Earnings predictability and the direction of analysts' earnings forecast errors.

ABSTRACT: Das et al. (1998) suggest that as earnings become less predictable, analysts issue increasingly optimistic forecasts to please managers and consequently gain, or at least limit the loss of, access to managers' private information. We reexamine the association between earnings forecast

error and earnings predictability because there is evidence suggesting that deliberate earnings forecast optimism is not an effective mechanism for gaining access to managers' information (e.g., Eames et al. 2002; Matsumoto 2002). We document associations between earnings level and both forecast error and earnings predictability. These associations suggest that earnings level may be an important control variable when examining the association between forecast error and earnings predictability. When we control for the level of earnings we find no significant association between forecast error and earnings predictability. Thus, we find no evidence that analysts intentionally issue optimistically biased earnings forecasts.

Keywords: analyst earnings forecasts; forecast error; forecast bias; earnings predictability.

Data Availability: Data are commercially available from the sources identified in the text.

I. INTRODUCTION

Das et al. (1998) (hereafter DLS) examine the association between Value Line analysts' earnings forecasts and earnings predictability and find that as earnings become less predictable analysts' earnings forecasts become increasingly optimistic. To explain this pattern, DLS extend the management relations hypothesis (Francis and Philbrick 1993) (hereafter FP) to the earnings predictability context and argue that analysts intentionally bias their forecasts to curry favor with management in order to gain access to management's information. (1) DLS posit that as earnings become less predictable, analysts issue increasingly optimistic earnings forecasts because the potential gain in forecast accuracy from access to managers' nonpublic information is greater when earnings are less predictable. This intentional bias improves the overall accuracy of forecasts when the gain in accuracy through access to private information outweighs the loss in accuracy from the optimistic bias. DLS's evidence that analysts intentionally issue optimistically biased earnings forecasts is of interest to regulators (interested in establishing an environment where information is fairly and efficiently transferred to all investors), investors (as they interpret earnings forecast for their investment decisions), firms (as managers try to meet or beat the earnings forecast), and forecasting agencies (as they face the scrutiny of regulators and investors).

Contrary to DLS's theory, recent research suggests that issuing intentionally optimistic earnings forecasts is not an effective means for pleasing managers and improving access to their private information. Jorge and Rees (2000) present interview and survey evidence from Spain that managers more commonly respond to unfavorable analyst reports by providing more, rather than less, information. In addition, optimistic earnings forecasts result in negative earnings surprises, and related negative market reactions, while accurate and pessimistic forecasts have been linked with positive market responses (Bartov et al. 2002; Kinney et al. 2002; Kasznik and McNichols 2002; Lopez and Rees 2002; Skinner and Sloan 2001). Burgstahler and Eames (2002) present evidence that managers avoid negative earnings surprises by both managing earnings upward and analysts' forecasts downward--results inconsistent with a managerial preference for forecast optimism. Matsumoto (2002) argues that managers prefer pessimistic forecasts to avoid adverse market reactions to negative earnings surprises, and presents results consistent with this position.

FP and Kim and Lustgarten (1998) report evidence of greater earnings forecast optimism in association with less favorable recommendations, and reason that forecast optimism repairs the damage to the manager-analyst relationship invoked by unfavorable recommendations. However, Eames and Glover (2002) and Eames et al. (2002) further examine the association and conclude that FP's and Kim and Lustgarten's (1998) results likely stem from omitting the correlated variable--earnings level--from their analyses. Without including earnings level in their analyses, Eames and Glover (2002) and Eames et al. (2002) replicate the FP and Kim and Lustgarten (1998) results. However, when they control for earnings level they find no association between forecast optimism and recommendations for Value Line analysts (Eames and Glover 2002) and greater forecast optimism in association with more favorable recommendations for broker analysts (Eames et al. 2002). (2) These findings, combined with recent research indicating that issuing intentionally optimistic earnings forecasts is not an effective method to curry favor with management, lead us to explore whether the level of earnings is also an important control variable when considering the relation between analyst forecast errors and earnings predictability. Specifically, we investigate whether DLS's results are driven by their failure to consider the level of earnings in their analyses.

For earnings level to be an important control variable in examinations of the association between forecast error and earnings predictability, there must be associations between earnings level and both forecast error and earnings predictability. Numerous studies report an inverse relation between forecast error and the level of reported earnings (e.g., Brown 2001; Eames et al. 2002; Eames and Glover 2002; Hwang et al. 1996). The association reflects both earnings shocks due to unanticipated events and earnings management. Unexpected positive (negative) earnings shocks result in higher (lower) earnings and will generally be associated with more forecast pessimism (optimism) (Eames et al. 2001). Earnings baths are associated with low earnings and forecast optimism, while management to beat the forecast will lead to higher earnings and greater forecast pessimism (Abarbanell and Lehavy 2002).

In addition to the association between forecast error and earnings level, we document an association between earnings level and earnings predictability. We find earnings predictability is highest for earnings observations near the cross-sectional median level of earnings and that earnings predictability declines dramatically as earnings increasingly deviates from the median. With an association between forecast error and earnings level, and between earnings level and earnings predictability, we anticipate that earnings level will be an important control variable in examining the association between forecast error and earnings predictability.

We replicate DLS's results using analyses and data similar to theirs. However, plots of earnings predictability and the level of earnings reveal that the association between these two variables differs above and below median earnings. When we partition our sample at median earnings and further control for the level of earnings by including earnings as an independent variable, we find that neither subsample yields significant evidence of an association between forecast optimism and earnings predictability. (3) These findings suggest that DLS's results indicating intentional analyst bias are spurious and attributable to a failure to adequately consider earnings level controls in their analyses.

This study makes several important contributions to the existing literature examining earnings forecast errors. First, we provide evidence that earnings level is an important control variable to consider when examining the association between forecast error and earnings predictability. This finding, in combination with studies demonstrating that earnings is an important control variable when examining the association between forecast error and recommendations (e.g., Eames et al. 2002), suggests that research on earnings forecast error in other contexts should consider controlling for the level of earnings. Second, we demonstrate that forecast-error research that assumes linear relationships can be seriously flawed. A simple plot of the underlying data reveals that DLS's theory and findings do not generalize to at least half of the firms (i.e., those with earnings above the median). Third, we demonstrate the appropriate "switch point" (from optimism to pessimism) in earnings-forecast-error research is the economy-wide earnings mean/median and not zero (as suggested by Brown [2001]). Fourth, we find no significant association between forecast errors and earnings predictability when we include a control for earnings in our model and we conclude that DLS's results are spurious.

II. DATA AND SAMPLE SELECTION

For consistency with DLS, we use Value Line (VL) annual earnings forecasts and realizations. The primary advantage of using VL data is that VL offers neither brokerage nor underwriting services, so VL forecast errors are not confounded by incentives for promoting security transactions. Each week VL issues annual EPS forecasts sequentially covering approximately one-thirteenth of the roughly 1,700 firms followed in the VL Standard Edition. (4) Thus in each quarter, VL issues a single annual EPS forecast for each firm followed. From the VL Estimates and Projections File, we obtain realized annual EPS and realized shares outstanding. We also obtain forecasts of annual EPS and shares outstanding, as well as share price, the Value Line Timeliness Rank, and the Value Line ValuGauge Relative Position measure, all as of the forecast date. The ValuGauge Relative Position measure provides VL's index of forecasting difficulty. Values range from 1 to nearly 1,600, with low numbers for earnings that are easy to forecast and high numbers for earnings that are difficult to forecast. We limit our sample to forecasts of annual earnings for December fiscal-year-end firms, issued over the four quarters preceding year-end, and obtain a sample of 29,888 forecasts of annual earnings.

To obtain annual forecasted and actual earnings we multiply forecasted EPS and actual EPS by forecasted and actual shares, respectively. (5) Because we obtain actual and forecasted earnings for a broad range of firm sizes, and for consistency with prior forecast error research (DLS; Eames et al. 2002; FP), we scale both earnings and forecast error by the market value of equity at the first fiscal quarter's forecast date, computed as the product of VL reported share price at the first quarter forecast date and number of shares outstanding at the prior year end. (6) We compute annual forecast error as actual earnings less forecasted earnings. To limit the affect of outliers, in those instances where we are not using the DLS method of analyzing average values over the period 1989 to 1993, we delete 456 firm-quarter forecast observations for which the absolute value of earnings exceeds 50 percent of market value or the absolute value of forecast error exceeds 20 percent of market value, resulting in a sample of 29,432 observations.

Table 1, Panel A, presents sample observations by year. Sample sizes for individual years range from a low of 1,639 observations in 1987 to a high of 2,832 observations in 1996. A total of 1,335 firms are represented in the sample. The average number of observations per firm is 20.05. The maximum and minimum number of observations for a firm are 52 and 1, respectively. Panel B of Table 1 provides an industry breakdown using two-digit SIC codes for our sample of 29,432 firm-quarter forecasts of annual earnings.

Given systematic differences in annual forecast errors across a year's four fiscal quarters (DLS; Kang et al. 1994), we conduct separate analyses of forecast errors by quarter. (7) Table 1, Panel C presents forecast error distributional statistics by quarter. Consistent with prior research, the average forecast error is optimistic (DLS; FP; O'Brien 1988; Lys and Sohn 1990), and declines as the year progresses (Richardson et al. 2001). This decline is expected because early quarters' earnings are available when late quarter annual earnings forecasts are issued.

III. METHOD

Model

To examine the association between forecast error and earnings predictability we model annual forecast error as a function of earnings predictability, earnings level, firm size, and Value Line Timeliness Rank. Omitting firm and year subscripts for brevity, we obtain:

FE = [[beta].sub.0] + [[beta].sub.1] x UNPRED + [[beta].sub.2] x EARNINGS + [[beta].sub.3] x SIZE + [[beta].sub.4] x TIMELY + [epsilon].

In the remainder of this section we define the variables and describe their measurement. (8) Table 1, Panel D, reports distributional statistics for the-model's independent variables. We measure timeliness ranking, unpredictability, and forecast error on a quarterly basis to reflect the transitory nature of these variables. We measure firm size and actual earnings on an annual basis because firm size is relatively stable and we focus on errors in forecasting annual earnings. We define forecast error, FE, as actual annual earnings minus forecasted annual earnings, and scale forecast error by the market value of equity.

Earnings Predictability (UNPRED)

Value Line assesses the predictability of earnings based on the standard deviation of the percentage change in quarterly earnings from one year to the next, over the preceding eight years, with the most recent years weighted more heavily than earlier years. Value Line makes special proprietary adjustments for comparisons around zero earnings and when the sign of earnings changes (Value Line 2000, 2001). Value Line ranks the computed measures from 1 to nearly 1,600, where the most unstable earnings stream receives a value near 1,600 and the most stable earnings stream receives a value of 1. These ranks serve as the basis for VL's two published measures of earnings predictability. The first measure, the ValuGauge Relative Position, is the rank from 1 to nearly 1,600 and is available in the VL Estimates file. The second measure, the Value Line Predictability Index, further groups the initial ranking into 20 ranked groups valued from 5 to 100, in increments of 5, where 100 represents the highest level of predictability and 5 represents the least predictable earnings. We use the more precise ValuGauge Relative Position as our measure of earnings unpredictability (UNPRED), but note that use of the Value Line Predictability Index yields identical inferences. We use the VL measures of earnings predictability for two reasons. First, DLS show the VL Predictability Index has greater explanatory power in their model than the alternative earnings predictability measures they derive from time, series models, and second Luttman and Silhan (1993, 1995) show the Value Line Predictability Index is inversely related to subsequent absolute VL forecast accuracy, confirming that earnings are more difficult to forecast for low predictability firms. (9)

Earnings Level (EARNINGS)

We measure EARNINGS as realized earnings for the year scaled by equity market value at the first fiscal quarter's forecast date. Accumulating evidence shows that forecast error varies with the level of earnings (e.g., Butler and Saraoglu 1999; Brown 2001) and several studies find increasing forecast optimism (pessimism) as the level of firm earnings declines below (increases above) the average earnings for all firms (Hwang et al. 1996; Eames et al. 2002; Eames and Glover 2002). This association between forecast error and the level of actual earnings is not surprising since forecast error equals actual earnings less forecasted earnings. Previous studies attribute this association to earnings management (Abarbanell and Lehavy 2002), as well as earnings shocks and a tendency of analysts' forecasts to exhibit excess regression to the mean (Eames et al. 2001).

Figures 1-3 illustrate the paired empirical relationships among earnings level, VL forecast error, and earnings predictability for first quarter forecast observations, respectively. We obtain similar relationships for the other quarters (not reported). For Figure 1, we sort first quarter firm-year observations into portfolios based on earnings level, with 250 observations in all portfolios except the highest earnings portfolio. Then, for each portfolio, we plot the mean and median error in first quarter forecasts of annual earnings on the portfolio's median annual earnings. The figure shows that VL analysts' forecasts exhibit increasing optimism for scaled earnings below the median earnings value (6.5 percent) and increasing pessimism for earnings above the median. (10) The observed association is similar to that observed by Eames et al. (2002) for Zacks broker-analyst data and confirms the association for a population of VL analysts not subject to business generating incentives. The mean and median forecast optimism observed in Table 1, Panel C is driven by the fact that forecast optimism at lower earnings levels dominates the pessimism at higher earnings levels. Much of the previous forecast error research regarding the management relations hypothesis focuses on the overall mean optimism in forecast error. However, the pessimism illustrated in Figure 1 is not predicted or explicable by the management relations hypothesis.

[FIGURES 1-3 OMITTED]

While we document a significant association between earnings and forecast error, it is important to control for earnings level only if it is also associated with earnings predictability. In Figure 2, we plot mean and median UNPRED on portfolio median scaled earnings. The V-shape in Figure 2 shows that extreme earnings are least predictable and earnings near the median are most predictable. A joint consideration of Figures 1 and 2 indicates the potential for differing associations between forecast error and earnings predictability above and below median scaled earnings of 6.5 percent.

In Figure 3 we create separate plots for scaled earnings observations above (Panel A) and below (Panel B) 6.5 percent, the median value of earnings (see Table 1, Panel D). (11)

For scaled earnings above 6.5 percent (Panel A), we find increasing forecast pessimism for increasingly unpredictable earnings. For earnings values below 6.5 percent (Panel B), we find increasing forecast optimism as earnings becomes more unpredictable. Figure 3 suggests that the basic association between VL earnings predictability and analyst forecast errors reverses above and below median earnings. While the pattern for earnings below 6.5 percent (Panel B) is consistent with the DLS theory, the results for earnings above 6.5 percent are not. Thus, Figure 3, Panel A suggests that DLS's results do not generalize to approximately half of the firms in our sample. We find the argument that analysts would attempt to curry favor with increasingly optimistic forecasts for increasingly unpredictable earnings only when managers report earnings below the median counter-intuitive and suggestive of model misspecification in the DLS study.

Firm Size (SIZE)

DLS control for firm size because it can proxy for the amount of public and private information available for a firm (Atiase 1987), which implies smaller forecast errors and greater earnings predictability for larger firms. We control for firm size measured as the common logarithm of market value, computed as the product of VL reported price per share on the first quarter forecast date and the number of shares outstanding at the end of the preceding fiscal year.

Timeliness Ranking (TIMELY)

To control for the previously identified positive association between forecast optimism and the Value Line Timeliness Rank and for consistency with DLS, we include the timeliness rank, TIMELY, in our model. The VL Timeliness Rank is VL's estimate of the firm's relative firm stock price performance over the next six to 12 months, on an integral scale of from 1 (highest) to 5 (lowest) (Value Line 2001). Rather than relying on analyst input, VL generates timeliness ranks via a proprietary computational algorithm, which ranks price performance relative to the other stocks VL follows.

Analyses of Five-Year Average Values and Individual Forecast Observations

DLS base their analysis on five-year average values for the period 1989 to 1993, to facilitate their estimation of earnings predictability and to reduce the likelihood of overstating significance as a result of serial correlation in the error term (DLS 1998, footnote 2). This averaging approach reduces the likelihood of overstating significance, but at the cost of losing information in the averaging process. The use of five-year average values assumes the variables are reasonably constant over the study period and that average values are interpretable in the context of the research question. However, for the 282 first-quarter observations for which values are available in all years from 1989 to 1993, not a single firm exhibits constant forecast error or predictability, while only 15 firms exhibit constant timeliness. Furthermore, 50 percent of firms exhibit firm-specific standard deviations for UNPRED, FE, and TIMELY in excess of 79, 1.4 percent, and .70, respectively, suggesting substantial intrafirm temporal variability in the data. (12) Moreover, it is difficult to interpret average values in our research context. An average timeliness rank of 3 may reflect a series of timeliness ranks never containing a 3 or all 3s. Similarly, an average forecast error of zero may result from a mix of large positive and negative individual errors.

For consistency with DLS's method and study period, we first present results based on five-year average values over the period 1989 to 1993. However, considering the difficulty of interpreting results based on five-year averages, our primary analyses focus on individual forecasts of annual earnings, considered separately by forecast quarter. Our primary analyses are based on forecast observations over the period 1987 to 1999. We expect to find no association between forecast error and earnings predictability after controlling for the level of earnings. Consequently, serial correlation in the errors in our analysis of individual earnings forecasts may bias against our prediction of no relation. (13)

IV. RESULTS

Analyses with Five-Year Average Observations

For analysis of the five-year average values, like DLS, we limit the sample to firms for which VL forecasts are available in all quarters from 1989 to 1993, and do not eliminate or truncate outliers. This yields a sample of 255 observations for each quarterly forecast horizon. (14) The distribution of industries across this sample mimics that of DLS, for example, utilities represent 19 percent of the DLS sample and 17 percent of our sample. (15) The distribution of firm-specific average forecast errors, however, differs substantially between the samples. Mean forecast errors for our sample vary from -0.012 in the first quarter to -0.001 in the fourth quarter, while similar values for the DLS sample are -0.030 and -0.015, respectively. (16) Furthermore, our forecast errors exhibit much less range than the DLS sample. For each quarter our maximum forecast error was 1.40 and minimums ranged from -0.26 in the fourth quarter to -0.39 in the third quarter. In contrast, DLS's maximum was 1.43, and minimum values range from -2.7 in the first quarter to -2.4 in the third and fourth quarters. Thus, the DLS sample includes extreme forecast errors that were 2.7 times the magnitude of market value. (17)

Table 2 displays the regression results for five-year average values, both with and without the EARNINGS control variable. Results for the model excluding EARNINGS are generally consistent with DLS's results (DLS 1998, Table 4). For the first two quarters, coefficient estimates for UNPRED are negative and significant at the 1 percent level. For the third quarter the coefficient is again negative, but significant at the 5 percent level. For the fourth quarter the coefficient is negative, but only modestly significant (10 percent level). We obtain less significance and lower [R.sup.2] values than DLS, and attribute these differences to the extreme observations in the DLS sample.

Including EARNINGS in the regression model leads to very different conclusions. First, the coefficient estimates for EARNINGS are uniformly positive and highly significant, consistent with prior research (e.g., Eames et al. 2002; Eames and Glover 2002). With EARNINGS in the model, the explanatory power increases substantially and all reported [R.sup.2] values exceed 82 percent. We now find a positive and significant coefficient on UNPRED for all quarters; a pattern that is inconsistent with DLS's prediction that as earnings predictability decreases forecast optimism increases.

Analysis of Individual Firm-Quarter Forecasts of Annual Earnings

Considering the serious limitations of analyses of firm-specific five-year average values, our primary analyses are based on firm-quarter forecasts of annual earnings. Reviewing Figures 1 to 3, it appears that the linear model and EARNINGS control considered in the five-year average analysis is not appropriate. Considering the switch in the sign of the association between forecast error and earnings predictability above and below the median earnings, we construct two subsamples by partitioning our data at the median level of EARNINGS, and separately analyze each subsample. (18)

Table 3, Panel A reports regression results for observations with scaled earnings greater than 6.5 percent. Excluding EARNINGS from the model, the coefficient on UNPRED is positive and significant for each quarter, indicating as unpredictability increases, forecast pessimism increases--just the opposite of DLS's hypothesized behavior. Adding EARNINGS to the model, the coefficients on UNPRED are inconsistent in sign and not significant in the final three quarters, while in the first quarter the coefficient is positive and significant, again inconsistent with the DLS hypothesis. Thus, for earnings levels above the median we find no evidence supportive of the DLS hypothesis.

Table 3, Panel B reports regression results for observations with scaled earnings less than 6.5 percent. Without EARNINGS in the model the coefficients on UNPRED are negative and highly significant for each quarter, indicating as unpredictability increases, forecast optimism increases. This pattern is consistent with DLS's hypothesized behavior. However, with EARNINGS in the model the coefficients on UNPRED are not significant in any of the last three quarters, and only modestly significant (10 percent level) in the first quarter. Thus, even when we focus on the portion of the earnings distribution seemingly most consistent with the DLS hypothesis, we find that prior evidence in DLS supporting the theory that analysts issue intentionally optimistic forecasts as earnings unpredictability increases is likely attributable to a failure to adequately consider underlying relationships among earnings level, forecast error, and earnings predictability.

V. SUMMARY AND CONCLUSIONS

Relying on the management relations hypothesis developed by Francis and Philbrick (1993), Das et al. (1998) suggest that as earnings become less predictable, analysts have an incentive to issue optimistic forecasts to curry favor with management so the analyst can obtain management's private information. However, recent articles in the academic and popular press suggest that issuing intentionally optimistic earnings forecasts is not an effective means of currying favor with management because optimistic earnings forecasts lead to negative earnings surprises, which are often associated with negative market reactions.

We re-examine the association between Value Line forecast error and earnings predictability for the period 1987 to 1999 using a more extensive data set than Das et al. (1998). We are able to replicate Das et al.'s (1998) results with our sample using similar techniques and for the same sample period. However, our results indicate that earnings is an important control variable when investigating the association between forecast error and earnings predictability. We document an association between earnings level and earnings predictability where firms with the most predictable earnings typically report earnings near the sample median earnings level. Plots of the association between forecast error and earnings predictability above and below median earnings suggest that Das et al.'s (1998) theory and results do not generalize to firms reporting earnings above the median. When we control for the level of earnings by including earnings level as an independent regression variable, we find no significant association between forecast error and earnings predictability. Our results suggest that Das et al.'s (1998) results are spurious and were caused by omitting earnings level from the analysis. We conclude that there is no significant evidence of an inverse association between Value Line forecast error and earnings predictability.

TABLE 1
Distribution Statistics

Panel A: Sample Forecasts by Year (a)

Year    Frequency   Percent

 1987     1,639        5.57
 1988     1,798        6.11
 1989     1,840        6.25
 1990     2,048        6.96
 1991     2,377        8.08
 1992     2,064        7.01
 1993     2,206        7.50
 1994     1,731        5.88
 1995     2,814        9.56
 1996     2,832        9.62
 1997     2,655        9.02
 1998     2,610        8.87
 1999     2,818        9.57
Total    29,432      100.0

Panel B: Industry Breakdown (a)

SIC
Code   Industry                              Observations   Percentage

 10    Metal Mining                                733          2.5
 12    Coal Mining                                 519          1.8
 20    Food and Kindred Products                   809          2.7
 26    Paper                                       995          3.4
 27    Printing                                    869          3.0
 28    Plastics, Pharmaceuticals, Chemical       2,514          8.5
 29    Petroleum Refining, Paving, Roofing         731          2.5
 33    Iron and Steel                              436          1.5
 35    Machinery and Equipment                   2,403          8.2
 36    Electrical and Electronics                1,441          4.9
 37    Transportation Equipment                  1,202          4.1
 38    Measuring Instruments                       500          1.7
 48    Communications                              861          2.9
 49    Utilities                                 3,594         12.2
 60    Depository Institutions                   1,799          6.1
 61    Nondepository Credit Institutions           786          2.7
 63    Insurance                                 1,516          5.2
 73    Business Services                           615          2.1
 80    Health Care                               1,081          3.7
       Other                                     6,028         20.5
                         Total                  29,432        100.0

Panel C: Annual Forecast Error as a Percentage of Market Value by
Forecast Quarter (b)

Quarter   Observations    Mean    Median   Std. Dev.

First        7,213       -0.008   -0.002     0.035
Second       7,268       -0.007   -0.001     0.031
Third        7,556       -0.005   -0.001     0.026
Fourth       7,395       -0.003   -0.000     0.019
  Total     29,432

Panel D: Distributional Statistics for Independent Variables:
29,432 Firm-Quarter Observations (c)

            Mean     Median    Std. Dev.    Min       Max

UNPRED     713.661   695.000   438.597      1.000   1551.000
EARNINGS     0.062     0.065     0.056     -0.475      0.480
SIZE         3.151     3.132     0.668      1.126      5.516
TIMELY       3.031     3.000     0.9228     1.000      5.000

(a) All values obtained from VL Estimates and Projections File.

(b) Forecast error defined as actual annual earnings less forecasted
earnings scaled by market value at the first quarter forecast date.
All values obtained from VL Estimates and Projections File. All mean
and median values are significant at the 0.001 level.

(c) All variables relate to firm-specific values over the period
1987 to 1999.

  UNPRED = Value Line predictability ranking at the forecast date =
           ValuGauge Relative Position measure (1 to 1,600 with 1
           being most predictable). For the years 1987 to 1999 the
           highest predictability ranking issued was 1,580, but due
           to the elimination of extreme forecast error and earnings
           values our sample maximum is less;

EARNINGS = annual earnings scaled by market value of common equity
           at the first fiscal quarter's forecast date;

    SIZE = common logarithm of the market value of equity, market
           value of equity = beginning of the year outstanding shares
           x Value Line reported price per share on the first fiscal
           quarter's forecast date; and

  TIMELY = Value Line timeliness ranking at the forecast date. The
           Value Line timeliness ranking ranges from 1 to 5, with a
           rank of 1 being the most favorable.

TABLE 2
Association between Five-Year Average Earnings Predictability and
Five-Year Average Forecast Error Using Firm-Quarter Forecasts of
Annual Earnings (1989-1993)

FE = [[beta].sub.0] + [[beta].sub.1] x UNPRED + [[beta].sub.2]
x EARNINGS + [[beta].sub.3] x SIZE + [[beta].sub.4] x TIMELY
+ [epsilon]

                             Without EARNINGS Control

Independent        1st Qtr      2nd Qtr      3rd Qtr     4th Qtr
Variable

Intercept          0.055        0.077        0.044       0.054
UNPRED            -0.435 ***   -0.383 ***   -0.350 **   -0.252 *
  (x [10
  .sup.4])
SIZE              -0.001       -0.004       -0.004      -0.006
TIMELY            -0.012       -0.016       -0.005      -0.007
EARNINGS
Adjusted           2.8          2.4          1.2         0.29
  [R.sup.2] (%)
F-statistic        3.39 **      3.08 **      2.06        1.24
Observations                   255 for each quarter

                                 With EARNINGS Control

Independent         1st Qtr       2nd Qtr       3rd Qtr       4th Qtr
Variable

Intercept          -0.047 ***    -0.046 ***    -0.062 ***    -0.065
UNPRED              0.095 **      0.133 **      0.173 ***     0.250 ***
  (x [10
  .sup.4])
SIZE                0.000        -0.003        -0.002        -0.004
TIMELY             -0.008 **     -0.005        -0.000         0.003
EARNINGS            0.872 ***     0.842 ***     0.846 ***     0.804 ***
Adjusted           90.7          88.0          87.7          82.8
  [R.sup.2] (%)
F-statistic       621.83 ***    468.15 ***    452.53 ***    306.28 ***
Observations                     255 for each quarter

*, **, *** Significance at the <0.10, <0.05, and <0.01 levels,
respectively.

All variables relate to firm-specific five-year average values
over the period 1989 to 1993.

      FE = average forecast error, forecast error = annual earnings
           less forecast earnings, all scaled by market value of
           common equity at the first fiscal quarter's
           forecast date;
  UNPRED = average earnings unpredictability, Value Line (VL)
           ValuGauge Relative Position measure of earnings
           predictability. This measure ranges from 1 to 1,600,
           where 1 denotes highly predictable earnings;
EARNINGS = average earnings scaled by market value of equity at
           the first fiscal quarter's forecast date;
    SIZE = common logarithm of the average market value of equity,
           market value of equity = beginning of the year
           outstanding shares multiplied by VL reported price per
           share on the first fiscal quarter's forecast date; and
  TIMELY = average VL timeliness ranking. The VL timeliness
           ranking ranges from 1 to 5, with a rank of 1 being the
           most favorable.

TABLE 3
Association between Earnings Predictability and Analysts'
Forecast Error

FE = [[beta].sub.0] + [[beta].sub.1] x UNPRED + [[beta].sub.2]
x EARNINGS + [[beta].sub.3] x SIZE + [[beta].sub.4] x TIMELY
+ [epsilon]

Panel A: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Greater than 6.5% (1987-1999)

                           Without EARNINGS Control

Independent           1st Qtr       2nd Qtr     3rd Qtr
Variable

Intercept            0.020 ***     0.009 ***     0.007 ***
UNPRED               0.106 ***     0.079 ***     0.040 ***
  (x [10.sup.4])
SIZE                -0.000        -0.001        -0.000
TIMELY              -0.007 ***    -0.003 ***    -0.002 ***
EARNINGS
Adjusted            10.6           4.7           2.9
  [R.sup.2] (%)
F-statistic        142.73 ***     60.01 ***     37.64 ***
Observation          3,603         3,629         3,738

Panel B: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Less than 6.5% (1987-1999)

                         Without EARNINGS Control

Independent           1st Qtr       2nd Qtr       3rd Qtr
Variable

Intercept           -0.019 ***    -0.017 ***    -0.013 ***
UNPRED              -0.154 ***    -0.125 ***    -0.096 ***
  (x [10.sup.4])
SIZE                 0.012 ***     0.010 ***     0.007 ***
TIMELY              -0.009 ***    -0.007 ***    -0.005 ***
EARNINGS
Adjusted            16.4          13.1           9.2
  [R.sup.2] (%)
F-statistic        236.32 ***    183.45 ***    130.30 ***
Observation          3,610         3,639         3,818

Panel A: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Greater than 6.5% (1987-1999)

                    Without
                    EARNINGS       With EARNINGS Control
                    Control

Independent         4th Qtr        1st Qtr       2nd Qtr
Variable

Intercept           0.006 ***     0.021 ***     0.024 ***
UNPRED              0.025 ***     0.024 ***     0.004
  (x [10.sup.4])
SIZE               -0.001 ***     0.001 **      0.000
TIMELY             -0.001 ***    -0.006 ***    -0.002 ***
EARNINGS                          0.392 ***     0.330 ***
Adjusted            1.7          30.1          25.8
  [R.sup.2] (%)
F-statistic        21.63 ***    389.09 ***    316.65 ***
Observation         3,621         3,603         3,629

Panel B: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Less than 6.5% (1987-1999)

                    Without
                    EARNINGS       With EARNINGS Control
                    Control

Independent         4th Qtr        1st Qtr       2nd Qtr
Variable

Intercept          -0.006 **     -0.037 ***    -0.039 ***
UNPRED             -0.058 ***    -0.023 *       0.015
  (x [10.sup.4])
SIZE                0.004 ***     0.009 ***     0.007 ***
TIMELY             -0.003 ***    -0.007 ***    -0.004 ***
EARNINGS                          0.436 ***     0.400 ***
Adjusted            5.9          39.4          39.1
  [R.sup.2] (%)
F-statistic        79.4 ***     586.63 ***    583.96 ***
Observation         3,774         3,610         3,639

Panel A: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Greater than 6.5% (1987-1999)

                     With EARNINGS Control

Independent           3rd Qtr       4th Qtr
Variable

Intercept           -0.014 ***    -0.004 **
UNPRED              -0.004         0.005
  (x [10.sup.4])
SIZE                 0.000        -0.009 **
TIMELY              -0.001 ***    -0.001 **
EARNINGS             0.204 ***     0.093 ***
Adjusted            16.3           6.2
  [R.sup.2] (%)
F-statistic        183.09 ***     60.52 ***
Observation          3,738         3,621

Panel B: Using Firm-Quarter Forecasts of Annual Earnings with
Scaled Earnings Less than 6.5% (1987-1999)

                     With EARNINGS Control

Independent           3rd Qtr       4th Qtr
Variable

Intercept           -0.027 ***    -0.014 **
UNPRED               0.005         0.009
  (x [10.sup.4])
SIZE                 0.005 ***     0.003 ***
TIMELY              -0.003 ***    -0.002 ***
EARNINGS             0.265 ***     0.146 ***
Adjusted            28.2          19.0
  [R.sup.2] (%)
F-statistic        376.39 ***    222.17 ***
Observation          3,818         3,774

*, **, *** Significance at the <0.10, <0.05, and <0.01 levels,
respectively.

All variables relate firm-year observations for the period 1987-1999.

      FE = forecast error, forecast error = annual earnings less
           forecast earnings, all scaled by the market value of
           common equity at the first fiscal quarter's forecast
           date;
  UNPRED = earnings unpredictability at the forecast date, VL
           ValuGauge Relative Position measure of earnings
           predictability. This measure ranges from 1 to 1,600,
           where 1 denotes highly predictable earnings;
EARNINGS = earnings scaled by market value of equity at the first
           fiscal quarter's forecast date;
    SIZE = common logarithm of the market value of equity, market
           value of common equity = beginning of the year outstanding
           shares * Value Line reported price per share on the first
           fiscal quarter's forecast date; and
  TIMELY = Value Line timeliness ranking at the forecast date. The
           Value Line timeliness ranking ranges from 1 to 5, with a
           rank of 1 being the most favorable.

We are grateful for the comments of Sandra Chamberlain, Suzanne Luttman, the editor, two anonymous reviewers, participants at the 2000 Western Regional Meeting of the AAA and the 2002 European Accounting Conference. We particularly thank Linda Bamber for her insightful recommendations. We thank the PricewaterhouseCoopers Foundation, the Ernst & Young Faculty Research Fund, and the Accounting Development Funds at Santa Clara University and Brigham Young University for financial support. We thank I/B/E/S and Value Line for making their data available.

Editor's note: This paper was accepted by Terry Shevlin, Senior Editor.

(1) The management relations hypothesis (Francis and Philbrick 1993) proposes that analysts intentionally issue earnings forecasts that exceed their true expectations in order to counter the negative impact of unfavorable recommendations and to curry favor with managers and consequently gain, or at least limit the loss of, access to their private information. DLS note anecdotal evidence that managers may limit contact with analysts who are considered negative toward the company (see Drucker and Sapsford [2002] for a recent example of such anecdotal evidence).

(2) Eames et al. (2002) interpret their observation of greater forecast optimism for buy recommendations and greater forecast pessimism for sell recommendations as consistent with their motivated reasoning hypothesis, as well as broker-analysts engaging in trade-boosting behavior.

(3) Partitioning the sample at mean earnings (as well as other points in that vicinity, see footnote 11) yields essentially identical results to partitioning at the median.

(4) VL forecasted and actual EPS values exclude, "non-recurring or one-time gains and losses, which are noted in the footnotes." (See Value Line 2001.)

(5) This technique effectively adjusts EPS forecasts for stock splits, stock dividends, and sales of new and treasury shares.

(6) This value approximates beginning-of-the-year market value of equity and is available from the VL database. Similar results are obtained when we analyze the data scaled by book value.

(7) At the suggestion of a reviewer we also considered within-quarter forecast horizons by including an additional independent variable measuring time from the forecast date to quarter end. In none of our analyses did this additional variable lead to material changes in coefficient estimates, observed significance levels, or inferences regarding the association between forecast error and earnings predictability.

(8) Our model is similar to DLS's with two notable exceptions. First, we include EARNINGS as an explanatory variable, while DLS do not. Second, while DLS include I/B/E/S analyst following (FOLLOWING) and the cross-products FOLLOWING x UNPRED and SIZE x UNPRED, we exclude these from our model because (1) SIZE and FOLLOWING are highly correlated, (2) the inclusion of these variables is atheoretical, (3) we obtain the same results whether we exclude or include these terms in the model, and (4) exclusion simplifies presentation of the results.

(9) DLS suggest that a limitation of the VL predictability measures is that the indices vary over time for a specific firm. We consider this variability an advantage because it is unlikely that firm-specific earnings predictability remains constant over multiple year periods, as DLS assume in their study. DLS (1998, 289) further warn that the VL predictability measure is "potentially subject to strategic behavior." However, this concern is unfounded as there is no strategic component of the measure because VL uses a purely mechanical computing process.

(10) Since F[E.sub.it] = [E.sub.it] - [F.sub.it], it appears there is an obvious algebraic association between F[E.sub.it], and [E.sub.it]. However, because [F.sub.it] must be viewed as a function of [E.sub.it] we do not necessarily expect a positive association between forecast error and earnings level, and such an association only exists under restrictive assumptions regarding the behavior of [F.sub.it] with respect to [E.sub.it]. To see this, we set F[E.sub.it] = [E.sub.it] - [F.sub.it] = [E.sub.it] - ([B'.sub.0] + [B'.sub.0][E.sub.it] + [e.sub.it]) = [B.sub.0] + [B.sub.1][E.sub.it] + [e.sub.it], differentiate with respect to [E.sub.it], obtain 1 - ([differential]F/[differential]E) = B.sub.1], and note that [B.sub.1] is positive only if [differential]F/[differential]E < 1, and [B.sub.1] = 1 only if [differential]F/[differential]E = 0. There is no basis for asserting either of these conditions a priori. For our sample [differential]F/[differential]E = 0.57 and is significantly different from zero at the 1 percent level. Thus, the association between forecast error and earnings is not algebraic, but rather, a systematic relationship between forecast errors and earnings requires systematic analyst behavior, such as an inability to forecast some component or portion of earnings.

(11) Inferences from the figures and subsequent analyses are identical for a range of earnings level breaks from at least 6 to 7 percent.

(12) Affleck-Graves and Mendenhall (1990) report that, for the period 1982 to 1986, 51.5 percent of followed firms maintained their timeliness rank for a period of 13 weeks or less, only 8 percent maintained their rank for at least a year, and in an average week 5.3 percent of followed firms changed their rank. For a sample of 5 weeks covering September and October 1998 we find that in the average week 3.2 percent of firms experience a change in timeliness rank.

(13) We conduct sensitivity analyses to address the potential impact of both time-series and cross-sectional correlation in the error term and find no evidence of a significant association between FE and UNPRED. To limit the effect of serially correlated errors, we individually estimate the model for each of the 52 quarters from 1987 to 1999. In Figure 3, and in analyses reported later, we demonstrate the importance of partitioning on EARNINGS and including EARNINGS in the model; thus, for the sake of brevity, we limit our discussion to analyses incorporating these controls. Among the resulting 104 (2 x 52 quarters) regressions, 79 yield coefficients on the UNPRED that are nonsignificant at the 5 percent level. Of the remaining, 25 coefficients, 9 are significantly positive and 15 are significantly negative at the 5 percent level. We also estimate the model across all firms and years while treating year and firm-specific effects first as fixed effects then second as random components (Kennedy 1992). Generally, we find no significant or consistent association between forecast error and UNPRED. Applying the random-effects model to earnings observations below 6.5 percent we find evidence of significant pessimism as earnings predictability decreases. This result is not consistent with DLS or the nonsignificant results in Table 3, Panel B.

(14) The DLS sample consists of 239 firm-specific five-year average observations for each quarter. Requiring that observations for all variables be available for all quarters of the 20-quarter sample period significantly limits the sample. A wide variety of factors, such as the temporary suspension of timeliness ranks, nonreporting of earnings predictability, changes in the firms covered, and lack of I/B/E/S International Inc. reporting in a single year significantly reduce the sample size. Without such stringent requirements, we would have obtained a sample of 820 five-year average observations.

(15) This 17 percent is higher than the 12.2 percent for utilities in our sample of firm-year observations over the period 1987 to 1999 (Table 1, Panel B) due to the substantial continuity in data required for the five-year sample and the relative stability of firms in the utilities industry.

(16) The mean forecast error reported in DLS is significantly higher than that reported in related research (e.g., Brown 2001; FP; Eames et al. 2002; Eames and Glover 2002).

(17) While we gather VL data from the same years as DLS, the extreme observations reported by DLS may have been corrected or dropped by VL.

(18) The dichotomous control is in the same spirit as Brown (2001) who considers a temporal analysis of earnings surprises and finds decidedly different results for profit and loss firms. It is readily apparent from an inspection of both Figures 1 and 2 that the direction of forecast error switches around median and mean earnings. Thus, when analyzing forecast error the appropriate "switch point" is median/mean earnings and not profit and loss.

REFERENCES

Abarbanell, J., and R. Lehavy. 2002. Biased forecasts or biased earnings? The role of earnings management in explaining apparent optimism and inefficiency in analysts' earnings forecasts. Journal of Accounting and Economics (2002 Conference).

Affleck-Graves, J., L. Davis, and R.R. Mendenhall. 1990. Forecasts of earnings per share: Possible sources of analyst superiority and bias. Contemporary Accounting Research 6 (Spring): 501-517.

Atiase, R. 1987. Market implications of predisclosure information: Size and exchange effects. Journal of Accounting Research 25 (Spring): 168-176.

Bartov, E., D. Givoly, and C Hayn. 2002. The rewards to meeting or beating earnings expectations. Journal of Accounting and Economics 22 (June): 173-204.

Brown, L. 2001. A temporal analysis of earnings surprises: Profits versus losses. Journal of Accounting Research 39 (September): 221-241.

Burgstahler, D., and M. Eames. 2002. Management of earnings and analysts' forecasts to achieve zero and small positive earnings surprises. Working paper, University of Washington.

Butler, K., and H. Saraoglu. 1999. Improving analysts' negative earnings forecasts. Financial Analysts' Journal 55 (May/June): 48-56.

Das, S., C. B. Levine, and K. Sivaramakrishnan. 1998. Earnings predictability and bias in analysts' earnings forecasts. The Accounting Review 73 (April): 277-294.

Drucker, J., and J. Sapsford. 2002. Analyze this: Unhappy firm bites back. Wall Street Journal (November 22): C1.

Eames, M., S. Glover, and K. Stice. 2001. The relation between earnings and earnings forecast error. Working Paper, Brigham Young University.

--, and --. 2002. Value Line timeliness ranks and Value Line earnings forecast errors. Working Paper, Santa Clara University.

--, --, and J. Kennedy. 2002. The association between trading recommendations and broker-analysts' earnings forecasts. Journal of Accounting Research 40 (March): 85-104.

Francis, J., and D. R. Philbrick. 1993. Analysts' decisions as products of a multi-task environment. Journal of Accounting Research 31 (Autumn): 216-230.

Hwang, L., C. Jan, and S. Basu. 1996. Loss firms and analysts' earnings forecast errors. The Journal of Financial Statement Analysis 1 (Winter): 18-29.

Jorge, M., and W. Rees. 2000. Three triangulated studies of financial analysts in Spain. Working paper, University of Glasgow.

Kang, S., J. O'Brien, and K. Sivaramakrishnan. 1994. Analysts' interim earnings forecasts: Evidence of the forecasting process. Journal of Accounting Research 32 (Spring): 103-112.

Kasznik, R., and M. McNichols. 2002. Does meeting expectations matter? Evidence from analyst forecast revisions and share prices. Journal of Accounting Research 40 (June): 727-759.

Kennedy, P. 1992. A Guide to Econometrics. Third edition. Cambridge, MA: MIT Press.

Kim, C., and S. Lustgarten. 1998. Broker-analysts' trade-boosting incentive and their earnings forecast bias. Working paper, Queen's College of the City University of New York.

Kinney, W., D. Burgstahler, and R. Martin. 2002. Earnings surprise "materiality" as measured by stock returns. Journal of Accounting Research 40 (December): 1297-1329.

Lopez, T., and L. Rees. 2002. The effect of meeting analyst forecasts and systematic positive forecast errors on the information content of unexpected earnings. Journal of Accounting, Auditing, and Finance 17 (Spring): 155-184.

Luttman, S., and P. Silhan. 1993. An empirical analysis of the Value Line earnings predictability index. Journal of Applied Business Research 9 (4).

--, and --. 1995. Identifying factors consistently related to Value Line earnings predictability. The Financial Review 30 (August): 445-468.

Lys, T., and S. Sohn. 1990. The association between revisions of analysts' earnings forecasts and security-price changes. Journal of Accounting and Economics 13 (December): 341-363.

Matsumoto, D. A. 2002. Management's incentives to avoid negative earnings surprises. The Accounting Review 77 (July): 483-514.

O'Brien, P. C. 1988. Analysts' forecasts as earnings expectation. Journal of Accounting and Economics 10: 53-83.

Richardson, S., S. Teoh, and P. Wysocki. 2001. The walkdown to beatable analyst forecasts: The roles of equity issuance and insider trading incentives. Working paper, University of Michigan.

Skinner, D., and R. Sloan. 2001. Earnings surprises, growth expectations and stock returns or don't let an earnings torpedo sink your portfolio. Working paper, University of Michigan.

Value Line. 2000. Glossary of investment terms. Available at: http://valueline.com/sup_glossa.html.

--. 2001. How to Invest in Common Stocks: The Complete Guide to Using the Value Line Investment Survey. New York, NY: Value Line Publishing Inc.

Submitted May 2002

Accepted January 2003

Michael J. Eames

Santa Clara Univervsity

Steven M. Glover

Brigham Young University