Underreaction to Industry-Wide Earnings and the Post-Forecast Revision Drift

Transcription

1 DOI: / X Journal of Accounting Research Vol. 00 No Printed in U.S.A. Underreaction to Industry-Wide Earnings and the Post-Forecast Revision Drift KAI WAI HUI AND P. ERIC YEUNG Received 15 February 2011; accepted 2 January 2013 ABSTRACT We test whether the post-forecast revision drift is mainly attributable to investors underreaction to industry-wide earnings news conveyed by analysts forecast revisions. We find a large drift associated with industry-wide earnings news but no drift associated with firm-specific earnings news. Consistent with the functional fixation hypothesis, we provide evidence that the postforecast revision drift is driven by investors underreaction to the higher persistence of industry-wide earnings. Although prior research has focused on differential persistence of earnings components stemming from managerial reporting discretion, we provide evidence suggesting that investors do not fully understand the differential earnings persistence attributable to industry fundamentals. 1. Introduction Revisions of earnings forecasts by individual financial analysts play a significant role in the dissemination of information about corporate earnings. Department of Accounting, Hong Kong University of Science and Technology; Samuel Curtis Johnson Graduate School of Management, Cornell University. Accepted by Douglas Skinner. We gratefully acknowledge the comments of Abbie Smith (prior editor), an anonymous referee, Linda Bamber, Sanjeev Bhojraj, Dan Dhaliwal, Gilles Hilary, Bin Ke, Karen Nelson, and workshop participants at the Chinese University of Hong Kong, Cornell University, Hong Kong University of Science and Technology, INSEAD, McMaster University, Nanyang Technology University of Singapore, and Rice University. 1 Copyright C, University of Chicago on behalf of the Accounting Research Center, 2013

2 2 K. W. HUI AND P. E. YEUNG Prior research has documented a large drift following analysts earnings forecasts, suggesting that investors are slow in processing earnings news embedded in analysts forecasts (e.g., Gleason and Lee [2003]). 1 However, it is not entirely clear why investors consistently underreact to the information in analysts earnings forecasts, especially given that this drift is of long duration and exists among large firms. Recent evidence indicates that analysts earnings forecasts convey significant news about industry-wide fundamentals (e.g., Piotroski and Roulstone [2004]). A natural conjecture is that investors consistently underreact to news embedded in analysts forecasts about industry-wide fundamentals (hereafter industry-wide earnings news ). Because industry-wide earnings tend to be more persistent than firm-specific earnings, the prediction of underreaction to industry-wide earnings news follows naturally from the functional fixation hypothesis, which suggests that investors fail to fully appreciate the differential persistence of earnings components (e.g., Sloan [1996]). To put it differently, we view earnings news as signals about industry and firm fundamentals and examine whether investors delayed reactions to analysts forecasts can be mainly attributed to their underreaction to industry-wide earnings news. 2 More specifically, the decomposition of analysts forecast revisions into industry-wide earnings news and firm-specific earnings news is motivated by two factors. First, we expect the post-forecast revision drift to be closely related to how investors process the news about industry-wide fundamentals because analysts forecast revisions provide rich industry-wide information (Piotroski and Roulstone [2004], Chan and Hameed [2006]). Unlike insiders who have access to firm idiosyncratic information, analysts are outsiders and therefore more likely to gain comparative advantages by focusing their efforts on obtaining and mapping industry-wide information into earnings (Hutton, Lee, and Shu [2012]). Second, we expect significant differential efficiency in pricing industrywide and firm-specific earnings news. Economic theory suggests that industry-wide earnings are more persistent than firm-specific earnings due to intense within-industry competition (e.g., Waring [1996]). We therefore expect that the industry-wide component of a forecast is more informative about future earnings than the firm-specific component, even though the 1 Prior research has documented robust evidence of drift in stock prices following analysts earnings forecasts (Givoly and Lakonishok [1980], Stickel [1991], Chan, Jegadeesh, and Lakonishok [1996], Gleason and Lee [2003], Zhang [2006]). Gleason and Lee [2003] suggest that the delayed market response to earnings news is one of the most perplexing anomalies to emerge from accounting-based capital market research over the past 20 years (p. 194). 2 Following the literature (Brown and Ball [1967], Ayers and Freeman [1997], Elgers, Porter, and Xu [2008]), industry-wide earnings represent the common component of earnings of all firms in the same industry, whereas firm-specific earnings are the deviations of individual firms earnings from the industry average. As in prior studies, we are less concerned about market-wide earnings because (1) market average earnings have little impact in crosssectional settings and (2) market-wide forces ultimately manifest themselves at the industry and firm levels.

3 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 3 forecast itself is presented and appears to investors as firm specific. If investors are functionally fixated on the forecasted earnings and do not fully appreciate the differential persistence of earnings components (e.g., Sloan [1996]), prices are likely to underreact to industry-wide earnings news. We test underreaction to industry-wide earnings news with a large sample of earnings forecasts in analyst industry reports. An industry report is issued by an analyst who follows multiple firms in the same industry and produces forecasts for these firms simultaneously based on common industry fundamentals (see the online appendix for an example). We use the forecasts in industry reports to minimize the measurement errors in deriving both industry-wide and firm-specific earnings news (e.g., Bhojraj, Lee, and Oler [2003]). 3 Because these industry analysts endogenously choose to follow a set of firms subject to common industry-wide shocks, aggregating earnings of these firms should yield a much cleaner measure of industry-wide earnings than grouping earnings of firms based on broad industry classifications. In addition, because these analysts use a top-down industry perspective to estimate firm performance, their forecasts reflect the efforts to map common industry fundamentals into earnings estimates. We collect analyst industry reports from Investext for firms in the manufacturing sector for the period To ensure that these forecasts were indeed disseminated widely to investors, we manually match these forecasts with the forecasts recorded by I/B/E/S through the 2006 Brokerage-Translation File. Our final sample covers 25,195 annual earnings forecast revisions for 4,098 firms. We define an analyst forecast revision as the earnings forecast in industry report minus the same analyst s prior earnings forecast (Gleason and Lee [2003]), scaled by lagged stock price. Industry-wide earnings news is measured as the average of forecast revisions for firms in the same industry report. Firm-specific earnings news is defined as the difference between a forecast revision and industry-wide earnings news. Our results are striking. We find a large drift associated with industrywide earnings news in analysts forecast revisions. On the other hand, we find no drift associated with firm-specific earnings news. These results indicate that, on average, the post-forecast revision drift is mainly driven by investors underreaction to industry-wide earnings news rather than firmspecific earnings news. More specifically, we document that the association between forecasted industry-wide earnings and future earnings is greater than the 3 Reducing measurement error in industry-wide and firm-specific earnings is a pressing issue in this line of inquires (e.g., Bhojraj, Lee, and Oler [2003]). For instance, although Ayers and Freeman [1997] find a post-earnings announcement drift associated with firm-specific earnings but not with industry-wide earnings, Elgers, Porter, and Xu [2008] do not find drift associated with either component. Elgers, Porter, and Xu [2008] point out that measurement errors in industry-wide earnings constructed based on Standard Industry Classification (SIC) codes are likely to significantly reduce the power of the empirical tests.

4 4 K. W. HUI AND P. E. YEUNG association between forecasted firm-specific earnings and future earnings. However, stock prices behave as if investors significantly underweight forecasted industry-wide earnings. In addition, we find that investors initial reactions to analysts forecast revisions place similar weights on industry-wide earnings news and firm-specific earnings news, which are followed by a drift associated with industry-wide earnings news. This drift is more pronounced when industry-wide earnings are more persistent than firm-specific earnings, consistent with the functional fixation hypothesis in which investors fail to fully appreciate the higher persistence of industry-wide earnings. Finally, we find that the drift disproportionally concentrates on subsequent earnings announcements, supporting the prediction that the drift is rooted in investors delayed reactions to information about future earnings other than risk-related considerations. We further exploit cross-sectional variation in the post-forecast revision drift. We find a more pronounced drift when the firm s earnings are tied more closely to industry-wide profits and when industry structure is likely to facilitate more persistent industry-wide economic rents. These crosssectional results corroborate the main finding that the source of the drift is investors underreaction to industry-wide earnings news. Furthermore, we find a more pronounced drift when arbitrage costs are higher, suggesting that the limits of arbitrage play an important role in explaining the survival of the drift. Our study contributes to the literature on the post-forecast revision drift (Givoly and Lakonishok [1980], Stickel [1991], Chan, Jegadeesh, and Lakonishok [1996], Gleason and Lee [2003], Zhang [2006]) by showing the information source of investors delayed reaction. We find that the drift is largely driven by investors underreaction to industry-wide earnings news and that there is little drift associated with firm-specific earnings news on average. Thus, investors functional fixation plays a more important role in explaining the post-forecast revision drift. 4 Our study also adds to the literature suggesting that investors likely misprice certain earnings components (e.g., Sloan [1996], Xie [2001], Hanlon [2005], Richardson et al. [2005], Hafzalla, Lundholm, and Van Winkle [2011]). Although prior research has focused on managerial reporting discretion as the primary reason for the differential persistence of earnings components, we provide evidence that investors do not seem to fully understand the differential persistence of earnings components attributable to economic forces. Finally, our study contributes to the literature on the value of analysts forecasts in conveying industry-level earnings information (Piotroski and Roulstone [2004], Chan and Hameed [2006]). Although prior research 4 The results of our study also suggest that the analyst-based post-earnings announcement drift (e.g., Doyle, Lundholm, and Soliman [2006], Livnat and Mendenhall [2006], Ayers, Li, and Yeung [2011]) may be related to investors underreaction to industry-wide information in analysts forecasts.

5 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 5 shows that analysts forecasts appear to increase the relative amount of industry information in stock prices, it does not examine whether prices incorporate industry-level news in an efficient manner. We extend this literature by showing evidence that investors underreact to industry-wide earnings news in analysts forecasts. The remainder of our paper is organized as follows: section 2 reviews the related literature and develops our main hypotheses, section 3 outlines the sample selection and provides descriptive statistics, section 4 reports results of hypothesis testing, section 5 discusses additional empirical analyses, and section 6 concludes. 2. Related Literature and Hypotheses 2.1 UNDERREACTION TO SIGNALS FROM SECURITY ANALYSTS Evidence that analysts forecasts are informative to investors dates back to the 1970s (Griffin [1976], Givoly and Lakonishok [1979, 1980], Elton, Gruber, and Gultekin [1981], Imhoff and Lobo [1984]). Subsequent research shows that the short-window price reaction around forecast revisions is incomplete (Givoly and Lakonishok [1980], Stickel [1991], Chan, Jegadeesh, and Lakonishok [1996]). Unlike these prior studies which mostly focus on changes in consensus forecasts, Gleason and Lee [2003] examine price reaction to individual analysts forecast revisions and document cross-sectional differences in the post-forecast revision drift. They find that analyst visibility and coverage mitigates the drift. In a similar vein, Zhang [2006] finds that the drift is greater when information about a firm is scarcer. Prior research, however, has not viewed earnings news as a signal on fundamentals and examined whether news about industry-wide or firm-specific fundamentals causes investors underreaction. A separate stream of literature also documents a drift following analysts stock recommendations (Elton, Gruber, and Grossman [1986], Stickel [1995], Womack [1996], Barber et al. [2001], Jegadeesh et al. [2004], Howe, Unlu, and Yan [2009]). Barber et al. [2001] show that a strategy of buying stocks with the most favorable recommendations and selling stocks with the least favorable recommendations generates significant abnormal returns. Jegadeesh et al. [2004] demonstrate that changes in quarterly analyst recommendations are robust predictors of future stock returns. Although Jegadeesh et al. [2004, p. 1118] recognize potential industry-related effects, they do not examine them in the recommendation setting. More recently, Howe, Unlu, and Yan [2009] analyze the information content of aggregate analyst recommendations and find modest evidence that analysts stock recommendations predict future industry returns. This result suggests that investors may underreact to information about industry fundamentals in analysts recommendations. The primary objective of our study is to examine whether and why investors fail to fully appreciate industry-wide earnings news in analysts reports.

6 6 K. W. HUI AND P. E. YEUNG 2.2 PRIOR EVIDENCE OF FUNCTIONAL FIXATION ON EARNINGS SIGNALS We hypothesize that investors functional fixation on earnings signals is an important explanation for the post-forecast revision drift. Two prominent forms of functional fixation bias on earnings signals in the literature are the post-earnings announcement drift and the accrual anomaly. 5 The former hypothesizes that investors are functionally fixated on naive random-walk earnings signals and fail to appreciate the more complex time-series properties of earnings (Jones and Litzenberger [1970], Foster, Olsen, and Shevlin [1984], Bernard and Thomas [1989, 1990], Abarbanell and Bernard [1992], Ayers, Li, and Yeung [2011]). The latter focuses on investors fixation on total earnings and failure to fully appreciate the differential persistence of earnings components (Sloan [1996], Xie [2001], Richardson et al. [2005]). Similar to the accrual anomaly literature, we argue that investors are functionally fixated on forecasted earnings and fail to fully appreciate the differential informativeness of industry-wide and firm-specific components in forecasted earnings. Although a forecast is presented and appears to investors as firm specific, the industry-wide component in this forecast is more informative than the firm-specific component EMPIRICAL PREDICTIONS Implications and Pricing of Forecasted Industry-Wide Earnings. Economic theory has long suggested that abnormal profits above or below the industry norm are likely to whittle away due to competition (Mueller [1977, 1986, 1990], Waring [1996]). Industry fundamentals that determine firm performance (e.g., consumer taste, production technology, and regulatory environment) are relatively long-lasting attributes. On the other hand, firms competitive positions within an industry are much more dynamic. Although more successful firms tend to lose their competitive edge due to new entrants and learning of other firms, unsuccessful firms tend to improve their performance by imitating others and taking corrective actions (Waring [1996]). Therefore, industry-wide economic rents are more sustainable than firm-specific profits. 5 More general forms of functional fixation unrelated to drifts date back to much earlier research on accounting choices (e.g., Ball [1972], Watts and Zimmerman [1986]). 6 A potential competing explanation for the post-forecast revision drift is the information availability hypothesis. Specifically, because investors can obtain industry-wide information from various sources but have only limited access to firm-specific information, they may, therefore, be more likely to underreact to firm-specific earnings news (Ayers and Freeman [1997]). However, it is unlikely that this hypothesis explains the drift because it views informational friction as the root cause of the drift. The firms followed by analysts have above-average information environment, and any temporary inefficiency due to informational friction should generally be short-lived. In contrast, the post-forecast revision drift lasts over 200 trading days and is more likely to be caused by persistent bias in processing earnings information.

7 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 7 Financial analysts forecasts contain significant industry-wide information (Piotroski and Roulstone [2004], Chan and Hameed [2006]), because analysts are outsiders who generally have less access to firm-level, idiosyncratic information than insiders and are therefore likely to focus their efforts on obtaining industry- and market-level information and mapping that into prices. 7 To the extent that analysts earnings forecasts are noisy measures of expected economic profits, the industry-wide component in forecasted earnings is more sustainable than the firm-specific component. We thus predict that the association between the forecasted industry-wide earnings and future earnings is greater than the association between the forecasted firm-specific earnings and future earnings: H1a: The association between the forecasted industry-wide earnings and future earnings is greater than the association between the forecasted firm-specific earnings and future earnings. 8 The functional fixation hypothesis suggests that investors are fixated on the forecasted total earnings and fail to fully appreciate the greater association between forecasted industry-wide earnings and future earnings. We thus predict that prices behave as if investors underweight the forecasted industry-wide earnings: H1b: Prices behave as if investors underweight the forecasted industry-wide earnings. Note that we do not predict that prices overweight the forecasted firmspecific earnings. Given the prior evidence of underreaction to forecasted earnings, we attribute this underreaction to underweighting the forecasted industry-wide earnings. This setting differs from the accrual anomaly setting in which prices on average do not underreact to annual earnings and investors appear to underweight cash flows but overweight accruals Initial and Delayed Reactions to the Information in Forecast Revisions. Our next set of predictions focus on price reactions to analysts forecast revisions. Forecast revisions contain news about both industry-wide and firmspecific earnings. If investors fail to fully understand the greater association between forecasted industry-wide earnings and future earnings, stock prices 7 Viewing analysts and insiders as competitors for value-relevant information is consistent with the finding that number of analysts following is inversely related to the proportion of the firm held by insiders (Moyer, Chatfield, and Sisneros [1989]). Prior evidence on financial analysts performance also suggests that analysts have industry-level preferences. For example, Clement [1999] and Jacob, Lys, and Neale [1999] show that analyst accuracy improves with industry specialization, although Gilson et al. [2001] show that the composition of analyst coverage changes after spin-offs and equity carve-outs. Hutton, Lee, and Shu [2012] find that analysts ability to forecast industry-level information is comparable to that of corporate insiders. 8 All of our hypotheses are stated in alternative form.

8 8 K. W. HUI AND P. E. YEUNG should initially underreact to industry-wide earnings news and thus place similar weights on both industry-wide and firm-specific earnings news: H2a: Prices behave as if investors place similar weights on industry-wide earnings news and firm-specific earnings news. During the post-forecast revision period, stock prices should self-correct and move toward the correct levels as information about industry fundamentals gradually arrives. If investors initially place similar weights on both news components and underreact to industry-wide earnings news, we predict a drift in stock prices associated with industry-wide earnings news: H2b: During the post-forecast revision period, there is a positive association between abnormal stock returns and industry-wide earnings news. Furthermore, if the post-forecast revision drift associated with industrywide earnings news represents a delayed response to predicted changes in future earnings, abnormal returns should be concentrated around information events that reveal the earnings changes, such as earnings announcements. Thus, we expect disproportionally large abnormal returns around subsequent earnings announcements: H2c: Abnormal stock returns associated with industry-wide earnings news are disproportionally concentrated on subsequent earnings announcements Underreaction and Persistence of Earnings Components. Our last hypothesis focuses on the link between the differential persistence of earnings components and the magnitude of the post-forecast revision drift. Recall our main argument is that industry-wide economic profits are more sustainable than firm-specific profits, which leads to a greater association between forecasted industry-wide earnings and future earnings than the association between forecasted firm-specific earnings and future earnings. Because investors fail to recognize the greater association between forecasted industry-wide earnings and future earnings, we observe a significant drift associated with industry-wide earnings news. Underlying our argument is an important link between the degree of underreaction and the differential persistence of earnings components. If the drift associated with industry-wide earnings news is driven by investors failure to fully appreciate the relatively higher persistence of industry-wide earnings, the more persistent are the industry-wide earnings relative to the firm-specific earnings the greater is the drift. We thus test the following prediction: H3: The post-forecast revision drift is more pronounced when industrywide earnings are more persistent than firm-specific earnings.

9 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 9 3. Sample Selection and Key Variable Definitions 3.1 SAMPLE OF FORECASTS IN ANALYST INDUSTRY REPORT Our sample starts with individual analysts industry reports in Investext. 9 We use the forecasts in industry reports to construct industry-wide earnings and industry-wide earnings news for two reasons. First, aggregating the earnings of all firms appearing in the same industry report should considerably reduce the measurement errors in industry-wide earnings arising from industry misclassification. Financial analysts endogenously choose to follow a set of firms and release forecasts of their earnings simultaneously in the same report. Because these firms share significant economic commonalities, it is more effective for an analyst to forecast their earnings simultaneously. 10 Second, we expect to obtain less noisy measures of industry-wide earnings news by aggregating earnings forecasts in an industry report. Because these analysts (1) self-select to specialize and focus their expertise on a given industry and (2) choose to adopt a top-down industry perspective to forecast firm performance, measures of industry-wide earnings news based on their reports should have sufficient power to capture analysts efforts to map market-, sector-, and industry-wide economic conditions into earnings estimates. 11 Because it is time-consuming to manually collect and process analyst industry reports from Investext, we focus on manufacturing industries and collect 58,796 industry reports for the period For each industry report, we collect the name of the brokerage house, the name(s) of the analyst(s), the date of the report, and the name of each firm in the report. To ensure that the forecasts in industry reports are widely disseminated, we manually match industry reports with I/B/E/S using the 2006 version of the Brokerage-Translation File maintained by I/B/E/S, which 9 Analysts self-select to follow certain industries due to their individual background and expertise. In the first part of industry reports, analysts typically discuss the expected overall performance of an industry based on industry fundamentals such as demand and supply conditions, cost of production and pricing powers, and potential structural shifts in the industry. In the second part, analysts make earnings forecasts and buy-sell recommendations for selected individual firms within that industry. The online appendix gives an example of an analyst industry report. 10 Bhojraj, Lee, and Oler [2003] show evidence that industry classification based on financial analysts (i.e., the GICS) is better in grouping more homogeneous firms for capital market research than SIC codes and North American Industry Classification System (NAICS) codes. Firms in analyst industry reports should be even more homogeneous than the GICS, because there are much fewer firms in each industry report. For instance, we derive our industry-wide earnings news based on, on average, only 13.2 firms in each industry report. In contrast, there are on average 26 firms in each six-digit GICS industry (Bhojraj, Lee, and Oler [2003]). 11 We note that random measurement errors in industry-wide earnings news bias against finding statistically significant drift associated with industry-wide earnings news (i.e., against our main conclusion).

10 10 K. W. HUI AND P. E. YEUNG records the names of analysts and their affiliated brokerage houses during Because we rely on the 2006 version of the Brokerage-Translation file, our sample period is the five-year window centered on We focus on analysts forecasts of upcoming annual earnings because all industry reports provide annual earnings forecasts, whereas only a subset of them provide quarterly earnings forecasts. Annual earnings forecasts have also been the primary focus of prior studies (e.g., Stickel [1991], Gleason and Lee [2003]). After merging with I/B/E/S data, we have 31,559 annual earnings forecasts for the manufacturing sector. We require a prior annual forecast issued by the same analyst to compute forecast revisions (Gleason and Lee [2003]), which reduces our sample to 28,029 forecasts. After merging with market data from the CRSP, we have 26,509 forecasts. We further delete the outliers of the continuous variables in the regression analysis at their top and bottom one-percentiles. Our final sample consists of 25,195 individual forecasts representing 4,098 firm-year observations. The forecasts in our sample come from 1,347 distinct analysts and 128 distinct brokerage houses. Table 1 provides descriptive statistics for our sample firms. Panel A shows that our sample includes about 800 firms per year. The average number of revisions per firm ranges from 5.83 (2004) to 6.42 (2008), which are much lower than Gleason and Lee [2003] simply because we focus only on forecasts in industry reports. Panel B of table 1 reports the number of revisions in our sample across firm size decile. The distribution is similar to that reported by Gleason and Lee [2003], which indicates that firms covered in industry reports are not tilted toward either large or small firms. Following Gleason and Lee [2003], in subsequent tests we compute sizeadjusted abnormal returns for each firm by subtracting the mean buy-andhold return of an equal-weighted portfolio of firms in the same NYSE (New York Stock Exchange) size decile over our holding period. If a stock is delisted during the return accumulation period, we obtain delisting returns following Shumway [1997] and assume the proceeds are reinvested to earn the average return of the matching size decile portfolio. 13 Panel C reports the distribution of the firms in our sample across two-digit SIC industry groups. 14 We find that our sample consists of higher proportions of chemicals (two-digit SIC = 28) and electronic equipment makers in comparison with the CRSP/Compustat/I/B/E/S population. 12 I/B/E/S stopped providing Brokerage-Translation File for academic research after For firms delisted during the future return period, we calculate the remaining return by taking CRSP s delisting return and then reinvesting the proceeds in the equally weighted reference portfolio. For firms delisted due to poor performance (delisting codes 500 and ), we use a 35% delisting return for NYSE/AMEX (American Stock Exchange) firms and a 55% delisting return for NASDAQ firms. 14 We use SIC codes to tabulate the industry distribution of our sample because analyst industry reports in Investext are organized by two-digit SIC code.

11 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 11 TABLE 1 Number of Forecast Revisions by Year and Firm Size Panel A: Number of Firms and Revisions per Year Number of Forecast Revisions per Firm Number Number of of Firms Revisions Average Median Std. Dev , , , , , Total 4,098 25, Panel B: Number of Revisions and Firms by Firm Size Decile Forecast Revisions Firms NYSE Size % of Number % of Decile Total Sample of Firms Total Firms % % % % % % % % 5 1, % % 6 1, % % 7 2, % % 8 3, % % 9 4, % % 10 9, % % Total 25, % 4, % Panel C: Industry Composition %ofcrsp, Two-Digit Industry Group No. of % of Compustat, SIC Code Name Firm-Years Sample & I/B/E/S 20 Food and kindred products % 4.59% 21 Tobacco products % 0.38% 22 Textile mill products % 0.62% 23 Apparel and other textile products % 1.89% 24 Lumber and wood products % 0.81% 25 Furniture and fixtures % 1.32% 26 Paper and allied products % 1.99% 27 Printing and publishing % 2.34% 28 Chemicals and allied products 1, % 21.59% 29 Petroleum and coal products % 1.29% 30 Rubber and misc. plastics products % 1.96% 31 Leather and leather products % 0.83% 32 Stone, clay, and glass products % 1.07% 33 Primary metal industries % 2.90% 34 Fabricated metal products % 2.65% 35 Industrial machinery and equipment % 12.36% 36 Electronic and other electric equipment 1, % 20.00% 37 Transportation equipment % 4.34% 38 Instruments and related products % 15.46% 39 Misc. manufacturing industries % 1.62% Total 4, % 100% Panel A of table 1 shows the number of firms and forecasts (and forecasts per firm) over our sample period. Panel B shows the distribution of forecast revisions and firms across NYSE size deciles. The threshold values for NYSE size deciles are based on beginning-of-year market capitalization for all NYSE firms. Panel C shows the distribution of firms in our sample and in the CRSP, Compustat, and I/B/E/S population across two-digit SIC industry groups within manufacturing industries.

12 12 K. W. HUI AND P. E. YEUNG 3.2 COMPUTATION OF KEY VARIABLES Our tests must partition forecasted earnings into industry-wide and firmspecific components. Following prior studies (e.g., Brown and Ball [1967], Ayers and Freeman [1997], Elgers et al., [2008]), industry-wide earnings represent the common component of the earnings of firms in the same industry, although firm-specific earnings are deviations of individual firms earnings from the industry average. Let F i,j,t denote the forecasted earnings of firm i in industry report j for year t in an analyst industry report and assume that there are N firms in the industry report, the forecasted industry-wide earnings for year t in report j are defined as: N IndF j,t = 1/N F i, j,t, (1) and forecasted firm-specific earnings of firm i in report j for year t are defined as: FirmF i, j,t = F i, j,t IndF j,t. (2) We scale forecasted and actual earnings per share by stock prices at the beginning of the fiscal year. 15 We also define industry-wide and firm-specific earnings news using analysts forecast revisions. Specifically, let Rev i,j,t represent an analyst forecast revision for firm i in report j for year t, the industry-wide earnings news (IndRev) isdefinedas: N IndRev j,t = 1/N Rev i, j,t. (3) Following Gleason and Lee s [2003] finding that price reactions are highly associated with the innovation in analysts own forecasts, we define an analyst forecast revision (Rev) as the difference between the forecast in the industry report (F ) and the prior forecast issued by the same analyst for the same fiscal period, scaled by stock price two days before the revision date. Note that industry-wide earnings news as defined in equation (3) is known to investors on the day that an industry report is released. So a trading strategy can be potentially implemented. We define firm-specific earnings news in an analysts forecast revision as the difference between the forecast revision and the industry-wide earnings news: FirmRev i, j,t = Rev i, j,t IndRev j,t. (4) Firm-specific earnings news captures the deviation of earnings news on individual firms from industry average earnings news. i=1 i=1 15 Our industry-wide earnings measures contain the impact of market-wide forces on each industry. Because the main variable of interest is abnormal stock returns, which exclude marketwide earnings information, including market-wide earnings does not bias our results.

13 4. Empirical Results 4.1 UNIVARIATE SORTING RESULTS UNDERREACTION TO INDUSTRY-WIDE EARNINGS 13 We first replicate prior findings of post-forecast revision drift in panel A of table 2. Column (1) shows that abnormal (size-adjusted) stock returns during the [ 1, +1] three-day window centered on the industry report date increase from the Low revision decile to the High revision decile. The spread between the average abnormal returns of High and Low deciles is 1.06% (t = 9.23). This evidence indicates that forecast revisions are informative to investors. Columns (2) (4) of panel A show average stock returns of revision deciles during various windows after the industry report date. Consistent with Gleason and Lee [2003], we find significant post-forecast revision drift in our sample. Abnormal stock returns increase from the Low revision decile to the High revision decile in these columns and the spreads between the High and Low deciles are significant (t 3.80). The total drift during the entire [+2, +240] period is 8.18% (i.e., the sum of the spreads between the average abnormal returns of High and Low), which is both statistically and economically significant. 16 Next, we provide new evidence that post-forecast revision drift is mainly attributable to industry-wide earnings news (in panel B) but not firmspecific earnings news (in panel C). Panel B reports abnormal stock returns sorted on industry-wide earnings news (IndRev). The patterns observed in this panel are very similar to those in panel A. Column (1) shows a significant spread between the average abnormal returns of High and Low IndRev deciles (0.79%, t = 7.04), suggesting that investors consider industry-wide earnings news informative. Columns (2) (4) show a significant drift associated with industry-wide earnings news. The total drift during the entire [+2, +240] period is 9.54%. In contrast, the evidence in panel C indicates that, although firm-specific earnings news (FirmRev) is informative to investors during the [ 1, +1] window, there is no drift associated with FirmRev. Specifically, we do not observe increases in stock returns from the Low FirmRev decile to the High FirmRev decile during any of the post-forecast revision windows. The spreads between High and Low FirmRev deciles are small in magnitude and statistically insignificant. Overall, the results in panels B and C are consistent with our prediction that while prices initially react to both industry-wide earnings news and firm-specific earnings news in analysts forecast revisions, they continue to drift only in the direction of industry-wide earnings news. 16 To compare the magnitude of the drift in our sample with that in Gleason and Lee s [2003] sample, we group positive and negative forecast revisions in our sample separately and find a 6.64% hedge portfolio return based on the positive-negative portfolios, slightly below the 8.0% reported by Gleason and Lee.

14 14 K. W. HUI AND P. E. YEUNG TABLE 2 Forecast Revision Drift Associated with Industry-Wide News and Firm-Specific News Panel A: Average Size-Adjusted Stock Returns (%) by Rev Decile (1) (2) (3) (4) Rev Decile [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Low High High Low (9.23) (7.86) (6.48) (3.80) Panel B: Average Size-Adjusted Stock Returns (%) by IndRev Decile IndRev Decile [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Low High High Low (7.04) (9.58) (7.72) (4.31) Panel C: Average Size-Adjusted Stock Returns (%) by FirmRev Decile FirmRev Decile [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Low High High Low (5.78) ( 1.22) ( 1.18) (0.35) Table 2 shows the average abnormal returns for decile portfolios sorted on forecast revision (Rev), industry-wide news (IndRev), and firm-specific news (FirmRev). Rev is defined as the difference between the forecast in an analyst industry report and the prior forecast issued by the same analyst based on I/B/E/S, scaled by stock price two days before the revision date. IndRev is defined as the average value of Rev for firms in an analyst industry report. FirmRev is defined as the difference between Rev and IndRev. Ret[ 1, +1] is cumulative abnormal (size-adjusted) returns during the [ 1, +1] three-day window centered on the analyst industry report date. Ret[+2, +60], Ret[+61, +120], and Ret[+121, +240] are abnormal (size-adjusted) stock returns cumulated over various windows relative to the analyst industry report date.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%. and 1% levels, respectively.

15 UNDERREACTION TO INDUSTRY-WIDE EARNINGS WEIGHTING FORECASTED INDUSTRY-WIDE AND FIRM-SPECIFIC EARNINGS In this section, we provide evidence on (1) whether the association between forecasted industry-wide earnings (IndF ) and future earnings is greater than the association between forecasted firm-specific earnings (FirmF ) and future earnings (H1a) and (2) whether stock prices behave as if they underreact to forecasted industry-wide earnings (H1b). We first estimate the following ordinary least square (OLS) regression to assess the actual association of forecasted earnings components and future earnings: FEarn i,j,t = a 0 + a 1 IndF j,t + a 2 FirmF i, j,t + ε 1i, j,t, (5a) where FEarn i,j,t represents future earnings of firm i that appears in industry report j (scaled by stock price at the beginning of year t). IndF i,j,t and FirmF i, j, t are defined in section 3.2. Our hypothesis H1a predicts a 1 > a 2. To be consistent with the one-year drift window (discussed below), future earnings (FEarn) should be the realized actual earnings during the one-year window after the industry report date. Because analysts forecasts are issued at various points in time, future earnings depend necessarily on when a forecast is issued (i.e., they are horizon-dependent). Specifically, we define FEarn as the earnings of year t if the forecast is issued early in the year t (i.e., the first quarter), average earnings of years t and t + 1 if the forecast is issued in mid-year (i.e., the second or the third quarter), and earnings of year t + 1 if the forecast is issued late in the year (i.e., during the fourth quarter). 17 To test hypothesis H1b, we follow the Mishkin test methodology (Abarbanell and Bernard [1992], Sloan [1996]) and estimate the implied weights on forecasted industry-wide earnings and forecasted firm-specific earnings in stock prices: Ret[ + 2, +240] (5b) = Multiplier(FEarn i, j,t α 0 α 1 IndF i, j,t α 2 FirmF i, j,t ) + ε 2i, j,t, where Ret[+2, +240] is (size-adjusted) abnormal return during the approximate one-year window after the industry report date, and Multiplier is the earnings-response-coefficient. If stock prices behave as if investors fail to fully appreciate the differential association between forecasted earnings components and future earnings, we expect α 1 = α 2 in equation (5b). Comparing estimated coefficients between (5a) and (5b), hypothesis H1b also predicts a 1 >α 1. We present the regression results in table 3. Column (1) shows the results of a baseline OLS regression that predicts future earnings (FEarn) with forecasted earnings (F ). We find that the estimated coefficient for 17 Results are similar if we use a simple average of the earnings of year t and year t + 1. More complicated horizon-dependent weighting algorithms on earnings of year t and year t + 1 (i.e., weighted by the forecasting horizon) also yield similar results.

16 16 K. W. HUI AND P. E. YEUNG TABLE 3 Mishkin Tests: Pricing of Industry-Wide and Firm-Specific Earnings Forecasts (1) (2) (3) (4) Future Future Future Future Earnings Abnormal Returns Earnings Abnormal Returns Intercept (4.50) (12.79) (2.34) (11.65) F (24.93) (25.87) IndF (23.60) (23.16) FirmF (16.38) (13.64) Multiple (38.99) (38.80) Observations 25,195 25,195 25,195 25,195 Adj. R % 5.79% 40.58% 5.83% F -test: IndF = FirmF (0.00) (0.79) χ 2 -test: Underpricing on F (0.00) χ 2 -test: Underpricing on IndF (0.00) χ 2 -test: Underpricing 0.10 on FirmF (0.75) Table 3 presents the results of Mishkin tests of whether stock prices behave as if investors underreact to industry-wide earnings. In columns (1) and (2), we present the results of the following two regressions: FEarn = a 0 + a 1 F + e 1 Ret[+2, +240] = Multiple(FEarn α 0 α 1 F ) + e 2 F is the earnings forecast for year t, scaled by stock price at the beginning of year t. FEarn is horizondependent future earnings, defined as actual earnings for year t if the forecast is issued in the first quarter of year t, average of actual earnings for years t and t + 1 if the forecast is issued in the second or the third quarter of year t, and actual earnings for year t + 1 if the forecast is issued in the fourth quarter of year t. This measure of future earnings is scaled by stock price at the beginning of year t. Ret[+2, +240] is (size-adjusted) abnormal returns for the period [+2, +240] after the forecast revision date. Multiple is the earnings multiple implicated in stock returns. In columns (3) and (4), we present the results of the following two regressions (i.e., equations (5a) and (5b)): FEarn = a 0 + a 1 IndF + a 2 FirmF + e 1 Ret[+2, +240] = Multiple(FEarn α 0 α 1 IndF α 2 FirmF ) + e 2 IndF (forecasted industry-wide earnings) is the average value of forecasted earnings per share (scaled by stock price at the beginning of year t) for all firms in an analyst industry report. FirmF (forecasted firmspecific earnings) is the difference between F and IndF.The χ 2 tests compare (1) a 1 and α 1 and (2) a 2 and α 2.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of OLS regressions are based on firm and year doubleclustered standard errors. forecasted earnings (F ) is 0.831(t = 24.93). Column (2) shows the implicit weight on forecasted earnings in stock prices. The estimated coefficient for F is (t = 25.87), significantly lower than that in column (1) (p < 0.01). In other words, the implicit association between forecasted earnings and future earnings in stock prices is significantly lower than the actual association between forecasted earnings and future earnings. Thus, the baseline regression results confirm an underreaction to analyst forecasts.

17 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 17 Column (3) shows the results of estimating equation (5a). We find that the estimated coefficient for forecasted industry-wide earnings (IndF )is (t = 23.60), whereas the estimated coefficient for forecasted firmspecific earnings (FirmF )is0.645(t = 16.38). An F -test indicates that these two coefficients differ significantly (p < 0.01). This evidence supports hypothesis H1a that the association between forecasted industry-wide earnings and future earnings is greater than the association between forecasted firm-specific earnings and future earnings. In other words, the industrywide component in forecasted earnings is more sustainable than the firmspecific component. Column (4) shows the implicit weights on forecasted industry-wide and firm-specific earnings in stock prices. The results indicate that the estimated coefficient for IndF is insignificantly different from the estimated coefficient for FirmF (p = 0.79). Thus, prices appear to place similar weights on the two forecasted earnings components. A chi-square test comparing the estimated coefficients for IndF between columns (3) and (4) indicates that stock prices significantly underweight forecasted industry-wide earnings (p < 0.01). These results support hypothesis H1b that the market does not appear to recognize the differential associations between earnings components and future earnings, and underreacts to forecasted industry-wide earnings PRICE REACTIONS TO INDUSTRY-WIDE AND FIRM-SPECIFIC NEWS In this section, we present results regarding whether the post-forecast revision drift is mainly attributable to investors underreaction to industrywide earnings news (H2a and H2b). We estimate the following two OLS regressions: CAR i, j,t = b 0 + b 1 IndRev i, j,t + b 2 FirmRev i, j,t + b k Controls + ε 3i, j,t,(6a) and CAR post i, j,t = β 0 + β 1 IndRev i, j,t + β 2 FirmRev i, j,t + β k Controls+ ε 4i, j,t,(6b) where CAR and CAR post are cumulative abnormal stock returns during the forecast revision window and various post-forecast revision windows. To maintain comparability with prior research, we follow the model used by Gleason and Lee [2003] and include three control variables that empirically explain abnormal stock returns. Log(MV ) is defined as the natural log of the market capitalization at the beginning of the year. BM is the book-tomarket ratio at the beginning of the year. Momentum is the market-adjusted returns over the prior six months Furthermore, a chi-square test comparing estimated coefficients for FirmF between columns (3) and (4) suggests that the implicit weight on forecasted firm-specific earnings in stock prices is similar to the actual weight (p = 0.75). 19 Statistical significance is based on firm and forecast date double-clustered standard errors.

18 18 K. W. HUI AND P. E. YEUNG TABLE 4 Price Reactions to Industry-Wide and Firm-Specific Earnings News During Various Windows Forecast Revision Window Post-Forecast Revision Windows (1) (2) (3) (4) [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Intercept (0.84) ( 1.62) ( 1.90) ( 1.62) IndRev (5.59) (4.35) (3.48) (2.65) FirmRev (7.27) ( 0.83) ( 0.28) ( 0.09) Log(MV ) ( 0.84) (0.80) (1.54) (1.08) BM (0.23) (2.76) (2.62) (2.73) Momentum (0.66) (3.15) (3.23) (2.00) F -test: IndRev = FirmRev (p-value) (0.24) (0.00) (0.00) (0.02) Observations 25,195 25,195 25,195 25,195 Adj. R % 1.29% 1.21% 1.11% Table 4 presents the regression results of (size-adjusted) abnormal stock returns cumulated over various windows relative to the analyst industry report date. IndRev is defined as the average of forecast revisions (Rev) of individual firms in an analyst industry report. Forecast revision (Rev) of a firm is defined as annual earnings forecast in an analyst report minus prior forecast issued by the same analyst, deflated by stock price two days before the revision date. FirmRev is the difference between Rev and IndRev. Log(MV ) is defined as natural log of the market capitalization at the beginning of the year. BM is the book value of equity plus the deferred tax divided by market value of equity at the beginning of the year. Momentum is the market-adjusted returns over the prior six months.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of regression analyses are based on firm and forecast date double-clustered standard errors. In equation (6a), we expect b 1 > 0andb 2 > 0 because prices initially react to both industry-wide earnings news and firm-specific earnings news. Hypothesis H2a predicts that investors place similar weights on industrywide earnings news and firm-specific earnings news (i.e., b 1 = b 2 ). Hypothesis H2b predicts a positive drift associated with industry-wide earnings news (i.e., β 1 > 0 in equation (6b)). We expect no drift associated with firmspecific earnings news (i.e., β 2 = 0). We present the regression results in table 4. Column (1) shows the results of price reactions to analysts forecasts during the three-day (i.e., [ 1, +1]) window centered on the industry report date. As expected, we find positive coefficients for IndRev (0.125, t = 5.59) and for FirmRev (0.103, t = 7.27), and the magnitudes are similar (p = 0.24). These results are consistent with hypothesis H2a in that prices behave as if investors place similar weights on industry-wide and firm-specific earnings news. Results in columns (2) (4) show evidence of delayed price reactions during various post-forecast revision windows. We find positive coefficients for IndRev in the regressions of cumulative abnormal returns during the

19 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 19 TABLE 5 Hedge Portfolio Returns (%) from Industry-Wide Earnings News Around Subsequent Earnings Announcements (1) (2) Number of Days Returns Average Around Subsequent During Daily Returns Quarterly Earnings Subsequent Number of Days Announcements Earnings in Earnings (1) (2) (1) /(2) vs. Total Announcements Announcements (Difference) (Ratio) Number of Days 60 days after /59 revision (3.81) 120 days after /119 revision (4.12) 240 days after /239 revision (4.68) Table 5 compares the magnitude of abnormal returns around subsequent earnings announcements and the magnitude of expected abnormal returns. Hedge portfolios are formed by longing the top decile industry-wide earnings news in analysts forecast revisions (IndRev) and shorting the bottom decile industrywide earnings news in analysts forecast revisions. IndRev is defined as the average of forecast revisions (Rev) of individual firms in analyst industry reports. Forecast revision (Rev) of a firm is defined as annual earnings forecast in an analyst report minus prior forecast issued by the same analyst, deflated by stock price two days before the revision date. In 60 days, the subsequent one quarterly earnings announcement is included. In 120 days, the subsequent two quarterly earnings announcements are included. In 240 days, the subsequent four quarterly earnings announcements are included.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. [+2, +60], [+61, +120], and [+121, +240] windows. The estimated coefficients are large ( 0.289) in comparison with the estimated coefficient in column (1) and statistically significant (t 2.65). These results provide strong evidence of a drift associated with industry-wide earnings news. On the other hand, we find no evidence of a drift associated with firm-specific earnings news. Specifically, the estimated coefficients for FirmRev are small and statistically insignificant ( t 0.83). F -tests in columns (2) (4) also indicate that the estimated coefficients for IndRev are significantly greater than the estimated coefficients for FirmRev (p 0.02). Overall the results in table 4 support hypotheses H2a and H2b that prices behave as if investors place similar weights on industry-wide and firm-specific earnings news and that the post-forecast revision drift is mainly driven by the underreaction to industry-wide earnings news. 4.4 RETURNS AROUND SUBSEQUENT EARNINGS ANNOUNCEMENTS If the post-forecast revision drift is due to a delayed market response to current information about future earnings and this underreaction is corrected when future earnings are released, future abnormal returns should cluster around subsequent earnings announcements (H2c). Conversely, if the price drift is due to omitted risk variables, we should not observe higher abnormal returns concentrated around the release of subsequent earnings news (e.g., Bernard and Thomas [1990], Sloan [1996]). Table 5 reports the results of tests that examine whether future abnormal returns cluster around earnings announcements. Specifically, we calculate the average of size-adjusted abnormal returns of hedge portfolios over

20 20 K. W. HUI AND P. E. YEUNG three-day windows centered on the next one- to four-quarterly earnings announcements. The hedge portfolios are formed by taking a long position in the firms in the top decile of industry-wide earnings news (IndRev) and taking a short position in the firms in the bottom decile of IndRev. Our results show that the average hedge returns are positive around all four subsequent earnings announcements, ranging from 0.553% to 0.925%. Under the null that abnormal returns do not cluster around earnings announcements, we expect hedge returns of 0.184% 0.477% around these earnings announcements. The difference between observed hedge returns and expected hedge returns is statistically significant (t 3.81) as well as economically significant (i.e., at least 1.94 times the expected hedge returns). The observed higher concentration of abnormal returns around subsequent earnings announcements suggests that at least a portion of the delayed price response is due to misperception about future earnings, which is corrected around subsequent earnings release dates. 4.5 EFFECTS OF PERSISTENCE OF EARNINGS COMPONENTS In this section, we show the results that link the degree of underreaction to the differential persistence of earnings components (i.e., H3). Specifically, we estimate the following two OLS regressions: CAR i, j,t = c 0 + c 1 Rev i, j,t + c 2 Rev i, j,t DiffPers i, j,t 1 + c k Controls + ε 5i, j,t, (7a) and CARpost i, j,t = γ 0 + γ 1 Rev i, j,t + γ 2 Rev i, j,t DiffPers i, j,t 1 + γ k Controls + ε 6i, j,t, (7b) where DiffPers is the difference between the persistence of industry-wide earnings and the persistence of firm-specific earnings. Persistence of industry-wide (firm-specific) earnings is defined as the estimated coefficient for industry-wide (firm-specific) earnings of year t 2 in predicting actual earnings of year t For the ease of interpreting the coefficients, we convert DiffPers into in-sample decile ranks bounded by [0, 1]. If investors fail to appreciate the differential persistence of earnings components on the industry report date, we expect an insignificant coefficient for Rev DiffPers (i.e., c 2 = 0) in equation (7a). During the post-forecast 20 More specifically, industry-wide and firm-specific earnings are defined by substituting the forecasted earnings in equations (1) and (2) with actual earnings of firm i in industry report j. Weuset 2 earnings to predict t 1 earnings to ensure that investors have this information before the analyst industry report. We estimate the persistence of industry-wide and firm-specific earnings by each industry-year. An industry is constructed based on multiple industry reports. Specifically, the union of firms that ever appeared in the industry reports for the same four-digit SIC industry make up an industry. For example, if firms A, B, and C appear in analyst 1 s report for industry j and firms B, C, and D appear in analyst 2 s report for industry j, industry j includes firms A, B, C, and D.

21 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 21 TABLE 6 Differential Persistence of Earnings Components and the Post-Forecast Revision Drift Forecast Revision Post-Forecast Revision Windows Window (1) (2) (3) (4) [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Intercept ( 0.18) ( 1.71) ( 2.10) ( 2.08) Rev (4.31) (0.14) (0.26) (1.16) Rev DiffPers (0.20) (3.01) (2.49) (1.89) DiffPers (1.78) (0.91) (0.05) (0.64) Log(MV ) ( 0.81) (0.84) (1.27) (1.31) BM (0.32) (2.83) (2.73) (2.83) Momentum (0.69) (3.13) (3.37) (2.35) Observations 25,195 25,195 25,195 25,195 Adj. R % 1.26% 1.14% 1.02% Table 6 presents the regression results of abnormal (size-adjusted) stock returns cumulated over various windows relative to the analyst industry report date. Forecast revision (Rev) of a firm is defined as annual earnings forecast in an analyst report minus prior forecast issued by the same analyst, deflated by stock price two days before the revision date. DiffPers is defined as the difference between the persistence of industrywide earnings and the persistence of firm-specific earnings. The persistence of industry-wide earnings is defined as the coefficient for industry-wide earnings of year t 2 (scaled by beginning stock price) in a regression predicting total earnings of year t 1 (scaled by beginning stock price). The persistence of firmspecific earnings is defined as the coefficient for firm-specific earnings of year t 2 (scaled by beginning stock price) in a regression predicting total earnings of year t 1 (scaled by beginning stock price). The union of firms that have ever appeared in the analyst industry reports for the same industry give the set of firms in that industry. For example, if firms A, B, and C appear in analyst 1 s report for industry j and firms B, C, and D appear in analyst 2 s report for industry j, industry j includes A, B, C, and D. DiffPers is converted into decile ranks (i.e., [0, 1]) in the regression. Log(MV ) is defined as natural log of the market capitalization at the beginning of the year. BM is the book value of equity plus the deferred tax divided by market value of equity at the beginning of the year. Momentum is the market-adjusted returns over the prior six months.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of regression analyses are based on firm and forecast date double-clustered standard errors. revision period, stock prices gradually reflect the implications of differential persistence of earnings components. We thus expect a significantly positive coefficient for Rev DiffPers (i.e., γ 2 > 0) in equation (7b). The results in column (1) of table 6 indicate an insignificant coefficient for Rev DiffPers (t = 0.20), suggesting that the market does not appear to appreciate the differential persistence of earnings components during the initial event window. On the other hand, we find significantly positive coefficients for Rev DiffPers (t 1.89) in columns (2) through (4), indicating that stock prices gradually correct initial underreaction to the differential persistence of earnings components. The magnitude of the estimated coefficients confirms that differential persistence of earnings components plays an important role in explaining the drift. For instance, the estimated coefficients

22 22 K. W. HUI AND P. E. YEUNG for Rev and Rev DiffPers in column (2) are and 0.516, suggesting that the drift in the top decile (DiffPers = 1) is 14 times larger than the drift in the bottom decile (DiffPers = 0). In short, the results in table 6 are consistent with hypothesis H3 that the post-forecast revision drift is more pronounced when industry-wide earnings are more persistent than firmspecific earnings Additional Analyses 5.1 FURTHER EVIDENCE ON THE CROSS-SECTIONAL VARIATION IN THE POST-FORECAST REVISION DRIFT Our first set of additional analyses concerns the cross-sectional variation in the drift associated with industry fundamentals. If the post-forecast revision drift is rooted in investors underreaction to industry-wide earnings news, we expect a more pronounced drift when a greater proportion of firm performance is explained by industry-wide profit (i.e., firm performance moves more closely with industry-wide earnings). Panel A of table 7 presents the results regarding the effects of performance co-movement on the post-forecast revision drift. We measure performance co-movement (Co-Movement) with two empirical proxies. The first proxy is Earnings Co-Movement, defined as the correlation coefficient between industry total earnings (i.e., the sum of earnings of all firms in the industry) and the firm s earnings, estimated over the prior eight years (with a minimum of three years). Our second proxy is Return Synchronicity, defined as the proportion of firm stock returns explained by the market and industry returns (Piotroski and Roulstone [2004]). 22 Higher Earnings Co- Movement or Return Synchronicity indicates that a greater proportion of firm performance is tied to industry-wide earnings, and we expect a more pronounced drift. Consistent with this expectation, we find positive coefficients for the interaction terms between forecast revision (Rev) and these two proxies (t 2.54) Our data confirm the prediction that firm-specific earnings are transitory and more likely to be mean reverting, whereas industry-wide earnings are not. We find a significantly negative first-order autoregressive coefficient for changes in firm-specific earnings ( 0.25, t = 3.86) in a pooled regression. In contrast, the first-order autoregressive coefficient for changes in industry-wide earnings is small and statistically insignificant (0.01, t = 0.18). 22 More specifically, Return Synchronicity is estimated as log(r 2 /(1 R 2 )), where R 2 is the R- squared from the regression RET i,t = a + b1market i,t + b2market i,t 1 + b3indret i,t + b4indret i,t 1 + ε i,t. In this regression, RET is firm s weekly return of week t, MARKET is the value weighted market return of week t, and INDRET is the value-weighted industry return of week t. We use the one-year period ending one month prior to industry report date (and require least 40 weeks) to estimate this regression. Any inefficiency in stock prices with respect to information should largely attenuate during this long estimation window. 23 For the ease of interpreting the estimated coefficients, all variables that interact with forecast revisions in the regressions discussed in section 5 are converted into in-sample decile rank and bounded within [0, 1].

23 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 23 TABLE 7 Earnings Co-Movement, Industry Structure, and Cross-Sectional Variation in the 60-Day Post-Forecast Revision Drift Panel A: Drift and the Co-Movement of Firm and Industry Performance (CoMovement) Panel B: Drift and Industry Structures with Higher Persistence of Industry-Wide Earnings (IndStru) (1) (2) (1) (2) Earnings Return Barrier- Industry Co-Movement Synchronicity to-entry Concentration Intercept Intercept ( 2.64) ( 2.73) ( 2.07) ( 1.51) Rev Rev ( 1.07) (0.10) ( 0.02) ( 0.03) Rev CoMovement Rev IndStru (4.33) (2.54) (2.82) (3.08) CoMovement IndStru (1.49) (4.45) ( 0.01) ( 0.99) Log(MV ) Log(MV) (0.94) (0.49) (0.84) (0.75) BM BM (2.98) (1.87) (2.82) (2.78) Momentum Momentum (3.10) (3.09) (3.24) (3.10) Observations 24,422 20,415 Observations 25,195 25,195 Adj. R % 1.83% Adj. R % 1.34% Panels A and B of table 7 present the regression results of abnormal (size-adjusted) stock returns cumulated over [+2, +60] three-month windows relative to the analyst industry report date. Forecast revision (Rev) for a firm is defined as the difference between the forecast in an analyst industry report and the prior forecast issued by the same analyst based on I/B/E/S, scaled by stock price two days before the revision date. CoMovement in panel A refers to the proxies for the co-movement of firm performance with industry profit. IndStru in panel B refers to the proxies for industry characteristics enabling higher persistence of industry-wide earnings. Earnings Co-Movement is defined as the correlation coefficient between the sum of earnings of all firms in the industry and the firm s earnings estimated over the prior eight years (or a minimum of three years). Return Synchronicity is estimated as log(r 2 /(1 R 2 )), where R 2 is the R-squared from the regression RET i,t = a + b1market i,t + b2market i,t 1 + b3indret i,t + b4indret i,t 1 + ε i,t. In this regression, RET is the firm s weekly return, MARKET is the value-weighted market return of week t, and INDRET is the value-weighted industry return of week t. We use the one-year period ending one month prior to industry report date (and require at least 40 weeks) to estimate this regression. Barrier-to-entry for an industry is defined as the sum of gross value (in billion dollars) of the cost of property, plant, and equipment for each firm in an industry weighted by its sales market share. Industry Concentration is defined as the Census six-digit NAICS Herfindahl-Hirschman index to which the firm belongs. We convert these proxies into decile ranks (i.e., [0, 1]) in our regression analysis. Log(MV ) is defined as natural log of the market capitalization at the beginning of the year. BM is the book value of equity plus the deferred tax divided by market value of equity at the beginning of the year. Momentum is the market-adjusted returns over the prior six months.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of regression analyses are based on firm and forecast date double-clustered standard errors. We also expect a larger drift when the industry structure deters entry. Entrants introduce competition, which in turn lowers the existing industry-wide economic rents (e.g., McAfee, Mialon, and Williams [2004]). Thus, an industry structure that deters entry is amenable to more sustainable industry-wide economic profits, which implies more persistent industry-wide earnings. If the post-forecast revision drift is driven by the

24 24 K. W. HUI AND P. E. YEUNG underreaction to more persistent industry-wide earnings, we predict a larger drift when the industry structure deters entry. We measure industry structure (IndStru) with two alternative proxies. The first proxy is Barrier-to-Entry, defined as the sum of gross value (in billion dollars) of the cost of property, plant, and equipment for each firm in an industry weighted by the firm s sales market share (Hui, Klasa, and Yeung [2011]). The second proxy is Industry Concentration, defined as the U.S. Census six-digit NAICS Herfindahl-Hirschman index to which the firm belongs (Ali, Klasa, and Yeung [2009]). 24 Industry concentration likely captures the extent to which an industry is insulated from potential competitions for two reasons. First, a concentrated industry by definition has fewer firms, which in itself could be a result of high barrier-to-entry. Second, firms in more concentrated industries tend to have larger research and development expenditures (Ali, Klasa, and Yeung [2009]), which creates high cost for potential entrants. 25 Panel B of table 7 presents the results regarding the effects of industry structure on the post-forecast revision drift. Consistent with our expectation, we find positive coefficients for the interaction terms between forecast revision (Rev) and these two proxies (t 2.82), indicating a more pronounced drift. Taken together, the results in panels A and B of table 7 corroborate our main results that suggest that the post-forecast revision drift is rooted in investors underreaction to news about industry-wide fundamentals. 5.2 EFFECTS OF TRADING FRICTION In this section, we examine whether trading friction related to arbitrage activities affects the magnitude of the post-forecast revision drift. If the source of the post-forecast revision drift can be traced to underreaction of investors, arbitrage activities of a subset of sophisticated investors should, to some extent, mitigate the mispricing. We thus examine whether the drift varies predictably with the proxies for higher trading friction, which increases the costs of arbitrage activities. We use three proxies for trading friction (Friction): the inverse of the level of stock price at the beginning of the year (Stock Price), the inverse of the dollar trading volume for the prior year (Trading Volume), and 24 We collect industry concentration ratios from the 2002 and 2007 Census of Manufactures publications. The Census of Manufactures is published during years in which a U.S. Census takes place. Following Ali, Klasa, and Yeung [2009], we assume that the U.S. Census industry concentration data are valid for the five-year windows surrounding a U.S. Economic Census year. Specifically, we assume that the values of the 2002 and 2007 industry concentration measures are valid for the 2004 and periods in our sample. 25 Consistent with more persistent industry-wide earnings in industries with high Barrierto-Entry (Industry Concentration), we find that the rank correlation between the persistence of industry-wide earnings and Barrier-to-Entry (Industry Concentration) in our sample is 0.14 (0.13). In contrast, the rank correlation between the persistence of firm-specific earnings and Barrierto-Entry (Industry Concentration) is 0.04 ( 0.10).

25 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 25 TABLE 8 Trading Friction and the 60-Day Post-Forecast Revision Drift Proxies for Market Friction (Friction) (1) (2) (3) Stock Price Trading Volume Idiosyncratic Risk Intercept ( 2.68) ( 1.90) ( 2.60) Rev (0.03) (0.31) ( 0.20) Rev Friction (2.23) (1.98) (2.36) Friction (1.24) (0.35) (2.79) Log(MV ) (1.50) (0.88) (2.33) BM (2.75) (2.79) (3.41) Momentum (3.37) (3.17) (3.15) Observations 25,195 25,195 25,195 Adj. R % 1.23% 1.51% Table 8 presents the regression results of (size-adjusted) abnormal stock returns cumulated over the [+2, +60] windows following the analyst industry report date. Forecast revision (Rev) for a firm is defined as the difference between the forecast in an analyst industry report and the prior forecast issued by the same analyst based on I/B/E/S, scaled by stock price two days before the revision date. Three proxies for market friction are stock price, trading volume, and idiosyncratic risk. Stock Price is the inverse of the level of stock price at the beginning of the year. Trading Volume is the inverse of the dollar trading volume for the prior year. Idiosyncratic Risk is estimated as the standard deviation of residuals from a market model regression of monthly stock returns on the monthly returns of the CRSP equal-weighted market index during the prior 48-month period (and require at least 20 months). We convert these proxies into decile ranks (i.e., [0, 1]) in our regression analysis. Log(MV ) is defined as the natural log of the market capitalization at the beginning of the year. BM is the book value of equity plus the deferred tax divided by market value of equity at the beginning of the year. Momentum is the market-adjusted returns over the prior six months.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of regression analyses are based on firm and forecast date double-clustered standard errors. Idiosyncratic Risk, which is defined as the standard deviation of the residuals from a market model regression of monthly stock returns on monthly CRSP equal-weighted stock return index (e.g., Mashruwala, Rajgopal, and Shevlin [2006]). We use the 48-month period ending one month prior to industry report date (and require at least 20 months) to estimate Idiosyncratic Risk. We expect that the post-forecast revision drift is more pronounced when trading friction is higher. The results in columns (1) (3) of table 8 indicate significantly positive coefficients for Rev Friction (t 1.98), consistent with a more pronounced drift when trading friction is greater. For firms in the bottom decile of each proxy of trading friction (i.e., Friction = 0), the small and insignificant coefficients for Rev suggest that the drift is insignificant. On the other hand, for firms in the top decile of each proxy of trading friction (i.e., Friction = 1), the large coefficients for Rev Friction indicate a strong drift. Our results in table 8, therefore, suggest that the survival of the

26 26 K. W. HUI AND P. E. YEUNG post-forecast revision drift is at least in part attributable to the trading friction in the market, which likely prevents sophisticated investors from trading aggressively against the initial underreaction GENERLIZABILITY OF THE RESULTS In this section, we provide evidence on the generalizability of our results to forecasts outside analyst industry reports. One concern of focusing on industry reports is that their releases might be triggered by significant industry-wide shocks. To shed light on the timing of the release of industry reports, we test whether the releases of industry reports systematically differ from the releases of all I/B/E/S forecasts. We find that the weekly frequency of the forecasts in industry reports (as a percentage of all forecasts in industry reports) tracks remarkably well with that of all forecasts in I/B/E/S. Thus, it is unlikely that systematic shocks cause the releases of industry reports. We also repeat our main analyses in a sample of 18,456 forecasts that do not appear in industry reports (i.e., stand-alone forecasts). We require that these stand-alone forecasts be I/B/E/S forecasts for the firms in our sample issued within five days after each industry report date (i.e., [+1, +5] window). We focus on stand-alone forecasts issued immediately after industry reports, because the industry environment is unlikely to change significantly in a short window. Panel A of table 9 shows that the post-forecast revision drift during the [+2, +60] window (= 1.91%, t = 4.49) is mainly attributed to the drift associated with industry-wide earnings news (= 2.19%, t = 5.28). In contrast, there is no drift associated with firm-specific earnings news (= 0.07%, t = 0.16). We also repeat the multiple regressions (6a) and (6b) in this sample. The results in panel B of table 9 show significantly positive coefficients for IndRev in the regressions of abnormal returns during the post-forecast revision windows (t 2.19). On the other hand, we find no evidence of drift associated with firm-specific earnings news ( t 0.61). 27 Thus, the results in table 9 provide the support for the generalizability of our main results to stand-alone forecasts. Finally, we test whether industry-wide earnings news plays an important role in explaining the post-forecast revision drift following all I/B/E/S forecasts. One caveat with this analysis is that the power of the tests may be low because of the noise in measuring industry-wide earnings news. Nevertheless, we create a measure of industry-wide earnings news on a particular date by requiring at least two forecasts that are issued for separate firms in 26 Further evidence suggests that the impact of transaction costs is primarily driven by the short ends. Specifically, we partition the sample by positive and negative revisions and find that the coefficient for Rev Friction is significantly positive in the negative revision subsample, but insignificant in the positive revision subsample. 27 Consistent with H3, nontabulated results further indicate that the post-forecast revision drift is more pronounced when industry-wide earnings are more persistent than firm-specific earnings.

27 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 27 TABLE 9 Stand-Alone Forecasts Analysis Panel A: Average [+2, +60] Size-Adjusted Stock Returns (%) by Rev, IndRev, FirmRev Decile (n = 18,456) Rev IndRev FirmRev Decile [+2, +60] [+2, +60] [+2, +60] Low High High Low (4.49) (5.28) (0.16) Panel B: Regressions of Abnormal Returns Forecast Revision Post-Forecast Revision Windows Window (1) (2) (3) (4) [ 1, +1] [+2, +60] [+61, +120] [+121, +240] Intercept (0.11) ( 0.62) ( 1.39) ( 1.75) IndRev (4.36) (2.77) (2.59) (2.19) FirmRev (4.39) ( 0.45) ( 0.61) ( 0.24) Log(MV ) ( 0.04) (0.19) (0.78) (1.45) BM ( 0.24) (2.67) (2.44) (2.63) Momentum (0.63) (1.98) (2.12) (2.55) F -test: IndRev = FirmRev (p-value) (0.48) (0.04) (0.05) (0.07) Observations 18,456 18,456 18,456 18,456 Adj. R % 1.20% 1.25% 1.38% Table 9 presents the results for a sample of stand-alone forecasts (n = 18,456), which are individual forecasts issued immediately after the analyst industry report date (i.e., [+1, +5]) but do not appear in any analyst industry report. Panel A presents the average abnormal returns for decile portfolios sorted on forecast revision (Rev), industry-wide earnings news (IndRev), and firm-specific earnings news (FirmRev) during the [+2, +60] three-month period after the analyst industry report date. Rev is defined as the difference between the forecast in an analyst industry report and the prior forecast issued by the same analyst based on I/B/E/S, scaled by stock price two days before the revision date. IndRev is defined as the average value of forecast revisions for all firms in the analyst industry report preceding the stand-alone forecast. Stand-alone forecast revision for a firm is defined as the difference between the stand-alone forecast and the prior forecast issued by the same analyst based on I/B/E/S, scaled by stock price two days before the revision date. FirmRev is defined as the difference between a stand-alone forecast revision and IndRev. Panel B presents regression of abnormal (size-adjusted) returns within various windows around and after analysts forecast revisions. Log(MV ) is defined as natural log of the market capitalization at the beginning of the year. BM is the book value of equity plus the deferred tax divided by market value of equity at the beginning of the year. Momentum is the market-adjusted returns over the prior six months.,, and indicate statistical significance of t-statistics in the parentheses at two-tailed 10%, 5%, and 1% levels, respectively. All statistical significances of regression analyses are based on firm and forecast date double-clustered standard errors.

28 28 K. W. HUI AND P. E. YEUNG the same eight-digit GICS industry. We then define industry-wide earnings news as the average of these forecasts, and firm-specific earnings news the deviation from the industry mean. The rest of the sample selection procedure follows Gleason and Lee [2003]. Nontabulated results show that the post-forecast revision drift during the [+2, +60] window (= 1.49%, t = 10.98) is mainly attributed to the drift associated with industry-wide earnings news (= 1.53%, t = 11.74). In contrast, there is no drift associated with firm-specific earnings news (= 0.22%, t = 1.59). Thus, the evidence from out-of-sample tests using all I/B/E/S forecasts supports our main results. 6. Conclusions Prior research has consistently documented a large drift following analysts earnings forecast revisions, indicating that investors are slow in processing earnings news in analysts forecast revisions. We develop and test the principal hypothesis that investors delayed reactions are mainly attributable to their underreaction to the news about industry-wide earnings. We test our predictions with a sample of earnings forecasts in analyst industry reports, in which an analyst issues forecasts for multiple firms in the same industry simultaneously. We find that the post-forecast revision drift is largely attributable to the drift associated with industry-wide earnings news. Consistent with the functional fixation hypothesis, we find evidence that investors behave as if they do not fully understand the higher persistence of industry-wide earnings and underreact to industry-wide earnings news. Patterns of stock returns around subsequent earnings announcements and cross-sectional variation in the drift further support the conclusion that the post-forecast revision drift is closely related to investors underreaction bias in processing industry-wide earnings news. Our study contributes to the literature on the post-forecast revision drift by showing the information source of investors delayed reaction. Our study also adds to the literature about investors mispricing of earnings components. Although the focus of prior studies is on managerial reporting discretion causing the difference in persistence of earnings components, we show evidence that investors do not seem to fully understand the differential earnings persistence attributable to economic forces. Finally, our study contributes to the literature on the value of analysts forecasts in conveying industry-level earnings information. Although prior research shows that analysts forecast revisions contain significant industry-wide information, we provide evidence that prices underreact to industry-wide earnings news in analysts forecasts. Although we focus on analysts earnings forecasts, our findings also have implications for the literature on investors underreaction to analysts stock recommendations (Stickel [1995], Womarck [1996], Barber et al. [2001], Jegadeesh et al. [2004], Howe, Unlu, and Yan [2009]). To the extent that analysts specialize in industries and their stock recommendations contain

29 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 29 useful industry-wide information, investors underreaction to recommendations is at least in part attributable to the industry-wide commonalities in analysts recommendations (Howe, Unlu, and Yan [2009]). Because the link between analysts earnings forecasts and stock recommendations has not been well specified (e.g., Bradshaw [2004]), investigating the extent to which functional fixation on industry-wide earnings can explain the drift associated with analysts recommendation is less than straightforward and we leave it for future research.

30 30 K. W. HUI AND P. E. YEUNG APPENDIX An Example of an Industry Analyst Report

31 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 31

32 32 K. W. HUI AND P. E. YEUNG

33 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 33

34 34 K. W. HUI AND P. E. YEUNG

35 UNDERREACTION TO INDUSTRY-WIDE EARNINGS 35