The Relative Accuracy of Analysts Disaggregated Forecasts: Identifying the Source of. Analysts Superiority

Transcription

1 The Relative Accuracy of Analysts Disaggregated Forecasts: Identifying the Source of Analysts Superiority Mark T. Bradshaw Boston College 140 Commonwealth Avenue Fulton 520 Chestnut Hill, MA Marlene Plumlee^ Department of Accounting David Eccles School of Business University of Utah 1645 E. Campus Center Dr. Salt Lake City, Utah Teri Lombardi Yohn Department of Accounting Kelley School of Business Indiana University 1309 East Tenth Street Bloomington, Indiana August 2011 FIRST DRAFT. Please do not circulate or quote without author s permission!! ^ Corresponding author: marlene.plumlee@utah.edu! "!

2 The Relative Accuracy of Analysts Disaggregated Forecasts: Identifying the Source of Analysts Superiority Abstract We examine how the superiority of analysts forecasts of disaggregated earnings relative to random walk (RW) forecasts of the same items contribute to analysts relative superiority over RW earnings forecasts. While prior research frequently finds that analysts earnings forecasts are more accurate than RW earnings forecasts, the source of this superiority (or lack thereof) is not well understood. Our findings suggest that analysts contributions to the forecasting process lies in their superior ability to forecast operating profit and non-operating items. We also find, however, that analyst forecast superiority in forecasting sales or operating expenses does not, on average, translate into superiority in forecasting operating profit or net income. Consistent with the economic importance of operating profit, we find that analysts operating profit forecast superiority is significantly associated with analyst net income forecast superiority. We also find that the analysts superiority across components of income varies by firm/year characteristics and across forecast horizons, consistent with analysts providing differential benefits. For example, longer-run analyst operating profit forecasts are superior to RW forecasts only for firms in the highest historical sales growth tercile and the largest size tercile. The results suggest that (1) analysts contribution to the forecasting process varies across firms, type of forecast, and forecast horizon and (2) analysts could improve their superiority over RW forecasts if they were able to translate accurate sales forecasts into accurate operating profit forecasts, especially for smaller, lower growth firms. Keywords: analyst forecasts; disaggregated forecasts; random walk; analysts superiority JEL Code: G17, M40, M41! #!

3 The Relative Superiority of Analysts and Time-series Disaggregated Forecasts 1 Introduction A tremendous amount of academic research has examined sell-side analysts forecasts of accounting earnings. 1 A long-held view about these forecasts is that they are more accurate than forecasts based on time-series models (Fried and Givoly 1982). However, several recent studies re-characterize this view as an overgeneralization, documenting that analysts superiority over even simple time-series forecasts is limited (e.g., Gao and Wu 2010; Allee 2010; Jung et al. 2011; Bradshaw et al. 2013). Time-series forecasts are relatively costless to implement, so the effort devoted to forecasting earnings along with market resources devoted to the closely watched earnings announcement season suggest that analysts must provide some value through their research activities. In this study, our ultimate interest is when and how analysts costly research efforts enhance market efficiency through more accurate expectations of future earnings. We address this issue by examining analysts disaggregated earnings component forecasts. Prior research documents that analysts superiority over time-series forecasts is due to both a timing advantage and an informational advantage (Brown et al. 1987b). Analysts have a timing advantage because the information in time-series forecasts is restricted to the period ended with the most recent earnings report, whereas analysts enjoy the advantage of a reduced forecast horizon with each day beyond the date of the most recent earnings report. 2 Also, analysts have an informational advantage because the information set available to them includes!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1 See reviews of this literature by Givoly and Lakonishok (1984), Schipper (1991), Brown (1993) including discussions by O Hanlon (1993), Thomas (1993), Brown (1993) and Zmijewski (1993) - and Ramnath, Rock and Shane (2008). #!We attempt to minimize the timing advantage by comparing analysts forecasts released contemporaneously with the announcement of earnings (i.e., the random walk forecast).! $!

4 everything impounded in the historical earnings series, plus other information yet to be reflected in earnings, such as accounting accruals or other information that predated the most recent earnings release or information arriving subsequent to that earnings release. We operationalize our analysis by focusing on the process by which analysts derive earnings forecasts, and investigate the contribution of disaggregated components of earnings - sales, operating expenses, and nonoperating expenses to the accuracy of the earnings forecasts. Lipe (1986) documents that disaggregated components of earnings provides incremental explanatory power for security returns relative to that provided by bottom-line earnings alone. Studies contemporaneous with Lipe (1986), however, conclude that disaggregation of earnings or inclusion of supplemental disclosures in analyses did not provide incremental information content for security prices. Nonetheless, subsequent research reinforces the findings in Lipe (1986) by documenting that profitability forecasts are improved by using disaggregated earnings components (Fairfield et al. 1996) and showing differential stock price reactions to revenue and expense surprises (e.g., Swaminathan and Weintrop 1991; Ertimur et al. 2003). We rely on this evidence to motivate a prediction that the accuracy of explicit forecasts of earnings components is likely associated with the accuracy of bottom-line earnings. Our study explores three questions. First, is the mere existence of disaggregated forecasts provided by analysts associated with more accurate earnings forecasts? Financial statement analysis texts typically describe the earnings forecast process as beginning with a sales forecasts, followed by expense forecasts that algebraically yield an earnings forecast (e.g., Lundholm and Sloan 2007; Penman 2009; Wahlen et al. 2011). However, this detailed analysis could actually be unnecessary for providing earnings forecasts because analysts could simply parrot forecasts of other analysts or the managers, or they may have little incentives to devote to such efforts! %!

5 (Williams 1996; Groysberg et al. 2010; Bradshaw 2011). Second, conditional on providing disaggregated forecasts, what forecasted components of earnings are most important to bottom-line earnings forecast accuracy? Algebraically, an error in the forecast of any component will directly impact the bottom-line earnings forecast. However, operating expenses are frequently driven by the level of sales, which implies that an inaccurate forecast of sales will trigger errors in expense forecasts. For example, cost of goods sold is routinely one of the largest expenses for a firm, and it is driven largely by sales volume. On the other hand, cost stickiness varies depending upon a firm s capacity utilization (Banker and Chen 2006). Baumgarten et al. (2011) examine the association between sales changes and analysts implicit forecasts of total expenses, and conclude that analysts are overoptimistic about costs decreasing with sales declines. We build on their results by examining operating and nonoperating expenses separately, and focus on cross-sectional variation in the association between disaggregated forecasts and earnings forecast accuracy. Finally, as alluded to above, we explore what factors are associated with analyst superiority over time-series forecasts. Bradshaw et al. (2013) find that random walk (RW) earnings forecasts outperform analysts earnings forecasts over longer horizons and for firms that are smaller, younger, and when analysts forecast negative or large changes in earnings. Similarly, we examine the associations among disaggregated forecasts and analysts superiority over RW forecasts over a number of factors, motivated both by prior research and our specific disaggregated forecast setting. What is of great interest to capital market research is the black box in which analysts derive their forecasts (Bradshaw 2011). However, empirical archival approaches to understanding the forecasting process are generally limited to examining correlations between! &!

6 hypothesized inputs into the forecasting exercise (e.g., information in past earnings, prices, and other information) and the outputs of that process (i.e., the earnings forecast and subsequent realized earnings). We are similarly limited in our ability to peer inside analysts actual thought processes and activities. However, we examine analysts disaggregated forecasts, which provide some indication of the actual process through which analysts derive bottom-line earnings forecasts. It is straightforward to expect that analysts with more accurate sales or expense forecasts ought to produce more accurate earnings forecasts. By analyzing such disaggregated forecasts, we obtain an insight, albeit indirect, into the use of internally generated inputs into the ultimate earnings forecast. We examine disaggregated forecast accuracy for both analysts and RW forecasts. Our analyses include forecasts made across various horizons, spanning from the current year (FY t+1 ) and the following two years (FY t+2 and FY t+3, respectively). We also examine whether analyst disaggregated forecast superiority varies by firm attributes. Consistent with prior research, we compute signed forecast errors as the difference between the actual disaggregated component and the corresponding forecast of the analysts or RW model. Although we provide some descriptive evidence on forecast errors, our primary analyses are based on analysts forecast accuracy, which is the unsigned forecast error. We quantify analysts forecast superiority (AFS) for each of disaggregated component and for total earnings. AFS is a direct measure of the forecast improvement, and hence analyst superiority, for each disaggregated item relative to a simple RW extrapolation. Inconsistent with our expectations, analysts sales forecasts are not uniformly more accurate that their forecasts of operating expenses across all three forecast horizons. In fact, while the unsigned FY t+1 analysts sales forecast error is not different from zero and the! '!

7 corresponding operating expense forecast error is statistically positive, in all other settings the analysts sales forecasts are less accurate than their operating expense forecasts. Analysts nonoperating item forecasts are more accurate than their forecasts of operating profit (which is the net of sales less operating expense forecasts). This difference in magnitude between analysts forecasts of operating profit and non-operating items likely is due to the difference in the average magnitude of operating profit and non-operating items. Consistent with our expectations, analysts forecast errors increase across forecast horizons across all the disaggregated components. We also document that, while analysts net income forecasts are more accurate than RW forecasts for FY t+1, this is not the case for FY t+2 or FY t+3. Furthermore, analysts sales and operating expense forecasts are more accurate than RW forecasts across all three forecast horizons, although the same cannot be said for analysts operating profit forecasts the cumulation of the sales and operating expense forecasts. In fact, when examining the cumulative forecasts (operating profit and net income forecasts), we find that over longer forecast periods RW forecasts consistently dominate analysts forecasts. We also find differences in the accuracy and relative accuracy of analysts disaggregated forecasts, based on firms types. Our study links previous findings on analyst superiority over time-series forecasts and on the incremental information contained in earnings components. We show how the differential accuracy of disaggregated forecasts is related to earnings forecast accuracy, which enhances our understanding of how analysts actually contribute value through their costly forecasting activities. The study proceeds as follows. The next section review relevant prior literature. We describe our data and formalize hypotheses in section 3. Primary tests appear in section 4, and! (!

8 section 5 concludes. 2 Prior research and motivation There are two primary streams of research that relate most directly to our study. The first includes studies that compare analysts forecast accuracy to time-series forecasts. The second includes studies that explore the incremental information content of disaggregated income statements over summary measures of earnings. 2.1 Analysts forecasts versus time-series models Many of the earlier studies that compare the accuracy of analysts earnings forecasts to RW forecasts typically document that analysts earnings forecasts are more accurate (e.g. Brown et al. 1987a, Brown et al. 1987b). The earliest studies provide evidence that annual earnings approximate a simple RW process (e.g., Little 1962; Ball and Watts 1972). Based on these studies, Brown (1993, 295) concludes that earnings follow a RW, something researchers have been in agreement about for decades. Subsequent research explores the explanation for why analysts forecasts are superior to time-series forecasts. The overall notion is that, because analysts have access to a broader information set including non-accounting information as well as information released after the prior fiscal year, it would be hard to understand how they would not be superior to time-series forecasts. Brown et al. (1987b) articulate this overall information advantage into two sources: an informational advantage (analysts better utilize information available on the date on which the RW forecast is made) and a timing advantage (analysts utilize information acquired between the date on which the RW forecast is made and the date on which the analysts forecast is made).! )!

9 Indeed, research finds that analysts forecast accuracy is negatively associated with forecast horizon (Kross et al. 1990; Lys and Soo 1995). O Brien (1988) argues that analysts superiority is explained by analysts use of time-series models and a broader information base, including information about industry and firm sales and production, general macroeconomic information, and other analysts forecasts. Consistent with the information environment argument, Kross et al. (1990) document a positive association between analysts forecast superiority and firm coverage in the Wall Street Journal. More recent research, however, (e.g. Bradshaw et al. 2013) employs a much broader sample of firms than typically included in the prior studies and considers three forecast horizons. They document that, contrary to the broadly held belief in the superior ability of analysts to forecast earnings, a RW model dominates analysts forecasts of earnings over longer horizons and for certain types of firms. This recent evidence motivates developing a deeper understanding of how analysts improve upon baseline forecasts that are as cheaply available as random walk forecasts. 2.2 Disaggregated earnings components The second stream of related research employs the disaggregated forecast data used in our study, although the research questions addressed differ from ours. For example, Ertimur et al. (2003) investigate market reactions to revenue and expense surprises, relying on I/B/E/S sales forecasts to estimate the sign and magnitude of the revenue/expenses surprises. 3 Specifically, the authors employ I/B/E/S analysts sales and net income forecasts to explore whether parsing net!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3 Ertimur et al. (2003) rely on sales and net income forecasts available from I/B/E/S to calculate their revenue and expense surprises. Specifically, revenue surprise is the difference between the analysts forecasts of revenues and Compustat reported revenues. Expenses (forecasted and actual) are calculated as the difference between revenue and income before extraordinary items (forecasted and actual), rather than relying on a direct forecast of expenses. Thus, expense surprise is the difference between the revenue surprise and the income before extraordinary items surprise.! *!

10 income into revenues and expenses provides additional information beyond net income alone. They document a differential market reaction to net income surprises due to revenue surprises relative to net income surprises due to expense surprises. They also find that investor reaction to revenue surprises is stronger for growth than for value firms. In related research, Ertimur et al. (2009) study the role of reputation in explaining analysts decisions to provide revenue and expense forecasts to I/B/E/S. The authors argue that, while all analysts privately produce these forecasts, not all of them make those forecasts available to I/B/E/S. Thus, the presence of these forecasts might be explained by analysts incentives. They document an inverse relation between analysts reputations and the decision to dissemination disaggregated forecasts. Based on their findings, the authors suggest that the issuance of disaggregated forecasts by analysts is a mechanism to facilitate reputation. In related research, Call et al. (2009) find that analysts earnings forecasts issued with cash flow forecasts are more accurate than those issued without cash flow forecasts. Similarly, Lee and Peterson (2011) show that revenue forecast accuracy and earnings forecast accuracy are positively related. Moreover, they confirm prior results on the negative association between forecast horizon and forecast accuracy, and show that analysts revenue forecast accuracy improves throughout the fiscal year. Overall, these studies suggest that the disaggregated forecasts are both informative in their own right, contribute to overall forecast accuracy, and might signal the quality of analysts research supporting earnings forecasts. 2.3 Motivation Our research questions are most similar to those addressed in the general literature on forecast accuracy, although our analyses rely on disaggregated forecasts available from I/B/E/S.! "+!

11 Specifically, we explore the differential accuracy of disaggregated forecasts, using RW disaggregated forecasts as a baseline expectation. Our interest lies in better understanding how and under what conditions analysts contribute value to the forecasting process. Most of the prior literature examines when and if analysts forecasts are superior to RW forecasts and focuses on aggregate earnings forecasts. In contrast, our goal is to better understand how analysts contribute to improved earnings forecasts. Specifically, we examine the increase in accuracy of disaggregated analysts forecasts relative to RW forecasts (analysts improvement) and consider the differential analysts improvement across disaggregated forecasts (e.g. sales versus expense forecasts). We also explore whether the improvement (in accuracy) of employing analysts sales, operating expenses, or non-operating expenses forecasts relative to RW forecasts provide the greatest benefit when forecasting the (aggregated) net income forecasts. We aim to enhance our understanding of the process by which analysts input improves forecasts of earnings. Relatedly, we also provide evidence on the settings where analysts provide the greatest improvement relative to RW forecasts. Our reasons for and expectations regarding the examination of disaggregated earnings forecast accuracy parallel the conclusions from prior research about why analysts are generally superior to time-series models. First, with regards to an information advantage, analysts superiority likely varies across disaggregated earnings components. Analysts may focus on specific items (e.g. sales) in the forecasting process and, consequently, spend less time on other items (e.g. expenses). Further, when analysts have relatively less specific information related to a component of earnings, their forecasts might default to historical reported values (i.e., a random walk). If descriptive of the forecasting process, we expect that disaggregated forecasts where analysts have an information advantage will deviate more from RW forecasts. Our primary focus! ""!

12 is thus on analysts information advantage as evidenced by our measure of analyst forecast superiority. Second, the differential persistence of disaggregated components necessarily affects analysts forecast superiority. Clearly, RW forecasts will perform better for earnings components that exhibit high persistence rather than low persistence. Evidence in Baumgarten et al. (2011) indicates that costs are more persistent than revenues. This suggests that analysts forecast superiority over time-series forecasts of operating expenses (and hence operating profits and bottom-line earnings) will be less than that for sales. Our analysis of RW disaggregated forecasts provides a better understanding of this relation. There are some caveats to our analysis. First, we ignore well-documented cross-sectional variation in analysts incentives. Although most research on analysts incentives focuses on individual analysts (e.g., O Brien et al. 2005), our analyses utilize the consensus, which filters some of the cross-sectional variation in individual analyst incentives, although incentives driven by firm-level factors (like growth and external financing) remain. We attempt to control for these in multivariate analyses. Second, we intentionally restrict our analysis to earnings forecasting, and are aware that this is only one of the many roles served by analysts. In addition to earnings forecasts, analysts provide price targets and stock recommendations, coordinate conferences that network institutional investors and managers, and provide numerous other client-related services. We believe that a better understanding of the overall value provided by analysts is rooted in the fundamental role of earnings forecasting, but acknowledge that other factors may be more important and may interact with the forecasting activities we examine. 2.4 Summary! "#!

13 We expect analysts superiority to decrease as the forecast horizon increases (Brown et al. 1987a, Bradshaw et al. 2011), consistent with prior literature. Thus, we compare the accuracy of analysts disaggregated forecasts to that of time-series forecasts, based on annual realizations over three forecast horizons (FY t+1, FY t+2, and FY t+3 ). Consistent with much of the prior research, we use RW forecasts as our time-series benchmark. More sophisticated time-series models of annual disaggregated values would impose data restrictions that would limit our sample and, more importantly, there is very little evidence that more sophisticated models produce earnings forecasts that more accurate than a RW model (Albrecht et al. 1977; Watts and Leftwich 1977; Brown et al. 1987a). We know of no evidence regarding the accuracy of disaggregated components across multiple time-series models. Thus, we default to the RW specification that is well-established in the literature for summary earnings. Furthermore, we expect the accuracy of the disaggregated values to vary by component. As noted previously, we expect analysts sales forecasts to be more accurate than their operating expense and nonoperating expense forecasts due to the relatively higher persistence of expenses (Baumgarten et al. 2011). Finally, we supplement our primary analysis of analyst forecast superiority across disaggregated components by examining cross-sectional variation based on factors such as firm size, sales growth, and disaggregated component forecast intensity, discussed in more detail below. 3 Data and sample 3.1 Data Our sample construction begins with the universe of I/B/E/S disaggregated annual! "$!

14 forecasts for the years 2003 through 2009, including realizations of each item. 4 We begin with 2003, as this is the first year in which I/B/E/S provides disaggregated forecasts and actual values in large numbers. We form three samples for each forecast horizons, FY t+1, FY t+2, and FY t+3. For each disaggregated forecast, we retain the consensus I/B/E/S forecast and the corresponding actual value to compute analyst forecasts errors. We employ Compustat values to serve as the RW forecast and compute RW forecast errors. To avoid the influence of outliers, we eliminate the top and bottom one percent of RW and analyst earnings forecast errors. These procedures result in 13,323, 10,507, and 5,115 firm-years for the FY t+1, FY t+2, and FY t+3 forecast horizons, respectively. We calculate RW and analyst forecasts errors as the actual realization less the forecast, and scale forecast errors by lagged total assets. RWFE_DI t+1 = CompustatActual t+1 - CompustatActual t (1) AFE_DI t+1 = I/B/E/S Actual t+1 I/B/E/S Forecast t+1 (2) RWFE refers to the RW forecast error, and AFE refers to the analyst forecast error. DI represents the three disaggregated forecasts provided by I/B/E/S sales (SALE), operating profit (OPPROF), and net income (NI) forecasts. All forecast errors are scaled by lagged total assets. Note that random walk forecast errors use the most recent realization as the forecast, whereas the I/B/E/S forecast reflects the consensus from I/B/E/S. Forecast errors for forecast horizons t+2 and t+3 are computed similarly, with the exception that the random walk forecasts use the most recent actual values as the forecast for all three horizons (i.e., Actual t ). 5 In addition to these explicit forecasts, we generate implicit forecasts for operating!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! %!,-./-! !50676!892-92:0;;!2:<!=042>?!#++$@!A.:<1!B42B!B40!901B!C2B34!A/-!32;3D;2B.:7!EFGFHFI!2:2;?1B1J! 02-:.:71!A/-0321B!0--/-1!.1!B40!23BD2;!>2;D0!-0K/-B0<!9?!EFGFHFI6!81!L.B4!02-:.:71!A/-0321B1M!EFGFHFI!-0K/-B1! 23BD2;!>2;D01!A/-!B40!< B0<!A/-0321B1M!L4.34!L0!D10!B/!32;3D;2B0!B40!2:2;?1B!A/-0321B!0--/-16!! &!For FY t+2 (FY t+3 ), the RW forecast is the realized value (e.g. sales) two (three) years prior to the targeted forecast year. Thus, the RW forecast error is the change in sales over the forecasted year(s).!! "%!

15 expenses (OPEXP), defined as SALE OPPROF, and non-operating expenses (NONOP), defined as OPPROF NI. OPEXP and NONOP are comparable to the (total) EXPENSES variable employed in Ertimur et al. (2003), who employ analysts revenue forecasts and calculate EXPENSES as the difference between the revenue and net income forecasts. Including OPEXP and NONOP allows us to parse the analysts forecasts of NI into three distinct items sales, operating expenses and non-operating expenses. Consistent with prior literature, positive (negative) forecast errors indicate that the forecast underestimates (overestimates) the actual value. Our primary focus in this study is accuracy, however, which is the absolute value of the forecast error. Moreover, we quantify the relative forecast accuracy of analysts and RW forecasts as analyst forecast superiority (AFS). AFS is the difference between the absolute value of the RW forecast accuracy and the analyst forecast accuracy. Thus, positive (negative) values of AFS indicate that the analyst forecast is more (less) accurate than the RW forecast. 4 Results 4.1 Descriptive statistics Table1 provides descriptive statistics. Panel A presents the number of firm-year observations for our primary sample of disaggregated analyst forecasts for each year and forecast horizon. There is a general increase in the number of disaggregated analyst forecasts over the sample period. In addition, there are more FY t+1 forecasts than FY t+2, and more FY t+2 forecasts than FY t+3 forecasts each year, consistent with prior findings on the decreasing availability of longer horizon forecasts. Panel B presents descriptive statistics on the firm-year characteristics of our sample, benchmarked against the I/B/E/S population. Firm-years for which disaggregated! "&!

16 analysts forecasts are available reflect smaller firms (in terms of total assets) but higher sales and greater profitability. The lower part of panel B reveals that the firm-years with disaggregated analyst forecasts have higher sales and profits than the I/B/E/S population, but also higher levels of expenses. These relations hold in general across all forecast horizons. [Please place Table 1 here] 4.2 Comparative accuracy of aggregated and disaggregated forecasts Our first question is whether the existence of disaggregated forecasts results in bottomline earnings forecasts that are more accurate, presumably due to the signal provided by explicit forecasts of disaggregated earnings components. Table 2 reports both the signed and absolute value of the analysts earnings per share (EPS) forecast errors for the I/B/E/S population and the subsample for which disaggregated forecasts are provided. The overall message from table 2 is that there is weak evidence that the provision of disaggregated forecasts results in higher forecast accuracy (i.e., lower absolute forecast errors), with supporting evidence concentrated in median tests. There are no significant differences between the mean analyst EPS forecast errors (AFE_EPS) across the full and restricted samples for FY t+1 and FY t+2 forecasts, but analysts FY t+3 EPS forecasts are more optimistically biased when disaggregated forecasts are provided than when they are not. Across all three horizons, the median absolute forecast error (AAFE_EPS) is smaller when analysts provide disaggregated forecasts than when they do not. These findings are generally consistent with Call et.al. (2009), who find that analysts earnings forecasts issued together with cash flow forecasts are more accurate than those that are not. Looking ahead to the results discussed later, when we compare the analysts forecast superiority over the random walk forecasts across the full and restricted samples (AFS_EPS), however, we! "'!

17 fail to document significant differences across the full and restricted samples for any forecast horizon. Our findings related to the accuracy and analysts forecast superiority could potentially be due to greater difficulty in forecasting firm-years in which disaggregated forecasts are provided. To provide insight into this possibility, we report the signed and the absolute value of the RW EPS forecast errors for the full and restricted samples. We find no evidence of differences in bias across the two groups, although we do document that, at the median for all forecast horizons, the restricted RW EPS forecasts are significantly less accurate than the full sample and that the mean absolute RW forecast error is significantly smaller for the disaggregated analyst forecast sample. While the results are not clear, the lack of improvement in EPS forecast accuracy for the restricted sample might reflect greater difficulty in forecasting those firm-years. In multivariate analyses later in the paper, we control for factors likely associated with more forecast difficulty (e.g., RW earnings forecast accuracy, size, etc.). [Please place Table 2 here] 4.3 Analyst forecast superiority across income components We next provide insight into our second research question. Conditional on analysts providing disaggregated forecasts, we examine the superiority of the analyst forecast relative to the RW forecast for each income component. This analysis is intended to provide insight into whether any specific income components appear most associated with ultimate earnings forecast accuracy. Table 3 reports the forecast accuracy and bias of the RW and analyst forecasts for those firm/years in which analysts provide disaggregated forecasts for each component of forecasted income.! "(!

18 Panel A documents that analyst forecasts of each component for FY t+1 are significantly more accurate than the analogous RW forecasts. This suggests that analyst forecasts of each income component are superior to RW forecasts. Interestingly, analysts appear to have the greatest superiority over RW forecasts when forecasting operating expenses and sales; however, they have the least superiority over RW forecasts with respect to operating profit. This is consistent with analysts providing value in forecasting sales and operating expenses but perhaps not fully understanding the interrelation between sales and operating expenses to translate those component forecasts into more accurate forecasts of operating profit. Analyst superiority at forecasting net income for FY t+1 is, therefore, driven by their superiority in forecasting nonoperating income and not by their superiority in forecasting operating profit. Analyst forecasts of sales, operating expenses, non-operating expenses, and net income for FY t+2 and FY t+3 are also superior to RW forecasts. In addition, the analyst forecast superiority for sales and operating expenses appears to increase over the longer horizons. However, we find that analyst forecasts of operating profit for FY t+2 and FY t+3 are significantly less accurate than RW forecasts at the mean and are not different for FY t+3 at the median. This is consistent with the results for FY t+1, and suggests that while analysts are more accurate at forecasting sales and operating expenses two- and three-years ahead, they are not more accurate at forecasting the difference between sales and expenses (operating profit). To provide some insight into the source of the accuracy differences in Panel A, Panel B of Table 3 reports the RW and analyst forecast bias for each of the income components. Analyst forecast superiority could be driven by less optimistically biased analysts forecasts and/or pessimistically biased RW forecasts, or some other combination. Given that the RW forecasts include no growth component, it is not surprising that the RW forecasts consistently! ")!

19 underestimate each component of income. The results for analyst forecasts yield some insight into the accuracy results in Panel A. Specifically, for FY t+1, analyst forecasts of sales are unbiased; however, analysts appear to underestimate operating expenses leading to optimistically biased forecasts of operating profit. This is consistent with analysts not fully considering the effect of forecasted sales on the variable and fixed components of operating expenses. For FY t+2 and FY t+3, analysts appear to overestimate sales and underestimate operating expenses (only at the median for FY t+2 ). This suggests that, for longer horizon forecasts, analysts appear to exhibit an optimistic bias for sales forecasts but continue to underestimate operating expenses. Again, this is consistent with analysts not fully incorporating cost structures into their forecasts, consistent with evidence in Baumgarten et al. (2011). Analysts appear to provide the greatest value in terms of forecast accuracy over RW forecasts with regards to either sales or operating expenses. However, the analysts do not appear to translate these superior forecasts into superior forecasts of operating profit. Analysts also tend to overestimate non-operating income over each forecast horizon. [Please place Table 3 here] To simultaneously examine the association between various disaggregated component forecasts and bottom-line forecast accuracy, Table 4 presents the results of multivariate regression analyses. The dependent variable is analyst forecast superiority over the RW forecasts of net income (AFS_NI). We regress AFS_NI on the rank of the analyst forecast superiority for each individual component of net income. We calculate the scaled decile rank for each of our disaggregated components for use in our analysis, similar to Abarbanell and Bushee (1998). Specifically, we rank the values of the variables into deciles (0 through 9) each year and divide the decile number by nine so that each variable observation takes on a value between zero and! "*!

20 one. This ranking process allows us to compare across the coefficients within a model. Model 1 includes all components of income. However, because AFS_SALE is highly correlated with AFS_OPEXP (0.84 for FY t+1, 0.91 for FY t+2, and FY t+3, untabulated), we also report the results of Model 2, which excludes AFS_OPEXP from the regression. The resulting ranked variables are indicated with an R prefix on the associated variable name. If the improvement in accuracy obtained from a disaggregated analyst forecast is related to an improvement in accuracy in the analysts net income forecasts, we expect a positive coefficient on the disaggregated forecast variable. We fail to document a significant association between the superiority of analysts sales or operating expense forecasts and analysts net income forecasts across any forecast horizon. In fact, the coefficients on RAFS_SALE and RAFS_OPEXP are either insignificant or significantly negative in each of the regressions. This suggests that, while analyst forecast superiority over RW forecasts is highest for sales and non-operating expenses (as evidenced in Table 3), the superiority of these line item forecasts does not translate into analyst net income forecast superiority. In contrast, we document a consistently positive association between analysts superiority in forecasting operating profit and their superiority at forecasting net income. In short, analyst forecasts of net income are only superior to a RW when analysts are able to combine the accurate forecasts of sales and operating expenses into a superior forecast of operating profit. In fact, the results suggest that providing superior forecasts of sales leads to significantly less superior forecasts of net income, after controlling for improved operating profit forecasts. Therefore, while much focus and effort has been on forecasting sales, the results suggest that translating superior sales forecasts into superior forecasts of operating profit is of great importance. Superior forecasts of operating profit, not superior forecasts of sales, drive! #+!