The Impact of Analysts Forecast Errors and Forecast Revisions on Stock Prices William Beaver, 1 Bradford Cornell, 2 Wayne R. Landsman, 3 and Stephen R. Stubben 3 April 2007 1. Graduate School of Business, Stanford University, Stanford, CA 94305. 2. California Institute of Technology, Pasadena, CA 91125 3. Kenan-Flagler Business School, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599. We thank I/B/E/S International for providing data on analysts earnings estimates, and the Center for Finance and Accounting Research, University of North Carolina for providing financial support. We also thank workshop participants at the 2005 Stanford Summer Camp, the University of Florida, and an anonymous reviewer for helpful comments. Corresponding author: William Beaver, william_beaver@gsb.stanford.edu.
The Impact of Analysts Forecast Errors and Forecast Revisions on Stock Prices Abstract We present a comprehensive analysis of the contemporaneous association between security returns, quarterly earnings forecast errors, and quarter-ahead and year-ahead earnings forecast revisions in the context of a fully specified model. We find that all three variables have significant pricing effects, indicating each conveys information content. The findings hold across years, across industries, and are robust to two procedures extending the event window. Findings also show that the fourth quarter is significantly different from the other three quarters. In particular, in the fourth quarter the relative importance of the forecast error is lower, while the relative importance of the quarter-ahead forecast revision increases. We find also a marked upward shift over time in the forecast error coefficients, even in the presence of the forecast revision variables, whose coefficient also exhibit a significant but less dramatic shift. This finding is consistent with the I/B/E/S data base reflecting an improved quality of earnings forecasts, as well as an improved measure of actual earnings.
1. Introduction One of the fundamental questions in finance and accounting is the impact of earnings surprises on stock prices. The question not only is important for evaluating theories that relate reported accounting numbers to firm value, but also has widespread implications for regulation and the law. For instance, in legal disputes related to financial reporting a central issue is how much the stock price would have been affected if the company released its correct earnings in place of allegedly inflated earnings. Proper analysis of that issue requires an appropriately specified model of the relation between earnings innovations and stock prices. Empirical studies of this question employ analysts earnings forecast data as proxies for market expectations and, thereby, to measure earnings surprises. In an early paper, Cornell and Landsman (1989) demonstrate that the earnings surprise should not be identified solely with analysts forecast errors. They stress that a properly specified model of residual returns must simultaneously take account of both earnings forecast errors and earnings forecast revisions. They present evidence to show that if the forecast revisions are excluded, the response coefficient on the forecast error is higher because forecast revisions are in part based on forecast errors. In this paper, we present a comprehensive analysis of the relation between stock returns, analysts forecast errors and analysts forecast revisions. Despite the fact that there has been extensive new research on the relation between analysts forecasts and stock prices, which we review below, much of this literature has not taken account of the combined impact of forecast errors and forecast revisions. As we show, this can lead to potentially misleading results. In addition, there have been improvements in the nature, quantity, and quality of the data used to measure forecast errors and forecast revisions. Whereas Cornell and Landsman (1989) is 3
based on only three years of data and a limited number of firms, our sample comprises 20 years of data covering a much greater number of firms. Moreover, with respect to the I/B/E/S data that we use in this paper, there has been an increasing effort over time by I/B/E/S to ensure a consistency between the forecast and the realization of earnings, as well as a consistency across analysts in the earnings number being forecast. This consistency is attained by ensuring the same earnings components are included (and excluded) in the actual and forecasted earnings. Presumably, the effects of these efforts could alter the observed relation between security returns, forecast errors and forecast revisions. More specifically, as the I/B/E/S database becomes more successful in providing an apples to apples comparison, the quality of the forecast error is expected to improve because it becomes a better proxy for unexpected earnings. The resulting reduction in measurement error should affect the estimated coefficients in regression models. In addition to improving the quality of the data, I/B/E/S has extended coverage over time making the data more comprehensive. This alteration in the composition of the data may also affect the empirical estimation of the security return model. We examine whether there has been an increase or decrease over time in the information content of forecast errors and forecast revisions to assess the extent to which changes in the nature of the data have affected the observed relations. Furthermore, it has been suggested that companies have come under added pressure to manage earnings and that this may affect the relation between residual returns, forecast errors and forecast revisions (Matsumoto, 2002; Abarbanell and Lehavy, 2003; Burgstahler and Eames, 2003). For example, it may lead to reduced information content of the earnings forecast error over time. Proper examination of whether this occurs requires a more fully specified model that takes account of both forecast errors and forecast revisions over time. 4
Finally, it also has been suggested that managers are increasingly actively managing analysts expectations to avoid negative earnings surprises, which may affect the relation between residual returns, forecast errors and forecast revisions (Brown, 2001; Matsumoto, 2002). Presumably, if this activity has increased over time and adversely affected the quality of both the forecast errors as well as the forecast revisions, the change in the relation between residual returns, forecast errors and forecast revisions over time should show up over time. Analysis of these questions requires a comprehensive examination of the relation between residual stock returns in the period surrounding quarterly earnings announcements, earnings forecast errors, and revisions in quarter-ahead and subsequent year-ahead analysts earnings forecasts during the period from 1984 to 2003. The length of the sample period permits us to examine whether changes in the properties of the earnings forecasts result in any perceptible trends in the coefficients on the forecast error and the forecast revisions. In addition, the growth in I/B/E/S coverage also permits us to control for potential mean differences in industry effects and to examine whether the observed relation is consistent across industries. Furthermore, the availability of an I/B/E/S actual earnings number, which was not provided when the database first became available, permits us to compare the properties of different specifications including forecast errors based on I/B/E/S actuals versus Compustat earnings. We also examine two important specification issues: the distinct nature of fourth quarter earnings and the measurement of the residual return interval. With respect to the first issue, we consider whether the relation between residual returns, forecast errors and forecast revisions differs during the fourth quarter for a variety of reasons that we discuss later in the paper. If this is so, failure to take account of the fourth quarter effect will lead to a misspecified model and, quite likely, biased coefficients. To study this possibility, we develop specifications that permit 5
the fourth quarter slope coefficients to differ from those of the interim quarters, and that take account of the intertemporal overlap in measurement of the quarter-ahead and year-ahead forecast revisions that occurs during the fourth quarter. With respect to the residual return window, models that incorporate both forecast errors and forecast revisions face a unique data problem. The problem arises because the forecast error, by definition, is observed at the time of the earnings announcement, but the forecasts revisions are not made available until a later date. This raises two issues. The first issue is that at the time of the earnings announcement the market must use the information in the forecast error to anticipate its long-run impact, and thereby its effect on analysts forecast revisions, without observing the revisions. Therefore, the residual return reflects both the forecast error and the forecast revisions expected at the earnings announcement date. However, by necessity, the model includes actual forecast revisions, which likely measure the market s expectations with error. To take account of this feature of the data, we extend the basic model in two ways. First, we extend the window over which the residual return is measured to the date at which the forecast revisions are observed. This assures us that the residual return will reflect both actual forecast errors and actual forecast revisions. A problem with this approach is that the window must be extended, on occasion, to more than two months after the earnings announcement to be sure the I/B/E/S consensus reflects forecasts made after the earnings announcement. By extending the return window, the coefficients on the forecast revisions will reflect information available subsequent to the earnings announcement. To counter this problem, the second approach turns to disaggregated data. Rather than using the I/B/E/S consensus forecasts, we employ the individual analysts forecasts to construct a custom consensus forecast following the 6
earnings announcement. In this way, we can shorten the window by using the subset of the individual forecasts that are available soon after the earnings announcement. The major findings are: First, in every model we estimate both the forecast error and the forecast revision coefficients are highly significant. In other words, neither the forecast error nor the forecast revisions dominate in that each provides information content not contained in the other. Second, based on twenty years of data, we find that, even in the presence of the forecast revision variables, the coefficient of the forecast error still increases substantially over time, with a marked shift in post 1991 period. Third, in contrast, the coefficients on the two forecast revisions exhibit a similar but less dramatic shift. We present evidence suggesting that the increase in the coefficients is attributable to joint effects of the improved quality of the I/B/E/S actual earnings and analysts earnings forecasts over the sample period. This finding is important because it indicates that the significance of the forecast revisions in explaining the cross-sectional variation in earnings announcement residual returns is not an artifact of measurement error in the forecast error. Rather, the significance of the forecast revision coefficients is a robust finding that holds up through time despite changes in database quality and changes in the institutional features of the earnings reporting environment. Findings from separate industry regressions indicate that although there are cross-industry differences in the magnitude of coefficients on the forecast error and the forecast revisions, the basic relation holds across all industry groups. Fourth, the results further support the view that the fourth quarter is different than other quarters. The evidence is consistent with the market reacting in the fourth quarter more strongly to the change in the next quarter forecast revision and less strongly to the forecast error. This finding suggests that a revision in the quarter-ahead forecast in the fourth quarter, which is the 7
forecast revision for the first quarter of the next fiscal year, conveys more information than earlier quarters forecast revisions, which refer to later quarters in the same fiscal year Fifth, findings from estimations that extend the announcement event window indicate the primary results are robust, but the impact of the forecast revisions, as compared to the forecast errors, increases. This supports the notion that when the market observes the actual forecast revisions prices are adjusted to take account of the difference between the forecast revisions that are observed and the forecast revisions that were expected at the time of the earnings announcement. These increased coefficients are also consistent with the forecast revisions reflecting information available after the earnings announcement. Consistent with these arguments, the subsequent move in stock price is correlated with the observed revisions, but not necessarily with the (earlier) forecast error. To summarize, our results emphasize the importance of using a properly specified model when assessing the impact of the release of earnings information on stock prices. Models that fail to include forecast revisions, fail to take account of the changing nature of the I/B/E/S data, or fail to adjust for fourth quarter effects will produce earnings response coefficients that to not correctly characterize the relation between reported earnings and firm value. The remainder of the paper is organized as follows. In the next section, we review the key findings of the research on the relation between analysts forecast errors and stock returns. Section three presents the research methodology and methods for measuring the variables. Section four describes the sample data. Section five presents the results and discusses their implications. The conclusions are summarized in the final section. 2. Prior Research 8
Using I/B/E/S consensus analyst forecast data, Cornell and Landsman (1989) study the pricing effects of earnings forecast errors and earnings forecast revisions in the period surrounding quarterly earnings announcements. The key finding of their study is that both the one-quarter-ahead and one-year-ahead forecast revisions have important explanatory power in addition to the earnings surprise. An important conclusion based on their findings is that a properly specified model of residual returns in response to the release of quarterly earnings must simultaneously take account of both earnings forecast errors and earnings forecast revisions. They present evidence to show that if the forecast revisions are excluded from the basic model, the coefficient on the forecast error is higher because the error serves as a proxy for the forecast revisions and must be interpreted accordingly. In the years following the Cornell and Landsman study, surprisingly few studies have used the more completely specified model. A notable exception is Liu and Thomas (2000), which models stock returns as a function of annual forecast errors, annual forecast revisions, and an estimated annual revision in terminal value. Liu and Thomas finds that both the forecast error and forecast revisions provide incremental explanatory power. This study differs from Liu and Thomas in several respects: (1) Whereas Liu and Thomas relates annual stock returns with earnings variables, we examine the shorter-term announcement effects of the earnings variables in the spirit of an earnings announcement event study. Given the variability of stock returns, our shorter horizon tests have considerably more power. (2) Liu and Thomas examines only annual earnings; our research design measures earnings variables for annual and interim quarters. Hence, our research designs permits us to address additional issues, such as the differential behavior of the fourth quarter. (3) Liu and Thomas reports results based on pooled crosssectional and time-series data and does not examine how the coefficients may have changed over 9
time. Further, year-by-year estimation permits the calculation of test statistics that are not affected by cross-sectional correlation in the data leading to less biased test statistics than those based on pooled estimation. (4) Liu and Thomas includes earnings variables, including revisions in long-term earnings forecasts and terminal values, that are based on the authors extrapolations and are not reported by I/B/E/S. Hence, the results reflect the joint effect of I/B/E/S reported variables and their extrapolations using I/B/E/S and other data. Although the number of studies that model stock returns as a function of both forecast errors and revisions is relatively small, there is a much larger literature on the properties of forecast errors and analysts forecasts. We briefly summarize key studies in both of this literature that provide some background to our study. A number of papers study the properties of earnings response coefficients using alternative earnings measures (Bradshaw and Sloan, 2002; Brown and Sivakumar, 2003; Lougee and Marquardt, 2004: Abarbanell and Lehavy, 2005). Of particular relevance to our study is Bradshaw and Sloan (2002), which documents that annual earnings response coefficients are higher when the forecast error is defined using I/B/E/S (i.e., Street earnings) rather than Compustat net income (i.e., GAAP earnings), and the difference in price response based on the two measures has increased over time. In particular, in the post- 1992 period there is a significant increase in the earnings response coefficient for I/B/E/S earnings forecast errors. Bradshaw and Sloan (2002) attributes these findings to analysts excluding over time an increasing number of special items from their earnings estimates, and to the increasing prevalence of special items, which predominately occur in the fourth quarter. The key distinction between our study and prior studies examining the properties of earnings response coefficients, including Bradshaw and Sloan (2002), is that we include in our regressions analysts forecast revisions for quarter-ahead and year-ahead earnings. Not only 10
does this permit us to study the price response to forecast revisions, but this also changes the interpretation of the coefficient on the forecast error. In particular, in a fully specified model the forecast revisions control for the future implications of the forecast error, which results in a coefficient on the forecast error that is not affected by the persistence of current earnings. This model allows us to examine if there is a shift in the earnings response coefficients in the presence of earnings forecast revisions for reasons other than a change in earnings persistence over time. Further, we are able to examine whether there has been a similar upward trend over time in the coefficients on the earnings revisions variables themselves. Neither is possible in the context of a model that contains only earnings forecast errors. One issue raised by Cornell and Landsman is whether the structural relation between the earnings variables and stock return in the fourth quarter could differ from that of the interim quarters. They raise the possibility that fourth quarter could differ because of the increased frequency of special items and because the fourth quarter result will reflect the effects of the audit process. If the fourth quarter is significantly different, and if this fact is not taken into account in the model specification, the estimated relation between stock returns and forecasts errors will not be properly measured. In the context of a model that includes only the earnings forecast error, Mendelhall and Nichols (1988) finds that the market reacts relatively less strongly to bad news in the fourth quarter because of the ability of managers to delay the reporting of bad news in earlier quarterly earnings, but which is effectively leaked to the market in earlier quarters. However, it is difficult to predict whether their results would hold in the presence of forecast revisions. Prior research examining the properties of analysts forecasts is substantial. Brown (1996) synthesizes a vast literature of the forecasting accuracy of analysts versus naïve 11
statistical models, concluding that analysts forecast outperform statistical models, that the forecast error has not increased over time, and that over subperiods of time analysts forecasts have been pessimistically, rather optimistically biased. Lys and Sohn (1990) find that even though security returns can predict a portion of the forecast revision, the analysts forecasts are incrementally informative. One key paper, Abarbanell and Bernard (2000), suggests that analysts forecasts do not fully reflect the implications of earnings forecast errors in their forecast revisions. Subsequently, Gleason and Lee (2003) document a post-revision price drift and suggest the market does not fully reflect the information content of the forecast revision. In particular, their evidence suggests that the market does not make a sufficient distinction between revisions that provide new information and those that merely move toward to consensus. Another strand of analyst research has examined the contention that mangers have increasingly guided analysts forecasts downward so that earnings meet or beat analysts forecasts (Brown, 2001; Matsumoto, 2002). Presumably, to the extent that this pressure on analysts has affected their forecasts, it could also to affect the relation between residual returns, forecast errors and forecast revisions. Other research related to analysts focuses on the suggestion that companies have faced increasing pressure over time to manage earnings and that this may have affected the relation between residual returns, forecast errors and forecast revisions (Matsumoto, 2002; Abarbanell and Lehavy, 2003; Burgstahler and Eames, 2003). In particular, successful earnings management could affect the earnings surprise coefficient over time as earnings management increases. In addition, if analysts do not fully incorporate the effects of earnings management in their forecast revisions, this also could affect their coefficients over time as earnings management increases. The only way to explore these issues fully is in the context of a model 12
that includes both forecast errors and forecast revisions, takes account of a possible fourth quarter effect, and then examines how the coefficients change over time and across industries. These streams of research motivate our interest in examining several issues: (1) Is the rise in the coefficient on the I/B/E/S forecast still observed in the presence of the forecast error revisions? (2) Is there a similar increase in the coefficients on the forecast revisions over time? (3) In a fully specified model, is the structural relation of the model in the fourth quarter different from that of the interim quarters and has that model also changed over time? (4) Are the findings robust with respect to the alternative specifications of the announcement window? 3. Research Design 3.1 The Model Based on a valuation model that expresses equity market value as the present value of future cash flows, Cornell and Landsman derives a model where change in equity value is equal to a linear function of the cash flow forecast error and a series of revisions in expectation about future periods cash flows. Assuming that cash flow forecast errors (changes in future expected cash flows) are collinear with the earnings forecast errors (forecast revisions), they then derive an empirical estimation equation that appears as equation (1) below. Subsequent to Cornell and Landsman, Ohlson (1995) and Feltham and Ohlson (1995) develop a characterization of equity market value as a linear function of equity book value and the present value of future expected abnormal earnings. Moreover, Feltham and Ohlson (1996) demonstrates an equivalency between the cash flow and abnormal earnings representations. Here, we present a valuation model based on the Feltham and Ohlson abnormal earnings formulation. Empirically the stream of future expected abnormal earnings is truncated at some 13
point and a terminal value is estimated. From this price levels equation, it is straightforward to derive an expression for the unexpected security return as a function of unexpected current earnings and the change in the future expected abnormal earnings, and in the case of truncation a change in expected terminal value. In particular, the model developed by Liu and Thomas (2000) expresses unexpected security returns as: [their Equation (10)] UR it = β 0 + β 1UE it + β 2RAE2 it +β 3RAE3 it + β 4RAE4 it +β 5RAE5 it + β 6RTERM it + e it, where UR is the expected stock return, UE is the earnings surprise with respect to current abnormal earnings, RAE is the revised expectations about future abnormal earnings for the next four accounting periods, and RTERM is the revision in the estimated terminal value at the end of the horizon. The Liu and Thomas model is developed in context of annual returns and annual revisions in future expected earnings. In our context, which is announcement period returns for quarterly announcements, UE is represented by the forecast error on current quarterly earnings. In the most general model, there would be separate estimates for each of the revisions in future quarterly earnings for a finite period and the revision in expected terminal value. Our estimating model is a parsimonious version of the Liu and Thomas model, which, as described in detail below, reflects the structure of the I/B/E/S analysts forecast data, including the availability of the data, the frequency with which the forecast variables are revised and the collinearity among the forecasted variables. In particular, our estimating equation is: AR it = a 0 + a 1 FE it + a 2 FRQ it + a 3 FRY it + e it (1) where AR is the unexpected security return, FE is the forecast error for current quarterly earnings, FRQ is the revision in the I/B/E/S consensus forecast for the next quarter, and FRY is the revision in I/B/E/S consensus forecast for the next fiscal year. 14
One set of potentially omitted variables is the revisions in the quarterly earnings for the remaining portion of the current fiscal year. For example, for the first quarter, FE is the first quarter forecast error, FRQ is the revision in the forecast for the second quarter, and the forecast revisions for the third and fourth quarter are omitted. There are several problems with including these additional variables. First, the number of observations for which the forecast revision is available more than one period ahead is limited. Second, the length of the remaining portion of the current fiscal year shrinks as for each of the later quarters (e.g., for the second quarter there are only two quarters remaining), and it unclear how one would incorporate the varying time horizon into estimating equation (1). Third, the revisions in forecasts for the remaining quarters are significantly correlated with one another. However, notwithstanding these difficulties, we conducted a complete specification for the first quarter for those observations where a forecast revision was reported. We found the overall explanatory power to be essentially the same as that of Equation (1). Hence, we rely on the more parsimonious form of the estimating equation. The main point to emphasize is that the coefficient on FRQ reflects the pricing multiplier that reflects the revisions for the remaining quarters as well. 1 Similarly, revisions in annual earnings beyond FRY are potentially omitted variables from Equation (1). As Liu and Thomas (2000) point out, the limited availability of I/B/E/S annual forecast revisions beyond one year results in a substantial reduction in number of observations. To require additional FRY for two-years hence would reduce the sample size by 65 percent and to further require FRY terms beyond two years would reduce the sample size by over 90 percent. 1 Alternatively, we could construct our own estimates of the forecast revision for the remaining quarters based upon some extrapolation of the FRQ. This is the approach employed by Liu and Thomas (2000) to project annual forecasts beyond those reported by I/B/E/S. We have chosen not to use such an approach here because then the findings would be a joint function of I/B/E/S data and our extrapolation procedure. Moreover, in conducting preliminary calculations over our interval of revision (two months as opposed to one year in Liu and Thomas), we found the extrapolated variables to be so highly correlated with the reported variable (FRQ) that no increase in 15
Moreover, regressions including these variables does not produce any increase in explanatory power. 2 Liu and Thomas (2000) constructs estimates of long-term earnings revisions based on the reported I/B/E/S short-term earnings forecasts and the I/B/E/S long-term growth rate. We examined the feasibility of using similar extrapolated variables. For our period of revision (two months versus one year), we found that the revision in long-term growth rates was zero for 68 percent of our observations. Because of this, the resulting extrapolations would produce revision variables that would either be zero or highly collinear with FRY. As a result, incorporation of these extrapolated variables would not add significantly to explanatory power and would only provide an illusion of additional variables that are in fact linear extrapolations of the FRY variable and a growth variable that is predominately zero. 3 As with the possibility of including additional terms for interim quarter forecasts revisions beyond FRQ, it is more straightforward to simply include only FRY and to interpret its coefficient accordingly namely, the coefficient also reflects the extent to which FRY is correlated with revisions in expected subsequent earnings. Further, there is no revision in terminal value calculation in Equation (1). Not only it is a purely extrapolated value in the sense that I/B/E/S does not report terminal value, but revisions in terminal value are greatly affected by revisions in the long-term growth rate, which was zero for a 68 percent of our observations. For this reason, we do not revision in terminal value in the estimating equation. Consistent with the standard approach in the literature, we measure analysts forecastbased variables using consensus forecasts in the I/B/E/S summary file. On the Thursday before explanatory power was provided. It is more straightforward to simply include only FRQ and to interpret its coefficient accordingly. 2 Not surprisingly, the coefficients on these variables are positive, much smaller than for FRY and closer to zero. As a result, the coefficient on FRY is slightly lower but the overall explanatory power remains the same. 3 One might dismiss 68 percent of the observations being zero as not being a sufficient reason for not using the growth rate. However, we feel the smaller propensity to update long-term forecasts is not a reflection of changing expectations and hence is a stale variable measured with considerable error. 16
the third Friday of each month, I/B/E/S calculates the consensus forecast as the mean or median of all outstanding estimates for a particular fiscal period. Forecasts are available for a variety of fiscal periods, including the current quarter, the next quarter, the current fiscal year, and the next fiscal year. Additional horizons are available, but analysts forecasts for these periods are less frequent. The ideal measurement of the response of security prices to earnings announcements and earnings forecast revisions would use a consensus forecast made just prior to an earnings announcement, and another made just after. However, consensus forecasts are compiled only monthly. Preannouncement forecasts, then, are the most recent consensus forecasts compiled before the earnings announcement date. Postannouncement forecasts are compiled the second month after the earnings announcement. Consensus forecasts for the first month after the earnings announcement are not used because they may contain individual forecasts issued both before and after the earnings announcement. As shown in the hypothetical example in figure 1, preannouncement forecasts are gathered on the last forecast date before the earnings announcement, March 19. In general, the time between the preannouncement forecast date and the earnings announcement will vary up to one month. Since the April forecast period might contain forecasts made both before and after the earnings announcement, postannouncement forecasts are instead gathered on May 21. Abnormal stock returns are calculated from the close of the preannouncement date, March 19, until one trading day after the earnings announcement, March 24, and abnormal stock returns over the extended window regressions described in section 3.3 are calculated until the end of the postannouncement period, May 21. 17
above, where Our initial tests are based on the Cornell and Landsman regression given by equation (1) i, t are indices referring to a sample firm and an announcement quarter. AR it = the abnormal stock return for firm i associated with quarterly earnings announcement t. AR it is measured from the close of the day of the announcement of the most recent I/B/E/S consensus forecast prior to the earnings announcement date (which we refer to as the last day of the preannouncement forecast period) through the trading day following the earnings announcement (see figure 1). The abnormal return is computed by subtracting the compounded daily mean return for the corresponding size decile, r dec, from the compounded daily firm return, r, over the period described above. That is, AR = Π (1 + r ) Π (1 + r it s s s dec s ). FE it = the forecast error for firm i and quarterly earnings announcement t. FE it, which is measured over the same time interval as AR it, is given by (EPS it E(EPS it θ 0 ))/P it, where EPS it is the realized quarterly earnings per share taken from I/B/E/S, E(EPS it θ 0 ) is the median preannouncement I/B/E/S consensus forecast of EPS it, and P it is the security price of firm i on the last day of the preannouncement forecast period (θ 0 refers to the set of information available on the preannouncement forecast date). 4 FRQ it = the forecast revision for firm i for quarter t+1, made subsequent to the earnings announcement for quarter t. FRQ it is given by (E(EPS i,t+1 θ 2 ) E(EPS i,t+1 θ 0 ))/P it, where E(EPS i,t+1 θ 0 ) is the preannouncement forecast of EPS for quarter t+1, and E(EPS i,t+1 θ 2 ) is the postannouncement forecast of EPS for quarter t+1. θ 2 refers to the set of 18
information available at the postannouncement date. As discussed above, this is the second, not the first I/B/E/S consensus forecast available after the earnings announcement. FRY it = the forecast revision for firm i for the next fiscal year. FRY it is given by (E(EPSY i,t+k θ 2 ) E(EPSY i,t+k θ 0 ))/ P it, where E(EPSY i,t+k θ 0 ) is the preannouncement forecast of EPS for the fiscal year which ends in quarter t+k, and E(EPSY i,t+k θ 2 ) is the postannouncement forecast of EPS for the fiscal year ending in quarter t+k. 5 Note that the number of quarters ahead for the subsequent fiscal year depends on the quarter of observation. For example, if the current quarter is the first quarter of the year, the subsequent fiscal year begins with k =4 and ends with k=7 quarters ahead, but if the current quarter is the third quarter of the year, the subsequent fiscal year is begins with k=2 and ends with k=5 quarters ahead. 3.2 Incorporation of by Year and by Industry Fixed Effects We estimate equation (1) several ways. These include (a) a pooled estimation with year and industry fixed effects, where year is determined by the quarter end date and industry is based on industry groupings used in Barth, Beaver, Hand, and Landsman (2005) (see table 1); (b) yearby-year estimations with industry fixed-effects; and (c) for each industry, year-by-year estimations. The fixed effects are included to capture sources of time dependence or cross sectional dependence of a particular form (i.e., a constant for a given year and a constant for a 4 All variables used to compute FE, FRQ, and FRY are adjusted for stock splits and stock dividends. 5 AR and the two forecast revisions, FRQ and FRY, are affected by the information revealed in the earnings release, θ 1. However, AR does not reflect information in the postannoucement period, θ 2. 19
given industry across years). We assess statistical significance of coefficients in the year-by-year estimations using Fama-MacBeth (1973) t-statistics and Z1 and Z2 statistics. 6 3.3 Measurement of Actual Earnings per Share Cornell and Landsman estimate equation (1) measuring earnings forecast errors using a Compustat measure of actual earnings per share, earnings per share before extraordinary items, which we hereafter refer to as the Compustat actual. Because I/B/E/S forecasts and I/B/E/S actual earnings are measured more similarly, i.e., exclude similar items, the I/B/E/S constructed forecast error is expected to be a better measure of earnings surprise. We assess whether this is the case directly by estimating equation (1) using FE_COMP in place of FE, where: FE_COMP it = the forecast error calculated as FE it, except EPS it is earnings before extraordinary items taken from Compustat, divided by shares outstanding taken from I/B/E/S. Even though the forecast errors measured using consistent I/B/E/S actuals likely have more explanatory power, it is still possible that the market derives additional insight from the information conveyed by the Compustat actuals. This may occur, for instance, if the Compustat actuals provide information about GAAP related variables, such as special items, that the market considers relevant, at least in some circumstances, but which are not included in the earnings measure reported by I/B/E/S. To examine this possibility, we estimate equation (2) which adds the term ADJ, the difference between the Compustat and I/B/E/S actuals: AR it = a 0 + a 1 FE it + a 2 FRQ it + a 3 FRY it + a 4 ADJ it + e it (2) 6 The Fama-MacBeth t-statistic = β /( stddev( β ) / ( N 1) ), where N is the number of years. Z1 equals 1 / N N j = 1t j / k j /( k j 2), where t j is the t-statistic for year j, k j is the degrees of freedom, and N is the number of years. Z2, which equals t /( stddev( t) / ( N 1), corrects for potential upward bias in Z1 arising from lack of independence of parameters across industries. See Barth (1994). 20
If the Compustat actuals provide additional information to the market, the ADJ coefficient, a 4, should be significantly different from zero. However, for the reasons indicated, we expect a 4 to be less than a 1. 3.4 Impact of the Fourth Quarter In their original paper, Cornell and Landsman conjecture that estimating the basic model across all four quarters was potentially misleading because the fourth quarter could be different than the other three quarters. They argue, for instance, that analysts might wait until year end to revise year-ahead forecasts and that the market might place more weight on annual forecast errors because annual financial results are audited. Although they produce some preliminary results to support those conjectures, it is based on a sample of only three years and uses Compustat actuals. There are reasons other than those suggested by Cornell and Landsman for believing that the fourth quarter may be unique. First, as one moves from the first to the fourth quarter, the forecasting horizon implicit for FRY becomes shorter. It is reasonable to expect that as the forecasting horizon becomes shorter the perceived precision and hence response coefficient would increase. This horizon is shortest at the time of the announcement of fourth quarter results, which is actually sometime within the first quarter of the next year. Second, the information environment, as well as the nature of quarterly earnings, may differ in the fourth quarter. For example, fourth quarter earnings contain more adjustments and special charges than the prior quarters, in part because of auditing of the annual financial statements. It is possible that these items are leaked to the market in earlier quarters (Mendelhall and Nichols, 1988), which could result in a lower response coefficient for the fourth quarter forecast error relative to the other quarters. Also, more information, in the form of a full set of financial statements, more 21
elaborate management discussion and analysis, and potentially more information gathering by analysts may also occur. As a result, the fourth quarter is more than simply another interim report. It is, in fact, the final quarter in the firm s annual financial statements. Similarly, the quarterly forecast revision, FRQ, is more than simply a forecast for a later quarter in the same fiscal year. It is, in fact, a forecast of the first quarter of the next fiscal year. Third, in addition to these substantive reasons, there are econometric reasons for separating the fourth quarter. For the first three quarters, there is no temporal overlap between FRQ and FRY. However, in the fourth quarter, FRQ is a component of FRY. Hence, the interpretation of the coefficients differs for the fourth quarter. To take account of these possibilities, we consider estimating a version of equation (1) that permits the intercept and FE, FRQ and FRY coefficients to differ for fourth quarter earnings announcements: AR it = a 0 + a 1 FE it + a 2 FRQ it + a 3 FRY it + a 4 D it + a 5 DFE it + a 6 DFRQ it + a 7 DFRY it + e it where D it is an indicator variable that equals one (zero) if the announcement is made in the fourth (interim) quarter, and DFE, DFRQ, and DFRY are interactions between D and the corresponding three variables. 7 In this model, the full impact of the forecast error and the quarter-ahead and year-ahead forecast revisions in the fourth quarter is α 1 + α 5, α 2 + α 3 + α 6 + α 7, and α 3 + α 7, respectively. The reason the full impact of the quarter-ahead revision is more complicated is that an increase in FRQ mechanically increases FRY. 8 To 7 In principle, coefficients for the explanatory variables could change each quarter as the forecast horizon approaches the end of the fiscal year. Findings from untabulated estimations of equation (1) indicate that coefficient point estimates for FE, FRQ, and FRY for the first three quarters are similar in magnitude, while those for the fourth quarter substantially differ from those of the three interim quarters. 8 Figure 3 illustrates the measurement of FE, FRQ, and FRY by quarter and the temporal overlap of FRQ and FRY that occurs in the fourth quarter. 22
account directly for the temporal overlap between FRQ and FRY in the fourth quarter, we estimate the following model: AR it = a 0 + a 1 FE it + a 2 FRQ it + a 3 FRY* it + a 4 D it + a 5 DFE it + a 6 DFRQ it + a 7 DFRY* it + e it (3) where FRY* equals FRY for announcement quarters 1, 2, and 3, and FRY FRQ for announcement quarter 4. In this model, the full impact of the quarter-ahead forecast revision for the fourth quarter is α 2 + α 6, while that for the year-ahead forecast revision for the fourth quarter is α 3 + α 7. 3.5 Extending the Event Window Ideally the forecast error and forecast revisions should be measured over the same time period, so that the market reacts to all three simultaneously. Because of the reporting lag and analyst aggregation issues, this ideal is not met. The aggregation problem arises because analysts do not release their forecasts simultaneously. A measure of a consensus forecast requires individual forecasts to be aggregated over time, and over time subsequent events that are unrelated to the earnings announcement may influence forecast revisions. Reporting lag refers to the time between an analyst forecast and its inclusion in the database. This becomes a problem when a forecast should be included in the current consensus but is not added to the database until after it is calculated. Cornell and Landsman address these problems by extending the measurement window for the forecast revisions. As a result, whereas FE is measured over the same time period as AR, the measurement periods of FRQ and FRY extend several weeks beyond the earnings announcement event window. This feature of the data raises the issue that at the time of the earnings announcement the market must use the information in FE to anticipate its long-run impact, and thereby its effect on analysts forecasts without observing the forecasts. 23
Therefore, AR reflects both the forecast error and the expected forecast revisions. Because the estimating equations include actual forecast revisions, FRQ and FRY, even if the market s expectation of the forecast revisions is unbiased, FRQ and FRY measure these expectations with error, thereby biasing their coefficients towards zero. 9 We address the non-simultaneous variable measurement issue by modifying the basic model in two ways. First, we extend the window over which the abnormal return is measured to the date at which the forecast revisions are observed. This assures us that the residual return will reflect both actual forecast errors and actual forecast revisions. Other things equal, with forecast revisions better aligned in time with the residual return, we would expect their coefficients to increase. However, the problem with this approach is that the window must be extended, on occasion, up to two months after the earnings announcement to make it unlikely that the postannoucement consensus I/B/E/S forecast is sensitive to preannouncement forecasts made by individual analysts. The example illustrated in figure 1 shows the event window runs from March 23 until May 21. The longer event window has the effect of causing both the stock return and the forecast revisions to reflect information that is unrelated to the earnings announcement. While this may increase the slope coefficients on the forecast revisions, the longer event window results in a regression that moves the research question away from discerning the informational effects of the earnings announcement. Therefore, we develop a second approach that utilizes disaggregated data. Rather than using the I/B/E/S consensus forecasts, we employ the individual analysts forecasts to construct a consensus forecast following the earnings announcement. The I/B/E/S detail file contains 9 Because FRQ and FRY are measured beyond the abnormal return event window, this also raises the possibility that forecast revisions are (at least partially) responding to the abnormal return, thereby creating a potential endogeneity issue. Cornell and Landsman (1989, p. 686) recognize this, but argue there is no economic reason to believe that the 24
individual analysts forecasts that can be combined to create custom consensus forecasts at any date and for any time interval. This permits us to shorten the postannoucement event window considerably relative to that associated with the consensus forecasts. The shorter event window mitigates the effects of the aggregation problem by shrinking the forecast periods and aligning them more closely to the earnings announcement. The timeline in figure 2 illustrates how the detail data are used to construct the forecast revisions and to compute the abnormal return over the extended event window. Each constructed forecast revision is computed as the median of analysts forecasts available 19 trading days or less after the earnings announcement less the median of analysts earnings forecasts made from 20 trading days through 1 trading day prior to the earnings announcement. We use median forecasts rather than mean forecasts to lessen the effect of stale forecasts, i.e., those which may be out of date. 10 The return window is computed from the day of the earnings announcement through 19 trading days after the earnings announcement. The forecast error is simply I/B/E/S actual earnings less the median earnings forecast made in the twenty trading days prior to the earnings announcement. As shown in the hypothetical example in figure 2, preannouncement forecasts are gathered over the twenty trading day period ending the trading day before the earnings announcement, March 22. Postannouncement forecasts are gathered over the twenty trading day period beginning on the earnings announcement date, or March 23 to April 20. Abnormal stock returns over the extended window are calculated from the close of the preannouncement date, March 22, until the end of the postannouncement period, April 20, a twenty-day window. information contained in analysts recommendations can be costlessly discerned by observing the change in price when earnings are released. 10 Note that I/B/E/S consensus forecasts likely suffer from effects of stale forecasts, as I/B/E/S includes all available forecasts to construct their consensus measure. For this reason, we use the consensus median rather than mean. 25
As we discuss below, one cost of using detail forecasts is that there is a significant reduction in sample size because we require new forecasts to be issued both before and after the announcement. Thus, the benefit of shortening the event window may be offset to some extent by the loss of precision associated with a smaller estimation sample. 4. Sample Data Firm-quarters included in the sample meet four criteria: 1. Median monthly earnings forecasts, actual earnings, and earnings announcement dates are available from the I/B/E/S summary forecast data file for quarters ending between 1984 and 2003. 2. Quarterly earnings are available on the Compustat file for the same period. Consistent with prior research, earnings is measured as income before extraordinary items and discontinued operations. 3. Daily security price and return data are available from the CRSP file for each earnings announcement event interval (defined above). 4. To mitigate the effects of outliers, for abnormal return, forecast error and forecast revisions, we treat as missing observations that are in the extreme top and bottom one percentile (Kothari and Zimmerman, 1995; Collins, Maydew and Weiss, 1997; Fama and French, 1998; Barth, Beaver, Hand, and Landsman, 1999). Table 1, panels A and B, report descriptive statistics and correlations among the variables used in the study; panel C reports annual descriptive statistics for the forecast errors using I/B/E/S actuals and Compustat actuals, and is discussed in detail in section 5 below. Panel A reveals that mean abnormal return is positive and, consistent with prior research (Abarbanell and Lehavy, 2003), mean forecast errors are negative. In addition, means for both forecast revisions 26