CFA Institute. CFA Institute is collaborating with JSTOR to digitize, preserve and extend access to Financial Analysts Journal.

Transcription

1 CFA Institute Double Surprise into Higher Future Returns Author(s): Alina Lerman, Joshua Livnat and Richard R. Mendenhall Reviewed work(s): Source: Financial Analysts Journal, Vol. 63, No. 4 (Jul. - Aug., 2007), pp Published by: CFA Institute Stable URL: Accessed: 23/07/ :42 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.. CFA Institute is collaborating with JSTOR to digitize, preserve and extend access to Financial Analysts Journal.

2 Financial Analysts Journal Volume 63 * Number 4?2007, CFA Institute Double Surprise into Higher Future Returns Alina Lerman, Joshua Livnat, and Richard R. Mendenhall Post-earnings-announcement drift is the well-documented ability of earnings surprises to predict future stock returns. Despite nearly four decades of research, little has been written about the importance of how earning surprise is actually measured. We compare the magnitude of the drift when historical time-series data are used to estimate earnings surprise with the magnitude when analystforecasts are used. We show that the drift is significantly larger when analystforecasts are used. Furthermore, we show that using the two models together does a better job of predictingfuture stock returns than using either model alone. O ne of the most puzzling characteristics of the stock market is that earnings surprises seem to predict future stock return performance. That is, when companies announce quarterly earnings that surpass market expectations, the stocks of those companies, on average, exhibit higher-than-normal return performance for weeks and even months following the earnings announcement. The opposite is true for stocks of companies whose earnings fall short of market expectations; they tend to perform poorly. This welldocumented phenomenon is normally referred to as either "post-earnings-announcement drift" or "the SUE (standardized unexpected earnings) effect." The drift was first noticed almost 40 years ago, but a constant stream of research from 1968 to the present has documented and confirmed the anomaly. Possible explanations for the drift include an inability of researchers to properly measure risk and various methodological problems, but today, most researchers interpret the body of evidence as indicating that the SUE effect simply represents a slow market response to the information contained in earnings announcements. Although the existing drift literature addresses many methodological concerns (some of which are discussed in the next section), one issue that seems to us to be critical has been largely ignored: What is the best way to measure the earnings surprise? If the drift represents a slow market reaction to the information in earnings announcements, the way in which that information is assessed may be vital Alina Lerman is a PhD candidate and Joshua Livnat is professor of accounting in the Stern School of Business at New York University, New York City. Richard R. Mendenhall is professor offinance at the University of Notre Dame, Notre Dame, Indiana. to the magnitude of the drift; that is, a better measure of earnings surprise may lead to a larger drift. Most SUE studies have estimated the earnings surprise as the difference between reported EPS and a time-series earnings forecast (usually deflated by price or past earnings variability). The specific timeseries process chosen does not seem to make much difference in the magnitude of the drift (see Foster, Olsen, and Shevlin 1984). Perhaps for this reason, researchers have not closely examined the issue of measuring the earnings surprise. But certainly, another prominent measure of earnings surprise does exist-namely, the difference between reported earnings and financial analysts' forecasts of earnings (in lieu of a time-series forecast). The purpose of the study reported in this article was to examine the magnitude of the drift when earnings surprises are assessed by using time-series forecasts, by using financial analyst forecasts, and by using both measures of surprise together. Prior Research and Motivation Ball and Brown (1968) took for granted that stock prices reflect virtually all pertinent information. To show the relevance of the earnings accrual system, they demonstrated a positive correlation between unexpected accounting earnings and contemporaneous stock returns. They were surprised to find not only a contemporaneous relationship between unexpected earnings and returns but a lagged one also. Given that Ball and Brown were relying on the concept of market efficiency to demonstrate the relevance of accounting earnings, we might imagine they were stunned to find evidence suggesting that changes in accounting earnings were able to predict future stock returns-that is, that the market may not be so efficient after all. July/August

3 Financial Analysts Journal Almost immediately following the study by Ball and Brown (1968), a stream of studies through the 1970s documented and confirmed the market's apparent slow response to earnings information. Ball (1978) summarized much of the early literature and pointed out limitations of then extant studies. He mentioned specifically the common practice of assuming that earnings had been announced by a certain time on the basis of the end of the fiscal quarter, Compustat's practice of changing earnings figures after the fact for those companies that later revise their earnings reports, and computational biases in estimating abnormal performance. He concluded that the market's apparent slow reaction to earnings information is most likely illusory and a result of researchers' inability to properly adjust stock returns for risk. Most of Ball's (1978) criticisms have now been addressed. For example, most studies since the early 1980s have used (nearly) exact earnings announcement dates from Compustat.1 Furthermore, Livnat and Mendenhall (2006) showed that Compustat's earnings-restatement policy has a trivial effect on the magnitude of the drift. Finally, Bernard and Thomas (1989) showed that the drift is robust to a wide range of risk-correction procedures. In light of the subsequent evidence, Ball (1992), like many drift observers, changed his mind and ascribed the SUE effect to "either market inefficiency or substantial costs of investors acquiring and processing information" (p. 319). The current view of the drift is that, although trading impediments, such as information-processing costs, may exist, the drift most likely represents the slow reaction of stock prices to earnings information. Bernard and Thomas (1990) hypothesized a specific form of the slow reaction to earnings information. That is, they hypothesized that at least a substantial subset of investors view earnings as following a seasonal random walk (SRW). In other words, many investors hold expectations that earnings "this quarter" will be the same as they were the same fiscal quarter of the previous year. By examining the pattern of stock returns around the four earnings announcements subsequent to the earnings surprise, the authors concluded that their results were consistent with many investors believing that earnings follow an SRW pattern. This finding is important because the SRW model understates the implications of earnings surprises for future earnings levels; that is, investors using this model will systematically underreact to earnings surprises, which could give rise to the drift. Although a reexamination was not their primary focus, Livnat and Mendenhall (2006) reexamined the primary results in the Bernard and Thomas (1990) paper for a later time period and with a subset of companies for which analysts' forecasts were available from Thomson Corporation's I/B/E/S. Livnat and Mendenhall's results suggest that the specific form of underreaction documented by Bernard and Thomas may be but one manifestation of a more general underreaction to earnings surprises. We took this suggestion as our basic premise: Postearnings-announcement drift is the manifestation of a general slow market response to earnings surprises. We hypothesized that superior, more complete measures of earnings surprise will produce larger drifts. Brown, Griffin, Hagerman, and Zmijewski (1987) showed that analyst forecast errors are a better measure of earnings surprise than are timeseries errors-at least as measured by initial stock market reactions. This finding makes sense because analysts have access to a broader and more timely set of information than simply the pattern of past earnings. Although analyst forecast errors may be a superior measure of surprise, Brown et al. also showed that these errors do not completely subsume time-series errors in explaining the immediate stock price reaction to earnings. The reason may be one or more of previously suggested sources of analyst bias. For example, analysts may be hesitant to make low earnings forecasts for fear of alienating company management, which might hinder the analysts' future ability to obtain information from the company. Analysts may fear alienating company management lest it risk existing (or potential) investment banking relationships between the company and the analysts' employers. Furthermore, some practitioners indicate that optimistic forecasts and buy recommendations increase trading commissions from their firms' brokerage clients (see Affleck-Graves, Davis, and Mendenhall 1990; Dugar and Nathan 1995). Finally, time-series errors calculated from Compustat data rely on earnings that reflect GAAP, but I/B/E/S and similar analyst- tracking services tend to use "Street" earnings figures that exclude some expenses required by GAAP. Whatever the reasons, Brown et al. showed that neither model subsumes the other as a measure of earnings surprise. We, therefore, estimated the magnitude of the drift by defining the earnings surprise on the basis of a time-series model, analyst forecasts, and a combination of the two. Methodology We first describe our methods of estimating earnings surprises and the sample. We used two approaches to estimate earnings surprises: the historical time-series model based on Compustat earnings data and the analysts' forecasts model based on I/B/E/S data , CFA Institute

4 Double Surprise into Higher Future Returns Estimation of Earnings Surprises. As in most previous drift literature, we used rolling windows of historical data to define the time-series measure of standardized unexpected earnings. Specifically, we defined SUE as errors from a seasonal random walk model with drift. For each company-quarter, we began by estimating the following model: Ej t = 8j,t + Ej,t_4 + Fjjt 1 where EjNt = quarterly diluted EPS before extraordinary items for company j in quarter t 61 t = a drift term to allow for the company's recent historical earnings growth ej t = the error term, with standard deviation STDj,t To compute the earnings surprise for quarter t, we used data from quarters t - 8 through t -1. Therefore, the time-series measure of earnings surprise, SUE, is defined as SUE1,t = E, i,-e1,t4 (2) STDj,t Our second approach was to estimate earnings surprises by using actual earnings versus expected earnings from I/B/E/S. We defined the standardized unexpected earnings using analysts' forecasts (SUEAF) as pibes SUEAFj,, = t STDF, F (3) where Elbes is the actual EPS reported in I/B/E/S and Fj,t is the mean of all quarterly forecasts of EPS made by analysts during the 90-day period prior to the disclosure of the actual earnings.2 The surprise was then scaled by the dispersion of analysts' forecasts, STDF Dispersion STDF was set to 0.01 if the forecast dispersion was zero or if the companyquarter had forecasts from fewer than two analysts. As mentioned, most prior research used models of earnings surprise based solely on historical time series. On the one hand, these models have the advantage of not requiring the researcher to obtain analyst forecast data. Furthermore, analysts who report to forecast providers, such as I/B/E/S, do not cover all companies. So, time-series models can be used to provide measures of earnings surprise for some companies not covered by analysts. These companies may be those that are most overlooked by investors and so may exhibit larger postearnings-announcement drift than widely covered companies. On the other hand, for those companies for which forecasts are available, analysts' earnings forecasts are more accurate than forecasts based on historical data and their errors are more highly associated with the stocks' contemporaneous price movements. So, for companies followed by analysts, analyst forecast errors may provide a better measure of earnings surprise and, therefore, be better predictors of drift. Sample. We obtained original Compustat data from Charter Oak Investment Systems, stock return data from CRSP, and analyst forecast data from the I/B/E/S Detail History File.3 As discussed, our sample was restricted to companyquarters for which Compustat data were available for the prior eight quarters and restricted to companies followed by at least one analyst in the 90-day period prior to the preliminary earnings release. Only U.S. companies were included in the sample. Additional selection criteria were as follows: * The earnings announcement date for quarter t was reported in both Compustat and I/B/E/S, and the two dates differed by no more than one calendar day. * The number of shares outstanding and the price per share were available from Compustat as of the end of quarter t. The calculated market value of equity as of the end of quarter t was greater than $100 million. * The book value of equity at the end of quarter t was available from Compustat. * The company's shares were traded on the NYSE, Amex, or NASDAQ. * Daily returns were available from the CRSP dataset from two days after the preliminary earnings announcement for quarter t through one day after the preliminary earnings announcement for quarter t + 1, if available, or 100 calendar days if unavailable. * Data were available to assign the company to one of the six Fama-French (1992) portfolios based on size and book value to market value (BV/MV). * Both SUE and SUEAF (as defined in Equations 2 and 3) could be calculated for quarter t. The final sample consisted of 129,888 companyquarter observations. The sample begins in the second quarter (Q2) of 1987 with 810 companies and ends in Q with 2,790 companies. Assignment to Quintiles. We grouped companies with fiscal quarters ending within a particular calendar quarter into quarter cohorts. For example, the first calendar quarter of 1999 includes all company-quarters whose fiscal quarters end from January through March For each quarter, we independently assigned the companies into quintiles on the basis of their SUE and SUEAF ranks.4 For use in regression analyses, we scaled the July/August

5 Financial Analysts Journal quintile ranks between 0 and 1 and then subtracted 0.5 to assign a score of 0 to the hypothetical median observation. Thus, the adjusted-quintile SUE and SUEAF ranks are between -0.5 and 0.5 and are labeled, respectively, QSUE and QSUEAF. Computing Buy-and-Hold Abnormal Returns. We first computed the buy-and-hold return (BHR) of the individual stock from two days after the preliminary earnings announcement date for quarter t through one day after the preliminary earnings announcement of quarter t + 1, if available, and 100 calendar days if unavailable. We then computed the BHRs of the equally weighted portfolios of companies with similar size and BV/MVs. The cut-off points to determine the size+bv/mvmatched portfolios and the daily returns of these matched portfolios came from Kenneth French's data library, based on a classification of the population into six (two size and three BV/MV) portfolios.5 The abnormal BHR (ABHR) is the buy-and-hold return of the individual stock minus the buy-andhold return of the size+bv/mv-matched portfolio. To avoid any bias from stocks that delisted after the preliminary earnings announcements but prior to the end of our drift period, we used the delisting return from CRSP if the stock was delisted prior to the subsequent earnings announcement or 100 days after the preliminary earnings announcement. If the delisting was a forced delisting from an exchange and CRSP had a missing delisting return, we assumed the return to be -100 percent. We assumed the stock earned the benchmark portfolio return after delisting. Statistical Tests. The aim of the statistical tests was to examine the extra return obtained by using SUE and SUEAF jointly as a measure of earnings surprise. First, we examined average subsequent-quarter ABHRs after sorting observa- tions independently by both SUE and SUEAF ranks. This step allowed us to compare for the entire sample period (73 quarters) the mean returns for the hedge portfolio on the basis of the joint distribution with the returns on the two separate hedge portfolios based on SUE and SUEAF individually. To consider the variability of these returns across quarters, we examined the performance of hedge portfolios formed on the basis of SUE and SUEAF by using a dynamic trading strategy. Specifically, for each quarter, we constructed an earnings surprise hedge portfolio with long positions in all stocks in the top 20 percent of SUE and short positions in those in the bottom 20 percent. We then used the same methodology to construct a hedge portfolio on the basis of SUEAF. Finally, for every quarter, we constructed a hedge portfolio that used both SUE and SUEAF-taking short (long) positions in stocks with both SUE and SUEAF in the bottom (top) 20 percent of their distributions. The mean difference across quarters in returns between the doublesurprise portfolio and each of the individualsurprise portfolios shows the greater ABHRs found by joint use of the two earnings surprises than found by each used separately. Finally, as in many prior drift studies, we performed a regression analysis in which the dependent variable was the ABHR and the independent variables were either or both earnings surprises. To assess the significance of regression coefficients, this study used average coefficients from quarterly regressions (a methodology similar to that of Fama and MacBeth 1973). Table 1 provides sample summary statistics. The median (50th percentile) amounts for both SUE and SUEAF are positive, with the former close to zero. The median ABHR for the entire sample is slightly negative, but its magnitude is only 30 bps. Table 1. Sample Summary Statistics, Q Q Percentile 50th Variable N 10th 25th (median) 75th 90th SUE 129, SUEAF 129, MV at t - 1 ($ millions) 129, , ,038.3 BV at t - 1 ($ millions) 128, ,695.0 ABHR (%) 129, Notes: Market and book value of equity are as of quarter-end. N is the number of company-quarters , CFA Institute

6 Double Surprise into Higher Future Returns Results Table 2 provides the number of observations, the mean ABHR, and the associated significance level of the mean ABHR for a 3 x 3 grouping of companies classified according to independent SUE and SUEAF distributions. We assigned companies to three groups on the basis of their SUE and SUEAF values, with the bottom 20 percent in one group, the middle 60 percent in another, and the top 20 percent in the third group. We expected the bottom group for each measure to have a negative drift, the top to have a positive drift, and the middle to have insignificant drift. The far-right column presents average ABHRs by time-series earnings surprise groups. These results indicate that through the next earnings announcement, the bottom 20 percent SUE portfolio had a mean drift of percent, the middle 60 percent SUE portfolio had a mean drift of 0.22 percent, and the top 20 percent SUE portfolio had a mean drift of 1.89 percent. These percentages imply an SUE-based average quarterly hedge portfolio return of 2.56 percent [1.89 percent - (-0.67 percent)]. The bottom panel presents average ABHRs by groups based on analyst forecast earnings surprise. Through the next earnings announcement, the bottom 20 percent SUEAF portfolio had a mean drift of percent, the middle 60 percent SUEAF portfolio had a mean drift of 0.27 percent, and the top 20 percent SUEAF portfolio had a mean drift of 2.62 percent. The result is an SUEAF-based average quarterly hedge portfolio return of 4.22 percent [2.62 percent - (-1.60 percent)]. Thus, as expected from prior literature, the drift (and correspondingly, the portfolio return) is larger when earnings surprises are based on analysts' forecasts (4.22 percent) than when they are based on time-series data (2.56 percent). Neither measure of the earnings surprise subsumes the other because only a third of the extreme (negative or positive) quintile companies is common to both SUE and SUEAF. And most importantly, the quarterly return on the combined SUE+SUEAF hedge portfolio is 5.86 percent [3.86 percent - (-2.00 percent)], which is larger than the return obtained using either measure of earnings surprise independently.6 The return is economically significant; it translates to about 23 percent annually. Note also that in absolute terms, the mean drift is higher for the long positions than for the short positions (3.86 percent versus 2.00 percent). These findings give credence to the applicability of such a strategy because, in practice, obtaining the returns on a long position is easier than obtaining the returns on a short position. Table 2. Distribution of Abnormal Returns: SUE+SUEAF Groups, Q Q SUEAF SUE Bottom 20% Middle 60% Top 20% Bottom 20 percent N 9,092 14,270 2,586 ABHR (%) Significance Middle 60 percent N 14,059 48,802 15,112 ABHR (Os) Significance Top 20 percent N 2,795 14,896 8,276 ABHR (%) Significance Notes: N shows the number of company-quarters in a particular group. Entries in boldface are statistically different from zero at the 5 percent level or better. July/August

7 Financial Analysts Journal The returns given in Table 2 are mean returns; therefore, they give equal weight to each companyquarter observation over the entire sample period. To consider the riskiness faced by an investor actually attempting to implement a hedge portfolio strategy, we next equal-weighted each quarter and examined the time series of returns. Panel A of Table 3 provides information about the average quarterly returns for the hedge portfolio based on SUE, the hedge portfolio base on SUEAF, and the hedge portfolio based on both SUE and SUEAF. The strategy consisted of holding short (long) positions in the bottom (top) 20 percent of the earnings surprise measure, where positions were closed each quarter. The data we obtained by equally weighting quarters are very similar to the data reported when observations were equally weighted (Table 2). Panel A of Table 3 shows that the average quarterly profit over 73 quarters for the SUE-based portfolio was 2.7 percent and for the SUEAF-based portfolio, 4.3 percent. The combined SUE+SUEAF-based portfolio, where short (long) positions were taken only for companies in the bottom (top) 20 percent of both SUE and SUEAF, yielded a return of 5.8 percenthigher than the returns from either portfolio alone. The difference between the combined portfolio and the SUE- (SUEAF-) based portfolio is 3.1 (1.5) percentage points. The three portfolio returns and the two differences are all statistically different from zero at the 1 percent significance level. Jegadeesh and Livnat (2006) showed that revenue surprises generally produce larger drifts than do expense surprises. Therefore, we added a measure of revenue surprise, SURGE. The calculation of SURGE was identical to that of SUE except that earnings were replaced by revenues. Panel B of Table 3 shows that whereas SURGE used alone led to a drift of only about 1.2 percent, when it was combined with SUEAF, it produced a drift of about 5.8 percent, which is statistically indistinguishable from that achieved by combining SUE with SUEAF. This combination, however, would require the investor to be more restrictive in picking investments than the SUE+SUEAF combination because combining SURGE and SUEAF leads to an average quarterly portfolio size (N = 162) that is nearly onethird smaller than when SUE and SUEAF are combined (N = 238). The last column of Panel B of Table 3 shows that using all three surprise variables produces a hedge return of almost 6.6 percent per quarter-almost 80 bps higher than either of the two-way sorts. But this gain comes at the cost of being far more selective than any of the other combinations (N = 74). The relative contributions of SURGE, SUE, and SUEAF to predicting future abnormal performance may be more apparent in the regression tests. We carried out the following three regressions, where Q applies to adjusted quintile ranks: BHRp,t =a+ b(qsuej,t) +tj, (4) BHRp,t =a+ b(qsueaf},t) +e1, (5) Table 3. Hedge Portfolio Returns, Q Q A. SUE and SUEAF SUE+ SUE SUEAF SUEAF Difference Difference Variable (1) (2) (3) (3-1) (3-2) ABHR (%) Standard deviation of ABHR (%) t-statistic Significance level Average number of companies B. SUE, SUEAF, and SURGE SURGE+ SURGE+ SURGE+ SUE+ Variable SURGE SUE SUEAF SUEAF ABHR(%) Standard deviation of ABHR (%) t-statistic Significance level Average number of companies Significance level of difference from Column Note: Entries in boldface are statistically different from zero at the 5 percent level or better , CFA Institute

8 Double Surprise into Higher Future Returns and BHRj,t = a + b (QsUE, t) + c (QSUEAFj,t (6) + jjst. Equations 4 and 5 were designed to examine the relationship between the drift (ABHR) and the earnings surprise as measured by either SUE or SUEAF. Equation 6 was designed to examine the joint effect of the two measures of earnings surprise on the drift. As discussed previously, QSUE and QSUEAF represent the adjusted quintile ranks scaled to be between -0.5 and 0.5. In such regressions of abnormal returns on adjusted quintile ranks, the slope coefficient can be interpreted as the return to a hedge portfolio that is long the most positive quintile and short the most negative quintile. As is standard in the Fama-MacBeth (F-M) methodology, we fit the regressions separately for each quarterly cohort. The time-series averages of the quarterly regression coefficients are given in the top six rows of Table 4. In all three regressions of earnings surprise on ABHR, the average coefficients on the measure of surprise (QSUE or QSUEAF) have the predicted positive sign and are statistically different from zero at the 1 percent significance level. In the multivariable regression, the average coefficients on QSUE and QSUEAF are not equal and are both statistically significant. In addition, for each of the quarterly multivariable regressions of SUE and SUEAF, we calculated the difference between the coefficient on QSUEAF and the coefficient on QSUE. We found the mean of these differences over the 73 quarters to be 2.3 and statistically different from zero at the 1 percent significance level. This finding indicates that these two measures of earnings surprise represent different forms of mispricing and are not interchangeable in the establishment of a hedge portfolio strategy. Furthermore, the sum of the average coefficients in the SUE+SUEAF regression, 5.256, is higher than the coefficient on either measure alone in the single-variable regressions. Consistent with this interpretation of the coefficients (the return to a hedge portfolio that is long the most positive quintile and short the most negative quintile), this finding indicates that the mean hedge portfolio return obtained by using both measures would be higher than the mean hedge portfolio return from using either measure alone. In the bottom six rows of Table 4, we add the contribution of revenue surprise, SURGE. The coefficient on SURGE is not significantly different from zero in any of these regressions, which suggests that SURGE does not add to the ability of SUE and SUEAF to predict future abnormal performance. SURGE's apparent contribution to predicting future abnormal returns documented in Table 3 comes from requiring investors to be more selective in picking investments. Investors are probably better off being more discerning with (i.e., taking positions only in the more extreme values of) SUE and SUEAF than combining them with SURGE. Table 4. Regression Results: ABHR on SUE, SUEAF, and SURGE and Combinations, Data for Q Q Measure Intercept QSUE QSUEAF QSURGE N R2 Significance SUE only (F-M) Significance (F-M t-statistic) SUEAF only (F-M) Significance (F-M t-statistic) SUE+SUEAF (F-M) Significance (F-M t-statistic) SUE+SURGE (F-M) Significance (F-M t-statistic) SUEAF+SURGE (F-M) Significance (F-M t-statistic) SUE+SUEAF+SURGE (F-M) Significance (F-M t-statistic) Notes: The sign on the coefficients is expected to be positive. For each of the 73 quarterly regressions using SURGE and either or both measures of the earnings surprise, the difference was calculated as the difference between the coefficient on QSURGE and the coefficient on either QSUE or QSUEAF. The mean of these differences is statistically different from zero at the 2 percent significance level for all regressions. "F-M" refers to the use of the Fama-MacBeth (1973) methodology. July/August

9 Financial Analysts Journal Sensitivity Analysis We carried out the following sensitivity tests. * Deciles. We repeated all the statistical analyses with deciles rather than quintiles. Our conclusions are qualitatively identical; indeed, many of the results were stronger when we used deciles. For the distribution of abnormal returns (equivalent to Table 2), the quarterly return on the combined hedge portfolio was found to be 7.6 percent and statistically significant. Similarly, the average quarterly return for the hedge portfolio based on SUE+SUEAF as recalculated every quarter (equivalent to Table 3) was found to be a statistically significant 7.3 percent. The drawback to using deciles rather than quintiles is that the number of observations is significantly reduced for all statistical procedures. * Extreme observations. We replicated all analyses after excluding company quarters in the top and bottom 0.5 percent of the distribution of cumulative abnormal returns. This procedure reduced the sample from 129,888 to 128,652. All results were qualitatively identical. * Number of analysts. Because there may be significant differences between companies covered by only one analyst and those covered by multiple analysts, we replicated the statistical analyses but used only observations with more than one available analyst forecast in the company-quarter. This screen reduced the sample from 129,888 to 107,732. All the results were qualitatively identical. * Negative book value. We replicated the analyses only for the company-quarters in which the book value of equity at the end of quarter t - 1 was available from Compustat and was positive. This screen reduced the sample from 129,888 to 105,086. All the results were qualitatively identical. * SUEAF scaled by price. We defined SUEAF as I/B/E/S actual EPS minus 1/B/E/S forecasted EPS, as in Equation 3, except that we scaled by price per share instead of the dispersion in analyst forecasts. The results were qualitatively similar to those reported in the tables. * Time periods. Finally, we examined the distribution of abnormal returns (equivalent of Table 2) and performed the hedge portfolio return analysis (equivalent of Table 3) for four subperiods: , , , and The results already discussed were present in all four subperiods. Conclusion We have shown that for companies followed by professional security analysts, using analyst forecast errors to define earnings surprise leads to greater predictability of future stock returns than does using time-series forecast errors. Return predictability can be further enhanced by combining analyst forecast errors with time-series errors. These findings proved to be robust to a range of specifications. Although analyst forecasts are not available for all companies, those companies for which they are available tend to be more liquid than other companies. Therefore, investors will, on average, find exploiting any mispricings among these stocks easier and less expensive than among others. This study can be beneficial for practitioners and academics alike. Professional investors can use the results to improve their selection of stocks. Instead of using the earnings surprise based on either analyst forecasts or time-series forecasts, they can, by using both measures, focus on a restricted set that has extreme surprises. Analysts can use both measures to identify companies where future earnings surprises are more likely to continue in the same direction as the current earnings surprise and can revise their earnings forecasts accordingly. Academics can use our results to infer whether the market underreaction to earnings surprises is greater when the earnings surprise is measured by using both time series and analyst forecasts of earnings than it is when each measure is used alone. Such analysis is likely to have implications for our understanding of how market participants process information. The authors gratefully acknowledge the Point in Time quarterly data provided by Charter Oak Investment Systems and theforecast data available through Thomson Corporation's IIB/EIS. These data were provided as part of a broad academic program to encouragearnings expectations research. This article qualifies for 1 PD credit CFA Institute

10 Double Surprise into Higher Future Returns Notes 1. For example, Rendleman, Jones, and Latane (1982); Foster et al. (1984); Bernard and Thomas (1989). 2. For each analyst, we included only the most recent forecast for a given company-quarter. 3. Payne and Thomas (2003) reported that the use of four decimal places in the I/B/E/S detail tape makes it less susceptible to rounding bias than data adjusted for stock splits. 4. Although some literature bases the ranks on the distribution in quarter t - 1, Bernard and Thomas (1990) reported that the drift is insensitive to whether the assignment of companies is based on the current quarter's SUE values or ex ante rankings, such as the distribution from quarter t The library is at faculty/ken.french/data-library.html. 6. It can be argued that what causes the increase in the drift is not the combination of SUE and SUEAF but the reduction in the number of observations that now have to fall into two extreme earnings surprise categories instead of just one. Therefore, we repeated the analysis of quarterly returns in Table 3 with short (long) positions in the lowest (highest) 8,600 observations for SUE and SUEAF separately. We obtained a hedge return of 3.66 percent for SUE and 5.09 percent for SUEAF, which is significantly smaller than the 5.80 percent for the joint SUE+SUEAF strategy with a similar number of observations. Thus, we found that the combination of SUE and SUEAF, not the reduction in the number of observations, causes the increase in returns. References Affleck-Graves, John, Larry R. Davis, and Richard R. Mendenhall "Forecasts of Earnings per Share: Possible Sources of Analyst Superiority and Bias." Contemporary Accounting Research, vol. 6, no. 2-1 (Spring): Ball, Ray "Anomalies in Relationships between Securities' Yields and Yield-Surrogates." Journal of Financial Economics, vol. 6, nos. 2-3 (June-September): "The Earnings-Price Anomaly." Journal of Accounting and Economics, vol. 15, nos. 2-3 (June-September): Ball, Ray, and Philip Brown "An Empirical Evaluation of Accounting Numbers." Journal of Accounting Research, vol. 6, no. 2 (Autumn): Bernard, Victor L., and Jacob K. Thomas "Post-Earnings- Announcement Drift, Delayed Price Response or Risk Premium?" Journal of Accounting Research, vol. 27, Supplement: "Evidence That Stock Prices Do Not Fully Reflect the Implications of Current Earnings for Future Earnings." Journal of Accounting and Economics, vol. 13, no. 4 (December): Brown, Lawrence D., Paul A. Griffin, Robert L. Hagerman, and Mark E. Zmijewski "An Evaluation of Alternative Proxies for the Market's Assessment of Unexpected Earnings." Journal of Accounting and Economics, vol. 9, no. 2 (July): Dugar, Amitabh, and Siva Nathan "The Effect of Investment Banking Relationships on Financial Analysts' Earnings Forecasts and Investment Recommendations." Contemporary Accounting Research, vol. 12, no. 1 (Fall): Fama, Eugene F., and Kenneth R. French "The Cross- Section of Expected Stock Returns." Journal of Finance, vol. 47, no. 2 (June): Fama, Eugene F., and James D. MacBeth "Risk, Return, and Equilibrium: Empirical Tests." Journal of Political Economy, vol. 81, no. 3 (May/June): Foster, George, Chris Olsen, and Terry Shevlin "Earnings Releases, Anomalies, and the Behavior of Security Returns." Accounting Review, vol. 59, no. 4 (October): Jegadeesh, N., and J. Livnat "Post-Earnings- Announcement Drift: The Role of Revenue Surprises." Financial Analysts Journal, vol. 62, no. 2 (March/April): Livnat, Joshua, and Richard R. Mendenhall "Comparing Post-Earnings Announcement Drift for Surprises Calculated from Analyst and Time-Series Forecasts." Journal of Accounting Research, vol. 44, no. 1 (March): Payne, Jess, and Wayne B. Thomas "The Implications of Using Stock-Split Adjusted I/B/E/S Data in Empirical Research." Accounting Review, vol. 78, no.4 (October): Rendleman, Richard J., Charles P. Jones, and Henry A. Latane "Empirical Anomalies Based on Unexpected Earnings and the Importance of Risk Adjustment." Journal of Financial Economics, vol. 10, no. 3 (November): July/August