The Implications of Using Stock-Split Adjusted I/B/E/S Data in Empirical Research

Transcription

1 The Implications of Using Stock-Split Adjusted I/B/E/S Data in Empirical Research Jeff L. Payne Gatton College of Business and Economics University of Kentucky Lexington, KY 40507, USA and Wayne B. Thomas Michael F. Price College of Business University of Oklahoma Norman, OK 73019, USA Published in The Accounting Review October, 2003: We are especially thankful to Bob Lipe and two anonymous reviewers whose comments were instrumental in developing the paper. We are also thankful for comments received from Jeff Abarbanell, Fran Ayres, Marlys Lipe, and G. Lee Willinger. The authors gratefully acknowledge the contribution of I/B/E/S Inc. for providing earnings per share forecast data, available through the Institutional Brokers Estimate System. I/B/E/S provides this data as part of a broad academic program to encourage earnings expectations research. Electronic copy of this paper is available at:

2 The Implications of Using Stock-Split Adjusted I/B/E/S Data in Empirical Research ABSTRACT: The purpose of this study is to highlight issues of interest to researchers employing the I/B/E/S earnings and forecast data. I/B/E/S has traditionally provided per share data on a split-adjusted basis, rounded to the nearest penny. In doing so, per share amounts are comparable over time. However, because not all prior forecasts and earnings per share amounts divide precisely to a penny, adjusting for stock splits and rounding to the nearest penny can cause a loss of information. Researchers are prohibited in many cases from determining the amounts actually reported in prior years, leading to misclassified observations. We obtain actual (unadjusted) earnings and forecast data from I/B/E/S and compare results to those generated using the adjusted I/B/E/S data. We replicate prior studies and find that conclusions are affected when using the actual I/B/E/S data. Key Words: Analysts Forecast Errors, Earnings, Earnings Changes, Earnings Announcement Returns, Dispersion, Misclassification. Data availability: All data are from public sources. Electronic copy of this paper is available at:

3 I. INTRODUCTION The purpose of this research is to investigate how using stock-split adjusted analysts forecasts data can influence inferences drawn in extant research. A large body of research employs analyst forecast data to investigate various financial accounting issues. Financial analysts forecast data provide observable measures of expected earnings by sophisticated users of financial accounting information (Brown et al. 1987, Brown 1993). As such, researchers use them to examine issues related to the valuation of a firm s securities, management s incentives to manipulate earnings, and the overall quality of a firm s information environment. In order to specify appropriate tests and draw appropriate inferences researchers must understand the conventions used in creating the forecast data. The procedures employed by data providers may impact research results (Abarbanell and Lehavy 2001). One of the primary providers of analyst forecast data is Institutional Brokers Estimate System (I/B/E/S). 1 I/B/E/S issues annual (from 1976) and quarterly (from 1984) forecasts. Traditionally, I/B/E/S has provided forecast data on an adjusted basis, rounded to two decimal places. 2 The adjustment controls for the effects of stock splits on per share amounts over time. Rounding to two decimals is presumably done so that all amounts are reported to the nearest penny. While it is important to be aware of both issues (split adjustment and subsequent rounding), the latter issue is potentially more troublesome to researchers. When reported amounts are divided by the split factor, forecast errors that were originally large can appear quite small. Further, because of the rounding, the researcher can recreate the actual reported amounts only if the adjusted amounts have no more than two decimal places. Otherwise, information is lost. For actual amounts rounded to zero by the stock-split adjustment, multiplying by the split

4 factor or scaling by some per share amount will not correct the rounding effect. The problem is best illustrated with the following examples. Suppose in year 1 that Company A and Company B have the follow earnings per share and forecast per share amounts: Example 1: After a 4-for-1 Stock Split in Year 2* Year 1 Year 1 (adjusted) Company A Earnings Forecast Forecast Error Company B Earnings Forecast Forecast Error * Amounts are rounded to two decimals, as done by I/B/E/S. In year 2, both companies have a 4-for-1 stock split. On an adjusted basis, I/B/E/S reports actual earnings per share in year 1 for Company A and Company B as.25. These numbers are the actual earnings per share divided by four and then rounded to the nearest penny. The adjusted forecast amount for each firm would be reported by I/B/E/S as.25. The researcher would then conclude that the forecast errors for both firms are zero. 3,4 The adjustment issue becomes more pronounced as the split factor increases. Suppose Company A and Company B have the following earnings per share and forecast per share amounts in year 1: Example 2: After a 64-for-1 Stock Split by Year N* Year 1 Year 1 (adjusted) Company A Earnings Forecast Forecast Error Company B Earnings Forecast Forecast Error

5 * Amounts are rounded to two decimals, as done by I/B/E/S. By year N, each firm has an adjustment factor of 64 for year 1 (e.g., six 2-for-1 stock splits since year 1). For Company A and Company B, I/B/E/S would report zero forecast error in year 1. In reality, Company A missed its forecast by almost 50 percent and Company B exceeded its forecast by nearly 50 percent. Using the adjusted I/B/E/S data, the researcher cannot distinguish these observations. When using the adjusted, rounded data provided by I/B/E/S, a researcher could misclassify prior years analysts forecast errors. Because stock splits are not random events, the misclassification can bias the empirical results. The purpose of our study is to first provide general evidence about the potential influence of rounding to two decimal places. The results indicate that stock split factors are of sufficient magnitude to affect reported adjusted amounts and the differences between adjusted and actual forecast data are systematically related to firm characteristics such as size and price-to-book ratios. Next, we use I/B/E/S actual data to replicate prior research that used the adjusted data. We find that many prior conclusions are affected by the rounding procedure. Furthermore, other studies with the potential to contribute to the accounting and finance literature may have been abandoned because the adjusted I/B/E/S data led to insignificant results when, in fact, a significant relation was present. The paper proceeds as follows. In section II we discuss the research design and report the results of our replications using both the adjusted and actual I/B/E/S forecast data. We conclude the paper in section III. II. EMPIRICAL ANALYSIS The potential for the rounding procedure to cause problems depends on the study s sample characteristics and research design. As shown in the introduction, the rounding 3

6 procedure employed by I/B/E/S is potentially problematic only for firms that report stock splits or stock dividends (hereafter both denoted as stock splits), otherwise there would be no subsequent rounding of previous amounts. In addition, the problem is exacerbated at zero forecast error, since the researcher cannot unadjust the data. 5 Therefore, conclusions are more likely to be affected in research settings that focus on zero forecast error, that involve firms that are more likely to have splits (e.g., larger firms, growth firms, well-performing firms, etc.), and/or that fail to consider the adjustment made by I/B/E/S. In the following sections, we provide general and specific evidence concerning the effects of the rounding procedure by replicating prior published research. For each replicated study, we follow as closely as possible the research design, sampling procedure, variable measurement, statistical tests, and controls for outliers. We determine if inferences change according to whether the adjusted or actual data are employed. We obtain analysts forecasts data from I/B/E/S for this study. 6 One dataset is stock-split adjusted and rounded to two decimals. These are the data provided to academics under I/B/E/S standard subscription agreement. We also obtain data from I/B/E/S for the same time period that is not stock-split adjusted via a special request. These data represent the analysts forecasts, earnings information, stock prices, etc., as they were originally entered into the I/B/E/S database. These amounts are not adjusted for subsequent stock splits. 7 The full sample includes 173,286 firm/quarter observations for the years General Evidence We first provide information on the distribution of stock split factors to demonstrate how sample observations are affected by stock splits. As shown in Table 1, the mean factor declines 4

7 from a little more than three in 1984 to just above one in In 1984 the top quartile of firms had split factors in excess of 3, with the top decile in excess of These values gradually decline, as expected, but still retain values greater than one until 1998 for the upper quartile and stay greater than one for the top decile of firms. Notice that the maximum split factor gets as large as 288 in This does not by itself indicate that the analysts forecast errors for stock split firms contain rounding errors. The influence of the stock split factor on rounding is determined by several factors including the level of reported earnings, analysts forecasts, and the related analysts forecast error. However, the results in Table 1 indicate that stock split factors are of sufficient magnitude to affect reported adjusted amounts as shown in Examples 1 and 2. [Place Table 1 about here] Next, we provide general evidence that use of the adjusted data versus actual I/B/E/S data can affect inferences. We focus this analysis on settings where reported analysts forecast errors equal zero as this represents the adjusted data that is most likely affected by the stock split adjustment. Forecast error is computed as earnings per share for the quarter as reported by I/B/E/S minus the mean forecast of earnings per share as of the final month of the quarter. adjusted data: Based on our previous discussion, we expect the following differences in the actual and 1) the adjusted data will overstate the percentage of observations with amounts equal to zero, 2) the overstatement of observations with amounts equal to zero will be greater in earlier years (because split factors are applied retroactively), 3) the distributions of the adjusted data will be tighter, and 4) firms that are more likely to have stock splits (e.g., larger firms and higher growth firms) will have more observations misclassified at zero. 5

8 As anticipated, the percentage of observations equal to zero is greater for the adjusted data than the actual data (see Panel A of Table 2). Examination of the first three columns shows that the proportion of zero adjusted forecast errors (AdjFE) is significantly greater than the proportion that actually occurred (ActFE), and these differences are greater in earlier years because of the retroactive application of the split adjustment. The statistics reported in the Ratio column, calculated as (AdjFE/ActFE)-1, indicate the degree of misclassification. This ratio begins at percent in 1984 and gradually declines to 7.40 percent in 1999, with an overall misclassification ratio of percent. This pattern is consistent with our expectation that as the stock split factor declines as time passes, the differential classification of AdjFE and ActFE will decline. [Place Table 2 about here] Researchers have also been interested in the trend of zero forecast error over time (e.g., Brown 1997, 2001; Matsumoto 2002). We calculate the trend as the estimated coefficient in a regression that uses the percentage of zero forecast errors each year (multiplied by 100) as the dependent variable and the year as the independent variable. The trend coefficient presented in Panel A shows that while there is an increasing trend in the percentage of zero observations for AdjFE (.505), the trend is statistically greater for ActFE (.821). In addition, while the percentage of zero AdjFE levels off around 1992, ActFE shows a continuing increase in the proportion of zero observations from 1992 to Untabulated results for the period show the trend coefficient is not significant for the adjusted data but it is significant for the actual data. Using the adjusted I/B/E/S data, one may conclude that the increasing pattern of zero forecast errors has leveled off in recent years when, in fact, it has continued to increase. 6

9 Panel A of Table 2 also reports the interquartile ranges for analysts forecast errors by year. The interquartile ranges indicate that not only are the number of observations at zero affected by stock split adjustment, so are the overall distributions. The distribution of forecast errors is noticeably tighter for the adjusted data. The differences in the distributions are particularly noticeable in earlier years In Panel B of Table 2 we demonstrate how the adjustment made by I/B/E/S could impact conclusions about the relation between certain firm-specific characteristics and earnings thresholds. Prior research has shown that larger firms and higher price-to-book firms are more likely to meet analysts forecasts (e.g., Elgers and Murray 1992; Skinner and Sloan 2002). Using data from the annual Compustat files, we sort the sample observations into groups based on their beginning of year market value of equity or price-to-book ratio. Those in the bottom third of the distribution are classified as small firms or low price-to-book firms. Those in the top third are classified as large firms or high price-to-book firms. We compare the distributions of adjusted versus actual data amounts for these groups. 11 The percentage of zero AdjFE and ActFE for small and large firms is shown in the first four columns in Panel B of Table 2. For both small and large firms, the percentage of observations with zero AdjFE is significantly higher that what actually occurred. Comparing AdjFE for small and large firms indicates that large firms were more likely to exactly meet analysts expectation. This pattern holds for all years. However, ActFE indicates that larger firms began to consistently report zero forecast error more often than small firms only in recent years ( ). For all years combined, larger firms are more likely to hit analysts forecasts, but the actual differential (10.47 percent vs percent) is much smaller than the results based on AdjFE (13.32 percent vs percent). The trend coefficients for AdjFE indicate no difference 7

10 for large and small firms (.457 versus.504), while the coefficients for ActFE indicate the trend for large firms is significantly greater than the trend for small firms (.945 versus.766; p <.05). The distributions for low and high price-to-book firms are shown in the next four columns of Panel B, Table 2. The AdjFE columns show strong evidence that high price-to-book firms are more likely to report earnings equal to analysts forecasts. The magnitude of the difference is reduced when observing the ActFE data, but it is still significant. Interestingly, the percentage of zero observations for AdjFE high price-to-book firms is stable over time and the trend coefficient (.039) is not significantly different from zero. However, when looking at the ActFE data, the proportion of high price-to-book firms reporting zero forecast error is increasing over time (.707; p <.01). The distribution and trend of AdjFE are similar to those of ActFE for low price-to-book firms, as these firms tend to have fewer stock splits. While the conclusion that high price-to-book firms are more likely to report zero forecast errors has not changed, the results demonstrate the potential effect that failure to consider the I/B/E/S adjustment procedure can have on conclusions, especially in the earlier years of the data. Overall, the results reported in Panel B demonstrate differences between adjusted and actual forecast errors that are systematically related to firm characteristics. Larger firms and higher price-to-book firms tend to represent successful, growing firms that are more likely to have stock splits. Thus, these firms are more likely to be affected by rounding split adjusted data. Researchers should be aware that cross-sectional comparisons of earnings thresholds may be affected by the correlation of numerous firm characteristics and the occurrence of stock splits. 8

11 Replication of Prior Research Distribution of Forecast Errors, Earnings, and Earnings Changes DeGeorge et al. (1999) investigate managements incentives regarding reported performance with respect to three earnings thresholds, analysts forecasts, prior year s reported earnings, and reporting a profit in the current period. They suggest that exactly meeting a threshold (i.e., analysts forecast error equal to zero) is likely the point at which managers modify reported earnings in a way that allows researchers to detect those actions. DeGeorge et al. (1999) do not attempt to identify the process that management uses to modify reported earnings. Instead, they look for abnormal amounts of reported performance at or around these thresholds as indicators that earnings targets are considered by management. Their results suggest that each threshold is an important determinant of reported earnings. DeGeorge et al. (1999) test the proportion of observations at each of the thresholds using adjusted I/B/E/S data. Because of the rounding procedure used by I/B/E/S, some of the observations around these thresholds are likely misclassified. This is especially true for observations at each earnings threshold (i.e., those observations with a zero amount). Additionally, DeGeorge et al. (1999) do not unadjust the I/B/E/S data by multiplying by the split factor or scaling by some per share amount (e.g., price, earnings per share), which increases the buildup of observations at and around zero (see footnote 4). To investigate the potential impact of the I/B/E/S rounding procedure on the results reported by DeGeorge et al. (1999), we construct a dataset for the same time period , include77 only companies with years ending in the months of March, June, September, and December, and remove all observations in the upper and lower decile of stock price to be consistent with their research design. 12 Based on our results in Table 2, we expect that the 9

12 distributions reported by DeGeorge et al. (1999) are likely affected by the rounding adjustment. Specifically, we expect their analyses to indicate more observations at zero for each threshold and a tighter distribution of observations around zero than actually occurred. The first two columns of results in Table 3 report the number of observations for $.01 forecast error intervals, ranging from $.15 to -$.15. The distribution for the adjusted forecast error is shown in the first column. As expected, the interval with the greatest number of observations is at zero forecast error. The distribution for the actual forecast errors in the second column is noticeably different. The adjusted data have 11,274 observations with zero forecast errors compared to 7,012 observations for the actual data; this difference is statistically significant at p <.05 using a chi-square test. 13 Compared to the distribution for adjusted forecast errors, fewer firms actually hit their forecasted target. The distribution is also much less peaked around zero forecast error. DeGeorge et al. (1999) derive a test statistic (see their appendix, pages 30-32) to determine if the proportion of observations at zero forecast error is statistically significantly different from the proportion of observations at $-.01. They report a test statistic of 6.61 (see their page 20) and suggest that a test statistic greater than two indicates significance. Using the adjusted (actual) data we calculate a test statistic of 7.40 (4.50). [Place Table 3 about here] Of the 11,274 cases where the firm appears to exactly meet its forecast based on the adjusted data, 4,262 (38 percent) are misclassified. The third column of Table 3 shows the forecast errors based on the actual data for the 4,262 misclassifications. The actual forecast error was less (greater) than zero 43.8 percent (56.2 percent) of the time. The actual forecast error ranged from a negative nine cents to positive twelve cents. Recall, based on the adjusted data, that all of these observations had zero forecast error. 10

13 The next two columns in Table 3 report the distributions for adjusted and actual changes in earnings per share. The distribution for actual changes in earnings is flatter at zero and across most intervals. Using the adjusted data, the number of observations that are reported as having no change in earnings is greater than the number of observations that actually had no change in earnings (2,768 vs. 1,790, p <.01). The DeGeorge et al. (1999) test statistic produces the following results for change in earnings per share: DeGeorge et al. (1999) = 5.63, our results using the adjusted data = 6.53, and the actual data = The adjusted data overstate the number of firms that just meet or slightly beat prior year s earnings. Overall, 978 of the 2,768 observations that have zero earnings change in the adjusted data are misclassified. Of the 978 misclassifications, 67.3 percent actually reported positive earnings changes and 32.7 percent had negative earnings changes. Examination of the final three columns in Table 3 shows the distributions for adjusted and actual earnings per share as reported by I/B/E/S and the distribution of actual earnings for those observations with an adjusted earnings level of zero. The two distributions are relatively similar for the intervals below zero. For the intervals at and above zero, the two distributions differ substantially. The adjusted data overstate the number of firms that just meet or slightly beat prior year s earnings (518 vs. 431, p <.01). The DeGeorge et al. (1999) test statistic produces the following results for earnings per share: DeGeorge et al. (1999) = 3.84, our results using the adjusted data = 3.06, and the actual data = Overall, the analysis reported by DeGeorge et al. (1999) is affected by the rounding procedure, especially for the zero value of each threshold. However, it is important to note that some of the differences are small, particularly for the earnings level threshold. Also applying the test statistic derived by DeGeorge et al. (1999) to our actual data produces conclusions that are 11

14 consistent with theirs. However, significant differences clearly exist between the classifications derived from the adjusted versus the actual data. For forecast errors, the adjusted data produces a significantly different classification of observations for 30 of 31 one cent intervals in Table 3. For earnings changes and earnings levels, the classification is significantly different from -$.09 to +$.15 and $.00 to $.15, respectively. We recommend using actual data to avoid drawing potentially erroneous conclusions when investigating earnings thresholds, especially in settings where firm characteristics such as size are important (see Panel B of Table 2). Inferior Returns to Growth Stocks Skinner and Sloan (2002) investigate the reason for the lower future returns of growth stocks relative to value stocks. They find that the market s negative reaction to negative earnings surprises is more pronounced for growth stocks. However, the market s positive reaction to positive earnings surprises does not differ between growth and value stocks. Skinner and Sloan (2002) argue that growth stocks receive a greater reaction to negative forecast errors because investors were too optimistic in the current period and this over optimism is corrected in subsequent periods as earnings forecast are not met. Their results are consistent with the inferior subsequent returns to growth stocks being attributable to expectational errors about future earnings performance, rather than being attributable to risk differences or methodological problems associated with calculating long-term abnormal returns. After controlling for the asymmetric price reaction to negative earnings surprises of growth firms, Skinner and Sloan (2002) find no difference in the subsequent returns of growth and value stocks. To test their primary hypothesis, Skinner and Sloan (2002) estimate the following model: R itτ = α + β 1 Growth it + β 2 Good itτ + β 3 Bad itτ + β 4 (Good itτ Growth itτ ) + β 5 (Bad itτ Growth itτ ) + ε itτ (1) 12

15 where, i = firm index; t = calendar quarter in which growth portfolio assignments are made; τ = twenty subsequent quarters over which returns and earnings surprises are tracked; R itτ = size-adjusted return for firm i in quarter t+τ, beginning two days after the announcement of earnings in quarter t+τ-1 and ending the day after the announcement of earnings in quarter t+τ; Growth it = growth quintile to which firm i is assigned in quarter t (zero = low growth quintile,, four = high growth quintile); Good itτ = indicator variable taking the value of one if the firm-quarter observation reports a positive earnings surprise in quarter t+τ and zero otherwise; and Bad itτ = indicator variable taking the value of one if the firm-quarter observation reports a negative earnings surprise in quarter t+τ and zero otherwise. Skinner and Sloan (2002) measure the earnings surprise as I/B/E/S earnings minus the median forecast of earnings in the final month of the fiscal quarter. The coefficient on Growth (β 1 ) estimates the relation between growth in quarter t and subsequent abnormal returns in quarters t+τ. The coefficients on the interaction terms (β 4 and β 5 ) estimate the differential price reaction to positive and negative earnings surprises, respectively, for growth and value stocks. Because of the rounding procedure used for the I/B/E/S data, the Good and Bad variables will be measured with error. Many of the observations that report zero forecast error actually had positive or negative forecast errors. A greater misclassification of positive forecast errors (a likely outcome given the analysis in Table 3) will lead to the average return at zero forecast error being overstated (i.e., too positive) when using the adjusted data. The overstated return at zero forecast error will cause the difference between firms classified as having zero forecast error and those that classified as missing (beating) the forecast to be overstated (understated). Furthermore, the potential for misclassification is greater in the growth portfolio because these firms are more likely to have stock splits (e.g., see Panel B of Table 2). Thus, the adjusted data have the potential to create an appearance of an asymmetric price response for just missing versus just beating the forecast, and the magnitude of the asymmetry could vary with growth. The greater 13

16 the frequency of stock splits, the disproportionately greater the misclassification of positive forecast errors to zero and the more overstated the return to zero forecast error. To determine the impact that the rounding procedure has on the results, we estimate equation (1) using both the adjusted and actual I/B/E/S data. We employ similar sampling procedures, variable measurements, and statistical tests as those in Skinner and Sloan (2002). The results are reported in Table 4. For convenience, we also tabulate the results reported by Skinner and Sloan (2002) (Panel A) along with ours (Panel B). [Place Table 4 about here] We first estimate the relation between growth quintiles and stock returns over the subsequent 20 quarters, without controlling for the asymmetric response to positive and negative earnings surprises. Similar to Skinner and Sloan (2002), we find a significantly negative relation between growth and subsequent returns. Next, we control for subsequent earnings news (Good and Bad) and the asymmetric price response for growth firms (Good x Growth and Bad x Growth), where Good vs. Bad are based on the adjusted I/B/E/S data. As in Skinner and Sloan (2002), the coefficient on Growth is no longer significant, while the coefficient on the interaction of Growth and Bad is significant. The negative price reaction to a negative earnings surprise appears more severe for growth stocks. After controlling for this, there is no evidence of inferior returns to growth stocks. The coefficient on the interaction of Growth and Good is not significant. These results lead Skinner and Sloan (2002) to conclude that the inferior subsequent returns of growth stocks are the result of expectational errors about future performance. In the final row of Table 4, we estimate equation (1) using the actual I/B/E/S data. There are two notable differences. First, the coefficient on growth remains significantly negative after controlling for any asymmetric price response. Second, the coefficient on the interaction of 14

17 Growth and Bad becomes less significant, going from t = (p =.0001) for the adjusted data to t = (p =.0818) for the actual data. The decrease in significance indicates that the differential market penalty associated with missing a forecast is not as severe for high and low growth stocks as suggested by the adjusted data. The significance of the interaction term Growth and Good increases from t = 1.10 (p =.2713) for the adjusted data to t = 2.54 (p =.0111) for the actual data. The reaction to good news is greater for high growth firms. Using the actual I/B/E/S data, we conclude that the inferior returns to growth stocks are not related to expectational errors about future performance. 14 Relation Between Analyst Uncertainty and Future Returns Prior research has shown a negative relation between the standard deviation of earnings forecasts and future returns (e.g., Ackert and Athanassakos 1997, Ang and Ciccone 2001). Ackert and Athanassakos (1997) argue that the standard deviation of earnings forecasts is a proxy for ex ante uncertainty, and analysts tend to be too optimistic when uncertainty is higher. If the market follows analysts optimism, then prices will also be too high. Ackert and Athanassakos (1997) therefore hypothesize and find that uncertainty (i.e., standard deviation of earnings forecasts) is negatively related to future returns. They argue that this finding would enable them to earn abnormal returns because this information about forecast dispersion is available to investors at the beginning of the holding period (p. 270). However, the standard deviation of earnings forecast is not totally an ex ante measure as reported on the I/B/E/S database; it is also an ex post measure because of future stock splits. I/B/E/S reports the standard deviation of earnings forecasts based on adjusted amounts. Therefore, a future stock split will reduce the level and standard deviation of forecasted earnings 15

18 per share in the current period (i.e., a future 2-for-1 split will reduce the current standard deviation by one-half, a future 3-for-1 split will reduce the current standard deviation by onethird, and so on). 15 Firms with future stock splits are more likely the ones that have higher future returns (Lakonishok and Lev 1987). In addition, firms with future stock splits are more likely to report lower standard deviation of earnings forecasts as reported by I/B/E/S. Putting these together, the prediction that firms with lower (higher) standard deviation of earnings forecasts will have higher (lower) future returns could be mechanically driven because the standard deviation is not purely an ex ante measure. Following the research design in Ackert and Athanassakos (1997), we use data from 1980 to 1991 to classify firms into quartiles each year based on their standard deviation of year t earnings forecasts made in June of year t-1, and then we measure raw returns and beta excess returns over the subsequent 20-month period beginning in July of year t-1 and ending February of year t+1. The results are reported in Table 5. The results in Panel B for the adjusted I/B/E/S data are similar to those in Panel A by Ackert and Athanassakos (1997). Firms with higher standard deviation of earnings forecasts have lower future returns. The difference between the low and high quartile s average monthly raw (beta excess) return is.84 percent (.78 percent), or an annualized return of percent (9.77 percent). These differences are statistically and economically significant. 16 These findings lead Ackert and Athanassakos (1997) to conclude that they can construct profitable portfolio strategies (p. 272) by creating a hedge portfolio that is long (short) in firms that have low (high) standard deviation of earnings forecasts as reported by I/B/E/S in June of each year. [Place Table 5 about here] 16

19 Next we repeat this analysis using the actual I/B/E/S data. In this analysis, the standard deviation is a truly ex ante measure. The results reported in Panel C do not support the negative relation between the standard deviation of earnings forecasts and future returns. In fact, the relation between actual standard deviation and future return is slightly positive. The difference between the low and high quartiles is -.05 percent for raw returns and -.14 percent for beta excess returns. While the beta excess return is statistically significant, it is only 1.69 percent on an annualized basis. 17 The third rows in Panels B and C highlight the source of the differences by showing the average stock split factor for each quartile. In Panel B, there is a significant negative relation between the split factor and standard deviation of earnings forecasts. The lowest (highest) quartile of standard deviation has an average split factor of (1.378). This is precisely what we would expect to find for adjustment factors greater than one. Firms that have more stock splits tend to have lower adjusted earnings amounts and therefore lower standard deviations of forecasts. In addition, firms that are more likely to have splits are more likely the ones that have higher stock price performance. As shown in Panel C, the relation between the ex ante measure of actual standard deviation and the stock split factors is no longer statistically significant. The lowest (highest) quartile has an average split factor of (3.719). Researchers must be aware that the standard deviation reported in the usual I/B/E/S database is affected by future stock splits and therefore is partially an ex post measure of stock performance. Prior research has also documented a positive relation between uncertainty and analysts forecast optimism (e.g., Ackert and Athanassakos 1997; Das et al. 1998). We find that this relation is reduced (but still statistically significant) when using the actual data instead of the adjusted data. The rank correlation between the standard deviation of earnings forecasts and 17

20 forecast error is (p <.01) for the adjusted data and reduces to (p <.01) for the actual data. But, the results in Table 5 suggest that the market understands that analyst uncertainty is positively related with analyst optimism and it adjusts analysts forecasts appropriately in setting expectations to prevent a relation with future returns. Market reactions to On-target Earnings Announcements Baber and Kang (2002) point out that split adjusting and rounding analyst forecast data can result in contaminated observations (i.e., observations reported as having had zero forecast error but actually had nonzero forecast error). Since splits occur for superior performers ex post, Baber and Kang (2002) speculate that the contaminated observations are predominantly positive forecast errors that likely have positive announcement period returns. Inclusion of these contaminated observations therefore overstates the announcement period returns of on-target firms. They do not use actual forecast data to test their speculation but instead rely on the association between announcement period returns and split factors. Baber and Kang (2002) find that on-target firms with larger split factors tend to have more positive announcement period returns. They interpret this as evidence of spurious results caused by rounding adjusted forecast data. 18 Using sampling procedures and variable measurements similar to Baber and Kang (2002), we replicate their study by observing the three-day abnormal return surrounding ontarget earnings announcements for firms sorted by split factor. 19 The first two columns in Table 6 present Baber and Kang s (2002) results. As shown in the next two columns, our replication of their study using the adjusted I/B/E/S data produces a similar positive relation between announcement period returns and split factors. Mean abnormal returns are negative (positive) in 18

21 the three lowest (highest) split factor categories. In the final four columns of Table 6, one can see that approximately 70 percent of contaminated observations [= 1,717/(721+1,717)] had positive forecast errors and a positive average announcement period return of 1.17 percent. These results are consistent with Baber and Kang s (2002) speculation that contaminated observations introduce a spurious positive stock performance into the analysis of on-target observations. [Place Table 6 about here] Next, we examine this relation using the actual data. Because we use actual forecast data, we can more directly assess the extent to which the positive relation between announcement period returns and split factors is due to contamination. Using the actual data, we find that the positive relation between split factors and announcement period returns remains. The average return is -.75 (-.39 percent) for observations with split factors less than (equal to) one versus.63 percent for observations with split factors greater than These results suggest that the complexity of the relation between on-target earnings announcements and returns extends beyond contamination. While it is beyond the scope of this paper, a number of firm and investor characteristics could explain why some on-target earnings announcement returns are positive while others are negative. The split factor (which is an ex post measure) proxies for one or more of these characteristics. For example, it could be that when firms announce earnings for the current period, they also release information about future earnings. Firms that have zero forecast error and reveal positive future earnings news will likely receive a more positive market reaction than will firms that have zero forecast error but release no (or negative) future earnings news. Firms that release positive future earnings news are also the firms that will, on average, perform better in the future and have future stock splits. Therefore, there is some reason to expect a positive relation between stock split factors and current announcement period returns. 19

22 Eliminating all firms that have factors greater than one to decontaminate the sample may bias results and lead to incorrect inferences because of the elimination of better performing firms. In addition to testing for a stock price reaction at the time earnings are announced, researchers may be interested in testing for subsequent price movements. For example, the negative average announcement period returns to on-target observations in Table 6 may lead to speculation that the market has not adequately rewarded the company for meeting its forecasts. In this case, one would predict positive future abnormal returns for on-target firms as the market eventually provides these rewards. However, firms that have higher stock price performance are more likely to have stock splits, and firms with more stock splits are more likely to have prior forecast errors misclassified at zero. We test whether the I/B/E/S rounding procedure could lead to a mechanical relation between higher future returns and current zero forecast errors. This is similar to our analysis in the prior section concerning the association between future returns and forecast dispersion, except that forecast error is used in place of forecast dispersion. Using the sample and variables generated from our replication of Skinner and Sloan (2002), we measure the future returns of firms that report zero forecast error. Future returns are measured over the four quarters following the announcement of a zero forecast error. For the adjusted I/B/E/S data, we find an average quarterly abnormal return of 1.59 percent (6.51 percent on an annualized basis). In contrast, when we use the actual I/B/E/S data to identify firms that actually had zero forecast error, we find an average quarterly return of only -.07 percent (-.28 percent on an annualized basis). The average quarterly return for all firms in the sample is.04 percent. With the adjusted I/B/E/S data, firms with zero forecast errors appear to earn positive future abnormal returns. But this potential trading strategy is an illusion because, in reality, firms with zero forecast error had future abnormal returns close to zero. 20

23 III. CONCLUSION The forecast data provided by I/B/E/S to the academic research community play an important role in research investigating issues related to financial accounting information. I/B/E/S traditionally provides these data on a split-adjusted basis, presumably to provide comparable time-series data for its customers. For academic researchers, it is more beneficial in many cases to know the actual historical amounts that are not adjusted for subsequent stock splits. This paper investigates potential ramifications to academic researchers from using I/B/E/S split-adjusted data versus data that have not been split-adjusted. The problem for academics is that after I/B/E/S divides an earnings amount by the split factor, the result is rounded to two decimal places. When decimal places are removed by rounding, the adjustment cannot be reversed. In those cases researcher cannot determine prior years actual amounts. The rounding procedure is especially problematic when analyzing cases where forecast errors equal zero. The typical (adjusted) I/B/E/S database can report a zero forecast error when, in fact, the error based on the unadjusted data is non-zero. No type of adjustment by the researcher can correct this misclassification. We obtain actual (unadjusted) data from I/B/E/S. We then replicate prior research using the adjusted and actual I/B/E/S data and find that some conclusions are affected. Research conclusions are more likely affected by the rounding procedure in samples that have stock splits (e.g., larger firms, higher price-to-book firms, better performers, etc.) and where the research question focuses on zero amounts (e.g., assessing the percentage of zero forecast errors over time; relating firm characteristics to the probability of zero forecast error; calculating the market s reaction to zero forecast error; inferring earnings management based on the distribution of earnings, earnings changes, and forecast errors around zero). The rounding procedure also 21

24 reduces variation in forecasts across analysts which results in a downward bias in forecast dispersion for all firms with subsequent stock splits. The bias will be an increasing function of the split factor, and again, the researcher using only adjusted data cannot fully eliminate this bias. To ensure that future research findings are not affected by this potential misclassification, we recommend that researchers request the actual data from I/B/E/S and create their own splitadjusted data without rounding to the nearest penny. If the actual data cannot be obtained, an alternative is to recalculate I/B/E/S consensus statistics using the detail I/B/E/S adjusted data. The detail data are rounded to four decimal places, which allows a more accurate recalculation of the actual amount using the split factor provided by I/B/E/S. Research that uses the adjusted summary data from I/B/E/S should present sensitivity test results that include only observations with a stock split factor equal to one to ensure the reported results are not affected by the rounding procedure. Additional analyses that exclude observations with stock split factors greater than two or three, large market value firms, or high price-to-book firms should also be considered. Our results suggest that some conclusions of published research are affected by using the adjusted I/B/E/S data. Furthermore, other research studies may have been abandoned or rejected because the adjusted I/B/E/S data led to statistically insignificant results when, in fact, a relation existed. Using forecast data that are not rounded will increase a researcher s ability to investigate and draw conclusions on issues of importance to the academic and practice community

25 References Abarbanell, J., and R. Lehavy 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, University of North Carolina, Chapel Hill. Ackert, L., and G. Athanassakos Prior uncertainty, analyst bias, and subsequent abnormal returns. The Journal of Financial Research 20 (Summer): Ang, J., and S. Ciccone Analyst forecasts and stock returns. Working paper, Florida State University and University of New Hampshire. Baber, W., and S. Kang The impact of split adjusting and rounding on analysts forecast error calculations. Accounting Horizons 16 (December): Brown, L. D., R. L. Hagerman, P. A. Griffin, and M. E. Zmijewski Security analyst superiority relative to univariate time-series models in forecasting quarterly earnings. Journal of Accounting & Economics 9 (April): Brown, L. D Earnings forecasting research: Its implications for capital markets research. International Journal of Forecasting 9 (November): Brown, L. D Analysts forecast errors: Additional evidence. Financial Analysts Journal 53 (November/December): Brown, L. D A temporal analysis of earnings surprises: Profits versus losses. Journal of Accounting Research 39 (September): DeGeorge, F., J. Patel, and R. Zeckhauser Earnings management to exceed thresholds. Journal of Business 72 (January): Das, S., C. Levine, and K. Sivaramakrishnan Earnings predictability and bias in analysts earnings forecasts. The Accounting Review 73 (April): Elgers, P., and D. Murray The relative and complementary performance of analyst and security-price-based measures of expected earnings. Journal of Accounting & Economics 15 (June/September): I/B/E/S Research Bibliography I/B/E/S, New York, New York. Lakonishok, J., and B. Lev Stock splits and stock dividends: Why, who, and when. The Journal of Finance 42 (September): Matsumoto, D. A Management s incentives to avoid negative earnings surprises. The Accounting Review 77 (July): Skinner, D. J., and R. G. Sloan Earnings surprises, growth expectations, and stock returns: Don t let an earnings torpedo sink your portfolio. Review of Accounting Studies 7 (June/September):

26 Endnotes 1 Over 1,500 researchers at over 400 institutions worldwide rely on I/B/E/S analyst forecast data for research (I/B/E/S 2000). The most recent issue of the I/B/E/S Research Bibliography contains 575 abstracts from published and on-going research that use the I/B/E/S database. We reviewed publications for the past four years ( ) in The Journal of Accounting Research, Journal of Accounting and Economics, The Accounting Review, Journal of Accounting, Auditing and Finance, and Review of Accounting Studies. We found at least 40 articles that used the I/B/E/S summary tapes or First Call in the past four years. 2 The data provided by I/B/E/S on the Summary files are rounded to two decimals. The data provided on the Detail files are round to four decimals, indicating that the rounding issues discussed in this paper are less severe if the Detail files are used. The primary analysts forecasts data provided by FirstCall are adjusted for stock-splits and rounded to two decimal places. 3 We use a 4-for-1 split to clearly illustrate the effect of stock splits on adjusted data. Smaller factors could lead to rounding errors as well. For the example shown, a stock split factor as low as 1.07 would cause the negative one cent forecast error to be reported as zero forecast error. 4 The split adjustment and subsequent rounding can also affect intervals other than zero. For example, when I/B/E/S reports an adjusted forecast error of $.01 for a firm that has a split factor of 4, the actual forecast error might have been $.02, $.03, $.04 or $.05. All of these actual forecast errors would round to $.01 after being divided by 4. There is no way for the researcher to determine using the I/B/E/S data which amount was the actual forecast error. Scaling by price per share reduces, but does not always eliminate, this rounding effect. 5 In addition, when the researcher is interested in measuring the actual forecast error per share (unscaled), it is impossible to determine that amount for many of the observations. Multiplying by the split factor can only approximate the actual forecast error (see footnote 4). 6 The actual and adjusted per share I/B/E/S data used in our study come from the Summary files. 7 As a sensitivity test, we replicate the adjusted data by applying the adjustment factor from the adjusted tapes to the actual data. For 90 percent of our sample, our replication of the adjusted data and the adjusted data provided by I/B/E/S agree. Excluding the remaining 10 percent of the observations where we could not replicate the adjusted I/B/E/S data does not affect any of our conclusions. Of the observations that could not be replicated, approximately 90 percent (i.e., nine percent of the total sample) differ by one cent. 8 As discussed below, one of the studies we choose to replicate employs annual forecast data over the period. Following that study s sampling procedures, we also employ annual forecast data since We use a version of I/B/E/S data created on March 2001 for our analysis. Therefore, stocks that had stock splits subsequent to 1999 will report a split factor greater than one in The stock split factor is not monotonically increasing as newer companies such as Dell and Microsoft, who have had numerous stock splits, were added to the I/B/E/S database after The average split-adjustment factor for small (large) firms is 1.53 (2.31). The average split-adjustment factor for low (high) price-to-book firms is 1.51 (2.35). The correlation between size ranking and price-to-book ranking is 18 percent, suggesting that these variables primarily provide a different ranking of the data. 12 To ensure that changes in the reported stock split factor between our dataset and that used by DeGeorge et al. (1999) are not driving our results, we replicated all analyses in this section using I/B/E/S data created in March 1998 and find qualitatively similar results. 13 Since the frequency of observations in one cell affects the frequency of observations in another cell, the Chisquare tests are not independent across forecast error intervals. 14 Skinner and Sloan (2002) also estimate equation (1) over different return intervals (see their Table 5). Our results using the adjusted data for these alternative return windows produce identical conclusions to theirs. Using the actual data, however, we find no evidence across these return intervals of an asymmetric response to bad news and no evidence that the inferior returns to growth stocks are related to expectational errors by the market. 15 It is even possible that a firm with a high enough split factor will report a zero standard deviation, when in fact the standard deviation was positive. For example, suppose a firm had three analysts that made forecasts of $.08, $.09, and $.10. In a subsequent period, if the firm had a 3-for-1 stock split, I/B/E/S would report the three forecasts as $.03, $.03, and $.03 and the standard deviation as $.00. Limiting the sample observations in Tables 1 and 2 to those with three or more analysts following, we find that the adjusted I/B/E/S data shows that 14.1 percent of the 24