Do Analysts Say Anything About Earnings Without Revising Their Earnings Forecasts?

Transcription

1 Do Analysts Say Anything About Earnings Without Revising Their Earnings Forecasts? Philip G. Berger Booth School of Business University of Chicago Charles G. Ham John M. Olin School of Business Washington University in St. Louis Zachary R. Kaplan John M. Olin School of Business Washington University in St. Louis October 2016 * We appreciate helpful comments from, and discussions with, Rich Frankel, Jared Jennings, Kevin Koharki, Xiumin Martin, Volkan Muslu (discussant), Doug Skinner, Sorabh Tomar and workshop participants at Carnegie Mellon, Washington University in St. Louis, the 2016 AAA Annual Meeting, and the 2016 EAA Annual Congress, as well as research assistance from Aadhaar Verma and Yuqing Zhou. Any remaining errors or omissions are ours. Berger acknowledges financial support from the University of Chicago s Booth School of Business.

2 Do Analysts Say Anything About Earnings Without Revising Their Earnings Forecasts? Abstract We identify a novel bias in analyst forecasts, after-revision bias, which we identify by examining an analyst s reports after his final earnings forecast of the quarter. We document that (i) qualitative predictions from the text of reports, (ii) share price target revisions, and (iii) revisions to next quarter s earnings forecast predict error in the current quarter s earnings forecast (CQE). Consistent with analysts maintaining a beatable benchmark for managers, we find analysts are more likely to disseminate positive (negative) news by revising a forecast other than the CQE (revising the CQE). Consistent with analysts minimizing deviations from the consensus, we demonstrate analysts revise the CQE (revise another forecast) more frequently when a revision to the CQE would move their forecast towards (away from) the consensus. Market returns are slow to impound the information in qualitative predictions and share price target revisions, as both predict earnings announcement window returns. Our results demonstrate that the value of the current quarter s earnings forecast to managers and investors distorts the flow of information into the forecast. Keywords: sell-side analysts, analyst incentives, earnings forecasts, forecast bias.

3 1. Introduction We study the means, determinants, and consequences of analysts disseminating earnings information without revising their current quarter earnings forecasts. Analysts may obtain information with earnings implications, but omit it from the current quarter forecast, if doing so enables them to satisfy key stakeholders. For instance, if revisions impose processing costs on clients and/or expose the analyst to reputation loss in the event of an error, these frictions may limit the frequency with which analysts revise their forecasts (Bernhardt et al. 2016). Similarly, incentives to maintain relationships with, and access to, company management (Lin and McNichols 1998; Lim 2001; Ke and Yu 2006) can affect the decision to issue a forecast, particularly one the firm may not beat. Analysts may compensate for the costs of communicating earnings information through the current quarter earnings forecast by instead disseminating information through other channels, at least partially explaining why analysts issue a preponderance of forecasts early in the quarter (Ivkovic and Jegadeesh 2004). We use two approaches to capture the information analysts disseminate after the current quarter s final earnings forecast. First, we use the sign of revisions to the share price target (SPT) and future quarter earnings (FQE) forecasts to capture news disseminated after the final forecast of the current quarter s earnings. Second, we use a textual approach to capture qualitative statements about earnings in analyst reports issued after the current quarter s final earnings forecast. Specifically, we define qualitative predictions as positive (negative) if the analyst uses a synonym for beat ( miss ) and a synonym for earnings expectations within the same sentence. 1 Collectively, we refer to the qualitative predictions, SPT revisions, and FQE revisions as non-earnings forecast signals. 2!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1 We examine the text of a subset of analyst reports published after the final forecast of the quarter and find numerous examples of analysts publishing explicit predictions that the company will beat or miss the analyst s own 1

4 We document that all non-earnings forecast signals strongly predict the analyst s own earnings surprise. Specifically, a qualitative prediction of miss (beat) increases the probability of the firm missing (meeting or beating) the earnings estimate by 5.5% (6.1%). When the analyst revises a share price target negatively (positively), the probability of the firm missing (meeting or beating) the earnings estimate increases 1.7% (4.3%). The FQE revisions have a lower, though still statistically significant, association with the probability of the firm missing or beating the analyst s prior forecast. 3 The preceding results imply that analysts do not always issue a revised earnings forecast when they have information with earnings implications ( after-revision bias ). The evidence from our textual analysis dismisses concerns that the issuance of non-earnings forecast signals is unintentional, because the analyst would have to not understand the meaning of his own words. Prior studies, which identify bias at the time of the forecast, have had difficulty distinguishing between intentional and unintentional bias (Francis 1997), because both intentional (incentivebased) and unintentional (behavioral-based) theories can predict under-reaction to contemporaneously released information.!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! forecast. Predicting performance relative to a previous forecast rather than revising the forecast implies the analyst does not always issue a revised earnings forecast when he has significant earnings information. For example, one week before the earnings announcement, a Wedbush Securities analyst published a report with the following phrase: We expect Garmin to report significant upside to our expectations for revenue, margins, and EPS given the tremendous growth in the PND market and strong results from its closest rival, TomTom. We validate our measure by hand-coding a subset of the sentences identified by our algorithm as containing qualitative predictions to ascertain whether the analyst report includes an explicit qualitative prediction that the firm will beat or miss earnings expectations. We document a significantly positive correlation between our algorithmic coding scheme and the hand-coded sentences, suggesting the text algorithm captures qualitative predictions, albeit with noise. 2 An advantage of the SPT and FQE setting is that revisions to these forecasts perfectly capture the intentions of the analyst, whereas the textual analysis will capture those intentions with error. An advantage of the textual setting is that the textual comments relate explicitly to the current quarter s earnings forecast, while the SPT and FQE revisions could relate to other information. 3 Our main specification compares the firm s actual earnings to the analyst s own forecast. We also document that non-earnings forecast signals are associated with the earnings surprise if we compare the firm s actual earnings to the consensus forecast. 2

5 Next, we investigate cross-sectional variation in after-revision bias to shed light on why analysts issue non-earnings forecast signals. 4 First, we examine whether analysts revise SPT and FQE forecasts for good news more or less frequently than for bad news. Incorporating good (bad) news into an SPT or FQE forecast, rather than the current quarter earnings (CQE) forecast, increases the likelihood that actual earnings meet or beat (miss) the earnings forecast. Because managers prefer to meet or beat forecasts (Richardson et al. 2004), and managers reward analysts who publish favorable forecasts by granting greater access to management (Lim 2001; Ke and Yu 2006; Brown et al. 2015), analysts have incentives to omit positive news from (incorporate negative news into) CQE forecasts. We find analysts revise SPT and FQE (CQE) forecasts more frequently for good (bad) news than bad (good) news. We also find positive SPT and FQE revisions provide more information content about the earnings surprise relative to negative SPT and FQE revisions. Our findings contribute to understanding the mechanism analysts use to walk-down forecasts they communicate positive (negative) news to their clients by omitting positive news from (incorporating negative news into) the earnings forecast. We also document that analysts with a buy recommendation walk-down their forecasts more aggressively, suggesting a positive disposition toward a company leads the analyst to incorporate more negative news into the CQE forecast. Consistent with this, we demonstrate that analysts with a buy recommendation (i) are more likely to issue negative CQE forecast revisions, and (ii) are more likely to issue positive SPT forecast revisions. Thus, analysts with a buy recommendation distort the flow of information into the CQE forecasts in a way that increases the probability of the firm beating these forecasts.!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4 We exclude the qualitative predictions from these analyses because the textual analysis is conducted on a small subset of the analyst reports, which limits the amount of cross-sectional variation we can exploit. 3

6 Second, we examine whether analysts respond to earnings information with a non-earnings forecast signal more frequently when the information moves the analyst forecast toward the consensus (herding) or away from the consensus (bold). Investors judge analyst forecasts relative to those of competing analysts. Prior research documents that analysts who issue herding forecasts are less likely to be fired, suggesting career consequences from issuing extreme forecasts (Hong et al. 2000; Clement and Tse 2005). We find that (i) analysts are more (less) likely to issue a non-earnings forecast signal for bold (herding) forecasts, and (ii) bold (herding) non-earnings forecast signals provide more information content about the earnings surprise. Our result that analysts suppress bold forecasts suggests that measures of forecast dispersion may understate the amount of disagreement, because analysts suppress their private information when it moves them away from the consensus. Third, we examine how the magnitude of earnings information after the initial CQE forecast affects the decision to issue a non-earnings forecast signal. When there is little information not already in the analyst s forecast, frictions may limit the analyst s incentive to revise the CQE forecast (Bernhardt et al. 2016). Conversely, incentives for accuracy may encourage the analyst to revise the CQE forecast when it omits substantial information (Clement and Tse 2005; Groysberg et al. 2011). Consistent with this prediction, we find that the absolute value of initial forecast error is lower when the analyst revises the SPT or FQE forecast relative to when the analyst revises the CQE forecast. Finally, we study the consequences of non-earnings forecast signals by examining the association between the issuance of non-earnings forecast signals and earnings announcement returns. If the market does not fully impound the earnings implications of non-earnings forecast signals at the time of their issuance, these signals may be associated with returns at the earnings 4

7 announcement date. Consistent with this, we document that when the analyst issues a qualitative prediction of miss (beat), average earnings announcement returns are 0.7% lower (0.5% higher) than when the analyst does not. The large predictable returns at earnings announcements suggest that the information in qualitative predictions is not fully impounded into returns prior to the earnings announcement. We document less (no) return predictability for SPT (FQE) forecast revisions. Our results provide several contributions. First, we demonstrate both the importance of CQE forecasts and the related danger of interpreting them as a proxy for market expectations. Finding that analysts hesitate to incorporate changes in their own CQE expectations into their CQE forecasts is consistent with analysts viewing CQE forecasts as prominent benchmarks, and with modifications to the benchmark imposing costs on managers or the analyst s clients. Our finding that analysts do not update their forecasts when they have updated their beliefs has critical implications for academic researchers using analysts CQE forecasts to proxy for investor expectations when (i) computing earnings response coefficients, (ii) measuring investor disagreement, and/or (iii) evaluating firm performance. Our results suggest a methodology to correct analyst forecasts for this omitted information. Finally, researchers hoping to infer the role of analysts by examining the timing of their forecast revisions may ignore the timing of qualitative analysis, which has substantial value to investors (Frankel et al. 2006; Chen et al. 2010). Second, we contribute to the literature on forecast bias by identifying a novel source of error, for which we have increased power to distinguish between intentional and unintentional bias. The ability to distinguish between intentional and unintentional bias is challenging at the time of the forecast because both intentional (incentive-based) and unintentional (behavioral- 5

8 based) theories can predict under-reaction to contemporaneously released information. Biases identified by the extant literature that condition on the existence of a forecast revision thus provide no natural counter-factual to identify causation. 5 By exploiting cross-sectional variation in after-revision bias, we show that the issuance of non-earnings forecast signals responds to analyst incentives, inconsistent with an unintentional explanation for the bias. 6 Specifically, in our textual analysis, when analysts make bullish (bearish) statements about future earnings, the firm tends to beat (miss) the unrevised forecast. If unintentional, the analyst has to not understand the meaning of his own words. 2. Literature Review and Hypothesis Development First, we review the literature on the consequences of meeting or beating analyst estimates as well as evidence on biases in these estimates. We also provide evidence on the properties of longer-horizon earnings forecasts and share price target forecasts, as we will ultimately argue these forecasts incorporate some of the information analysts omit from their short-horizon forecasts. Second, we review the prior literature on the incentives of analysts, with a particular focus on their incentives to issue accurate forecasts and beatable forecasts. We conclude by motivating our research question, which asks whether analysts outputs after the final current quarter earnings forecast predict error in the forecast.!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 5 Prior studies document predictable errors in analyst forecasts, including: (i) forecasts under-react to past earnings information (Abarbanell and Bernard 1992), (ii) forecasts under-react to returns (Lys and Sohn 1990), (iii) forecast revisions have a positive correlation with subsequent forecast errors (Shane and Brous 2001), (iv) forecasts underincorporate information from other analyst forecasts (Stickel 1992), and (v) forecasts under-incorporate information from firm characteristics (Hughes et al. 2008). We elaborate further on the state of the literature as to the cause of the biases in section 2. 6 In our setting, after the initial forecast of the quarter, the analyst has three options: (i) issue a non-earnings forecast signal, (ii) revise the current quarter earnings forecast, or (iii) issue no forecasts. We exploit scenarios (ii) and (iii) as counter-factuals to identify the determinants and consequences of after-revision bias. 6

9 2.1 Meeting or Beating Analyst Forecasts and Predictable Errors in Analyst Forecasts A lengthy stream of literature evaluates whether firms manage earnings to exceed analysts forecasts of earnings, and the consequences of meeting or beating these expectations. DeGeorge et al. (1999) provide evidence that a higher fraction of firms just meet or beat analyst forecasts than just miss the threshold, suggesting managers take actions to reach this benchmark. Evidence documenting analysts walk-down forecasts during the quarter suggests expectations management may drive a portion of the propensity for firms to meet or beat analyst forecasts (Kang et al. 1994; Matsumoto 2002; Cotter et al. 2006). Other studies document firms may use real earnings management to meet or beat estimates (Roychowdhury 2006; Bhojraj et al. 2009). Corroborating this empirical evidence, Graham et al. (2005) survey managers and find (i) a majority of managers would sacrifice value increasing projects to avoid missing expectations, and (ii) analyst expectations are the most important financial reporting threshold to exceed. Studies document that firms face negative consequences upon failing to meet or beat earnings forecasts. Firms that miss estimates have negative share price performance (Skinner and Sloan 2002; Bartov et al. 2002; Bhojraj et al. 2009), suggesting capital market consequences upon missing expectations. In addition, managers suffer career penalties for failing to meet or beat earnings forecasts (Matsunaga and Park 2001; Farrell and Whidbee 2003). While most studies investigate the importance of meeting the consensus forecast, recent research documents that the percentage of forecasts met has better predictive power for earnings announcement returns, suggesting that investors evaluate firm performance relative to multiple benchmarks (Kirk et al. 2014). Despite the importance of exceeding earnings expectations to managers, numerous studies document information available to analysts at the time they issue earnings forecasts 7

10 predicts forecast error. Past earnings surprises (Abarbanell and Bernard 1992) and past returns (Abarbanell 1991) have a positive correlation with forecast error. Firm characteristics such as book-to-market ratios, dividend payout ratios (Hughes et al. 2008; So 2013), accruals (Bradshaw et al. 2001), and book-tax differences (Weber 2009) also predict forecast errors. Individual forecasts under-weight information from the consensus (Chen and Jiang 2006) and the sign of the analyst s revision to the current quarter s earnings forecast has a positive correlation with the error in the revised forecast (Shane and Brous 2001; Raedy et al. 2006). Although short-horizon forecasts contain significant biases, they still provide more accurate forecasts of future earnings than time-series models (Brown et al. 1987; Bradshaw et al. 2012). In contrast, long-horizon forecasts provide less accurate forecasts than time-series models (Bradshaw et al. 2012) and contain significant bias (DeBondt and Thaler 1990). Prior studies on share price targets have documented: (i) share price target revisions have large announcement window returns (Brav and Lehavy 2003), (ii) forecasts of share price appreciation contain some information about future returns, but also contain significant bias (Bradshaw et al. 2013), and (iii) analysts do not exhibit persistent ability to forecast share price appreciation (Bradshaw et al. 2013). 2.2 Analyst Incentives and Explanations for Forecast Bias Explanations for analyst forecast bias fall under two broad categories: (i) intentional bias: analysts have incentives to issue forecasts that satisfy key stakeholders such as managers or clients, and (ii) unintentional bias: analysts make errors transforming information into forecasts of earnings. Under intentional explanations for forecast bias, analysts face a trade-off between their incentives to issue accurate forecasts and the benefit to biasing forecasts to please key 8

11 stakeholders (Lim 2001). Research demonstrates forecast accuracy is not valued as highly as some other analyst forecast attributes. Groysberg et al. (2011) find no evidence analyst compensation relates to forecast accuracy after conditioning on institutional investor votes. Bagnoli et al. (2008) investigate the attributes investors value most highly using survey evidence and find that investors rank written reports and industry insight highest, above the importance of earnings forecasts. Groysberg et al. (2011) also find analyst compensation increases with the analyst s rating on investor surveys, suggesting analysts have incentives to provide information institutional investors demand. Survey evidence confirms analysts believe their primary objective is to provide information their clients perceive to be valuable (Brown et al. 2015). Research demonstrates a number of channels through which analysts can increase the value of their information by potentially biasing forecasts. First, analysts value access to management (Soltes 2014; Brown et al. 2015), suggesting analysts might be willing to bias forecasts to procure access (Lim 2001; Ke and Yu 2006). Confirming the credibility of managerial threats to withhold access, Soltes (2014) documents that managers refuse to interact with analysts who issue negative recommendations. Brown et al. (2015) also provide survey evidence suggesting analysts bias their earnings forecasts to obtain access to management. Second, if forecast accuracy is evaluated relative to the accuracy of other analysts, this creates incentives for analysts to omit information from forecasts in certain circumstances. Hong et al. (2000) demonstrate that inexperienced analysts are more likely to be fired for inaccurate forecasts and bold forecasts, whereas being bold and accurate does not improve the analysts career outcomes. They argue this creates incentives for analysts to forecast strategically. Clement and Tse (2005) further demonstrate that herding forecasts, or forecasts that move the 9

12 forecast toward the consensus, contain less new information than bold forecasts, which they interpret as further evidence of strategic forecasting. Third, the analyst could bear some cost when revising the forecast (i.e., frictions). Empirical evidence that analysts revise their forecast infrequently, a median of once per quarter, is consistent with frictions impeding the flow of information into forecasts. Frictions can impact forecast frequency through several channels, including: (i) the analyst incurs costs from adjusting the model and/or explaining the changes in the model to clients. 7 Evidence that analysts make mistakes preparing their models suggests the risk of error is non-negligible (Green et al. 2016). (ii) The analyst endogenizes the processing cost his revisions impose on clients and only issues revisions from which a client would obtain a benefit that exceeds the client s processing cost. 8 For the frictions explanation to be plausible in our setting, in at least some instances frictions must prevent a revision to the CQE forecast while not preventing either the publication of a report or the revision of another forecast. Fourth, when analysts publish forecasts, they are broadly disseminated through I/B/E/S, Bloomberg, and other financial networks. Thus, updating information in earnings forecasts potentially allows non-clients to capture some of the informational rents created by an analyst s research. Inserting information into the text of the report could limit the number of non-clients!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 7 Costs also arise because analysts receive questions about the materials they disseminate. By issuing a report the analyst alerts clients to his report, and these clients might follow up with questions or demands for additional analysis. 8 When an analyst revises a forecast the reports and/or notes are circulated to clients. If the information content of the reports is low, the client may feel the cost incurred from processing the information was not worth the benefit from obtaining the information in the report. The client may then penalize the analyst on his survey evaluations for wasting his time. Investors can condition their decision to process an analyst report on (i) the brokerage of the analyst, (ii) the company covered, and/or (iii) the timing of the report. Investors will often find it difficult to decide whether to process information based on the content of the information, because obtaining the content involves incurring costs processing the information (Sims 2003). 10

13 consuming the information. 9 This should increase the benefit clients obtain from an analyst s research by limiting the number of informed investors (Grossman and Stiglitz 1980). 2.3 Research Question Ultimately, analysts must provide valuable information to their clients. Empirical evidence documents that the final revision to the current quarter earnings forecast often occurs near the prior quarter earnings announcement date (Ivokovic and Jegadeesh 2004). If the analyst disseminates earnings information only through revisions to the current quarter earnings forecast, this implies that in many firm-quarters the analyst goes three months without providing earnings information to clients. Not disseminating information over such a long time frame likely decreases the value of his research to clients (Groysberg et al. 2011). Alternatively, the analyst may disseminate earnings information via other channels. In other words, the analyst may respond to the incentives to omit information from the CQE forecast, yet incorporate the information into other forecasts. Thus, we study whether the aforementioned non-earnings forecast signals have information content about earnings. RQ: Do analysts non-earnings forecast signals after the final current quarter earnings forecast provide information content about the earnings surprise? If so, why do analysts issue nonearnings forecast signals after the final current quarter earnings forecast?!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 9 Prior research investigates tipping by studying whether institutional investors trade on analyst recommendations before they have been publicly disseminated (Irvine et al. 2007). Although providing information to clients before publishing it offers the clients a significant information advantage, it exposes the analyst to potential discipline from his brokerage, many of which proscribe the behavior. This form of selective disclosure is also against FinRa regulations. 11

14 3. Sample Selection and Descriptive Statistics 3.1 Revision Sample Selection The sample begins with all analyst-firm-quarters on the I/B/E/S unadjusted detail file over the period 1999Q1 to 2015Q2. The start date is the first year share price targets are available on I/B/E/S and the end date is the last quarter with data availability. We keep analystfirm-quarters for which (i) the analyst issued a quarterly earnings forecast in the current and prior quarters, (ii) the firm s actual earnings figure and earnings announcement date are available for the current and prior quarters, (iii) the firm s share price, outstanding shares, and the cumulative factor to adjust price are available on CRSP for the prior quarter s earnings announcement, and (iv) the cumulative factor to adjust price is available on CRSP for the current quarter. This includes 2,293,778 analyst-firm-quarters. We drop analyst-firm-quarters (i) without a current quarter earnings forecast after the prior quarter earnings announcement date and before the current quarter earnings announcement date (917,272), (ii) without a share price target issued either with the final current quarter earnings forecast or in the 365 days before the final current quarter earnings forecast date (359,227), (iii) without a next quarter earnings forecast either with the final current quarter earnings forecast or in the 365 days before the final current quarter earnings forecast date (149,856), (iv) with a stock split or a stock dividend after the prior quarter earnings announcement date and before the current quarter earnings announcement date (13,648), and (v) without CRSP or Compustat data to calculate control variables (134,402). The final revision sample includes the remaining 719,373 analyst-firm-quarters corresponding to 8,504 unique analysts and 7,509 unique firms. These sample selection procedures are detailed in Appendix A. 12

15 3.2 Revision Sample Descriptive Statistics Panel A of Table 1 reports descriptive statistics for the 719,373 analyst-firm-quarters in the revision sample. The analyst revises a share price target after the final current quarter earnings forecast in 13.1% of the analyst-firm-quarters, including 8.3% upward revisions and 4.8% downward revisions (thus 63.5% of the share price target revisions are revised upwards). The analyst revises the next quarter earnings forecast after the final current quarter earnings forecast in 11.7% of the analyst-firm-quarters, including 6.0% upward revisions and 5.6% downward revisions (thus 51.8% of the next quarter earnings forecast revisions are revised upwards). The analyst revises either a share price target or next quarter s earnings forecast in 19.5% of analyst-firm-quarters. The high proportion of positive SPT and FQE revisions relative to negative SPT and FQE revisions suggests analysts more frequently disseminate positive news via non-earnings forecast signals after the final current quarter earnings forecast. We also report descriptive statistics for a number of control variables. RET_QTR, the firm s return over the ninety days before the analyst s final current quarter earnings forecast, has a negative median, suggesting quarterly earnings forecasts tend to be issued when prior economic news is negative. REV_CQE is an indicator variable set equal to one if the analyst revises the current quarter earnings forecast after the first current quarter earnings forecast following the prior quarter earnings announcement date, and to zero otherwise. NEG_CQE (POS_CQE) is an indicator variable set equal to one if REV_CQE equals one and the revision is negative (positive), and to zero otherwise. These variables help us evaluate the determinants of an analyst s decision to (i) revise the current quarter earnings forecast, (ii) revise the share price target or next quarter earnings forecast, or (iii) revise no forecasts. The analyst revises the current quarter earnings forecast in 32% of the analyst-firm-quarters in our sample and 59.6% of 13

16 these revisions are negative (12.9% upward revisions and 19.1% downward revisions). The predominance of negative CQE revisions contrasts with the results for SPT and FQE revisions, which are predominantly positive. Finally, the firm meets or beats the analyst s forecast of earnings (MEET_BEAT) in 70.6% of analyst-firm-quarters. 3.3 Text Sample Selection The following sample selection procedures are detailed in Appendix A. The qualitative prediction variables (MISS_TEXT, BEAT_TEXT) require the text of analyst reports. We collect analyst reports from ThomsonOne Banker for 175 randomly selected analysts in the revision sample detailed above. 10 The 175 randomly selected analysts correspond to 65,539 analyst-firmquarters from the revision sample. We search for the analyst name from the I/B/E/S recommendation file in ThomsonOne Banker and if we cannot identify the analyst by name in ThomsonOne Banker, we discard the analyst. We are unable to identify 33 of the 175 analysts in ThomsonOne Banker. The text sample is thus comprised of the 53,765 analyst-firm-quarters for the remaining 142 analysts we can identify in ThomsonOne Banker. We collect 44,237 analyst reports corresponding to the 142 analysts and 958 unique firms. We match each analyst report to an analyst-firm-quarter based on its issue date (classifying a report as belonging to an analyst-firm-quarter if the report is issued after the prior quarter earnings announcement date and before the current quarter earnings announcement date). We also drop analyst reports (i) issued before the analyst s final current quarter earnings forecast, and (ii) issued more than 30 days before the current quarter earnings announcement!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 10 We collect the analyst reports over three iterations. In the first iteration, we collect all analyst reports for 15 analysts randomly selected from the revision sample for the year In the second (third) iteration, we randomly select 60 (100) analysts with at least twenty firm-quarters of observations for a minimum of three firms in the revision sample. 14

17 date. The final text sample includes the remaining 7,593 analyst reports corresponding to 130 unique analysts and 509 unique firms. 3.4 Text Sample Descriptive Statistics Panel B of Table 1 reports descriptive statistics for the text sample of 53,765 analystfirm-quarters. In this sample of firms, the firm meets or beats the analyst s forecast in 70.6% of the observations, identical to that in the revision sample. The mean value of analysts positive qualitative predictions (BEAT_TEXT) is and the mean value of analysts negative qualitative predictions (MISS_TEXT) is 0.084, suggesting analysts make positive qualitative predictions more frequently than negative qualitative predictions. 11 The revisions to the share price target, next quarter earnings, and current quarter earnings forecasts are comparable to those in the revision sample, but the analysts in the text sample are slightly more active. 4. Research Design and Empirical Results 4.1 The Earnings Surprise and Non-Earnings Forecast Signals The Analyst s Own Earnings Surprise We first seek to understand whether research an analyst publishes after the final current quarter earnings forecast predicts the analyst s own forecast error. We address this question by regressing the analyst s own earnings surprise on information from research the analyst publishes after his final current quarter earnings forecast, but before the current quarter earnings announcement date. We estimate the following model. EarningsSurprise = β 0 + β 1 (NonEarningsForecastSignals) + Σβ i (Controls) + ε (1)!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 11 To validate our qualitative prediction variables, a research assistant read a subsample of 1,905 sentences in which the textual algorithm classified the analyst as having issued a qualitative prediction. The research assistant assessed whether the analyst clearly indicated an expectation that the firm would beat or miss expectations. We find a 0.46 (0.48) correlation between the research assistant s coding of the sentences and the textual algorithm for those sentences classified as beat (miss), suggesting we capture the intentions of the analyst, albeit with noise. 15

18 We use two measures of EarningsSurprise: (i) an indicator variable set equal to one (zero) if the firm meets or beats (misses) the analyst s own forecast (MEET_BEAT), and (ii) the analyst s earnings surprise, equal to the firm s actual reported earnings per share less the analyst s final quarterly earnings per share forecast, scaled by price (CURR_SURP). We use three measures of NonEarningsForecastSignals. First, we capture the analyst s revisions to the share price target. We set NEG_SPT (POS_SPT) equal to one if the analyst issues a negative (positive) revision to the share price target after the final current quarter earnings forecast date and before the current quarter earnings announcement date, and to zero otherwise. Second, we capture the analyst s revisions to the next quarter earnings forecast. We set NEG_FQE (POS_FQE) equal to one if the analyst issues a negative (positive) revision to the next quarter earnings forecast after the final current quarter earnings forecast date and before the current quarter earnings announcement date, and to zero otherwise. Third, we capture the analyst s qualitative non-earnings forecast signals by analyzing the text of analyst reports issued after the final current quarter earnings forecast date and before the current quarter earnings announcement date. MISS_TEXT (BEAT_TEXT) is a count variable equal to the number of times the analyst uses a synonym for miss ( beat ) and a synonym for earnings expectations within the same sentence. 12 The synonyms for miss include miss, fall short, below, lower, worse, downside, and underperform. The synonyms for beat include beat, exceed, outperform, better, higher, above, top, and upside. The synonyms for earnings expectations include earnings, EPS, estimates, and expectations.!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 12 Our inferences remain unchanged if MISS_TEXT (BEAT_TEXT) is a count variable that equals the number of times the analyst uses a synonym for miss ( beat ) and a synonym for earnings expectations within five words of each other. Our inferences also remain unchanged if MISS_TEXT (BEAT_TEXT) is an indicator variable equal to one if the analyst uses a synonym for miss ( beat ) and a synonym for earnings expectations within the same sentence or within five words of each other. 16

19 We also include a series of control variables. PREV_SURP is the prior quarter earnings surprise, equal to the firm s actual prior quarter earnings per share less the analyst s final prior quarter earnings per share forecast, scaled by price. RET_QTR is the firm s return for the period beginning ninety days before the final current quarter earnings forecast date and ending at the final current quarter earnings forecast date. MVE is the natural log of the market value of equity. BTM is the book value of equity scaled by the market value of equity. ROA is the firm s return on assets, calculated as earnings before interest and taxes scaled by total assets. MVE, BTM, and ROA are calculated as of the most recent fiscal year-end before the prior quarter earnings announcement. FOLLOW is the natural log of analyst following. #DAYS is the natural log of the number of days between the analyst s final current quarter earnings forecast and the current quarter earnings announcement date. #FORE is the natural log of the number of current quarter earnings forecasts the analyst issued in the previous quarter. All variable definitions are detailed in Appendix B. Panel A of Table 2 reports the results for the revision sample from estimating equation (1), which regresses the analyst s own earnings surprise on the analyst s non-earnings forecast signals. The dependent variable is MEET_BEAT in columns (1)-(3) and CURR_SURP in columns (4)-(6). In column (1), we estimate equation (1) without any control variables. We find positive coefficients on POS_SPT (β=0.043) and POS_FQE (β=0.018), both significant at the 1% level. We also find a significantly negative coefficient on NEG_SPT (β=-0.017) and an insignificant coefficient on NEG_FQE. The results indicate SPT and FQE forecasts convey economically significant information about the probability of a firm meeting or beating the analyst s own earnings forecast. For example, an analyst-firm-quarter with a positive share price 17

20 target revision will meet or beat the analyst s earnings forecast 6% more frequently than an analyst-firm-quarter with a negative share price target revision. In column (2), we include control variables for factors previously shown to affect the probability of a firm meeting or beating earnings expectations such as firm characteristics, prior returns, and the prior quarter earnings surprise. We find that three (four) of the four non-earnings forecast signals are statistically significant at the 5% (10%) level, but we observe substantial attenuation in the coefficient estimates. For instance, the coefficient estimate on POS_SPT decreases from β=0.043 to β=0.026 after including the control variables. It is unclear whether the coefficient estimates with or without control variables included in the model better capture the information analysts convey to clients with non-earnings forecast signals because investors may not incorporate predictable errors into analyst forecasts. In column (3), we also include analyst and quarter fixed effects and obtain similar inferences to those in column (2). In columns (4)-(6), we estimate analogous models to those in columns (1)-(3), but the dependent variable is CURR_SURP, the analyst s signed earnings surprise. In all three specifications, the coefficients on POS_SPT, NEG_SPT, and POS_FQE remain statistically significant at the 5% level, whereas the coefficient on NEG_FQE is insignificant in two specifications and significantly negative in one specification. Thus, the evidence remains consistent with the analyst s non-earnings forecast signals providing information content for the analyst s own earnings surprise. Panel A of Table 3 reports the results for the textual sample from estimating equation (1), which regresses the analyst s own earnings surprise on the analyst s qualitative predictions. We estimate analogous models to those presented in Panel A of Table 2, except the non-earnings forecast signals are the qualitative predictions, BEAT_TEXT and MISS_TEXT. Thus, in column 18

21 (1), we regress MEET_BEAT on BEAT_TEXT and MISS_TEXT without control variables. The coefficient on MISS_TEXT is significantly negative (β=-0.048) and the coefficient on BEAT_TEXT is significantly positive (β=0.050). The results indicate qualitative predictions also convey economically significant information about the probability of a firm meeting or beating the analyst s earnings forecast. For example, an analyst-firm-quarter with a positive qualitative prediction will meet or beat the analyst s earnings forecast 9.8% more frequently than an analystfirm-quarter with a negative qualitative prediction. In column (2), we include the same series of control variables as in Panel A of Table 2. We find similar coefficient estimates on MISS_TEXT and BEAT_TEXT after including these control variables and both coefficient estimates remain statistically significant at the 1% level. In column (3), we also include analyst and quarter fixed effects and find similar coefficient estimates and inferences. In columns (4)-(6), we estimate analogous models to those in columns (1)-(3), but the dependent variable is CURR_SURP, the analyst s signed earnings surprise. The coefficient on MISS_TEXT (BEAT_TEXT) remains negative (positive) and significant at the 1% level in all three specifications The Consensus Earnings Surprise In our main analysis, we examine firm performance relative to the analyst s own forecast, because an investor evaluating an analyst s research product will judge the analyst on the information in the forecasts he publishes. In this section, we examine performance relative to the consensus forecast to examine whether non-earnings forecast signals provide incremental information relative to the market s expectation of earnings. To evaluate the information in non-earnings forecast signals, we replace the analyst s own earnings surprise in model (1) with 19

22 the consensus earnings surprise calculated using earnings forecasts issued after the previous quarter s earnings announcement. We estimate the following model: ConsensusEarningsSurprise = β 0 + β 1 (NonEarningsForecastSignals) + Σβ i (Controls) + ε (2) We use two measures of ConsensusEarningsSurprise: (i) an indicator variable set equal to one (zero) if the firm meets or beats (misses) the consensus earnings forecast on the day of the analyst s final current quarter earnings forecast (MEET_BEAT_CON), and (ii) the consensus earnings surprise, equal to the firm s actual earnings per share less the consensus earnings per share forecast on the day of the analyst s final current quarter earnings forecast, scaled by price (CURR_SURP_CON). In Panel B of Table 2, we estimate analogous models to those in Panel A of Table 2, but the dependent variable is the consensus earnings surprise rather than the analyst s own earnings surprise. In column (1), we regress MEET_BEAT_CON on the analyst s non-earnings forecast signals without control variables. We find positive coefficients on POS_SPT (β=0.084) and POS_FQE (β=0.034), and negative coefficients on NEG_SPT (β=-0.045) and NEG_FQE (β=-0.018), all significant at the 1% level. Moreover, the coefficients are all economically larger than the corresponding coefficients in Panel A of Table 2. In column (2), we include control variables and in column (3) we also include analyst and quarter fixed effects. We find similar results, all eight coefficients are statistically significant with the sign predicted by after-revision bias and seven of the eight coefficients have a larger magnitude coefficient than the corresponding coefficient in Panel A of Table 2. In columns (4)-(6), we continue to find all coefficients are significant and ten of the twelve coefficient estimates are larger than the corresponding coefficient in Panel A of Table 2. 20

23 In Panel B of Table 3, we estimate analogous models to those in Panel A of Table 3, but the dependent variable is the consensus earnings surprise rather than the analyst s own earnings surprise. Across all six columns, we find that both BEAT_TEXT and MISS_TEXT have coefficients statistically significant at the 1% level with coefficient magnitudes larger than corresponding estimates using the analyst s own earnings surprise as the dependent variable. Collectively, the evidence indicates that analysts non-earnings forecast signals also provide information content about the consensus earnings surprise. 4.2 When Do Analysts Issue Non-Earnings Forecast Signals? Next, we exploit cross-sectional variation in the frequency of non-earnings forecast signals to provide evidence on the incentives that generate after-revision bias. Specifically, we test three hypotheses: (i) analysts issue non-earnings forecast signals more for positive news to avoid increasing earnings expectations ( catering to management ), (ii) analysts issue non-earnings forecasts more frequently for news that would have moved the analyst s forecast away from the consensus ( strategic herding ), and (iii) analysts issue non-earnings forecast signals more frequently for smaller magnitude information ( frictions ) Do Analysts Issue Non-Earnings Forecast Signals More Frequently for Good News? Incorporating good news into an SPT or FQE forecast, and bad news into a CQE forecast, increases the likelihood that management meets or beats the earnings forecast. Given analysts incentives to both publish forecasts that managers will meet or beat (Ke and Yu 2006; Brown et al. 2015) and communicate information to clients (Groysberg et al. 2011; Brown et al. 2015), mapping good and bad news into different forecasts may help analysts satisfy multiple stakeholders. 21

24 We report forecast frequencies for positive and negative forecast revisions in Panel A of Table 4. Row (1) indicates that analysts issue a positive non-earnings forecast signal (i.e., SPT or FQE forecast revision) in 12.2% of analyst-firm-quarters, whereas they issue a negative nonearnings forecast signal in 8.9% of analyst-firm-quarters. The 3.3% difference is statistically significant, indicating that analysts are more likely to issue positive than negative non-earnings forecast signals. Rows (2) and (3) indicate this relation holds separately for SPT and FQE forecast revisions. Row (4) documents that analysts issue a positive CQE forecast revision in 12.9% of analyst-firm-quarters, whereas they issue negative CQE forecast revisions in 19.1% of analyst-firm-quarters. The 6.2% difference is also statistically significant and indicates that analysts are more likely to revise CQE forecasts downwards. In rows (5)-(8), we exclude analyst-firm-quarters in which the analyst revises the current quarter earnings forecast (REV_CQE=1). The frequency of both positive and negative non-earnings forecast signals increases relative to rows (1)-(3), but the ratio of positive to negative revisions also increases, thus corroborating the above results. The evidence is consistent with analysts responding to incentives to issue forecasts that managers will meet or beat. Conditional on positive news, analysts are less likely to revise the CQE forecast so earnings will be compared to a static benchmark. Conditional on negative news, managers will be compared to a benchmark that varies with performance. The systematic correlation between the sign of the news and the forecasts analysts choose to revise suggests one mechanism analysts use to walk-down forecasts is omitting positive news from the CQE forecast, while responding to positive news by revising another forecast. 22

25 4.2.2 Do Analysts Issue Non-Earnings Forecast Signals More Frequently for Bold News? Second, we examine whether analysts issue SPT and FQE forecast revisions more frequently when a corresponding revision to the CQE forecast would move the analyst s CQE forecast towards the consensus (herding) or away from the consensus (bold). Prior research suggests investors judge analyst forecasts relative to those of competing analysts, and analysts face career penalties for issuing more inaccurate forecasts than their peers (Hong et al. 2000; Groysberg et al. 2011). We argue these incentives may lead analysts to avoid issuing extreme forecasts relative to those of their peers. 13 The results are reported in Panel B of Table 4. Row (1) indicates that analysts issue a negative non-earnings forecast signal (i.e., SPT or FQE forecast revision) in 9.4% (8.3%) of analyst-firm-quarters when the analyst s final CQE forecast is below (above) the consensus forecast. Row (2) indicates that analysts issue a positive non-earnings forecast signal in 11.1% (13.2%) of analyst-firm-quarters when the analyst s final CQE forecast is below (above) the consensus forecast. Thus, analysts are more likely to issue non-earnings forecast signals if a corresponding revision to the CQE forecast would move the analyst s forecast away from the consensus. This finding is opposite to that predicted by the theory that analysts non-earnings forecast signals are a function of their forecast error. The consensus forecast contains substantial information about future earnings, omitted from the analyst s own forecast. If non-earnings forecast revision frequency increased in the prevalence of news omitted from the analyst s earnings forecasts, we would predict results opposite to those we find (i.e., more herding than bold non-earnings forecast signals). Rows (3)-(6) indicate that the same relations hold for the!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 13 In a private conversation, an analyst at a prestigious bulge bracket firm commented that issuing a forecast which departs substantially from consensus draws attention. The analyst commented that without (with) a convincing reason for the deviation, the attention can damage (enhance) the analyst s reputation. 23