Friday Distraction in Financial Markets: Evidence from Earnings Announcements And Bid-Ask Spreads

Transcription

1 Friday Distraction in Financial Markets: Evidence from Earnings Announcements And Bid-Ask Spreads Jonathan Cohen * Northwestern Department of Economics May 2015 ABSTRACT Do Friday-related distractions affect financial markets? I demonstrate the documented tendency of long-term stock price drift following Friday earnings announcements relative to those on other weekdays is due to investor inattention, as DellaVigna and Pollet (2009) suggest, rather than firm selection, as Michaely et al. (2015) suggest, by finding an attenuation of this tendency following the publishing of DellaVigna and Pollet s paper even after controlling for firm characteristics. I then examine the documented tendency of bid-ask spreads to widen on Fridays as a potential effect of inattention. I exploit an exogenous market structure change to identify the causal effect of market maker Friday inattention on bid-ask spreads and find no evidence in support of this channel. * I am greatly indebted to Professor Benjamin Iverson for advising and supporting me throughout the research process; Professors Terrence Hendershott of the Haas School of Business, University of California and Pamela Moulton of the School of Hotel Administration, Cornell University for publicly providing the NYSE Hybrid activation data; Professor William Rogerson, Sarah Muir Ferrer, and the rest of the faculty and students in the MMSS program for providing a superb intellectual atmosphere; and, most importantly, my family for making possible my academic career. 1

2 2 DONNA: What's Take Out the Trash Day'? JOSH: Friday. DONNA: I mean what is it? JOSH: Any stories we have to give the press that we're not wild about we give all in a lump on Friday. -The West Wing Take Out The Trash Day Season 1 Episode 13 * *

3 3 1. Introduction According to the weak and semi-strong forms of the efficient markets hypothesis, security prices reflect all past prices and publicly available information, respectively (Fama). At the same time, there is a large body of literature detailing predictable risk-adjusted returns based both on past price movements (Jegadeesh and Titman (1993) and publicly available corporate announcements (Dichev and Piotroski (2001), Hirshleifer and Teoh (2003)). In practice, there is a vast amount of information released in financial markets each day. In distinguishing the signal and the noise, investors must selectively direct limited cognitive resources towards information processing (Kahneman and Tversky (1981)). Since the psychology literature suggests that distraction can affect both short-term encoding and long-term recall, it is plausible that information stored during high distraction periods may be overlooked for a long portion of time (Craik (2014)). Investors may even believe they cannot profitably act on the information once they incorporate it with delay. And even though automated trading now comprises more than three-fourths of NYSE and Nasdaq volume, the algorithms are often programmed to trade according to other orders, namely those of human investors subject to attention constraints (MacGregor (2010)). Documented sources of inattention s impact on asset prices include firm size (Peng (2005)), media strikes (Peress (2014)), unrelated earnings announcements (Hirshleifer et al. (2009)), and timing of earnings announcement (Francis et al. (1992)). Along these lines, DellaVigna and Pollet (2009; hereafter DP) focus on the effect of investor distraction on Fridays due to the upcoming weekend and find less immediate price response and more drift following Friday earnings announcements. Michaely et al. (2015; hereafter MRV), reproduce the reduced immediate response to a broader range of corporate

4 4 announcements, but, contrary to DP, attribute the differential response to firm heterogeneity and selection bias rather than investor inattention due to Fridays. As of now, the question of whether there is an effect of Friday inattention on financial markets is unresolved. This paper reconciles the DP and MRV findings, provides further evidence of the Friday inattention effect, and proceeds as follows. Section 2 outlines the theory and existing literature surrounding return predictability following Friday earnings announcements. There is mixed evidence whether firms strategically release announcements during low attention periods, which are usually characterized as after trading hours or on Fridays, and whether investors react differently to these announcements. Several papers find no effect when comparing announcements for a given firm on Friday to those on another weekday. Section 3 describes the data and methodology used for testing the existence of Friday earnings announcement effect due to inattention. I primarily follow the procedures employed by DP, though there are some issues due to data revisions and potential measurement error. I reproduce the DP results in order to justify the validity of my data. I then exploit the publishing of DP as an exogenous shock to investor attention on Fridays to test for a change in the Friday earnings announcement anomaly. Section 4 demonstrates a change in the Friday earnings announcement effect following the publishing of DP. These results contribute to the Friday earnings announcement literature by connecting conflicting evidence based on looking within firms or across firms with respect to price response following Friday earnings announcements. This result is broadly related to McLean and Pontiff (2015), which finds that cross-sectional return predictability documented in peer-reviewed academic journals significantly decreased after the study s publication. As Cochrane (1999) points out, such an exercise is a joint test of statistical biases, unobserved

5 5 systematic risk factors, and the extent to which investors learn from the publication. The third point, regarding investors ability to pay attention to results in an academic paper, is particularly relevant to the focus of this paper because investor inattention is one of the main explanations of Friday post-earnings drift. In order to test the Friday inattention effect outside the scope of strategic corporate announcements, I turn to the effect of Friday market maker inattention on bid-ask spreads. Section 5 examines the theory and existing literature surrounding interday differences in volume and bid-ask spreads. Lower volume on Fridays is well-documented, though the empirical relation between daily volume and bid-ask spreads is not conclusive. This leaves open the possibility of identifying differences in bid-ask spreads on Fridays separately from differences in volume on Fridays. Section 6 provides a brief background on the NYSE Hybrid implementation, a change in NYSE market structure that the existing literature suggests alleviated market maker attention constraints. Section 7 describes the data and methodology used to test for a Friday inattention effect through higher bid-ask spreads. Because actual trade and quote data is increasingly unwieldy, I employ a recently developed approach to estimate intraday bid-ask spreads using daily data. I exploit the NYSE Hybrid implementation in order to test whether the difference in Friday bidask spreads relative to spreads of NYSE stocks compared to similar Nasdaq stocks changes following a relaxation of market maker attention constraints. Section 8 demonstrates that while there are higher bid-ask spreads on Fridays and an attenuation following the Hybrid implementation, this trend mirrors that of similar Nasdaq stocks that are relatively unaffected by Hybrid. To my knowledge, no section of the literature has tried

6 6 to identify the relative importance between investors and market makers of a potential Friday inattention effect. I propose a test for future research that exploits an exogenous market structure change in order to identify an effect of Friday market maker inattention on bid-ask spreads. Section 9 concludes.

7 7 2. Friday Earnings Announcements Ball and Brown (1968) first provided evidence of delayed investor reaction to earnings announcements, which is often referred to as the post-earnings drift anomaly, and Jones and Litzenberger (1970); Rendleman, Jones, and Latané (1982); and Bernard and Thomas (1990) subsequently argue this can be leveraged into a profitable risk-adjusted investment strategy. One such strategy would be to construct a zero cost portfolio by financing the purchase of stocks about to release earnings announcements by shorting stocks in between earnings announcements. Chordia and Shivakumar (2006) find the systematic price momentum factor from Carhart s (1997) four-factor pricing model is largely captured by the systematic component of such a portfolio. In terms of practical implementation, Mendenhall (2004) and Ng, Rusticus, and Verdi (2008) argue transaction costs associated with the necessarily high turnover of such a portfolio remove any risk-adjusted returns. More recent studies have suggested the post-earnings drift anomaly might be attributable to information quality ((Francis et al. (2005), Vega (2006)), opinion divergence (Garkinkel and Sokobin (2006)), and inflation risk (Chordia and Shivakumar (2005)). Finally, Sadka (2006) demonstrates momentum as characterized by the time-varying component of firm-level liquidity risk explains a large part of the anomaly. In line with Chordia and Shivakumar, Sadka finds this liquidity factor explains a large portion of risk-adjusted returns to momentum in Fama and French s (1992) three-factor pricing model. The Friday post-earnings drift literature investigates whether there is an additional abnormal response of stock prices following Friday earnings announcements relative to earnings announcements on other days. The Friday News Dump, or propensity for government to release bad news on Friday evenings, has been a well-accepted practice in political circles for

8 8 several decades 1. The conventional wisdom is that officials seek to minimize the impact of unflattering information by disclosing it late on Friday evenings while highlighting the impact of flattering information by disclosing during the day on an earlier weekday. Similarly, strategic managers decide when to release the firm s quarterly earnings announcement and may try to do so during low attention periods. For example, Chen and Mohen (1994) find that 50% of firms vary announcement timing, and the variation is most often due to unexpected earnings level. As their name indicates, investors are invested in the outcomes of firms earnings announcement, probably more so than constituents in the affairs of their political representatives. Thus it is unclear if firms could benefit from the same kind of strategic timing. Patell and Wolfson (1982) note the general negativity of news released after the market closes as evidence of intraday strategic release, while Penman (1987) and Damodaran (1989) note the general negativity of news released later in the following fiscal quarter and on Mondays or Fridays as evidence of interday strategic release. This strategic tendency could be motivated by either opportunistic information concealing in the style of the aforementioned Friday News Dump or by a more benign desire to allow for the distribution of information. In the former, the potential benefits of strategic release timing are due to inattention constraints. In the latter, the potential benefits are due to information dissemination constraints. Bagnoli et al. (2005) exploit information diffusion innovations of the 1990 s for example, the advent of the Internet and 24/7 news coverage to distinguish the relative importance of each factor. They find evidence suggesting investors are now able to access and process earnings announcements made after trading hours just as well as those during trading hours. Additionally, while they find muted immediate reaction to negative Friday 1 The website and Twitter are devoted to the topic, and the White House, Pentagon, and NSA have all tended to set the release unflattering documents to late Friday afternoon (Clarke)

9 9 earnings releases, they show the negative price response occurs in the days preceding the announcement. This is consistent with firms releasing negative information on Fridays to guide investor expectations rather than exploit inattention. Recent studies have sought to use more novel measures as proxies for investor demand for public information around earnings announcements. Drake et al. (2012) use Google search data and show the price response of earnings releases is preempted for announcements with high information demand. DeHaan et al. (2015) confirm lower attention after trading hours by looking at earnings-related news articles, analyst revisions of future earnings forecasts, EDGAR 8-K downloads, and Google search volume. While these measures confirm inattention is higher outside of trading hours, they indicate no difference in attention on Fridays. Doyle and Magilke (2009) more carefully examine the managerial motives associated with releasing negative news after trading hours or on Fridays by conducting firm-level tests on firms that switch their disclosure timing. They find no evidence of managers opportunistically switching negative announcements to after trading hours or Friday. Instead, they find evidence these switching decisions are largely driven by released information s complexity. However, DeHaan et al. find the median firm does not schedule their earnings announcement until two weeks after the end of the fiscal quarter, which, according to the Financial Executive Research Foundation, is roughly four days after the time most firms have finished the close process (Foundation (2013)). With the exception of Bagnoli et al. (2005), the literature largely focused on managerial incentives and firm release timing decisions. DP present a model and provide empirical evidence of systematic subdued short-term response and longer-term drift to earnings announcements on Fridays differing from analyst consensus expectations. In their model, risk-averse investors

10 10 receive a signal about a firm s future profitability, but, depending on the day of week, a fraction of investors ignore the signal. In turn, while firms cannot manipulate the signal, they determine on which day of the week it is released. In equilibrium, prices underreact to earnings surprises. Consistent with the model, they find cumulative excess returns from the announcement day to the next trading day following a large positive earnings surprise are 15% lower on Fridays relative to other weekdays. Similarly, they find excess returns in the window from two trading days to seventy-five trading days following the announcement are 70% higher on Fridays relative to other weekdays. 2 MRV look at a broader class of corporate announcements including dividend changes, seasoned equity offering, stock repurchases, mergers, and earnings and, in focusing solely on two-day abnormal returns relative to various market models, reproduce the DP result of reduced immediate response for Friday announcements. However, similar to Doyle and Magilke s method of looking within firms to test for strategic release, they look within firms for evidence of delayed investor response. When accounting for unobserved firm heterogeneity and selection bias by employing fixed effects, the Friday effect of muted immediate response disappears. My focus is on reconciling the competing theories in the Friday earnings announcement literature regarding investor inattention and firm selection. 3. Friday Earnings Announcement Data and Summary Statistics 3.1. Data I primarily follow the procedure of DP. After finding that earnings announcement dates could be reliably calculated after January 1995, DP use earnings announcements from then until 2 Contrary to Bagnoli et al., DP find no anticipatory response of stocks prior to Friday announcements.

11 11 June I collect all quarterly earnings announcements January 1995 to April 2015 from the Institutional Brokers Estimate System (I/B/E/S). For each announcement (814,312 observations), I collect all one quarter ahead analyst earnings forecasts from I/B/E/S. DP s final sample is of earnings announcements for which at least one analyst makes a forecast at most 90 days before the announcement, and their earnings announcement date imputation method allows for at most a five day difference from the announcement date recorded in I/B/E/S. Accordingly, I keep only announcements for which I/B/E/S records the forecast as occurring at most 5 days after or 95 days before (275,294 observations). In order to directly account for stock splits made over time, I adjust the earnings estimates and actual earnings announcements using the CRSP cumulative adjustment factor 3 for December 31, 2014 and merge these announcements with Compustat to cross-check the announcement date in I/B/E/S with that in Compustat (Livnat and Mendenhall (2006)). As a result, I keep only announcements for which there is there is daily return data in The Center for Research in Security Prices (CRSP) and a recorded announcement date in both I/B/E/S and Compustat, where the difference between the two is at most five days. I impute the announcement date to be the earlier one. I drop announcements due to a missing announcement date in Compustat or a discrepancy greater than five days (6,750 observations) along with those for which the announcement date is greater than three months after the fiscal quarter end (404 observations). Of those, I impute dates for 6,747. I end up with 238,230 announcements in the entire sample period and 138,942 in the DP sample for which there are forecasts made at most 90 days prior to the imputed announcement date. 3 The formula for adjusting unadjusted prices is, where is the adjusted price, is the raw price, and is the factor

12 12 To get a measure of consensus analyst forecast, DP use the median forecast made within 30 days of the announcement, using the most recent forecast if a given analyst makes multiple forecasts in that window. After more data cleaning procedures, DP end up with 143,583 observations. However, ostensibly due to changes, deletions, and revisions to I/B/E/S data (Glushkov (2009)), I find only 78,956 of the 138,942 announcements have forecasts within that window. MRV use the median analyst forecast from the I/B/E/S Summary file. In addition to issues related to unclear I/B/E/S methodology in producing Summary files ((Glushkov (2007)), this measure incorporates all outstanding estimates. To exclude stale estimates while retaining sample size, I broaden the range of acceptable forecasts to within 90 days of the announcement and otherwise define the consensus analyst forecast following DP. I also count the number of unique analysts making at least one forecast within the 90 day window. Of these announcements, I drop those occurring on Saturday or Sunday (85 observations), those for which either the consensus earnings forecast or actual earnings are greater in absolute value than the share price (1583 observations), and those corresponding to penny stocks (1524 observations). Following Kothari (2001) and DP letting, be the earnings per share announced in quarter for company,, the corresponding consensus analyst forecast, and, the share price of company 5 days before the announcement in quarter I define the earnings surprise, as,,,, (1) In order to obtain cumulative abnormal returns following each earnings announcement, I follow DP and use the Capital Asset Pricing Model (CAPM) market model as a benchmark. This equilibrium pricing model assumes investors are compensated with expected returns only for the

13 13 non-diversifiable risk associated with their assets and that a security s correlation with returns on the market portfolio is the only such risk. In calculating daily abnormal returns, DP assume the only non-diversifiable risk investors are compensated for is correlation with the market portfolio. Letting, be the net stock return of company and, be net value-weighted market return from CRSP on day, I obtain, in quarter for company from the regression,,,, ε, (2) for days from 300 to 46 4, where is the announcement date in quarter for company and ε, 0, (3) Using the, for each announcement, I calculate the returns for company in excess of those predicted by the CAPM following the announcement in quarter. Following DP, I calculate abnormal returns by subtracting only the expected return according to CAPM. That is, even though it can empirically be the case, 0 (4) for some company for days from 300 to 46, where is the announcement date in quarter for company, the abnormal return for a day u in quarter t for company k is defined as 5 4 DP specified this timeframe without any explanation. It roughly corresponds to a year s time starting about two and a half months before the announcement date and seems to include sufficient and relevant data in order to estimate the market beta without being correlated with returns associated with the current announcement. 5 I choose to follow DP for consistency, though an alternative would have been to account for systematic violations of the CAPM and calculate deresidualized abnormal returns,,,,,

14 14,,,, (5) The cumulative abnormal return (CAR) over time period, for stock in quarter is,, 1, 1, 1, 1. (6) If the company releases its earnings report after trading hours, the subsequent trading day is the first opportunity investors have to trade in the exchange on that information. Therefore the time at which the announcement is made is necessary to accurately calculate cumulative abnormal returns following the announcement. After the publishing of DP, I/B/E/S released reliable timestamp data for announcements made after January For these announcements, I denote those occurring between 12:00 a.m. and 9:29 a.m. as morning announcements, those occurring between 9:30 a.m. and 4:00 p.m. as trading hours announcements, and those occurring between 4:01 p.m. and 11:59 p.m. as evening announcements, where all times are denoted in EST. I therefore define,,,_ CAR in quarter for company from days to without timestamp date,,, CAR in quarter for company from days to with timestamp date For announcements without timestamp data, I calculate the cumulative abnormal returns,,,_ according to the rule in Figure 1. Figure 1 Return Window Announcement day Short-term 0 1 Long-term 2 75

15 15 Following an earlier draft of MRV that examined long-term CAR with timestamp data, I calculate the cumulative abnormal returns,,, according to the rule in Figure 2. Announcement Figure 2 Return Time Window Morning or Trading Hours Announcement day Short-term 0 0 Long-term 1 75 Evening Announcement day + 1 Short-term 1 1 Long-term 2 76 For simplicity, I denote,, and,,_ with and without timestamp data, respectively. as the long-term cumulative abnormal returns There are 233,231 announcements in the total sample of which 184,159 have timestamp data. In the DP sample period, there are 136,446 announcements compared to 143,583 in DP. I also collect several characteristics of each company in each quarter that may be related to the degree of attention investors allocate to that stock. For each company, I use Compustat to calculate the quarter-end market capitalization by multiplying the number of common shares outstanding by the quarter-end share price. I also collect Standard Industrial Classification (SIC) codes from Compustat. For each security, I collect the institutional ownership as a percentage of shares outstanding from the Thomson-Reuters Institutional Holdings Database on Form 13F. Following Grinstein and Michaely (2005), I restrict the institutional ownership data to the unit

16 16 interval by imputing 0% for companies not included in the 13F (33,040 observations) and 100% for observations greater than 100% (7,357 observations). 6 Figure 3 shows the distributions of earnings announcements by day of the week for the entire sample. Companies are less likely to release earnings announcements on Mondays and significantly less likely to do so on Fridays. Figure 3 All Total Monday Tuesday Wednesday Thursday Friday Number Fraction Table 1 shows two-sample t-tests between Friday and non-friday earnings announcements for the entire sample. Columns (1) and (2) present summary statistics with standard deviations in parentheses. Column (3) presents the difference between columns (1) and (2), and the standard errors for these differences are in parentheses. Similar to the findings in DP, Friday announcers are more likely to release a negative surprise, have smaller market capitalizations, and release later in the quarter. Additionally, Friday announcers have less analyst coverage and institutional ownership. 6 Missing data is often because the institutional holders do not meet the 13F filing requirements, while institutional holdings >100% are often because of fiscal/calendar year differences between different institutional investors

17 17 Table 1 Differences Between Announcements on Fridays and Other Weekdays Friday Non-Friday Difference (1) (2) (3) Earnings Surprise ** (.1634) (.1211) (.0011) Ln(Market Cap) *** (2.543) (2.301) (.020) Year *** (5.82) (5.59) (.05) Month 1 in Quarter *** (.498) (.489) (.004) Month 2 in Quarter *** (.492) (.483) (.004) Month 3 in Quarter *** (.193) (.156) (.001) Analyst Following *** (5.20) (5.58) (.05) Institutional Ownership *** (.315) (.318) (.318) N 13, , ,231 Summary statistics for entire sample of matched announcements from June 1995 to April Standard deviations are in parentheses for columns (1) and (2). Standard errors are in parenthesis for column (3). (* significant at 10%, ** significant at 5%, *** significant at 1%) To measure the relative surprise of an announcement, DP rank the announcements by their corresponding earnings surprise, and sort them into deciles. To compare the immediate and longer-term abnormal returns for Friday and non-friday announcement, they use the OLS specification,,,_,,,,,,,, (7) where, is a dummy variable indicating whether the corresponding earnings announcements of company in quarter was in the top two deciles of earnings surprises that year and, is a dummy indicating whether the announcement occurred on Friday. In order to account for a potential asymmetric effect of large positive and negative earnings surprises, the sample includes

18 18 only observations in either the top two or bottom two deciles of earnings surprises in its corresponding year. Intuitively, we expect, or the marginal difference in abnormal returns for announcements with earnings surprises in the top two deciles relative to those in the bottom two deciles, to be positive. If a company surprises the market with good news, then we expect higher returns due to the revelation of positive information. Under the null hypothesis of no selection that is, firm announcements are homogenous across weekdays, or the marginal difference in abnormal returns for Friday announcements relative to other weekdays, should equal zero. Since Table 1 shows Friday announcements are, on average, slightly worse in terms of earnings surprise than announcements on other weekdays, we expect this difference to be slightly negative. However, under the null hypothesis of no Friday inattention that is, stock prices respond to large earnings surprise regardless of the earnings announcement s timing, or the marginal difference in abnormal returns for announcements with surprises in the top two deciles on Fridays relative to those on other weekdays, should not be statistically different from zero. If, as DP find, stock prices following a Friday earnings surprise in the top two deciles in the quarter exhibit muted immediate response and more long-term drift than those on other, weekdays, we expect to be negative when,,_ is the dependent variable and positive, when,,_ is the dependent variable. Specification (7) allows for the abnormal response to depend on a set of control variables,. The standard controls include indicators for month and year to allow for seasonality and long-term trends, respectively, along with market capitalization and month in following quarter indicators to account for differences in company size and announcement times between Friday and non-friday announcers. Specifically, information diffusion may be easier for large

19 19 companies, and abnormal returns following announcements during different times in the following quarter may be capturing attention effects distinct from those with respect to Fridays. The ten market capitalization indicators are constructed from sorted deciles of the difference between the log market capitalization of company and the average log market capitalization of other companies announcing in quarter. All of the controls, are fully interacted with the indicator variable for an announcement in the top two deciles of earnings surprises that quarter, to allow for differential responses to the control variables depending on whether the announcement is in the top two deciles. Because DP do not account for analyst following or institutional ownership as control variables, I deliberately exclude them from, in the initial reproduction exercise even though they differ significantly between Friday and non-friday announcers. When controlling for analyst following or institutional following in later regressions, I construct indicators for those variables using the same procedure as that for constructing log market capitalization indicators. 7 Industry fixed effects are constructed from two-digit SIC codes. 8 Some specifications include firm fixed effects. In those, I exclude the other company characteristics as controls because those are relatively constant over time. The benefit of firm fixed effects is that they control for time-invariant characteristics of firms that can affect the delayed response of Friday announcements, even if these unobservable characteristics are correlated with the observable characteristics. 9 Standard errors are robust to heteroskedasticity and clustered by day of announcement to allow for correlation of returns on the same day. 7 The results in all subsequent regressions are similar if I include the additional controls as continuous variables as opposed to indicator variables, but this method emphasizes the importance of relative rather than absolute characteristics, allows for nonlinearities in the characteristics effects, and is consistent with the previous DP methodology. 8 The following is an example of hierarchical SIC code classification of businesses: 2024 refers to Ice Cream and Frozen Desserts, 202 refers to Dairy Products, and 20 refers to Food and Kindred Products. 9 The primary drawback is that I must assume these characteristics are fixed over time. Since the main tests will focus on announcements around a four-year window, this is a relatively reasonable assumption.

20 20 Table 2 reports results of specification (7) during the DP sample period in which the short-, term cumulative abnormal return without timestamp data,,_ is the dependent variable. I am unable to reproduce the immediate muted response seen in the literature. Table 2 Immediate Response in DP Sample Period (1) Top Two Surprise Decile *** (0.0044) Friday ** (0.0019) Friday x Top Two Surprise Deciles.0014 (0.0028) Constant *** (0.0029) Standard Controls (Interacted) X Observations 54,638 R-squared The sample period is from January 1995 to June The, dependent variable is,,_, the cumulative abnormal returns on the day of and day after the announcement. The standard controls include indicators for year, month, market capitalization, and month in the following quarter of the announcement. Each control is interacted with an indicator for an announcement being in the top two deciles of earnings surprise in that year. Robust standard errors clustered by date are in parentheses. (* significant at 10%, ** significant at 5%, *** significant at 1%) The immediate abnormal returns are based on just two trading days and therefore depend on getting the announcement date precisely correct. Even though I followed DP s announcement date procedure, any measurement error even if it is mean zero will bias the estimate of and attenuate it towards zero 10. Any additional mean zero measurement error in the data collection and estimation process will further attenuate the estimate towards zero. 10 This assumes there is no difference in pre-announcement returns between Friday announcers and non-friday announcers, which is supported by DP but not Bagnoli et al. (2005).

21 21 Because the long-term abnormal returns are compounded over 74 trading days, the regression on long-term abnormal returns will be more robust to slight measurement error in the announcement day. Table 3 reports results of specification (7) during the DP sample period in, which the long-term cumulative abnormal return without timestamp data,,_ is the dependent variable. Column (1) shows I am able to reproduce the differential longer-term drift on Fridays similar to DP, and the magnitude is comparable to what they find. 11 The coefficient on top two surprise deciles is not statistically different from what DP s estimate 12, though its noisiness further suggests the presence of measurement error in my data collection process. Similar to MRV, Columns (2) through (6) show the Friday effect disappears as additional firm characteristic controls for analyst following, institutional ownership, industry or firm fixed effects are introduced. Therefore, I focus on longer-term drift in the following sections. Because the following sections involve time periods for which I have timestamp data, I exclusively use the long-term cumulative abnormal returns with timestamp data as a measure of long-term drift and, for simplicity, suppress the notation to, They find a magnitude between.02 and.05 depending on the specification. 12 Approximately.04 13, All results that follow are similar if I naively use,,_

22 22 Table 3 Delayed Response in DP Sample Period (1) (2) (3) (4) (5) (6) Top Two Surprise Deciles ** ** (0.189) (0.186) (0.190) (0.191) (0.189) (0.0159) Friday ( ) ( ) ( ) ( ) ( ) ( ) Friday x Top Two Surprise Deciles * * (0.0109) (0.0109) (0.0109) (0.0110) (0.0110) (0.0128) Constant ** 0.473** (0.191) (0.188) (0.192) (0.190) (0.188) (0.0119) Standard Controls (Interacted) X X X X X Institutional Ownership X X Analyst Following X X Industry Fixed Effects X X Firm Fixed Effects X Number of Firms 7,612 7,612 7,612 7,612 7,612 7,612 Observations 53,869 53,869 53,869 53,863 53,863 53,869 R-squared The sample period is from January 1995 to June The dependent variable is,,_, the cumulative abnormal returns on the two days to seventy-five days after the announcement date. The standard controls include indicators for year, month, market capitalization, and month in the following quarter of the announcement. Institutional ownership and analyst following controls are constructed as deciles of within-quarter deviations from average levels. Each control, except for the firm fixed effects, is interacted with an indicator for an announcement being in the top two deciles of earnings surprises in that year. Robust standard errors clustered by date are in parentheses. (* significant at 10%, ** significant at 5%, *** significant at 1%)

23 Methodology DP attribute the anomalous behavior of prices following Friday earnings announcements to investor inattention, while MRV claim it is due to firm heterogeneity correlated with the strategic choice to announce on Fridays. Under an exogenous shock to investor inattention, DP would predict a change in the Friday earnings response, while MRV would predict a change only if it changes the composition of firms releasing on Fridays. If firm fixed effects were added so that firm specific factors that may be correlated with the decision, then MRV would not predict a change. I exploit the publishing of DP s paper Investor Inattention and Friday Earnings Announcements in the April 2009 issue of The Journal of Finance as an exogenous shock to investor attention. Stefano DellaVigna posted the first working paper 14 on his website on May 23, 2004; Social Science Research Network (SSRN) posted the initial working paper 15 on January 15, 2005 and a revised working paper 16 on September 24, 2005; the NBER 17 distributed the working paper in October 2005; and The Journal of Finance published the online version on March 13, 2009 in its April 2009 issue. By various measures based on citations, practitioner surveys, and subsequent research response, The Journal of Finance is the most prestigious academic finance journal 18, and Investor Inattention and Friday Earnings Announcements has been cited by articles in academic journals 495 times as of May 18, 2015 according to Google Scholar. McLean and Pontiff use the publishing date of the final paper as the academic paper s treatment time in their evaluation of the effect of predictable return finance papers, though because DP posit the predictable returns following Friday earnings announcements are generated See Chan, Chang, and Chang (2013)

24 24 by inattention to easily available publicly information, I need to more carefully consider whether this is the correct methodology. Ideally, I could use Google search data as a proxy for attention to the paper, but the publicly available Google Trends time-series data does not exist before 2004 and is very sparse for years prior to To my knowledge, the only instances of popular media reporting are a Slate Moneybox blog post on September 17, , a Street Sleuth column in The Wall Street Journal on September 24, , and a post in The New York Times The Lede blog on November 4, In terms of the practical investor applications of the paper s content, earlier drafts of the paper, and thus the limited popular media reporting, did not include any mention of the theoretical risk-adjusted returns a trading strategy based on the Friday earnings announcement anomaly could make. The initial draft, titled Strategic Released of Information on Fridays: Evidence from Earnings Announcements, focused on selection of negative earnings surprise into Fridays, and also examined whether this dynamic affects when the President signs legislation. It was not until the final publication version that DP demonstrated a zero-investment portfolio constructed by buying companies with large positive surprise earnings announcements on Fridays in the previous month and selling companies with large negative surprise earnings announcements on other weekdays in the previous month earns 4.62% monthly returns in excess of the Carhart four-factor model, and the online publication in The Journal of Finance highlights this finding in the openly accessible abstract

25 25 Considering all the evidence 22, I follow McLean and Pontiff and use the publication date as the treatment date. My initial regression is a difference-in-difference approach, taking specification (7) as the baseline, in a four-year window 23 around the March 13, 2009 publishing of DP in order to evaluate the paper s effect on attention. Specifically, I use the OLS specification, (8),,,,,,,,,,,,,,,,,,,,,,,,,,,,, where the other variables are defined as before and, is a dummy variable indicating whether the announcement of company in quarter occurred after March 13, 2009, which is when DP was published in The Journal of Finance. The main coefficient of interest is that on the triple interaction term,,, indicating announcements in the top two earnings surprise deciles made after the paper on a Friday. Under the null hypothesis of no Friday attention effect due to the publishing of DP, we would expect there to be no change in the Friday effect following the publishing of DP and thus, to equal zero. On the other hand, if prior abnormal returns were truly due to inattention and the publishing of DP represented a significant attention shock to which investors respond, we would expect, to be of similar magnitude and opposite sign as that of,. 22 Another potential explanation is that practitioners do not trust academic research until it has been fully vetted in the peerreview process. 23 Including announcements two years before and two years after ensures sufficient sample size so that I can estimate firm fixed effects. The results are similar if I use wider timeframes.

26 26 Similar to specification (7), we expect, to be positive and, to be negative. If, as I found for the DP sample period, there is long-term drift for Friday announcements relative to other weekday announcements, we would expect, to be positive. The interaction terms including, measure whether there was a change in the marginal effects of the variables with which, is interacted. The controls in, are interacted with, and, to allow for differences in how characteristics affect abnormal returns for companies in the top two earnings surprise deciles or after the paper s publishing. They are additionally interacted with the interaction term,, to allow the marginal effect of characteristics on abnormal returns to change following the paper s publishing. 24 It is important to note that because I lack precise measures of investor attention, I cannot reject the null hypothesis that abnormal returns following Friday announcements were never due to inattention and any change following the paper s publishing is simply due to irrational investor overreaction to irrelevant news, even if I estimate a positive, and negative, Friday Earnings Announcements Response to Paper 4.1. Persistence of Abnormal Returns 24 There are 4,030 firms with announcements in the four-year window, so I do not interact the firm fixed effects with either the top two decile or post paper indicators. I also do not interact the year dummies with the post paper indicator due to the redundancy of doing so. 25 Though I present additional robustness checks suggesting the abnormal returns are not due to other non-diversifiable risk factors, this story of investor reaction to irrelevant news would nevertheless have novel implications for the behavioral finance literature concerning investor overconfidence and overreaction (See Daniel et al. (1998) for and Hong and Stein (1999) for models of boundedly rational investors overreacting to news).

27 27 Given the concerns about information leakage, I first need to verify that the abnormal longterm price response following Friday announcements persisted in the two years prior to the paper s publishing. Table 4 reports specification (7) with, as the dependent variable for a sample in the two years prior to the publishing of DP on March, The effect appears to persist and, if anything, strengthen. Interestingly, the significance of the coefficient of interest remains even after adding additional controls and firm fixed effects, suggesting the abnormal returns following Friday announcements are robust to firm selection. 26 Table 4 Delayed Response in Two Years Prior to Paper s Publishing (1) (2) (3) (4) (5) Top Two Surprise Deciles e (0.0579) (0.0604) (0.0591) (0.335) (0.0763) Friday (0.0169) (0.0162) (0.0164) (0.0157) (0.0248) Friday x Top Two Surprise * ** * * Deciles (0.0324) (0.0347) (0.0347) (0.0350) (0.0424) Constant *** *** *** *** (0.0888) (0.0957) (0.0999) (0.351) (0.111) Standard Controls (Interacted) X X X X X Institutional Ownership X X Analyst Following X X Industry Fixed Effects X X Firm Fixed Effects X Number of Firms 3,303 3,303 3,303 3,303 3,303 Observations 9,923 9,923 9,923 9,923 9,923 R-squared The sample period is from March 13, 2007 to March 13, The dependent variable is, as defined in Section 3.1. The standard controls include indicators for year, month, market capitalization indicators, and indicators for month in the following quarter of the announcement. Institutional ownership and analyst following controls are constructed as deciles of within-quarter deviations from average levels. Industry fixed effects are constructed by two-digit SIC codes. Each control except for the firm fixed effects is interacted with an indicator for an announcement being in the top decile of earnings surprise in the year. Robust standard errors clustered by date are in parentheses. *** p<0.01, (* significant at 10%, ** significant at 5%, *** significant at 1%) 26 In Appendix 10.2, I test whether this is due to the characteristics of Friday announcers (Friday announcers in the top two deciles of earnings surprises) relative to non-friday announcers (Friday announcers in the bottom two deciles of earnings surprises) changes following the initial dissemination and reporting of the paper. I do not find this to be the case.

28 Selection Into Friday Announcements Around Paper In additional to the standard Gauss-Markov assumptions necessary to derive the efficiency and unbiasedness of OLS estimates, a difference-in-difference identification strategy requires a parallel trends assumption. In other words, this assumes that had DP not alerted investors to the Friday earnings announcement anomaly, the difference in abnormal returns between announcements in the top two earnings surprises deciles on Friday and other days, conditional on the included controls, would not have changed. A potential issue is if the composition of Friday announcers relative to non-friday announcers changes around the paper s publishing. In this case, the estimate of, may simply be reflecting different characteristics of top two surprise decile Friday announcers before and after the paper rather than a different investor response. To formally test this, I use the OLS specification,,,,,, (9) where, is a component of the vector, corresponding to characteristics of company in quarter, and the other variables are defined as above. Under the null hypothesis of no change in selection into releasing on Fridays, should equal zero. Table 5 presents the results of specification (9) and suggests the month in the following quarter in which the announcement is made is the only characteristic that significantly changed between announcements on Fridays and other weekdays after the paper s publishing.

29 29 Table 5 Change in Selection into Friday Announcement Around Paper s Publishing (1) (2) (3) (4) (5) (6) (7) Dependent Variable Earnings Surprise Ln(Market Cap) Month 1 in Quarter Month 2 in Quarter Month 3 in Quarter Analyst Following Institutional Ownership Post Paper ** *** *** *** *** ( ) (0.0204) ( ) ( ) ( ) (0.0536) ( ) Friday *** *** *** *** *** *** *** ( ) (0.0568) (0.0130) (0.0130) ( ) (0.149) ( ) Friday x Post Paper *** ** *** ( ) (0.0788) (0.0180) (0.0180) ( ) (0.207) (0.0116) Constant *** 16.51*** 0.488*** 0.475*** *** 6.348*** 0.601*** ( ) (0.0142) ( ) ( ) ( ) (0.0374) ( ) Observations 49,291 49,291 49,291 49,291 49,291 49,291 49,291 R-squared The sample period is from March 13, 2007 to March 13, Standard errors in parentheses. (* significant at 10%, ** significant at 5%, *** significant at 1%) The main test of earnings response change following the paper s publishing restricts the sample to earnings surprises in the top two and bottom two deciles of earnings surprises in the announcement s corresponding year, and the coefficient of interest, compares Friday announcement in the top two deciles before the announcement relative to those after. Therefore, I also need to check whether the difference in characteristics of announcements in the top two and bottom two deciles between Fridays and other weekdays changed following the paper s publishing. To formally test this, I use the OLS specification, (10),,,,,,,,,,,,,,,,, Under the null hypothesis of no change in selection into Friday announcements for those with earnings surprises in the top two deciles for the corresponding year,, should

30 30 equal zero. Table 6 presents the results of specification (10) for the two years before and after the paper s publishing. The estimate of, is significant in Column (2) and Column (6), providing some evidence that Friday announcements in the top two earnings surprise deciles are more likely to have larger market capitalizations and analyst followings in the two years following the paper s publishing relative to the two previous years. This could potentially lead to a violation of the parallel trends assumption, so I add analyst following indicators to the set of fully interacted standard controls,. Table 6 Change in Selection into Friday Announcement for Top Two and Bottom Two Surprise Deciles Around Paper s Publishing (1) (2) (3) (4) (5) (6) (7) Earnings Surprise Ln(Market Cap) Month 1 in Quarter Month 2 in Quarter Month 3 in Quarter Analyst Following Institutional Ownership Top Two Surprise Deciles *** *** ** *** ** ** ** ( ) (0.0440) (0.0102) (0.0103) ( ) (0.103) ( ) Friday * ** ** ** ( ) (0.0449) (0.0103) (0.0105) ( ) (0.105) ( ) Post Paper *** *** * *** *** ( ) (0.111) (0.0256) (0.0259) (0.0113) (0.260) (0.0175) Post Paper x Friday ** ( ) (0.154) (0.0356) (0.0360) (0.0158) (0.362) (0.0243) Post Paper x Top Two Surprise Deciles ** 0.374** (0.0107) (0.173) (0.0398) (0.0404) (0.0177) (0.406) (0.0273) Top Two Surprise Deciles x Friday *** *** ( ) (0.0631) (0.0145) (0.0147) ( ) (0.148) ( ) Post Paper x Top Two Surprise Deciles x Friday ** *** (0.0146) (0.235) (0.0541) (0.0548) (0.0240) (0.550) (0.0370) Constant *** 16.17*** 0.423*** 0.518*** *** 5.515*** 0.528*** ( ) (0.0310) ( ) ( ) ( ) (0.0727) ( ) Observations 19,823 19,823 19,823 19,823 19,823 19,823 19,823 R-squared The sample period is from March 13, 2007 to March 13, 2009 and is restricted to announcements in the top two and bottom two deciles of earnings surprises in the corresponding year. Standard errors in parentheses. (* significant at 10%, ** significant at 5%, *** significant at 1%)