Managerial Incentives and the Language in Management Forecast Press Releases

Transcription

1 Managerial Incentives and the Language in Management Forecast Press Releases Stephen Baginski + University of Georgia Elizabeth Demers * INSEAD Chong Wang Naval Postgraduate School Julia Yu # Nanyang Technological University February 2012 Please do not Quote Without the Authors Permission *We thank seminar participants at ESSEC, the 2011 American Accounting Association Annual Meeting, the 2011 European Accounting Association Annual Meeting, the 7 th Interdisciplinary Workshop on Intangibles, Intellectual Capital & Extra-Financial Information in Warsaw, Ben Ayers, Stephen V. Brown (AAA discussant), John Hassell, Karol Marek Klimczak (EAA discussant), Claudine Mangen, and Eric Yeung for helpful comments. We appreciate the programming assistance provided by Rajiv Jayaraman of Knolscape as well as Rashid Ansari. We thank Mary Boldrini, Pascale Gadroy, and Cécile Maciulis for their research assistance, and Craig Carroll for processing our text files using the Diction software. Demers is grateful for the research funding provided by the INSEAD Alumni Fund. + Terry College of Business, University of Georgia, Athens, GA , baginski@terry.uga.edu * INSEAD, Boulevard de Constance, Fontainebleau Cedex, France, liz.demers@insead.edu Naval Postgraduate School, Monterey, CA 93943, cwang@nps.edu # Nanyang Business School, Nanyang Technological University, S3-01A-07 Nanyang Avenue, Singapore , juliayu@ntu.edu.sg

2 2 Abstract Using a sample of 2,085 voluntarily-provided unbundled management earnings forecasts in 1,372 unique press releases during the period, we find significant incremental pricing effects for linguistic sentiment included in management forecast press releases. Our setting is distinct from those examined in prior linguistic studies in that forecasts are voluntarily issued, incentive-driven disclosures that are unaudited, unstructured, and likely noisier than earnings realizations. The pricing effects that we document are strong relative to the accompanying hard management forecast news and stronger when the linguistic sentiment is provided in a pure forecast setting and not bundled with an earnings release. Sentiment and the unexpected earnings conveyed by the hard forecast are positively correlated, suggesting that managerial language is being used to reinforce the information contained in the hard forecasts. Consistent with many management incentives to strategically delay bad news and inflate good news, and with cheap talk models that entail a cost to lying, we document that investors react more strongly to net negative relative to net positive sentiment (incremental to a stronger reaction to bad hard forecast news). Net optimism is decreasing in the probability of litigation, the likelihood of financial distress, and in expected political costs, but greater for annual forecasts, for more frequent forecasters, and post-regulation Fair Disclosure. After controlling for hard news and the stronger pricing effects of net negative sentiment, our capital market tests find little evidence that investors view sentiment as being biased when such situational incentives are in play, except for some evidence of the discounting of the typically more optimistic sentiment in annual forecasts. Furthermore, our findings are consistent with the threat of litigation and more frequent forecasting rendering sentiment more credible. Finally, we detect a positive interaction effect between sentiment and hard forecast news, consistent with the use of language to reinforce rather than overturn the hard news, and with the market s ascribing credibility to managerial linguistic sentiment. In summary, aside from the finding that net pessimistic language is strategically delayed (or that net pessimistic language is more credible), most of our evidence is not consistent with managers using non-credible language to inflate prices, and taken as a whole, is consistent with the language in management forecasts playing a significant role in the price formation process. JEL Classifications: G14; D82; M41 Keywords: Management forecasts; linguistic sentiment; management incentives

3 3 I. Introduction A growing body of literature seeks to understand the market price relevance of language contained in corporate filings, incremental to the simultaneously released quantitative or hard news. To our knowledge, extant research on the price relevance of management s linguistic sentiment is limited to the context of mandatory filings, and prior empirical studies generally do not consider conditions under which language might be driven by managerial incentives. 1 We contribute to the literature by examining the role of language in the largely unstructured, voluntary disclosure context of management forecast press releases. Using a Factiva search, we identify a sample of 2,085 voluntarily-provided unbundled management earnings forecasts in 1,372 unique press releases during the period from which we derive linguistic measures of sentiment. 2 We hypothesize and find that linguistic tone is associated with situational incentives, language reinforces the news in the hard forecast surprise, linguistic sentiment is more important in the price formation process when it accompanies noisy, unstructured forecasts relative to more well-defined earnings realizations, and the market responds differentially to net optimistic versus net pessimistic sentiment but otherwise views linguistic sentiment as credible and does not discount language that is issued in the context of situational incentives. Overall, we conclude that language is unlike other bundled disclosures that have been used strategically to offset accompanying hard news, but rather that language plays an important, directionally consistent role in the price formation process when issued with voluntary management forecasts. Management forecast press releases are fundamentally different from mandatory filings that contain relatively well understood, more clearly defined, lower noise earnings or other (e.g., restatement) information. In the voluntary forecast setting, managers convey expectations that are inherently less 1 Prior financial linguistic studies have examined the price relevance of soft information in the context of mandatory filings such as earnings announcements (Davis, Piger and Sedor (2012); Demers and Vega (2011)), restatement announcements (Mangen and Durnev (2010)), IPO prospectuses (Balakrishnan and Bartov (2010)), and MD&A and other elements of the 10K reports (Davis and Tama-Sweet (2011); Feldman, Govindaraj, Livnat and Segal (2010); Li (2010)). 2 Following the literature, we define unbundled management forecasts as those that are issued in isolation (i.e., not bundled with quarterly or annual earnings announcements). As described in Section III, we derive linguistic measures of sentiment using both the finance-oriented dictionaries of Loughran and McDonald (2011) and the Diction 6.0 algorithm.

4 4 reliable than audited or externally reviewed realizations. Furthermore, prior research has identified situations where management forecasts appear to be strategic. For example, Rogers and Stocken (2005) link management incentives to management forecast bias when the bias is difficult to detect, and show that a rational securities market adjusts for this bias. Waymire (1984) and Rogers and Van Buskirk (2011) document that the sign of forecast news disproportionately conflicts with the sign of accompanying earnings realization news, suggesting a potential strategic role for supplying multiple earnings-related signals. Other research detects larger price reactions to bad forecast news, indicating either that investors view managerial forecasts of good news to be less credible (Hutton, Miller and Skinner (2003)) or that managers strategically delay bad news (Kothari, Wysocki and Shu (2009)). The uniqueness of the management forecast press release setting and the potential effects of management incentives on strategic behavior and disclosure credibility provide a unique opportunity to test whether language is priced in the forecast setting, whether it is differentially priced in the forecast setting relative to the earnings release setting, and whether management incentives affect both linguistic sentiment and the market s assessment of linguistic sentiment credibility. 3 Demers and Vega (2011) argue that linguistic sentiment may be qualitatively similar to "cheap talk" (i.e., it is costless to convey and difficult to verify, even ex post), and thus, absent strong management incentives to build a reputation for credible communication of linguistic tone and institutional factors to enhance its credibility, one would expect that linguistic sentiment would not be priced. However, they document an incremental pricing effect for sentiment in quarterly earnings releases. We provide 3 Several additional benefits are gained by using a management forecast press-release setting. First, with respect to providing additional context-specific evidence that language is priced, management forecast press releases arguably provide a less confounded and thus more powerful setting for at least two reasons: a) because management forecasts are often issued in relative isolation (or one can sample them, as we have, to obtain this quality), the incremental effect of language and cross-sectional differences in that effect are detected with greater precision, and the interpretation of the quasi-experiment suffers far less from the potential correlated omitted variables problem; and b)because the voluntary nature of, and costs associated with, public forecast disclosure increase the likelihood that information conveyed, both numerical and linguistic, is intended to change investor beliefs rather than to merely accompany mandated disclosures. Second, unlike mandatory releases, certain management forecast characteristics such as forecast horizon and frequency that are potentially associated with the incentives for linguistic optimism vary across observations. Third, when we test the delay of bad sentiment, Kothari et al. (2009) favor the use of management forecasts instead of earnings announcements because the forecast setting provides an unconditional test of the delay hypothesis.

5 5 additional evidence that sentiment is significantly positively related to both management earnings forecast release announcement period and post-announcement drift returns incremental to the hard forecast surprise, and that sentiment offers significant explanatory power for returns relative to the hard forecast news in the management forecast setting. Most interestingly, we document that the information content of linguistic sentiment is significantly greater when it accompanies the noisier and less well-structured forecasts relative to settings where well-understood (and eventually audited) earnings are the main news event in the announcement. The finding of a strong role for linguistic sentiment in a forecast setting (and a stronger role when the forecast is not bundled with an earnings release) calls into question whether sentiment exhibits the substantial incentive-related biases that characterize other voluntary disclosures. This motivates us to examine whether various situational incentives help to explain the linguistic sentiment in forecasts, and whether investors ultimately price managerial sentiment conditional on those incentives. We find that, consistent with managers having incentives to adjust market expectations with the issuance of management forecasts, there is a positive correlation between the sign of sentiment and the sign of unexpected earnings conveyed by the hard forecast. The results indicate that managerial language is being used, on average, to reinforce, rather than to overturn, the information contained in the hard forecasts. Furthermore, these findings suggest that the incentives driving sentiment conveyance are not the same as the incentives that drive the strategic bundling of negatively correlated news discovered by Waymire (1984) (disproportionately more good news disclosures with bad news management forecasts) and Rogers and Van Buskirk (2009) (disproportionately more good news management forecasts with bad news earnings). We next consider the overriding management incentive to maximize their firm s security price, which could lead to overoptimistic language or a delay of bad news. We hypothesize and find that the mean price reaction to net negative sentiment is greater in absolute magnitude than the mean price reaction to net positive sentiment. These results suggest either that investors view negative sentiment as more

6 6 credible (as in Hutton et al (2003)) or that managers strategically delay negative sentiment similar to the delay of bad hard news documented by Kothari et al (2009). We extend our investigations to examine whether net optimistic language is associated with numerous other situational incentives that prior literature has documented to be important in the context of voluntary disclosures. We find that sentiment is decreasing in the probability of litigation, suggesting that managers pay attention to language s potential legal exposure as documented by Rogers and Van Buskirk (2011). Contrary to expectations, sentiment is found to be decreasing rather than increasing in the likelihood of financial distress. This result is consistent with managers credibly expressing their firm s negative financial condition rather than fulfilling a self-serving bias by trying to put a positive spin on a bad situation. We find no evidence of managers using less optimistic sentiment to avoid product-marketrelated costs, however sentiment is found to be decreasing in firm size, consistent with the argument put forth by Li (2010) that political costs dampen optimistic language. We detect more optimistic language in forecasts of annual results, consistent with the tendency for managers to use interim forecasts more often to convey bad news (e.g., Skinner (1994)) and the difficulty of assessing longer-horizon forecast quality. More frequent forecasters are also more optimistic, which we interpret as managers greater willingness to be optimistic when credibility has been enhanced through repeated forecasting in the manner suggested by Stocken (2000) in his analysis of a multi-period disclosure setting. Finally, in supplemental tests, we find that linguistic sentiment is more positive in the years after the passage of Regulation Fair Disclosure. 4 After controlling for hard news and the stronger pricing effects of net negative sentiment, our capital market tests find little evidence that investors view sentiment as being biased when management incentives are in play, except for some evidence of the discounting of the typically more optimistic sentiment in annual forecasts. In fact, we detect stronger price reactions to sentiment when litigation risk 4 Kothari et al. (2009) attribute a post-reg FD decrease in the magnitude of bad news as evidence of less strategic delay of bad news in the post-reg FD period.

7 7 is higher, consistent with the notion that litigation costs enhance credibility. 5 The sentiment of more frequent forecasters (which we found to be more optimistic) is not discounted by the market, consistent with the notion that frequent disclosure enhances the credibility of good news. Furthermore, we detect a positive interaction effect between sentiment and hard forecast news, consistent with language being used to credibly reinforce, rather than to overturn, the accompanying hard news. In summary, aside from the finding that net pessimistic language is strategically delayed, most of our evidence is not consistent with managers using non-credible language to inflate prices. Supplementary tests investigating the association between linguistic sentiment and the ex post hard forecast error further corroborate this; language does not offer incremental explanatory power for managers ex post hard forecast bias. Taken as a whole, our evidence indicates that the language accompanying managerial forecasts is credible and directionally consistent with the hard news, and that it plays an important role in the price formation process. Our study extends the literature on the pricing of linguistic sentiment in several ways. First, we provide evidence of pricing effects for linguistic sentiment in a purely voluntary setting. The forwardlooking, unaudited, and ex post less verifiable nature of a management forecast implies potentially less credible language relative to the language in mandated filings. However, we find strong pricing effects for linguistic sentiment, more so when the linguistic sentiment is provided in a pure forecast setting and not bundled with an earnings release. Second, we document that either investors consider negative tone to be more credible than positive tone, or that managers strategically delay bad news. The effect is incremental to the corresponding effect for bad hard forecast news and is consistent with management incentives to increase share prices. Third, we document the effects of other management incentives on linguistic sentiment. Managers suppress optimism when the probability of litigation and political costs are higher and when facing financial distress, and they convey more optimism when they are frequent forecasters, when they issue annual forecasts, and in post-reg FD periods. With the exception of 5 In a concurrent working paper, Bonsall, Bozanic and Fischer (2012) also find a stronger price reaction to soft talk in earnings press releases when litigation risk is greater.

8 8 discounting of annual forecasts, we find no evidence that investors view these tendencies as bias. In fact, language is priced more strongly when expected litigation costs are higher. We also extend the literature on the pricing of management earnings forecasts. Prior research considers the effects of verifiable concurrent disclosures and external attributions on management forecast pricing (e.g., Hutton et al. 2003; Baginski et al. 2004). We extend this research to the realm of linguistic tone and document the incremental pricing and interaction effects of linguistic sentiment. In Section II, we develop our hypotheses about the association of language and prices. In Section III, we describe our sample and language data. We present our empirical design and results in Section IV, additional tests in Section V, and conclusions in Section VI. II. Theory and Hypotheses The Incremental Price-Relevance of Linguistic Sentiment The price-relevance of language is plausible if language is a sufficiently reliable signal of future earnings or dividends. However, language is relatively costless to provide, difficult to verify, and likely linked to future earnings and dividends in a fairly noisy way. The lack of verifiability, in particular, calls the informativeness of cheap talk into question (Crawford and Sobel (1982); Benabou and Laroque (1992); Dye and Sridhar (2004)). The language we investigate is issued in a management forecast press release. Although more ex post verifiable than language, the verifiability of hard management forecast news is nevertheless still difficult for several reasons. First, management forecasts are also imprecise, more often expressed in ranges, in minimums or maximums, or rounded to the nearest nickel (Baginski and Hassell (1997); Bamber and Cheon (1998); Bamber, Hui and Yeung (2010)). Second, the potential for unforeseen events between the forecast date and the realization of earnings confounds the ability of market participants to assess whether managers had truthfully revealed their expectations. Third, management has discretion over accruals at the subsequent earnings release. Notwithstanding these considerations, research into the

9 9 information content of management forecasts has long held that institutional arrangements (e.g., legal liability, the existence of information intermediaries and regulators) and reputational consequences create incentives for credible management forecasting (King, Pownall and Waymire (1990)). Consistent with this, numerous studies document that both precise and imprecise management forecasts have information content for security prices. 6 Likewise, linguistic sentiment has been shown to be incrementally informative for security prices when released with an earnings announcement (Davis et al (2012); Demers and Vega (2011)) or in other mandatory filings (e.g., IPO prospectuses and restatement announcements). Earlier studies also document an association between linguistic sentiment and future earnings, and between sentiment and the uncertainty of future earnings, suggesting a link to valuation fundamentals as the reason for the information content of language. Further, if linguistic sentiment predicts earnings, subsequent earnings realizations provide a degree of verifiability to sentiment, which also potentially enhances its information content. 7 The usefulness of linguistic sentiment in a purely voluntary disclosure setting such as unbundled management earnings forecasts is, however, not implied by prior research findings in which the sentiment is released with mandated, audited, and thus likely more reliable hard earnings-related information. In a management forecast press release, the quantitative forecast is less precise, less verifiable, and thus ultimately less reliable than a realization conveyed in an earnings press release. Thus, two potentially noisy signals are both being voluntarily released in the management forecast, the quantitative management forecast and linguistic sentiment. The relative weight to be placed on language in this context is not clear. A standard Bayesian learning model with two independent, noisy signals would suggest that an increase in one signal s noise should lead to a greater reliance on the other signal. Thus, holding constant the reliability of linguistic 6 Hirst, Koonce and Venkataraman (2008) provide an excellent review of this literature. 7 Although we primarily rely on the notion that economic-driven conditions exist that are sufficient to induce some level of truth-telling, the information content of linguistic sentiment is also suggested by a general aversion to lying (Gneezy (2005); Hurkins and Kartik (2009)). Further, models in behavioral economics (e.g., Mullainathan, Schwartzstein and Shleifer (2008)) and experimental results (e.g., Bertrand, Karlan, Mullainathan, Shafir and Zinman (2010)) suggest that uninformative material can affect choice, while archival evidence also supports the conclusion that uninformative disclosures matter (Michels (2012)).

10 10 sentiment (which we already assume to be lower than the quantitative management forecast), one would expect the effects of linguistic sentiment to be stronger in the context of unbundled management forecasts relative to other settings involving less noisy quantitative signals such as bundled or unbundled earnings announcements. On the other hand, when moving from the mandatory settings in prior research to a voluntary setting where management incentives play a greater role and regulation plays less of a role, the amount of noise in the two signals could be correlated. For example, the noise in linguistic sentiment could also increase if managers have incentives to bias all of the information, both quantitative and linguistic, in the management forecast press release. Thus, the predicted price relevance of language in the purely voluntary disclosure setting is unclear, motivating an empirical analysis of the following hypotheses: H1: Price response to a management forecast press release is positively associated with the linguistic sentiment expressed in the press release (incremental to the quantitative forecast news). H2: The incremental information content of linguistic sentiment when issued with unbundled management forecasts differs from its incremental information content when it is issued in conjunction with earnings announcements. H1 predicts that managerial incentives to credibly convey private information will lead to investor reliance on management s linguistic sentiment. H2 s non-directional prediction is motivated by the alternative directional possibilities that either linguistic sentiment becomes more important to investors in the context of management forecast press releases that are presumed to be noisier than earnings releases, or that linguistic sentiment becomes less important in a management forecast venue containing potentially overall less credible, incentive-driven voluntary disclosures.

11 11 Management Incentives and the Credibility of Linguistic Sentiment Prior research has identified situations where management forecasts appear to be strategic, thus raising the possibility that linguistic sentiment is also conveyed in a strategic manner. For example, Hutton et al (2003) provide empirical evidence consistent with the idea that the overarching management incentive to increase stock prices leads to a skeptical view of good news management forecasts. Relatedly, Kothari et al (2009) document that managers strategically delay bad news. Waymire (1984) and Rogers and Van Buskirk (2011) document that the sign of management forecast news disproportionately conflicts with the sign of accompanying, more reliable earnings realizations, suggesting a potential strategic role for supplying multiple earnings-related signals. Rogers and Stocken (2005) argue that managers will have incentives to bias hard forecast news, and will do so when the bias is difficult to detect. The uniqueness of the management forecast press release setting and the potential effects of management incentives on strategic behavior and forecast credibility provide an opportunity to test whether management incentives affect the managers use of linguistic sentiment and the market s assessment of linguistic sentiment credibility. The Incentive to Use Management Forecasts to Adjust Expectations Ajinkya and Gift (1984) provide evidence that managers use forecasts to adjust market expectations, and King et al (1990) provide the economic underpinnings of the expectations adjustment hypothesis. If the intent of a management forecast press release is indeed to adjust market expectations, then linguistic sentiment could be used to reaffirm the quantitative forecast news, thereby increasing the likelihood and/or magnitude of adjustment to this news. Alternatively, the quantitative forecast news could be used to confirm linguistic sentiment, consistent with the finding of Hutton et al (2003) that good news is more credible when it is issued with more verifiable information. In either case, two conforming signals are likely to cause a more pronounced adjustment of expectations than two conflicting signals, and thus the notion of conforming news is consistent with the expectations adjustment hypothesis.

12 12 Accordingly, we predict that linguistic sentiment is directionally consistent with the accompanying quantitative forecast news: H3: Linguistic sentiment is positively associated with hard forecast news. The Incentive to Disclose Good News (and Delay Bad News) Hutton et al (2003) argue that, given few incentives to release bad news and several incentives to increase share prices, management s bad news forecasts will be viewed as inherently more credible than good news forecasts. If investors perceive that linguistic sentiment is similarly used to increase share prices, then net negative sentiment will be viewed as more credible than net positive sentiment, and thus will be associated with a stronger price reaction. Cheap talk theories suggest that language inflation occurs, and this phenomenon exists regardless of the costs of lying (Kartik (2009)). Although a rational receiver of information is not deceived and thus discounts the language, the sender nevertheless does not remove the inflation from language because the sender fully expects the receiver to apply discounting. i Vidal (2003) modifies the model in Morgan and Stocken (2003) to establish an equilibrium in which investors react more to bad news than to good news in the presence of the optimistic cheap talk tendency. This leads us to test the following hypothesis regarding managerial incentives to inflate language and increase share prices: H4: The mean price response to net negative sentiment in a management forecast press release is, in absolute magnitude, greater than the mean price response to net positive sentiment. Evidence in favor of H4 is also subject to an alternative explanation, that managers delay bad news. As discussed by Verrecchia (2001), managers have several incentives to delay bad news disclosure. For example, Hermelin and Weisbach (2007) formally model bad news delay as a function of career concerns. Kothari et al (2009) argue that news arriving randomly to managers would result in symmetrically distributed stock returns unless managers delay the disclosure of bad news to a threshold where the costs or difficulty of further delay require bad news disclosure. They argue that evidence of a more negative price reaction for bad news management forecasts relative to good news is evidence that

13 13 bad news has been delayed, not that it is differentially credible. Because neither Rogers and Stocken (2005) nor Kothari et al (2009) detect that bad news is more ex post accurate than good news, stronger price reactions to net negative sentiment could be evidence of management s incentives to strategically delay bad news. Under either interpretation, evidence in favor of H4 is consistent with managerial incentives to increase share price leading to strategic linguistic sentiment disclosure. Situational Incentives for Strategic Communication In addition to the aforementioned general incentives to disclose good news (or to delay bad news) and to use voluntary disclosure to move market expectations, there are some further contexts in which managers may wish to strategically communicate with the stakeholders of the firm. The predicted relations between each such situational variable and linguistic sentiment are discussed in turn. Litigation. Although managers have incentives to increase stock prices with optimistic disclosures, fear of litigation based on voluntary disclosures that are ex post over-optimistic tempers the tendency towards over-optimism in voluntary disclosures (Skinner (1994); Kasznik and Lev (1995); Baginski, Hassell and Kimbrough (2002); Rogers and Stocken (2005)). Consistent with this, Rogers and Van Buskirk (2011) document that managers use of optimistic language increases litigation risk by showing that plaintiffs target optimistic statements in their lawsuits and that, controlling for a firm s economic conditions, sued firms have unusually linguistically optimistic earnings announcements. Accordingly, we expect that if optimistic linguistic bias can be detected, then managers will avoid increased legal costs by tempering their optimism. Product market competition. Optimism is also costly because it signals high industry profitability, which in turn may encourage entry into the industry by potential competitors. Therefore, firms in more concentrated (i.e., less competitive) industries are apt to downward bias their voluntary disclosures (Newman and Sansing (1993)), leading us to expect that managers of these firms will temper their linguistic optimism.

14 14 Political costs. Li (2010) argues that political costs temper the use of optimistic language, and he documents this effect in the Management Discussion and Analysis section of the annual report. We expect a similarly tempered use of linguistic sentiment in the context of management forecasts. Financial distress. Rogers and Stocken (2005) argue that managers whose firms are in financial distress have an incentive to bias their forecasts upwards as a means of convincing investors of their ability to restore their firms to financial health. Accordingly, we expect linguistic optimism to be positively associated with financial distress. Disclosure horizon. Managers choose their forecast horizon. A common finding in prior research is the tendency for annual (interim) forecasts to be used to convey good (bad) news (e.g., Skinner 1994). Predominately good news annual forecasts do not necessarily represent a bias, but they are consistent with Kothari et al. s (2009) hypothesis of earlier disclosure of good news. Managers ability to hide bias in an annual forecast is also enhanced since annual forecasts have expected lower accuracy, and any errors can be explained more readily by appeal to the length of time during which unexpected intervening events have occurred to render the manager s forecast less accurate. In the spirit of Rogers and Stocken (2005), we therefore expect that, with the relatively less verifiable language in annual forecasts, these releases will contain more optimistic language. Existence of a repeated game. Repeated games enhance management s disclosure credibility. Stocken (2000) examines the credibility of a manager s disclosure of non-verifiable information in a repeated cheap talk game and concludes that the manager almost always truthfully reveals private information under certain sufficiency conditions, including an accounting report that can be used to assess credibility and a longer-run assessment of management credibility. Given their desire to increase share prices, managers who have established credibility via frequent voluntary disclosures are expected to be more willing to use optimistic language, trusting that the market will treat it as credible. Taken together, the preceding predictions lead us to generally hypothesize that situational incentives affect linguistic sentiment: H5: Linguistic sentiment is associated with situational incentives.

15 15 The tension underlying this hypothesis is considerable because the effectiveness of strategic disclosure depends upon the difficultly of detecting the strategic communication (Rogers and Stocken 2005), whether managers believe that markets will rationally discount their biases, and whether managers have sufficient incentives to build a positive voluntary disclosure reputation. If linguistic sentiment satisfies the difficult to verify ex post condition, then we expect that a rational securities market would ignore cheap-talk linguistic sentiment. However, if linguistic sentiment is sufficiently verifiable ex post because investors observe the object of linguistic sentiment s prediction, realized earnings, then the costs of biasing disclosures are significant, managers find benefit in building a credible disclosure reputation, managers will engage in credible linguistic expression, and markets will not discount language. Evidence of Ex Post Bias and the Market Pricing of Sentiment as a Function of Incentives An efficient market response to potentially incentive-driven language will depend upon whether the incentives lead management to provide misleading (i.e., ex post biased) sentiment. If language is ex post optimistically biased, then the market should discount the weight applied to language in settings where managers have incentives to increase share prices with such optimism. We investigate these issues by providing tests of the following null hypotheses: H6a: Linguistic sentiment is unrelated to ex post management forecast error. H6b: Price reaction to linguistic sentiment is not conditioned on management incentives to either confirm the hard forecast news or to strategically alter linguistic sentiment. Main Sample Determination III. Sample and Language Data We use the Factiva database to individually identify and download candidate management earnings forecasts. We follow Baginski et al. (2004) by using business newswires Dow Jones Business News ( DJBN ) and Press Release Newswire ( PRN ) to search for the following word strings expects

16 16 earnings, expects net, expects income, expects losses, expects profits, and expects results in addition to three parallel lists where expects is replaced alternatively by forecasts, predicts, and sees ). This search yielded 6,180 candidate earnings forecasts (3,577 for DJBN and 2,603 for PRN) for the period 1997 through 2006, downloaded in batches of 100 announcements per.txt file. We next created individual.txt files for each candidate management forecast article (i.e., converting 62 files into 6,180.txt files), and extracted firm identifiers for the companies underlying each respective Factiva article in order to attempt to match the candidate observations into the CRSP, Compustat, and First Call databases. In order to be included in our initial sample, each candidate management forecast from the Factiva search and extraction process had to match up, within a three-day window surrounding the Factiva date, with a management forecast from the First Call Company Issued Guidance database. Finally, we delete observations having fewer than 100 words in the forecast announcement and observations for which the data is not available to measure all of our regression variables. This process yielded a total of 3,828 forecast observations. Additional details related to the sample determination and matching procedures are summarized in Appendix A. 8 Following the prior literature (e.g., Anilowski, Feng and Skinner (2007); Rogers, Skinner and Van Buskirk (2009)), we define bundled management forecasts as those falling within two days of an earnings announcement date. As in Hutton et al (2003) and Rogers et al (2009), we discard bundled forecasts in order to create a sample of non-bundled management forecasts. Bundled forecasts have properties that distinguish them from unbundled forecasts on the dimensions of forecast tone and presence of conflicting hard earnings news (Rogers and Van Buskirk 2009) as well as incremental information content (Atiase, Li, Supattarakul and Tse (2005)). Discarding bundled forecasts is further justified on the basis that it is 8 The requirement that our press release observations match with First Call earnings forecasts results in the loss of some press releases. Chuk, Matsumoto and Miller (2009) document that excluded observations are likely to be made by firms that: a) have less analyst forecast coverage; b) have poor prior performance; c) do not issue an accompanying earnings release; and d) do not provide an EPS forecast (e.g., provide a revenue or cash flow forecast instead). Because we require analyst forecasts for our observations, sample to avoid bundling with earnings announcements, have other data requirements that skew our sample toward larger firms, focus on earnings forecasts, and sample primarily after 1997 (which is suggested by Chuk et al (2009) as a means of mitigating bias), we do not believe that our sample would be meaningfully different had First Call matches not been required. However, the data required by our research design does result in a larger firm sample bias, consistent with that characterizing many large sample capital markets studies.

17 17 not possible to cleanly distinguish which aspects of the language in the bundled press releases pertain to the historical earnings announcement versus the forward-looking management forecast, while the properties of the historical earnings announcements have been examined at length in prior studies. The resulting sample of 2,085 unbundled forecasts consists of 1,372 distinct press releases in which the management forecast is the main event as in Hutton et al. (2003). Alternative Sample Definitions Management forecast press releases commonly include both an annual and an interim forecast. We address the potential confounding effect of these simultaneously released signals by reporting our results using four alternative sample definitions. The first of the four samples is simply the full sample of 2,085 management forecasts. The second sample treats the press release as the unit of observation, with the interim forecast taken as the primary signal when both interim and annual forecasts are provided. 9 This second sample contains 1,372 observations. The third and fourth samples consist of, respectively, the 749 annual and the 1,336 interim forecasts, a dichotomy that is useful for at least three reasons. First, because annual and interim forecasts are often packaged, splitting the sample and analyzing them separately leads to greater independence of the linguistic sentiment measures. Second, the annual and interim forecasts may be somewhat redundant to one another, 10 an observation that is supported by the extremely high rank correlations and sign agreements between the simultaneously released annual and interim forecasts in our sample (reported later). Third, several past studies have documented that annual and interim forecasts have different properties, generating, e.g., different price reactions (e.g., Pownall and Waymire 1989). Of greatest importance to our study, annual forecasts have longer horizons, a characteristic that generally yields higher forecast errors and increases the likelihood that intervening 9 We treat the interim forecast as the primary signal because prior research has established that interim forecasts have greater price relevance than annual forecasts. Consequently, we treat the annual forecast as merely an update of expected annual results given the forecast of the interim period. As will be discussed throughout our empirical tests, our results are not affected by using different approaches to handle the multiple simultaneously released forecasts. 10 One can view the annual forecast sample as potentially containing an additional interim forecast in the press release that is subsumed by the annual forecast. Likewise, one can view the interim forecast sample as potentially containing an annual forecast that is merely the implication of the interim forecast for annual earnings.

18 18 events might explain the forecast errors (i.e., rather than management s lack of forecasting ability). Thus, management bias, if any, is more likely to be found in an annual forecast rather than in an interim forecast that is followed shortly by an earnings release. Clustering of observations and the existence of multiple forecasts in a single release cause potential statistical inference problems. For each of our empirical tests, we describe the procedures that we undertake to mitigate these problems. Data Sources We identify and extract text passages of the management forecasts from DJBN and PRN within the Factiva database as described above. The non-linguistic characteristics of the management forecasts (e.g., the numerical estimate, annual versus quarterly, etc.) are derived from the First Call Management Issued Guidance database. Market prices and returns data are provided by the Center for Research in Security Prices (CRSP), while the Compustat database is our source for accounting data. Measuring the Language Constructs Evidence from prior studies (Loughran and McDonald 2011; Demers and Vega 2011) suggests that generic linguistic algorithms such as Diction or General Inquirer may yield noisy measures of positive and negative linguistic tone in the context of financially-oriented text passages. Consequently, for our primary tests we adopt the Loughran and McDonald (2011) (L&M) financeoriented dictionaries (i.e., word lists), for capturing positivity and negativity in the management forecast textual passages, measures that L&M refer to as Fin-Pos and Fin-Neg, respectively. 11 Rogers et 11 We also follow the cleansing procedure in Demers and Vega (2011). Corporate press releases on the newswire services often include several paragraphs at the end of the announcement that are not part of the body of the announcement that is of interest to our study. Specifically, the releases typically include a company-standard paragraph that describes the firm, often using very flattering language. In addition, most of the articles tend to include some form of Safe Harbor disclaimer paragraph related to the forward-looking information included in the press release. These latter paragraphs vary somewhat across firms, but are generally boilerplate in nature and are presumed to be drafted by the company s legal advisers. Finally, the press releases typically end with company contact information such as a listing of the corporate website address, their investor relations contact names and

19 19 al. (2011) argue, however, that more general language measures such as Diction might be more relevant in the context of communications with non-financial audiences such as attorneys, judges, and others. Accordingly, and also in order to facilitate the comparison of our results with the growing body of accounting research that has principally relied upon the Diction algorithms, we also replicate all of our analyses using the Diction set of linguistic measures. 12 Also following the prior linguistic literature (e.g., Davis et al (2012); Demers and Vega (2011)), our principal measure of tone is defined as the difference between Fin-Pos and Fin-Neg (optimism and pessimism) for the L&M- (Diction-) based measures, a variable that we label as NetPositivity. Finally, following Rogers, Van Buskirk and Zechman (2011), we combine the two sentiment measures. Because we are dealing with only two measures, we rank each in the cross-section, and use the average rank for our tests. 13 Similar to most prior studies, our NetPositivity variable does not attempt to explicitly measure the unexpected portion of sentiment. Although some prior studies have examined the time-series properties of sentiment in earnings releases, leading them to adopt the change in net optimism as a proxy for the unexpected or news component of linguistic sentiment, we have opted not to do so for several reasons. First, as Loughran and McDonald (2011) point out, measuring unexpected sentiment in this way imposes a considerable amount of structure on the linguistic parameters; specifically, it presumes a considerable amount of processing capability on the part of investors in the cross-section. While this is conceivable in the context of regularly recurring, mandated, quarterly earnings announcements, it is much numbers, and/or information related to upcoming conference calls. Since all of these paragraphs contain textual and numerical data that may include optimism-, pessimism-, and/or certainty-related language that does not form part of the content portion of management s press release per se, based upon manual review and the identification of keyword strings, we developed algorithms to cleanse the.txt files of such non-announcement content. These cleansed.txt files are ultimately used as the basis for the linguistic characteristics extracted from the firm s earnings forecast press releases. 12 Notwithstanding the potential noisiness of the measures derived from Diction in the context of financial text, it is a well-established language processing algorithm that has been used extensively in prior research to measure the sentiment in earnings announcements, corporate annual reports, Federal Reserve Board Chairmen s speeches, and other economic and political communications. See Davis et al (2012), Demers and Vega (2011), Davis and Tama- Sweet (2011), or the listing provided at for a more extensive summary of published academic studies using the Diction software. 13 At least three variables are required in order to perform a proper factor analysis. Harman (1976) notes that twovariable factor analysis yields factor weightings that are not unique, and the two-variable description of a common factor is quite arbitrary.

20 20 less likely to hold in the context of sporadically issued, non-standardized management forecasts. Second, the primary study to take this more refined approach, Demers and Vega (2011), reports that their basic findings are unaffected by the use of net optimism rather than the change in net optimism. Finally, a sufficient time series of management forecasts is not available for the majority of our sample firms, making it impossible to model the time-series behavior of sentiment in the context of management forecasts. The consequences of not specifying an expectations model for our sentiment variable would generally be to introduce noise into the measure, thereby reducing the power of our cross-sectional tests. As we discuss below, however, the results in relation to the NetPositivity variable are generally quite strong and significant in conservative testing that employs clustered standard errors. IV. Empirical Results In this section, we present descriptive statistics followed by the empirical research design and test results for each of our hypotheses in turn. Descriptive Statistics Table 1 presents descriptive statistics for the forecasts, linguistic sentiment, and other regression variables used in our tests. 14 In Panel A, consistent with prior research (e.g., Rogers and Van Buskirk (2011)), the unbundled forecasts in our sample are, on average, bad news, as evidenced by the negative means and medians for abnormal announcement returns (AR3), management forecast surprises (FSURP), and the post-announcement drift (AR60). The mean and median of NetPositivity are similarly slightly negative for the L&M proxy, although the Diction proxy mean is slightly positive with a median of zero. The management forecasts are optimistic ex post when compared to realized earnings, as captured by the variable ExPostForecastOpimism that is defined as the price-deflated difference between forecasted and actual EPS. Panel B shows that the majority of forecasts (64.1%) are annual. Panel C shows that the forecasts are highly dispersed across years, with an initial increase in forecast incidence over time 14 The distributions presented in Table 1 are after winsorizing every variable at both the 1 st and 99 th percentiles.