Market (In)Attention and Earnings Announcement Timing

Transcription

1 Market (In)Attention and Earnings Announcement Timing Ed dehaan Graduate School of Business Stanford University Terry Shevlin Merage School of Business University of California at Irvine Jacob Thornock Foster School of Business University of Washington February 20, 2014 Ed dehaan acknowledges support from the Stanford Graduate School of Business. Jake Thornock acknowledges support from the Foster School of Business. Terry Shevlin acknowledges support from the Merage School of Business. Wall Street Horizon kindly provided data on earnings announcement timing. We thank Malcolm Bain for his generous assistance with the RavenPack data. Abridged version.

2 Market (In)Attention and Earnings Announcement Timing Abstract We revisit a long-standing but still unresolved question: do managers attempt to hide bad earnings news by announcing earnings during periods of low market attention? Or conversely, do managers highlight good news by reporting during periods of higher attention? We posit three conditions that must be met to answer this question in the affirmative: (i) managers must change their earnings announcement timing frequently enough that to do so would not immediately attract unwanted attention to bad news; (ii) there must be variation in market participants attention that is predictable to the manager ex-ante; and (iii) we must observe that managers tend to release more negative earnings news during periods of lower expected attention. The first and second conditions have not been directly examined. Prior studies examining the third condition have produced mixed results. We examine three specific times during which prior research has speculated that market attention is lower: after trading hours, on Fridays, and on busy days when numerous other firms are reporting earnings. We find that earnings announcement timing are highly variable, which supports the first necessary condition. Using four measures of market attention, we find that attention does appear to be lower after hours and on busy reporting days. However, we find that attention is the same or even higher on Fridays, which is inconsistent with the second condition. Finally, we find that unexpected earnings are lower in all three settings. In sum, the results are consistent with managers strategically reporting bad news during times when they expect that attention is limited, and conversely, reporting good news in periods of higher attention. However, given that attention is the same or higher on Fridays than other days, it is unlikely that managers are able to effectively hide bad news by reporting immediately prior to the weekend. Instead, the preponderance of strategically reporting bad news on Fridays is possibly due to managers incorrectly perceiving attention as lower on Friday. Keywords: Market participant attention, earnings release timing, strategic disclosure. 1

3 1. Introduction In this paper, we revisit a long-standing but still unresolved question: do managers attempt to hide bad earnings news by announcing during periods of low market attention? Or, conversely: do managers attempt to highlight good earnings news by announcing earnings during periods of high market attention? We posit three conditions that must be met to sufficiently answer these questions. First, to be able to hide bad news, managers must change their earnings announcement timing frequently enough that to do so would not immediately attract unwanted attention. A deviation from a long-standing pattern of earnings release timing could attract attention to the very news the manager is trying to hide. Second, there must be variation in market participants attention that is predictable to the manager ex-ante random variation in attention would not allow for strategic timing of bad or good news. Third, we must observe that managers do tend to release more negative (positive) earnings news during periods of lower (higher) market attention, after controlling for other correlated factors. The first and second of these conditions have not been directly examined. Prior studies examining the third condition have produced mixed results. Managers have several incentives to alter the timing of their negative (positive) earnings releases to periods of low (high) perceived attention. First, managers likely prefer to release good (bad) earnings news during periods where it might have the highest (lowest) price impact. Even if the initial under- or over-reaction is temporary, managers will likely strategically time their earnings release such that the news is gradually reflected in prices to reduce the likelihood of sudden price drops and trading halts that can adversely affect the company. Second, even if there is no pricing impact, managers may face reputational consequences for negative news that is publicized more, or may wish to draw attention to good earnings news to improve their 1

4 managerial reputation. Finally, to the extent that investors more frequently buy firms that catch their attention, as suggested by Barber and Odean (2008), there may be liquidity implications for differences in investor attention. Managers could alter their earnings announcement times in attempt to avoid or attract attention for any of the above reasons, even if the probability of success is low, because the direct cost of switching earnings announcement timing is likely also low. 1 We construct a novel database of precise earnings announcement dates and times to investigate the timing of earnings announcements by hour and by weekday.specifically, we use the combination of four databases, Compustat, RavenPack, Wall Street Horizon, and IBES, to create a dataset of 192,485 confirmed earnings announcement dates and 151,106 confirmed earnings announcement times. We corroborate the date and time of each earnings announcement by at least two sources. As a result, our sample is larger and likely more accurate than samples used in previous studies of strategic earnings release timing. 2 We discuss the sample in detail in Section 3. We find that firms frequently change the timing of their earnings releases: within any given year, 81.6% of firms change their quarterly earnings announcement weekday at least once, and 25.6% change whether they report before, during, or after market hours. Although just 7.6% of all earnings announcements happen on Fridays, 51.4% of firms have at least one Friday announcement during our sample period. Managers intent is unobservable, but the majority of these changes in earnings announcement times are likely made for administrative or other benign 1 For instance, the direct cost of moving a press release from a low to high attention time of day is likely near zero as neither EDGAR nor the major press wires charge different amounts depending on the time of the announcement. If an action is sufficiently cheap, a manager will undertake the action even if the expected benefit is low. 2 Most of the samples in prior research use private or hand-collected databases, so we do not make direct comparisons to particular papers. However, our sample is larger and appears to be more comprehensive than the samples in prior studies. Further, our measurements of earnings release dates and times are likely more accurate than prior research because we use several sources to validate the exact timing. 2

5 reasons. The high frequency of benign changes is precisely the camouflage needed for managers to occasionally switch their earnings announcement timing for strategic purposes without raising alarm. Thus, we interpret this evidence as consistent with the first condition that changes in earnings release timing are prevalent enough to allow for hiding bad news. We turn next to the second necessary condition, which is that lulls and peaks in market attention must be ex-ante predictable in order for managers to shift bad (good) news into times where the market is paying less (more) attention. We examine three specific times during which prior research has speculated that market attention differs: before versus after the close of regular trading hours; on Mondays through Thursdays versus Fridays; and on slow versus busy news days, based on the total number of firms that are reporting earnings. Prior research has inferred that attention is lower on Fridays (DellaVigna and Pollet 2009) and on busy reporting days (Hirschleifer et al. 2009), primarily based on evidence that price responses are muted on Fridays and busy days. We raise several concerns about drawing a causal link between attention and stock price responses. First, the interpretation of lower earnings response coefficients (ERCs) as evidence of reduced attention assumes that market frictions and/or behavioral biases prevent other undistracted arbitrageurs and algorithmic traders from eliminating under-reactions by distracted investors. Second, to the extent that ERCs do in fact capture attention, it derives primarily from equity investors; the attention of other important market participants known to improve market efficiency, such as analysts and the business press, is ignored using an ERC methodology. Finally, Kothari (2001) explains several shortcomings in ERC methodologies and discusses the possibility for correlated omitted variables. Given these concerns, we argue that additional investigation is warranted to further evaluate variation in attention for earnings announcements. 3

6 Rather than inferring variation in attention based on price responses, we employ four empirical proxies to investigate temporal variation in market participants attention: (i) the number of news articles written specifically about the firm s earnings; (ii) the speed with which analysts incorporate the earnings news into future earnings forecasts; (iii) investor downloads of 8-Ks likely containing the earnings news; and (iv) abnormal Google search volume. The advantage of these attention measures is that they are user-oriented, related to information processing, and measured on a timely basis. Our findings are consistent with attention being lower (higher) after (before) market close and on busy (slow) reporting days. Specifically, announcements after trading hours (on the busiest reporting days) are associated with a 7% (12%) decrease in news articles, a 7% (4%) decrease in the speed with which analysts update forecasts, a 19% (30%) decrease in EDGAR downloads, and a 0% (2%) decrease in Google searches, after controlling for firm fixed effects, non-stationary firm characteristics, and the sign and magnitude of the earnings news. However, our analyses generally indicate that attention is actually no different on Fridays than other weekdays, which is inconsistent with prior research that has interpreted lower Friday ERCs as being indicative of lower attention. We also tie together our first two conditions by showing that the act of switching earnings release timing (condition 1) does not draw unwanted attention to bad news (condition 2). Specifically, we find that the aforementioned variation in attention still holds for those firms that switch their earnings release timing in the current quarter as compared to firms that maintain their earnings announcement timing from the prior quarter. However, we again fail to find compelling results for differences in attention for Friday versus non-friday announcers. In sum, the data are consistent with the second condition that market attention is predictably lower after hours and on busy reporting days, but the same does not hold for Fridays. 4

7 Next, we examine the validity of the third necessary condition: that managers tend to release worse (better) earnings news after market close, on Fridays, and on busy reporting days (before market close, on Mondays through Thursdays, and on slow reporting days). Prior research has examined whether earnings news is worse after hours and on Fridays, but the results to date have been mixed. Early work by Patell and Wolfson (1982) and Damodaran (1989) find that earnings news tends to be worse on Fridays and after hours, but Doyle and Magilke (2009) find that such differences disappear when the empirical methodology is improved to examine only within-firm variation. Using a larger sample of precise earnings release times, we find statistically and economically significant declines in unexpected earnings (i.e., IBES earnings less analyst consensus) for firms that choose to report earnings after hours, on Fridays, and on the busiest reporting days. In analyses that tie the first condition with the third, we also find evidence consistent with managers both switching to periods of low expected attention to hide bad news, as well as with switching to periods of higher expected attention to highlight good news. We make several contributions to the academic literature. First, although it can be inferred from prior research that the existence of earnings announcement timing changes is nonzero, to date there has been limited evidence on the prevalence, timing, and dimensions of the changes in earnings release timing. We contribute to the literature by using robust data to quantify how often changes occur in general (e.g., across all firms, 15% change their before/during/after hours announcement timing from the previous quarter) as well as how often changes occur within a given firm (e.g., within each firm, 25.6% change their before/during/after hours announcement timing at least once within a given year, and 72.6% of firms have at least one change within our sample period). Further, we characterize changes in quarter-over-quarter 5

8 earnings announcement timing along several dimensions: by hour, by weekday, and in relation to the timing choices of other firms. We also provide evidence consistent with changes in earnings announcement timing not drawing the attention of market participants, an assumed condition that underlies prior research on strategic earnings timing, but which has not been explicitly tested to date. Our second contribution is to provide new and direct evidence on the existence of predictable variation in attention to earnings announcements. We employ state-of-the-art proxies for market participants attention to investigate how and when attention to earnings releases appears to decline, and find evidence that both confirms and conflicts with the inferences from prior research. First, our evidence that market attention is lower after hours is new to the literature. Second, our evidence that market attention is lower on busy reporting days confirms Hirshleifer et al. s (2009) inference based on evidence from stock price responses. Finally, our evidence that attention to earnings announcements is no different on Fridays as compared to Monday Thursdays conflicts with DellaVigna and Pollet (2009). Finally, we contribute to the literature that investigates the determinants of managers earnings release timing choices. Prior studies have produced mixed evidence as to whether managers strategically time earnings releases based on the sign and magnitude of the earnings news. Using a dataset that is larger and likely more precise than the data in previous studies, and using a comprehensive set of empirical tests, we find that earnings news does tend to be worse when released after hours, on Fridays, and on busy reporting days. We also predict and find evidence that managers appear to highlight good news during periods of high perceived attention, a finding that is new to this literature. 6

9 In sum, we conclude that managers are likely able to limit attention to negative news by announcing earnings after hours and on busy days. However, since we find that attention is no different on Friday than other weekdays, we conclude that the preponderance of bad news on Fridays is driven by either managers having incorrect beliefs about market attention, or by some other motivation. Section 2 discusses prior literature and develops our research hypotheses. We discuss our sample in Section 3 and detail our empirical analysis in Section 4. Section 5 concludes. 2. Prior Research & Hypothesis Development 2.1 Hiding bad news and highlighting good news In today s information-rich environment, the idea that public companies can hide an earnings release is potentially difficult to grasp. At the same time, though, limited attention theory predicts that the enormous amount of information available in recent years likely makes it easier for managers to hide bad news among the clutter (Hirshleifer et al 2009; Lim and Teoh 2010). Moreover, the cost of changing earnings release timing is likely very low, so a manager will likely do so even if s/he expects that it has only a small chance of being effective at altering attention. In this section, we provide several reasons why managers likely would want to limit (attract) attention to bad (good) earnings news, and why managers might have some expectation of getting away with it. One incentive for managers to reduce (attract) attention to bad (good) news is that doing so might influence market price responses, at least in the short term. Anecdotal evidence indicates that inattention is associated with prices that do not fully impound all available public information. For instance, Huberman and Regev (2001) examine the case of the biotechnology company EntreMed: a May 1998 article in New York Times caused EntreMed s stock price to 7

10 rise over 330% in one day and remain high through the rest of the year, despite the fact that this same good news was reported in Nature, New York Times, and other sources more than five months earlier. Lim and Teoh (2010) speculate that many so-called market anomalies are potentially attributable to variation in market participants attentiveness. Moreover, even if managers understand that bad news will soon be fully revealed in prices, they likely prefer gradual rather than sudden price declines because the latter can lead to negative media attention, price drops, and trading halts. Further, prior research shows that severe price drops increase the likelihood of shareholder litigation (Donelson et al. 2012). Even if there is no pricing benefit of strategic earnings release timing, other potential benefits are also possible. First, Farrell and Whidbee (2002) find that forced CEO turnover after poor performance is higher among firms with greater media coverage, consistent with news dissemination affecting labor outcomes for managers. Thus, managers likely want to limit the dissemination of bad news so as to protect their reputations and career prospects. Equivalently, managers likely want to maximize dissemination of good news to boost their reputations. Second, Barber and Odean (2008) discuss how some investors only consider purchasing stocks that have first caught their attention. A logical extension of this argument is that investors will be less (more) likely to purchase stocks about which they have heard negative (good) news. Even if prices are unaffected by individual traders, there is plausibly a liquidity implication of news dissemination. As previously noted, the idea that managers limit exposure to bad news by reporting during periods of lower market attention involves three conditions: (i) managers must change their earnings announcement timing frequently enough that to do so does not immediately attract unwanted attention; (ii) there must be ex ante predictable variation in market participants 8

11 attention; and (iii) we must observe that earnings news is more negative during periods of expected lower market attention, after controlling for other correlated factors. The second and third of these conditions apply equivalently to managers being able to highlight reporting good news in periods of high attention. In the following three subsections, we develop hypotheses regarding each of these conditions Prevalence of Earnings Announcement Timing Changes Microsoft schedules its earnings release to occur at the same time and weekday each quarter Thursday afternoon just after 4pm Eastern Time (all quoted times hereafter are in U.S. Eastern Time). Costco presents another interesting case in earnings release timing it releases earnings consistently at 1:00am. If announcement times are in general as highly persistent as they are at Microsoft and Costco, a manager s attempt to limit exposure to negative news by switching the announcement timing will likely have the opposite result: the act of changing timing will draw unwanted attention. Thus, we argue that changes in earnings announcement timing must be sufficiently prevalent that a switch does not immediately stand out as anomalous. There are a number of reasons to expect that earnings announcement times are sticky. First, managers have a preference for being consistent with precedent in their voluntary disclosures (Graham et al. 2005). This preference is most likely driven by an expectation that changing the announcement timing between periods will be interpreted as signaling something about the firm s earnings news. Several research firms make efforts to forecast in advance the predicted date of earnings release, and there is evidence that markets respond negatively when firms miss an expected reporting date (Chambers and Penman 1984, Bagnoli et al. 2002, Duarte- 9

12 Silva et al. 2013). 3 This finding suggests that investors form expectations not only for the earnings news, but also for the timing of the earnings news release. Thus, it is probable that managers prefer to maintain consistency in their reporting to avoid returns volatility from varying announcement timing. Indeed, Graham et al. (2005) find that 53% of surveyed executives give no preferential treatment to disclosing good or bad news faster (page 63). There are also a number of reasons to expect that earnings announcement times are highly variable. Academic research has also found evidence consistent with managers both accelerating and delaying earnings announcement timing for strategic purposes (Skinner 1994, 1997; Donelson et al. 2012; Kross 1981, Chambers and Penman 1984; Kross and Schroeder 1984; Begley and Fischer 1998; Bagnoli et al. 2002). However, the most common reasons for changing the timing of earnings releases are likely benign, simply reflecting calendar conflicts or administrative conflicts. For example, a firm that regularly releases earnings on a Monday may have to change the day for common holidays that fall on Monday. Our discussions with investor relations experts also revealed the administrative conflicts, such as executive and analyst availability, are common reasons for changing the earnings release timing. As mentioned above, frequent and benign changes in earnings release timing are precisely the camouflage needed for managers to change the earnings release timing for strategic reasons. We can infer from prior research that changes in earnings announcement timing do indeed occur (e.g., Doyle and Magilke 2009). However, prior research does not quantify the extent and dimensions of changes in earnings release timing. For most of the past studies, examining the extent of earnings release timing changes was not the primary goal of the paper, and as a result, the studies provide very little context surrounding the frequency of switches. We 3 For example, Zacks Investment Research forecasts the expected report date, which is released in advance to the public. Wall Street Horizon sells proprietary data on expected earnings announcement times. 10

13 put forth three predictions about consistency in earnings announcement timing across periods where variation in market attention is thought to exist: H1a: Quarter after quarter, firms release earnings news at the same time of day. H1b: Quarter after quarter, firms release earnings news on the same weekday. H1c: Quarter after quarter, firms release earnings news on similarly busy reporting days Variation in Market Participants Attention Prior research has posited that, at some point, the limited capabilities of humans to acquire and process information prevents them from absorbing the complete set of public information (Lim and Teoh 2010). A large and expanding literature has begun to explore the determinants and implications of what is often referred to as market inattention or distraction (see Lim and Teoh (2010) for a review). Patell and Wolfson (1982), among others, speculate that earnings releases that occur in the evening might receive less attention than similar news that is released earlier in the day. The premise for this speculation is presumably that market participants are not working once the market closes, or that they are distracted by other issues. However, Patell and Wolfson (1982) also note that the opposite could be true: that earnings releases in the evening could receive greater attention as more time is allowed for dissemination and interpretation before trading begins the following day. To our knowledge, no previous paper has explicitly examined differential attention to earnings released after hours versus earlier in the day. We therefore investigate the following hypothesis (written in the null form): H2a: Market participant attention to earnings news is the same after market hours as during or before market hours. 11

14 DellaVigna and Pollett (2009) and Damodaran (1989), among others, theorize that market participants are distracted just before the weekend and, as a result, earnings released on a Friday receive less attention than similar earnings released on Monday through Thursday. However, because relatively few firms release earnings on Friday (in our sample, just 7.6 percent occur on a Friday), it is plausible that each individual announcement receives more attention than announcements on other days of the week. Further, the idea that market participants are systematically inattentive on Fridays is inconsistent with the work patterns for the typical investment banker, money manager, or analyst, who often work long hours and weekends. Consistent with this idea, in untabulated analysis we find no difference in the frequency of analyst forecast revisions that occur on Fridays versus other weekdays. Thus, there are reasons both for and against expecting lower market attention immediately prior to the weekend. DellaVigna and Pollet (2009) find that ERCs are lower, PEAD is higher, and trading volume is lower on Fridays, which is consistent with lower attention. However, as discussed above, inferring a causal link between lower attention and muted stock price responses requires controversial assumptions about the earnings-returns relation and also that the numerous other known determinants of ERCs are adequately controlled in empirical analysis. 4 Further, since the majority of trading activity is driven by computer algorithms that are likely unaffected by distraction on Fridays, there are reasons to question whether lower Friday ERCs are driven by behavioral biases. Thus, it is plausible that the differential market responses for Friday announcements observed by DellaVigna and Pollet (2009) are caused by other factors, and we argue that further investigation of attention on Fridays is warranted. 5 In our primary tests, we use 4 The literature on the determinants of ERCs is vast. Kothari (2001) discusses the literature on how expectations about growth, discount rates, earnings persistence, and a number of other firm fundamentals impact ERCs. 5 Melessa (2012) also finds evidence that lower Friday ERCs are likely not driven by lower attention, but rather by macroeconomic uncertainty. 12

15 alternative, user-based measures of market participant attention to investigate H2b (written in the null form): H2b: Market participant attention to earnings news is the same on Friday as it is Monday through Thursday. Similar to DellaVigna and Pollett (2009), Hirshleifer et al. (2009) interpret lower stock market responses and greater PEAD on days with numerous earnings announcements as indicating that attention to individual firm announcements is lower on days with high information flow. A plausible alternate is that, since earnings season is highly anticipated and report dates are often known in advance, market participants arrange their schedules such that they can pay adequate attention to numerous firms announcements. It is therefore plausible that there is equal or even greater attention paid to earnings releases on busy days relative to periods of low information flow, in which case the differential market responses on busy days are driven by non-attention causes. As with above, we use alternative measures of market inattention to investigate H2c (written in the null form): H2c: Investor attention to earnings news is the same on days with many earnings releases as it is on days with few earnings releases Strategic Timing of Earnings News Several studies have examined whether managers choose an earnings release time depending on the content of the news. Patell and Wolfson (1982) use a small sample of earnings releases for 96 firms during the late seventies and find that period-over-period earnings changes as well as stock returns tend to be lower after trading hours. Damodaran (1989) finds similar results for Fridays earnings announcements on Fridays are more likely to contain bad news than for other weekdays. Doyle and Magilke (2009) use a slightly larger sample of earnings 13

16 announcement changes (ranging from about 1,300 to 2,000 observations in their primary analyses) in a more recent time period ( ) to re-assess the question of whether managers report bad news after hours and on Fridays. The primary innovation in Doyle and Magilke (2009) over the previous research is to examine strictly within-firm changes in earnings release patterns. In contrast to prior papers, the study finds little evidence that unexpected earnings (as measured relative to analyst consensus) are lower after hours or on Fridays. There are several potential reasons for the mixed prior results on whether managers report bad news after hours and on Fridays. First, times have changed the information environment is clearly different in the internet age than in prior eras, which implies that the manner of information dissemination has also changed. Indeed, we find that the timing of earnings releases has changed dramatically over the past decade now, almost no firms announce earnings during market hours (see Figure 1, Panel B). Second, the studies use differing research designs and sample sizes. Finally, obtaining data on the exact timing of earnings releases has been difficult we find that up to 10% of earnings announcement dates and up to 22% of earnings announcement times are mismatched during our sample period. Earnings release timing is voluntary. Thus, if managers prefer to hide (highlight) bad (good) earnings news, they likely alter their earnings announcement timing in an attempt to take advantage of any possible differences in market participants attention. Specifically, we predict that managers will time bad (good) news for release in periods when they expect attention to be lower (higher). We also note that, if the cost of changing announcement timing is sufficiently low, managers will likely do so even if the expected probability of receiving different attention is low. Our hypotheses in the null form are as follows: H3a: Earnings news is no different after market hours than it is earlier in the day. 14

17 H3b: Earnings news is no different on Friday than it is on Monday through Thursday. H3c: Earnings news is no different on days with many earnings releases than it is on days with few earnings releases. 3. Sample Construction Because the mixed results in the prior literature could be partially attributed to challenges in measuring earnings release timing, our objective in creating our sample is to construct a sample of earnings announcement days and times that is as accurate as possible. To do so, we obtain earnings announcement dates and times from four independent sources Compustat, IBES, RavenPack and Wall Street Horizon and retain only those observations for which the earnings announcement days and times can be validated by at least two sources. Information on the sample selection is shown in Table 1. We create separate samples for earnings announcement days and for earnings announcement times in order to maximize the sample sizes. To create our sample of earnings announcement days, we begin by intersecting Compustat and CRSP to identify all US public company quarterly earnings releases from 2000 through 2011, for a total of 233,827 observations. 6 We remove 19,166 observations for which IBES data are unavailable. To better ensure that the earnings announcement dates are accurate, we eliminate another 22,106 observations for which the Compustat and IBES earnings announcement dates differ. We drop 70 observations (less than 0.04% of the sample) that have an earnings announcement on a Saturday or Sunday. The remaining sample of 192,485 firmquarters is used for our descriptive analysis of earnings announcements by day. For our statistical tests, we drop 47,151 observations that do not have the control variables used in the regression 6 We limit to our analyses to common stocks (i.e., CRSP shrcd 10 and 11), which removes non-typical securities such as REITs, ADRs, and publicly-traded partnerships. We begin our sample period in 2000 as news data from RavenPack are not available prior to

18 tests below. 7 To reduce the effects of anomalous observations, we drop 3,652 observations that have an absolute earnings surprise that exceeds the stock price, that have a stock price below $1 at period-end, or that announce earnings more than 90 days after the fiscal period-end (which is the longest SEC reporting deadline during our sample period). Finally, we drop 17,677 observations that do not have news coverage data. The final sample for statistical tests of earnings announcements by date is 124,005 firm-quarters. To create our sample for earnings announcement times, we further require accurate data on the time of day of the earnings release, which leads to a different sample size from the samples described above. We use three different data sources to validate earnings announcement times. Our first source of the announcement time stamp is IBES. 8 We require that the earnings release time is validated by at least one of two additional sources, RavenPack and Wall Street Horizon. RavenPack s database comprises news articles that appear in Wall Street Journal (all editions), Dow Jones Newswires, and Barron s. For each earnings announcement date, we locate in RavenPack the firm s press release or the first news article written specifically about the firm s earnings. 9 The time stamp of the press release or first news article is our second source of the earnings announcement time. Our third source of earnings announcement times is from Wall Street Horizon, although these data are only available after For our descriptive analysis of earnings announcement times, we require that the before/during/after-hours timing classification can be corroborated by at least two data sources. This criteria eliminates 41,379 of the 192,485 observations with a valid earnings announcement 7 Specifically, we drop observations with missing data for UE, SIZE, BTM, LEV, NUMEST, CAR, FQ4, and REPLAG, all of which are defined in Appendix A and discussed below. 8 IBES has an usual spike of earnings announcement times of 00:00:00. We set these likely errors to missing. 9 Specifically, we retain only earnings press releases and news articles that RavenPack assigns a relevance score of 100, which indicates that the article is dedicated to only that particular firm. 10 Wall Street Horizon reports an accuracy rate of 99.95%, as per Accessed 8/17/

19 date, leaving a sample of 151,106 observations for our descriptive analysis of intraday announcements. As discussed below, our statistical analysis primarily focuses on a binary classification of before versus after the close of trading hours (as opposed to a ternary classification of before/during/after-hours). Thus, for our statistical tests we only require that before versus after market close timing classification can be corroborated by at least two data sources. Of the 124,005 observations with sufficient data for the statistical tests, 120,345 have corroborated timing information. Although we cannot observe the actual error rate among these dropped observations, a paper by Dong et al. (2011) finds that during , 87% of IBES time stamps are late compared to the actual press release time, with the average delay being 105 minutes. Given that 34.6% of our sample firms report earnings immediately prior to market opening, retaining these observations would likely classify many before-hours announcements as being during-hours announcements. 4. Empirical Design and Results All of the variables discussed below are defined in Appendix A. All continuous variables are winsorized at 1% and 99%. Sample sizes vary across tests depending on data availability Analysis of Hypotheses 1 Changes in Earnings Announcement Timing The timing of each earnings announcement is identified along three dimensions. The first is time of day, measured consistently based on U.S. Eastern Time. Regular U.S. trading hours are identified as 9:30AM to 4:00PM. The second dimension is the day of the week. The third dimension is how busy the earnings announcement day is. Consistent with Hirschleifer et al. (2009), we proxy for how busy a reporting day is by taking the simple count of all earnings 17

20 announcement per day on Compustat, sorted into deciles by calendar year. We label this variable EAFREQ. 11 Panel A of Figure 1 presents summary information on the intra-day distribution of earnings announcements. 34.6% of announcements happen immediately before U.S. market trading hours, while 45.4% happen immediately after trading hours. Just 7% of announcements happen during trading hours. Panel B shows that these percentages change over time the proportion of firms announcing during trading hours decreases monotonically from 21.2% in 2000 to just 3.2% in Panel C shows considerable variation in the fraction of earnings announcements that happen each weekday. The frequency of earnings announcements per day increases monotonically from 14.4% on Mondays to 30.5% on Thursdays, followed by a sharp decline to just 7.6% on Fridays. Hypothesis H1a, H1b, and H1c predict that firms tend to announce their earnings at the same time of day, same weekday, and similarly busy days each period. Panels A, B, and C of Table 2 present quarter-over-quarter transition matrices of announcement times. Between any two consecutive quarters: 15% of firms change their before/during/after trading hours timing; 52.8% of firms change their reporting weekday; and 80.5% change their EAFREQ decile. Panel D of Table 2 presents frequencies of announcement timing changes within each fiscal year. Within any given year: 25.6% of firms have at least one change in before/during/after market timing; 81.6% of firms have at least one change in announcement weekday; and 98.8% of firms have at least one change in EAFREQ decile. Untabulated analysis shows that, over the entire 11 A concern is that managers are not able to predict busy earnings days ex ante, in which case it would be impossible to strategically time earnings news based on EAFREQ. In untabulated tests, we find that a regression of EAFREQ on binary variables for the weekday, calendar week, and calendar day explains about 90% of the variation in EAFREQ. Further, out-of-sample tests predict EAFREQ to within two deciles of the actual EAFREQ for 98.7% of observations. We interpret these results as being consistent with managers being able to predict busy versus slow earnings reporting days. 18

21 sample period, 72.6% of our sample firms have at least one change in before/during/after hours timing and nearly 100% of firms have at least one change in EAFREQ decile. Further, 51.4% of firms have at least one Friday earnings announcement during our sample period, which is unexpected given that only 7.6% of all earnings announcements happen on a Friday. In summary, the results in Table 2 indicate that changes in earnings announcement timing happen frequently, to the extent that one could conclude that strategic changes in the earnings release pattern do not draw the attention of market participants (statistical tests linking changes in announcement timing and attention are provided below). We interpret these data as being inconsistent with the null in H1a, H1b, and H1c Analysis of Hypotheses 2 Differences in Market Attention H2 predicts that market participants exhibit predictable variation in attention. As discussed in Da et al. (2011), despite significant academic interest in the concept of market attention, it is a complicated and multifaceted construct that is difficult to empirically measure. We define attention simply as a user taking notice of a piece of information, and we employ four measures of market participants attention (collectively ATTN) to evaluate this difficult-toobserve construct, each of which has been vetted in prior research (e.g., Barber and Odean 2008; Da et al. 2011; Drake et al 2012a,b; Drake et al 2013; Zhang 2008). 12 Our ATTN proxies have several strengths. First, each of the four measures has the advantage of being focused on the user, in that they measure the actions and/or choices of information users if a user is taking action with respect to an earnings release, we assume that he/she is paying attention to it. Second, each measure is related to either the production of information (as in the case of news articles and analyst revisions) or the acquisition of information (as in the case of 8-K downloads and Google 12 We note that our four proxies are not necessarily intended to be independent of one another, but are rather four related proxies for the same underlying construct of market participants attention. 19

22 searches), which allows us to speak to what information market participants pay attention to. Third, these variables are measured on a timely basis, which makes them superior to other less timely measures of attention, such as quarterly advertising expense (e.g., Chemmanur and Yan 2009). Fourth and finally, unlike measures such as stock returns or trading volume, our ATTN variables do not rely assumptions of market equilibrium an assumption that is debatable in tests of behavioral biases such as limited attention. In the following subsections, we first discuss each of our ATTN proxies for market before discussing our empirical tests of H2a, H2b, and H2c. Sample sizes for each proxy vary based on data availability Attention Proxy: News Articles The first proxy for attention is the number of news articles written specifically about a firm s earnings announcement. Barber and Odean (2008) and Li et al. (2011) both find evidence consistent with market activity being a function of news coverage. The underlying theory is that, ceteris paribus, the number of published articles correlates with the breadth of earnings news dissemination (i.e., fewer articles mean that fewer readers learn about the earnings news) as well as the depth of news dissemination (i.e., even those market participants who learn about the earnings news will acquire less information when fewer news articles are available). Our news count measure, NEWSCOUNT, is the natural log of the number of earnings-specific news articles that appear in the RavenPack database within the 24-hour period starting with the firm s earnings release. NEWSCOUNT is available for years , Attention Proxy: Analyst Updating Speed 13 Firm-quarters with zero news coverage are eliminated from all statistical tests because, as discussed in Section 3, we use the publication time of first news story to help identify the time of the earnings release. 14 We also consider subtracting from NEWSCOUNT the average number of earnings-specific articles that appear over the previous 7 weeks, but the average is zero for nearly all firm-quarters. Untabulated analysis using such a measure of abnormal NEWSCOUNT produces qualitatively and quantitatively similar results. 20

23 Our second proxy for market attention is the speed with which equity analysts impound earnings news into their future forecasts, in the spirit of Zhang (2008). Analysts typically update their future forecasts immediately after an earnings release to incorporate the most recent financial results. Our intuition underlying this proxy is that greater analyst attention is associated with faster forecast updates. During times when analysts are distracted, we assume that it will take them longer to update their future forecasts. We use the IBES detail file to collect data on all the analyst forecasts, j, that are updated within 30 days of the firm s earnings announcement. These data are available for all years in our sample. We then calculate the number of weekdays between the earnings announcement and forecast update. The inverse of the average of these lags measures how quickly analysts update their forecasts following the earnings announcement, which is our proxy for analyst attention. Specifically, our measure of analyst speed, ANALYST_SPD, is calculated as follows: 15 "#$"_"# = 1 log 1 + ""#$%&'"#$%"#$%&'("#$%&. (1) Attention Proxy: Edgar 8-K Downloads The third measure of market attention is based on the number of 8-K downloads from EDGAR. These data represent the activity of EDGAR users to acquire the 8-K, which is the regulatory filing that accompanies an earnings press release and other significant events. 16 We measure investor EDGAR activity on earnings release days relative to that on other days. Thus, 15 As ANALYST_SPD is based on daily data, this metric is potentially biased upwards for earnings announcements that occur after market hours relative to those that occur before or during market hours. We adjust for this bias by subtracting 1/3 rd of a day (8 hours) from ANALYST_SPD for all after-hours earnings announcements. 8 hours seems to be a reasonable adjustment as the vast majority of earnings announcements happen either just before or just after market hours, or roughly 8 hours apart. Results using no adjustment or a 12-hour adjustment produce similar results. 16 For brevity, we refer the reader to Drake et al. (2013) for a detailed description of the data, including the sample criteria, sample period and analyses on the timing and common determinants of user acquisition of EDGAR data. 21

24 this measure is a proxy for attention to firm-specific earnings news. We compute abnormal EDGAR downloads as follows: "#$% = log "#$% log "#$% (2) The first term is the natural log of the sum of EDGAR 8-K downloads for two days around the earnings announcement (days 0, 1), while the second term is the natural log of the trailing average EDGAR 8-K downloads for the same two weekdays over the preceding seven weeks. EDGAR data are available for the period Attention Proxy: Google Search Volume Our final measure of market attention is Google search volume that occurs around the earnings announcement for a given firm s stock ticker, following Da et al. (2011) and Drake et al. (2012). The intuition underlying this proxy is that as a market participant searches the internet for firm-specific information, they are paying attention to the stock. Data on weekly Google search volume for firms stock tickers is obtained from Google Trends, following the methodology in Drake et al. (2012). Google Trends reports a normalized measure of the number of searches for a given search term in the Google search engine. We calculate abnormal Google search volume as follows: ""#$%& = log "#$ log "#$ (3) where w is the week of the earnings announcement. 17 We use a two-week event period as data are on a weekly basis and using a one-week period would not capture any post-announcement attention from Friday announcers (e.g., if earnings announcement attention to a Friday 17 For brevity, we refer readers to Drake et al. (2012) for a detailed description the Google Trends index. Note that we follow Da et al. (2011) in using weekly Google search data in order to maximize the sample size. 22

25 announcement extends through the following Monday, a one-week event window would ignore the Monday attention). Google search data are available for the years Tests of H2 Panel A of Table 3 provides summary statistics for our ATTN proxies. Panel B shows statistically significant but qualitatively low correlations between our ATTN variables. The low correlations are similar to those found by Da et al. (2011) and are likely due to the proxies capturing different aspects of market attention. NEWSCOUNT likely most directly correlates with media attention. ANALYST_SPD is likely a direct measure of analyst attention, and Da et al. (2011) find that GOOGHITS is likely most directly related to retail investor attention. EDGAR likely captures earnings-related attention from multiple types of market participants Regression Tests of H2 To empirically test for predictable differences in market participants attention, we employ regression analysis that allows us to simultaneously consider all three earnings announcement timing choices, AFTER, FRIDAY, and EAFREQ: 19 ATTN = β 0 + β 1 AFTER + β 2 FRIDAY + β 3 EAFREQ +Σβ k CONTROLS + Σβ k YEAR + Σβ k FIRM + ε. (4) AFTER is a binary variable equal to one for earnings announcements from 4pm to midnight; all other times are coded as zero. FRIDAY is a binary variable equal to one for earnings announcements on Friday. As discussed above, EAFREQ is the decile of the number of market- 18 Our tests are agnostic as to whether variation in attention can alter stock prices responses (i.e., ERCs). Still, untabulated analysis we estimate a standard ERC model of two-day abnormal returns regressed on ranked unexpected earnings and an interaction between unexpected earnings and ATTN, along with typical control variables. We find that all four ATTN interaction variables are significantly positive, which is consistent with higher ATTN leading to larger ERCs. 19 In untabulated analysis we perform separate, univariate tests of within-firm changes in ATTN between quarters when there is a change AFTER, FRIDAY, or EAFREQ. Results are generally consistent with the regression results discussed herein. 23