Attracting attention in a limited attention world - The University of ...
ACCOUNTING WORKSHOP “Attracting attention in a limited world: An exploration of the forces behind positive extreme earnings surprises” By Allison Koester Foster School of Business University of Washington Russell Lundholm Sauder School of Business University of British Columbia Mark Soliman* Foster School of Business University of Washington
Thursday, April 14, 2011 1:30 – 2:50 p.m. Room C06
*Speaker Paper Available in Room 447
Attracting attention in a limited attention world: An exploration of the forces behind positive extreme earnings surprises
Allison Koester Ph.D. Candidate Foster School of Business University of Washington [email protected]
Russell Lundholm Professor of Accounting Sauder School of Business University of British Columbia [email protected]
Mark Soliman Associate Professor of Accounting Foster School of Business University of Washington [email protected]
December 2010
Acknowledgements The paper has benefitted from comments by David Burgstahler, Jared Jennings, Jon Karpoff, Shiva Rajgopal, Devin Shanthikumar, and Terry Shevlin, as well as from workshop participants at the University of Utah, University of Washington, the 2010 UBCOW Conference, and the 2011 FARS mid-year meeting.
0
Attracting attention in a limited attention world: An exploration of the forces behind positive extreme earnings surprises
Abstract Why do large positive earnings surprises occur? The literature often treats large earnings surprises as the exogenous event that precipitates subsequent stock price drift, but analyst expectations and earnings realizations are the result of conscious decisions made by analysts and managers. While neither analysts nor managers have an obvious incentive for an extreme deviation between expected and realized earnings to occur, this phenomenon is regularly observed. In this paper we explore three hypotheses as to why large positive earnings surprises occur: a) managerial manipulation of analyst expectations and/or earnings realizations; b) analyst inattention; and c) natural variation in firm performance. We find support for all three hypotheses. Further, if managers created the earnings surprise to attract attention, they were successful. There is a significant increase in analyst following, the percentage of firm shares held by institutional investors, and short and long-term trading volume following large positive earnings surprises.
Keywords: Large Earnings Surprise, Attracting Attention, Analyst Forecast Accuracy, Earnings Management JEL classification: M4
1
I. INTRODUCTION In this paper we investigate the causes, and to a lesser degree, consequences of large positive earnings surprises. Instances where a firm‟s quarterly earnings per share (EPS) far exceeds the analyst consensus forecast are interesting because, ex ante, it seems that neither managers nor analysts have an incentive for this outcome to occur.
While positive earnings
surprises are typically met with stock price increases, the earnings/return relation is S-shaped, meaning the marginal stock price reaction is diminishing in earnings surprise magnitude (Freeman and Tse 1992; Kinney et al. 2002; Burgstahler and Chuk 2009). Why, in some circumstances, did management sit on such big earnings news, contradicting the existing literature‟s finding that managers tend to immediately reveal good news to investors (see Kothari et al. 2009 and references therein)? Why did analysts fail to revise their EPS forecasts upward prior to the firm‟s earnings announcement? Was information available that the analysts failed to incorporate into their forecasts, or did analysts not have the basic inputs necessary to revise their forecasts? Finally, did neither managers nor analysts anticipate that an unexpectedly good quarter was materializing, and the positive unexpected earnings were truly a surprise to all parties involved? The accounting literature has focused considerable attention on the middle of the earnings surprise distribution since this is where most of the observations reside. In our sample, 36.5 percent of firm-quarter observations report actual earnings per share (EPS) within a two cent range of the analyst consensus forecast estimate. The relatively large number of cases with small positive earnings surprises and the relatively few cases with small negative earnings surprises is offered as evidence consistent with earnings management (Burgstahler and Dichev 1997; Burgstahler and Eames 2006; Burgstahler 2010).
2
Bartov et al. (2002) and Kasznik and
McNichols (2002) offer a compelling reason for this pattern: firms who consistently “meet or beat” the consensus analyst forecast are rewarded with a valuation premium. This is one reason management communicates regularly with analysts. In contrast to these studies, we are interested in cases where communication between managers and analysts is relatively ineffective and actual EPS far exceeds the analyst consensus forecast. To put this in some perspective, the median earnings surprise in the top decile of quarterly earnings surprises is one percent of price, which translates to an earnings surprise greater than 140 percent of the forecasted EPS (untabulated). In absolute terms, this is a huge earnings surprise. Our research question is simple: why did the large positive earnings surprise occur? Did management fail to communicate or did analysts fail to update their forecasts? The null hypothesis for this study is that there is no way to identify firms that will have extreme earnings surprises; by definition, ten percent of the firms will always be in the top (i.e., “extreme") earnings surprise decile and there is nothing “surprising” about this. However, prior work has shown that being in the extreme deciles precedes rather dramatic abnormal returns (Bernard and Thomas 1990; Abarbanell and Bernard 1992; Livnat and Mendenhall 2006; Doyle et al. 2006), suggesting there may be a reason (other than chance) as to why these firm-quarter observations fall in the extreme earnings surprise deciles.1 The only other research focusing explicitly on extreme earnings surprises is Kasznik and Lev (1995). They examine firms‟ disclosure activity prior to large earnings surprises in the fourth quarter in 1988, 1989 and 1990. They focus mostly on large negative surprises, finding that firms who issue earnings warnings tend to be larger, in high tech industries, and have issued 1
Specifically, Bernard and Thomas (1990) find that post-earnings announcement returns are eight percent higher for firms in the top decile of earnings news relative to firms in the bottom decile (Table 2); Abarbanell and Bernard (1992) document a one-year hedge portfolio return between the top and bottom deciles of earnings surprise of eight percent (Figure 1); and Doyle et al. (2006) document a three-year hedge portfolio return of between the top and bottom deciles of earnings surprise of twenty-four percent (Table 2).
3
management guidance previously in the quarter. These disclosures appear to be driven by litigation concerns. Possibly for this reason, bad news warnings outnumber good news warnings two to one. Nonetheless, given the harm suffered by shareholders who sell their shares prior to a large positive earnings surprise, the authors conclude by wondering why managers did not disclose their firms‟ large positive earnings surprise earlier. We explore three reasons why a firm might have a positive extreme earnings surprise (PEES). Our first hypothesis explores the incentives of managers. PEES firms tend to be smaller firms with less analyst coverage and lower institutional investor ownership (Doyle et al. 2006). We hypothesize that managers of these neglected stocks might be attempting to garner investor attention by withholding positive information from analysts in order to later announce “surprising” positive earnings results and garnish attention from the press. Investor relations departments of small firms are constantly struggling to attract the attention of analysts and hoping that those analysts highlight them in their reports to investors. In short, management‟s goal is to be in the Earnings Surprise column in the Wall Street Journal and that this is an effective way to get analysts to take a closer look at their company. Odean (1999) proposes that investors limit their investment decisions to stocks that have recently caught their attention. Barber and Odean (2008) test this proposition by examining attention-based decision making in the context of equity purchases and sales for retail investors. Given the large number of stocks available to possibly invest in and the information search costs of researching so many firms, the authors propose that investors only consider purchasing stocks that have recently caught their attention, and that this effect will be greater for individual investors who have fewer information search resources than institutional investors. The authors find that individual investors display attention-driven buying behavior – individuals are net
4
buyers when stocks are mentioned in the news, on high-volume days, and on the day after both extremely negative and extremely positive one-day returns, and find that attention-driven buying is similar for both large and small capitalization stocks. Busse and Green (2002) may provide the most compelling direct evidence of attention-based investing, using TAQ data to document abnormally high trading volume when investor attention to a stock increases. As an example, when Maria Bartiromo mentions a stock during the CNBC television show Midday Call, average trading volume for this stock increases nearly five-fold during the few minutes following the mention. Our first hypothesis is that managers, aware that increases in investor attention are associated with increases in investment, use large positive earnings surprises to garner investor attention. While it is always difficult to establish managerial intent, we find that PEES firms tend to have “neglected” firm characteristics. In particular, in the year prior to a PEES, PEES firms experience return on assets above the industry median yet stock price increases below the industry median, suggesting that PEES firms experienced good performance news that was not fully appreciated by the market. PEES firms also have larger increases in return on assets during the four quarters following the PEES. Further, these same variables do not predict negative extreme earnings surprises. Finally, these firms are less likely to have issued EPS guidance for the quarter in which the PEES occurred, indicating that the resulting earnings “surprise” may be the result of management withholding favorable information until the earnings announcement date. We also show that management‟s attempt to attract attention is ultimately successful. We document a significant increase in the number of analysts following the firm, the percentage of shares held by institutions, and short-term and long-term trading volume subsequent to a PEES quarter.
5
Our second hypothesis is that PEES occur because analysts are not paying attention to the positive developments happening at the firm. Analysts may not be sufficiently interested in the stock to put in the effort necessary to stay abreast of management communications and other firm and industry events that have the potential to make the firm‟s upcoming earnings more predictable. If this hypothesis is descriptive, the variables that proxy for analyst costs and benefits should predict both large positive and large negative earnings surprises. We find that relative to non-PEES firms, PEES firms have lower analyst following, and the analysts who are following the firm are busier (defined as following a greater number of other firms) and issue individual earnings forecasts with greater disagreement between analysts forecasting for the same firm.2 Our third and final hypothesis – the natural variation in firm performance hypothesis – serves as a benchmark for the other two hypotheses. It could simply be that neither party in the earnings announcement game knew anything about the impending large positive earnings surprise; rather, exogenous shocks to the firm‟s operating performance generated the unexpectedly good earnings realizations. Consistent with this hypothesis, we find that PEES are more likely for firms with volatile prior quarterly return on equity and higher operating leverage, both of which make accurately forecasting quarterly earnings a more challenging endeavor. Recognizing that these three hypotheses are not mutually exclusive, we attempt to “apportion blame” between the three hypotheses by analyzing the changes in the odds ratio for different sets of independent variables. We examine how a one unit change in binary variables 2
As discussed in greater detail later, managers and analysts may have conflicting incentives to forecast and disclose negative earnings surprises. While analysts have incentives to forecast negative events as soon as possible, managers may have incentives to delay the dissemination of bad information as long as possible (Kothari et al. 2009). We compare the determinants of large positive earnings surprises to the determinants of large negative earnings surprises in our empirical analysis to further distinguish between our three hypotheses. We predict that while the latter two hypotheses likely predict both positive and negative large earnings surprises, the managerial manipulation hypothesis will predict large positive but not large negative earnings surprise.
6
and a one standard deviation change in continuous variables influence the odds of observing a PEES. We find that such a change in the managerial manipulation proxies doubles the odds of a PEES firm-quarter, after controlling for firm size, institutional ownership, and all the other hypothesized variables in the model. Similarly, a one standard deviation change in the analyst inattention variables doubles the odds of a PEES occurring, after controlling for all other variables in the model. The natural variation variables turn out to be less important: a one standard deviation change in these proxies increases the odds of a PEES occurring by a factor of only 1.11, after controlling for all other variables. Finally, a one unit or one standard deviation change in all the variables in the full model yields an impressive nine-fold increase in the odds of a positive extreme earnings surprise. The paper proceeds as follows. Section II provides hypothesis development and variable measurement. Section III presents the results and Section IV concludes.
II. HYPOTHESIS DEVELOPMENT AND VARIABLE MEASUREMENT Defining a Positive Extreme Earnings Surprise Before developing our hypotheses we describe how “surprising” it is for firm-quarter observations to be in the top earnings surprise decile. We define quarterly earnings surprise (SURPRISE) as the IBES actual EPS for a firm-quarter less the IBES analyst consensus EPS forecast for the firm-quarter, divided by the firm‟s stock price per share at the end of the fiscal quarter, where the IBES actual is unadjusted for subsequent splits and the IBES consensus EPS forecast is the most recent median forecast preceding the firm‟s earnings announcement date. 3 We use EPS figures unadjusted for subsequent stock splits based on findings that using split3
Note that this is the definition of an analyst forecast error commonly used in the literature. We choose to label this measure „earnings surprise‟ because labeling the deviation as an analyst forecast error implies that it is the analysts who are the cause of the difference between actual and forecasted EPS.
7
adjusted data can potentially distort both time-series and cross-sectional characteristics of earnings surprises (Baber and Kang 2002; Payne and Thomas 2003). The unsplit-adjusted database gives the earnings per share that was actually reported in the company‟s earnings announcement, which is the news that the market observed at that time. We rank firm-quarter observations by SURPRISE in each calendar quarter and focus on observations in the top decile. The indicator variable PEES_D is set equal to one when a firmquarter is in the top decile of SURPRISE for that calendar quarter, and set equal to zero otherwise. By sorting observations within each calendar quarter to determine when a PEES has occurred, as opposed to sorting the entire pooled sample over the time period examined, we control for changing economic conditions. Without this calendar quarter restriction, quarters with large unexpected improvements in economy-wide performance would dominate the PEES sample, and the answer to why firms have large positive earnings surprises would revolve around predicting when the economy will experience unexpected improvements. The SURPRISE value cutoff for the top decile changes each quarter, with the lowest cutoff value of 0.0032 in the first quarter of 1998 and the highest cutoff value of 0.0159 in the third quarter of 1984 (untabulated). Pooling over all quarters, the median SURPRISE value cutoff is 0.01 (i.e., a quarterly earnings surprise equal to 1 percent of stock price).4
4
Doyle et al. (2006) compare the earliest available IBES actual EPS in the unsplit-adjusted database with the actual press release found through LexisNexis for fifty firms and find that the data from the two sources matched in all fifty cases. In addition, Doyle et al. (2010) match 1,000 random quarterly observations from 1997 through 2000 from the IBES unsplit-adjusted database to the press release issued by firms via a Lexis-Nexis search and find that IBES corresponds perfectly with the press release in 915 cases. Given this high level of documented accuracy, we do not believe that large systematic errors in the IBES database are driving our results.
8
Managerial Manipulation Hypothesis This hypothesis is based on the premise that managers of “neglected” firms want to attract investor attention with large positive earnings surprises. Feeling neglected is not an irrelevant concern; if the amount of attention a firm receives influences its stock price, then the manager of a neglected firm has an incentive to draw attention to the firm. In Merton‟s limited attention CAPM (1987) model, investors only form diversified portfolios from the set of firms that they know about. And, as previously discussed, Odean (1999) proposes that investors do not evaluate each of the thousands of stocks actively trading in public markets when making their investment decisions. Rather, investors limit their investment decisions to stocks that have recently caught their attention. If a firm is neglected, it is less likely to be included in many of these portfolios, which increases its cost of capital. Further, the greater the firm‟s idiosyncratic risk, the more it stands to benefit from an increase in investor recognition. Lehavy and Sloan (2008) provide some evidence consistent with this model, finding that increases in institutional ownership are associated with decreases in expected returns. More directly, Brown et al. (2009) find that the bid-ask spread and the probability of informed trade both decline in the year following large positive earnings surprises, and interpret this as evidence that the firm has increased its visibility to the market. Similarly, Irvine (2003) finds that firms experience a significant increase in liquidity following the initiation of analyst coverage of a stock, and the effect increases with the analyst‟s recommendation. Barber and Odean (2008) find that individual investors tend to buy stocks that grab their attention – i.e., those mentioned in the news, with high daily trading volume, with an extreme daily return, etc. These same investors do not exhibit attention-based selling behavior because they tend to only sell stocks that they already own.
9
Management has various means to attract investor attention. For instance, Bushee and Miller (2007) find that small firms who hire an investor relations consulting firm enjoy increased media coverage, increased analyst following and increased institutional ownership. In a similar vein, Grullon et al. (2004) find that firms with greater annual advertising spending have a larger number of both individual and institutional investors. Our first hypothesis considers whether managers create large positive earnings surprises in an attempt to attract investor and analyst attention. Similar to managers attempting to put their best foot forward by strategically highlighting the prior period earnings amount that provides the lowest possible benchmark when announcing current period earnings (Schrand and Walther 2000), we hypothesize that managers attempt to gain attention by announcing a very large favorable earnings surprise at a time when the firm‟s strong earnings performance allows managers to put their best foot forward. Did management intentionally withhold information or intentionally guide analyst forecast downward in order to create a positive earnings surprise? Such behavior will be effective if it causes investors to investigate the firm further and favorably change investors‟ beliefs about firm value. Further, manufacturing a positive earnings surprise is not an easy strategy for fundamentally bad firms to imitate; the bad firms would have to guide analysts to extremely low expectations, and would have to withstand the subsequent increased attention and scrutiny by investors. Besides being neglected (as measured by analyst coverage and institutional ownership), does the manager have a reason in this particular quarter to seek investor attention?5 Teoh and Hwang (1991) provide an analytical model in which a firm's strategic timing of disclosure is informative to investors in revealing firm type (high or low). Firm type refers to the 5
In this paper, we do not use direct measures of earnings management as a way to test if managers are exceeding benchmarks to seek attention. The primary reason for this exclusion is that 1) measures of earnings management are extremely noisy and imprecise and 2) managers may use other methods of achieving this objective that are not captures through tradition earnings management measures; for instance, they may alter the definition of earnings through pro forma reporting (Doyle et al. 2003).
10
propensity of the firm to receive good news, with high-type firms having a higher propensity for good news than low-type firms. Firm type is known by the managers of the firm but is infeasible to directly and credibly disclose to investors. The model predicts that high-type firms prefer to withhold favorable information in period t while low-type firms prefer to immediately disclose favorable information in period t. This strategic disclosure occurs because a firm that receives favorable information in period t and expects additional favorable information in period t+1 can signal its confidence in its future favorable information by waiting until after period t to disclose the favorable information. In contrast, a firm that receives favorable information in period t but does not expect additional favorable information in period t+1 will immediately disclose its period t favorable information to prove it is not the worst of the low-type firms in period t. Because high-type firms are more likely to receive future good news than low-type firms, a separating equilibrium is obtained through the timing of disclosure of favorable information (i.e., when investors observe a firm disclosing favorable information about period t in period t, investors infer the firm is of low-type), even in the absence of exogenous disclosure costs. The authors' prediction that hightype firms can use the withholding of favorable information to signal their high type is consistent with our prediction that managers of PEES firms will wait to reveal positive and unexpected earnings information to investors at the earnings announcement. Because we cannot directly observe management‟s intentions, we look for indirect evidence to assess this hypothesis. We proxy for managerial intent to attract attention with three variables. All variables are defined in detail in the Appendix. The first variable is the change in mean industry-adjusted return on assets (CH_ROA) from four quarters prior to the PEES quarter (t-4 to t-1) to the four quarters following the PEES quarter t (t+1 to t+4), where return on assets
11
(ROA) is measured as net income divided by total assets (Compustat item niq divided by item atq), less the industry median for the same period. Quarter t is defined as the quarter in which a PEES occurs, and industries are defined using the Fama-French 48 industry classification (Fama and French 1997). This variable was used in Lang and Lundholm (1993) to proxy for the private information managers have about the firm‟s expected future performance. The idea is that managers know how much expected good or bad news the future holds, and this will influence how much attention the managers want from the analyst community. We predict a positive association between CH_ROA and PEES_D. The second variable is designed to capture situations where the manager might feel that the firm‟s strong past earnings performance was not adequately rewarded by the stock market. We set the binary variable NOLOVE_D equal to one if both of the following are true (and set equal to zero otherwise): i) the firm‟s average ROA in periods t-4 through t-1 is greater than the industry median ROA over the same period, and ii) the firm‟s prior year market-adjusted buyand-hold stock return is less than the industry median market-adjusted return over the same period. The prior year stock return is measured beginning 252 trading days and ending two trading days before the firm‟s quarterly earnings announcement date, and returns are marketadjusted by subtracting the return on a value-weighted market portfolio from a firm‟s raw return. We predict a positive association between NOLOVE_D and PEES_D. The third variable, labeled GUIDE_D, is a binary variable set equal to one if management provided any type of EPS guidance for period t earnings prior to the earnings announcement date, and set equal to zero otherwise. If management provides guidance it is unlikely the manager is attempting to manufacture an extreme earnings surprise, as there are negative
12
consequence to managers who miss their own earnings forecast (Lev and Kasznik 1995). Consequently, we predict a negative association between GUIDE_D and PEES_D.
Analyst Inattention Hypothesis Our second hypothesis focuses on analysts and the quality of their forecasts, as it is possible large earnings surprises occur because analysts made particularly poor forecasts. Forecast accuracy has been shown to increase with the analyst‟s general experience, firmspecific experience, employment at a top-ten brokerage house, and the number of forecast revisions; and decrease with the number of firms and industries an analyst follows and the staleness of the forecast (Mikhail et al. 1997, 1999; also see Ramnath et al. 2008 for an excellent literature review). Assuming that the analyst labor market allocates the most time and talent to the firms that are most valuable to the analysts, it could be that large earnings surprises are associated with analysts and firms where the benefit of being more accurate is simply too low to justify the cost. Consistent with the idea that analysts rationally trade off the costs and benefits of expending effort on forecast accuracy, Alford and Berger (1999) find that both analyst following and forecast accuracy increase with the firm‟s trading volume (and thus trading commissions).
Similarly, Lang and Lundholm (1996) find that forecast accuracy increases as
the cost of collecting information declines, where the cost of collection is measured by the firm‟s disclosure activity. Our second hypothesis is that large earnings surprises occur for firms where analyst attention is low (i.e., firms where analysts expect the cost of being more accurate to exceed the benefit). We use five variables to proxy for analyst inattention. The first is the number of analysts issuing forecasts included in the most recent IBES analyst consensus EPS forecast prior to the
13
earnings announcement date (ANALYST_CNT).
This is a good summary measure of the
equilibrium supply of analyst effort allocated to the firm. We expect ANALYST_CNT to be negatively associated with PEES_D. The next two variables relate to properties of the consensus forecast itself. FORECAST_STALE is the number of days between the most recent IBES analyst consensus EPS forecast and the earnings announcement date. All FORECAST_STALE values are positive by construction as we require the analyst consensus estimate date to be before the earnings announcement date. The greater the period of time between the two dates, the greater the chance that new information regarding the earnings realization is not incorporated into the consensus forecast. We therefore expect this variable to be positively associated with PEES_D. The standard deviation of individual analyst EPS forecasts that comprise the consensus forecast, labeled FORECAST_SD, has been used in prior research as a measure of analyst uncertainty (Baron et al. 1998; Burgstahler and Chuk 2009). We expect this variable to be positively associated with PEES_D. If there is only one forecast in the consensus, which occurs 24 percent of the time, FORECAST_SD is set to zero; there is no disagreement among analysts in this case. The next two variables relate to properties of the analysts themselves. ANALYST_BUSY is defined as the average number of other firms each analyst is following in the quarter. As another measure of the supply of analyst attention, we predict ANALYST_BUSY will be positively associated with PEES_D. Finally, ANALYST_EXP is the average number of years the analysts who contribute forecasts to the consensus in the quarter have been providing earnings forecasts captured by IBES. Given that prior research has found that more experienced analysts are more accurate, we expect ANALYST_EXP to be negatively associated with PEES_D (Ramnath et al. 2008).
14
The interpretation of the managerial manipulation variables sometimes overlaps with the interpretation of the analyst inattention variables. This is not surprising, as the managerial manipulation hypothesis posits that managers feel their firm is neglected and attempt to manufacture investor attention through PEES, while the analyst inattention hypothesis measures how neglected the firm actually is. If extreme positive earnings surprises are simply due to analyst inattention, however, then the analyst inattention variables should predict both positive and negative extreme earnings surprises. This, however, should not be the case for the managerial inattention variables. Accordingly, we create the binary variable NEES_D that is set equal to one when the earnings surprise is in the lowest decile (i.e., largest negative SURPRISE values) for the calendar-quarter, and set to zero otherwise. By contrasting the relation between our independent variables and PEES_D versus NEES_D, we can partially distinguish between the managerial manipulation and analyst inattention hypotheses.6
Natural Variation Hypothesis Our third hypothesis is that a PEES is neither managers‟ nor analysts‟ fault; rather, the firm‟s operations are simply volatile and generate high-variance earnings from quarter to quarter. We proxy for this effect with the variable SD_ROE, defined as the standard deviation of return on equity computed over the four quarters prior to the PEES quarter t (t-4 to t-1), where return on equity (ROE) is measured as net income divided by book value of equity (Compustat item niq divided by item seqq). In addition, the earnings of firms with high operating leverage (i.e., high fixed costs as a percent of total costs) are inherently difficult to forecast because most of these firms‟ profit (or lack of profit) occurs in the last few days of the quarter. For these firms, neither
6
Note that by construction, when PEES_D = 1, NEES_D = 0 and vice versa. Consequently, a variable that predicts PEES_D = 1, will weakly predict NEES_D = 0, and vice versa.
15
the firm nor the analyst may know what the quarter will hold until the last few days of the quarter.
Following Darrat and Mukherjee (1995), we measure operating leverage, labeled
OP_LEV, as the slope coefficient from regressing the change in earnings before interest and taxes (Compustat items niq + xintq + txtq) on the change in sales (Compustat item revtq), estimated via rolling regressions using data from quarters t-20 to t-1. Both SD_ROE and OP_LEV are expected to be positively associated with PEES_D. As both variables capture the uncertainty about the firm‟s upcoming earnings results, both variables should predict NEES_D=1 as well as PEES_D=1 (similar to the analyst inattention variables). Finally, we include two control variables in all regression specifications. The percent of institutional ownership, INST_OWN, measured as the percentage of outstanding shares held by institutional investors at the end of the prior quarter and is obtained from the Thomson-Reuters Institutional 13f Holdings database. This variable is another measure of the market‟s interest in the firm apart from analyst coverage. Finally, we include MVE (the natural log of the firm‟s market value (Compustat items prccq*cshoq)) as our second control variable because prior research finds that larger firms disclose more and have more analyst coverage (Lang and Lundholm 1993, 1996). Further, Vuolteenaho (2002) argues that small firms have greater idiosyncratic risk and consequently have more to gain by increasing their investor recognition. Both control variables are expected to be negatively correlated with PEES_D.
16
III. RESULTS Descriptive Statistics Our initial sample consists of all firm-quarters between July 1984 and July 2008 with sufficient data to compute the variable SURPRISE for the quarter.7 Our initial sample includes 257,832 firm-quarter observations, with the final number of observations included in each regression specification ranging from 130,080 to 243,350 based on which independent variables are required.8 Panel A of Table 1 presents the descriptive statistics for all the variables in our study. During the typical firm-quarter, management does not provide any type of EPS guidance, as only 15.1 percent of the observations have GUIDE_D=1. Because CH_ROA is industry median adjusted, the median value is zero and the mean value of -0.002 is very close to zero. On average, 5.2 analysts provide an EPS estimate (ANALYST_CNT) and the consensus forecast is issued 20.5 days before the earnings announcement (FORECAST_STALE). The average analyst follows 12 firms (ANALYST_BUSY) and has been issuing forecasts as recorded in IBES for 5.3 years (ANALYST_EXP). Institutional ownership averages 47 percent (INST_OWN) and a dollar
7
The earliest period in which IBES reports analyst earnings forecast consensus data is March 1984. Our analysis begins with July 1984 because we require a minimum of 500 earnings surprise observations per calendar quarter to compute SURPRISE. 8 We re-estimate the four regressions presented in Table 3 (when PEES_D is the dependent variable) after conducting the following four sensitivity checks. 1) The data used to construct the variable GUIDE_D comes from Thomson First Call‟s Company Issued Guidance database. Because Chuk, Matsumoto and Miller (2009) provide evidence that this database‟s coverage may be unreliable prior to 1998, we restrict the sample to post-1997 firmquarter observations; 2) The majority of firm-quarter observations in our sample (85.53 percent) have fiscal quarterends in March, June, September, and December. In our main analysis we group observations by calendar quarter (i.e., combining observation with January, February, and March fiscal quarter-ends, April, May, and June fiscal quarter-ends, etc.) to insure a sufficiently large quarterly sample pool. We eliminate observations with non-March, June, September, or December quarter-ends to ensure these observations are not driving our results; 3) To consider any possible confounding effects of Regulation Fair Disclosure‟s implementation in August 2000, we restrict the sample to post-2000 firm-quarter observations; and 4) To consider whether systematic differences between negative and positive earnings surprise observations are driving our results, we eliminate all negative earnings surprise firmquarter observations. There are no qualitative or quantitative changes in the results for any of these four sensitivity checks.
17
change in sales results in a 0.298 dollar change in earnings before interest and taxes, on average (OP_LEV). Panel B of Table 1 shows descriptive statistics for the subsamples where PEES_D=1, NEES_D=1 and the remaining middle 80 percent of the earnings surprise distribution.
The
results indicate that SURPRISE exhibits two of the same asymmetries documented in Abarbanell and Lehavy (2003). Large negative earnings surprises are more extreme than large positive earnings surprises, as evidenced by a median SURPRISE value of 0.010 for the PEES sample and a median SURPRISE value of -0.026 for the NEES sample. In untabulated results we also find that there is a greater number of small positive surprises (49.24 percent of firm-quarter observations with 0.00 < SURPRISE > < > > < > < < > >
Middle 80 Percent Sub-sample Mean -0.001 0.168 -0.002 0.272 5.846 18.172 12.136 5.389 0.026 0.486 6.367 0.052 0.293
S.D. 0.005 0.374 0.028 0.445 5.234 21.573 7.055 2.761 0.042 0.286 1.700 0.207 0.675
P50 0 0 0 0 4.000 14.000 11.625 5.276 0.010 0.484 6.243 0.013 0.240
(B)
> > > > > < < > < > > <
Thursday, April 14, 2011 1:30 – 2:50 p.m. Room C06
*Speaker Paper Available in Room 447
Attracting attention in a limited attention world: An exploration of the forces behind positive extreme earnings surprises
Allison Koester Ph.D. Candidate Foster School of Business University of Washington [email protected]
Russell Lundholm Professor of Accounting Sauder School of Business University of British Columbia [email protected]
Mark Soliman Associate Professor of Accounting Foster School of Business University of Washington [email protected]
December 2010
Acknowledgements The paper has benefitted from comments by David Burgstahler, Jared Jennings, Jon Karpoff, Shiva Rajgopal, Devin Shanthikumar, and Terry Shevlin, as well as from workshop participants at the University of Utah, University of Washington, the 2010 UBCOW Conference, and the 2011 FARS mid-year meeting.
0
Attracting attention in a limited attention world: An exploration of the forces behind positive extreme earnings surprises
Abstract Why do large positive earnings surprises occur? The literature often treats large earnings surprises as the exogenous event that precipitates subsequent stock price drift, but analyst expectations and earnings realizations are the result of conscious decisions made by analysts and managers. While neither analysts nor managers have an obvious incentive for an extreme deviation between expected and realized earnings to occur, this phenomenon is regularly observed. In this paper we explore three hypotheses as to why large positive earnings surprises occur: a) managerial manipulation of analyst expectations and/or earnings realizations; b) analyst inattention; and c) natural variation in firm performance. We find support for all three hypotheses. Further, if managers created the earnings surprise to attract attention, they were successful. There is a significant increase in analyst following, the percentage of firm shares held by institutional investors, and short and long-term trading volume following large positive earnings surprises.
Keywords: Large Earnings Surprise, Attracting Attention, Analyst Forecast Accuracy, Earnings Management JEL classification: M4
1
I. INTRODUCTION In this paper we investigate the causes, and to a lesser degree, consequences of large positive earnings surprises. Instances where a firm‟s quarterly earnings per share (EPS) far exceeds the analyst consensus forecast are interesting because, ex ante, it seems that neither managers nor analysts have an incentive for this outcome to occur.
While positive earnings
surprises are typically met with stock price increases, the earnings/return relation is S-shaped, meaning the marginal stock price reaction is diminishing in earnings surprise magnitude (Freeman and Tse 1992; Kinney et al. 2002; Burgstahler and Chuk 2009). Why, in some circumstances, did management sit on such big earnings news, contradicting the existing literature‟s finding that managers tend to immediately reveal good news to investors (see Kothari et al. 2009 and references therein)? Why did analysts fail to revise their EPS forecasts upward prior to the firm‟s earnings announcement? Was information available that the analysts failed to incorporate into their forecasts, or did analysts not have the basic inputs necessary to revise their forecasts? Finally, did neither managers nor analysts anticipate that an unexpectedly good quarter was materializing, and the positive unexpected earnings were truly a surprise to all parties involved? The accounting literature has focused considerable attention on the middle of the earnings surprise distribution since this is where most of the observations reside. In our sample, 36.5 percent of firm-quarter observations report actual earnings per share (EPS) within a two cent range of the analyst consensus forecast estimate. The relatively large number of cases with small positive earnings surprises and the relatively few cases with small negative earnings surprises is offered as evidence consistent with earnings management (Burgstahler and Dichev 1997; Burgstahler and Eames 2006; Burgstahler 2010).
2
Bartov et al. (2002) and Kasznik and
McNichols (2002) offer a compelling reason for this pattern: firms who consistently “meet or beat” the consensus analyst forecast are rewarded with a valuation premium. This is one reason management communicates regularly with analysts. In contrast to these studies, we are interested in cases where communication between managers and analysts is relatively ineffective and actual EPS far exceeds the analyst consensus forecast. To put this in some perspective, the median earnings surprise in the top decile of quarterly earnings surprises is one percent of price, which translates to an earnings surprise greater than 140 percent of the forecasted EPS (untabulated). In absolute terms, this is a huge earnings surprise. Our research question is simple: why did the large positive earnings surprise occur? Did management fail to communicate or did analysts fail to update their forecasts? The null hypothesis for this study is that there is no way to identify firms that will have extreme earnings surprises; by definition, ten percent of the firms will always be in the top (i.e., “extreme") earnings surprise decile and there is nothing “surprising” about this. However, prior work has shown that being in the extreme deciles precedes rather dramatic abnormal returns (Bernard and Thomas 1990; Abarbanell and Bernard 1992; Livnat and Mendenhall 2006; Doyle et al. 2006), suggesting there may be a reason (other than chance) as to why these firm-quarter observations fall in the extreme earnings surprise deciles.1 The only other research focusing explicitly on extreme earnings surprises is Kasznik and Lev (1995). They examine firms‟ disclosure activity prior to large earnings surprises in the fourth quarter in 1988, 1989 and 1990. They focus mostly on large negative surprises, finding that firms who issue earnings warnings tend to be larger, in high tech industries, and have issued 1
Specifically, Bernard and Thomas (1990) find that post-earnings announcement returns are eight percent higher for firms in the top decile of earnings news relative to firms in the bottom decile (Table 2); Abarbanell and Bernard (1992) document a one-year hedge portfolio return between the top and bottom deciles of earnings surprise of eight percent (Figure 1); and Doyle et al. (2006) document a three-year hedge portfolio return of between the top and bottom deciles of earnings surprise of twenty-four percent (Table 2).
3
management guidance previously in the quarter. These disclosures appear to be driven by litigation concerns. Possibly for this reason, bad news warnings outnumber good news warnings two to one. Nonetheless, given the harm suffered by shareholders who sell their shares prior to a large positive earnings surprise, the authors conclude by wondering why managers did not disclose their firms‟ large positive earnings surprise earlier. We explore three reasons why a firm might have a positive extreme earnings surprise (PEES). Our first hypothesis explores the incentives of managers. PEES firms tend to be smaller firms with less analyst coverage and lower institutional investor ownership (Doyle et al. 2006). We hypothesize that managers of these neglected stocks might be attempting to garner investor attention by withholding positive information from analysts in order to later announce “surprising” positive earnings results and garnish attention from the press. Investor relations departments of small firms are constantly struggling to attract the attention of analysts and hoping that those analysts highlight them in their reports to investors. In short, management‟s goal is to be in the Earnings Surprise column in the Wall Street Journal and that this is an effective way to get analysts to take a closer look at their company. Odean (1999) proposes that investors limit their investment decisions to stocks that have recently caught their attention. Barber and Odean (2008) test this proposition by examining attention-based decision making in the context of equity purchases and sales for retail investors. Given the large number of stocks available to possibly invest in and the information search costs of researching so many firms, the authors propose that investors only consider purchasing stocks that have recently caught their attention, and that this effect will be greater for individual investors who have fewer information search resources than institutional investors. The authors find that individual investors display attention-driven buying behavior – individuals are net
4
buyers when stocks are mentioned in the news, on high-volume days, and on the day after both extremely negative and extremely positive one-day returns, and find that attention-driven buying is similar for both large and small capitalization stocks. Busse and Green (2002) may provide the most compelling direct evidence of attention-based investing, using TAQ data to document abnormally high trading volume when investor attention to a stock increases. As an example, when Maria Bartiromo mentions a stock during the CNBC television show Midday Call, average trading volume for this stock increases nearly five-fold during the few minutes following the mention. Our first hypothesis is that managers, aware that increases in investor attention are associated with increases in investment, use large positive earnings surprises to garner investor attention. While it is always difficult to establish managerial intent, we find that PEES firms tend to have “neglected” firm characteristics. In particular, in the year prior to a PEES, PEES firms experience return on assets above the industry median yet stock price increases below the industry median, suggesting that PEES firms experienced good performance news that was not fully appreciated by the market. PEES firms also have larger increases in return on assets during the four quarters following the PEES. Further, these same variables do not predict negative extreme earnings surprises. Finally, these firms are less likely to have issued EPS guidance for the quarter in which the PEES occurred, indicating that the resulting earnings “surprise” may be the result of management withholding favorable information until the earnings announcement date. We also show that management‟s attempt to attract attention is ultimately successful. We document a significant increase in the number of analysts following the firm, the percentage of shares held by institutions, and short-term and long-term trading volume subsequent to a PEES quarter.
5
Our second hypothesis is that PEES occur because analysts are not paying attention to the positive developments happening at the firm. Analysts may not be sufficiently interested in the stock to put in the effort necessary to stay abreast of management communications and other firm and industry events that have the potential to make the firm‟s upcoming earnings more predictable. If this hypothesis is descriptive, the variables that proxy for analyst costs and benefits should predict both large positive and large negative earnings surprises. We find that relative to non-PEES firms, PEES firms have lower analyst following, and the analysts who are following the firm are busier (defined as following a greater number of other firms) and issue individual earnings forecasts with greater disagreement between analysts forecasting for the same firm.2 Our third and final hypothesis – the natural variation in firm performance hypothesis – serves as a benchmark for the other two hypotheses. It could simply be that neither party in the earnings announcement game knew anything about the impending large positive earnings surprise; rather, exogenous shocks to the firm‟s operating performance generated the unexpectedly good earnings realizations. Consistent with this hypothesis, we find that PEES are more likely for firms with volatile prior quarterly return on equity and higher operating leverage, both of which make accurately forecasting quarterly earnings a more challenging endeavor. Recognizing that these three hypotheses are not mutually exclusive, we attempt to “apportion blame” between the three hypotheses by analyzing the changes in the odds ratio for different sets of independent variables. We examine how a one unit change in binary variables 2
As discussed in greater detail later, managers and analysts may have conflicting incentives to forecast and disclose negative earnings surprises. While analysts have incentives to forecast negative events as soon as possible, managers may have incentives to delay the dissemination of bad information as long as possible (Kothari et al. 2009). We compare the determinants of large positive earnings surprises to the determinants of large negative earnings surprises in our empirical analysis to further distinguish between our three hypotheses. We predict that while the latter two hypotheses likely predict both positive and negative large earnings surprises, the managerial manipulation hypothesis will predict large positive but not large negative earnings surprise.
6
and a one standard deviation change in continuous variables influence the odds of observing a PEES. We find that such a change in the managerial manipulation proxies doubles the odds of a PEES firm-quarter, after controlling for firm size, institutional ownership, and all the other hypothesized variables in the model. Similarly, a one standard deviation change in the analyst inattention variables doubles the odds of a PEES occurring, after controlling for all other variables in the model. The natural variation variables turn out to be less important: a one standard deviation change in these proxies increases the odds of a PEES occurring by a factor of only 1.11, after controlling for all other variables. Finally, a one unit or one standard deviation change in all the variables in the full model yields an impressive nine-fold increase in the odds of a positive extreme earnings surprise. The paper proceeds as follows. Section II provides hypothesis development and variable measurement. Section III presents the results and Section IV concludes.
II. HYPOTHESIS DEVELOPMENT AND VARIABLE MEASUREMENT Defining a Positive Extreme Earnings Surprise Before developing our hypotheses we describe how “surprising” it is for firm-quarter observations to be in the top earnings surprise decile. We define quarterly earnings surprise (SURPRISE) as the IBES actual EPS for a firm-quarter less the IBES analyst consensus EPS forecast for the firm-quarter, divided by the firm‟s stock price per share at the end of the fiscal quarter, where the IBES actual is unadjusted for subsequent splits and the IBES consensus EPS forecast is the most recent median forecast preceding the firm‟s earnings announcement date. 3 We use EPS figures unadjusted for subsequent stock splits based on findings that using split3
Note that this is the definition of an analyst forecast error commonly used in the literature. We choose to label this measure „earnings surprise‟ because labeling the deviation as an analyst forecast error implies that it is the analysts who are the cause of the difference between actual and forecasted EPS.
7
adjusted data can potentially distort both time-series and cross-sectional characteristics of earnings surprises (Baber and Kang 2002; Payne and Thomas 2003). The unsplit-adjusted database gives the earnings per share that was actually reported in the company‟s earnings announcement, which is the news that the market observed at that time. We rank firm-quarter observations by SURPRISE in each calendar quarter and focus on observations in the top decile. The indicator variable PEES_D is set equal to one when a firmquarter is in the top decile of SURPRISE for that calendar quarter, and set equal to zero otherwise. By sorting observations within each calendar quarter to determine when a PEES has occurred, as opposed to sorting the entire pooled sample over the time period examined, we control for changing economic conditions. Without this calendar quarter restriction, quarters with large unexpected improvements in economy-wide performance would dominate the PEES sample, and the answer to why firms have large positive earnings surprises would revolve around predicting when the economy will experience unexpected improvements. The SURPRISE value cutoff for the top decile changes each quarter, with the lowest cutoff value of 0.0032 in the first quarter of 1998 and the highest cutoff value of 0.0159 in the third quarter of 1984 (untabulated). Pooling over all quarters, the median SURPRISE value cutoff is 0.01 (i.e., a quarterly earnings surprise equal to 1 percent of stock price).4
4
Doyle et al. (2006) compare the earliest available IBES actual EPS in the unsplit-adjusted database with the actual press release found through LexisNexis for fifty firms and find that the data from the two sources matched in all fifty cases. In addition, Doyle et al. (2010) match 1,000 random quarterly observations from 1997 through 2000 from the IBES unsplit-adjusted database to the press release issued by firms via a Lexis-Nexis search and find that IBES corresponds perfectly with the press release in 915 cases. Given this high level of documented accuracy, we do not believe that large systematic errors in the IBES database are driving our results.
8
Managerial Manipulation Hypothesis This hypothesis is based on the premise that managers of “neglected” firms want to attract investor attention with large positive earnings surprises. Feeling neglected is not an irrelevant concern; if the amount of attention a firm receives influences its stock price, then the manager of a neglected firm has an incentive to draw attention to the firm. In Merton‟s limited attention CAPM (1987) model, investors only form diversified portfolios from the set of firms that they know about. And, as previously discussed, Odean (1999) proposes that investors do not evaluate each of the thousands of stocks actively trading in public markets when making their investment decisions. Rather, investors limit their investment decisions to stocks that have recently caught their attention. If a firm is neglected, it is less likely to be included in many of these portfolios, which increases its cost of capital. Further, the greater the firm‟s idiosyncratic risk, the more it stands to benefit from an increase in investor recognition. Lehavy and Sloan (2008) provide some evidence consistent with this model, finding that increases in institutional ownership are associated with decreases in expected returns. More directly, Brown et al. (2009) find that the bid-ask spread and the probability of informed trade both decline in the year following large positive earnings surprises, and interpret this as evidence that the firm has increased its visibility to the market. Similarly, Irvine (2003) finds that firms experience a significant increase in liquidity following the initiation of analyst coverage of a stock, and the effect increases with the analyst‟s recommendation. Barber and Odean (2008) find that individual investors tend to buy stocks that grab their attention – i.e., those mentioned in the news, with high daily trading volume, with an extreme daily return, etc. These same investors do not exhibit attention-based selling behavior because they tend to only sell stocks that they already own.
9
Management has various means to attract investor attention. For instance, Bushee and Miller (2007) find that small firms who hire an investor relations consulting firm enjoy increased media coverage, increased analyst following and increased institutional ownership. In a similar vein, Grullon et al. (2004) find that firms with greater annual advertising spending have a larger number of both individual and institutional investors. Our first hypothesis considers whether managers create large positive earnings surprises in an attempt to attract investor and analyst attention. Similar to managers attempting to put their best foot forward by strategically highlighting the prior period earnings amount that provides the lowest possible benchmark when announcing current period earnings (Schrand and Walther 2000), we hypothesize that managers attempt to gain attention by announcing a very large favorable earnings surprise at a time when the firm‟s strong earnings performance allows managers to put their best foot forward. Did management intentionally withhold information or intentionally guide analyst forecast downward in order to create a positive earnings surprise? Such behavior will be effective if it causes investors to investigate the firm further and favorably change investors‟ beliefs about firm value. Further, manufacturing a positive earnings surprise is not an easy strategy for fundamentally bad firms to imitate; the bad firms would have to guide analysts to extremely low expectations, and would have to withstand the subsequent increased attention and scrutiny by investors. Besides being neglected (as measured by analyst coverage and institutional ownership), does the manager have a reason in this particular quarter to seek investor attention?5 Teoh and Hwang (1991) provide an analytical model in which a firm's strategic timing of disclosure is informative to investors in revealing firm type (high or low). Firm type refers to the 5
In this paper, we do not use direct measures of earnings management as a way to test if managers are exceeding benchmarks to seek attention. The primary reason for this exclusion is that 1) measures of earnings management are extremely noisy and imprecise and 2) managers may use other methods of achieving this objective that are not captures through tradition earnings management measures; for instance, they may alter the definition of earnings through pro forma reporting (Doyle et al. 2003).
10
propensity of the firm to receive good news, with high-type firms having a higher propensity for good news than low-type firms. Firm type is known by the managers of the firm but is infeasible to directly and credibly disclose to investors. The model predicts that high-type firms prefer to withhold favorable information in period t while low-type firms prefer to immediately disclose favorable information in period t. This strategic disclosure occurs because a firm that receives favorable information in period t and expects additional favorable information in period t+1 can signal its confidence in its future favorable information by waiting until after period t to disclose the favorable information. In contrast, a firm that receives favorable information in period t but does not expect additional favorable information in period t+1 will immediately disclose its period t favorable information to prove it is not the worst of the low-type firms in period t. Because high-type firms are more likely to receive future good news than low-type firms, a separating equilibrium is obtained through the timing of disclosure of favorable information (i.e., when investors observe a firm disclosing favorable information about period t in period t, investors infer the firm is of low-type), even in the absence of exogenous disclosure costs. The authors' prediction that hightype firms can use the withholding of favorable information to signal their high type is consistent with our prediction that managers of PEES firms will wait to reveal positive and unexpected earnings information to investors at the earnings announcement. Because we cannot directly observe management‟s intentions, we look for indirect evidence to assess this hypothesis. We proxy for managerial intent to attract attention with three variables. All variables are defined in detail in the Appendix. The first variable is the change in mean industry-adjusted return on assets (CH_ROA) from four quarters prior to the PEES quarter (t-4 to t-1) to the four quarters following the PEES quarter t (t+1 to t+4), where return on assets
11
(ROA) is measured as net income divided by total assets (Compustat item niq divided by item atq), less the industry median for the same period. Quarter t is defined as the quarter in which a PEES occurs, and industries are defined using the Fama-French 48 industry classification (Fama and French 1997). This variable was used in Lang and Lundholm (1993) to proxy for the private information managers have about the firm‟s expected future performance. The idea is that managers know how much expected good or bad news the future holds, and this will influence how much attention the managers want from the analyst community. We predict a positive association between CH_ROA and PEES_D. The second variable is designed to capture situations where the manager might feel that the firm‟s strong past earnings performance was not adequately rewarded by the stock market. We set the binary variable NOLOVE_D equal to one if both of the following are true (and set equal to zero otherwise): i) the firm‟s average ROA in periods t-4 through t-1 is greater than the industry median ROA over the same period, and ii) the firm‟s prior year market-adjusted buyand-hold stock return is less than the industry median market-adjusted return over the same period. The prior year stock return is measured beginning 252 trading days and ending two trading days before the firm‟s quarterly earnings announcement date, and returns are marketadjusted by subtracting the return on a value-weighted market portfolio from a firm‟s raw return. We predict a positive association between NOLOVE_D and PEES_D. The third variable, labeled GUIDE_D, is a binary variable set equal to one if management provided any type of EPS guidance for period t earnings prior to the earnings announcement date, and set equal to zero otherwise. If management provides guidance it is unlikely the manager is attempting to manufacture an extreme earnings surprise, as there are negative
12
consequence to managers who miss their own earnings forecast (Lev and Kasznik 1995). Consequently, we predict a negative association between GUIDE_D and PEES_D.
Analyst Inattention Hypothesis Our second hypothesis focuses on analysts and the quality of their forecasts, as it is possible large earnings surprises occur because analysts made particularly poor forecasts. Forecast accuracy has been shown to increase with the analyst‟s general experience, firmspecific experience, employment at a top-ten brokerage house, and the number of forecast revisions; and decrease with the number of firms and industries an analyst follows and the staleness of the forecast (Mikhail et al. 1997, 1999; also see Ramnath et al. 2008 for an excellent literature review). Assuming that the analyst labor market allocates the most time and talent to the firms that are most valuable to the analysts, it could be that large earnings surprises are associated with analysts and firms where the benefit of being more accurate is simply too low to justify the cost. Consistent with the idea that analysts rationally trade off the costs and benefits of expending effort on forecast accuracy, Alford and Berger (1999) find that both analyst following and forecast accuracy increase with the firm‟s trading volume (and thus trading commissions).
Similarly, Lang and Lundholm (1996) find that forecast accuracy increases as
the cost of collecting information declines, where the cost of collection is measured by the firm‟s disclosure activity. Our second hypothesis is that large earnings surprises occur for firms where analyst attention is low (i.e., firms where analysts expect the cost of being more accurate to exceed the benefit). We use five variables to proxy for analyst inattention. The first is the number of analysts issuing forecasts included in the most recent IBES analyst consensus EPS forecast prior to the
13
earnings announcement date (ANALYST_CNT).
This is a good summary measure of the
equilibrium supply of analyst effort allocated to the firm. We expect ANALYST_CNT to be negatively associated with PEES_D. The next two variables relate to properties of the consensus forecast itself. FORECAST_STALE is the number of days between the most recent IBES analyst consensus EPS forecast and the earnings announcement date. All FORECAST_STALE values are positive by construction as we require the analyst consensus estimate date to be before the earnings announcement date. The greater the period of time between the two dates, the greater the chance that new information regarding the earnings realization is not incorporated into the consensus forecast. We therefore expect this variable to be positively associated with PEES_D. The standard deviation of individual analyst EPS forecasts that comprise the consensus forecast, labeled FORECAST_SD, has been used in prior research as a measure of analyst uncertainty (Baron et al. 1998; Burgstahler and Chuk 2009). We expect this variable to be positively associated with PEES_D. If there is only one forecast in the consensus, which occurs 24 percent of the time, FORECAST_SD is set to zero; there is no disagreement among analysts in this case. The next two variables relate to properties of the analysts themselves. ANALYST_BUSY is defined as the average number of other firms each analyst is following in the quarter. As another measure of the supply of analyst attention, we predict ANALYST_BUSY will be positively associated with PEES_D. Finally, ANALYST_EXP is the average number of years the analysts who contribute forecasts to the consensus in the quarter have been providing earnings forecasts captured by IBES. Given that prior research has found that more experienced analysts are more accurate, we expect ANALYST_EXP to be negatively associated with PEES_D (Ramnath et al. 2008).
14
The interpretation of the managerial manipulation variables sometimes overlaps with the interpretation of the analyst inattention variables. This is not surprising, as the managerial manipulation hypothesis posits that managers feel their firm is neglected and attempt to manufacture investor attention through PEES, while the analyst inattention hypothesis measures how neglected the firm actually is. If extreme positive earnings surprises are simply due to analyst inattention, however, then the analyst inattention variables should predict both positive and negative extreme earnings surprises. This, however, should not be the case for the managerial inattention variables. Accordingly, we create the binary variable NEES_D that is set equal to one when the earnings surprise is in the lowest decile (i.e., largest negative SURPRISE values) for the calendar-quarter, and set to zero otherwise. By contrasting the relation between our independent variables and PEES_D versus NEES_D, we can partially distinguish between the managerial manipulation and analyst inattention hypotheses.6
Natural Variation Hypothesis Our third hypothesis is that a PEES is neither managers‟ nor analysts‟ fault; rather, the firm‟s operations are simply volatile and generate high-variance earnings from quarter to quarter. We proxy for this effect with the variable SD_ROE, defined as the standard deviation of return on equity computed over the four quarters prior to the PEES quarter t (t-4 to t-1), where return on equity (ROE) is measured as net income divided by book value of equity (Compustat item niq divided by item seqq). In addition, the earnings of firms with high operating leverage (i.e., high fixed costs as a percent of total costs) are inherently difficult to forecast because most of these firms‟ profit (or lack of profit) occurs in the last few days of the quarter. For these firms, neither
6
Note that by construction, when PEES_D = 1, NEES_D = 0 and vice versa. Consequently, a variable that predicts PEES_D = 1, will weakly predict NEES_D = 0, and vice versa.
15
the firm nor the analyst may know what the quarter will hold until the last few days of the quarter.
Following Darrat and Mukherjee (1995), we measure operating leverage, labeled
OP_LEV, as the slope coefficient from regressing the change in earnings before interest and taxes (Compustat items niq + xintq + txtq) on the change in sales (Compustat item revtq), estimated via rolling regressions using data from quarters t-20 to t-1. Both SD_ROE and OP_LEV are expected to be positively associated with PEES_D. As both variables capture the uncertainty about the firm‟s upcoming earnings results, both variables should predict NEES_D=1 as well as PEES_D=1 (similar to the analyst inattention variables). Finally, we include two control variables in all regression specifications. The percent of institutional ownership, INST_OWN, measured as the percentage of outstanding shares held by institutional investors at the end of the prior quarter and is obtained from the Thomson-Reuters Institutional 13f Holdings database. This variable is another measure of the market‟s interest in the firm apart from analyst coverage. Finally, we include MVE (the natural log of the firm‟s market value (Compustat items prccq*cshoq)) as our second control variable because prior research finds that larger firms disclose more and have more analyst coverage (Lang and Lundholm 1993, 1996). Further, Vuolteenaho (2002) argues that small firms have greater idiosyncratic risk and consequently have more to gain by increasing their investor recognition. Both control variables are expected to be negatively correlated with PEES_D.
16
III. RESULTS Descriptive Statistics Our initial sample consists of all firm-quarters between July 1984 and July 2008 with sufficient data to compute the variable SURPRISE for the quarter.7 Our initial sample includes 257,832 firm-quarter observations, with the final number of observations included in each regression specification ranging from 130,080 to 243,350 based on which independent variables are required.8 Panel A of Table 1 presents the descriptive statistics for all the variables in our study. During the typical firm-quarter, management does not provide any type of EPS guidance, as only 15.1 percent of the observations have GUIDE_D=1. Because CH_ROA is industry median adjusted, the median value is zero and the mean value of -0.002 is very close to zero. On average, 5.2 analysts provide an EPS estimate (ANALYST_CNT) and the consensus forecast is issued 20.5 days before the earnings announcement (FORECAST_STALE). The average analyst follows 12 firms (ANALYST_BUSY) and has been issuing forecasts as recorded in IBES for 5.3 years (ANALYST_EXP). Institutional ownership averages 47 percent (INST_OWN) and a dollar
7
The earliest period in which IBES reports analyst earnings forecast consensus data is March 1984. Our analysis begins with July 1984 because we require a minimum of 500 earnings surprise observations per calendar quarter to compute SURPRISE. 8 We re-estimate the four regressions presented in Table 3 (when PEES_D is the dependent variable) after conducting the following four sensitivity checks. 1) The data used to construct the variable GUIDE_D comes from Thomson First Call‟s Company Issued Guidance database. Because Chuk, Matsumoto and Miller (2009) provide evidence that this database‟s coverage may be unreliable prior to 1998, we restrict the sample to post-1997 firmquarter observations; 2) The majority of firm-quarter observations in our sample (85.53 percent) have fiscal quarterends in March, June, September, and December. In our main analysis we group observations by calendar quarter (i.e., combining observation with January, February, and March fiscal quarter-ends, April, May, and June fiscal quarter-ends, etc.) to insure a sufficiently large quarterly sample pool. We eliminate observations with non-March, June, September, or December quarter-ends to ensure these observations are not driving our results; 3) To consider any possible confounding effects of Regulation Fair Disclosure‟s implementation in August 2000, we restrict the sample to post-2000 firm-quarter observations; and 4) To consider whether systematic differences between negative and positive earnings surprise observations are driving our results, we eliminate all negative earnings surprise firmquarter observations. There are no qualitative or quantitative changes in the results for any of these four sensitivity checks.
17
change in sales results in a 0.298 dollar change in earnings before interest and taxes, on average (OP_LEV). Panel B of Table 1 shows descriptive statistics for the subsamples where PEES_D=1, NEES_D=1 and the remaining middle 80 percent of the earnings surprise distribution.
The
results indicate that SURPRISE exhibits two of the same asymmetries documented in Abarbanell and Lehavy (2003). Large negative earnings surprises are more extreme than large positive earnings surprises, as evidenced by a median SURPRISE value of 0.010 for the PEES sample and a median SURPRISE value of -0.026 for the NEES sample. In untabulated results we also find that there is a greater number of small positive surprises (49.24 percent of firm-quarter observations with 0.00 < SURPRISE > < > > < > < < > >
Middle 80 Percent Sub-sample Mean -0.001 0.168 -0.002 0.272 5.846 18.172 12.136 5.389 0.026 0.486 6.367 0.052 0.293
S.D. 0.005 0.374 0.028 0.445 5.234 21.573 7.055 2.761 0.042 0.286 1.700 0.207 0.675
P50 0 0 0 0 4.000 14.000 11.625 5.276 0.010 0.484 6.243 0.013 0.240
(B)
> > > > > < < > < > > <
