The impact of split adjusting and rounding on analysts' forecast error calculations.

SYNOPSIS: This study finds that analysts' forecast data files, commonly used by accountants and financial analysts to estimate market expectations about earnings announcements, contain inaccurate historical data for companies that split their common stock. These inaccuracies result because stock

split adjustments are made retrospectively and split-adjusted data are rounded. Moreover, because well-performing firms are more likely to execute stock splits, the consequences of the stock split problem are systematic, potentially distorting both time-series and cross-sectional characteristics of forecast errors. The analysis also demonstrates that the problem can influence interpretations of security price reactions to earnings announcements. To illustrate this point, we report evidence suggesting that errors induced by rounding split-adjusted data alter conclusions about how investors interpret earnings that meet, but do not exceed, the consensus forecast.

INTRODUCTION

The use of analysts' forecasts of earnings per share (BPS) is pervasive among practitioners and academics. A typical use presumes that the consensus forecast indicates EPS anticipated by market participants, and that a difference between reported EPS and the consensus forecast reflects the surprise in accounting earnings disclosures. The importance of determining the surprise component in earnings is well known, and thus, the properties of analysts' forecast data are investigated frequently in both the academic and the practice-oriented literature (e.g., Kahn and Rudd 1999; Abarbanell and Lehavy 2000).

This paper describes a potential inaccuracy from using historical analysts' forecast data for firms that split their common stock. More specifically, in commonly used analysts' forecast data files, both forecasted and actual earnings per share are rounded to the nearest cent after making retroactive and cumulative stock split adjustments. (1) The consequence of this practice is that comparisons of actual and forecast earnings for firms that execute stock splits are less precise than comparisons for firms that do not split their stock. (2)

For example, assume that the consensus analysts' forecast EPS is $0.10 and the actual EPS is $0.09, and that the firm subsequently executes a 2-for-1 stock split. Both the forecast and the actual EPS are reported as $0.05 after rounding to the nearest cent. Thus, data from the forecast files erroneously indicate that earnings "meet" the consensus forecast, even though actual earnings reported at the disclosure date differed from the consensus expectation.

Using the quarterly EPS database compiled by First Call during 1993-1999, we find that approximately 22 percent of the observations have retroactive split adjustments that potentially compromise the accuracy of the forecast errors. Consistent with evidence reported in prior studies (Asquith et al. 1989; Lakonishok and Lev 1987), we find that observations affected by the split adjustment differ systematically from observations not affected by splits. Split-adjusted observations show greater accounting performance, pre- and post-announcement period stock price performance, sales growth, and systematic risk, but lower book-to-market and debt-to-asset ratios. Thus, errors induced by rounding split-adjustment data do not occur randomly. While the systematic loss of accuracy in forecast errors can be tolerated in many situations, it can have a material impact in others, especially when the analysis focuses on small forecast errors.

Investigating Stock Splits and Stock Returns

To illustrate how stock splits can affect empirical relations, we examine stock returns associated with "on-target" earnings announcements, defined as earnings announcements where EPS exactly meets the consensus earnings forecast. Meeting the consensus forecast is often presumed to be "good news" (Levitt 1998; Dechow et al. 2000). However, three-day abnormal returns at the earnings announcement are significantly negative for the on-target observations of firms that do not subsequently split their stock. The lack of subsequent splits indicates that reported earnings are actually on-target for this group of observations. In contrast, for observations where actual and forecasted earnings are split-adjusted and rounded (hereafter "split-rounded") so that we cannot be certain that reported earnings actually equaled the forecast, the three-day abnormal returns at the earnings announcement are significantly positive. Thus, conclusions about whether investors interpret "meeting the EPS target" as good news or bad new s depend on the extent that the sample is "contaminated" by split-rounded observations. Our explanation for this result is that many of the apparently on-target observations on a split-rounded basis are in fact not on-target, and misclassification occurs more often for positive rather than negative forecast errors.

To test this explanation, we hand-collect data reported at the time of the earnings announcement for a random sample of observations that have zero forecast errors after being adjusted by split factors greater than 1.0 and rounded to $0.01. Consistent with our explanation, we find that 48 percent of these observations originally reported positive forecast errors compared to 14 percent that reported negative forecast errors. The remaining 38 percent reported zero forecast errors and earned a negative return during the three-day period that encompasses the earnings announcement. Such results suggest that split-rounding creates classification errors that distort associations between announcement period returns and earnings disclosures for apparently on-target observations.

To investigate further, we regress three-day announcement period returns for on-target observations on the split factor while controlling for characteristics of the firms that execute stock splits. We find a statistically significant association between the returns and the split factor. Failure to find that firm characteristics explain the differences in returns across split factors supports our interpretation that rounding error induces a positive relation between security returns and the split factor for apparently on-target observations.

Implications of Split-Rounding Forecast and Actual Data

Three examples illustrate how understanding the consequences of split-rounding is relevant to accounting practice and scholarship:

1. Materiality judgments can depend on associations between unexpected accounting earnings and security returns (Kinney et al. 2000). SEC Staff Accounting Bulletin No. 99, Materiality, recommends "when...management or the independent auditor expects (based, for example, on a pattern of market performance) that a known misstatement may result in a significant positive or negative market reaction, that expected reaction should be taken into account when considering whether a misstatement is material" (SEC 1999). This statement suggests that an assessment of materiality can involve a historical analysis of stock returns and analysts' forecast errors. Our analysis indicates that correlations between earnings and security returns from prior periods can be compromised or distorted by split-rounded data.

2. Comparisons of actual with forecasted errors are relevant to public policy debate. For example, former SEC Chairman Levitt (1998) expresses concern "that the motivation to meet Wall Street earnings expectations may be overriding commonsense business practices [and]...the zeal to satisfy consensus earnings estimates...may be winning the day over faithful representation." Chairman Levitt also asserts "missing an earnings projection by a penny can result in a loss of millions of dollars in market capitalization." Investigating whether firms manage earnings to meet earnings expectations and the consequences of missing earnings expectations typically involves historical analyses of how reported earnings compare with consensus forecasts (Degeorge et al. 1999; Bartov et al. 2002). At least two of this study's findings are directly germane to designing and interpreting these investigations:

* We demonstrate that a spuriously high number of split-rounded observations apparently meets forecasted EPS, which can bias uninformed users of the data toward concluding that firms manage earnings to meet the consensus analysts' forecast.

* The analysis implies that relating forecast errors of a specific magnitude--such as cases where firms meet or miss the consensus forecast by one cent--with security price reactions at the earnings disclosure date is not straightforward for split-rounded observations.

3. Trading strategies developed and employed by practitioners are commonly tested using the history of EPS forecasts (Herzberg 1998) or on how reported earnings compare with analysts' forecasts (Ye 1999; Mozes 2000). For example, a study published recently in the practitioner-oriented Journal of Investing (Brown et al. 1998) uses the First Call forecast database for 1989-1998 to demonstrate that a specific model (designated EPM) predicts earnings better than the consensus forecast. The authors also report results suggesting that abnormal returns can be achieved by buying stocks when the EPM exceeds the consensus estimate. Portfolio managers and financial analysts presumably use analysts' forecast databases to implement, modify, or back-test trading strategies. Thus, the accuracy of the forecast data is relevant for investors who use trading rules developed using historical data that are split-rounded.

DELL COMPUTER AND THE STOCK SPLIT/FORECAST ERROR PROBLEM

Table 1 reports the forecasted and actual EPS for Dell Computer for the six-year period from the third quarter of 1993 through the second quarter of 1999. Forecasted EPS, actual EPS, and the forecast error (which equals actual EPS -- forecasted EPS) are from two sources. Columns (4) through (6) report these items as they appeared in the financial press (the Dow Jones Newswire, Wall Street Journal, or New York Times) at the time of the earnings disclosure. Columns (7) through (9) are from the I/B/E/S Summary file or the First Call file, depending on the source cited in the news article as the basis for comparing actual with expected EPS. To ensure that recent EPS are comparable with EPS reported prior to stock splits, both I/B/E/S and First Call report EPS on a split-adjusted basis. That is, the forecast service provides split-adjusted forecasted and actual EPS, computed as the forecasted and actual earnings reported at the time of the earnings announcement deflated by the split adjustment factor in column (3) . Notice that Dell executed six 2-for-1 common stock splits during the period encompassed by the data.

Consider an investigation that requires distinguishing quarters where Dell exceeds earnings expectations, quarters where Dell reports on-target earnings, and quarters where Dell reports earnings less than the consensus expectation. Column 6 indicates that actual EPS reported at the time of the earnings announcement exceeds the consensus forecast EPS for 18 of the 24 quarters, actual EPS is less than forecast only once, and reported EPS is on-target in five quarters. In contrast, forecast errors computed using split-adjusted EPS (column 9) indicate erroneously that EPS exceeds the consensus forecast seven times, EPS is never less than the forecast, and EPS is on-target for 17 of 24 quarters. In all, 12 of the 24 quarters are classified incorrectly using the analysts' forecast data files.

IMPLICATIONS OF THE LOSS OF ACCURACY FROM SPLIT ROUNDING

In general, the loss of accuracy that results from rounding split-adjusted data varies directly with the size of the cumulative split factor. At least four points need to be considered when using and/or evaluating earnings forecast data affected by stock splits:

1. A spuriously high number of observations apparently meets forecasted EPS. The spurious concentration toward "meeting" the forecast can induce bias toward concluding that firms manage earnings to meet the analysts' forecast.

2. We expect distortions of distributions of analysts' forecast errors that potentially compromise interpretations of statistical tests of how analysts perform relative to each other and relative to mechanical forecasting techniques (Diether et al. 2002; Payne and Thomas 2002).

3. Because stock splits typically indicate superior performance ex post, we suspect that errors in assessing forecast accuracy can be correlated with firm characteristics such as security returns, firm size, and growth.

4. The consequences tend to be more substantial as time passes. That is earlier, rather than more recent, observations are more likely to be affected by the rounding problem because split-adjustments are more likely to apply to earlier observations. For example, only 6.7 percent of 1999 EPS observations in the First Call database are split-adjusted whereas about 35.4 percent of 1993 EPS observations are split-adjusted.

PRELIMINARY EVIDENCE ABOUT ROUNDING ERROR AND STOCK SPLIT FREQUENCIES

We examine the consequences of the stock split problem by considering security price reactions to on-target earnings disclosures, identified as disclosures where actual EPS meet, but do not exceed, the most recent consensus EPS forecast prior to the earnings announcement date (EAD). To be included in our analysis, the consensus forecast must be within 90 calendar days of the EAD for the first three quarters and within 100 calendar days for the fourth quarter, since annual (fourth quarter) earnings announcements tend to occur an average often days later than announcements for other quarters. (3)

We focus on on-target observations for three reasons. First, since unexpected earnings for on-target observations are zero by construction, stock returns associated with earnings announcements can be interpreted without controlling for unexpected earnings. Second, we need not be concerned about the choice of the deflator to normalize forecast errors across firms, or about spurious relations between security returns and unexpected earnings that can result from the deflator used to scale earnings forecast errors (Christie 1987). Third, although there is little a priori reason to expect the security price response to on-target earnings to be positive or negative, both academics and practitioners typically characterize on-target earnings as good news (Levitt 1998; Dechow et al. 2000). Thus, we have a basis for expecting non-negative security price reactions to on-target earnings announcements.

Table 2 reports the distribution of 1993-1999 quarterly earnings announcements recorded by First Call partitioned according to cumulative split factors and according to whether earnings are "on-target." Column 2 in the table indicates that 77.6 percent of the quarters are unaffected by stock splits, but 21.6 percent have stock split factors greater than 1. Column 3 indicates the number of "apparent" on-target EPS, where split-rounded actual EPS equals the split-rounded consensus forecast. Recall that the loss of accuracy from rounding means we cannot be certain about whether observations with split factors greater than 1 are actually on target. Column 4 indicates the relative frequency of (apparently) on-target EPS. Notice that the relative frequency of observations with on-target earnings varies directly with the split factor, increasing from 16.7 percent for observations with split factors of 1 to 31.4 percent for observations with split factors greater than 4. One explanation for this association between t he split factor and apparently on-target observations is that the split-rounded data induces a bias toward a zero forecast error. (4) Evidence for a random sample of apparently on-target observations suggests that this explanation applies. As discussed later, we find that only 38 of 100 apparently on-target split-rounded observations are actually on-target prior to the split-adjustment.

SECURITY PRICE RESPONSES TO ANNOUNCEMENTS OF ON-TARGET EARNINGS

The following analyses focus on observations where actual EPS and consensus forecast of EPS are equal according to the First Call data files. We begin with the 13,916 observations with apparent on-target earnings displayed in Table 2. Of these, 1,562 lack stock returns on the daily CRSP file. Comparing earnings announcement dates (EADs) from First Call with those from Compustat indicates EAD differences of more than one trading day for 5.5 percent of the sample (763 firm-quarters). (5) Omitting these observations yields a final sample of 11,591 quarterly on-target earnings announcements. There are 3,106 observations (26.7 percent) with potentially inaccurate zero forecast errors (observations with split-factors > 1) and 8,485 "clean" on-target observations (split factor [less than or equal to] 1). Notice that split factors < 1 do not distort the on-target classification. That is, if split-rounded actual EPS equals the split-rounded forecasted EPS and the split factor < 1, then both actual and forecast EPS as reported at the announcement date also must be identical.

We follow conventional procedures and estimate the market model for 170 to 21 days prior to the EAD and use the parameter estimates to compute abnormal returns and cumulative abnormal returns (CAR). Because the actual EAD potentially differs by one day from the EAD identified by the forecast data file, we cumulate security returns over three days (days -1, 0, and +1) to ensure that the event window encompasses the earnings disclosure. (6) Z-statistics for mean portfolio excess returns are used to test whether returns during the three-day event period are reliably different from zero (Brown and Warner 1985). (7)

Figure 1 displays cumulative three-day returns for all on-target observations and for subsamples delineated according to the split-factor. Mean return for the observations unaffected by splits (split factor = 1, column 2) is -0.296 percent, which is significantly less than zero (Z-statistic = -2.55). (8) Mean return for the potentially contaminated samples increases monotonically as the split factor increases. In particular, cumulative three-day excess returns increase from -0.015% (Z = -0.14) for observations with 1 < split factor [less than or equal to] 2, to +0.578% (Z = 2.69) for observations 2 < split factor [less than or equal to] 3, and to +1.456% (Z = 3.89) for observations with split factor >4. Mean return for all 3,106 observations that subsequently execute stock splits (split factor >1) is 0.269% (Z = 3.20). Finally, mean excess return during the earnings announcement period for all on-target observations, irrespective of stock splits, is -0.148% (column 8) with the associated Z-statistic = -1.57, which is not significantly different from zero at the 5 percent level (two-tailed).

Interpretations of Security Price Reactions to Earnings Disclosures

Consider the implications of these results for evaluating how financial statement users interpret announcements of on-target earnings. Results for the sample not affected by stock splits indicate that the mean return is significantly negative, yet the combined stock price reaction to the potentially contaminated on-target sample (split factor >1, column 7) is significantly positive at the 0.05 level. Thus, conclusions about how investors respond to on-target earnings announcements depend on whether the sample firms split their stock in subsequent periods.

The analysis in Figure 1 is not intended to be a comprehensive investigation of security price reactions to on-target earnings disclosures, as we cannot determine from the First Call forecast data how many of these observations are actually on-target as-reported. Even so, we attribute the systematic relationship between split factor and security returns at the EAD to split-rounding rather than to substantive differences between firms that subsequently split their stock and firms that do not.

We perform two procedures to investigate this explanation. First, we randomly select 100 split-rounded on-target EPS observations and, using the actual and forecasted EPS reported by the Dow Jones News Service (DJNS), ascertain how many are actually on-target. (9) Table 3, which summarizes this investigation, indicates in row 1 that only 38 of the 100 apparently on-target EPS observations are actually on-target. Of the remaining 62 observations that are not on-target, 48 (77 percent) are positive earnings surprises, ranging from $0.01-0.08, and 14 (23 percent) are negative earnings surprises, ranging from $0.01-0.06. Thus, split-rounding causes misclassification of over 60 percent of the apparently on-target observations, with a bias toward a higher number of misclassifications for positive earnings surprises.

Observe that the mean return of 0.65 percent to the random sample of apparently on-target split-rounded observations reported in row 4 is positive, as is the mean return of 0.27 percent for the population of 3,106 such observations in Figure 1. The mean return to the random sample is not significantly different from zero, however, perhaps due to the lack of power associated with a small sample.

Now consider how returns for the subsamples that comprise the random sample help us to understand whether the positive return for the 3,106 observations is due to split-rounding or to differences in factors that distinguish firms that execute stock splits. If the positive return in Figure 1 for observations with the split factor >1 is due to factors associated with the decision to split, then the return to the 38 truly on-target observations in the random sample reported in Table 3, row 1 should be positive because these firms also split their stock subsequently. We find the opposite, however. The mean (median) return is -1.21 percent (-0.16 percent), similar to the negative returns for the 8,463 observations unaffected by splits in Figure 1, column 2.

Next, notice in Table 3 that the random sample of 100 observations that appear to be on-target based on the analysts' forecast data files includes 48 observations where the firm actually reported a positive earnings surprise, and only 14 observations where the firm actually reported a negative earnings surprise. If the random sample is representative of the population, then almost half of the observations with split factors > 1.0 in Figure 1 actually reported positive earnings surprises that, post-split, are misclassified as on-target. The high frequency of misclassifying positive earnings surprises, which typically experience positive stock market responses, contributes to the positive stock return in Figure 1.

Overall, analysis of the random sample reveals: (1) a positive mean return for 48 misclassified positive forecast errors, (2) a negative mean return for the 38 correctly classified forecast errors (with a modest significance level), and (3) only 14 misclassified negative forecast errors. Taken together, these results imply that the relation between announcement period returns and the split factor in Figure 1 is attributable to split-rounded data, rather than to differences between firms that do and do not subsequently split their stock.

Using Regression Analysis to Investigate Security Return Differences

Regression analysis is another way to investigate whether rounding errors or firm characteristics cause the positive mean return for the split-rounded observations in Figure 1. Prior studies indicate that firms that execute stock splits differ systematically from firms that do not (Grinblatt et al. 1984; Lakonishok and Lev 1987). This study verifies systematic differences between splitting and non-splitting firms that comprise the data used in this study. We find that 10,759 firm quarter observations with split factor> 1 have greater excess returns during the month of the announcement and during the pre- and post-announcement periods. These observations have greater accounting performance-mean income before depreciation, interest, and taxes divided by the mean total assets for the prior four quarters-during the year prior to the earnings announcement, and greater annual sales growth. We find that observations with split factor > 1 also have a lower mean book-to-market ratio, a higher mean long-term debt-to-to tal assets ratio, and a lower mean market-model equity beta. All comparisons are highly significant statistically. (10)

We use the firm characteristics identified above as control variables in a regression of three-day excess returns on the cumulative split factor for the on-target sample. If cumulative returns are biased positively by rounding errors, then we expect a positive coefficient on the split factor variable. On the other hand, if the link between returns and the split factor in Figure 1 arises because firms that split their stocks differ from those that do not, then the coefficient on the split factor will be insignificant because the control variables explain the differences in returns. Table 4 reports results of this investigation. (11) Observe that the estimate on the split-factor is positive and statistically significant in these regressions after controlling for firm characteristics. This result further supports our interpretation that the positive mean return displayed in Figure 1 for firms with split-factor > 1 is attributable to split-rounding data and not to factors that distinguish firms that execute stock splits.

In sum, although we cannot definitively rule out all competing explanations for the results in Figure 1, the evidence in Tables 3 and 4 supports our interpretation:

* price reactions to on-target earnings depend on the extent of subsequent stock splits;

* the link between stock returns and splits is due to misclassifying observations as on-target.

Finally, the evidence apparently contradicts suppositions advanced or investigated by both academics and policymakers that meeting, but not beating, the consensus forecast is interpreted favorably by stock market participants (Levitt 1998; Dechow et al. 2000).

CONCLUDING REMARKS

The purpose of this paper is to direct attention to potential dangers of using analysts' forecast data that are both rounded and retroactively adjusted for stock splits. Distortions caused by rounding split-adjusted observations ("split-rounding") potentially undermine comparisons of reported earnings with analysts' forecasts. We illustrate the consequences of using rounded split-adjusted data by focusing on on-target earnings announcements where split-rounded actual and forecasted earnings are identical. Results of our analysis indicate that forecast errors computed from these data commonly used by academics and practitioners are biased toward reporting more on-target earnings announcements than actually occur. This bias results because split-rounded EPS undermines the precision of earnings forecast errors.

Our investigation of security price reactions to earnings announcements that appear to be on-target further demonstrates the potential relevance of split-rounding for practice and for interpreting empirical analysis. We also find that the consequences of split-rounding are predictably correlated with factors often used to explain cross-sectional differences in forecast errors or in security price reactions to earnings announcements. Thus, the split-rounding problem can induce spurious relations or obscure relations that are otherwise detectable. We emphasize, however, that the extent to which inferences from any investigation are affected by the loss of accuracy from stock splits depends on the composition of the particular sample.

Addressing the split-rounding problem is not straightforward. Because the consequences are systematic, simply omitting split-rounded observations can yield a biased sample. To ensure the integrity of the analysis, the investigator can analyze both the full sample that includes split-rounded observations and the clean sample that excludes split-rounded observations, and demonstrate that results and inferences are robust between the two samples. If results differ, then the investigator needs to obtain unadjusted data from either the data service or by hand-collecting all or part of the split-rounded observations potentially affected by rounding error. Another approach is to obtain individual analysts' forecast data that are not rounded and compute the consensus forecasts more precisely than two decimals.

[FIGURE 1 OMITTED]

TABLE 1

Forecast and Actual EPS Reported by First Call and Those Published by
the Financial Press for Dell Computer Inc.


                                  Published EPS ($)
Fiscal    Fiscal     Adj.                            Forecast
 Year      Qtr.     Factor    Forecast    Actual      Error
 (1)       (2)       (3)        (4)        (5)         (6)

 1993       3         64        0.05       0.26        0.21
 1993       4         64        0.34       0.39        0.05
 1994       1         64        0.42       0.42           0
 1994       2         64        0.72       0.65       -0.07
 1994       3         64        0.85       0.93        0.08
 1994       4         64        0.96       1.36        0.40
 1995       1         64        1.08       1.26        0.18
 1995       2         64        1.22       1.31        0.09
 1995       3         32        0.70       0.75        0.05
 1995       4         32        0.70       0.70           0
 1996       1         32        0.73       0.84        0.11
 1996       2         32        0.86       1.15        0.29
 1996       3         32        1.10       1.56        0.46
 1996       4         16        0.83       1.01        0.18
 1997       1         16        0.93       1.08        0.15
 1997       2          8        0.54       0.59        0.05
 1997       3          8        0.65       0.69        0.04
 1997       4          8        0.76       0.81        0.05
 1998       1          4        0.42       0.44        0.02
 1998       2          4        0.50       0.57        0.07
 1998       3          2        0.27       0.28        0.01
 1998       4          2        0.31       0.31           0
 1999       1          1        0.16       0.16           0
 1999       2          1        0.18       0.18           0

                     EPS Reported by
                 I/B/E/S/First Call ($)
Fiscal                               Forecast
 Year     Forecast         Actual     Error
 (1)        (7)             (8)        (9)

 1993       0.00            0.00          0
 1993       0.01            0.01          0
 1994       0.01            0.01          0
 1994       0.01            0.01          0
 1994       0.01            0.01          0
 1994       0.02            0.02          0
 1995       0.02            0.02          0
 1995       0.02            0.02          0
 1995       0.02            0.02          0
 1995       0.02            0.02          0
 1996       0.02            0.03       0.01
 1996       0.03            0.04       0.01
 1996       0.03            0.05       0.02
 1996       0.05            0.06       0.01
 1997       0.06            0.07       0.01
 1997       0.07            0.07          0
 1997       0.08            0.09       0.01
 1997       0.10            0.10          0
 1998       0.11            0.11          0
 1998       0.13            0.14       0.01
 1998       0.14            0.14          0
 1998       0.16            0.16          0
 1998       0.16            0.16          0
 1999       0.18            0.18          0

"Financial press" refers to Dow Jones News Wire, the Wall Street
Journal, and the New York Times. From 1995Q1 to 1999Q2, both published
and actual forecasts (columns 4 and 5) are expressed with respect to
First Call's expected and actual EPS. From 1993Q3 to 1994Q4, both
published and actual forecasts (columns 4 and 5) are expressed with
respect to I/B/E/S expected and actual earnings. For 1995Q1-1999Q2 EPS
data (columns 7 and 8) are from First Call. From 1993Q3 to 1994Q4, EPS
data are from I/B/E/S.

TABLE 2

Distribution of Cumulative Split Factors and "Apparent" On-Target
Quarterly EPS Observations Recorded in First Call during 1993-1999

                                                    "Apparent"
                                 Firm-    % of All  On-Target
                                Quarters   Sample    Quarters
Cumulative Split Factor           (1)       (2)        (3)

Split factor < 1                   644      0.8          42
 (reverse split)
Split factor = 1                60,711     77.6      10,163
1 < Split factor [less than or
 equal to] 2                    12,161     15.6       2,538
2 < Split factor [less than or
 equal to] 3                     2,123      2.7         435
3 < Split factor [less than or
 equal to] 4                     1,739      2.2         467
4 < Split factor                   861      1.1         271

Total                           78,239      100      13,916

                                "Apparent"
                                On-Target
                                    %
Cumulative Split Factor         (3) + (1)

Split factor < 1                    6.5
 (reverse split)
Split factor = 1                   16.7
1 < Split factor [less than or
 equal to] 2                       20.8
2 < Split factor [less than or
 equal to] 3                       20.5
3 < Split factor [less than or
 equal to] 4                       26.8
4 < Split factor                   31.4

Total                              17.8

"Apparent" on-target quarters are those where the split-adjusted EPS is
identical to consensus forecast EPS at the date closest to, but not
before, the fiscal quarter end and issued within 90 calendar days prior
to the earnings announcement date (EAD).

TABLE 3

Excess Returns around the Earnings Announcement Date for 100 Randomly
Selected Split-Rounded On-Target Observations

                                               True
                                             Forecast
                                             Error ($)     Mean Excess
                              Number of        Mean          Return
                             Observations   [Max./Min.]   (Z-statistic)
                                 (1)            (2)            (3)

On-target earnings                38           0.00          -1.21%
                                             [0.0/0.0]      (-1.70*)
Negative earnings surprises       14          -0.0228        -0.42%
                                           [-0.01/-0.06]    (-0.52)
Positive earnings surprises       48          0.0200          2.48%
                                            [0.08/0.01]      (3.59**)
All sample                       100          0.0064          0.65%
                                           [0.08/-0.06]      (1.15)


                                Meadian
                             Excess Return
                               (Wilcoxon
                             Z-statistic)
                                  (4)

On-target earnings              -0.16 %
                               (-0.23)
Negative earnings surprises     -0.52 %
                               (-0.25)
Positive earnings surprises      2.56%
                                (2.75**)
All sample                       0.63%
                                (0.97)

*, ** Significantly different from zero at the 10 percent and 5 percent
level, respectively, two-tailed.

"Negative (Positive) earnings surprise" observations are where actual
EPS is reportedly below (above) the consensus forecast according to Dow
Jones News Service. Both the actual and the expected EPS are based on
I/B/E/S prior to 1995, and based on First Call after 1995, according to
Dow Jones News Service.

TABLE 4

Regression of Three-Day CAR Centered on EAD on Split Factor and Firm
Characteristics for On-Target Sample

[t-statistics]


                                  Model 1        Model 2
                                 (excludes      (includes
                               split factor)  split factor)

Intercept                         -0.021         -0.021
                                 [-3.91] **     [-3.92] **
Cumulative split factor (a)         NA            0.002
                                                 [4.33] **
12-month excess return              NA             NA
  up to month -1 (b)
1-month excess return in            NA             NA
  month 0 (month of EAD
  excluding the 3-day EAD (b)
12-month excess return              NA             NA
  from month +1 to +12 (b)

Pretax ROA (c)                     0.016          0.015
                                  [1.75]         [1.62]
Sales Growth (d)                  -0.001         -0.001
                                 [-0.26]        [-0.52]
Book-to-Market (e)                 0.020          0.021
                                  [5.33] **      [5.48] **
Debt-to-Total Assets (f)           0.010          0.011
                                  [2.00]         [2.29] **
Equity Beta (g)                    0.001          0.001
                                  [0.65]         [0.28]
Forecast Dispersion (h)            1.193          0.742
                                  [2.46] **      [1.50]
Log of Total Assets (i)            0.001          0.001
                                  [0.87]         [0.36]
[R.sup.2]                          0.006          0.009

                                       Model 3
                               (includes split factor,
                                  and pre- and post-
                                    EAD Returns)

Intercept                              -0.021
                                      [-4.00] **
Cumulative split factor (a)             0.002
                                       [4.94] **
12-month excess return                 -0.272
  up to month -1 (b)                  [-8.77] **
1-month excess return in               -0.046
  month 0 (month of EAD               [-7.00] **
  excluding the 3-day EAD (b)
12-month excess return                  0.054
  from month +1 to +12 (b)             [2.62] **

Pretax ROA (c)                          0.010
                                       [1.11]
Sales Growth (d)                        0.003
                                       [1.39]
Book-to-Market (e)                      0.009
                                       [2.36] **
Debt-to-Total Assets (f)                0.010
                                       [2.15] *
Equity Beta (g)                         0.001
                                       [0.85]
Forecast Dispersion (h)                 0.43
                                       [0.87]
Log of Total Assets (i)                 0.001
                                       [1.25]
[R.sup.2]                               0.027

*, ** Significantly different from zero at the 10 percent and 5 percent
level, respectively, two-tailed.

(a) Cumulative split-factor is the factor to adjust for stock splits
retroactively, taken from the forecast data file.

(b) Excess returns are computed net of the CRSP value-weighted index.

(c) Pretax ROA = sum of pretax operating income (earnings before
depreciation interest, and taxes) for the prior four quarters divided by
the mean total assets of the prior four quarters.

(d) Sales growth = growth rate of sales over the same quarter last year.

(e) Book-to-market = book value of equity divided by market value of
equity at the quarter-end.

(f) Debt-to-total assets = long-term debt divided by total assets as the
quarter-end.

(g) Equity beta = market model beta used to compute daily excess returns

(h) Forecast dispersion = standard deviation of forecasts scaled by
split-adjusted stock price per share.

(i) Logsize = log of total assets at the quarter-end.

For ROA, sales growth, book-to-market, debt-to-assets, and equity beta,
the top and the bottom 1 percent of the distribution are excluded.

Submitted: February 2002

Accepted: August 2002

(1.) To our knowledge, rounding occurs in both the I/B/E/S consensus database and the First Call databases. We understand that, in principle, a researcher can construct "unrounded" consensus forecasts using the I/B/E/S Detail History file that contains raw individual analysts' forecasts, but that this cannot be done with the First Call data. We do not have access to the Zack's database.

(2.) The problem is also noted recently in Diether et al. (2002) and in Payne and Thomas (2002).

(3.) Results are comparable when forecast errors are computed as the difference between actual earnings and the most recent consensus forecast prior to the end of the quarter. Philbrick and Ricks (1991) find that using the actual EPS from the forecast data, as opposed to using actual EPS reported (for example) in Compustat, minimizes the error in FE attributable to differences in how EPS is defined. Thus, to ensure consistency in computing FE, actual EPS also is from First Call.

(4.) Since earlier observations are more likely to be affected by larger split adjustment factors, an alternative explanation for the Table 2 pattern of increasing on-target frequency for larger split factors is that, for whatever reason, on-target frequency decreases over time. An examination of our data (for split factor = 1 observations) and Bartov et al. (2002) indicate that both on-target and above-target frequencies increase, not decrease, over time, however.

(5.) For First Call EAD, we use the "System_date." Verifying First Call's EAD with Compustat reduces the possibility of measuring returns over the wrong event window.

(6.) We obtain comparable results using one- and two-day returns.

(7.) The Z-statistic is computed for day t as [R.sub.t]/[sigma], where [R.sub.t] is the average return to the portfolio of firms on day t, and [sigma] is the sample standard error of the mean portfolio returns during the estimation period (t = -170 to -21). The Z-statistic follows the standard normal distribution.

(8.) A firm-quarter observation with split factor of 1.0 can have rounding error if the firm subsequently has a reverse stock split followed by a regular stock split. No such instances occur in our sample, however.

(9.) For each earnings announcement, we specifically search for the "Earnings Surprise Summary" reported by DJNS, where both the actual and the consensus EPS and the source of the forecast (I/B/E/S or First Call) are reported.

(10.) For comparisons of accounting returns, sales growth, book-to-market, debt-to-assets, and equity beta, we omit the top and bottom 1 percent of the distributions in order to mitigate the effects of outliers.

(11.) Results of Model 3 including ex post stock returns need to be interpreted with caution due to the use of explanatory variables not available at the earnings announcement date.

REFERENCES

Abarbanell, J., and R. Lehavy. 2000. Differences in commercial database reported earnings: Implications for inferences concerning analyst forecast rationality, the association between prices and earnings, and firm reporting discretion. Working paper, The University of North Carolina at Chapel Hill and University of California, Berkeley.

Asquith, P., P. Healy, and K. Palepu. 1989. Earnings and stock splits. The Accounting Review 64: 387-403.

Bartov, E., D. Givoly, and C. Hayn. 2002. The rewards to meeting or beating earnings expectations. Journal of Accounting and Economics 33 (2): 173-204.

Brown, L., J. Guo, and M. Herzberg. 1999. Enhancing earnings predictability using individual analyst forecasts. Journal of Investing 8: 15-24.

Brown, S., and J. Warner 1985. Using daily returns: The case of event studies. Journal of Financial Economics 14: 3-31.

Christie, A. 1987. On cross-sectional analysis in accounting research. Journal of Accounting and Economics 9: 231-258.

Dechow, P., S. Richardson, and I. Tuna. 2000. Are benchmark beaters doing anything wrong? Working Paper, University of Michigan.

Degeorge, F., J. Patel, and R. Zeckhauser. 1999. Earnings management to exceed thresholds. Journal of Business 71: 1-33.

Diether, K., C. Malay, and A. Scherbina. 2002. Differences of opinion and the cross-section of stock returns. Journal of Finance 57: 2113-2141.

Grinblatt, M., R. Masulis, and S. Titman. 1984. The valuation effects of stock splits and stock dividends. Journal of Financial Economics 13: 461-490.

Herzberg, M. 1998. Implementing EVA/EBD analysis in stock selection. Journal of Investing 7:45-53.

Kahn, R., and A. Rudd. 1999. Modeling analyst forecast behavior. Journal of Investing 8: 7-14.

Kinney, W., D. Burgstahler, and R. Martin. 2000. The materiality of earnings surprises. Working paper, The University of Texas at Austin.

Lakonishok, J., and B. Lev. 1987. Stock splits and stock dividends: Why, who, and when. Journal of Finance 42: 913-932.

Levitt, A. 1998. The numbers game. Speech delivered at the NYU Center for Law and Business, New York, NY, September 28.

Mozes, H. 2000. The role of value in strategies based on anticipated earnings surprise. Journal of Portfolio Management 26: 54-62.

Payne, J., and W. Thomas. 2002. The implications of using stock split adjusted I/B/E/S data in empirical research. Working paper, University of Okiahoma.

Philbrick, D., and W. Ricks. 1991. Using Value Line and I/B/E/S analyst forecasts in accounting research. Journal of Accounting Research 29: 397-417.

U.S. Securities and Exchange Commission (SEC). 1999. Materiality. SEC Staff Accounting Bulletin No. 99. August 12. Washington, D.C.: Government Printing Office.

Ye, J. 1999. Excess returns, stock splits, and analyst earnings forecasts. Journal of Portfolio Management 25: 70-76.

We thank Larry Brown, chris Jones, Krishna Kumar, Reuven Lehavy, K. Sivaramakrishnan, Karen Taranto, Kumar Visvanathan, Joanna Wu, two anonymous reviewers, and workshop participants at Texas A&M University, University of Cincinnati, University of Nebraska--Lincoln, University of Oregon, and Virginia Commonwealth University for helpful comments. We also thank First Call[TM], and Stanley Levine in particular, for providing the analysts' forecast data. We are particularly grateful to Robert Lipe for his input.