Additional Evidence on Analysts’ Decision to Issue Disaggregated Earnings Forecasts: Strategic Biasing

Hypotheses Development

It has been well established in the literature that analysts strategically bias their forecasts to curry favour with management and to establish and maintain investment banking relationships with clients. Francis and Philbrick (1993) suggest that Value Line analysts issue overly optimistic earnings forecasts in an effort to gain access to management's private information. Lin and McNichols (1998) and Michaely and Womack (1999) show that affiliated analysts (analysts whose employers provide underwriting services to their clients) issue more optimistically biased stock recommendations than unaffiliated analysts. In a recent study, Jackson (2005) suggests that analysts also issue overly optimistic forecasts to increase the trading volume and therefore, commissions, for their brokerage houses. Understandably, analysts who engage in the type of strategic biasing described above are not likely to make public the details of their biased forecasts. For instance, an overly optimistic bias in disaggregated earnings forecasts is more likely to be detected by those who follow the sales projections through examining expected changes in economic and industry conditions, and the firm's market share, backlog and strategy, etc. In this case, the aggregated forecasts may delay the identification of this bias because outside observers cannot readily determine whether the optimistic earnings forecast resulted from intentional bias by the analyst or good faith estimates of lower expenses. Furthermore, some evidence (Ertimur et al., 2003) suggests that investors can use sales forecasts to assess the quality of earnings forecasts. Absent a sales forecast, this quality of earnings forecast assessment is not immediately obvious. In short, these strategic analysts are less likely to disaggregate their forecasts submitted to I/B/E/S. Other analysts, who may not have easy access to management or are more concerned with their professional reputation, may choose to focus on and showcase their forecast accuracy. These analysts are more likely to issue disaggregated forecasts (Ertimur et al., 2011; Marks, 2008). Because higher bias also translates into higher forecast error, ceteris paribus, our first hypothesis (stated in alternative form) is as follows:

As pointed out above, in the most recent year examined in the current study, 85% of the sample analysts either always, or never, disaggregated their forecasts when submitting them to I/B/E/S. Despite the significant upward trend over time towards disaggregation, a sizable number of analysts disaggregated for some firms, but not others, in the same year. These analysts are identified in the middle columns of Table 1 and we refer to them as selective disaggregators. If these analysts are attempting to balance reputation building (by being more accurate) with developing a closer relationship with management and investment banking clients (by strategically biasing their forecasts), we can test our predictions with respect to strategic disaggregation by focusing exclusively on this group. These analysts, when providing less biased and more accurate forecasts, will be able to justify their positions by providing additional corroborating evidence such as forecasts of sales and expenses. Alternatively, when they strategically bias their forecasts, they may find it difficult to justify their forecasted numbers when accompanied by the disclosure of additional data (sales forecasts) that make the bias more obvious. Consequently, these strategic analysts are less likely to disaggregate their biased forecasts. Furthermore, their disaggregated forecasts are likely to be less optimistically biased when compared to their own non-disaggregated forecasts. Because higher bias also translates into higher forecast error, ceteris paribus, our second hypothesis (stated in alternative form) is as follows:

The decision to disaggregate, if driven by strategic biasing, may also be reflected at the firm-level in a given year. For example, the analyst may strategically choose to not disaggregate a forecast for a particular firm in a given year when there is a need to bias the forecast to promote a security issuance of an investment banking client. For instance, defamed and banned ‘all-star’ telecom analyst Jack Grubman issued more optimistic forecasts and changed his recommendation to ‘buy’ for AT&T in order to attract underwriting business for his employer, Solomon Smith Barney-Citi Group, in the initial public offering of ATT Wireless tracking stock (AWE). Though this strategic biasing received considerable media coverage and public attention because of its blatancy and the stature of the analyst, other similar occurrences by lesser known analysts in other contexts may be less obvious. Therefore, we analyse an additional measure of the degree of analyst disaggregation at the analyst-firm level by examining the persistence of an analyst's disaggregation for a firm over the years of his/her coverage. If an analyst disaggregated forecasts for a given firm for all future years after initially issuing a disaggregated forecast for that firm, we refer to that analyst as a persistent disaggregator. Alternatively, an analyst who does not disaggregate all future forecasts after initially issuing a disaggregated forecast for a particular firm is referred to as non-persistent disaggregator. Because the actions of non-persistent disaggregator analysts are more likely to be strategic and optimistically biased, our third hypothesis (stated in alternative form) is as follows:

Test of Hypothesis 1

Hypothesis 1 predicts that analysts who do not disaggregate their forecasts are expected to be more optimistically biased and less accurate than analysts who disaggregate. To test this hypothesis, it is necessary to construct a measure of analysts’ level of disaggregation. Because analysts can disaggregate for a firm in one year and not in other years, or they can disaggregate for one firm and not for other firms in the same year, and because the proportions disaggregated can vary across analysts, the data are analysed by developing a categorical classification scheme based on analysts’ overall disaggregation behaviour. Another consideration is that the number of years that a given analyst's earnings forecasts appear in the I/B/E/S database can vary significantly from a few years to a relatively long tenure. We initially examined a sample that required analysts to have provided forecasts in at least three of the 11 years covered by the data and then we incrementally ratcheted the minimum requirement up to at least six years in the 11 years examined. In the Sensitivity Tests section of the paper, we test the robustness of our findings to these additional (more restrictive) analyst inclusion criteria. Because the largest sample size is achieved when the least restrictive analyst inclusion criterion is applied, we present results here for the sample that includes all analysts that had a minimum of three years of forecasts in I/B/E/S. We classify selected analysts into one of the following four categories:

Because we are examining differences in bias and accuracy of analysts’ forecasts based on the extent of disaggregation of issued forecasts, the four categories of analysts are constructed such that the level of disaggregation increases monotonically when moving from category 1 to category 4. Categories 1 and 4 represent two extremes of the disaggregation categorization because they include analysts who either never disaggregated their forecasts (category 1) or always disaggregated their forecasts (category 4) over the sample years. Because partial disaggregators are non-homogeneous in terms of the pattern of their disaggregation behaviour, some judgment was required to classify them into meaningful categories for analysis. PD1 includes analysts who we judged to be less prevalent disaggregators than those included in the PD2 group. To test Hypothesis 1, we employ both univariate and multivariate (regression-based) tests to arrive at our conclusions.

Multivariate tests  In addition to the univariate tests described above, we also performed multivariate regression analyses to control for other factors identified in the prior literature that have an effect on analyst forecast bias and accuracy. Das et al. (1998) find that bias is positively associated with the level of difficulty in predicting the firm's future earnings. As a proxy for the information environment of the firm, size has been shown to be negatively associated with the level of bias (e.g., Brown, 1997, 2001; Das et al., 1998). Brown (2001) finds that firms reporting losses receive more optimistically biased earnings forecasts. Richardson et al. (2004) show that analysts issue optimistic earnings forecasts at the beginning of the year and then gradually revise down their estimates to a level that the firms can beat later in the year. Consequently, there appears to be a negative relation between the timeliness of earnings forecasts and measures of forecast bias. The same study also shows that high book-to-market (low growth) firms receive more optimistically biased earnings forecasts. In addition, previous studies have found that the level of investment in technology (e.g., the ratio of research and development expenditures to total assets) is positively associated with pessimistic earnings forecasts (Matsumoto, 2002; Richardson et al., 2004). At the analyst level, analysts with longer tenure should be better forecasters because of their experience, although they may be indifferent with respect to the decision to bias their forecasts. Lastly, regardless of other factors, some forecasts may be less accurate simply because they are less timely (Sinha et al., 1997). To control for these factors, we included the following variables in our multivariate regression models:

• 1

  TIMELINESS is a measure of closeness of the forecast to the earnings announcement date, measured as the number of months prior to the actual earnings announcement date that the earnings forecast is issued. As a result, this variable takes on a negative value with higher values associated with more timely forecasts.

• 2

  LOSS is a dummy variable equal to 1 if actual EPS as reported on I/B/E/S for the fiscal year is negative, and 0 otherwise.

• 3

  Because some firm-years may be harder to forecast than others, we define DIFFICULTY as the standard deviation of the most recent forecasts made by all analysts for that firm in a given year, normalized by the mean of these forecasts, as a control for differences in difficulty in predicting a firm's earnings.

• 4

  Because large firms face less information asymmetry than small firms, we use SIZE, measured by the log of the market value of owners’ equity at the start of the current fiscal year, as a proxy to control for this effect.

• 5

  We use B/M, the ratio of book value to market value, computed as the book value of common equity at the start of the fiscal year divided by market capitalization at the start of the fiscal year, to control for differences in errors and biases induced by firms with different future growth opportunities.

• 6

  Difference in forecasting accuracy induced by experience in the industry is controlled by the TENURE variable, measured by the number of years an analyst has appeared in the I/B/E/S database on the forecast date (Clement, 1999).

• 7

  To control for the potential biasing effects of the level of investment in technology, we include a dummy variable TECH, which takes a value of 1 if the firm belongs to the Information Technology and Health Care sectors (including 15 high tech industries) of the Global Industry Classification Standard (GICS) industry classification, and 0 otherwise.

• 8

  We control for differences in forecasting earnings for the remaining (non-Tech) industries and across the 11 years that cover various stages of economic expansion and contraction, by adding dummy variables for industry effects and year effects.

To test for differences in BIAS and FORECAST ERROR in a multivariate setting, we consider two regression models that differ only in how the level of analyst disaggregation is measured. The specifications for these two models for BIAS and FORECAST ERROR are as follows:

()

()

()

()

In models (1a) and (2a), we use PD1, PD2 and FD as dummy variables corresponding to the analyst disaggregation categories described earlier. A directional test of the significance of the difference in the coefficients of these dummy variables provides a test of the relationship between the level of disaggregation and BIAS and FORECAST ERROR while controlling for the variables discussed above. In models (1b) and (2b), we construct a single ordinal disaggregation variable, LEVEL_D, which takes a value of 1, 2, 3 or 4 for analysts classified into ND, PD1, PD2 or FD categories, respectively. A test of significance of the coefficient for this variable provides a basis for testing the hypothesis.

Tests of BIAS  Panel A and Panel B of Table 3 present results of the estimation of models (1a) and (1b), respectively. The F-statistics for the two models and the t-statistics for the significance of coefficients for all the control variables (except TENURE) are significant at conventional levels of significance (0.05 or better). In both models, all the control variables, except for DIFFICULTY, have the expected signs. For model (1a), the estimated coefficients for PD1, PD2, and FD are significantly different from zero at the conventional levels of significance (0.05 or better) and monotonically decreasing from −0.006 to −0.010 to −0.016, respectively. The F-statistics for the tests that PD1=PD2 is 4.85 (p= 0.028), that PD2=FD is 9.38 (p < 0.001), and that PD1=PD2=FD is 8.5 (p < 0.001), all statistically significant at conventional levels of significance (0.05 or better). For model (1b), disaggregation is measured using an ordinal variable, LEVEL_D. The estimated coefficient for this variable is -0.005, which is negative and statistically significant (p= 0.010). These results corroborate our univariate analysis findings that the level of disaggregation is negatively associated with the level of FORECAST BIAS. Collectively, these multivariate results, along with the univariate results presented earlier, provide strong support for Hypothesis 1.

Table 3. MULTIVARIATE TESTS OF DIFFERENCES IN BIAS AND FORECAST ERROR ACROSS ANALYST DISAGGREGATION CATEGORIES

Presented below are the results from the regression of BIAS and FORECAST ERROR on analysts, firm and industry control variables and levels of analyst forecast disaggregation. BIAS is defined as (forecasted one-year ahead EPS – actual EPS) ÷|actual EPS| and the FORECAST ERROR is defined as |forecasted one-year ahead EPS – actual EPS|÷|actual EPS|. Panel A presents results when forecast disaggregation is measured using PD1, PD2 and FD categorical dummy variables. Panel B presents regression results when forecast disaggregation is measured using LEVEL_D, an ordinal variable that takes values of 1, 2, 3, and 4 for analysts belonging to categories ND, PD1, PD2, and FD, respectively.
Panel A: Categorical measurement of forecast disaggregation categories
()
()
	Predicted sign	Dependent variable
		BIAS		FORECAST ERROR
		Coefficient	p-value	Coefficient	p-value
Intercept		0.045	(0.01)	0.240	(0.01)
PD1	−	−0.006	(0.01)	−0.010	(0.01)
PD2	−	−0.010	(0.01)	−0.011	(0.01)
FD	−	−0.016	(0.01)	−0.020	(0.01)
TIMELINESS	−	−0.010	(0.01)	−0.015	(0.01)
LOSS	+	1.430	(0.01)	1.330	(0.01)
TECH	−	−0.017	(0.01)	−0.010	(0.01)
DIFFICULTY	+	−0.000	(0.01)	0.000	(0.77)
SIZE	−	−0.007	(0.01)	−0.019	(0.01)
B/M	+	0.008	(0.01)	0.004	(0.01)
TENURE	?	−0.0004	(0.11)	−0.002	(0.01)
INDUSTRY EFFECT		Yes		Yes
YEAR EFFECT		Yes		Yes
Adj R2		23.10%		27.22%
N		242,334		242,334
Test		F-statistic	p-value	F-statistic	p-value
PD1= PD2		4.85	0.028	0.68	0.41
PD1= FD		16.94	<0.0001	28.50	<0.0001
PD2= FD		9.38	<0.0001	29.16	<0.0001
PD1=PD2= FD		8.50	<0.0001	16.15	<0.0001

Panel B: Ordinal measurement of forecast disaggregation categories
()
()
	Predicted sign	Dependent variable
		BIAS		FORECAST ERROR
		Coefficient	p-value	Coefficient	p-value
Intercept		0.0508	(0.01)	0.2400	(0.01)
LEVEL_D	−	−0.0050	(0.01)	−0.0070	(0.01)
TIMELINESS	−	0.0000	(0.01)	0.0005	(0.01)
LOSS	+	1.4300	(0.01)	1.3300	(0.01)
TECH	−	−0.0170	(0.01)	−0.0090	(0.01)
DIFFICULTY	+	−0.0000	(0.01)	0.0000	(0.77)
SIZE	−	−0.0070	(0.01)	−0.0190	(0.01)
B/M	+	0.0080	(0.01)	0.0040	(0.01)
TENURE	?	−0.0003	(0.16)	−0.0020	(0.01)
INDUSTRY EFFECT		Yes		Yes
YEAR EFFECT		Yes		Yes
Adj R2		23.10%		27.25%
N		242,334		242,334

• PD1, PD2 and FD are dummy variables corresponding to analyst disaggregation categories, PD1, PD2 and FD, respectively. LEVEL_D is an ordinal disaggregation variable, which takes a value of 1, 2, 3 or 4 for analysts belonging to ND, PD1, PD2 or FD categories, respectively. Analyst disaggregation categories are (1) ND representing analysts who never issued disaggregated forecasts for any firm they followed during the entire sample period, (2) PD1 representing analysts who issued disaggregated forecasts for some, but not all, of the firms they followed each year and analysts who provided no sales forecast for some of the sample years but did issue sales forecasts for some of their firms in the other sample years, (3) PD2 representing analysts who issued disaggregated forecasts for all of the firms they followed for some, but not all, of the sample years, and (4) FD representing analysts who issued disaggregated forecasts for all the firms they followed for all the sample years. TIMELINESS is measured by the number of months prior to the actual earnings announcement date that the earnings forecast is issued. LOSS is a dummy variable equal to 1 if actual EPS as reported on I/B/E/S for the fiscal year is negative, and 0 otherwise. DIFFICULTY is the standard deviation of the most recent EPS₁ forecasts made by all analysts for that firm in a given year, normalized by the mean of these forecasts. SIZE is measured by the log of market value of the firm equity at the start of the current fiscal year. B/M is the book value of common equity at the start of the fiscal year divided by market capitalization at the start of the fiscal year. TENURE is measured by the number of years an analyst has appeared in the I/B/E/S database. TECH is a dummy variable, which takes a value of 1 if the firm belongs to Information Technology and Health Care sectors (including 15 high tech industries) of the GICS industry classification, and 0 otherwise. INDUSTRY EFFECT represents results from the test of the hypothesis of the equality of coefficients for the dummy variables constructed on the basis of the second level of GICS industry classification. YEAR EFFECT represents results from the test of the hypothesis of the equality of coefficients for the year dummy variables.

Test of Hypothesis 2

Hypothesis 2 makes predictions about the BIAS and FORECAST ERROR of selective disaggregators who disaggregate earnings forecasts for some, but not all, firms in a given year. As per this hypothesis, selective disaggregators are likely to be less biased and more accurate in their disaggregated forecasts when compared to their aggregated forecasts. These analysts are identified in the middle columns of Table 1. The number of analysts falling into this group ranges from a low of 625 (15%) in 2008 to a high of 1,397 (34%) in 2001. In order to provide a strong test for differences in bias and accuracy of these analysts on the disaggregation dimension, we compared the BIAS and FORECAST ERROR of selective disaggregators for the set of firms for which they issued aggregate forecasts versus the set of firms for which they issued disaggregated forecasts. We first computed the mean BIAS and mean FORECAST ERROR for each analyst for the two sets of firms in a given year and then pooled these data across all years. Table 4 presents the results of a matched-pair t-test of the differences in BIAS and FORECAST ERROR between the two sets of firms. By analysing the data in this fashion, each analyst is essentially serving as his/her own control for differences that may exist amongst analysts with respect to factors affecting forecasting ability. As the results in Table 4 show, the difference in mean BIAS and the difference in mean FORECAST ERROR of aggregated and disaggregated forecast of an average selective disaggregator is 0.059 and 0.085, respectively. Both of these differences are positive and significantly different from zero at the 5% level of significance (p= 0.030 and 0.010, respectively). Consistent with the predictions of Hypothesis 2, these results provide evidence that the selective disaggregators are more optimistically biased and less accurate when they do not disaggregate their forecasts than when they do disaggregate.

Multivariate tests  Based on the earlier discussion of potentially confounding variables that may affect analysts’ propensity to bias their forecasts, we also conducted multivariate tests to compare the BIAS and FORECAST ERROR of selective disaggregators when they disaggregate their forecasts versus when they do not disaggregate. Specifically, we regressed BIAS and FORECAST ERROR on DISAGGREGATE, a dummy variable that takes a value of 1 if the analyst disaggregated the forecast, and 0 otherwise, after appropriate controls. We included dummy variables to control for the ANALYST EFFECT, INDUSTRY EFFECT and the YEAR EFFECT, along with other variables identified earlier that may affect BIAS or FORECAST ERROR. The specifications for the two estimated regression models are as follows:

()

()

The results presented in Table 5 from the estimation of model (3a) with BIAS as the dependent variable show that the adjusted R² is 0.238 and the model is significant at conventional levels (0.05 or better). All the control variables, except DIFFICULTY, have the predicted signs and are statistically significant at conventional levels of significance (0.05 or better). The coefficient for DISAGGREGATE is -0.006, which is negative and significant (p < 0.01).

Table 5. MULTIVARIATE TESTS OF DIFFERENCES IN BIAS AND FORECAST ERROR OF SELECTIVE DISAGGREGATORS

Presented below are the results from the regression of BIAS and FORECAST ERROR on analysts, firm and industry control variables and analysts’ decision to disaggregate the issued forecast. BIAS is defined as (forecasted one-year ahead EPS – actual EPS) ÷|actual EPS| and the FORECAST ERROR is defined as |forecasted one-year ahead EPS – actual EPS|÷|actual EPS|. Analysis is confined to the sample of selective disaggregators who disaggregated for some, but not all, of the firms that they followed in a given year.
()
()
	Predicted sign	Dependent variable
		BIAS		FORECAST ERROR
		Coefficient	p-value	Coefficient	p-value
Intercept		0.085	(0.01)	0.250	(0.01)
DISAGGREGATE	−	−0.006	(0.01)	−0.005	(0.01)
TIMELINESS	−	−0.010	(0.01)	−0.016	(0.01)
LOSS	+	1.499	(0.01)	1.325	(0.01)
TECH	−	−0.032	(0.01)	−0.005	(0.20)
DIFFICULTY	+	−0.020	(0.01)	−0.002	(0.02)
SIZE	−	−0.008	(0.01)	−0.019	(0.01)
B/M	+	0.006	(0.01)	0.003	(0.01)
TENURE	?	−0.001	(0.05)	−0.002	(0.01)
ANALYST EFFECT		Yes		Yes
INDUSTRY EFFECT		Yes		Yes
YEAR EFFECT		Yes		Yes
Adj R2		23.84%		27.90%
N		81,760		81,760

• DISAGGREGATE is a dummy variable that takes the value of 1 if the analyst issues disaggregated forecasts for a firm in a given year, and 0 otherwise. TIMELINESS is the number of months prior to the actual earnings announcement date that the earnings forecast is issued. LOSS is a dummy variable equal to 1 if actual EPS as reported on I/B/E/S for the fiscal year is negative and 0 otherwise. DIFFICULTY is the standard deviation of the most recent EPS₁ forecasts made by all analysts for that firm in a given year, normalized by the mean of these forecasts. SIZE is the log of market value of the firm equity at the start of the current fiscal year. B/M is the book value of common equity at the start of the fiscal year divided by market capitalization at the start of the fiscal year. TENURE is measured by the number of years an analyst has appeared in the I/B/E/S database. TECH is a dummy variable, which takes a value of 1 if the firm belongs to Information Technology and Health Care sectors (including 15 high tech industries) of the GICS, and 0 otherwise. INDUSTRY EFFECT represents results from the test of the hypothesis on the equality of coefficients for the dummy variables constructed on the basis of the second level of GICS industry classification. YEAR EFFECT represents results from the test of the hypothesis on the equality of coefficients for the year dummy variables.

Table 5 also presents results from the estimation of model (3b) with FORECAST ERROR as the dependent variable. The adjusted R² is 0.279 and the model F-statistic (not reported) is significant at conventional levels of significance (0.05 or better). All the control variables have the expected signs and are statistically significant at conventional levels of significance (0.05 or better). The coefficient for the test variable, DISAGGREGATE, is -0.005, which is negative and significant (p < 0.01). Overall, both sets of regression results are consistent with the univariate results reported above and support the predictions of Hypothesis 2 that selective disaggregators’ disaggregated forecasts will be associated with lower levels of forecast errors and less optimistic bias than their non-disaggregated forecasts.

Test of Hypothesis 3

To test Hypothesis 3, we developed definitions of persistent and non-persistent disaggregators. This posed somewhat of a challenge because the data are not balanced, and the number of analysts and the number of firms covered by each analyst often change from year to year. For purposes of this analysis, we defined a persistent disaggregator (P) at the analyst-firm level as an analyst who meets the following two criteria. First, the analyst must have continued to issue sales forecasts for a particular firm every year subsequent to the year the first sales forecast was generated for that firm. This condition was imposed to ensure that an analyst was committed to issuing disaggregated forecasts once he/she started that practice. Second, we also required that the analyst followed the firm for a minimum of three years subsequent to the year of the initial sales forecast generated for that firm. This is done to minimize the possibility that this group included those analysts who may have begun issuing disaggregated forecasts for a firm toward the end of the period examined in this study and then stopped doing so in the later years for which data were unavailable. This effectively limited persistent disaggregators to those analysts who had issued disaggregated forecasts for their firms for a minimum of four consecutive years.

Because non-persistent disaggregators can vary with respect to the extent of non-disaggregation, we developed three mutually exclusive sub-categories at the analyst-firm level as follows: (a) analysts who never disaggregated the forecasts of the firm (NP0), (b) analysts who never disaggregated a forecast of the firm after issuing just one sales forecast (NP1), and (c) analysts who disaggregated the forecasts of the firm more than once but did not meet the strict definition of a persistent disaggregator (NP2). Thus, NP0, NP1, NP2 and P are categories that include analysts who are progressively increasing in the degree of their forecast disaggregation persistence.

Table 6 presents the mean BIAS and the mean FORECAST ERROR of analyst-firm forecasts pooled over the years for which the forecasts were made by each of the four forecast persistence categories defined above. The table is arranged starting with the least persistent group of disaggregators (NP0) and ending with the most persistent group (P). Our predictions, based on Hypothesis 3, are that the mean BIAS (FORECAST ERROR) of the analyst-firm categories will monotonically decrease (increase) from the least persistent disaggregators (NP0) to the most persistent disaggregators (P). The results presented in Table 6 show that the mean BIAS for NP0, NP1, NP2, and P are 0.080, 0.058, 0.036 and 0.007 respectively. The magnitude of BIAS decreases monotonically as the level of persistence increases. The results of Duncan's multiple range test confirms that the differences in BIAS are statistically significant across all disaggregation categories at conventional levels of significance (0.05 or better). We obtain similar results for FORECAST ERROR. As predicted, the mean FORECAST ERROR decreases from 0.186 to 0.163 to 0.146 to 0.095 (and thus, the mean forecast accuracy increases) monotonically as the level of disaggregation increases from NP0 to NP1 to NP2 to P, respectively. Duncan's multiple range test confirms that the differences in FORECAST ERROR are significantly different across all categories at conventional levels of significance (0.05 or better). To summarize, the results presented in Table 6 suggest that analysts who provide disaggregated forecasts for a firm in some years, but not others, are more biased and less accurate than analysts who issue disaggregated forecasts on a more consistent basis. This finding provides support for Hypothesis 3, and the argument that strategic biasing is an important determinant of analysts’ disaggregation decision.

Sensitivity Test 1

Tests of Hypothesis 1, using the univariate and multivariate analyses presented in Panel A of Tables 2 and 3, show that the level of forecast bias decreases monotonically as the degree of disaggregation increases from Non-disaggregators (ND) to Full disaggregators (FD). However, if the decision to disaggregate earnings forecasts is tied to specific industries such as the high tech industry where losses are frequently reported and the demand for sales forecasts is higher, the documented relationship between forecast bias and the degree of disaggregation may also be driven by analyst industry coverage. To this end, we examined analyst disaggregation behaviour across industry groupings. Panel A of Table 7 presents a breakdown of Technology, Non-technology and Unclassifiable industry groupings by disaggregation categories. The proportions of analysts belonging to ND, PD1, PD2, and FD categories for technology firms are 13%, 16%, 36%, and 48%, respectively. This evidence of the largest (smallest) concentration of FD (ND) analysts for the technology firms is consistent with the expected source of bias. To ensure that the reported results are robust and not driven by this potential underlying industry effect, we partitioned the original sample of analysts into two subsamples—a technology subsample and a non-technology subsample. Regression models (1a) and (1b) were estimated separately for each subsample and the results and inferences were qualitatively similar for both the estimations. For brevity, only the results from the estimation of model (1b) are presented in Panel B of Table 7. The results show that for both of the subsamples, the coefficient of the disaggregation variable, LEVEL_D, is negative and significantly different from zero at conventional levels of significance (0.05 or better). Although not reported here, we also estimated a regression model with an ordinal measure of the degree of disaggregation as the dependent variable. We find that the analyst forecast bias is negatively associated with the degree of disaggregation after controlling for analysts’ industry coverage using a high-tech dummy variable. In short, these tests suggest that although the decision to disaggregate is related to industry coverage, our finding that analysts avoid issuing disaggregated forecasts when their forecasts are optimistically biased is robust.

Table 7. SENSITIVITY TEST 1: EXAMINATION OF ANALYST DISAGGREGATION BEHAVIOUR BY TECHNOLOGY VS. NON-TECHNOLOGY INDUSTRY GROUPINGS

Presented below are the results of analyst disaggregation behaviour across Technology and Non-technology industry groupings. Panel A presents percent of analysts belonging to each industry grouping, by analyst category. Panel B presents results from the regression of BIAS on analysts, firm and industry control variables and levels of analyst forecast disaggregation, for the two industry groupings. BIAS is defined as (forecasted one-year ahead EPS—actual EPS) ÷|actual EPS|.
Panel A: Industry compostion of firms by analyst disaggregation categories
Analyst category	Percent technology	Percent non-technology	% of firms with no GICS code	Number of analyst-firms
ND	13%	61%	27%	30,409
PD1	16%	64%	21%	53,593
PD2	36%	47%	17%	184,247
FD	48%	41%	11%	22,651

Panel B: Regression results for the technology and the non-technology industry groupings
()
	Predicted sign	Dependent variable: BIAS
		Technology firms		Non-technology firms
		Coefficient	p-value	Coefficient	p-value
Intercept		0.0030	(0.01)	0.050	(0.01)
LEVEL_D	−	−0.0060	(0.01)	−0.004	(0.01)
TIMELINESS	−	−0.0004	(0.01)	−0.000	(0.01)
LOSS	+	1.4300	(0.01)	1.460	(0.01)
IFFICULTY	+	−0.0240	(0.01)	−0.000	(0.77)
SIZE	−	−0.0040	(0.01)	−0.007	(0.01)
B/M	+	0.0070	(0.01)	0.007	(0.01)
TENURE	?	−0.0007	(0.07)	−0.002	(0.42)
INDUSTRY EFFECT		Yes		Yes
YEAR EFFECT		Yes		Yes
Adj R2		19.91%		25.09%
N		79,387		162,947

• LEVEL_D is a single level of disaggregation variable, which takes a value of 1, 2, 3 or 4 for analysts belonging to ND, PD1, PD2 or FD categories, respectively. Analyst disaggregation categories are (1) ND representing analysts who never issued disaggregated forecasts for any firm they followed during the entire sample period, (2) PD1 representing analysts who issued disaggregated forecasts for some, but not all, of the firms they followed each year and analysts who provided no sales forecast for some of the sample years but did issue sales forecasts for some of their firms in the other sample years, (3) PD2 representing analysts who issued disaggregated forecasts for all of the firms they followed for some, but not all, of the sample years, and (4) FD representing analysts who issued disaggregated forecasts for all the firms they followed for all the sample years. TIMELINESS is the number of months prior to the actual earnings announcement date that the earnings forecast is issued. LOSS is a dummy variable equal to 1 if actual EPS as reported on I/B/E/S for the fiscal year is negative. DIFFICULTY is the standard deviation of the most recent EPS₁ forecasts made by all analysts for that firm in a given year, normalized by the mean of these forecasts. SIZE is measured by the log of market value of the firm equity at the start of the current fiscal year. B/M is the book value of common equity at the start of the fiscal year divided by market capitalization at the start of the fiscal year. TENURE is measured by the number of years an analyst has appeared in the I/B/E/S database. TECH is a dummy variable, which takes a value of 1 if the firm belongs to Information Technology and Health Care sectors (including 15 high tech industries) of the GICS, and 0 otherwise. INDUSTRY EFFECT represents results from the test of the hypothesis on the equality of coefficients for the dummy variables constructed on the basis of the second level of GICS industry classification. YEAR EFFECT represents results from the test of the hypothesis on the equality of coefficients for the year dummy variables.

Sensitivity Test 2

The number of years that an individual analyst contributes earnings forecasts to the I/B/E/S database in our sample varies from one year to a longer tenure. When we classified analysts into four groups of disaggregation behaviour (ND, PD1, PD2, and FD), we included only those analysts in the sample who had contributed forecasts to I/B/E/S for at least three of the 11 years examined. To ensure that the inclusion of short tenure (transitory) analysts in our sample does not create any systematic bias, we re-examined BIAS and FORECAST ERROR by re-defining the four disaggregation categories (ND, PD1, PD2, and FD) using more restrictive criteria of including only those analysts who had provided at least (a) four, (b) five or (c) six years of earnings forecasts. The univariate results under these three alternative definitions are presented in Panels A–C of Table 8. As we move from the most relaxed to the most restrictive criterion of analyst inclusion, the number of analysts decreases from 4,959 (Panel A), to 3,622 (Panel B) to 2,638 (Panel C). Because drastic reductions in sample sizes can make it difficult to reliably estimate multivariate models, we tested each set in an essentially identical manner to those reported in Panel A of Table 2 for the BIAS and FORECAST ERROR variables. For the BIAS variable, the predicted and the actual ranking based on Duncan's multiple range test are identical for the samples with four and five years of analysts’ forecasts as the inclusion criterion. For the six year analysts’ forecast inclusion criterion, the rankings of BIAS of the PD1 and PD2 analysts are statistically lower than the BIAS of the ND analysts, but are not significantly different from one another. However, the relative ranking of the two extreme categories (ND and PD) is the same as in the prior tests.

With respect to FORECAST ERROR, results presented in Panel A of Table 8 show that for the samples with four years of analysts’ forecasts as the inclusion criterion, ND category analysts issued the least accurate forecasts. PD1 and PD2 category analysts were more accurate than ND category analysts but were indistinguishable from each other in forecast accuracy, and the FD category analysts were significantly more accurate than the analysts belonging to the other three categories. When we examined the samples that include only those analysts who provided at least five (Panel B) or six (Panel C) years of forecasts, differences in FORECAST ERROR remain only for the analysts belonging to the two extreme categories (ND and FD). The difference in forecast accuracy across all four analyst categories is no longer detectable. This is possibly because of decreased variance in the forecasting experience of the smaller number of remaining analysts.

Sensitivity Test 3

In this study, we labelled analysts’ forecasts for a firm as disaggregated in a particular year if the analyst issued at least one sales forecast for that firm during that year. Because one forecast is an arbitrary absolute measure and does not take into consideration the total forecasts issued by the analysts, we also used an alternative measure to re-define analysts’ level of forecast disaggregation. In this section, we test the sensitivity of our findings to this alternative measure of disaggregation. Specifically, for each analyst-firm-year observation, we computed the ratio of the number of sales forecasts issued by the analyst in that year divided by the number of EPS forecasts issued by the analyst in the same year. If the computed ratio was more than 0.5, we considered that analyst to have issued disaggregated forecasts for the firm in that year; if the ratio was equal to zero, we considered that analyst to have not issued disaggregated forecasts for the firm in that year. If the ratio was between 0 and 0.5, the corresponding analyst-firm-year observation was excluded from the analysis. We estimated the following two regression models to test the sensitivity of our original sample findings to this stricter definition of disaggregation:

()

()

where ALT_DISAGGREGATE is a dummy variable which takes a value of 1 if the forecast is disaggregated and 0 if the forecast is not disaggregated, on the basis of our alternative definition of disaggregation. All other variables are the same as previously defined.

The results from the estimation of (4a) and (4b) are presented in Table 9. In both of the estimations, the results are qualitatively similar to those reported earlier. Specifically, the signs of all the variables are as expected and significant at conventional levels of significance (0.05 or better). The only exception is DIFFICULTY in model (4b) with FORECAST ERROR as the dependent variable. However this control variable was not significant in our earlier estimation (reported in Table 3) either. The estimated coefficient of ALT_DISAGGREGATE is negative and significant in both of the regression models, consistent with our hypothesis. These results provide corroborating evidence to our earlier finding that the act of issuing disaggregated forecasts is associated with less biased and more accurate earnings forecasts. We considered even more restricted definitions of disaggregation in which the analyst was considered to have issued a disaggregated forecast for the firm in a particular year only if the ratio of sales forecasts to EPS forecasts was greater than 0.5. The results (not reported here) were qualitatively similar.

Table 9. SENSITIVITY TEST 3: MULTIVARIATE TEST OF ALTERNATIVE DEFINITION OF DISAGGREGATION

Presented below are the results from the regression of BIAS and FORECAST ERROR on analysts, firm and industry control variables and the alternative definition of analyst forecast disaggregation. BIAS is defined as (forecasted one-year ahead EPS – actual EPS) ÷|actual EPS| and the FORECAST ERROR is defined as |forecasted one-year ahead EPS – actual EPS|÷|actual EPS|. ALT_DISAGGREGATE is a dummy variable that takes on a value of 1 if the ratio of number of sales forecasts to the total number of EPS forecasts of an analyst exceeds 50% in a given firm-year.
()
()
	Predicted sign	Dependent variable
		BIAS		FORECAST ERROR
		Coefficient	p-value	Coefficient	p-value
Intercept		0.056	(0.01)	0.253	(0.01)
ALT_DISAGGREGATE	−	−0.010	(0.01)	−0.009	(0.01)
TIMELINESS	−	−0.010	(0.01)	−0.015	(0.01)
LOSS	+	1.440	(0.01)	1.339	(0.01)
TECH	−	−0.017	(0.01)	−0.013	(0.01)
DIFFICULTY	+	−0.000	(0.01)	0.000	(0.91)
SIZE	−	−0.007	(0.01)	−0.019	(0.01)
B/M	+	0.008	(0.01)	0.005	(0.01)
TENURE	?	−0.001	(0.01)	−0.008	(0.01)
INDUSTRY EFECT		Yes		Yes
YEAR EFFECT		Yes		Yes
Adj R2		22.99%		27.36%
N		244,645		244,645

• TIMELINESS is the number of months prior to the actual earnings announcement date that the earnings forecast is issued. LOSS is a dummy variable that equals 1 if the actual EPS for the fiscal year is negative, and 0 otherwise. DIFFICULTY is the standard deviation of the most recent EPS forecasts made by all analysts for that firm in a given year, normalized by the mean of these forecasts. SIZE is the log of market value of the firm equity at the start of the current fiscal year. B/M is the book value of common equity at the start of the fiscal year divided by market capitalization at the start of the fiscal year. TENURE is the number of years an analyst has appeared in the I/B/E/S database. TECH is a dummy variable, which takes on a value of 1 if the firm belongs to Information Technology and Health Care sectors (including 15 high tech industries) of the GICS industry classification, and 0 otherwise. INDUSTRY EFFECT represents results from the test of the hypothesis on the equality of coefficients for the dummy variables constructed on the basis of the second level of GICS industry classification. YEAR EFFECT represents results from the test of the hypothesis on the equality of coefficients for the year dummy variables.