Herding in analysts recommendations: The role of media

Transcription

1 Herding in analysts recommendations: The role of media Bart Frijns Department of Finance Auckland University of Technology Thanh D. Huynh Department of Banking and Finance Monash Business School Monash University Version: August 6, 2016 Abstract: This study investigates the impact of media on analysts herding behavior when making stock recommendations. We find three main results. First, for firms with high news coverage, price reactions following analysts recommendation revisions that are away from the consensus are weaker than to those closer to it, indicating that the market recognizes analysts tendency to issue bold recommendations when the firm is intensively covered in the spotlight. Second, when the firm has negative media sentiment, the market reacts strongly to recommendation revisions that are away from the consensus consistent with the notion that the market believes that analysts have an incentive to herd following negative news sentiment. Third, disagreement in the media is associated with higher tendency to herd among analysts. These findings are robust to the confounding effect of news flows on returns as well as to alternative explanations. In addition, we find that experienced analysts tend to herd more following news with a negative tone, while those with investment bank affiliations and those covering high trading volume stocks tend to herd less after negative news. Keywords: security analyst, herding, news sentiment, news coverage, news dispersion B. Frijns ( bfrijns@aut.ac.nz) is from Department of Finance, Auckland University of Technology, New Zealand and T. D. Huynh ( thanh.huynh@monash.edu) is from Department of Banking and Finance, Monash University, Australia. We thank participants at the NZ Finance Colloquium 2016 and Victoria University of Wellington seminar for helpful comments. We gratefully acknowledge the support of SIRCA in supplying and explaining Thomson Reuters News Analytics data to us. All errors are our own.

2 Collective fear stimulates herd instinct, and tends to produce ferocity toward those who are not regarded as members of the herd. B. Russell 1 Introduction Security analysts tend to herd by issuing recommendation revisions that are close to the consensus (Welch, 2000; Jegadeesh and Kim, 2010). Scharfstein and Stein (1990), Trueman (1994) and Banerjee (1992) theoretically show that herding behavior can be rational due to analysts incentive structures or reputational concerns. What is relatively less known from the literature is, however, external factors that impact the herding decision of analysts. 1 This question is important because it has been shown that analysts play an important role in equity markets and their stock recommendations are a valuable source of information for investment decisions. For example, Loh and Stulz (2011) show that analysts stock recommendations can exert a significant impact on stock prices while Amiram, Owens, and Rozenbaum (2016) find that their outputs could reduce information asymmetry in the market. Since herding can jeopardize the information content of analysts outputs (Trueman, 1994), understanding what factors influence their herding behavior and, more importantly, whether the market is aware of such incentives is undoubtedly imperative. A growing strand of literature also acknowledges the important role of news media in financial markets. For instance, Tetlock (2007) and Tetlock, Saar-Tsechansky, and Macskassy (2008) show that negative news can predict future firms fundamentals as well as stock returns of S&P 500 companies. Tetlock (2010) finds that firm-specific news helps reduce information asymmetry in the market. Dzielinski and Hasseltoft (2014) document that dispersion in news sentiment is negatively related to future returns. Although the impact of firm-specific news on stock returns is well-documented, its impact on the behavior of market experts such as security analysts, whose recommendations can be influential in the market (Loh and Stulz, 2011; Ramnath, Rock, and Shane, 2008), is still not known. Given the importance of news as a source of market information, this study examines whether the media also influences analysts behavior. Specifically, we address the following questions: does the media have an impact on the herding behavior of analysts when making stock recommendations? And, 1 Hirshleifer and Teoh (2003) provide an excellent survey of the literature on herding behavior among security analysts. 1

3 if so, what types of analysts are more influenced by the media? By doing so, our study contributes to the growing literature of public news releases by being the first to document that news coverage, news sentiment, and media disagreement about the company s stock can simultaneously affect analysts herding behavior. We further contribute to the literature on analysts herding by showing that the impact of the media on herding depends on the characteristics of analysts. Using a comprehensive dataset of 2,256,238 news stories over the period from 2003 to 2012, we document a strong association between past news sentiment and future herding behavior of analysts. To detect herding among analysts, we employ the model developed by Jegadeesh and Kim (2010). This model is suitable for our research questions because it uses market price reactions to recommendation revisions to distinguish between herding effects and information effects. Our novel findings are that analysts tend to herd less when media coverage of a particular stock is high, while they tend to herd more when the tone of the media is negative and when there is a wide dispersion in the tone of the media. These results hold after a battery of robustness tests such as removing recommendations that could potentially piggyback on major corporate news (Loh and Stulz, 2011), press releases (Li, Shen, and Wu, 2015), as well as controlling for possible influence of some analysts on the media. Furthermore, we find that the media does not influence the recommendation revisions of all analysts equally. Specifically, experienced analysts tend to herd more following news that has a negative tone, while affiliated analysts and those covering high turnover stocks tend to herd less following negative news sentiment. These results are most consistent with the incentive hypothesis put forth in the literature and demonstrate that analyst recommendation revisions are conditional on prior news sentiment of the stock. Since news and the media are important sources of information, researchers and practitioners have long been interested in how they affect stock prices and the behavior of market participants. With the advent of modern news analytics technology, such as Thomson Reuters News Analytics, that quantifies the tone of a large number of news articles, it is therefore not surprising that research in this area has recently received considerable attention. Fang and Peress (2009), for example, examine the relation between media coverage and the crosssection of stock returns and find that stocks not covered by the media earn 3% per year more than those that are covered by the media. Antweiler and Frank (2006) use corporate news stories from the Wall St Journal (WSJ) and document short-run reversals after a news 2

4 publication. Tetlock (2007) analyzes the tone of words in the WSJ s Abreast of the Market and finds that negative words predict negative future stock returns. Similarly, Tetlock et al. (2008) show that negative words from the Dow Jones News Service and the WSJ can predict earnings and returns on S&P 500 firms. Engelberg (2008) considers the tone of a firm s earnings announcement and finds that this tone has stronger return predictability than quantitative financial measures such as earnings surprises. Similarly, Ahmad, Han, Hutson, Kearney, and Liu (2016) analyze the tone for 20 large U.S. firms and find that it contains fundamental information about the firm value. In addition to news tone, the literature has also examined the effect of other news characteristics on stock returns. For instance, Dzielinski and Hasseltoft (2014) find that news dispersion can predict negative future returns. Whereas prior research focuses on investigating the effect of a particular feature of news, our study examines how the three prominent aspects of news can simultaneously influence analyst herding behavior. First, news coverage, defined as the number of news articles covering the firm s stock over the quarter prior to the analyst recommendation revision, can affect the herding tendency of analysts via two competing hypotheses. On the one hand, public news can reduce information asymmetry in the market (Tetlock, 2010), as high news coverage may narrow the information gap among analysts. This information hypothesis posits that, because analysts information is correlated, they issue similar recommendations, and therefore predicts a positive association between news coverage and analysts herding. 2 On the other hand, when a stock attracts high media attention, analysts recommendations are more likely to receive attention as well. 3 Because being covered in the media enhances analysts career prospects (Rees, Sharp, and Twedt, 2015), they have an incentive to issue recommendations that are away from the consensus to stand out. This incentive hypothesis thus predicts a negative effect of news coverage on analysts tendency to herd. When stocks are intensively covered in the media, we find a significantly weaker market reaction to recommendation revisions that are away from the consensus. This finding suggests that the market recognizes that analysts tend to herd less following high media coverage of a stock consistent with the incentive hypothesis. Second, we examine the impact of news sentiment, measured by the tone of the news, on analysts herding tendency. Since news sentiment can affect the market s opinion about 2 According to the information hypothesis, analysts are not herding. Rather, they take similar actions because they have access to similar information set. 3 Barber and Odean (2008) use news coverage as a proxy for market s attention on the stock. 3

5 the stock (Tetlock, 2007), we hypothesize that it also has an impact on the tendency to herd among analysts. In addition, the impact of negative news sentiment could be different from that of positive sentiment. Prior research has documented that negative news sentiment coincides with downturn market, when analysts private information is also less precise (Garcia, 2013; Zhang, 2006; Loh and Stulz, 2014). This suggests that they could rely more on the consensus recommendation, and therefore have a stronger tendency to herd. 4 Following times of negative news sentiment, we find that the market reacts much more strongly to recommendation revisions that are away from the consensus than to those closer to it. This finding is consistent with the notion that the market recognizes analysts stronger tendency to herd when the stock s news sentiment is negative. The last feature of news that we examine is media disagreement about the firm s stock. When media disagreement is strong, the market is uncertain about the future prospect of the firm (Dzielinski and Hasseltoft, 2014). Investors thus tend to seek the opinion of analysts when making stock selections (Amiram, Landsman, Owens, and Subben, 2014). Because increasing attention to the analyst s recommendation revision comes at the time when their own information is also less precise (Loh and Stulz, 2014), they have a stronger incentive to herd toward the consensus. Motivated by Dzielinski and Hasseltoft (2014), we proxy for media disagreement by news dispersion, defined as the standard deviation of news tone scores of all articles related to the company s stock. Dzielinski and Hasseltoft (2014) show that this news dispersion measure is a good proxy for market disagreement. By allowing analysts herding tendency to be conditional on media disagreement, we find results that are consistent with the above prediction: when media disagreement is strong, market s reactions to bold recommendation revisions are stronger than to those that are closer to the consensus. This result suggests that the market is aware that analysts herd more when media disagreement is higher. Having documented the impact of news on the herding tendency among analysts, a second objective of our study is to examine which types of analysts are more influenced by the media. Motivated by the literature (for example, Chen and Jiang (2006)), we examine three primary characteristics of analysts: experience, investment banking affiliation, and analysts covering stocks with high trading volume. The incentive hypothesis posits that, because experienced analysts are more concerned about maintaining their reputation and 4 We discuss alternative hypotheses in detail in Section 2. 4

6 since they are more reluctant to issue negative information, they have more incentives to herd than other analysts following negative news sentiment. 5 Investment banking affiliation can represent another incentive to herd. 6 Because analysts, whose brokerage firm has an advisory relation with the company, have an interest in maintaining an optimistic view of the stock, the incentive hypothesis predicts that they are less likely to herd following negative news sentiment. Similarly, the incentive hypothesis also predicts that analysts, who cover stocks with high trading volume (which are typically associated with high trading commissions as argued by Chen and Jiang (2006)), are less likely to herd following negative news sentiment. Our results are consistent with all three predictions. We find that experienced analysts are more likely to herd following times of negative news sentiment. Affiliated analysts and those covering high trading volume stocks are, however, less likely to herd when the news sentiment is negative. 7 Overall, our study makes a number of important contributions to the literature on news analytics in accounting and finance, as well as the literature on analyst herding. First, Jegadeesh and Kim (2010) show that the market is aware of the analyst s incentive to herd when issuing stock recommendation revisions. What is still unknown in the literature, however, is the external determinants of this awareness. This research question is crucial because after all, investors are the ultimate users of analysts stock recommendations (Loh and Stulz, 2011). Thus, investigating when investors are able to correctly adjust prices for herding bias will extend our understanding as to when the consequences of herding are less severe and leading to lower pricing errors. To the best of our knowledge, we are the first to point out that the media is an important external determinant. 8 We find that the market is aware of analysts stronger incentive to herd when the media coverage is low, when the media sentiment is negative, and when disagreement in the media is strong. 5 An alternative hypothesis is that experienced analysts, who have been in the job for a long time, are more likely to be overconfident than inexperienced analysts. Thus, since overconfident analysts tend to overweight their private information, they do not herd toward the consensus in their recommendations. Another hypothesis is that experienced analysts are more likely to have skills and more private information;. They therefore do not herd based on public information. The prediction of these two hypotheses are in contrast to the incentive hypothesis. 6 Malmendier and Shanthikumar (2014) find that investment banking affiliation is an important factor that influences analysts recommendations while bank reputation, institutional ownership, and all-star status are not. 7 Section 4.4 discusses alternative explanations including the potential influence of some analysts on news sentiment. 8 Rees et al. (2015) is perhaps the closest study to ours, though their study, which is not about herding among analysts, focuses on comparing the characteristics of analysts who were featured in the media versus those who were not. 5

7 Second, in both analyst herding literature and the literature on news analytics in accounting and finance, our study is one of the first to simultaneously examine three prominent features of news flows on analysts herding tendency. Investigating the effects of all news characteristics simultaneously is important because the intensity of news coverage and the strength of media disagreement can affect an analyst s interpretation of the current news sentiment. This joint examination allows us to test whether the effect of news sentiment is subsumed once we control for the other news features, thereby providing the literature with a more complete picture on the role of the media in influencing analysts incentives to herd. Third, we contribute to the analyst herding literature by examining the types of analysts that are more influenced by the media when making recommendation revisions. Finally, we employ a large sample of firm-specific news with over 2.25 million news articles between 2003 and This compares favorably with 2,024 articles in Rees et al. (2015) and other prior studies on news analytics in accounting and finance. Tetlock et al. (2008) argue that one should examine all types of news by using as large a sample of news as possible because it limits the scope for dredging for anomalies. The next section formally develops the hypotheses for this study. Our study relies on the comprehensive news data from Thomson Reuters, which is described in Section 3. Section 4 reports results for the impact of media on analysts herding behavior and examines which types of analysts are more influenced by the media. We conclude in Section 5. 2 Hypothesis development We set two main goals for our study. The first is to examine the impact of the media on the herding behavior of analysts. The second is to examine which types of analysts are more influenced by the media. For the first objective, we investigate three main features of news that are likely to affect analysts recommendations, namely news coverage (defined as the number of news articles covering the company s stock over the past quarter), news sentiment (defined as the general tone of the media about the stock), and media disagreement (defined as the dispersion of news tone in the media). Prior research (e.g., Tetlock, 2010; Bushee, Core, Guay, and Hamm, 2010) has shown that news coverage can reduce information asymmetry in the market. High news coverage could 6

8 thus narrow the information gap among analysts as some private information of analysts has been revealed to the public. The information hypothesis posits that, analyst do not herd, rather, they issue similar recommendations because analysts have access to similar information. This hypothesis therefore predicts a positive relationship between news coverage and the similarity in stock recommendations of analysts. On the contrary, the incentive hypothesis posits that intensive news coverage indicates that the company s stock has attracted high attention from the media and the market. Rees et al. (2015) show that being covered in the media significantly benefits the analyst s careers. Thus, during times of high news coverage, analysts are more willing to issue bold recommendations that are away from the consensus so that their opinion could gain media attention. But even when the chance that the analyst s recommendation revision is covered in the media is very low in the midst of intensive news coverage, it is still less costly for the analyst to be wrong because her bold recommendation revision does not attract as much attention as that during the low coverage period. As a result, she still has a higher incentive to issue a bold recommendation revision. Indeed, Barron, Byard, and Kim (2002) find that the consensus among analysts forecasts decreases after an earnings announcement, suggesting that they put more weight on their private information when producing reports after the news. The incentive hypothesis predicts that analysts incentives to herd are lower when media coverage is high. Since the incentive hypothesis and the information hypothesis have opposite predictions, evidence on news coverage can provide an indication on whether analysts herd or they simply act based on similar information set. Hypothesis on news coverage (H1a): if analysts herd when making stock recommendations, the incentive to herd is reduced when the firm is intensively covered in the media. Alternative hypothesis of news coverage (H1b): the tendency to issue recommendation revisions toward the consensus is higher when the firm is intensively covered in the media. Since news sentiment affects the market s opinion about the stock as shown in Tetlock (2007) and others, we hypothesize that it may also have an influence on analysts recommendation revisions. In fact, Rogers and Grant (1997) analyze 187 sell-side analysts reports between 7

9 1993 and 1994 and find that about half of them use information not contained in annual reports. This finding suggests that analysts also rely on other external sources of information when producing recommendations. In this article, we point out that public news content, specifically the tone of the news coverage, is one of them. The impact of news sentiment on stock prices is, however, different for pessimism and optimism. Tetlock (2007) finds that media pessimism is a stronger predictor of future stock returns than optimism. Garcia (2013) documents that the effect of news sentiment is most pronounced during economic recessions. Loh and Stulz (2014) show that analysts find it difficult to evaluate short-term fundamentals during economic downturns, suggesting that the information from the consensus recommendation becomes more important. Similarly, Zhang (2006) finds that analysts forecasts are less accurate during down markets, indicating that they are more likely to be wrong. These studies together suggest that negative news sentiment could affect analysts recommendation revisions more than positive sentiment, and that it is risky for analysts to issue bold recommendations during negative sentiment periods. Furthermore, since analysts tend to herd more for downgrades (Jegadeesh and Kim, 2010, and others), they are even more reluctant to deviate away from the consensus when the firm has had negative news sentiment. This line of argument is similar in spirit to the incentive hypothesis: during negative sentiment periods, exaggerating differences from the consensus would make the analyst s revision more noticeable, and hence it would be more costly for her reputation and career if her recommendation later turns out to be wrong. By herding toward the consensus, her reputation is less likely to be damaged because, after all, she is not the only wrong in the crowd. In fact, Hong, Kubik, and Solomon (2000) find that being bold and wrong would hurt the analyst s career outcomes while being bold and correct does not significantly improve her career prospects. Another mechanism that is also in line with the incentive hypothesis is that analysts are reluctant to convey negative information to the market because it would hurt their business relationships with the company. This reluctance, however, could be alleviated when news sentiment is also negative because the analyst is not alone in issuing pessimistic recommendation. Indeed, Conrad, Cornell, Landsman, and Rountree (2006) show that if there are enough analysts issuing negative opinions on the stock, this may relieve the analyst from her conflict of interests with the firm. As a result, she is more likely to herd following times of pessimism in the media. On the effect of news sentiment, the predictions of the incentive hypothesis and the information hypothesis are similar. The information hypothesis 8

10 states that analysts issue similar stock recommendations following negative news sentiment because negative news could be more informative than positive news. 9 Similar to Welch (2000), without observing the real intention of analysts when making recommendations, we cannot differentiate the two hypotheses. Although it is not the goal of our study to test these two competing hypotheses, the findings from Hypothesis 1 can suggest which hypothesis is supported. Since Jegadeesh and Kim (2010) find that analysts have an incentive to herd when making stock recommendations, we state our second hypothesis as follows: Hypothesis on news sentiment (H2): analysts have a stronger incentive to herd following negative news sentiment of the firm. When media disagreement is high and investors are uncertain about the future prospect of the firm, they tend to seek the opinion of analysts (Loh and Stulz, 2014 and Amiram et al., 2014). Specifically, Loh and Stulz (2014) argue and show that during periods of high uncertainty, analysts information about the firms fundamentals is also less precise. Similarly, Amiram et al. (2014) find that analysts become more reluctant to make forecasts when uncertainty is high, perhaps because management provides less guidance. Navone and Zapatero (2015) find that analysts are averse to uncertainty and some may stop covering firms with high uncertainty. Because the increasing importance of the analyst s recommendation comes at the time when their accuracy is reduced, we hypothesize that they have even more incentives to herd when media disagreement is strong. Motivated by Dzielinski and Hasseltoft (2014), we proxy for disagreement by news dispersion, which is the standard deviation of news sentiment scores of all news articles related to the firm obtained from TRNA. Dzielinski and Hasseltoft (2014) show that, compared with analysts forecast dispersion, news dispersion is a better proxy for market uncertainty and a stronger predictor of future stock returns. An alternative hypothesis is that disagreement among investors as well as among analysts could be reflected in the media, thereby causing the news dispersion to be high. In this case, we should observe less herding behavior of analysts as analysts disagree about the future prospect of the firm. Finally, the information hypothesis posits that, if disagreement among journalists reflects disagreement among analysts, then the stronger the news dispersion the lower the tendency to herd. If the information hypothesis is supported, the impact of news 9 This line of argument also acknowledges the role of the media. 9

11 dispersion on herding should become insignificant once we remove news articles featuring the analysts recommendations themselves. Because the incentive hypothesis and the information hypothesis have contradicting predictions, the evidence on news dispersion impacts can help distinguish between herding and mutual immitation based on common information. We therefore have the following hypotheses: Hypothesis on news dispersion (H3a): when disagreement in the media is strong. the incentive to herd among analysts increases Alternative hypothesis on news dispersion (H3b): the tendency to herd among analysts decreases when disagreement in the media is strong. Previous research has shown that analysts are a heterogenous group in which herding behavior is conditional on their characteristics and incentive structures. To examine which type of analysts are more influenced by the media, we follow Chen and Jiang (2006) and examine three prominent characteristics of analysts, namely experience, investment banking affiliation, and those covering stocks with high trading volume. Our objective is to examine how these characteristics affect analysts herding behavior following negative news sentiment, which is the most important feature of the media. The first analysts attribute is experience. Experienced analysts may be more likely to herd than young analysts because being wrong is more costly to their reputation that they have been building for years. Since times with pessimistic media sentiment are particularly sensitive, experienced analysts are more likely to herd following the negative news sentiment of the firm. Indeed, Loh and Stulz (2014) find that it is faster for young analysts to build their reputation during bad economic times, suggesting that young analysts could herd less when the sentiment is negative. Another reason for experienced analysts to herd more during negative news sentiment periods is due to the pressure from their connections with the company. Experienced analysts tend to have established connections and career interests with companies they cover. In normal times, this connection prevents them from downgrading their stock recommendations. During times of negative news sentiment, since both the media and the market agree that the company s future prospects are dim, experienced analysts are 10

12 relieved from the pressure and more free to follow the consensus. This line of argument is, again, consistent with the incentive hypothesis. 10 Another alternative hypothesis is the overconfidence hypothesis, which posits that overconfident analysts overestimate the precision of their private signal about the company s stock, and therefore should not be affected by media sentiment. This overconfidence hypothesis arises from biased self-attribution in which analysts do not know their true ability and are more confident about their own private information and analysis but not about public information (Daniel, Hirshleifer, and Subrahmanyam, 1998; Gervais and Odean, 2001). Since experienced analysts are more likely to be overconfident in their job over time than inexperienced analysts, the overconfidence hypothesis predicts no relation between media sentiment and the herding behavior of experienced analysts. An alternative hypothesis is that experienced analysts are more likely to be skilled in their job and gaining more solid networks, which, in turn, gives them more private information. 11 Experienced analysts, therefore, tend to overweigh their private information and underweigh public information. Thus, this hypothesis, which we call the ability hypothesis, has a similar prediction to the overconfidence hypothesis. The second attribute is analysts investment banking affiliation. Analysts, whose brokerage firm has an advisory relation with the company, are rewarded for optimistic recommendations, which help generate underwriting businesses and trading commissions (Hong and Kubik, 2003). Since negative news sentiment is typically associated with pessimistic consensus, the incentive hypothesis predicts that affiliated analysts are less likely to herd following negative news sentiment. Finally, we follow Chen and Jiang (2006) and employ high trading volume as a proxy for stocks with potentially high commissions. The incentive hypothesis predicts that, analysts covering those stocks earn their commissions from generating interests in the stock, they are less likely to downgrade their stock recommendations, which are more likely during 10 The literature offers mixed predictions on the relation between experience and analysts herding tendency. The model of Scharfstein and Stein (1990) predicts that young and inexperienced analysts tend to herd more whereas, in the model of Prendergast and Stole (1996), those analysts are more likely to issue bold recommendations in order to make their names stand out from the crowd. The empirical findings are equally inconclusive. Hong et al. (2000) and Clement and Tse (2005) find that young analysts are less likely to herd. In contrast, Zitzewitz (2001), Chen and Jiang (2006) and Jegadeesh and Kim (2010) document that experienced are more concerned about their reputations and therefore herd more. 11 Prior research has shown that analysts with higher ability are more likely to stay longer in the profession (e.g., Stickel, 1992; Mikhail, Walther, and Willis, 1999; Hong et al., 2000). 11

13 pessimistic periods. Accordingly, we hypothesize that those analysts have lower incentive to herd following negative media sentiment. Hypothesis on analysts characteristics and herding (H4): analysts with different characteristics have different incentives to herd following negative news sentiment. There is no obvious reason to expect that analysts with the above mentioned attributes would have an opposite herding incentive to other analysts following times of high news coverage and media disagreement. If anything, the incentive hypothesis predicts that those analysts would have a stronger incentive to herd during times of strong media disagreement and low news coverage. For example, when the firm has already had intensive news coverage, the recommendation revision of analysts with an investment banking affiliation does not attract more attention than that of non-affiliated analysts. Since affiliated analysts are now more free, they are even more likely to issue bold recommendations that are away from the consensus. As a result, we hypothesize that affiliated analysts are less likely to herd than other analysts following intensive news coverage of the firm. The impact of media disagreement goes in the other direction. When the media disagreement about the firm is strong and investors are uncertain about the future of the firm, they tend to seek the opinion of affiliated analysts for a stock recommendation, with the belief that those analysts are more likely to possess private information. This makes it more costly for affiliated analysts to issue bold recommendations that stand out of the crowd. As a consequence, affiliated analysts are more likely to issue recommendations that are closer to the consensus following times of strong media disagreement. 3 Data The data employed in this study are the intersection of IBES, Thomson Reuters News Analytics (TRNA), and U.S. stock returns. In this section, we describe these data and provide summary statistics for our sample. 12

14 3.1 Analyst recommendations We obtain analyst recommendations from IBES. We use IBES s stopped recommendation file to remove revisions that are made after stopped dates. We employ daily returns for all U.S. stocks with share codes of 10 and 11 (excluding closed-end funds, Real Estate Investment Trusts, trusts, American Depository Receipts, and foreign stocks) from the Center for Research in Security Prices (CRSP) and accounting information from Compustat. IBES standardizes analyst recommendations ( Strong Buy, Buy, Hold, Sell, and Strong Sell ) and converts them to numerical scores where 1 is Strong Buy, 2 is Buy, etc. For ease of interpretation, we reverse IBES s recommendation scores so that 1 corresponds to a Strong Sell and 5 corresponds to a Strong Buy. Following Jegadeesh and Kim (2010), stocks must also satisfy the following criteria: (1) there should be at least one analyst who issues a recommendation for the stock and revises the recommendation within 180 calendar days; (2) at least two analysts, other than the revising analyst, should have active recommendations for the stock as of the day before the revision. A recommendation is considered active for up to 180 days after it is issued or until the IBES stopped file records that the analyst has stopped issuing recommendations for that stock; and (3) the firm should be covered in CRSP and TRNA. 12 Table 1 reports summary statistics for analyst recommendations between 2003 and After taking the intersection of various databases, we have 5,348 unique firms in our sample. The sample starts with 3,888 firms in 2003 and, by 2012, the number decreases to 3,129 firms. This decline in number of firms coincides with the global financial crisis during which many firms ceased to exist. The total number of unique analysts issuing stock recommendations is 4,279. This number varies year by year ranging from 1,391 to 2,041 analysts. Consistent with Barber, Lehavy, McNichols, and Trueman (2001), analysts are reluctant issue Strong Sell (on average about 4% of the recommendations between 2003 and 2012 and less than 8% in any given year) or Sell recommendations (on average about 8% over the whole sample period and less than 10% in any given year). The proportions of Buy and Strong Buy recommendations are much higher with averages of 25% and 18.6% of the total recommendations, respectively, indicating analysts tendency to give positive recommendations. We label each 12 The last requirement is consistent with the practice in the literature on news analytics in accounting and finance (for example, Tetlock (2007), Tetlock et al. (2008), Tetlock (2011), Loughran and McDonald (2011), Rees et al. (2015), and Li et al. (2015)). 13 The sample period is constrained by the availability of TRNA s data. 13

15 recommendation revision as an upgrade, downgrade, or reiteration by comparing the revised recommendation with the previous active recommendation for the stock by the revising analyst. Analysts have slightly higher tendency to issue upgrades than downgrades; nearly 40% of the total revisions are upgrades while about 37% of the revisions are downgrades. 3.2 Media coverage, sentiment, and news dispersion Our study takes advantage of the modern news analytics technology of TRNA, which is available for the period January 2003 to December Even though TRNA s services are used frequently by industry practitioners, it has only recently started to gain grounds in academic literature. For instance, Hendershott, Livdan, and Schurhoff (2015), Heston and Sinha (2014), and Sinha (2012) employ TRNA to study the informativeness of institutional trading and the market s reaction to firm-specific news. Li, Shen, and Wu (2015) show that the coverage of TRNA is comprehensive and that 92.7% of Compustat s earnings announcements (the most common source of this information for the U.S. market) are covered in TRNA. 14 Thomson Reuters collects and analyzes firm-level news from major news sources such as Dow Jones Newswires, the Wall St Journal, Reuters, and other regional newspapers. Specifically, Thomson Reuters quantifies the tone of each relevant news item using a proprietary algorithm and provides sentiment scores that we use in this study. For each news article, we compute its sentiment score as follows: NewsT one = positive negative (1) where positive and negative are the probabilities of the news being positive and negative, respectively, as calculated by Thomson Reuters (i.e., positive + neutral + negative = 1). These sentiment scores are computed based on a hybrid system of textual analysis methods including lexical analysis, linguistic parsing, and machine learning together with experts annotation of words. 15 Thomson Reuters also provides us with a relevance score of the news item, which measures how relevant the news item is to firm i. Intuitively, it is calculated by 14 See Hendershott et al. (2015), Heston and Sinha (2014), and Li et al. (2015) for a detailed description TRNA and its superiority over other conventional methods of quantifying news tone such as the bag-of-word approach, which simply counts the number of negative words without accounting for context and grammar. The Thomson Reuters White Paper outlining TRNA s algorithm is available upon request. 15 For the human annotation process, Thomson Reuters white paper states that each word is annotated by three experts who are working from a carefully prepared script. The words are presented to each annotator in a random order and the consensus taken as the final word score. 14

16 comparing the relative number of occurrences of the firm with the number of occurrences of other firms within the text of the news item. If the firm is mentioned in the headline, then the news item is deemed highly relevant to the firm and the relevance score is 1. For stories that mention multiple firms, the firm with the most mentions will have the highest relevance score. Thomson Reuters white paper states that the relevance score allows the distinction between the three cases: (1) when the relevance score is greater than 0.8, the company is one of the determinant players in the article; (2) when the relevance score is between 0.8 and 0.2, the firm is one of several mentioned substantively in the article ; and (3) when the relevance score is less than 0.2, the company is a minor player in the article. To ensure that the news item is highly relevant to the firm, we require that the relevance score to be greater than ,17 We compute the average tone (or sentiment) score of the media coverage for stock i (denoted by T one i,t ) as the average of all NewsT one s over the past quarter. We also calculate news tone dispersion for each stock, StdT one i,t, as the standard deviation of NewsT one over the past quarter. 18 We call this measure media disagreement, which Dzielinski and Hasseltoft (2014) show to be good proxy for market disagreement. Henceforth, we use the terms media disagreement and news dispersion interchangeably. Table 2 provides summary statistics for news flow. In total, there are over 2.25 million unique news stories for firms that met our sample requirements between 2003 and As each article could mention multiple firms in the content, the final dataset that is a result of merging across databases has a total of 9,396,270 observations. On average, 70% of the firms have news coverage in a given year, with the coverage getting more comprehensive towards the end of the sample. The media tends to generally have a positive tone, with an average T one score of About 75% of the stories have a T one score that is greater than zero, leaving about 25% of messages with a negative tone. The average news dispersion is 41%, suggesting that journalists have moderate disagreement in the tone of their coverage. 16 Our results are robust to using news items with 100% relevance scores. 17 Some large news items with lengthy body text are received by TRNA in parts. We only include the final take of the complete news in which Thomson Reuters provides the quantitative news scores for the whole article. By doing this, we avoid double counting large news items. 18 In unreported robustness tests, we also compute all the news measures over the past month and our results do not qualitatively change. 15

17 4 Results 4.1 Analyst recommendations and news In this section, we examine the relation between analysts recommendations and various news measures. These results provide a first indication of the relation between analyst recommendations and media coverage. Specifically, we consider the following regressions: Y i,j,t = b 0 + b 1 NewsCoverage i,t 1 + b 2 T one i,t 1 + b 3 StdT one i,t 1 + b Controls + ɛ i,j,t (2) where Y t = {Revision i,j,t, Consensus i,t, Deviation j,i,t }; Revision t is the difference between the new recommendation level and the previous recommendation level of analyst j for stock i; Consensus i,t is the consensus recommendation for stock i, which is computed as the average recommendation level of all analysts (Hong et al., 2000 and Jegadeesh and Kim, 2010); and Deviation j,i,t = (NewRec j,i,t Consensus i,t ) is the difference between the new recommendation level and the existing consensus (before the new recommendation of the analyst. NewsCoverage i,t 1 is the log of one plus the total number of news articles over the past quarter prior to the revision date; T one i,t 1 is the average of news sentiment scores of all news articles for a stock over the past one quarter prior to the revision date t; StdT one i,t 1 is news dispersion, which is the standard deviation of news tone scores for firm i over the past one quarter prior to the revision date. Controls i,t are past returns, dummy for lead analysts of Cooper, Day, and Lewis (2001), and average trading volume. Cooper et al. (2001) argue that lead analysts recommendations could influence the revisions of other analysts. We thus follow their methodology to identify lead analysts whose revision is more likely to be followed by other analysts, but lead analysts are less likely to revise following revisions by other analysts. Cooper et al. s (2001) methodology is briefly described in the Appendix A. Consistent with Jegadeesh and Kim (2010) and other prior studies, we pool all recommendations each a quarter and estimate Equation (2) using Fama and MacBeth s (1973) approach. Newey-West corrected standard errors with 30 lags are computed. Table 3 reports the coefficient estimates from Equation (2). We observe a positive relation between news coverage and recommendation revisions, suggesting that, on average, analysts increase their recommendations when a stock is covered more in the media. We also observe 16

18 a positive relation between the tone of the news and recommendation revisions, i.e. positive news coverage in the past results in increased recommendations. Lastly, we observe that when the disagreement in the media is high, analyst tend to revise their recommendations downwards. These results demonstrate that analyst recommendations are affected by the news provided by the media. When we look at the impact of news on the consensus recommendation, we observe that there is no effect. However, we do observe that tone and media disagreement have an positive and negative impact on the consensus, respectively. Finally, if we consider the impact of news on the deviation between new analyst recommendation and the consensus, we note that when stocks are covered more in the news, analysts tend to deviate from the consensus more. Similarly, we observe that analysts tend to deviate more from the consensus when news tone is positive. However, a larger disagreement in media results in deviations that are less far away from the consensus. These results clearly demonstrate that news affects deviations from the consensus, but may not directly indicate that news affects analyst herding. Analysts may deviate from the consensus due to e.g. private information signals they receive. Thus the observation that positive news tone leads to larger deviations from the consensus can be interpreted as either anti-herding behavior, or can be caused by private information that analysts may receive. The last three columns of panel B repeat regression (2), but replaces T one with NegNews, which takes the value of one if T one < 0 or zero otherwise (positive news sentiment is thus the opposite regression). The coefficient on NegNews is negative and statistically significant in all regressions. These results suggest that, following negative media sentiment of the firm, analysts tend to revise their recommendations downwards and closer to the existing consensus suggesting stronger herding behavior. Furthermore, the consensus recommendation level is also lower following negative news sentiment of the firm. 4.2 Market-based tests of herding To separate private information from herding, we employ the market-based test of herding developed by Jegadeesh and Kim (2010). An advantage of this approach is that it uses the market s reactions to recommendation revisions to detect herding. This allows us to distinguish between the effects of private signals and herding behavior. To understand this 17

19 point, it is worthwhile to briefly summarize the intuition of Jegadeesh and Kim (2010). 19 Without loss of generality, assume that a firm is releasing an important piece of good news during a conference call in which two analysts attend. Due to differences between the analysts in their pace of writing reports and conducting analysis, the faster analyst issues an upgraded recommendation level (e.g., from hold to buy) before the other analyst. Since market prices incorporate all available public information, by the time the second analyst completes her analysis, the stock price has already increased to reflect the information revealed through the first analyst s upgrade. Thus, absent further news from the firm, the second analyst does not need to revise her recommendation based on the information she received, as this is already incorporated into the market price. Based on this intuition, Jegadeesh and Kim (2010) show that the market s reactions to analyst recommendations can serve as a useful detection tool for herding behavior. 20 A major advantage of this market-based test is that it is easier to address the endogeneity concerns where the effect of news flows on stock returns can potentially confound the relation between analysts herding and news flows. Using market reaction as a detection tool, these confounding effects can be easily controlled for in a Fama-MacBeth regression framework (Tetlock, 2010, 2011). As in Jegadeesh and Kim (2010), after a recommendation revision for stock i on date t, we compute the H-day buy-and-hold abnormal returns, ABR i (t, t + H), as follows: t+h t+h ABR i (t, t + H) = (1 + R i,τ ) (1 + R m,τ ), (3) τ=t τ=t where R i,τ and R m,τ are the return on stock i and the value-weighted index return, respectively. As suggested by Jegadeesh and Kim (2010), we examine the windows H = {0, 1, 2, 21, 42, 126}. The model of Jegadeesh and Kim (2010) tests whether the market recognizes analysts tendency to herd by examining the stock price reaction following recommendation revisions. Specifically, we run the following regression to examine the presence of herding behavior 19 Jegadeesh and Kim s (2010) Appendix B describes this stylized model in detail. 20 Note that this intuition does not apply to earnings forecasts. In contrast to recommendations, earnings forecasts can be revised based on stale information because analysts are concerned about how far away their forecasts are from realized earnings. The revised forecast could reflect the analysts information about the firm s potential earnings and thus convey new information regardless of whether analysts herd or anti-herd. Thus, one cannot infer herding from the market s reaction to analysts revised forecasts. This is another advantage of our study, which focuses on herding in recommendations. 18