. Prior studies also have investigated a variety of factors that explain the magnitude of the market response to analyst fore cast revisions. Clement and Tse (2003) find that the response varies with forecast accuracy. Barniv and Cao (2006) document that analyst characteristics and innovation explain investors' reaction s to forecast revision s . Livnat and Zhang (2012) find that a significant percentage of analyst forecast revisions are issued promptly after a broad set of corporate public disclosures and that investors perceive these prompt revisions as more valuable than non - prompt revisions. Their findings indicate that investors ability to interpret public disclosures, especially less structured or non - financial disclosures, than 19 2.6. Development of Hypotheses As summarized above, research on the role of financial analyst s as information future operations. In addition, studies on are informative and are associated with reduced information asymmetry. Overall, both financial analysts and social media contribute to the efficient functioning of financial market s . However, it is u nclear whether social media disclosures increase or decrease the need for financial analysts as information intermediaries, and whether the information on social media is correlated with the information used by analysts. Therefore, investigating the impac t of social media disclosures on the intermediation role of financial analysts is necessary to fully understand the consequences o f social media for companies and investors. Below, I discuss my predictions related to the relation between social media usage and analyst coverage, forecast accuracy, forecast dispersion , and the market reaction to forecast revisions . The effect of social media disclosure on the demand for intermediation by financial analysts is ambiguous. Social media platforms undeniably redu ce the costs a firm must bear to disseminate information and the costs information users must bear to gather information. This unique benefit of social media encourages firms to directly communicate with investors and thus would suggest a reduced demand fo r the intermediation role of financial analysts. This assumes that investors are able to process the information released on social media to develop earnings forecasts. However, according to information overload theory, too much information can make it mo re difficult to understand an issue and to use information to make decisions. Although t ask 20 performance initially improves as more i nformation is available , Speier, Valacich, and Vessey (1999) find that as the amount of information begins to exceed the dec i processing capacity , performance eventually declines. The combination of more information and limited information processing capacit y leads to information overload, which reduces decision making effectiveness. Information technology creates i nformation overload because ideas are disseminated instantly and frequently (Evaristo, Adams, and Curley, 1995; Hiltz and Turoff, 1985). If social media disclosures produce excessive information, the disclosure could be treated as noise even when it contai ns information . Information overload can increase investor information analysis and interpretation costs, and, therefore, increase the need for an information intermediary. Moreover, disclosures must be interpreted and analyzed to have information content. Social media disclosure has high flexibility in format, but low comparability in content compared to traditional SEC filings. Therefore, interpreting and judging the relevanc e and value of social media disclosure could be challenging to investors , thus s uggesting an increased demand Bhushan (1989a; 1989b ) and Lang and Lundholm (1993) argue that voluntary disclosure lowers the cost of information acquisition for analysts and hence increases the number of firms analysts follow. Overall , I expect greater use of social media to result in an increased demand for intermediation by financial analysts. As analysts are more likely to initiate coverage of firms for which investors have a high demand for intermediation, I predict a positive asso ciation between the use of social media and the number of analysts following a firm. Thus, my first hypothesis is as follows: H 1 : The degree of corporate use of social media is positively associated with the number of a nalysts following the firm. 21 Another important question is whether social media disclosures undermine or reinforce the earnings forecasting abilities of financial analysts. Mosaic theory in finance refers to an analyst gleaning many different pieces of information to construct a sensi ble narrative and then deciding whether to recommend a trade (Pozen, 2005). This involves collecting public, non - public, material, and immaterial information about a company in order to determine the underlying va lue of the company's securities (Caccese, 1 997; Davidowitz, 201 4 ). Following Mosaic theory, skilled analysts with industry knowledge will interpret, analyze, and combine immaterial information with material information. T herefore, even information that is immaterial on its own contributes to reachi ng a conclusion. Given that even immaterial information is useful to financial analysts, Mosaic theory suggests that disclosure through social media is a valuable additional information source. Therefore, expanded disclosure through social media potentiall y enables financial analysts to create valuable new information, such as superior forecasts and reinforces the intermediation role. A s ambiguity and uncertainty among analysts concerning the future performance of the company decreases, the level of disagre second and third hypotheses are as follow: accuracy. H3: The degree of corporate use of socia l media is negatively associated with dispersion of they believe a nalyst s provide new information about the industry, firm, and macro economy, as well as informative interp re ta t ions of financial statements and public disclosures ( Beaver, Corne ll, Landsman, and Stubben, 2008; Clement, Hales, and Xue, 20 1 1 ; Baginski, Hassell, and Wieland, 2011). Livnat and Zhang (2012) find that investors especially retation of less structured or non - financial 22 disclosures . Social media disclosure s are unstructured, implying that when a firm uses social media, analyst forecast revisions will be particularly useful to investors because there is more info rmation to inter pre t . I also expect analysts to work harder because there is more information to interpret. Thus, my fourth hypothesis is : H 4 : The degree of corporate use of social media is positively associated with the market response to analyst forecasts revisi ons . 23 3. DATA AND RESEARCH DESIGN Chapter 3 discusses sample selection and research design. I begin with a discussion of the sample used in th e analyses . I then discuss the social media disclosure s used to test my hypotheses, followed by the measures of ana lyst coverage and properties of analyst forecasts that are my dependent variables . Then, I present the empirical model used to test H1 , which predict s a positive association between the degree of corporate use of social media and the number of analysts fol lowing the firm. Next, I present the empirical model used to test H2 ( H3 ) , which predict s a positive (negative) association between the corporate use of social media . Lastly, I present the empirical model used to test H4 , which predict s a positive association between the degree of corporate use of social media and the market response to analyst forecast revisions. 3.1. Data To investigate the effect of social media disclosure on analyst follo wing , the properties of analyst earnings forecasts, and the market reaction to analyst forecast revisions, I analyze the social media activity of S&P 500 firms from the first calendar quarter of 2012 through the fourth quarter of 2014 , a period that includ es the April 2, 2013, SEC rule permitting the initial disclosure of material nonpublic information on social media . The initial sample consists of firm - quarter s with data available on Compustat, Thomson Reuters Financial, I/B/E/S, EDGAR, and Twitter. I ext ract a nalyst forecast and management forecast data from I/B/E/S, financial data from Compustat Quarterly and Segment files, and institutional ownership from Thomson Reuters Financial. I obtain complete historical Twitter data from CrimsonHexagon, one of th e official resellers of Twitter data. I winsorize all independent and dependent variables at the top and bottom one percent. As 24 shown in Table 1, excluding observations without information needed to estimate the control variables reduces my initial sample from 6,000 to 4,9 74 firm - quarter observations . 3.2. Measures of Social Media Disclosure Social media research in accounting is still evolving, and measures of social media activity are not yet well established. Previous studies count the number of tweets duri ng a specific short window period around news events such as earnings announcement dates and product recall announcements (Lee et al., 2015; Blankespoor et al., 201 3 ). In contrast, this study distinguishes itself from most studies that focus on specific ev ents by focusing on the level of Twitter disclosure activity of firms and interactions among stakeholders and potential investors aggregated at the firm level. I create three measures that capture disclosure about a firm on Twitter the number of tweets on financial topics by the firm, the number of tweets on non - financial topics by the firm , and the number of tweets about the firm by the public . To measure the number of financial tweets released by the firm, I count the number of tweets that originate fr om the Twitter account linked W ebsite and Dictionary , which I augmented by an analysis o f the frequency with which various words appear in SEC 10 - K filings. I first restrict the augmented word list to those words used more than 100,000 times in 10 - Ks filed from 1994 to 2014. From that subset, four individuals with work experience as financia l accountants individually identified the words they considered financial performance . Any words determined to be financial - related by at least two of the four individuals were retained in the dictionary . Second, I measure the n umber of nonfinancial tweets by counting the number of tweets from the firm that do not contain 25 financial keywords. Finally , to measure the amount of financial information shared by social media users other than the firm, I coun t the number of tweets twitter account and C ashtag. 3.3. Measures of A nalyst C overage and P roperties of F orecasts Analyst coverage is measured as the number of analysts who comprise the most recent I/B /E/S consensus quarterly earnings forecast prior to the quarterly fiscal period ending date. Following prior literature ( e.g., Schipper, 1991; Brown, 1993), I calculate the forecast error for firm i in quarter t as the absolute value of forecast EPS less a ctual EPS , scaled by actual EPS, where forecast EPS is based on the last consensus quarterly earnings forecast before the financial period ending date of the I/B/E/S Summary data: Analyst forecast dispersion is computed as the standard deviation from the last consensus quarterly earnings forecast before the financial period ending date on the I/B/E/S Summary data file. 3.4. Re search Design 3.4.1. Analyst Following Model To examine the effect of social media disclosure on the intermediary role of financial analysts, I estimate the following regression model: 12 Following it = + Financial_Tweets it + Non - financial_Tweets it + Crowd_Tweets it + News it + Size it + BtoM it + Loss it + ROA it + it + Intangible_Asset it + R&D it + + 12 The twelve industry classi fications are defined in appendix B. T ime fixed effects are controlled using twelve indicator variables , one for each of the twelve calendar quarters in my sample period . 26 + Institution al it + 8 - K it + Management _Qtr it + Industry Fixed Effects + Time Fixed Effects + (1) Equation (1) is estimated using a negative binomial count - data model with indust ry and time indicators . 13 1995; Lehavy et al. , 2011), I define analyst following , Following , as the number of analysts that comprise the last consensus quarterly earnings forec ast before the financial period ending date on the I/B/E/S Summary data file for firm i in quarter t . Following Bhushan (1989) and Lehavy et al. (2011), I interpret this measure as a proxy for the collective effort of the financial analyst community in the analysis of an individual firm. The variables of interest, Financial_Tweets , Non - f inancial_Tweets , and Crowd_Tweets , capture the amount of information released on social media by firm i in quarter t and information shared by others about the firm over the quarter . My first hypothesis predicts that the degree of social media usage by the firm is associated with the demand and/or supply of information from financial analyst s . Support for H1 implies that , the coefficient on Financial_Tweets , will be positive. The degree of information shared by the crowd is a good proxy for the degree of public attention and public demand for information about a firm. Therefore, I also expect Crowd_Tweets to be posit ively associated with Following, i.e., is expected to be positive. 13 Rock, Sedo, and Willenborg (2001) show that when analyzing count data such as analy st coverage (i.e. nonnegative integer data), the negative binomial model is more appropriate than the OLS or Poisson models and better captures the true underlying data generating process. It also addresses the econometric issues associated with truncation (zero value) and over - dispersion (lower standard error) in the data. 27 3.4.1.1. Control V ariables for Analyst Following Model In addition to the social media variables of interest, I include control variables identified by the prior literature as explaining a nalyst following and the properties of analyst forecasts . First, I control for the volume of fact - based articles contain ing financial news provided by formal news organizations each quarter to ensure my variables of interest are not just capturing the over all volume of available information. 14 By controlling for the volume of financial information released in other traditional media outlets, I can investigate whether my variables of interest are associated with the demand and/or supply of information from fi nancial analyst s . If a greater volume of coverage by traditional media also result s in an increased demand for intermediation by financial analysts , analysts are more likely to initiate coverage of firms that have a high demand for intermediation . Therefor e, I predict a positive association between the volume of fact - based articles contain ing financial news provided by formal news organizations and the number of analysts following a firm. Additional control variable s include firm size , book - to - market rat io, return on assets, leverage, and an indicator variable for losses. 15 Previous studies document that firm size is the most important explanator of analyst following, with larger firms having greater following (Bhushan , , 1990; Bre nnan and Hughes , 1991; Lang and Lundholm , 1996; Barth, Kasznik, and McNichols , 2001 ). To control for size, I include the natural log of market value. Following prior work, I include book - to - market , an inverse proxy for growth (Smith 14 I use the , which provides full access to all available Fact - based articles by formal news organizations, such as CNN, New York Times, Wall Street Journal, etc. using web searches. I retain each article containing an official name of a specific company and at least one of the words from the financial key words dictionary developed for financial tweets. 15 Size, book - to - market ratio, lever age , return on assets, and the loss indicator variable are measured at the end of the quarter t . 28 and Watts, 1992; Barth et al. , 2001; Lehavy et al. , 2011). T o control for firm performance, I include return on assets, a loss indicator variable, and financial leverage . I also include controls for the level of business complexity and the degree of information asymmetry betwe en a firm and its market participants. Barth et al. (2001) find analysts have greater incentives to follow firms with larger intangible assets , which are more difficult for investors to value . They find analyst coverage is significantly greater for firms w ith larger R&D expenses relative to their industry peers. I include both intangible assets scaled by total assets and the dollar value of research and development expenditures as controls . To control for the effect of business complexity, I include the num ber of reported business and geographic segments as of the ending date of the previous fiscal year (Bradshaw , 2009 ; Lehavy et al. , 2011). Prior research also documents that analyst coverage is associated with the number of institutional investors (Bhushan, 1989; O'Brien and Bhushan, 1990; Brennan and Subrahmanyam, 1995; and Frankel, Kothari, and Weber , 2006). To control for the level of institutional holdings, I include the percentage of institutional ownership as of the quarter ending date ( Ljungqvist, Mar ston, Starks, Wei, and Yan 2007; Bae , Stulz, and Tan, 2008 ) . I include the number of Form 8 - K filings issued over the quarter to control for the amount of information distributed through SEC filings, which I expect to be positively associated with the num ber of analysts following a firm. Finally, following Lehavy et al. (2011), I include the number of management forecast s issued each quarter by the firm as a proxy for the firm discretionary disclosure (Nagar , Nanda, and Wysocki, 2003; Cotter , Tuna, and W ysocki, 2006) . On one hand, m anagement forecast s may increase analyst following because there is more information to interpret and an increase d demand for the intermediary role of financial analysts. O n the other hand, earnings forecasts provided by manage ment may preempt or substitute for information processing by financial analysts because there is already a benchmark 29 future performance. Therefore, the impact of management forecast s on analyst following is an empirical question. 3.4.2. Forecast Pr operties Model To examine the effect of social media disclosures on the properties of analyst earnings forecasts (H2), I estimate the following OLS regression model: Forecast_Properties it = + Financial_Tweets it + Non - financial_Tweets it + Crowd_Tweets it + News it + Following it + Horizon it + Size it + BtoM it + Loss it + ROA it + it + Intangible_Asset it + R&D it + + + Institutional it + 8 - K it + + Management_ Qtr it + Industry Fixed Effects + Time Fixed Effects + (2) Model (2) is est imated using ordinary least - squares (OLS) regression with industry and time indicators to control for industry and time fixed effects . 16 This model is estimated separately for Forecast_Error and Forecast_Dispersion . 3.4.2.1. Control V ariables for Forecast Propertie s Model Bagnoli, Levine, and Watts (2005) find that news released by a firm through traditional media outlets has a significant influence on analysts forecasting activity. In addition, Huang and Mamo (2016) find th at company information disseminated via n ew s media outlet s influence s analysts earnings revisions. Therefore, I include a control variable for the natural log of the volume of articles contain ing financial news provided by traditional news organizations over the 16 Industry fixed effects are controlled using 12 indicator variables that each represent one industry division. Classification of industry divisions is discuss ed in appendix B. T ime fixed effects are controlled using quarterly time indicator variables 1 through 12 to capture the twelve quarters . 30 quarter to ensure my variables of interest are not proxying for the overall volume of information available. I also include analyst following , measured as the number of analysts who compromise the most recent I/B/E/S consensus quarterly earnings forecast prior to the quarterly financial p eriod ending date to proxy for the time, effort, and resources analysts devote to gathering and analyzing information about the firm. A nalyst coverage is expected to be negatively related to the level of information asymmetry and, therefore, Forecast_Error and Forecast_Dispersion finds that recent forecasts are more accurate. Horizon is included in the model to control for the amount of time elapsed between the forecast date and the related earnings announcement date. P rior studies find that larger firms have richer information environments and potentially smaller Forecast_Error and Forecast_Dispersion (Bhushan , , 1990; Brennan and Hughes , 1991; Lang and Lundholm , 1996; Barth et al. , 2001). I include Size as of the e nding date of the quarter to control for the impact of the general information environment . I include ROA because p rior research also concludes that more profitable firms have higher analyst following and, therefore, lower information asymmetry. I include a control for leverage as of the end of each quarter because Thomas (2002) presents evidence that highly leveraged firms have less accurate and more highly dispersed forecasts. I also include Loss , an indicator variable equal to 1 if the firm reported a q uarterly loss because firms suffering loss es may have a different information environment due to stakeholder dynamics. The valuation of growth opportunities is more difficult than the valuation of assets in place, and book to market , a n inverse proxy for g rowth opportunities is expected to be positively associated with Forecast_ Error and Forecast Dispersion (Smith and Watts, 1992). 17 17 Book - to - market is calculated as book value of equity divided by market value of equity as of the end of quarter t . 31 I also inclu de controls for the level of business complexity and the degree of information asymmetry between a firm and its market participants. Previous studies suggest that as forecast complexity increases, analyst forecast accuracy deteriorates (Haw, Jung, and Ruland, 1994; Duru and Reeb, 2002; Lehavy et al. , 2011). To control for forecast complexity, my model include s inta ngible assets scaled by total assets and the dollar value of research and development expenditures. Also, t o control for the effect of business complexity, I include the number of business and geographic segments (Bradshaw et al. , 2009, Lehavy et al. , 2011 ). Previous studies also find that institutional ownership is associated with higher analyst forecast accuracy and lower dispersion because firms with high levels of institutional holdings tend to have a richer information environment (Brennan and Subrahm anyam, 1995; Frankel et al., 2006) . Therefore, to control for the effect of institutional ownership, I include a control for the level of institutional holdings (Ljungqvist et al. , 2007; Bae et al. , 2008 ; Lehavy et al., 2011 ) . I include the number of Form 8 - K filings issued over the quarter to control for the amount of information released through SEC filings . Finally, following Lehavy et al. (2011), I include the number of management forecast s issued over the quarter by the firm to control for the amount of information conveyed through other form s of voluntary disclosure. 3.4.3. Market Reaction to Analyst Forecast Revisions Model To examine forecast revisions (H4 ), I estimate the followin g OLS regression model: CAR(0,1) j i t = + Mean_ AFRevise it + Mean_ AFRevise it * Log_Financial_Tweet it + Log_Financial_Tweet it + News it + Size it - 1 + BtoM it + Leverage it + Total_Revise it + 8 - K it + Managem ent_ Ind it + (3) 32 Consistent with prior research (Green , Jame, Markov, a nd Subasi, 2014; Huang, Zang, and Zheng, 2014), the dependent variable, CAR(0,1) , is abnormal returns cumulated over the two - day window beginning on the date that the forecast revision is released. I use market - adjusted returns. The daily abnormal returns are calculated as the firm - specific return less the CRSP value - weighted return . I compute the average analyst forecast revision, Mean_ A FRevise ijt , as follows. For each individual sell - side analys t , the analyst forecast revision by analyst j for firm i at time t is measured as ( AF i, j, t - AF i, j, t - 1 ) , where AF i, j, t - 1 is the most recent earnings forecast by analyst i for firm j prior to AF i , j, t , based on the I/B/E/S detail data . Both analyst forecasts and stock price are adjusted for stock spl its, consistent with Payne and Thomas (2003). Each revision is then scaled by one - month prior stock price. If there are multiple individual analyst forecast r evisions for firm j on day t , I use the average analyst forecast revision on day t . Th us, my vari able measures the average t . Prior research, for example Loh and Stulz (201 8 ), exclude days when multiple analysts issu e forecasts . However, they note this may result in bias to the extent revisions are clust ered on day s with news releases . Rather than eliminating these forecast revisions, I use the average analyst forecast revisions for firm j on day t. I also winsorize the top and bottom 1% of each independent and dependent variable to mitigate outlier effe cts. To capture social media activity between analyst forecasts, Log_Financial_Tweet is calculated as the log of one plus the sum of daily firm financial - related tweets between the prior analyst forecast and the current analyst forecast. 18 In Eq. (3), the m ain variable of i nterest is the interaction term, AFRevise * Log_Financial_Tweet . Th is variabl e capture s the impact of social media disclosure on the market 18 Because the length of time between the prior analyst forecast and the current analyst forecast is not fixed, the sum of daily firm financial - related tweets over the revision periods is high ly skewed . Considering this , I use logged value s of the variable instead of the raw values. 33 response to a forecast revision. A positive and significant coefficient , implies that the higher and the market response to the revision . In addition, the higher the level of forecast revision, the greater the association between social media activity and market response. 3.4.3.1. Control V ariables for Market Reaction to Analyst Forecast Revisions Model Nicholas and Wieland (2009) document that popular press news influence s the market reaction to analyst forecast revisions. To control for the impact of information via traditional media and press releases, I include a control variable , News , equal to the log of the volume of financial news about the firm released by formal news organizations between the previous and the current a nalyst forecast revision date s . I also include various controls identified by prior literature as potentially affecting the sensitivity of price to analyst forecast revisions. To control for the influence of analyst coverage, I include the number of analy sts who issue a revision on the forecast revision da te , Total_Revise . Brennan, Jegadeesh , and Swaminathan (1993) find that stocks with greater analyst coverage react faster to market - wide common information . F ollowing Gleason and Lee (2003) and Bonner , Hug on, and Walther (2007), I control for firm characteristics such as siz e ( SIZE ) , b ook - to - m arket ratio ( B to M ) , and leverage ( Leverage ) . 19 I also control for the t otal number of Form 8 - Ks filed by each firm between the prior analyst forecast and the current a nalyst forecast revision . Finally, following Lehavy et al. (2011) , to control for other voluntary disclosures by the firm, I include an indicator variable equal to one if there is at least one management forecast of EPS issued between the previous and curr ent forecast revision ( Management_Ind ) . 19 Size is measured as of quarter t - 1 , where quarter t is the quarter in which the analyst forecast revision is released. Book - to - Market and leverage are measured as of the beginning of the quarter in which the analyst forecast r evision is released. 34 4. E MPIRICAL RESULTS In Chapter 4, I provid e descriptive statistics for the complete sample and a profile analysis that compares the characteristics of firms whose financial tweet volume in the first qu arter of 2014 is in the top quartile of the sample with the characteristics of firms in the bottom quartile. I then present the main empirical results of multivariate tests of H1 through H4. 4.1. Descriptive S tatistics Descriptive statistics for the key variabl es used in this study are reported in T able 2. Of the S&P 500 firms, 65 firms do not have official Twitter accounts linked to their official company W ebsite s . I remove the se 65 firms from my sample , leaving 435 unique firms and 4,974 quarterly observations . The mean and standard deviation of financial tweets by the firm per firm - quarter are 143 and 35, respectively. The minimum and maximum values per firm - quarter are 9 2 and 211, respectively, implying that there is significant variation across firms in the amount of financial information shared via social media. The mean and standard deviation of non - financial tweets by the firm per firm - quarter are 267 and 79, respectively. Minimum and maximum values of non - financial tweets per firm - quarter are 135 and 392. The mean and standard deviation of tweets about the firm that contain a Cashtag ( Crowd_Tweets ) is 5,24 6 and 2,734 per firm - quarter, respectively. 20 There is also significant variation across firms in the amount of financial information shared via other so urces such as traditional media , SEC filings, and discretionary disclosure. The mean and standard deviation of the logged values of financial news by traditional media per firm - quarter are 4.7 and 2. 2 , respectively which are 110 and 9 in raw values . The me an and standard deviation of the number of Form 8 - K filings per firm - quarter is 3.6 and 2. 4 , respectively. On average, there are 1.25 management forecast s p er firm - quarter , and the standard deviation is 0.58. 20 Table 2 presents the values of Crowd_Tweets scaled by 100. 35 Descriptive statistics also include general information on the analyst forecasts. 21 The average analyst following is 18.3 2 analysts , and the range is from 3 to 38 analysts. The mean value s of f orecast e rror and dispersion are 13.7% and 0.05, respectively. The average forecast horizon i s 12.84 days, with a standard deviation of 12.48 days. Finally, in terms of firm characteristics, the mean and standard deviation of the logged values of firm size are 4.2 and 0.43, which indicate s firm size do es not vary significantly, reflective of the fact that sample firms are in the S&P 500. The variable, BtoM , has a mean of 0.47 and standard deviation of 0.40. On average, only 5.8% of firms reported a loss during the sample period, reflective of the fact that sample firms are in the S&P 500. The aver age and standard deviation of ROA per firm - quarter for sample period are 0.02 and 0.02, respectively. The mean and standard deviation of the leverage ratio are 0. 6 2 and 0.2. The mean value of Intangible Asset and R&D are 0.72 and 107.07 million , respective ly and indicate that sample firms have a high percentage of intangible assets compared to total assets and that they spend a significant amount on research and development. The average number of reported business segments and geographic segments of sample firms are 3.3 and 1.2. On average, 71.5 the quarter ending date of the sample period . 4.2. Profile A nalysis In T able 3, I compare the firm characteristics of the quartile of firms that released the largest number of financial tweets in the first quarter of 2014 to the characteristics of the quartile of firms with the lowest number of financial tweets in the first quarter of 2014. I select 2014 because Figure 1 shows that the average number of tweets per fir m and the average number of 21 T he mean value of analyst following is greater and the mean values of forecasts error and dispersion are similar to or smaller than values of previous studies, e.g., Lang and Lundholm (1996) and Lehavy et al. (2011), reflective of the fact that my sample firms are in the S&P 500. 36 C ashtag tweets by the public are generally increasing throughout the sample period. Thus, 2014 reflects the most mature stage of social media usage. The results of the t - test (Wilcoxon test) show that the mean (median) values of the volume of nonfinancial information released on Twitter, the volume of tweets about a firm by the public, the volume of financial news about a firm covered by traditional media, analyst following, size, and leverage are significantly larger for the fir ms in the top quartile. On the other hand, earnings forecast dispersion, book - to - market ratio, business segments, geographic segments, and the degree of institutional ownership are significantly smaller. Figure 1. Number of Tweets by Quarter, 2012 - 2014 a a Crowd_Tweets are in 100s. 4.3. Correlation A nalysis Table 4 reports pairwise correlations among the variables. The Twitter variables are highly correlated with each other. The correlation between Financial_Tweets and Non - f inancial_Tweets is 0.7 4 , implying that social media usage is a firm - level choice that is reflected in the volume of 37 both financial and nonfinancial tweets. 22 News is positively correlated with all three Twitter variables. Each of the three Twitter variables is positively correlated with Fol lowin g. Error and Dispersion are negatively correlated with both Financial_Tweets and Non - f inancial_ Tweets , implying that forecast errors and dispersion are smaller as the volume of information released through Twitter increases . In contrast, the Crowd_Twe ets variable is significantly and positively correlated with Error and Dispersion . One interpretation is that there is a positive correlation between the heterogeneity of analyst and investor beliefs, with the latter reflected in a higher volume of crowd t weets. 4.4. Regression R esults 4.4.1. Analyst Following Table 5 presents the multivariate regression results from the estimation of equation (1) , which tests the H1 prediction that analyst following is positively associated with social media usage. 23 The coefficient on Financial_Tweets is significant and positive (p<0.05) , indicating a positive association between social media usage and analyst following . T he coefficient on Non - financial_Tweets is significant and negative (p<0.01) , after controlling for Financial_Twee ts . Given the high correlation between these two variables, I re - estimate the model without Financial_Tweets (untabulated) and find that the coefficient on Non - financial_Tweets is positive and significant. This suggests that analysts find both financial an d nonfinancial tweets to increase demand for information. The coefficient on Crowd_Tweets is in significant. One possible interpretation is that the crowd may not provide additional information . 22 Variance inflation factors (VIFs) for all variables in all models are no larger than 3, indicating multicollinearity is not a concern. 23 The m ultivariate regression for the analysis of analyst following adopt s the negative binomial model following Rock, Sedo, and Willenborg (2001) to address the econometric issues associated with truncation (zero value) and over - dispersion (lower standard error) in the data. The p seudo R 2 from the negative binomial model is not comparable to the adjusted R 2 . Therefore, I do not compare the explanatory power of my model to the adjusted R 2 of previous studies on the variability of analyst following around its mean . 38 Similar to prior research (Bhushan , , 1990; Brennan and Hughes , 1991; Lang and Lundholm , 1996; Barth et al. , 2001; Lehavy et al. , 2011), I find Size is significantly and positively (p<0.01) associated with Following . Consistent with Barth et al. (2001) , I document analyst following is smalle r for firms with higher growth (p<0.01) . Consistent with , 1990; Brennan and Subrahmanyam, 1995; and Frankel et al., 2006), I find that institutional ownership is positively (p<0.01) associated with analyst following. As predi cted, I also find that the volume of fact - based articles contain ing financial news provided by formal news organizations is positively and significantly (p<0.01) associated with analyst following, while the impact of the volume of manag ement forecasts on analyst following is insignificant. 4.4.2. Forecast Error Table 6 presents the multivariate regression results of estimating equation (2) with forecast error as the forecast property of interest. 24 Consistent with H2 , I find that the coefficien t on Financial_Tweets is significant and negative (p<0.1 0 ) . This supports the hypothesis that social media usage is positively associated with analyst forecast accuracy. This finding provides evidence that financial information delivered via social media p rovide s incremental information to analysts in addition to that provided by traditional media. The coefficient on Non - f inancial_Tweets is not significant. The coefficient on Crowd_Tweets is significant and positive (p<0.01) . There are several possible inte rpr et ations. First, on average, the crowd may provide misleading , meaningless information. Second, p rocessing of divergent information involves more screening, evaluating, and interpreting (Schick, Gordo, and Haka, 1990) and these tweets may cont ribute to information 24 The multivariate regression model for the analysis of forecast error has an adjusted R 2 of 0.216, which indicates that my model explains about 22% of the variability of analyst forecast error around its mean. Lehavy et al . (2011), Dhaliwal, Radhakrishna n, Tsang, and Yang (2012), and Lang and Lundholm (1993) report adjusted R 2 of 0.05, 0.12, and 0.38, respectively. Compared to the previous studies, my model has a decent level of explanatory power. 39 overload and a decline in forecasting performance (Agnew and Szykman, 2005). Alternatively, the volume of crowd t weets As expected, I also find analyst coverage is negativ ely (p<0.05) associated with Forecast_Error . Similar to prior research (Bhushan , , 1990; Brennan and Hughes , 1991; Lang and Lundholm , 1996; Barth et al. , 2001; Lehavy et al. , 2011), I find Size is significantly and negatively (p<0. 05) associated with Forecast_Error . This implies that larger firms have richer information environments . Firms with greater R&D expenses have lower forecast accuracy (p<0.01) . T his finding is consistent with previous research document ing that analyst forec ast accuracy deteriorates as forecast complexity increases (Haw et al., 1994; Duru and Reeb, 2002; Lehavy et al. , 2011). While News is positively and significantly (p<0.01) associated with analyst following, I find that it does not have a significant impac t on forecast accuracy. This finding may indicate that the amount of information available about a specific firm increase s the demand for analysts as information intermediar ies, but that there is significant redundancy among articles provided by traditiona l media and that analysts do not view the redundancy to be informative. Management Forecasts is not significantly associated with analyst following . However , it is negatively and significantly (p<0.05) associated with analyst forecast error. This finding i mpl ies that financial analysts view management forecast s as value relevant voluntary disclosure s . 4.4.3. Forecast Dispersion 25 Table 7 presents multivariate regression results on forecast dispersion. Inconsistent with H3 , I do not find a significant negative assoc iation between firm social media disclosure and forecast dispersion. Th e combined finding s indicate that s ocial media usage is associated with 25 The multivariate regression model for the analysis of forecast dispersion has an adjusted R 2 value of 0.2 35 , which indicates that my model explains about 2 4 % of the variability of analyst forecasts dispersion around its mean . Lehavy et al. (2011), and Lang and Lundholm (1993) report adjusted R 2 of 0 .20 , and 0 . 42 , respectively. 40 improved forecast accuracy , but the additional information does not lead to a decrease in the heterogeneity of an alyst beliefs. As predicted , I find Following is negatively and significantly (p<0.01) associated with Forecast_Dispersion . T his finding is consistent with the prediction that analyst coverage is negatively related to the level of information asymmetry. N ews is not significantly associated with forecast dispersion . T his c ould be due to the redundancy of information or diversified information from media not reduc ing uncertainty in the prediction of future performance. I do not find significant associations between other control variables represent ing forecast complexity (i.e. R&D, Busin e ss_Seg, Geo_Seg ) and Forecast_Dispersion . 4.4.4. Market Reaction to Forecast Revisions 26 Table 8 reports descriptive statistics for the key variables used in the analysis of the mar ket reaction to forecast revisions. There are 25,835 revisions with complete data. The average number of revisions per firm quarter is 4.95. The mean and standard deviation of 2 - day cumulative abnormal returns , ( CAR(0,1) ) , in response to average forecast r evisions are 0. 00 5 % and 2.5 % , respectively. This indicates that, on average, there is a positive market response to analyst forecast revisions. The minimum and maximum values of CAR (0,1) are - 36.1 % and 29.7 % , respectively, implying that there is significa nt variation across firms in the direction and amount of financial information captured in analyst forecast revisions. The mean and standard deviation of Mean_AFRevise are - 0.001 and 0.011, respectively. The minimum and maximum values of News are 0 and 3.0 19, respectively which are 0 and 21 in raw values, implying that the amount of 26 The multivariate regression model for the analysis of the market response to analyst forecast revisions h as an adjusted R 2 of 0. 008 , which indicates that my model explains about 0.8 % of the variability in abnormal two - day abnormal ret urns in response to the forecast revision release . In contrast, the adjusted R 2 in Green et al. (2014) is 0. 23 . However, it is difficult to make a direct comparison between the explanatory power of the two models because Green et al. (2014) classify foreca st revisions into only two categories (upward revisions and downward revisions), while my study uses t he mean analyst forecast revision on day t scaled by stock price at the end of the prior month. 41 traditional media coverage of a financial news in between analyst forecast revisions varies significantly. The mean value s of Size , BtoM , Leverage are 9. 878, 0.468, and 0 .600, respectively. On average, 1.83 analysts issue revisions on the forecast revision date . T he average number of Form 8 - K s filed over the revision period is 0.263 , and about 10% o f sample firms issue at least one management forecast between the previous and current forecast revision date s. Table 9 reports multivariate regression results for the market reaction tests . Consistent with To examine the effect of analyst for ecast revision s on stock price discovery in the social media e ra , I first run the model with out the social media and social media interaction variables, Log_Financial_Tweet and AFRevise * Log_Financial_Tweet . I find a significant market reaction to analyst f orecast revisions , indicat ing that analysts remain a n important information intermediary in the social media disclosure era. In the second model, I include Log_Financial_Tweet , but not AFRevise * Log_Financial_Tweet . I again find that analyst forecast revisi ons have a significant association with two - day returns . The primary test of my fourth hypothesis is the third model , which includes t he mean forecast revision , t he financial tweets variable, and the interaction term . Both the mean revision (p<0.05) and the interaction term are positive and significant (p<0.1 0 ) . The significance of the interaction term using the two continuous variables implies that the higher the level of social media activity, the greater the association between analyst forecast revisio ns and the market response s to the revisions . In addition, the larger the forecast revision, the greater is the association between social media activity and the market response. The coefficient on the number of tweets is again insignificant. These findin gs suggest social media disclosures complement rather than substitute for the intermediary role of financial analysts, consistent with mosaic theory (Pozen , 2005) . 42 In addition to the variables of interest, I find News (p<0.05) and Size (p<0.01) are signif icantly positively and negatively associated with the market response, respectively. The coefficients on News are consistently positive and significant for all three specifications. This finding is in line with the significant and positive association with News and analyst following. Although News is not significant in either the Forecast_Error or Forecast_Dispersion models, these findings imply that investors still value information available from financial news articles by traditional news providers. Nega tive and significant coefficients on Size imply that bigger firms stock prices are less responsive to news releases because they have richer information environments , implying that more of the information in the release has already been impounded in price . 43 5. SUPPLEMENTAL ANALYS E S In Chapter 5, I perform several additional analyses to validate the main results and to provide enhanced perspectives about the main findings reported in Chapter 4 . I begin by estimating the analyst following, analyst forecas t error, and analyst forecast dispersion models for subsamples of firms in consumer - oriented versus non - consumer - oriented industries to further rule out concerns that the results of my main analyses are influenced by particular industr ies . I also examine t he impact of social media on the market response to analyst forecast revisions using a three - d ay Cumulative Abnormal Return s (CAR s ) window to show that the findings of my main analyses are robust to specification of the length of the event window. I conclu de by examining the impact of social media on the market response to analyst forecast revisions for the subsample s of firms that do versus do not issue management forecast s of EPS between the previous and current forecast revisions. 5.1. Analysis of Analyst Following, Analyst Forecast Errors, and Analyst Forecast Dispersion for Subsamples of Companies in Consumer - Oriented and Non - Consumer - Oriented Industries U nlike traditional disclosure channels that focus on investors, disclosure via social media focus es jo intly on investors and consumers, so often includes both fi nancial information and advertising . While the differen ce in audience can increase the risk of misinterpretation, advertising has the potential to engage investors as well ( Madsen & Niessner, 2016 ) . Therefore, it is worth investigating whether the impact of social media activity on analyst following and properties of analyst forecasts var ies if firms are consumer - oriented or not. In my main analyses, I include industry fixed effects to isolat e variance attributable solely to industry idiosyncrasy . In this analysis, to provide deeper understanding of the impact of 44 information shared through social media on analyst following and properties of analyst forecasts , I investigate whether the findings of the main analyses are sensitive to firms purpose for social media communication. To do so, I split the main sample into two subsamples consisting of observations in the two consumer - oriented industr ies ( Retail Trade and Services ) and observations in t he other non - consumer - oriented industries. I then re - estimate the main analyses for the two subsamples, separately . Table 10 , columns (1) and (2), present the results from estimation of the analyst following model in Table 5 for subsample s that include obs ervations from consumer - oriented and non - consumer - oriented industries . The impact of financial social media disclosure ( Financial_Twe e ts ) o n analyst following is significantly positive ( p<0. 1 0 for consumer - oriented firms and p<0.05 for non - consumer - oriente d firms ) , regardless of the industr y composition of the sub - sample s . The economic significance of Financial_Twe e ts is greater in the non - consumer - oriented subsample . One interpretation is that the firms in these industries provide more effective investor r elation s information via social media disclosure. Non - f inancial_Tweets is negative and significant (p<0.01) , but only for the subsample of firms in consumer - oriented industries . One interpretation is that , regardless of industries, financial social media d isclosure provides value relevant information to financial analysts. However, s ocial media disclosure is often context specific (Bla n kespoor , 2018) and a differen ce in the audience can increase the risk of misinterpretation ( Madsen & Niessner, 2016 ). T his may in turn reduce the incentives of financial analysts to follow a firm with greater nonfinancial social media disclosure in consumer - oriented industries. Together these findings indicate that social media disclosure provide s a channel for companies to co mmunicate with different groups of stakeholders at the same time. Crowd_Tweets is consistently not significant. News and Size have a positive association (all p<0.01) with analyst following for both subsamples , 45 as was true of the main analysis. Management forecasts is positive and significant (p<0.05) only for the subsample of non - consumer - oriented firms . Table 11 estimates the forecast error model presented in Table 6 for the two subsample s . With some exceptions , the results are qualitatively similar to those in Table 6 . As is true of the main analysis, Financial_Tweets is negative and significant (all p<0.1 0 ) for both subsamples. This indicates that financial information disclosed through social media provide value relevant information for financial anal ysts , regardless of whether the firm is in a consumer - oriented or non - consumer - oriented industry . Similar to the results from the main analyses, Non - f inancial_Tweets is not significantly associated with analyst forecast error s for the non - consumer - oriented subsample . However, Non - f inancial_Tweets is negatively and significantly (p<0.1 0 ) associated with analyst forecast error for the subsample of consumer - oriented industries. T his indicates that for the firms in the consumer - oriented industries , nonfinancial social media disclosures by the firm provide analysts with value relevant information. Crowd_Tweets is positive and significant ( p<0.01 for consumer - oriented firms and p<0.05 for non - consumer - oriented firms ) . This implies that information from the crowd m ay mislead or add noise to the information mosaic of analysts. A Size is negatively and significantly ( p<0.05 for consumer - o riented firms and p<0.1 0 for non - consumer - oriented firms ) associated with forecast error, while Loss (p<0.05 for consumer - oriented firms and p<0.01 for non - consumer - oriented firms) and ROA (p<0.1 0 for consumer - oriented firms and p<0.01 for non - consumer - ori ented firms) are positively and significantly associated with forecast error. These findings indicate that firms with superior performance have richer information environment s . Also , many of the control variables are 46 consistent with the results in the main model . An exception is Intangible Asset , which is significant only in the subsample of firms in consumer - oriented industries . Table 12 presents the analyst forecast dispersion analyses for th e two subsample s . None of the three types of social media discl osure is significantly associated with analyst forecast dispersion in either subsample . In contrast, the results in Table 11 indicate that forecast errors are lower, the higher the volume of Financial_Tweets . T ogether, t hese findings imply that s ocial medi a usage is associated with improved forecast accuracy , but that the additional information does not lead to a decrease in the heterogeneity of analyst beliefs. Overall, the subsample results are qualitatively similar to those reported in the primary analys es . Thus , I continue to find that financial information released by firms through Twitter is relevant to financial analysts , regardless of industr y. However, the significance of some control variables differs between two subsamples . 5.2. Analysis of the Market Response to Analyst Forecast Revision s using 3 - Day CAR s I also consider the sensitivity of my main results to the length of my event window . Gleason and Lee (2003) and Clement and Tse (2003) examine the market response to analyst forecast revisions using 3 - Day Cumulative Abnormal Return s (CAR s ) . In contrast, I use 2 - Day CAR s in my main tests. In this sub - section, I present market response results using 3 - day CAR s . Table 13 presents results from the estimation of the model in Table 9 with 3 - Day CAR s as th e dependent variable . The se results show that t he impact of analyst forecast revisions ( Mean_AFRevise ) is consistently significantly positive (p<0.01 for columns (1) and (2), and p<0.05 for column (3)) in all three specifications. T he coefficient on Log_Fi n ancial_Tweet is negative and significant (p<0.05) . Similar to the main analyses, the interaction term between analyst forecast revisions and social media disclosure ( Mean_AFRevise * Log_Fiancial_Tweet ) is positively and significantly (p<0.1 0 ) associated w ith the market response. Also similar to the main 47 analyses, the coefficient on the amount of financial information provided by registered formal news organizations ( News ) is positive and significant (p<0.01) , while the coefficient on Size is negative and s ignificant (p<0.01) . The c oefficients on the number of analysts who issu e a revision on the forecast revision date ( Total_Revise ) are now statistically significant (p<0.1 0 ) . 5.3. Analysis of the Market Response to Analyst Forecast Revision s for Subsamples with and without Prior Management Forecasts Previous studies find not only that stock price s significantly respond to management forecasts ( Baginski and Hassell, 1990; Rogers and Stocke n , 2005 ; Hirst, Koonce, and Venkataraman, 2008) but also that prior earning s forecasts by management influence subsequent financial analyst forecast revision s (Baginski and Hassell, 1990) . To control for the confounding effect of management forecast s , my main analyses include an indicator variable ( Management_Ind ) that equals 1 w hen there is at least one management forecast of earnings in the period between the previous analyst forecast revision and the current analyst forecast revision. T o more fully understand the potential influence of management forecast s on the market respons e to financial analyst earnings forecasts, I reexamine the market response to analyst forecast revisions for two subsample s which consists of observations that have a value of 1 and 0 for Management_Ind . Table 14 reproduces the results presented in Table 9 using th es e subsample s . Column (1) and column (2) present the result s for the subsample s with and without management forecast between the previous analyst forecast and the current analyst forecast revision, respectively. Analyst forecast revisions ( Mean _AFRevise ) and the interaction term between analyst forecast revisions and social media disclosure of financial information ( Mean_AFRevise * Log_Fi n ancial_Tweet ) are positively and significantly associated with 2 - day window CAR s for both subsamples ( all p< 0.0 5 for column (1) and all p<0.1 0 for column (2) ) . 48 Mean_AFRevise*Log_Financial_Tweet , the interaction term between the forecast revision and the volume of financial tweets, are positively and significantly (all p<0.1 0 ) associated with the market response to forecast revisions , regardless of whether a management forecast was issued between the previous analyst forecast and the current analyst forecast revision. In addition, the impact of the control variables does not vary much across the two subsamples. Fo r example, News is positively and significantly associated with the market response (all p<0.1 0 ) for both subsamples. This indicates that investors still acquire value relevant information from traditional news organizations. After noting that the coeffi cient on the interaction term is larger in the subsample of observations with no confounding release of a management forecast during the revision period, I construct a formal test of whether financial tweets have a larger impact on forecast revisions in th e absence of concurrent management forecast . I include a 3 - way interaction term of Mean_AFRevise*Log_Finanacial_Tweet*Management _ Ind in the model and re - estimate the results for the full sample. The 3 - way interaction term is significant and negative (p<0.0 5) , which indicates that financial tweets have a larger impact on the market response to the forecast revisions in the absence of concurrent management forecast. I nvestors rely more on financial information provided by firms through social media when there is less information disclosed by management in other format s such as management forecast s . 49 6. CONCLUSION This paper studies the effect of social media disclosure on the demand for financial analysts as information intermediaries . I measure the amount of social media disclosure by the firm as the number of tweets on financial and nonfinancial topics from the Twitter account and the amount of social media disclosure about the firm by the public as the number of tweets that contain . I find that analyst following is larger and forecast errors are smaller, the larger the number of financial tweets by the firm . A nalyst following is smaller, the larger the number of nonfinancial tweets by the firm. F orecast errors are larger, the larg e r the number of tweets by the public , while t he volume of tweets by the public is not significantly associated with analyst following. Forecast dispersion is unassociated any of the three social media disclosure measures. Collectively, the findings sugges t that only financial social media disclosure by the firm provides timely, value - relevant information to analysts. I also provide evidence on the relation between social media disclosures and the market response to analyst forecast revisions. The coefficie nts on the revision variable and the interaction between the revision variable and the log of the number of financial tweets variable are both positive and significant. The significance of the interaction term implies that the higher the level of social me the market response to the revisions . In addition, the higher the level of a forecast revision, the greater is the association between social media activity and market respo nse. The coefficient on the number of tweets is insignificant, consistent with the previous impounding of information in those tweets. 50 Results of this study also suggest that even though S&P 500 firms have significant media attention , interpretation of s ocial media information by analysts is valuable. Prior studies show that social media is an efficient conduit for disseminating information to financial markets and affects investor behavior. However, there has been little scrutiny of how information on so cial media affects the behavior and beliefs of sophisticated information intermediaries. To my knowledge, this is also the first study to examine concurrently the influence of social media disclosure s by firms and the public on financial analyst following and the properties of analyst earnings forecasts. In closing, I mention two limitations of the study. First, this study provides descriptive associations, but does not establish a causal relation between social media disclosure and either analyst followin g or the properties of analyst earnings forecasts. Proving causality would require knowledge of whether and how an individual analyst improves his (her) forecasting process owing to social media disclosure. However, the results from the two - day market reac tion tests are consistent with analysts using social media disclosures to form and revise their forecasts. Second, although I control for the number of news media articles about the firm, the number of Form 8 - K filings, and the release of management earnin gs forecasts , the documented association between social media disclosure and analyst following and the properties of analyst earnings forecasts might be due to other information sources, as opposed to the increased volume of information on Twitter. Despit e these caveats , overall, this paper increases our understanding of how social media affects sophisticated information intermediaries such as financial analysts. In particular, this study provides evidence that both financial analysts and social media disc losures contribute to the flow of information. 51 APPENDICES 52 APPENDIX A Variable Definitions and Data Sources 53 APPENDIX A. Variable Definitions and Data Source s Variable Definition Data Source Financial_Tweets Number of financial tweets from official Twitter account that are sent during the fiscal quarter and contai n financial key words . Twitter via CrimsonHexagon Non - financial_Tweets Number of non - financial tweets from a that are sent during the fiscal quarte r and do not contain financial key words . Twitter via CrimsonHexagon Crowd_Tweets C asthtag (i.e., $Ticker) sent by any account during the fiscal quarter . Crowd_Tweets is sc aled by 100. Twitter via CrimsonHexagon Following N umber of analysts who comprise the most recent I/B/E/S consensus quarterly earnings forecast prior to the quarterly fiscal period ending date . I/B/E/S Summary Forecast Error Absolute d ifference between I/B/E/S actual reported earnings and the most recent I/B/E/S quarterly earnings median consensus forecast prior to the quarterly fiscal period ending date , scaled by actual reported earnings . I/B/E/S Summary Forecast Dispersion Standard deviati on of the individual analyst forecasts in the most recent I/B/E/S quarterly earnings median consensus forecast prior to the quarterly fiscal period ending date . I/B/E/S Summary CAR(0,1) Abnormal daily returns cumulated over the two - day window beginning on the dat e that the forecast revision is released. Market - adjusted daily abnormal return s are calculated as the firm - specific return s less the CRSP value - weighted return s . CRSP Mean_AFRevise expected earnings on Day t , defined as the m ean analyst forecast revision on day t scaled by one - month prior stock price . For each individual sell - side analyst i, firm j and on day t , analyst forecast (AF) revision is measured as ( AF i, j, t - AF i, j, t - n ) , where AF i, j, t - n is the latest earnings forecast by analyst i for firm j prior to AF i , j, t. If there are multiple individual analyst I/B/ E/S Detail , CRSP 54 forecast revisions for firm j on day t , I use the average of the day t revisions . A nalyst forecasts and stock price are adjusted for stock splits. Log_Financial_Tweet Log of one plus the total number of the Financial Tweets between the prior analyst forecast and the current analyst forecast revision. Twitter via CrimsonHexagon Mean_AFRevise* Log_Financial_Tweet Interaction b etween Mean_AFRevise and Log _Financial_Tweet . CRSP, Twitter via CrimsonHexagon News Log of one plus the total n umber of financial news articles during the fiscal quarter that contain the name and financial key words . Obtained from CrimsonH source option , which provides full access - based articles by formal news organizations, such as CNN, New York Times, Wall Street Journal, In the market response tests, News is measured as the Log of o ne plus the total n umber of financial news articles that financial key words between the previous and the current analyst forecast revision dates . News via CrimsonHexagon Size Log of market value as of the ending date of the f iscal quarter. For the market response test s , m arket capitalization is as of quarter t - 1, where quarter t is the quarter in which the analyst forecast revision is released . Compustat Quarterly B to M B ook value of equity as of the end of the fiscal q uarter , divided by m arket value of equity as of the end of the quarter. In the market response test s , B to M is measured as of the beginning of the quarter in which the analyst forecast revision is released . Compustat Quarterly Loss I ndi c a tor v a ri a ble e q u a l t o 1 if quarterly net income is negative a nd 0 oth e r wis e. Compustat Quarterly ROA Return on assets , as of the end of the fiscal quarter. Calculated by dividing net income by total assets . Compustat Quarterly Leverage Financial leverage, calculated as th e ratio of total liabilities to total assets as of end of the fiscal quarter . In the market Compustat Quarterly 55 response test s , Leverage is measured as of the beginning of the quarter in which the analyst forecast revision is released . Intangible Asset T otal intangible assets scaled by total assets , measured as of end of a quarter . Compustat Quarterly R&D Quarterly research and development expense . M issing R&D expenses are set to be zero. Compustat Quarterly Bus_Seg Number of reported business segm ents as of the ending date of the previous fiscal year . Compustat Segments Geo_Seg Number of reported geographic segments as of the ending date of the previous fiscal year . Compustat Segments Institutional Level of institutional holdings. Percentag e institutions as of the calendar quarter ending date. Thomson Reuters Financial 13F data Horizon Number of days elapsed between the forecast date and the related earnings announcement date. I/B/E/S Summary 8 - K Number of 8 - K filings by a firm in each quarter. In the market response tests, 8 - K is measured as the number of 8 - K s filed by each firm between the prior analyst forecast and the current analyst forecast revision. EDGAR Management_Qtr Number of manage ment forecasts of EPS issued per quarter. I/B/E/S Guidance Management_Ind I ndi c a tor v a ri a ble e q u a ls t o 1 if there is at least one management forecast of EPS issued between the period of the previous and current forecast revisions, 0 oth e r wis e. I/B/E/S Gui dance Total_Revise The number of analysts who issu e a revision on the forecast revision da te . Total number of revisions is used to calculated Mean_AFRevise I/B/E/S Detail 56 APPENDIX B Industry Classification s 57 APPENDIX B . Industry Classification s 27 R ange of SIC Codes Division 0100 - 0999 Agriculture, Forestry and Fishing 1000 - 1499 Mining 1500 - 1799 Construction 1800 - 1999 not used 2000 - 3999 Manufacturing 4000 - 4999 Transportation, Communications, Electric, Gas and Sanitary service 5000 - 5199 Wholesal e Trade 5200 - 5999 Retail Trade 6000 - 6799 Finance, Insurance and Real Estate 7000 - 8999 Services 9100 - 9729 Public Administration 9900 - 9999 Nonclassifiable 27 See https://www.osha.gov/pls/imis/sic_manual.html and http://www.ehso.com/siccodes.php 58 APPENDIX C T ables 59 Table 1. Sample Selection The final sample consists of 4,9 74 firm - qu arter observations o n 435 S&P 500 firms over the period 2012 2014. Firm - Quarters Firms Firm quarter observations in 2012, 2013, and 2014 6,000 500 Less observations from firms without official Twitter accounts linked to their official company W ebsit e (780) (65) Less observations missing the data necessary to calculate the control variables ( 237 ) (0) Less: observations with missing data needed to estimate dependent variables (9) (0) Final sample used in the analyses 4,974 435 60 Table 2. Descr iptive Statistics This table reports descriptive statistics for the sample 4,9 74 firm - quarter observations o n 435 S&P 500 firms over the period 2012 2014 . Variable definitions are available in Appendix A. Variable Obs Mean Std. D ev. Min Median Max Financial_Tweets 4,974 143.07 5 35.34 1 9 2 . 000 133.000 211. 000 Non - financial_Tweets 4,974 267.35 9 7 8.869 135. 000 275.000 39 2 . 000 Crowd_Tweets 4,974 52.45 7 27.3 40 30. 000 37. 0 00 11 8 . 000 News 4,974 4.700 2.187 1.099 4.595 12.089 Following 4,974 18.317 7.72 8 3.000 1 9 .000 38.000 Error 4,974 0.13 7 0.28 6 0.000 0. 040 2.111 Dispersion 4,974 0.048 0.055 0.000 0.030 0.320 Horizon 4,974 12.840 12.480 8.000 12.000 31.000 Size 4,974 4.24 1 0.425 3.471 4. 288 5.411 BtoM 4,974 0.466 0.401 - 1.593 0.3 22 6.861 Loss 4,974 0.058 0.234 0.000 0.000 1.000 ROA 4,974 0.017 0.017 - 0.038 0.01 2 0.080 Leverage 4,974 0.618 0.200 0.146 0. 661 1. 65 9 Intangible Asset 4,974 0.71 7 0.134 0.309 0.596 0.9 14 R&D 4,974 107.070 306.388 0.000 31.804 1933.000 Bus iness_Seg 4,974 3.336 2.64 3 1.000 3 .000 15.000 Geo_Seg 4,974 1 .206 1 .207 1 .000 1 .000 1 2.000 Institutional 4,974 0.715 0.135 0. 309 0.725 1 .000 8 - K 4,974 3.599 2.3 59 0.000 3 .000 1 3 .000 Management_ Qtr 4,974 1.246 0.584 1.000 1.000 9.000 61 Table 3. Profil e Analysis This table reports mean and median firm characteristics for firms in the top and bottom quartile of Financial_Tweet volume in the first quarter of 2014. The numbers in parentheses denote the medians and z - stat s from the Wilcoxon rank - sum test s of quartile differences . The symbols ***, **, and * denote significance at 1%, 5%, and 10%, respectively. Variable s are defined in Appendix A. Financial_Tweets Financial_Tweets = Low 25% volume = High 25% volume Mean value of Mean value of t - test Variable (Median value of) (Median value of) (Wilcoxon test) Non - financial_Tweets 236 . 7 00 3 45.251 - 8.239 *** (0.000) (3 62 . 00 0) ( - 14.222) *** Crowd_Tweets 59 . 490 96 . 92 5 - 2.515 *** (32 . 0 00 ) (45 . 000 ) ( - 5.770) *** News 6.138 7.028 - 6.086 *** (6.216) (6.939) ( - 6.406 ) *** Following 17.023 20.059 - 2.957 *** (16. 000 ) (20.000) ( - 3.218) *** Error 0.117 0.123 - 0.249 (0.086) (0.070) (0.668) Dispersion 0.055 0.045 1.523 * (0.034) (0.030) (1.760) * Horizon 11.881 12.068 - 2.131 ** (1 2 . 00 0) (1 2 . 00 0 ) ( - 1.419) Size 4.196 4.439 - 4.310 *** (4.136) (4.332) ( - 4.102) *** BtoM 0.458 0.384 1.910 ** (0.396) (0.305) (2.231) ** Loss 0.063 0.052 0.493 (0.000) (0.000) (1.283) ROA 0.016 0.017 - 0.359 (0.014) (0.013) ( - 0.078) Leverage 0.580 0.669 - 3.597 *** (0.563) (0.676) ( - 3.947) *** Intangible Asset 0. 3 13 0. 3 12 0.035 (0. 3 61) (0. 3 38) ( - 0.459) R&D 62.767 171.775 - 2.522 *** (0.000) (0.000) ( - 0.656) 62 T able 3 ( c Financial_Tweets Financial_Tweets = Low 25% volume = High 25% volume Mean value of Mean value of t - test (Wilcoxon test) Variable (Median value of) (Median value of) Business_Seg 3.194 2.732 1.617 * (3.000) (1.000) (2.124) ** Geo_Seg 3.685 2.675 3.146 *** (3.000) (2.000) (3.147) *** Institutional 0.681 0.640 2.513 *** (0.699) (0.647) (2.925) *** 8 - K 3.468 3.925 - 1.805 ** (3. 00 0) (3. 00 0) ( - 1.024) Management _Qtr 1.245 1.000 0.842 (1.000) (1.000) 0.977 Observation 124 123 63 Table 4. P airwise Correlations among Variables Used in the Analysis This table reports P airwise correlations for the complete sample of 4,9 74 firm - quarter observations o n 435 S&P 500 firms over the period 2012 2014 . The coefficients in bold italic s are significant at least at the 5% level. Variable definitions are provided in Appendix A. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (1 ) Financial_Tweets 1.000 (2) Non - financial_Tweets 0.735 1.000 (3) Crowd_Twee ts 0.083 0.122 1.000 (4) News 0.272 0.276 0.290 1.000 ( 5 ) Following 0.165 0.167 0.163 0.367 1.000 ( 6 ) Error - 0.075 - 0.066 0.031 0.013 - 0.138 1.000 ( 7 ) Dispersion - 0.071 - 0.069 0.050 - 0.0 42 - 0.023 0.290 1.000 ( 8 ) Horizon - 0.034 0.009 - 0.375 0.044 - 0.005 0.020 - 0.004 1.000 ( 9 ) Size 0.156 0.136 0.288 0.455 0.334 - 0.140 0.064 - 0.047 1.000 ( 10) BtoM - 0.081 - 0.124 - 0.031 - 0.040 - 0.032 0.203 0.274 0.017 - 0.137 1.000 ( 11 ) Loss - 0.033 - 0.028 0.022 0.006 0.001 0.285 0.099 0.015 - 0.092 0.128 (1 2 ) ROA 0.053 0.083 0.024 0.077 0.113 - 0.212 - 0.105 - 0.002 0.136 - 0.460 (1 3 ) Leverage 0.080 - 0.017 - 0.014 0.054 - 0.17 7 0.022 0.027 - 0.005 - 0.031 0.063 (1 4 ) Intangible Asset 0.008 - 0.046 - 0.029 0.268 - 0.002 - 0.168 - 0.293 - 0.002 0.034 - 0.024 (1 5 ) R&D 0.019 0.015 0.229 0.271 0.216 - 0.053 - 0.028 - 0.021 0.457 - 0.13 9 (1 6 ) Business_Seg - 0.060 - 0.069 0.029 0.018 - 0.073 - 0.044 - 0.021 - 0.005 0.121 0.070 (1 7 ) Geo_Seg - 0.100 - 0.089 0.007 - 0.040 0.082 - 0.045 0.001 0.005 0.060 - 0.052 (1 8 ) Institutional - 0.115 - 0.097 - 0.213 - 0.243 0.030 0.085 0.031 0.104 - 0.360 - 0.079 (1 9 ) 8 - K 0.033 - 0.010 0.040 0.031 0.033 0.086 0.145 - 0.003 0.074 0.242 ( 20) Management _Qtr - 0.062 - 0.062 - 0.089 0.026 0.024 - 0.089 - 0.032 - 0.102 - 0.022 - 0.129 64 Table 4 ( c ont d) (11) (12) (13) (14) (15) (16) (17) (18) (1 9 ) (2 0 ) (1 ) Financial_Tweets (2) Non - financial_Tweets (3) Crowd_Twee ts (4) News ( 5 ) Following ( 6 ) Error ( 7 ) Dispersion ( 8 ) Horizon ( 9 ) Size ( 10) BtoM ( 11 ) Loss 1.000 (1 2 ) ROA - 0.448 1.000 (1 3 ) Leverage 0.044 - 0.275 1.000 (1 4 ) Intangible Asset - 0.038 0.034 - 0.135 1.000 (1 5 ) R&D - 0.002 0.110 - 0.141 0.093 1.000 (1 6 ) Business_Seg - 0.028 - 0.048 - 0.013 0.015 0.175 1.000 (1 7 ) Geo_Seg 0.004 0. 082 - 0.229 0.104 0.188 0.108 1.000 (1 8 ) Institutional 0.046 0.032 - 0.119 0.087 - 0.154 - 0.078 0.021 1.000 (1 9 ) 8 - K 0.056 - 0.196 0.206 - 0.114 - 0.050 0.013 - 0.059 - 0.061 1.000 ( 20) Management _Qtr - 0.043 0.201 0.041 - 0.111 - 0.101 0.031 - 0.133 0.039 0.056 1.000 65 Table 5. Analyst Following Regression Anal ysis This table reports the results of the negative binomial regression of analyst following on Twitter related variables and controls . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance t he readability of the results, the Twitter related variables and R&D are scaled by 1000 and 10,000 respectively . The numbers in parentheses denote robust standard errors. Variable definitions are available in Appendix A. VARIABLES (1) (2) Financial_Tweets 135.331 ** (55.255) Non - f inancial_Tweets - 0.076 *** (0.022) Crowd_Tweets - 0.002 (0.001) News 0.034 *** 0.035 *** (0.006) (0.006) Size 0.274 *** 0.303 *** (0.036) (0.038) BtoM 0.091 *** 0.071 *** (0.054) (0.0 0 5) Loss - 0.033 - 0.036 (0.049) (0.049) ROA 1.245 * 1.288 * (0.717) (0.722) Leverage - 0.635 *** - 0.640 *** (0.066) (0.065) Intangible Asset - 0.04 8 - 0.049 (0.058) (0.058) R&D 0.0 50 0.030 (0.034) (0.035) Business_Seg - 0.01 3 *** - 0.014 * ** (0.004) (0.004) Geo_Seg - 0.004 - 0.005 (0.004) (0.004) Institutional 0.529 *** 0.499 *** (0.091) (0.091) 8 - K - 0.004 - 0.004 (0.005) (0.005) Management _Qtr 0.004 0.002 (0.013) (0.013) Constant 1.611 *** 1.501 *** (0.194) (0.196) Observations 4,974 4,974 Pseudo R - squared 0.0 90 0. 092 66 Table 5 ( c ont d) Industry Indicators YES YES Time Indicators YES YES 67 Table 6. Analyst Forecast Error Regression Analysis This table reports the results of the ordinary least square regression of anal yst forecast error s on Twitter related variables and controls . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . T he Twitter related variables and R&D are scaled by 1000 and 10,000 respectively . The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES ( 1 ) ( 2 ) Financial_Tweets - 25.466 * (14.262) Non - f inancial_Tweets 0.003 (0.005) Crowd_Tweets 0.002 *** (0.000) News - 0.001 - 0.001 (0.003) (0.003) Following - 0.004 ** - 0.004 ** (0.000) (0.000) Horizon - 0.002 - 0.001 (0.002) (0.002) Size - 0.069 *** - 0.055 ** (0.024) (0.022) BtoM 0.109 *** 0.106 *** (0.034) (0.033) Loss 0.060 *** 0.053 *** (0.0 21) (0.020) ROA 1.342 *** 1.191 ** (0.498) (0.471) Leverage 0.044 * 0.037 (0.026) (0.026) Intangible Asset - 0.054 ** - 0.052 ** (0.022) (0.022) R&D 0.045 *** 0.047 *** (0.015) (0.015) Business_Seg - 0.003 - 0.002 (0.002) (0.002) Geo_Seg 0.001 0.002 (0.002) (0.002) Institutional 0.123 *** 0.124 *** (0.047) (0.047) 8 - K 0.008 *** 0.007 *** (0.002) (0.002) Management _Qtr - 0.009 ** - 0.009 ** (0.004) (0.004) 68 Table 6 ( c Constant 0.141 0.168 (0.102) (0.115) Observations 4,974 4,974 Adjusted R - squared 0. 214 0.216 Industry Indicators YES YES Time Indicators YES YES 69 Table 7. Analyst Forecast Dispersion Regression Analysis This table reports the results of ordinary least square regression of forecas t dispersion on Twitter related variables and controls . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance the readability of the results, the Twitter related variables and R&D are scaled b y 1000 and 10,000 respectively . The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES ( 1 ) ( 2 ) Financial_Tweets 8.781 (16.880) Non - f inancial_Tweets - 0.004 (0.004) Crowd_Tweet s 0.000 (0.000) News 0.001 0.001 (0.002) (0.001) Following - 0.001 *** - 0.001 *** (0.000) (0.000) Horizon 0.002 0.003 (0.002) (0.002) Size 0.004 0.003 (0.001) (0.011) BtoM 0.045 ** 0.045 ** (0.018) (0.018) Loss 0.039 0.038 (0.025) (0.025) ROA - 0.003 - 0.032 (0.410) (0.425) Leverage - 0.016 - 0.018 (0.015) (0.016) Intangible Asset - 0.042 *** - 0.042 *** (0.012) (0.012) R&D - 0.005 - 0.006 (0.006) (0.007) Business_Seg 0.002 0.002 (0.002) ( 0.002) Geo_Seg 0.001 0.001 (0.002) (0.002) Institutional 0.007 0.006 (0.016) (0.015) 8 - K - 0.000 - 0.000 (0.001) (0.001) Management _Qtr 0.003 0.003 (0.003) (0.003) Constant - 0.015 - 0.014 70 Table 7 ( c (0.069) (0.072) Observations 4,974 4,974 Adjusted R - squared 0. 224 0.2 35 Industry Indicators YES YES Time Indicators YES YES 71 Table 8 . Descriptive Statistics for Variables Used in the Market Response Analysis This table reports descriptive statistics for the variables used in the market response analys i s . Variable definitions are provided in Appendix A. Variable Obs Mean Std. Dev. Min Median Max CAR(0,1) 25,835 0.005 2.467 - 36.101 - 0.001 29.734 Mean_AFRevise 25,835 - 0.001 0.011 - 1.151 - 0.001 0.208 Mean_ AFRevise*Log_Financial_Tweet 25,835 0.000 0.010 - 0.393 0.000 0.400 Log_Financial_Tweet 25,835 1.474 1.469 0.000 1.386 8.304 News 25,835 1.243 1.187 0.000 1.946 3.019 Size 25,835 9.878 1.075 5.438 9.774 13.348 BtoM 25,835 0.468 0.365 - 0.441 0.378 2.567 Leverage 25,835 0.600 0.195 0.081 0.588 1.652 Total_Revise 25,835 1.830 2.141 1.000 4.000 27.000 8 - K 25,835 0.263 0.440 0.000 0.000 1.000 Management_Ind 25,835 0.099 0.299 0.000 0.000 1.000 72 Table 9 . Regression A nalysis of the Market Response to Anal yst Forecast Revisions Using 2 - Day CARs This table reports the results of the ordinary least square regression analysis of the market response to analyst forecast revisions . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - ta iled tests , respectively . To enhance the readability of the results, the dependent variable, CAR(0,1) , is multiplied by 100. The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES (1) (2) (3 ) Mean_AFRevise 12.492 * * * 12.564 * * * 7.756 * * ( 3.284 ) ( 3 . 284 ) ( 3 .692) Mean_AFRevise* 7.351 * Log_Financial_Tweet (4.204) Log_Financial_Tweet - 0.010 - 0.010 (0.011) (0.010) News 0.013 * 0.018 ** 0.018 ** (0.007) (0.0 07) (0.007) Size - 0.049 *** - 0.052 *** - 0.051 *** (0.016) (0.016) (0.016) BtoM 0.009 0.004 0.009 (0.047) (0.047) (0.047) Leverage 0.123 0.123 0.115 (0.080) (0.080) (0.079) Total_Revise - 0.004 - 0.005 - 0.005 (0.003) (0.003) (0.003) 8 - K 0.038 0.040 0.040 (0.053) (0.053) (0.053) Management _ Ind - 0.009 - 0.007 - 0.009 (0.081) (0.082) (0.081) Constant 0.376 ** 0.421 ** 0.413 ** (0.162) (0.164) (0.160) Observations 25,835 25,835 25,835 Adjusted R - squared 0.005 0.006 0.008 73 Table 10 . Analyst Following Regression Analysis for Subsamples of Consumer - Oriente d Industries and Non - Consumer - Oriented Industries This table reports the results of the negative binomial regression of analyst following on Twitter related and control variables for subsamples of consumer - oriented industries and non - consumer - oriented industries . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance the readabi lity of the results, the Twitter related variables and R&D are scaled by 1000 and 10,000 respectively . The numbers in parentheses denote robust standard errors. Variable definitions are available from Appendix A VARIABLES (1) (2) Consumer - Ori ented Non - Consumer - Oriented Financial_Tweets 12 2.512 * 346.949 ** (72 .312) (66.464) Non - f inancial_Tweets - 0.2 32 *** - 0.070 (0.073 ) (0.068) Crowd_Tweets - 0.002 - 0.001 (0.002 ) (0.001) News 0.030 *** 0.041 *** (0.01 0) (0.007) Size 0.41 7 *** 0.212 *** (0.052 ) (0.041) BtoM - 0. 047 *** 0.274 *** (0.00 8) (0.006) Loss - 0.043 - 0.008 (0.06 2) (0.051) ROA 2.471 *** - 0.675 (0.8 63) (0.733) Leverage - 0.557 *** - 0.880 *** (0.09 3) (0.072) Intangible Asset - 0.118 - 0.001 (0.09 0) (0.061) R&D 0.06 3 - 0.060 (0.0 51) (0.038) Business_Seg - 0.003 - 0.025 *** (0.005) (0.004) Geo_Seg 0.002 - 0.006 (0.00 7 ) (0.00 5 ) Institutional 0.2 89 * 0.670 *** (0.152 ) (0.106) 8 - K 0.004 - 0.010 * (0.009 ) (0.005) Manag ement _Qtr 0.00 1 0.030 ** (0.019 ) (0.015) Constant 2.585 *** 0.731 *** 74 Table 10 ( c (0.2 56) (0.201) Observations 906 4,068 Pseudo R - squared 0.087 0.105 Industry Indicators YES YES Time Indicators YES YES 75 Table 11 . An alyst Forecast Error Regression Analysis for Subsamples of Consumer - Oriented Industries and Non - Consumer - Oriented Industries This table reports the results of the ordinary least square regression of forecast error on Twitter related and control variable s for subsamples of consumer - oriented and non - consumer - oriented industries . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . T he Twitter related variables and R&D are scaled by 1000 and 10,000 respe ctively . The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES (1) (2) Consumer - Oriented Non - Consumer - Oriented Financial_Tweets - 27.012 * - 25.20 3 * (16.102) (14.42 2 ) Non - f inancial_Tweets - 0.0 12 * 0.00 5 (0.007) (0.00 5 ) Crowd_Tweets 0.005 *** 0.002 ** (0.002) (0.00 1 ) News - 0.003 0.000 (0.005) (0.00 3 ) Following - 0.006 *** - 0.00 1 ** (0.001) (0.000) Horizon 0.002 - 0.00 2 (0.003) (0.00 2 ) Size - 0.061 ** - 0.048 * (0.031) (0.0 28) BtoM 0.151 *** 0.05 5 (0.053) (0.037) Loss 0.049 ** 0.059 *** (0.025) (0.021) ROA 1.040 * 1.439 *** (0.603) (0.489) Leverage 0.050 0.035 (0.039) (0.029) Intangible Asset - 0.115 *** - 0.041 (0.010) ( 0.033 ) R&D 0.099 *** 0.03 7 ** (0.027) (0.017) Business_Seg - 0.002 - 0.002 (0.003) (0.00 2 ) Geo_Seg 0.001 0.00 2 (0.003) (0.00 2 ) Institutional 0.167 *** 0.101 ** (0.061) (0.0 50 ) 8 - K 0.008 *** 0.007 *** 76 Table 11 ( c Management _Qtr - 0.009 * - 0.009 ** (0.005) (0.00 4 ) Constant 0.201 - 0.097 (0.159) (0.101) Observations 906 4,068 Adjusted R - squared 0.223 0.216 Industry Indicators YES YES Time Indicators YES YES 77 Table 12 . Analyst Forecast Dispersion R egression Analysis for Subsamples of Consumer - Oriented Industries and Non - Consumer - Oriented Industries This table reports the results of ordinary least square regression s of forecast dispersion on Twitter related and control variables for subsamples of consumer - oriented and non - consumer - oriented industries . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance the readability of the results, the Twitter related variables and R&D are scaled b y 1000 and 10,000 respectively . The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES (1) (2) Consumer - Oriented Non - Consumer - Oriented Financial_Tweets 6.213 9.597 (24. 135) (17.6 78 ) Non - f inancial_Tweets - 0.003 - 0.004 (0.007) (0.005) Crowd_Tweets 0.000 0.000 (0.000) (0.000) News 0.001 0.00 2 (0.002) (0.001) Following - 0.001 *** - 0.001 *** (0.000) (0.000) Horizon 0.002 0.00 3 (0.003) (0.00 2 ) Size 0.002 0.00 5 (0.021) (0.01 4 ) BtoM 0.052 * 0.04 1 ** (0.031) (0.020) Loss 0.032 0.039 (0.036) (0.028) ROA - 0.036 - 0.028 (0.587) (0.445) Leverage - 0.014 - 0.02 5 (0.024) (0.01 8 ) Intangible Asset - 0.043 ** - 0.041 *** (0. 021) ( 0.014 ) R&D - 0.005 - 0.006 (0.012) (0.008) Business_Seg 0.003 0.002 (0.004) (0.00 2 ) Geo_Seg - 0.002 0.00 1 (0.003) (0.00 2 ) Institutional 0.033 - 0.00 6 (0.024) (0.01 7 ) 8 - K - 0.001 - 0.000 78 Table 12 ( c (0.002) (0.00 1 ) Management_Qtr - 0.001 0.003 (0.004) (0.003) Constant 0.005 - 0.04 5 (0.092) (0.08 4 ) Observations 906 4,068 Adjusted R - squared 0.217 0.233 Industry Indicators YES YES Time Indicators YES YES 79 Table 1 3 . Regression A nalysi s of the Market Response to Analyst Forecast Revisions Using 3 - day CAR s This table reports the results from ordinary least square s regression analysis of the market response to analyst forecast revisions using 3 - day CARs . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance the readability of the results, the dependent variable, CAR( - 1 , 1 ) , is multiplied by 100. The numbers in parentheses denote robust standard errors. Variable definitions ar e provided in Appendix A. VARIABLES (1) (2) (3) Mean_AFRevise 22.579 *** 20.353 *** 15.092 ** (7.682) (7.153) (6.296) Mean_AFRevise* 8.135 * Log_Financial_Tweet (4.822) Log_Financial_Tweet - 0.033 ** - 0.032 ** (0.013) (0 .013) News 0.012 * 0.031 ** * 0.032 ** * (0.008) (0.010) (0.010) Size - 0.049 *** - 0.078 *** - 0.077 *** (0.019) (0.021) (0.021) BtoM - 0.001 - 0.002 0.004 (0.055) (0.063) (0.063) Leverage - 0.148 - 0.099 - 0.099 (0.111) (0.133) (0.134 ) Total_Revise - 0.005 * - 0.006 * - 0.006 * (0.003) (0.004) (0.004) 8 - K 0.121 ** 0.095 0.096 (0.052) (0.061) (0.061) Management _ Ind 0.034 0.017 0.0136 (0.081) (0.095) (0.095) Constant 0.497 *** 0.735 *** 0.720 *** (0.187) (0.212) (0.212) Observations 25,835 25,835 25,835 Adjusted R - squared 0.005 0.005 0.006 80 Table 1 4 . Regression A nalysis of the Market Response to Analyst Forecast Revisions for Subsamples with and without Prior Management Forecasts Using 2 - Day C ARs This table reports results of the ordinary least square regression of the market response to analyst forecast revisions on Twitter variables and control variables for subsamples with and without concurrent management forecasts (MF) . The symbols ***, **, and * denote significance at 1%, 5%, and 10% based on two - tailed tests , respectively . To enhance the readability of the results, the dependent variable, CAR(0, 1 ) , is multiplied by 100. The numbers in parentheses denote robust standard errors. Variable definitions are provided in Appendix A. VARIABLES (1) (2) (3) With MF Without MF Full Sample Mean_AFRevise 10.032 ** 6.715 * 7.756 ** (4.821) (4.014) (3.692) Mean_AFRevise* 6. 0 35 * 8.474 * 6.513 * Log_Finanacial_Tweet ( 3 .6 5 3) (5.1 32) (3.895) Log_Finanacial_Tweet - 0.008 - 0.013 - 0.010 (0.015) (0.013) (0.010) Mean_AFRevise* - 1.452 * * Log_Finanacial_Tweet * Management _ Ind (0. 7 26) News 0.022 * 0.017 * 0.018 ** (0.013) (0.010) (0.007) Size - 0.065 *** - 0.032 * - 0.051 *** (0.025) (0.019) (0.016) BtoM - 0.013 0.022 0.009 (0.065) (0.053) (0.047) Leverage - 0.008 0.141 0.115 (0.135) (0.101) (0.079) Total_Revise - 0.006 - 0.005 - 0.005 (0.006) (0.005) (0.003) 8 - K - 0.021 0.054 0.040 (0 .068) (0.057) (0.053) Management _ Ind - 0.009 (0.081) Constant 0.513 ** - 0.083 0.411 ** (0.250) (0.192) (0.160) Observations 10,647 15,188 25,835 Adjusted R - squared 0.008 0.008 0.009 81 REFERENCES 82 REFERENCES Abdel - Khalik, A. R., & Ajinkya, B. B. (1982). Returns to informational advantages: The case of analysts' forecast revisions. The Accounting Review , 57 (4), 661 - 680. Agnew, J. R., & Szykman, L. R. (2005). Asset allocation and information overload: The influence of information display, asset choice, and investor experience. The Journal of Behavioral Finance , 6 (2), 57 - 70. Agrawal, A., Chadha, S., & Chen, M. A. (2006). Who is afraid of reg FD? The behavior and performance of The Journal of Business , 79 (6), 2811 - 2834. Antweiler, W., & Frank, M. Z. (2004). Is all that talk just noise? The information content of internet stock message boards. The Journal of Finance , 5 9 (3), 1259 - 1294. Ashbaugh, H., Johnstone, K. M., & Warfield, T. D. (1999). Corporate reporting on the internet. Accounting Horizons, 13 (3), 241 - 257. Asq uith, P., Mikhail, M. B., & Au, A. S. (2005). Information content of equity analyst reports. Journal of F inancial E conomics , 75 (2), 245 - 282. Bae, K. H., Stulz, R. M., & Tan, H. (2008). Do local analysts know more? A cross - country study of the performance of local analysts and foreign analysts. Journal of Financial Economics , 88 (3), 581 - 606. Baginski, S. P., & Hassell, J. M. (1990). The market interpretation of management earnings forecasts as a predictor of subsequent financial analyst forecast revision. The Accounting Review , 65 (1), 175 - 190. Baginski, S. P., Hassell, J. M., & Wieland, M. M. (2011). An examination of the effects of management earnings forecast form and explanations on financial analyst forecast revisions. Advances in Accounting , 27 (1), 17 - 25. Bagnoli, M., Levine, S., & Watts, S. G. (2005). Analyst estimation revision clusters and corpor ate events, Part I. Annals of Finance , 1 (3), 245 - 265. Baik, B., Cao, Q., Choi, S., & Kim, J. M. (2016). Local Twitter Activity and Stock Returns. Working Paper . Seoul National University, Texas Tech University, Sungkyunkwan University, & Rutgers Universit y . 83 Bailey, W., Li, H., Mao, C. X., & Zhong, R. (2003). Regulation fair disclosure and earnings information: Market, analyst, and corporate responses. The Journal of Finance , 58 (6), 2487 - 2514. Barnes, N. G., Lescault, A. M., & Wright, S. (2013). Fortune 5 00 are bullish on social media: Big companies get excited about Google+, Instagram, Foursquare and Pinterest. Center for Marketing Research; University of Massachusetts: Dartmouth, MA, USA . Barnes, N. G., and Daubitz, C. (2017) . Time for r eevaluation? Soc ial m edia and the 2016 Inc. 500 . Available at https://www.umassd.edu/cmr/socialmediaresearch/2017inc500 Barniv, R., & Cao, J. (2006). Analyst f orecast r evision and m arket p rice d isc overy f ollowing a ccounting r estatement. Working Paper . Kent State University & Florida Atlantic University . Barth, M. E., Kasznik, R., & McNichols, M. F. (2001). Analyst coverage and intangible assets. Journal of A ccounting R esearch , 39 (1), 1 - 34. Bartov, E., Faurel, L., & Mohanram, P. S. (2017). Can T witter h elp p redict f irm - l evel e arnings and s tock r eturns ? The Accounting Review , 93 (3), 25 - 57. Beaver, W., Cornell, B., Landsman, W. R., & Stubben, S. R. (2008). The impact of analysts' forecast errors and forecast revisions on stock prices. Journal of Business Finance & Accounting , 35 (5 6), 709 - 740. Beyer, A., Cohen, D. A., Lys, T. Z., & Walther, B. R. (2010). The financial reporting environment: Review of the recent literature. Journal of A ccounting and E conomics , 50 (2), 296 - 343. Bhagwat, V., & Burch, T. (2016). Pump it up? Tweeting to manage investor attention to earnings news. Working Paper . University of Oregon & University of Miami . Bhushan, R. (1989). Collection of information about publicly traded firms: Theory and evidence. Journal of Accounting and Economics , 11 (2 - 3), 183 - 206. Bhushan, R. (1989). Firm characteristics and analyst following. Journal of Accounting and Economics , 11 (2 - 3), 255 - 274. Blankespoor, E., Miller, G. S., & White, H. D. (201 3 ). The role of dissemination in market liquidity: The Accounting Review , 89 (1), 79 - 112. Blankespoor, E. (2018). Firm communication and investor response: A framework and discussion integrating social media. Accounting , Organizations and Society . Forthcoming reporting. Accounting Horizons , 16 (3), 233 - 243. Bollen, J., Mao, H., & Zeng, X. (2011). Twitter mood predicts the stock market. Journa l of Computational Science , 2 (1), 1 - 8. 84 Bonner, S. E., Hugon, A., & Walther, B. R. (2007). Investor reaction to celebrity analysts: The case of earnings forecast revisions. Journal of Accounting Research , 45 (3), 481 - 513. Bradshaw, M. T. (2009). Analyst inf ormation processing, financial regulation, and academic research. The Accounting Review , 84 (4), 1073 - 1083. Bradshaw, M., Wang, X., & Zhou, D. (2016). recommendation revisions . Working P aper , Boston Col lege, Ohio State University, & Baruch College. Brennan, M. J., & Hughes, P. J. (1991). Stock prices and the supply of information. The Journal of Finance , 46 (5), 1665 - 1691. Brennan, M. J., Jegadeesh, N., & Swaminathan, B. (1993). Investment analysis and the adjustment of stock prices to common information. The Review of Financial Studies , 6 (4), 799 - 824. Brennan, M. J., & Subrahmanyam, A. (1995). Investment analysis and price formation in securities markets. Journal of F inancial E conomics , 38 (3), 361 - 381. Brown, P., Foster, G., & Noreen, E. (1985). Security analyst multi - year earnings forecasts and the capital market . Studies in Accounting Research , 2 3 , Amer ican Accounting Ass ociatio n. Sarasota, FL. Brown, L. D. (1993). Earnings forecasting research: Its implications for capital markets research. International J ournal of F orecasting , 9 (3), 295 - 320. b lack b s ell s ide f inancial a nalysts. Journal of Accounting Research , 53(1), 1 - 47. Caccese, M. S. (1997). Ethics and the financial analyst. Financial Analysts Journal , 53 (1), 9 - 14. Cade, N. L. (2018). Corporate social media: How two - way disclosure channels influence investors. Accounting, Organizations and Society , Forthcoming Cao, Y., Keskek, S. , Myers, L. A. , & Tsang , A. (2014). The effect of media competition on analyst forecasts properties: Cross - country evidence. Working P aper . The Chinese University of Hong Ko ng, University of Arkansas, University of Tennessee, & York University. Chava, S., Kumar, P., & Warga, A. (2009). Managerial agency and bond covenants. The Review of Financial Studies , 23 (3), 1120 - 1148. Chen, H., De, P., Hu, Y. J., & Hwang, B. H. (2014). Wisdom of crowds: The value of stock opinions transmitted through social media. Review of Financial Studies , 27 (5), 1367 - 1403. Chen, H., Hwang, B. H., & Liu, B. (201 8 ). The consequences for financial markets . Work ing Paper . City University of Hong Kong, Cornell University, & Florida State University. 85 Clement, M. B., & Tse, S. Y. (2003). Do investors respond to analysts' forecast revisions as if forecast accuracy is all that matters? The Accounting Review , 78 (1), 2 27 - 249. Clement, M. B., Hales, J., & Xue, Y. (2011). Understanding analysts' use of stock returns and other analysts' revisions when forecasting earnings. Journal of Accounting and Economics , 51 (3), 279 - 299. Cotter, J., Tuna, I., & Wysocki, P. D. (2006). Expectations management and beatable targets: How do analysts react to explicit earnings guidance? Contemporary Accounting Research, 23 (3), 593 - 6 24. Crowley, R., Huang, W., & Lu, H. (2018). Discretionary d isclosure on Twitter. Working Paper . Singapore Ma nagement University, Hong Kong Polytechnic University, & University of Toronto. Culnan, M. J., McHugh, P. J., & Zubillaga, J. I. (2010). How large U.S. companies can use T witter and other social media to gain business value. MIS Quarterly Executive , 9 (4), 243 - 259. Das, S. R., & Chen, M. Y. (2007). Yahoo! for Amazon: Sentiment extraction from small talk on the web. Management Science , 53 (9), 1375 - 1388. Davidowitz, A. S. (2014). Abandoning the m osaic t heory: Why the m osaic t heory of s ecurities a nalysis c on stitutes i llegal i nsider t rading and w hat to d o about It. Wash. UJL & Pol'y , 46 , 281 - 303 . Deloitte Touche Tohmatsu . (2015). Managing s ocial m edia r isks to r eputation r isk: A h ot t opic on the b oard a genda. Available at https://www2.deloitte.com/content/da m/Deloitte/lu/Documents/risk/lu - managing - social - media - risks - reputation - risk - 03032015.pdf Dhaliwal, D. S., Radhakrishnan, S., Tsang, A., & Yang, Y. G. (2012). Nonfinancial disclosure and analyst forecast accuracy: International evidence on corporate social responsibility disclosure. The Accounting Review , 87(3), 723 - 759. Duru, A., & Reeb, D. M. (2002). International diversification and analysts' forecast accuracy and bias. The Accounting Review , 77(2), 415 - 433. Elliott, W. B., Grant, S. M., & Hodge, F. D. (2018). Investor reaction to $ Firm or # CEO use of Twitter for negative disclosures. Working Paper . University of Illinois at Urbana - Champaign & University of Washington. Elliott, W. B., Grant, S. M., & Hobson, J. L. (2017). Trader participation in disc losure: Implications of interactions with management. Working Paper . University of Illinois at Urbana - Champaign & University of Washington. 86 Ernst & Young. (2014). Ensuring s ocial m edia is on the b a genda. Available at https://www.ey.com/gl/en/issues/governance - and - reporting/audit - committee/ey - ensuring - social - media - is - on - the - boards - agenda Ettredge, M., Richards on, V. J., & Scholz, S. (2002). Dissemination of information for investors at corporate W ebsites . Journal of Accounting and Public Policy , 21 (4 - 5), 357 - 369. Evaristo, R., Adams, C., & Curley, S. (1995). Information load revisited: A theoretical model. IC IS 1995 proceedings , 18. Frankel, R., Kothari, S. P., & Weber, J. (2006). Determinants of the informativeness of analyst research. Journal of Accounting and Economics , 41 (1 - 2), 29 - 54. Fried, D., & Givoly, D. (1982). Financial analysts' forecasts of earn ings: A better surrogate for market expectations. Journal of Accounting and Economics , 4(2), 85 - 107. Gans, J. S., Goldfarb, A., & Lederman, M. (2017). Exit, t weets and l oyalty . Working Paper . University of Toronto. Givoly, D., & Lakonishok, J. (1979). Th e information content of financial analysts' forecasts of earnings: Some evidence on semi - strong inefficiency. Journal of Accounting and Economics , 1 (3), 165 - 185. Givoly, D., & Lakonishok, J. (1980). Financial analysts' forecasts of earnings: Their value to investors. Journal of Banking & Finance , 4 (3), 221 - 233. Gleason, C. A., & Lee, C. M. (2003). Analyst forecast revisions and market price discovery. The Accounting Review , 78 (1), 193 - 225. Gomez arrasco, P., & Michelon, G. (2017). The p ower of s takehol ders' v oice: The e ffects of s ocial m edia a ctivism on s tock m arkets. Business Strategy and the Environment , 26 (6), 855 - 872. Green, T. C., Jame, R., Markov, S., & Subasi, M. (2014). Access to management and the informativeness of analyst research. Journal o f Financial Economics , 114 (2), 239 - 255. Griffin, P. A. (1976). Competitive i nformation in the s tock m arket: An e mpirical s tudy of e arnings, d ividends and a nalysts' f orecasts. The J ournal of F inance , 31 (2), 631 - 650. Grossman, S. J., & Stiglitz, J. E. (198 0). On the impossibility of informationally efficient markets. The American E conomic R eview , 70 (3), 393 - 408. Guan, Y ., Lu, H., & Wong, M. F. (2012). Conflict - of - interest reforms and investment bank Journal of Accounting, Auditin g & Finance , 27 (4), 443 - 470. 87 Hales, J., Moon, J., & Swenson, L. (2018). A n ew e ra of v oluntary d isclosure? Empirical e vidence on h ow e mployee p ostings on s ocial m edia r elate to f uture c orporate d isclosures. Working Paper . Georgia Institute of Technology, Georgia State University, & University of Wisconsin, Milwaukee. Haw, I. M., Jung, K., & Ruland, W. (1994). The accuracy of financial analysts' forecasts after mergers. Journal of Accounting, Auditing & Finance , 9 (3), 465 - 483. Healy, P. M., Hutton, A. P. , & Palepu, K. G. (1999). Stock performance and inte rmediation changes surrounding sustained increases in disclosure. Contemporary A ccounting R esearch , 16 (3), 485 - 520. Heflin, F., Subramanyam, K. R., & Zhang, Y. (2003). Regulation FD and the financial inf ormation environment: Early evidence. The Accounting Review , 78 (1), 1 - 37. Hiltz, S. R., & Turoff, M. (1985). Structuring computer - mediated communication systems to avoid information overload. Communications of the ACM , 28 (7), 680 - 689. Hirst, D. E., Koonc e, L., & Venkataraman, S. (2008). Management earnings forecasts: A review and framework. Accounting Horizons , 22 (3), 315 - 338. Hope, O. K. (2003). Accounting policy disclosures and analysts' forecasts. Contemporary Accounting Research , 20 (2), 295 - 321. Hsu , L., & Lawrence, B. (2016). The role of social media and brand equity during a product recall crisis: A shareholder value perspective. International J ournal of R esearch in Marketing , 33 (1), 59 - 77. Huang, A. H., Zang, A. Y., & Zheng, R. (2014). Evidence o n the information content of text in analyst reports. The Accounting Review , 89 (6), 2151 - 2180. Huang, W., Lu, H., & Su, B. (2016). Green d isclosure? Social m edia and p rosocial b ehavior . Working P aper . Hong Kong Polytechnic University & University of Toron to. Huang, A., & Mamo, K. (2016). Do a nalysts r ead the n ews? Working paper . University of Waterloo. Irani, A. J., & Karamanou, I. (2003). Regulation fair disclosure, analyst following, and analyst forecast dispersion. Accounting Horizons , 17 (1), 15 - 29. Imhoff, E. A., & Lobo, G. J. (1984). Information content of analysts' composite forecast revisions. Journal of Accounting Research , 22 (2), 541 - 554. revisions. Journal of Financial Economics , 73 (3), 433 - 463. 88 Jacob, J., Lys, T. Z., & Neale, M. A. (1999). Expertise in forecasting performance of security analysts. Jo urnal of Accounting and Economics , 28 (1), 51 - 82. Jame, R., Johnston, R., Markov, S., & Wolfe, M. C. (2016). The value of crowdsourced earnings forecasts. Journal of Accounting Research , 54 (4), 1077 - 1110. Jegadeesh, N., Kim, J., Krische, S. D., & Lee, C. (2004). Analyzing the analysts: When do recommendations add value? The J ournal of F inance , 59 (3), 1083 - 1124. Jones, D. (2013, June 11). Survey finds social media gap between investors, companies. IR Web Report . Available at http://irwebreport.com/20130611/iros - vs - investors - social - media/ Jung, M. J., Naughton, J. P., Tahoun, A., & Wang, C. (2018). Do firms strategically disseminate? Evidence from corporate us e of social media. The Accounting Review . 93 (4), Forth c oming Kane, G. C., Alavi, M., Labianca, J., & Borgatti, S. (2014). Integrating s ocial n etworks and i nformation s ystems: A review and framework for research. MIS Quarterly , 38 (1), 275 - 304. Kaplan, A. M., & Haenlein, M. (2010). Users of the world, unite! The challenges and opportunities of s ocial m edia. Business H orizons , 53 (1), 59 - 68. Klein, A. (1990). A direct test of the cognitive bias theory of share price reversals. Journal of Accounting and Econo mics , 13 (2), 155 - 166. KPMG. (2015). Managing the r isks of s ocial m edia. Available at https://assets.kpmg.com/content/dam/kpmg/pdf/2015/06/managing - risks - s ocial - media.pdf Kross, W., Ro, B., & Schroeder, D. (1990). Earnings expectations: The analysts' information advantage. The Accounting Review , 65 (2), 461 - 476. Lang, M., & Lundholm, R. (1993). Cross - sectional determinants of analyst ratings of corporate disclosures. Journal of A ccounting R esearch , 31 (2), 246 - 271. Lang, M. H., & Lundholm, R. J. (1996). Corporate disclosure policy and analyst behavior. The Accounting review , 71 (4), 467 - 492. Lehavy, R., Li, F., & Merkley, K. (2011). The effect of annual re port readability on analyst following and the properties of their earnings forecasts. The Accounting Review , 86 (3), 1087 - 1115. Lee, L. F., Hutton, A. P., & Shu, S. (2015). The role of social media in the capital market: E vidence from consumer product reca lls. Journal of Accounting Research , 53 (2), 367 - 404. Lev, B., & Gu, F. (2016). The E nd of A ccounting and the P ath F orward for I nvestors and M anagers . John Wiley & Sons. 89 Libby, R., & Emett, S. A. (2014). Earnings presentation effects on manager reporting choices and investor decisions. Accounting and Business Research , 44 (4), 410 - 438. Livnat, J., & Zhang, Y. (2012). Information interpretation or information discovery: W hich role of analysts do investors value more? Review of Accounting Studies , 17 (3), 612 - 641. Ljungqvist, A., Marston, F., Starks, L. T., Wei, K. D., & Yan, H. (2007). Conflicts of interest in sell - side research and the moderating role of institutional investors. Journal of Financial Economics , 85 (2), 420 - 456. Loh, R. K., & Stulz, R. M. (20 18). Is s ell s ide r esearch m ore v aluable in b ad t imes? The Journal of Finance , 73 (3), 959 - 1013. Loughran, T., & McDonald, B. (2014). Measuring readability in financial disclosures. The Journal of Finance , 69 (4), 1643 - 1671 . Lys, T., & Sohn, S. (1990). The association between revisions of financial analysts' earnings forecasts and security - price changes. Journal of Accounting and Economics , 13 (4), 341 - 363. Madsen, J., & Niessner, M. (2016). Is investor attention for sale? The role of advertising in financi al markets. Working Paper . University of Minnesota & Yale School of Management. Mao, Y., Wei, W., Wang, B., & Liu, B. (2012, August). Correlating S&P 500 stocks with Twitter data. In Proceedings of the F irst ACM I nternational W orkshop on H ot T opics on I nt erdisciplinary S ocial N etworks R esearch , 69 - 72. ACM. Miller, G. S., & Skinner, D. J. (2015). The evolving disclosure landscape: How changes in technology, the media, and capital markets are affecting disclosure. Journal of Accounting Research , 53 (2), 221 - 239. Mohanram, P. S., & Sunder, S. V. (2006). How has regulation FD affected the operations of financial analysts? Contemporary Accounting Research , 23 (2), 491 - 525. Nagar, V., Nanda, D., & Wysocki, P. (2003). Discretionary disclosure and stock - based ince ntives. Journal of A ccounting and E conomics , 34(1 - 3), 283 - 309. of analyst services? Working Paper . Cornell University & University of Georgia . O' B rien , P. C. (1 988). Analysts' forecasts as earnings expectations. Journal of A ccounting and Economics , 10 (1), 53 - 83. O'Brien, P. C., & Bhushan, R. (1990). Analyst following and institutional ownership. Journal of Accounting Research , 28 ( Studies on Judgment Issues in Ac counting and Auditing) , 55 - 76. 90 Payne, J. L., & Thomas, W. B. (2003). The implications of using stock - split adjusted I/B/E/S data in empirical research. The Accounting Review , 78 (4), 1049 - 1067. Pozen, D. E. (2005). The mosaic theory, national security, an d the freedom of information act. The Yale Law Journal , 115 ( 3) , 628 - 679. Prokofieva, M. (2014). Twitter - based dissemination of corporate disclosure and the intervening effects of firms' visibility: Evidence from Australian - listed companies. Journal of Inf ormation Systems , 29 (2), 107 - 136. PWC . (2017). w orld: Enhancing t rust and t ransparency t hrough c ommunication. Available at https://www.pwc.com/us/en/cfodirect/publications/point - of - view/enhancing - trust - transparency - through - technology.html Ramnath, S., Rock, S., & Shane, P. (2008). The financial analyst forecasting literature: A taxonomy with suggestions for furthe r research. International Journal of Forecasting , 24 (1), 34 - 75. Rickett, L. K. (2016). Do financial blogs serve an infomediary role in capital markets? American Journal of Business , 31 (1), 17 - 40. Rock, S., Sedo, S., & Willenborg, M. (2000). Analyst follo wing and count - data econometrics. Journal of Accounting and Economics , 30 (3), 351 - 373. Rogers, J. L., & Stocken, P. C. (2005). Credibility of management forecasts. The Accounting Review , 80 (4), 1233 - 1260. Rosen, J. (2006). The people formerly called the audience. PressThink, Available at http://archive.pressthink.org/2006/06/27/ppl_frmr.html Ruiz, E. J., Hristidis, V., Castillo, C., Gionis, A., & Jaimes, A. (2012, February). Correla ting financial time series with micro - blogging activity. In Proceedings of the F ifth ACM I nternational C onfere nce on Web S earch and D ata M ining , 513 - 522. ACM. Securities and Exchange Commission. (2008). Commission guidance on the use of company W ebsites. SEC Release , 34 - 58288. Available at: https://www.sec.gov/rules/interp/2008/34 - 58288.pdf Securities and Exchange Com mission. (2013). Report of investigation pursuant to s ection 21 (a) of t he s ecurities e xchange a ct of 1934: Netflix, Inc., and Reed Hastings (Release No. 69279). Available at: https://www.sec.gov/litigation/investreport/34 - 69279.pdf Schipper, K. (1991). Analysts' forecasts. Accounting H orizons , 5 (4), 105 - 121 . Schick, A. G., Gordon, L. A., & Haka, S. (1990). Information overload: A temporal approach. Accounting, Organizations and Society , 15 (3), 199 - 220. 91 Smith, C. W., & Watts, R. L. (1992). The investmen t opportunity set and corporate financing, dividend, and compensation policies. Journal of F inancial Economics , 32 (3), 263 - 292. Speier, C., Valacich, J. S., & Vessey, I. (1999). The influence of task interruption on individual decision making: An informat ion overload perspective. Decision Sciences , 30 (2), 337 - 360. Sprenger, T. O., Tumasjan, A., Sandner, P. G., & Welpe, I. M. (2014). Tweets and trades: The information content of stock microblogs. European Financial Management , 20 (5), 926 - 957. Tang, V. W. (2018). Wisdom of c v ariation in the i nformativeness of t hird p arty g enerated p roduct i nformation on Twitter. Journal of Accounting Research . 56 (3), 989 - 1034 Trinkle, B. S., Crossler, R. E., & Bélanger, F. (2015). Voluntary disclosures via social media and the role of comments. Journal of Information Systems , 29 (3), 101 - 121. forecasts and earnings announcements. Journal of Financial Economics , 64 (3), 373 - 396. W omack, K. L. (1996). Do brokerage analysts' recommendations have investment value? The J ournal of F inance , 51 (1), 137 - 167. Yang, J. H., & Liu, S. (2017). Accounting narratives and impression management on social media. Accounting and Business Research , 47 (6), 673 - 694. Yang, J. H., Liu, S., & Zhou, D. (2016). Voluntary f inancial d isclosure on s ocial m edia: Does c orporate g overnance m atter? Working Paper . University of Reading. Yeldar, R. (2012). The value of extra - financial disclosure. What investors and analysts said. Radley Yeldar . Available at: https://www.globalreporting.org/resourcelibrary/The - value - of - extra - financial - disclosure.pdf Zhang, X., Fuehres, H., & Gloor, P. A. (2011). Predicting stock market indicators through T witter Procedia - Social and Behavioral Sciences , 26 , 55 - 62.