An Empirical Examination of the Divergence between Managers and Analysts Earnings Forecasts

Transcription

1 An Empirical Examination of the Divergence between Managers and Analysts Earnings Forecasts SOMNATH DAS College of Business Administration University of Illinois at Chicago Chicago, IL KYONGHEE KIM Trulaske College of Business University of Missouri at Columbia Columbia, MO SUKESH PATRO College of Business Northern Illinois University DeKalb, IL This version: November 2012 This paper has benefited from the comments and suggestions of Chih-Ying Chen, Harry Evans, Dan Givoly, Bin Ke, Hennock Louis, Shail Pandit, Ram Ramakrishnan, Amy Sun, Jeff Wong, Jingjing Zhang, Tracey Zhang and seminar participants at McGill University, Northern Illinois University, Pennsylvania State University, Singapore Management University, the University of Illinois at Chicago, the University of Missouri at Columbia, the 2010 Annual Meetings of the American Accounting Association, and the 2011 Annual Meetings of the Financial Management Association.

2 An Empirical Examination of the Divergence between Managers and Analysts Earnings Forecasts ABSTRACT Contrary to the popular emphasis on managers guiding or walking-down analyst expectations, analyst forecasts routinely diverge from managers earnings forecasts. In this paper, we document this regularity and examine how this forecast divergence affects investor response to earnings announcements. Our results suggest that such divergence results in a significant discounting of earnings surprises measured by analyst forecast errors. This discounting is incremental to similar discounting arising from transitory earnings associated with discretionary accruals or extreme earnings. Additional tests support the notion that the information content of analyst forecast-based earnings surprises is lower when forecast divergence is large, and that the discounting is not due to changes in the firm s discount rate arising from earnings uncertainty. This role of forecast divergence at the time of earnings announcements is consistent with its use by investors to adjust their reliance on the consensus analyst forecast as an earnings benchmark.

3 1. Introduction The documentation of the relative superiority of analyst forecasts to time-series forecasts (Fried and Givoly 1982; Brown et al. 1987; Brown and Kim 1991), led capital market researchers to generally accept analyst forecasts of earnings per share (EPS) as the representative proxy for investor expectations of earnings. In recent decades, firms, via the issuance of management earnings guidance, have played an increasingly significant role in influencing both investor expectations and analyst estimates of EPS. 1 In spite of the overlapping roles of analyst forecasts and management earnings guidance, the interplay of the two forecasts in informing the market has received relatively limited attention (Brown and Zhou 2012). Arguments based on the timeliness of analysts forecasts, and evidence that analyst forecasts incorporate the information contained in management forecasts, have provided support for the continuing role of the analyst consensus EPS as the dominant earnings benchmark. However, also implicit in this use of analysts forecasts is the assumption that management earnings forecasts, having exhausted their informational capacity via their impact at issuance on both stock prices and analyst estimates, have no role to play at earnings announcements. In this paper, we examine the validity of this assumption. Specifically, we examine how the divergence between consensus analyst forecasts and prior management forecasts (henceforth, forecast divergence) influences the efficacy of consensus analyst forecasts as benchmarks for investor expectations of earnings. Our examination is motivated, first, by evidence that analysts are not always efficient in incorporating information and that their forecasts demonstrate systematic biases (e.g., Elgars and Lo 1994; Abarbanell and Bushee 1997; Bradshaw, Richardson and Sloan 2001; Lim 2001; Feng and McVay 2010). To the extent that forecast divergence is due to analysts inefficiency in 1 In fact, managers explicitly state that they provide earnings guidance to influence analysts and investors expectation of future earnings (Graham et al. 2005). 1

4 incorporating information in management forecasts and/or investor perception of such inefficiency, market participants are likely to put less weight on the consensus analyst forecast in the presence of forecast divergence when calibrating their earnings expectations. A second, the evidence that investors perception of earnings informativeness is a motivating factor for the issuance of management forecasts (Lennox and Park 2006) suggests that investors are likely to use management forecasts, directly or indirectly, to calibrate their expectations of future earnings. Such reliance by investors on management forecasts is supported by recent studies of managers comparative advantage, relative to analysts, in processing certain types of information pertaining to future earnings (Hutton et al. 2012; Brown and Zhou 2012). The evidence in these studies pointing to the nature and quality of information provided by management forecasts also suggests that divergence of the consensus analyst forecast from the management forecast is likely to reduce investor reliance on the consensus forecast as an earnings benchmark. Our examination is also in part motivated by the prevalence of forecast divergence for the universe of management earnings forecasts in the First Call Historical Database, 69 percent (50 percent) of the last consensus analyst forecasts of quarterly EPS just prior to earnings announcement differs (by over one cent) from the last preceding management earnings forecast. Combined with the above arguments about the relative strengths and weaknesses of the two types of forecasts, the prevalence of forecast divergence suggests that management forecasts may be a viable competing benchmark of earnings expectations. However, we find no support for this alternative to the conventional wisdom in the existing literature. Nevertheless, this does not preclude an information role for management forecasts at the earnings announcements, as investors may calibrate their expectations conditional on the degree of divergence. Further, since the representative investor expectations remain unobservable to researchers, it is a priori unclear 2

5 whether the consensus analyst forecast will be as representative a proxy for investor expectations when it diverges from the management earnings forecasts as when there is convergence between the two forecasts We measure forecast divergence as the absolute value of the stock price-scaled difference between the last management forecast of EPS and the final consensus analyst forecast issued after the last management forecast and before the earnings announcement for the fiscal quarter. We examine the impact of forecast divergence in the period surrounding the earnings announcement and find that the return response to analyst forecast errors is decreasing in the magnitude of divergence and is on average 31% lower in the presence of divergence. This attenuating effect is independent of whether forecast divergence is due to the consensus analyst forecast being higher or lower than the management earnings forecast. An immediate concern is that these results arise from issues with the validity of our construct of forecast divergence. To address this concern, we examine whether factors such as (i) the staleness of management forecast due to its timing, (ii) the imprecision in management forecasts when managers issue guidance in a range versus a point, and (iii) measurement errors in general, 2 contribute to the attenuating effect documented above. In each of these situations, we find that the attenuating effect of forecast divergence persists and is statistically significant, suggesting that the discounting cannot be attributed to common perceptions about the source of divergence. Having confirmed the construct validity of our measure of forecast divergence, we next turn to an investigation of the source of the associated attenuating effect. In doing this, we rely on Kothari s (2001) taxonomy of the determinants of earnings response coefficients. Kothari (2001) suggests changes in earnings response coefficients can arise from noise in the earnings 2 We examine this by re-estimating the effect of forecast divergence using it in non-parametric form. We find that using decile (quintile) ranks does not alter our primary finding. 3

6 benchmark (the consensus analyst forecast) or variations in the quality of reported earnings, i.e., from the numerator of the valuation function. They can also arise from changes in the firm s discount rate, i.e., from the denominator of the valuation function. We begin with reported earnings as a potential competing source for the lowered return response to analyst forecast errors. The existence of transitory components, such as an extreme change in earnings and/or large discretionary accruals, introduces noise into reported earnings, making earnings less useful in predicting future cash flows. This can reduce investors reliance on such earnings and also have an adverse impact on analysts ability to make accurate forecasts. Further, transitory earnings can also contribute to forecast divergence, because such information may not be available to analysts prior to the earnings announcement. These arguments suggest that the moderating effect of forecast divergence could be driven by such transitory components in earnings. Consistent with prior studies, we find that the return response to analyst forecast errors is smaller when earnings contain extreme news or a large amount of discretionary accruals. However, we also find that forecast divergence weakens the return response to analyst forecast errors even when earnings do not contain extreme news or large discretionary accruals. This result suggests that the attenuating effect of forecast divergence is not explained by variation in the quality of reported earnings. Several prior studies find that earnings uncertainty, reflecting the discount rate implicit in the firm s valuation, impacts the return response to earnings news (Imhoff and Lobo 1992; Atiase and Bamber 1994; Clement et al. 2003). This finding suggests that the attenuating effect of forecast divergence could be attributable to cross-sectional variation in earnings uncertainty. These studies rely on analyst forecast dispersion as a proxy for earnings uncertainty. For example, Clement et al. (2003) argue that a lower discount rate arising from a reduction in 4

7 analyst forecast dispersion (earnings uncertainty) is the reason for a positive market reaction to management forecasts that confirm existing analyst forecasts. 3 In a similar spirit, we examine whether the discounting effect of forecast divergence is attributable to high analyst forecast dispersion or the absence of confirmatory management forecasts, and find that these factors do not explain the attenuating effect of forecast divergence. We also examine whether the effect of forecast divergence is due to changes in the cost of capital by conditioning our analysis on whether the earnings surprise is positive or negative. If our results are due to a concomitant increase in the cost of capital with forecast divergence, the attenuating effect should weaken the reaction to positive surprises and strengthen the reaction to negative surprises. However, we find a significant negative attenuating effect for both subgroups of earnings surprises. To further examine if forecast divergence reflects earnings uncertainty, we next relate abnormal trading volume around the earnings announcement to forecast divergence. Prior studies suggest that dispersion in investor expectations of future cash flows prior to the earnings announcement (i.e., pre-announcement information asymmetry) is positively associated with trading volume reaction to earnings news (Karpoff 1986; Kim and Verrecchia 1991; Atiase and Bamber 1994). If forecast divergence is attributable to pre-announcement information asymmetry (hence earnings uncertainty), then this argument suggests that the trading volume reaction to the magnitude of analyst forecast errors will increase in forecast divergence. In contrast, if forecast divergence reflects lower investor reliance on the consensus analyst forecast as an earnings benchmark, it should have a negative impact on the trading volume reaction to the analyst forecast errors (Beaver 1968). Consistent with prior studies, we find that abnormal trading volume is positively related to the absolute value of analyst forecast errors (Bamber 3 Note that confirmatory forecasts carry no incremental cash flow information. Further supporting their argument, Clement et al. (2003) document a decrease in analyst forecast dispersion after such confirmatory forecasts. 5

8 1986, 1987; Landsman and Maydew 2002). We also find a significant negative impact of forecast divergence on this relation. This result incrementally suggests that the discounting is unlikely due to earnings uncertainty, and instead, is more consistent with forecast divergence reflecting lowered investor reliance on consensus analyst forecasts. We conduct several additional tests to examine the contribution to our results from confounding factors such as the bundling of earnings announcements with management earnings forecasts, the relative accuracy of management earnings forecasts, and bad news contained in management forecasts. In addition, to account for unobserved time-invariant firm characteristics as potential contributors to forecast divergence, we re-estimate the impact of forecast divergence using a firm-fixed effects model. The attenuating effect of forecast divergence is robust to the impact of these factors. Taken together, our findings suggest that the attenuating impact of forecast divergence on the return response to analyst forecast errors is likely attributable to investors relying less on the consensus analyst forecast as an earnings benchmark and not due to issues related to the quality of reported earnings or earnings uncertainty. This evidence on the lowered efficacy of the consensus analyst forecast as representative of investor expectations constitutes new evidence on the consensus forecast as an earnings benchmark and is the central contribution of this paper. This finding also suggests that management forecasts have an indirect informational role at the time of earnings announcements. Our analysis complements Clement et al. (2003) who examine the consequences of convergence in management and analyst forecasts, by focusing on the complementary situation when forecasts diverge. Our focus on the case of divergence permits us to disentangle the source of the discounting. Specifically, we are able to attribute the attenuating effect of forecast 6

9 divergence to the consensus analyst forecast not being as representative of investor expectations as when there is no divergence, rather than a mere shift in the discount rate arising from changes in underlying uncertainty of future earnings. This distinction between underlying uncertainty and the representativeness of the consensus forecast as potential sources of decreases in the earnings response coefficient is a novel contribution of this paper. Finally, our results also complement (1) recent findings in Bradshaw et al. (2012) that analyst forecasts are not as superior to time-series forecasts of earnings as has been widely held, (2) findings in recent studies that examine settings in which managers have relative information advantages (Brown and Zhou 2012; Hutton et al. 2012), and (3) Dhole et al. (2010) who focus on identifying settings in which the market relies more on management forecasts than analyst forecasts in determining meet/beat premiums. The rest of this paper is organized as follows. The next section reviews the related literature to develop our central hypothesis. Section 3 describes the selection of the sample and provides variable descriptions. Section 4 discusses the results. Sections 5 reports tests of competing explanations for the impact of forecast divergence. Section 6 reports results of robustness tests. Section 7 summarizes and concludes the paper. 2. Review of Related literature Examination of the properties of the analyst forecasts as earnings benchmarks began by comparisons to the time-series properties of earnings. The superiority of analyst forecasts has primarily been attributed to the broader information set used by analysts (Brown et al, 1987), and their timing advantage, i.e., their forecasts reflect information that arrives after the previous period s earnings numbers have been released. Early research provided evidence that is consistent with such superiority of analyst forecasts and formed the basis for their dominant use 7

10 as earnings benchmarks by capital market researchers. 4 However, recent work by Bradshaw et al. (2012) re-evaluates this practice and finds that time-series forecasts perform better in a variety of settings. Management forecasts, an increasingly important information source that helps form earnings expectations, have not been a significant part of the above work. One reason for this is that much of the debate over earnings benchmarks predates the 1990s during which the incidence of management forecasts sharply increased to a point where such forecasts are now recognized as an integral component of the expectation-setting process. A more enduring reason is that management forecasts, relative to analyst forecasts, have a timing disadvantage. Specifically, management forecasts are made during the fiscal quarter allowing both stock prices and analysts to react to the information contained therein. Evidence of significant price reactions to management forecasts (Ajinkya and Gift 1984; McNichols 1989, Skinner 1994) and revisions of forecasts by analysts following management forecasts (Cotter et al. 2006) also provides the implicit conclusion that in the intervening period between the management forecast and the announcement of earnings the information content of management forecasts has been fully utilized. Recent research on management forecasts in conjunction with prior work on analyst forecasts is bringing this equilibrium about earnings benchmarks under refreshed scrutiny. First, prior research on analyst forecasts recognizes systematic inefficiencies in the processing of earnings and non-earnings information by analysts. These inefficiencies, most commonly consistent with analyst underreaction, arise from a combination of sources including analyst 4 See Bradshaw et al. (2012) for a survey of this evidence. 8

11 incentives and behavioral biases. 5 Such inefficiencies underline the notion that the consensus forecast, while superior to other metrics as an earnings benchmark, is not without its problems. Second, recent research on management forecasts has examined their informational value relative to analyst forecasts. Hutton et al. (2012) find that analysts have an advantage over managers in processing macroeconomic information, whereas managers have a similar advantage with firm-level information. Brown and Zhou (2012) find that managers are comparatively efficient at incorporating information from past earnings changes, accruals, stock returns and analyst forecast errors, and that managers issue earnings forecasts to mitigate analyst inefficiency in processing these pieces of information. These studies help sharpen the contribution made by management forecasts in informing investor expectations about earnings, and point to the need of a deeper examination of the interplay between analyst forecasts and management forecasts. This paper examines the interplay between analyst forecasts and management forecasts at earnings announcements, an issue that has received particularly scant attention in the literature. While the studies cited above discuss various relative strengths of management forecasts, we agree with the general view in the literature that such advantages are likely not sufficient to supplant analyst forecasts as a relatively superior earnings benchmark. 6 However, we also believe that these advantages of management forecasts will lead to investors reliance on them albeit in a less direct way. Our study is aimed at uncovering the full extent of such reliance and addressing the gap created by the paucity of empirical tests of the role of management forecasts around earnings announcements. We expect that although management forecasts do not replace analyst forecasts as earnings benchmarks they moderate their efficacy in this role. Specifically, 5 The evidence on these issues is extensive and not unmixed. See Ramnath et al. (2008) for a recent and comprehensive review. 6 We formally confirm this view at the outset of our empirical analysis (discussed further in Section 4). 9

12 we expect that when analyst forecasts diverge from management forecasts, investors faced with this divergence are likely to rely less on the consensus analyst forecast than when there is convergence between the two types of forecasts. One examination of the impact of management forecasts at earnings announcement is a recent study by Dhole et al. (2010) who find premiums in earnings announcement returns for meeting/exceeding management forecasts when analyst forecasts are not convergent with them. Based on the evidence, the authors conclude that management forecasts, despite their inherent timing disadvantage, seem to serve an important role as in shaping investor expectation of future earnings when analyst forecasts differ from the management forecasts. While related, our study differs in intent and approach from Dhole et al. (2010). Whereas Dhole et al. (2010) uncover specific circumstances in which management forecasts compete with analyst forecasts via the existence of meet/beat premiums, we examine the moderating impact of forecast divergence on the efficacy of the consensus analyst forecast as an earnings benchmark and possible sources of this impact. Thus, the two papers examine complementary aspects of the broad issue of forecast divergence. 3. Sample Selection, Variable Measurement and Descriptive Statistics 3.1 Sample Selection We use data from First Call, COMPUSTAT and CRSP for U.S. firms with fiscal periods ending during 1996 through We use management forecasts of quarterly earnings from the First Call Company Issued Guidelines dataset 7 and restrict the sample to point forecasts and range forecasts where a point estimate can be calculated using the mid-point of lower and upper bounds. For multiple management forecasts in a given firm fiscal quarter, we retain only the last 7 Given the longer horizons of management forecasts of annual earnings, annual forecasts are likely to have more stale information and divergence of such forecasts from the analyst consensus is likely less relevant for investors when evaluating earnings news. Therefore, we limit our sample to quarterly earnings forecasts. 10

13 forecast before actual earnings are announced. This yields an initial sample of 35,125 management forecasts of quarterly earnings. For this sample of forecasts, we obtain corresponding analyst consensus estimates from the First Call Summary Statistics dataset. To measure divergence between managers forecasts and consensus analyst forecasts, we require that at least one consensus analyst estimate be available after the management forecast is issued. This restriction reduces the sample to 28,174 quarterly management forecasts. Availability of reported earnings in the First Call Actual database reduces the sample to 27,513 observations. We further remove observations if stock price at the beginning of the quarter is $1 or less, or if key financial or market data are not available in CRSP and COMPUSTAT, yielding a base sample of 20,738 quarterly management forecasts Variable Measurement Dependent and Independent Variables: Our main dependent variable is the 3-day cumulative abnormal return around earnings announcements (Ab_RET) measured as the daily return less the return on the corresponding value-weighted CRSP size-decile portfolio. Our two key independent variables are the divergence between managers' and analysts' forecasts of EPS (FDIV) and the earnings surprise measured by analyst forecast errors (AFE). We measure forecast divergence as the absolute value of the difference between the last management forecast of EPS (MF) and the final consensus analyst forecast (AF) issued after the last management forecast and before the earnings announcement for a fiscal quarter, with the difference scaled by the stock price at the beginning of the fiscal quarter (FDIV). 9 In regression analyses, we multiply this variable by 100 for expositional convenience. To allow for differences in the effect of forecast divergence when the final analyst consensus is higher than the management forecast 8 The number of observations in some regression analyses is smaller due to additional data requirements. 9 In First Call, management earnings forecasts are not adjusted for stock splits, while consensus analyst forecasts are. Therefore, we adjust management earnings forecasts using the adjustment factor provided by First Call. 11

14 versus when it is lower, we also sub-group FDIV based on whether AF > MF or AF < MF. Consistent with the literature, we measure analyst forecast errors as reported earnings minus AF, scaled by stock price at the beginning of the fiscal quarter. Control Variables: We include in our analysis variables that have a documented effect on the market's response to earnings news. Collins and Kothari (1989) identify growth opportunities and systematic risk as significant determinants of the market response to earnings news. 10 Firm size is also known to play a similar role (Freeman 1987; Collins et al. 1987). We use the marketto-book value of equity (MTB), beta from the CAPM using monthly stock returns (BETA), and market capitalization (MKTCAP) to control for the effect of growth opportunities, systematic risk and firm size, respectively. Firms with persistent earnings tend to have larger earnings response coefficients (ERCs) (Collins and Kothari 1989). We use an indicator variable for negative earnings in one or more of the prior four fiscal quarters (LOSS) as an inverse measure of earnings persistence. We also use the amount of special items in reported earnings scaled by the market value of equity (SPCL-ITEM) because such items are expected to decrease the persistence of earnings (Bradshaw and Sloan 2002). Finally, we include the number of analyst forecasts (ESTNUM) to control for the effect of the information environment on ERCs (Collins and Kothari 1989) and the inverse of lagged stock price (INVERSE-PRICE) used to deflate analyst forecast errors (Cheong and Thomas, 2011). All control variables are measured at the beginning of the fiscal quarter and winsorized at the top and bottom 1 percent level. Detailed variable descriptions are reported in the Appendix. 3.3 Descriptive Statistics 10 Specifically, they document that earnings response coefficients are positively associated with growth opportunities and negatively associated with systematic risk. 12

15 Table 1, Panel A shows the frequency of quarterly management forecasts by year. Consistent with prior studies, the number of in-sample management forecasts steadily increases during our sample period , with a significant jump from 898 in 2000 to 2125 in The percentage of firm quarters with perfect agreement between management and analyst forecasts increases from around 32 percent to 43 percent in the period 1996 to 2003 and thereafter falls back to around 34 percent in 2006 and The fraction of relatively optimistic analyst forecasts (AF>MF) is substantially higher towards the end of the sample period (over 48 percent), while before 2002 this number is generally below 30 percent. The proportion of pessimistic analyst forecasts monotonically increases to its peak of 37 percent in 2001 and decreases thereafter to 17 percent in We also examine forecast divergence for the 48 Fama-French industry groups and find that it is not concentrated in particular industries (result not tabulated). Overall, the data shows that forecast divergence is widely prevalent during our sample period. INSERT TABLE 1 HERE Panel B of Table 1 reports descriptive statistics of the dependent and key independent variables. The pooled distribution of FDIV has a mean (median) of (0.0316) percent. FDIV_DUM, an indicator variable that breaks out the sample into sub-groups with and without forecast divergence, has a mean of percent compared to approximately 69 percent for the First Call universe. On average, the final consensus analyst forecast is recorded 40 days after the management forecast is announced (HORIZON_LENGTH). The sample firms are generally large with mean (median) market capitalization of $4,875 million ($1,097 million) and an average of about 9 analysts following each firm. The positive mean (median) earnings surprise (AFE) reflects, on average, pessimism in last consensus analyst forecasts (p < 0.01) and is consistent 13

16 with the positive earnings announcement return (Ab_RET). The mean (median) Ab_RET at 0.27 percent (0.27 percent) is statistically significant at the 1 percent level (result not tabulated). 4. Empirical Results 4.1 Analyst Forecasts versus Management Forecasts as Earnings Benchmarks We begin our analysis with a comparison of the final consensus analyst forecast and the most recent management forecast in terms of their accuracy and the extent to which they explain investor reaction to earnings announcements. Panel A of Table 2 reports the distributions of forecast accuracy measured by absolute values of the difference between reported earnings and analyst (management) forecasts. The mean (median) forecast error is 0.17 percent (0.07 percent) for consensus analyst forecasts and 0.24 percent (0.09 percent) for management forecasts. These are significantly different (p < 0.01). The smaller magnitude of analyst forecast errors is expected on account of the final consensus forecast being significantly closer to the announcement of earnings. INSERT TABLE 2 HERE To place the proposed role of forecast divergence in an appropriate context, we next examine which earnings forecast investors rely more on by regressing 3-day cumulative announcement returns (Ab_RET) on each type of forecast error. A Vuong s (1989) test comparing the R 2 in Panel B of Table 2 shows the two have significantly different explanatory power (p < 0.01) percent for AFE versus 1.55 percent for MFE. The result is similar when we limit the sample to firm quarters with forecast divergence. We also regress Ab_RET on AFE and MFE in the same regression and find that for both the full sample and the sub-sample of firm-quarters with forecast divergence, the coefficient on AFE is positive and significant, while 14

17 the coefficient on MFE is insignificantly different from zero (Panel C). 11 Taken together, the results in Table 2 confirm the existing view that the most recent consensus analyst forecast is a better proxy of investors earnings expectation. We use this unambiguous support for the superiority of analyst forecasts in measuring investors earnings expectation as the basis for our test design. Specifically, given that the management forecast is, on average, not a viable competing earnings benchmark, we examine whether investors evaluation of the earnings surprise measured by analyst forecast errors (AFE) is influenced by the divergence of the consensus forecast from the management forecast, i.e., by forecast divergence. 4.2 Influence of Forecast Divergence on Return Response to Analyst Forecast Errors To test our central hypothesis that investors lower their reliance on analyst forecasts when analyst forecasts diverge from management forecasts, we examine whether the stock return response (Ab_RET) to the analyst forecast error (AFE) is conditional on forecast divergence (FDIV). Specifically, we estimate equations (1) and (2) below: Ab_RET = a 0 + a 1 AFE + a 2 AFE*FDIV + a 3 FDIV + Σ AFE*Control Variables + Σ Control Variables + ε 1...(1) Ab_RET = b 0 + b 1 AFE + b 2 AFE*FDIV*(AF>MF)+ b 3 AFE*FDIV*(AF<MF) + b 4 FDIV + Σ AFE*Control Variables + Σ Control Variables +ε 2... (2) We expect the coefficient on AFE*FDIV to be negative and significant. To test whether the effect of FDIV differs depending on whether analyst estimates of EPS are above or below management estimates, in equation (2), we condition FDIV accordingly, using indicator variables for whether AF is greater or less than MF. Panel A of Table 3 reports regression results for the simple versions of equation (1). Model 1 is the base model where Ab_RET is regressed on AFE. Model 2 which includes the interaction between AFE and forecast divergence in indicator form 11 This is equivalent to a model with AFE and FDIV as explanatory variables. In such a model, we find that AFE is positive and significant whereas FDIV is insignificant. 15

18 (FDIV_DUM) shows that the coefficient on AFE is positive and significant (5.946, p < 0.01) and the coefficient on the interaction AFE*FDIV_DUM is negative and statistically significant ( , p < 0.01). 12 This magnitude of the coefficient corresponds to a 31 percent (1.822/5.946) mean discounting of the associated earnings surprise (Model 2). The result is qualitatively stronger when divergence is measured in continuous form (Model 3). The coefficient on AFE*FDIV is (p < 0.01). Supporting this result, inclusion of AFE*FDIV increases the regression R 2 by 16 percent (F= , p < 0.01, result nor tabulated). Finally, consistent with forecast divergence being prior public information, its direct effect on returns is statistically insignificant whether it is measured in indicator (FDIV_DUM) or continuous form (FDIV). INSERT TABLE 3 HERE In Panel B of Table 3, we estimate equations (1) and (2) with the control variables discussed in Section 3 and interactions of AFE with each of the control variables. The results show that the impact of forecast divergence is robust to the inclusion of the control variables. The coefficient on AFE is positive and significant (5.923, p < 0.01), while the coefficient on AFE*FDIV is negative and significant (-2.459, p < 0.01). In Model 2, we estimate the moderating effect of forecast divergence separately when analyst forecasts are greater (less) than management forecasts, FDIV*(AF>MF) (FDIV*(AF<MF)). While the moderating effect of forecast divergence continues to hold regardless of whether analyst forecasts are larger or smaller than management forecasts, the effect is greater when forecast divergence reflects optimism in analyst forecasts. The coefficient on AFE*FDIV (AF>MF) of (p < 0.01) and on AFE*FDIV (AF<MF) of (p < 0.01) are different at the 1 percent level. Overall, the results in Table 3 provide strong support for the primary thesis of this paper consensus analyst 12 All reported t-stats for least squares regressions are robust to heteroscedasticity and clustering at the firm level. 16

19 forecasts are less efficient in measuring investor expectation of earnings when they diverge more from management earnings forecasts. 4.3 Construct Validity of Forecast Divergence Because the impact of forecast divergence is an attenuating one, a natural concern is whether the results arise from the invalidity of our construct of forecast divergence. Specifically, measurement errors arising from the format in which the management forecast is made (range vs. point) might introduce noise into our measure of forecast divergence. Or forecast divergence itself may be the result of the management forecast being dated or stale. Although these effects should bias us away from finding a significant impact of forecast divergence, we re-estimate the discounting effect of forecast divergence after controlling for these factors. 13 To examine the staleness aspect, we interact the number of days between the management forecast date and the consensus analyst forecast date (HORIZON_LENGTH) with AFE*FDIV (alternatively with AFE*FDIV conditioned on whether AF>MF or AF<MF). We find that while HORIZON_LENGTH weakens the discounting effect of forecast divergence, the discounting effect remains significant (p < 0.01 level). As an alternative to HORIZON_LENGTH, we use an indicator variable for preannouncements defined as managers forecasts announced after the fiscal quarter end. We find qualitatively similar results. Next, to account for the effect of potential measurement errors in management forecasts, we partition the sample based on the format of management forecasts: range versus point estimates. While the discounting effect of forecast divergence is larger for point forecasts than for range forecasts, it is significant for both formats. Finally, to account for a potential measurement error associated with the skewed distribution of forecast divergence, we re-estimate the impact of forecast divergence using it in non-parametric form. Specifically, we use the decile (quintile) rankings of forecast divergence. 13 These results, not tabulated for the sake of brevity, are available from the authors upon request. 17

20 We find that the significance of the impact of forecast divergence remains unaltered. These results suggest that the effect of forecast divergence is not just an outcome of stale management forecasts or measurement errors in forecast divergence. 5. Alternative Sources of the Impact of Forecast Divergence In this section, we examine potential sources that can mimic and explain the impact of forecast divergence. The moderating impact of forecast divergence on the valuation of earnings surprises can arise from a combination of (1) cash flow effects due to the transitory components in reported earnings, and (2) discount rate effects arising from changes in underlying earnings uncertainty. 5.1 Forecast Divergence and Transitory Earnings Transitory earnings weaken the return-earnings relation (Kothari and Zimmerman 1995) and may occur for various reasons including unusual gains or losses resulting in an extreme earnings change, agency-motivated managerial use of discretionary accruals, and accounting conservatism (Kothari 2001). By definition, transitory earnings are difficult to predict for analysts, and hence likely to increase forecast divergence. In the following subsections, we examine two sources of transitory earnings (TRANS-EARN), discretionary accruals and extreme earnings news, to examine whether they explain the moderating effect of forecast divergence. We report the results in Table 4. INSERT TABLE 4 HERE Discretionary Accruals 18

21 We use the absolute value of discretionary accruals as a first measure of TRANS-EARN. 14 HIGH-ACCRUAL, an indicator variable coded as 1 if the magnitude of discretionary accrual is in the top quintile of the sample, is interacted with AFE and AFE*FDIV. Results in the second column of Table 4 show that the coefficient on the interaction between AFE and TRANS-EARN is negative and significant (-1.612, p = 0.05), supporting prior findings that the return response to earnings news is lower when the transitory component of earnings is larger. The coefficient on AFE*FDIV at remains negative and significant (p < 0.01) suggesting that the discounting effect of FDIV is incremental to the effect of discretionary accruals. The coefficient on AFE*FDIV*TRANS-EARN is (p = 0.06). The insignificant discounting effect of forecast divergence when earnings quality is low ( = 0.949, p = 0.17) suggests that high discretionary accruals mitigate the effect of forecast divergence. A potential explanation of this result is that investors rely more on analyst forecasts when earnings are more transitory, making the effect of forecast divergence less salient. When we partition FDIV into AF>MF and AF<MF, we find results qualitatively similar to those for unconditional FDIV (column 3) Extreme Earnings News Prior studies document that the return-earnings relation is non-linear (Freeman and Tse 1992; Das and Lev 1994) and weakens at extreme values of earnings surprises that are generally expected to be transitory. In this sub-section, we estimate transient earnings (TRANS-EARN) with an indicator variable for extreme earnings (EXTREME-EARNINGS). Specifically, we use a EXTREME-EARNINGS to indicate the top or bottom decile of earnings surprises and interact the dummy with AFE, and similarly with the interaction term AFE*FDIV. The results are reported in the fourth and fifth columns of Table Discretionary accruals are residuals from the modified Jones model estimated by firm-quarter and two-digit SIC code. We require the sample to have at least 10 observations for each firm quarter and two-digit SIC code. The data requirements for discretionary accruals reduce the sample size to 12,118 observations. 19

22 Consistent with prior studies, the results in Table 4 show that extreme values of AFE are significantly discounted the coefficient on AFE*TRANS-EARN is (p < 0.01) in the model with divergence in unconditional form (Model 1) and (p < 0.01) in the model with divergence conditioned on whether AF>MF or AF<MF. However, the effect of FDIV remains significant the coefficient on AFE*FDIV is (p < 0.01) in Model 1 and the coefficients on AFE*FDIV*(AF<MF) and AFE*FDIV*(AF>MF) in Model 2 are (p < 0.01) and (p < 0.01), respectively. Interestingly, the effects of extreme earnings surprises and forecast divergence do not reinforce each other. AFE*FDIV*TRANS-EARN has a coefficient of (p < 0.01) suggesting that, similar to the case of high discretionary accruals, investors rely more on analyst forecasts when they perceive the earnings to be less persistent, making the effect of forecast divergence less salient. 15 Overall, the results in Table 4 suggest that although there is a discounting of the earnings surprise when it contains a large transitory component, a significant portion of the discount due to forecast divergence is incremental to these effects. 5.2 Forecast Divergence and Earnings Uncertainty Investor response to the information content of earnings surprises is also a function of the implicit discount rate arising from the underlying uncertainty of future earnings. In the following three sub-sections, we examine whether the attenuating effect of forecast divergence is attributable to earnings uncertainty Confirming versus Non-confirming Management Forecasts Clement et al. (2003) document a positive stock price response to confirmatory management forecasts and a subsequent reduction in analyst forecast dispersion. They interpret 15 Results using divergence conditional on its direction (Model 2) are similar. The F tests reported at the bottom of Table 5, Panel B show that, although significantly weaker, the discounting effect of forecast divergence continues to hold even under extreme earnings news. 20

23 this as evidence of a decrease in earnings uncertainty caused by the convergence of management and analyst earnings forecasts. Based on Clement et al. (2003), we measure earnings uncertainty by the existence or absence of confirming management forecasts. We also use analyst forecast dispersion at the time of earnings announcement as an alternative measure of earnings uncertainty (Imhoff and Lobo 1992). Using these measures, we examine if the attenuating effect of forecast divergence is attributable to these two measures of earnings uncertainty. INSERT TABLE 5 HERE Table 5 reports results using confirming versus non-confirming management forecasts. In the regression with the full sample, the coefficient on AFE is positive and significant at the p < 0.01 level. The coefficient on the interaction of AFE and CONFIRM-MF, an indicator variable for firm quarters with confirming management forecasts, is positive and significant (2.852, p < 0.01). This result suggests that the reduction in uncertainty of future earnings due to confirming management forecast leads to not only a positive market response when confirming management forecasts are announced (Clement et al. 2003), but also a stronger market response to earnings news when earnings are announced. With respect to the effect of forecast divergence, the coefficient on AFE*FDIV in Model 1A remains negative and significant (-2.367, p < 0.01). These results continue to hold when we condition forecast divergence on whether AF>MF or AF<MF (Model 1B). Also, this moderating effect of forecast divergence remains similar when we estimate the effect separately for each sub-sample of non-confirming management forecasts (Models 2A and 2B) and confirming management forecasts (Models 3A and 3B). These results suggest that the discounting effect of forecast divergence is different from the effect arising from the existence or absence of confirming management forecasts Analyst Forecast Dispersion 21

24 Prior literature has often used analyst forecast dispersion as a proxy for the uncertainty about future earnings or the degree of consensus among analysts or market participants (see for example, Imhoff and Lobo 1992; Diether, Malloy, and Scherbina 2002). Table 6 reports the effect of forecast divergence on the return-earnings relation after controlling for analyst forecast dispersion. HIGH-AF-DISPERSION is an indicator variable coded as 1 if the standard deviation corresponding to the last consensus analyst forecast deflated by beginning-of-quarter stock price is in the top quintile of the distribution, and 0 otherwise. In Model 1 of Table 6, the coefficient on AFE*HIGH-AF-DISPERSION is negative and significant (-3.468, p < 0.01), consistent with a lower return response to analyst forecasts errors when analyst forecast dispersion is high. More importantly, the discounting due to forecast divergence persists - the coefficient estimate on AFE*FDIV in Model 1 is (p < 0.01). Interestingly, while analyst forecast dispersion is negatively associated with the market response to earnings news, it mitigates the discounting effect of forecast divergence as suggested by the positive coefficient on AFE*FDIV*HIGH-AF- DISPERSION (3.792, p < 0.01). Model 2 (Table 6) shows the result persists when we partition forecast divergence depending on whether AF > MF or AF < MF. INSERT TABLE 6 HERE Cost of Capital Effects Prior theoretical (Barry and Brown 1985, Merton 1987) and empirical work (Botosan 1997; Clarkson and Thomson 1990) show a link between a firm s information costs/risk and its cost of capital. Clement et al. (2003) extend such arguments to the context of voluntary disclosure with confirming management forecasts contributing to a lower discount rate. In a similar way, forecast divergence could have the opposite effect and contribute to an increase in the discount rate. Ceteris paribus, an increase in the firm s cost of capital should have a 22

25 downward effect on the investor response to earnings surprises. Thus, an increase in the cost of capital should lower the reaction to positive earnings surprises but strengthen the reaction to a negative earnings surprise. To test the potential effect of cost of capital, we examine the moderating effect of forecast divergence on the earnings surprise conditional on whether the surprise is positive (POSAFE) or negative (NEGAFE). Because it is unlikely that small degrees of forecast divergence lead to an increase in the discount rate, we replace FDIV with a dummy variable to indicate divergence in the top quintile, 0 otherwise (HIGH-FDIV). 16 INSERT TABLE 7 HERE If the effects of forecast divergence are due to an increase in the cost of capital, the coefficient on the interaction term POSAFE*HIGH-FDIV should be negative and that on NEGAFE*HIGH-FDIV should be positive. However, results in Table 7 show that both interaction terms have significant negative coefficients (p<0.01). Partial support for the discount rate argument could be found if the magnitude of the coefficient on the interaction with positive earnings surprises (-3.324) is larger than for the interaction with negative earnings surprises ( ). However, the coefficients are not significantly different from each other (F=0.14, p=0.71). Thus, we are unable to find any discernible influence of cost of capital in the moderating effect of forecast divergence. Overall the results in the previous three sub-sections do not support earnings uncertainty as an explanation for the discounting effect of forecast divergence. Therefore, we infer that the attenuating effect of forecast divergence is more likely due to lowered investor reliance on analyst forecasts in the presence of forecast divergence Forecast Divergence and Trading Volume 16 By doing this, we increase the likelihood of finding support for the competing discount rate explanation. 23

26 To further examine whether larger forecast divergence reflects greater pre-announcement earnings uncertainty or a lower information content of analyst forecast errors, we examine the trading volume response to analyst forecast errors. The former hypothesis predicts a positive association between trading volume response to the magnitude of analyst forecast errors and forecast divergence, whereas the latter hypothesis predicts a negative association. Abnormal trading volume (AVOL) is measured following Landsman and Maydew (2002) (details in the Appendix). To avoid problems arising from extreme values in the distribution of AVOL, we use the percentile rankings as the dependent variable instead of using it in raw form. 17 The results are reported in Table 8. Consistent with prior studies we find a strong positive association between AVOL and the magnitude of the earnings surprise ( AFE ). Model 1 also shows a significant negative effect of forecast divergence on the association between AFE AVOL (p < 0.01). As with announcement period returns, the main effect of FDIV on trading volume is insignificant. Model 2 shows that the impact of forecast divergence continues to persist (p < 0.05) when we control for the competing impact of the control variables discussed earlier. INSERT TABLE 8 HERE These results are supportive of forecast divergence being associated with the lower information content of analyst forecast errors, and inconsistent forecast divergence reflecting greater pre-announcement earnings uncertainty. 6. Robustness Tests In this section, we conduct robustness tests to account for potential confounding factors that can impact the return response to analyst forecast errors and potentially the effect of forecast 17 Using median regressions yields slightly weaker but qualitatively similar results. 24