Earnings Announcement Clustering. and Analyst Forecast Behavior

Transcription

1 Earnings Announcement Clustering and Analyst Forecast Behavior Matthew Driskill Fisher School of Accounting University of Florida Gainesville, FL (352) (office) January 15 th, 2016 ** Please do not quote without author s permission ** I would like to thank my dissertation committee, Jenny Tucker (chair), Marcus Kirk, David Reppenhagen, and Mike Ryngaert. I would also like to thank Will Ciconte, Joost Impink, Justin Leiby, Michael Mayberry, Kathy Rupar, and Jim Vincent for their assistance and helpful comments on this project.

2 ABSTRACT This study investigates whether earnings announcement clustering creates a limited attention effect that impairs analyst forecast performance. I find that concurrent, sameday earnings announcements within an analyst s coverage portfolio occur quite frequently and negatively affect analyst forecasting behavior. Specifically, busy analysts are less likely to issue a forecast in the week after that firm s earnings announcement. When busy analysts do issue a first-week forecast, it is less timely, incorporates less information from the earnings announcement, and is less accurate following large earnings surprises. Busy analysts are also more likely to issue forecasts later in the quarter, and these forecasts are no less bold or accurate relative to their nonbusy peers. These results are consistent with limited analyst attention impairing initial forecast responsiveness, resulting in an initial analyst underreaction, and subsequent efforts by analysts to increase forecasting activity and informativeness later in the quarter. The results have implications for the literature on limited attention, analysts role in facilitating price discovery, and the cycle of analyst information production. Keywords: earnings announcements, analysts, forecasts, analyst performance, timeliness, accuracy, analyst information production, limited attention. Data Availability: Data publicly available from the sources identified in the paper. JEL Classification: M41

3 1. Introduction Bounded rationality posits that limits on the information-processing capacities of individuals, brought about by task complexity and/or time constraints, may impair fully rational decision-making and thereby lead to sub-optimal outcomes (Simon, 1955, 1972; Hirshleifer, 2001). Building upon this theory, Hirshleifer & Teoh (2003) suggest that cognitive limitations and high information loads can create a limited attention effect which impairs investors ability to fully process earnings information upon arrival, potentially resulting in delayed price responses and underreactions. More recently, Hirshleifer, Lim, & Teoh (2009) find that firms which announce earnings on days with more (less) competing earnings announcements have smaller (larger) magnitude immediate earnings-announcement window [-1, +1] price responses and experience larger (smaller) magnitude post-earnings announcement drift, and attribute this result to limited investor attention with respect to firms that announce earnings on high-news days. This paper applies a similar logic to analysts by investigating the relationship between earnings announcement clustering and analyst forecast behavior. Specifically, within a given analyst s coverage portfolio, I investigate the effect of concurrent (that is, two or more), same-day earnings announcements on the timing and magnitude of subsequent analyst forecast activity. About 75% of all firms covered by analysts announce their quarterly earnings within a four-week window from roughly three to six weeks (days 16-40) after the beginning of a typical thirteen-week calendar quarter (i.e., calendar quarters beginning January 1 st, April 1 st, July 1 st, and October 1 st ). Given that earnings announcements represent important public information disclosures, this suggests that a substantial portion 1

4 of all information arriving to analysts occurs in a relatively small window of time. In fact, since 1999 nearly 45% of all firms covered by analysts announce earnings concurrently with another firm in that analyst s portfolio. This earnings announcement clustering has also increased over time, from 36% in 1999 to 53% in 2014, and in the peak of earnings season (days 16-40), analysts have concurrent earnings announcements within their coverage portfolios just over 50% of the time. Thus not only are concurrent earnings announcements within an analyst s coverage portfolio quite frequent, but that frequency is growing over time. In addition to periods of increased information load, prior research also finds that investors value more timely analyst forecasts and associate timeliness with leader characteristics and better analysts, suggesting that, all things equal, analysts have incentives to produce forecasts in a timely fashion (e.g., Cooper, Day, & Lewis, 2001; Mozes, 2003; Clement & Tse, 2003; Keskek, Tse, & Tucker 2014). Similar to the increasing prevalence of concurrent earnings announcements, analysts appear to be responding to these public information disclosures in an increasingly rapid fashion. Zhang (2008) documents a substantial increase in analyst forecast timeliness after earnings announcements over the period from 1996 to 2002, and this trend has continued to intensify over time. In 1999, 35% of all first forecasts in a quarter (that is, disregarding subsequent quarterly revisions by analysts) were issued on day 0 or day 1 following an earnings announcement. By 2014, 63% of all first forecasts in a quarter were issued on day 0 or day 1 following an earnings announcement. The presence of concurrent earnings announcements within an analyst s coverage portfolio, which suggests increased information processing demands, coupled with 2

5 market expectations to produce ever more timely analyst forecasts, together suggest the possibility of a limited analyst attention effect. Using concurrent, same-day earnings announcements within an analyst s coverage portfolio as a proxy for limited analyst attention, I examine whether the presence of concurrent earnings announcements affects analyst forecast behavior in terms of forecast likelihood, timeliness, boldness, and accuracy in the week after the earnings announcement. I then examine whether this initial forecast behavior influences subsequent forecast behavior later in the quarter in terms of subsequent forecast likelihood, boldness, and accuracy (after the first week), and whether investors differentially price forecasts issued by busy and non-busy analysts. Over my sample period from 1999 to 2014, I find that busy analysts are less likely to produce a forecast in the week following an earnings announcement compared to nonbusy analysts, suggesting that the presence of concurrent earnings announcements impairs an analyst s ability to respond to an earnings announcement in a relatively timely fashion (i.e., issue a first-week forecast). Of analysts who do choose to issue a first-week forecast, I find that busy analysts issue less timely forecasts relative to non-busy analysts, and that these forecasts are less bold relative to the prevailing analyst consensus estimate, suggesting that these initial forecasts by busy analysts incorporate less new information from the earnings announcement. While I find no main effect results with respect to the initial forecast accuracy of busy analysts, busy analysts are less accurate than non-busy analysts following firm earnings announcements with large earnings surprises, suggesting that forecasts issued by busy analysts are less informative following large earnings surprises. In additional testing I find that busy analysts are more likely to issue forecasts later in the quarter, regardless of whether or not they issued a first-week forecast, and that 3

6 these subsequent forecasts are no less bold or accurate than the subsequent forecasts of their non-busy peers. While this suggests that busy analysts may catch up with nonbusy analysts in terms of subsequent forecast boldness and accuracy, investors appear to price these forecasts at a discount relative to their non-busy peers. Collectively these results suggest a pattern of analyst behavior which resembles the pattern of investor behavior documented by Hirshleifer et al. (2009). When confronted with heightened information processing demands, analysts appear to initially underreact to earnings announcement information and then increase their forecasting activity at a later date when the constraints have loosened. My results thus inform literature documenting underreactions to widely available and salient information disclosures, limited attention in particular, by using a unique setting to test the effects of information load, information processing, and performance. My results also provide additional evidence on the role that analyst play in facilitating market (in)efficiency and price discovery. Zhang (2008) finds that firms with timely (day 0, 1) analyst forecasts display larger magnitude immediate price reactions and less subsequent post-earnings announcement drift, and attributes these results to analyst forecasts mitigating price drift. Given that prior research finds that analyst forecasts throughout the quarter generate significant market reactions, and that a substantial portion of price drift in a given quarter occurs in short windows around subsequent analyst forecasts (Gleason & Lee, 2003), this suggests that earnings announcement clustering, and the effect it has on analyst forecast behavior, may play a prominent role in terms of the speed with which prices impound the information contained in earnings announcements. 1 1 I do not examine leader and follower analysts, but a number of papers have investigated the differential forecast characteristics and market reactions to these analysts (e.g., Cooper, Day & Lewis, 4

7 Finally, while research has previously documented earnings announcement clustering (e.g., Ramnath, 2002; Thomas & Zhang, 2008), my study documents the frequency of concurrent earnings announcements in analyst's portfolios and the effect that this clustering has on analyst forecast behavior, particularly with respect to the timing and the production cycle of analyst forecasts. If the majority of public information from publicly-traded companies tends to arrive to the market in a compressed timeframe, and if analyst face incentives to interpret and disseminate this information in a timely manner, the most efficient use of analyst resources is likely to delay private information acquisition and processing with respect to a given firm to a later date, when the analyst has more time to do so. This provides support for the findings of Chen, Cheng, & Lo (2010) and Keskek et al. (2014) who find that analyst information production appears to be more concerned with quickly disseminating public information after public disclosures (such as earnings announcements) and more focused upon generating private information in the weeks and months after and prior to earnings announcements. The paper proceeds as follows. Section 2 provides a review of the relevant literature and hypothesis development. Section 3 discusses the sample, tests and research design. Section 4 provides results. Section 5 provides additional tests. Section 6 summarizes and concludes. 2. Prior research and hypothesis development The (semi-strong) efficient markets hypothesis (EMH) holds that stock prices reflect all public information without bias and that new information is priced 2001; Gleason & Lee, 2003). While these papers find that the market reacts more strongly to forecasts by leader analysts, forecasts by follower analysts are still found to generate significant price impacts (Shroff, Venkataraman & Xin, 2014). 5

8 instantaneously, thus effectively assuming information processing as unbiased and instantaneous. Despite the general resilience of the efficient market hypothesis for conceptualizing the relationship between information and prices, the last few decades have seen the gradual emergence of empirical results and theories from behavioral finance challenging the hypothesis (e.g., Barberis, Schliefer, & Vishny, 1998; Daniel, Hirshleifer & Subrahmanyam, 1998; Hirshleifer, 2001). Recognizing the vastness of information and variation in acquiring or processing information due to variation in complexity, and the heuristics which might arise from these scenarios, some of the more effective challenges to EMH have focused on the assumption of instantaneous and unbiased information processing. One of the more modest departures from the classic rational choice theory underpinning EMH is the theory of bounded rationality. While rational choice theory acknowledges that actors are limited by the information available to them, bounded rationality posits that individuals may be further subject to cognitive limitations and time constraints, and that these forces impair fully rational decision-making (Simon, 1955, 1972; Hirshleifer, 2001). Stated simply, information acquisition and processing can be difficult and time consuming, and therefore may be impaired depending on the difficulty of the task or the time available to complete the task. Bounded rationality can lead to limited investor attention where investors are unable to process information immediately, completely, and without bias due to the presence of other, distracting information and that this limited investor attention can manifest as an underreaction where information is gradually impounded into stock prices over time (Hirshleifer & Teoh, 2003; DellaVigna & Pollet, 2009; Hirshleifer et al., 2009). 6

9 While much previous research regarding information processing has focused on how investors price information, I apply limited attention theory to an analyst s coverage portfolio: limited analyst attention may arise due to the confluence of increased information processing demands brought about by concurrent earnings announcements (cognitive constraints due to increased information load) coupled with the analyst desire/market expectation of timely forecasts (time constraints). Assuming that interpreting the information contained in an earnings announcement is somewhat complex (i.e., not instantaneous), a small expected forecast response window suggests that the presence of concurrent (two or more, three or more, etc.) earnings announcements, which may essentially double or triple the effective workload for an analyst, represents an unique setting to test whether and how an increased workload affects the attention and cognitive resources that the analyst is able to employ with respect to a given earnings announcement when the expected response time approaches immediate. Given an increased workload, one potential analyst response would be to not produce a forecast in the days after the earnings announcement at all, and simply generate a forecast later in the quarter when the cognitive and time constraints have loosened. Another potential response would be to generate a forecast in the days after the earnings announcement, yet given the constraints involved, one might expect the forecast to be either less timely reflecting more effort by the analyst to generate private information or more conservative or less accurate reflecting less effort by the analyst to generate 7

10 private information or that the analyst may trade-off some combination of being both less timely and accurate. 2 While much initial research on analyst performance focused on analyst accuracy (e.g., O Brien, 1990; Stickel, 1990, 1992; Mikhail, Walther, & Willis, 1997; Sinha, Brown, & Das, 1997; Clement, 1999; Jacob, Lys, & Neale, 1999), subsequent research suggests that analyst accuracy may mainly be of importance to analysts in the sense of avoiding excessive inaccuracy and possible termination (Mikhail et al., 1999; Hong, Kubik, & Solomon, 2000; Hong & Kubik, 2003). 3 More recent research finds evidence that analysts and investors may value forecast timeliness, forecast boldness or general forecast quality more (Mozes, 2003; Clement & Tse, 2003, 2005; Keskek et al., 2014). Despite this, consensus analyst forecast estimates are widely accepted as a proxy for market expectations and therefore establish the information content of earnings announcements in terms of earnings surprises. Combining the predictions of limited attention theory with commonly accepted indicators of analyst performance, I investigate the following hypotheses in relation to the likelihood, timeliness, boldness and accuracy of analysts initial forecast behavior: H1: Limited analyst attention negatively affects the likelihood of an analyst issuing a forecast in the first week after the firm earnings announcement. H2a: Among analysts who issue a forecast in the first week after earnings, limited analyst attention negatively affects analyst forecast timeliness. H2b: Among analysts who issue a forecast in the first week after earnings, limited analyst attention negatively affects analyst forecast boldness. 2 I do not directly investigate herding, but herding should work against me finding results by allowing busy analysts to quickly incorporate the information of other analysts into their forecasts. 3 More recently, the survey results of Brown, Call, Clement & Sharp (2015) suggest that accuracy is a relatively low priority for security analysts. 8

11 H2c: Among analysts who issue a forecast in the first week after earnings, limited analyst attention negatively affects analyst forecast accuracy. I further investigate subsequent analyst forecast behavior throughout the remainder of the quarter after the first week following earnings, though I do not present formal hypotheses. If the limited attention effect I document in this setting is valid, one would expect the forecast performance of busy analysts to improve relative to non-busy analysts once the constraints brought about by concurrent earnings announcements have loosened. 3. Sample and research design 3.1 Sample The primary data source for the sample is adjusted I/B/E/S quarterly earnings announcement and forecast data. I start by downloading a quarterly window of consecutive (qt-1 and qt) firm earnings announcement dates and earnings results, and then download all qt quarterly EPS forecasts issued between the qt-1 and qt earnings announcement dates. Since some analysts who cover a firm may not issue a qt quarterly EPS forecast during qt, I also capture the existence of any analyst activity in the quarter prior to the qt-1 earnings announcement and after the qt earnings announcement, and append these inactive qt analysts to the sample of active qt analysts. My sample begins in 1999 because of the widespread introduction of I/B/E/S timestamps in this year. 4 Beginning the sample at this time also corresponds to the period of increased forecast timeliness following earnings announcements documented by Zhang (2008). I use the I/B/E/S actuals file for firm identifiers, earnings dates, times and values, the I/B/E/S detail estimates file for individual analyst forecast dates, times and values. I exclude firm- 4 I/B/E/S timestamps are highly prevalent beginning in 1999, representing nearly 100% of the sample. I delete the observation if the timestamp is missing (dehaan, Shevlin, & Thornock, 2015). 9

12 quarters where the number of days between earnings announcements is less than 60 or more than 120 days, and I also exclude firm-quarters where the earnings announcement occurs more than 90 days after the end of the fiscal quarter (e.g., dehaan et al., 2015). I do so to avoid firms that change fiscal quarter dates, have restatements or postpone earnings announcements due to financial reporting quality concerns. This aggregated file form the basis of my tests of forecast characteristics (e.g., first-week forecast likelihood, first-week forecast timeliness, first-week forecast boldness and accuracy). I calculate analyst coverage, consensus analyst estimates (to establish unexpected earnings), and analyst dispersion for qt-1 from the I/B/E/S details file from all analysts covering the firm during qt-1. I calculate analyst consensus qt earnings estimates (to establish forecast accuracy) from all analysts who issue a qt quarterly EPS forecast during qt-1. I capture the presence of managerial guidance in concert with the qt-1 earnings announcement from the I/B/E/S guidance file. I use trading days from CRSP to avoid counting holidays and weekends in the calculation of forecast lag. I construct continuous trading-day forecast lag using I/B/E/S timestamps for both the earnings announcement and all forecasts issued between the qt-1 and qt earnings announcement dates. Prior research finds analysts forecasts on nontrading days are highly infrequent and often are intentionally issued at these dates to minimize market attention and curry favor from management (Rees, Sharp, & Wong, 2015). Since roughly 40% of all quarterly EPS forecasts issued by analysts in qt occur on day 0 or day 1 following the qt-1 earnings announcement, I/B/E/S timestamps allow more precise measurement of analyst forecast lag across shorter (intraday) timeframes. 5 I also 5 The majority of analyst forecasting activity appears to occur between 7am and 7pm, which would correspond to a 12-hour work day being typical for analysts. 10

13 capture price information from CRSP in order to calculate deflated earnings and forecast variables (dispersion, absolute earnings surprise, forecast boldness, and forecast error). I delete analyst observations where the analysts issues more than six forecasts in the quarter, and I delete observation with stock prices under $5. I capture firm characteristics (firm size, book-to-market, 2-digit SIC, fiscal quarter) from the Compustat quarterly fundamentals file and merge these characteristics with my aggregate file after deleting firms lacking positive book value. Finally, I delete the extreme 1% of observations on qt-1 individual forecast accuracy in case these outlier values are due to data errors, and then winsorize all continuous variables at the 99% level. In total, the H1 sample includes all first forecasts issued by active analysts between the qt-1 and qt earnings announcement dates, as well as a unique observation for each inactive analyst (242,559 observations), for a total of 1,339,317 observations. The sample for H2 includes all first-week forecasts following the qt-1 earnings announcement date, for a total of 833,241 observations. In supplemental tests, I examine the 592,109 analyst forecasts that occur after the first week following the qt-1 earnings announcement date. 3.2 Research design Figure 2 provides an overview of the research design by presenting analyst qt forecast behavior during the quarter following the qt-1 earnings announcement as a decision tree. I consider the first week after the qt-1 earnings announcement to represent the analyst s first decision in the quarter: whether or not to produce a first-week forecast. In line with prior research, I consider qt EPS forecasts issued in the first week (five 11

14 continuous trading days) following the qt-1 earnings announcement because prior research considers these forecasts as reasonably timely, specifically reflecting the information contained in the earnings announcement (Chen et al., 2010; Keskek et al., 2014). 6 This first-week window also captures over half (53%) of all analyst forecast activity in the quarter, and partially controls for the timing element of analyst forecast attributes (analyst forecasts and revisions tend to drift over the quarter, and more recent forecasts tend to be more accurate). In terms of variable construction, I draw upon prior research by Clement & Tse (2003, 2005) on analyst timeliness and accuracy by scaling all continuous variables from 0 to 1, thus facilitating the economic interpretation of regression coefficients. I amend the Clement & Tse approach to explicitly control for firm and earnings announcement characteristics, in addition to analyst characteristics as they do, by scaling all non-categorical variables from 0 to 1 per quarter rather than per firmquarter or firm-year, according to the following form: 7 Characteristic_Scaledijt-1 = (Raw_Characteristicijt-1 Raw_Characteristic_mint-1) / (Raw_Characteristic_maxt-1 Raw_Characteristic_mint-1). where a raw characteristic (e.g., firm size, brokerage size, absolute earnings surprise, etc.) corresponds to individual analyst i, for firm j in quarter t-1 is scaled against the minimum and maximum values of the raw characteristic in quarter t-1. This compares an individual 6 While highly infrequent for an analyst to issue two forecasts in the week following the q t-1 earnings announcement, I also delete any revisions in this window, keeping only the analyst s first forecast. 7 Clement & Tse (2003, 2005) scale variables by firm-year since they are examining annual earnings forecasts over the course of the year leading up to the year-end earnings announcement; I examine quarterly earnings forecasts, but scale on a quarter rather than firm-quarter basis. Kim, Lobo & Song (2011) scale by quarter as well. I do not scale continuous trading-day forecast lag due to the ease of interpreting daily timeliness. 12

15 analyst relative to all other analysts who produce any forecast for any firm during the quarter, rather than simply other analysts who produce a forecast for that same firm. 8 My hypotheses test the effect of limited analyst attention (BUSY) on first-week analyst forecast behavior in terms of forecast likelihood, forecast timeliness, forecast boldness, and forecast accuracy. I indicate whether the qt quarterly EPS forecast is issued in the first week following qt-1 earnings, forming the basis of my first-week forecast likelihood variable (H1). The continuous forecast lag from each observation leads to the formation of my first-week forecast lag variable (H2a). I capture first-week forecast boldness as the difference between the analyst s first-week forecast and the consensus qt analyst estimate on the eve of the qt-1 earnings announcement (H2b). I calculate forecast accuracy as the difference between the analyst s first-week qt forecast and the eventual actual qt earnings as my accuracy variable (H2c). I capture limited analyst attention (BUSY) by the presence (1) or lack of (0) concurrent, same-day earnings announcements for a given analyst on the same day. The basic regression equation for these tests is specified as follows: Prob(FWFCASTijt = 1), FWFCAST_LAGijt = f (β0 + k αk*yeardummyk + β1*busyijt-1 + β2*sizejt-1 + β3*btmjt-1 + β4*covjt-1 + β5*dispjt + β6*bsizeit-1 + β7*apsizeit-1 + Β8*AINDit-1 + β9*afexpijt-1 + β10*days_elapsedijt-1 + β11*auejt-1 + β12*bnewsjt-1 8 There are advantages to both approaches. While scaling by firm-quarter implicitly controls for firm and earnings announcement characteristics, it can also suffer from issues related to sampling size. It requires at least 2 analysts to follow a firm in order to generate comparative statistics, and thus disregards the 10% of firms followed by a single analyst, and it may present further issues for firms with modest analyst following. As an example, suppose a firm is followed by 2 analysts (which also occurs in roughly 10% of the sample). In scenario 1, Analyst A works at a brokerage house that employs 100 analysts; Analyst B works at a brokerage house that employs 10 analysts. In scenario 2, Analyst A works at a brokerage house that employs 50 analysts; Analyst B works at a brokerage house that employs 49 analysts. Both scenarios would generate a coefficient of 1 for Analyst A and 0 for Analyst B, despite the fact that the raw differences are quite substantial. Scaling by quarter rather than firm-quarter alleviates this difficulties. 13

16 + β13*lossjt-1 + β14*guidejt-1 + β15*qtr4jt-1 + εijt). (1) where each observation corresponds to individual analyst i, firm j in quarter t or t-1. 9 Detailed variable descriptions are provided in the appendix. I control for firm, analyst and earnings announcement determinants of analyst forecast timeliness and boldness. I use Compustat data to calculate firm size (SIZE) and book-to-market (BtM). These variables capture primary firm characteristics associated with firm complexity and information environment, and are commonly associated with stock returns and firm valuation (e.g., Fama & French, 2015). I employ analyst coverage (COV) and analyst dispersion (DISP) as additional proxies for firm-level information environment and information uncertainty (Zhang, 2006a; Zhang, 2008). These variables tend to capture similar information, but analyst coverage can be more directly linked to possible herding activity by analysts when analyst dispersion is included. I control for analyst characteristics previously identified in the literatures on analyst accuracy and analyst timeliness, in particular brokerage size (BSIZE), analyst portfolio size (APSIZE), the number of industries covered by an analyst (AIND), analyst firm-specific experience (AFEXP), and days elapsed since the analyst s previous qt quarterly EPS forecast (DAYS_ELAPSED) (Clement, 1999; Jacob et al., 1999; Clement & Tse, 2003, 2005) Earnings announcement variables are selected from prior research on analyst forecast timeliness (Stickel, 1989; Zhang, 2008). I use absolute qt-1 unexpected earnings scaled by stock price at the end of qt-1 (AUE) to control for the amount of new information in the qt-1 earnings announcement. I include categorical variables for the 9 The moment of the q t-1 earnings announcement marks the shift from q t-1 to q t, hence, all variables are constructed from the q t-1 fiscal quarter and earnings announcement, and are then used to model the characteristics of a q t forecast issued within the first week following this q t-1 earnings announcement. 14

17 presence of a loss (LOSS) or missing analyst expectations (BNEWS) in qt-1. I also include categorical variables for qt-1 earnings announcements accompanied by managerial guidance (GUIDE) or that mark a fiscal year end (QTR4). For my multivariate regressions (H2), I provide additional variables to reflect ways in which limited attention might differentially interact with particular firm, analyst or earnings announcement characteristics. This provides further insights into whether busy analysts are uniformly affected by increased workload when producing forecasts, or whether they forecast differentially based on varied information environments. In particular, I capture how an analyst with limited attention might interact with two additional elements of the forecasting environment: the information environment and the possibility of herding. Earnings signals that convey little new information for example, actual earnings results that reflect and meet market expectations may provide little new information to the analyst ex post, and thus have minimal effect on analyst forecast timing or boldness. On the other hand, I would expect busy analysts covering firms that announce large earnings surprises to possibly underreact to such earnings surprises to a greater extent than non-busy analysts, as larger earnings surprises should reflect larger (and likely more uncertain) information content. I interact my busy analyst variable with earnings surprise as a result (BUSY*AUE). I apply a similar rationale to analyst dispersion, a proxy for ex ante information uncertainty, and interact this with my proxy for limited analyst attention (BUSY*DISP). Firms with more complicated information environments should place larger cognitive demands upon analysts, and one might expect a differential effect with respect to busy analysts relative to non-busy analysts. 15

18 I also investigate possible herding activity in the form of interactions between busy analysts and analyst coverage and analyst portfolio size. For example, if an analyst is the only analyst following a given firm, herding is mechanically impossible. On the other hand, firms followed by numerous analysts should allow for possible herding activity in response to heightened information processing demands, and thus one might expect busy analysts to be more likely to engage in such herding behavior as an efficient coping strategy (BUSY*COV). Finally, interacting busy analysts with analyst portfolio size may further capture heightened information processing demands upon analysts (BUSY*APSIZE) by capturing analysts who are not only experiencing the acute effects of concurrent earnings, but also have larger coverage portfolios in general, and thus may devote less attention to each firm in their portfolio, beyond the acute effect of same-day earnings announcements. I report these interactions in all multivariate regressions. After assessing the initial forecast likelihood and timeliness effects (H1 and H2a), I investigate forecast boldness (H2b) and accuracy (H2c). Since prior research documents a trade-off between analyst forecast timeliness and forecast accuracy, I further control for the timeliness of the analyst forecasts when assessing forecast boldness and accuracy (Cooper et al., 2001). For example, follower analysts may incorporate information from leader analysts, and thus may show differential performance in terms of accuracy. I indicate the first firm forecast (FFF) issued by an analyst following a firm as a categorical variable, and scale the amount of time between the first firm forecast and the individual analyst forecast (HERD_LAG), scaled from 0 to 1 as with the other continuous variables. The regression below mimics the basic regression model (1) above, but additionally includes FFF and HERD_LAG: 16

19 FWFCAST_BOLDijt, FWFCAST_ACCijt, = f (β0 + k αk*yeardummyk + β1*busyijt-1 + β2*sizejt-1 + β3*btmjt-1 + β4*covjt-1 + β5*dispjt + β6*bsizeit-1 + β7*apsizeit-1 + Β8*AINDit-1 + β9*afexpijt-1 + β10*days_elapsedijt-1 + β11*fffijt + β12*herd_lagijt + β13*auejt-1 + β15*bnewsjt-1 + β15*lossjt-1 + β16*guidejt-1 + β16*qtr4jt-1 + εijt). (2) For H2b, I define individual forecast boldness (FWFCAST_BOLD) as the absolute value of the difference between analyst i s qt EPS forecast issued in the week after the earnings announcement and the consensus qt analyst estimate prior to the qt-1 earnings announcement, deflated by firm stock price at the end of qt-1: FWFCAST_BOLDijt = (Forecastijt Prior_Consensusjt-1)/Pricejt-1. I then scale these results from 0 to 1 as with other non-categorical explanatory variables. For H2c, I define individual forecast accuracy (ACC) as the absolute forecast error for qt (Actual Earningsjt Forecastijt), deflated by price at the end of qt-1, then scaled from 0 to 1 per the following: FWFCAST_ACCijt = (ACC_maxjt ACCijt) / (ACC_maxjt ACC_minjt). 10 Finally, I include year-fixed effects and cluster standard errors by firm for all regressions (Peterson, 2008; Zhang, 2008). 4. Results I begin my analysis with basic descriptive statistics on the nature of earnings announcement clustering and analyst forecast responses. Earnings announcements tend to be highly clustered in calendar time. Panel A of Figure 1 indicates that over the sample 10 Distinct from the previous variable scalings, the equation for ACC is amended such that a 0 value corresponds to inaccuracy and a value of 1 corresponds to accuracy. This facilitates interpretations and follows Clement & Tse (2003, 2005). 17

20 period from 1999 to 2014, about 75% of all earnings announcements covered by analysts occur within a four-week window from roughly three to six weeks after the beginning of a typical thirteen-week calendar quarter. Panel B of Figure 1 indicates analyst forecast lag in the aftermath of an earnings announcement for all forecasts. The majority of quarterly analyst forecasts are issued in the days immediately after an earnings announcement. Further, earnings clustering and the timeliness of analyst forecasts after earnings announcements are also increasing over time. The increasingly concentrated arrival of public information also corresponds to a general increase in the size of analyst coverage portfolios from roughly 10 in 1999 to over 15 by 2014, which creates more opportunity for earnings announcement overlap on an individual analyst basis (Table 1). Table 1 also suggests that the improved analyst responsiveness documented by Zhang (2008) over the period from 1996 to 2002 has continued to increase over time. From the sample of all first forecasts by analysts issued during the quarter, mean forecast lag has declined from about 1.43 days in 1999 to 0.91 days in 2014 (Table 1). 11 Concurrent (two or more) earnings announcements on the same day are quite common in the sample, rising from about 35% of all earnings announcements covered by analysts in 1999 to 53% in For the entire sample, 45% of all earnings announcements covered by analysts occur concurrently with another earnings announcement covered by that same analyst. 12 In untabulated results, of this 45%, about 24% of observations occur concurrently with exactly one other same-day earnings announcement, 12% of observations occur concurrently with exactly two other same-day 11 If one considers all first forecasts issued in the first month, mean forecast lag has declined from about 15 days to 6.7 days. 12 So, for example, if an analyst covers five firms, and two of the five firms in the analyst s portfolio announce earnings on the same day, then 40% of the firms that analyst covers occur on the same day. 18

21 earnings announcements, and 8% of observations occur concurrently with three or more same-day earnings announcements. In terms of concurrent earnings announcements (four or more) over a three-day window [-1, +1], the numbers reflect a similar pattern: 29% of observations occur concurrently with three other three-day earnings announcements. More specifically, 11% of observations occur concurrently with three other three-day earnings announcements, 7% of observations occur concurrently with four other threeday earnings announcements, 4.5% of observations occur concurrently with five other three-day earnings announcements, and the remaining 6.5% of observations occur concurrently with six or more other three-day earnings announcements. Panel A of Table 2 provides the means of raw firm, analyst and earnings announcement characteristics from all analysts covering firms during qt, corresponding to the H1 sample. Panel B of Table 2 includes the means of scaled firm, analyst and earnings announcement characteristics conditional on an analyst being busy (BUSY = 0, 1). Busy analysts tend to work at larger brokerage houses, have larger coverage portfolios, cover more industries and have more firm-specific experience. Thus busy analysts appear to share some characteristics associated with better forecast performance working for larger brokerage houses and having more firm-specific experience and some characteristics associated with weaker forecast performance having larger coverage portfolios and covering more industries. Table 3 presents Pearson/Spearman correlations of all regression variables. The concurrent earnings announcement categorical variable (BUSY) is positively correlated with continuous trading-day forecast lag (FWFCAST_LAG) and negatively correlated with forecast boldness (FWFCAST_BOLD) and accuracy (FWFCAST_ACC). These 19

22 correlations provide some preliminary evidence that concurrent earnings announcements negatively affect analyst forecasting behavior. BUSY is positively correlated with SIZE, BtM, analyst coverage (COV) and analyst dispersion (DISP); as in univariate mean comparisons, it is positively correlated with brokerage size (BSIZE), analyst portfolio size (APSIZE), the number of industries an analyst covers (AIND) and analyst firmspecific experience (AFEXP). BUSY is positively correlated with absolute unexpected earnings (AUE) and missing analyst consensus expectations (BNEWS). Finally, BUSY is negatively correlated with the presence of a loss (LOSS), guidance (GUIDE) and fourth quarter earnings announcements (QTR4). Table 4 provides results for H1, the likelihood of an analyst issuing a first-week forecast (FWFCAST = 1) given the presence (or lack) of concurrent earnings announcements (H1). The coefficient on my variable of interest (BUSY) is significant and negative, as predicted, suggesting that concurrent earnings announcements impair an analyst s ability to issue a forecast in the week following the earnings announcement. Firm size (SIZE) and DAYS_ELAPSED since analyst i s last forecast for firm j have substantial positive associations with the likelihood of a first-week forecast. All four analyst-specific characteristics brokerage size (BSIZE), firm-specific experience (AFEXP), analyst portfolio size (APSIZE) and industry coverage (AIND) are negatively associated with the likelihood of a first-week forecast, with BSIZE representing a particularly substantial effect. Absolute earnings surprise magnitude (AUE) is negatively associated with first-week forecast likelihood, whereas bad news (BNEWS) and losses (LOSS) are positively associated with first-week forecast likelihood. As expected, the presence of managerial guidance (GUIDE) accompanying 20

23 the earnings announcement is positively associated with the likelihood of a first-week forecast. Fiscal-year-end earnings announcements (QTR4) are also positively associated with first-week forecast likelihood. A few of these results appear to contradict early research by Stickel (1989) on the determinants of analyst responsiveness in the two weeks after an earnings announcement. Stickel (1989) finds that larger unexpected earnings (AUE) and firms with more analyst coverage (COV) generate more analyst forecasting activity, and motivates these results by hypothesizing that analysts have greater incentives to produce timely forecasts under these circumstances. 13 I find the opposite with respect to both: AUE and COV are negatively related to the likelihood of a first-week forecast. Stickel (1989) also finds firm size (SIZE) to produce less analyst forecast activity, whereas I again find the opposite: SIZE substantially increases the likelihood of a first-week forecast. My results corroborate the findings of Stickel (1989) with respect to analyst dispersion (DISP) and missing earnings expectations (BNEWS). Some of these contradictory results may speak to the complex relationship between analyst ability and willingness to issue forecasts (essentially, analyst forecast supply) and investor demand for such information. For example, while one might expect firms with more analyst coverage (COV) to have lower information acquisition costs, more potential for herding, and to require less analyst effort in terms of producing a forecast, given the duplicative nature of analyst forecasts for firms covered by many analysts in terms of producing new information, firms with more analyst coverage may lead to fewer forecasts per analyst covering the firm due to this lack of analyst incentive and investor demand for forecasts from firms with richer information environments. My results with respect to AUE suggests that higher 13 Stickel (1989) also uses a two-week response window rather than a one-week response window. 21

24 information uncertainty and more difficult information processing responsibilities decreases the likelihood of a first-week forecast. Further tests appear to support my result and contradict the Stickel (1989) result, as firms with larger earnings surprises experience more subsequent forecast activity due to the rapidity and uncertainty with which their earnings expectations change over the course of a quarter. The first column of Table 5 provides results for first-week forecast timeliness (FWFCAST_LAG) given the presence (or lack) of concurrent earnings announcements (H2a). The presence of concurrent, same-day earnings announcements (BUSY) is significant and in the predicted positive direction that is, increasing the overall firstweek forecast lag suggesting that concurrent earnings announcements impair an analyst s ability to forecast in a timely fashion if they choose to issue a first-week forecast. The coefficients on analyst characteristics are as expected with respect to BSIZE, APSIZE, AIND, and AFEXP, though APSIZE is insignificant. BSIZE and AFEXP are associated with more timely forecasts, while APSIZE and AIND are associated with less timely forecasts. 14 Finally, the coefficient magnitude on BUSY (0.06) is comparable to those of APSIZE (0.02), AIND (0.08), and AFEXP (-0.07). The largest coefficient magnitude of all is on COV (-0.35), followed by BtM (0.33) and BSIZE (-0.31), suggesting that these characteristics have the most explanatory power with respect to determining the timeliness of analyst response in the week after an earnings announcement. Interestingly, while BSIZE and COV decrease the likelihood of an analyst issuing a first-week forecast (Table 4), if the analyst chooses to issue a firstweek forecast, BSIZE and COV increase the timeliness of the analyst forecast (Table 4). 14 Before the inclusion of BUSY, APSIZE loads significantly as found in prior research. The coefficient becomes insignificant with the inclusion of BUSY, suggesting that BUSY may capture some of the effect previously associated with APSIZE. 22

25 While both H1 and H2a captures elements of analyst forecast timeliness, these results highlight the distinction between whether or not to issue a first-week forecast, versus the subsequent timeliness of that forecast if one does. Column two of Table 5 provide results of basic regression model (1) with four additional interactions designed to capture any differential effects stemming from information environment (BUSY*DISP, BUSY*AUE), as well as the possibility of herding (BUSY*COV) and potential desire to herd (BUSY*APSIZE). BUSY*COV loads in the expected negative direction (-0.03), but insignificantly, suggesting that the possibility of herding appears to modestly improve the forecast timeliness of busy analysts relative to non-busy analysts. BUSY*DISP loads in the expected positive direction (0.04) but also insignificantly, suggesting that higher ex ante information uncertainty has a modestly more negative effect on the forecast timeliness of busy analysts relative to non-busy analysts. The most notable interaction effect is the coefficient on BUSY*APSIZE (0.33), rivaling that of COV, BtM, and BSIZE, and thus indicating a substantial interactive effect between BUSY and APSIZE. Analysts who are both busy and who have large portfolios are particularly susceptible to delayed forecast timeliness. Considering that the mean forecast lag is about one day, the size of this interaction represents about 33% of the average forecast lag. In terms of interpreting the size of this effect, multiplying this coefficient by a twelve-hour analyst workday indicates that analysts who are both busy and have very large portfolios issue first-week forecasts about four hours later, on average, than analysts with either small portfolios and/or no concurrent earnings announcements. The BUSY*AUE interaction is completely 23

26 insignificant, suggesting that earnings surprise has no differential effect on the timeliness of first-week forecasts of busy analysts relative to non-busy analysts. Table 6 provides results on forecast boldness (FWFCAST_BOLD), and thus can be interpreted as the amount of information which an analyst impounds into the qt EPS forecast as a result of the qt-1 earnings announcement (H2b). While somewhat modest in size (-0.004), as predicted, busy analysts forecast more conservatively relative to nonbusy analysts. DISP (0.401) and AUE (0.178) represent the most economically meaningful coefficients by a substantial margin, as one might expect when modeling changes in forecast boldness relative to an earnings announcement. The R-squared for both regressions are around 33.42%, which suggests that the basic model has substantial explanatory power, but also that the proposed interactions have little incremental effect. None of the interaction variables load in a significant fashion, due to large standard errors. 15 That said, both BUSY*DISP and BUSY*AUE load in the expected direction (0.012 and respectively), suggesting that busy analysts are insignificantly more susceptible to higher ex ante and ex post earnings information uncertainty. Similarly loading in the expected direction, BUSY*APSIZE loads negatively (-0.003), suggesting that analysts who are both busy and have large coverage portfolios deviate less from consensus analyst expectations, though insignificantly so. Table 7 provides results for forecast accuracy (FWFCAST_ACC) given the presence (or lack) of concurrent earnings announcements (H2c). I find that busy analysts are insignificantly less accurate than non-busy analysts, suggesting that the presence of concurrent earnings announcements has little effect on analyst accuracy. In terms of 15 The coefficients on three of the four interaction variables (BUSY*COV, BUSY*DISP, and BUSY*AUE) are larger than some significant coefficients of main effects (e.g., BSIZE, FFF, and BNEWS). 24

27 control variables, DISP (-0.385) and AUE (-0.251) again represent the most economically meaningful coefficients, in agreement with the results above. These results are consistent with prior research which finds that large earnings surprises tend to generate larger forecast errors and more subsequent forecast activity and post-analyst-revision drift in annual analyst forecasts (Zhang, 2006b). Interaction variables are in the expected directions, as in Table 6, but only BUSY*AUE is significant at the 5% level (-0.020). While the size of this coefficient is modest compared to DISP and AUE, it represents one of the larger coefficients in the model, as only BtM (-0.056) and COV (0.039) are larger. This suggests that busy analysts are less accurate with respect to larger earnings surprises relative to non-busy analysts, thus providing further evidence consistent with the information uncertainty story in Zhang (2006b). It also suggests that busy analysts are perhaps more likely to revise these forecasts later in the quarter, given their relatively higher degree of inaccuracy relative to non-busy analysts. 16 The R-squared for both regressions is over 41%, again indicating that the models have substantial explanatory power. Taken together, the evidence suggests that busy analysts generally perform more poorly than non-busy analysts with respect to a variety of indicators of analyst performance over the first week following the earnings announcement. More specifically, their initial ability to process public information and generate private information appears to be slightly impaired under the constraint of increased information load relative to nonbusy analysts. The stronger results appear to be with respect to the timing aspects of analyst forecasting, both in terms of economic and statistical significance, suggesting that analysts are more likely to sacrifice timeliness to achieve accuracy. The weaker results 16 I investigate this claim in section 5 of the paper. 25