Interactions between Analyst and Management Earnings Forecasts: The Roles of Financial and Non-Financial Information Lawrence D. Brown Seymour Wolfbein Distinguished Professor Department of Accounting Fox School of Business Temple University (404)520-0489 ldbrown@temple.edu Ling Zhou Assistant Professor of Accounting Anderson School of Management University of New Mexico (504)875-0304 zhoul@unm.edu November, 2012 We thank Rajiv Banker, Sudipta Basu, Eric Press, Richard Willis, an anonymous reviewer for the 2012 AAA annual meeting and workshop participants at Georgia State University, Temple University and the 2012 AAA annual meeting.
Interactions between Analyst and Management Earnings Forecasts: The Roles of Financial and Non-Financial Information Abstract We examine interactions between analyst earnings forecasts and management earnings forecasts by investigating: (1) managers comparative efficiency relative to analysts at incorporating financial statement information (past earnings changes and accruals), and non-financial statement information (stock returns) into their earnings forecasts; (2) the extent to which analyst inefficiencies in incorporating financial versus non-financial information into their earnings forecasts relate to managers propensity to issue earnings forecasts; and (3) the role of financial versus non-financial information at improving analysts forecasts made after management forecasts. We show that: (1) managers have a greater comparative advantage over analysts with respect to incorporating non-financial information than financial information into their earnings forecasts; (2) analysts failure to incorporate non-financial information rather than financial information into earnings forecasts is highly associated with managers propensity to issue earnings forecasts; and (3) after observing management forecasts, analysts improve their earnings forecasts more by incorporating non-financial than financial information into them. 1
1. Introduction Interactions between Analyst and Management Earnings Forecasts: The Roles of Financial and Non-Financial Information Prior research has shown that management s voluntary disclosures help investors to process information more efficiently (e.g., Kimbrough 2005), but it leaves unanswered a key question: what information do voluntary disclosures assist the market to understand? We address this question by examining the interactions between analyst and management earnings forecasts with respect to financial and non-financial information. Specifically, we investigate the: (1) comparative efficiency of analysts and managers at incorporating financial information versus non-financial information into their forecasts; (2) type of information (financial or non-financial) that better explains managers propensity to issue their own forecasts; and (3) type of information (financial or non-financial) that better explains how analysts improve their forecasts after managers issue forecasts. We show that non-financial information is more important than financial information at explaining the interaction between analyst and management earnings forecasts in three contexts. First, managers comparative advantage versus analysts is greater with respect to non-financial information (past returns), than financial information (past earnings changes and past accruals). Second, a stronger association between analyst forecast errors and past returns (but not past earnings changes and past accruals) prompts managers to issue forecasts. Third, analysts improve their post-management earnings forecasts more by utilizing past returns information than either past earnings change information or past accruals information. Our results are consistent with: (1) managers are aware of their comparative information advantage versus analysts at interpreting how past returns impact future earnings; (2) managers 2
awareness prompts them to make forecasts when analysts are more inefficient than usual in incorporating past returns information into their forecasts; (3) analysts are aware of managers comparative information advantage is with respect to non-financial information so they extract managers comparative information advantage from managers forecasts to improve their own post-management earnings forecasts. Our paper makes several contributions to the literature. First, we examine the comparative roles that financial and non-financial statement information play in analysts and managers earnings forecasts. Second, we show that managers greatest information advantage versus analysts is their understanding of how past returns, our proxy for non-financial statement information, impact future earnings. Third, we validate our finding of the important role of nonfinancial information by showing it in a series of sequential, related acts: (1) managers better utilize non-financial information than analysts; (2) managers rely more on how analysts utilize non-financial information versus financial information when deciding whether or not to issue their own earnings forecasts; and (3) analysts rely more on how managers process non-financial versus financial information to improve their own post management forecasts. The rest of our paper is organized as follows. Section 2 reviews related literature. Section 3 develops hypotheses. Section 4 describes the construction and summary statistics of our sample. Section 5 contains results of our empirical tests. Section 6 concludes. 2. Literature Review Our paper is based upon and integrates four literatures in the earnings forecasting area, determinants of: (1) analyst forecast errors; (2) management forecast errors; (3) factors inducing managers to issue forecasts; and (4) factors enabling analysts to improve their post-management 3
forecasts. We combine the first two literatures by focusing on managers versus analysts comparative advantage in incorporating financial and non-financial information into their earnings forecasts, but consider the third and fourth literatures mentioned above separately. 2.1 Determinants of analysts and managers earnings forecast errors Analysts are important information intermediaries in capital markets who facilitate both information generation and dissemination. However, analysts do not efficiently incorporate many types of publicly available information into their earnings forecasts, leading to predictable errors. As examples, Abarbanell (1991) and Lys and Sohn (1990) show analysts underreact to information in past stock returns; Elgers and Lo (1994) find analysts underreact to information in their own past forecast errors, earnings changes and stock returns; Lev and Thiagarajan (1993) and Abarbanell and Bushee (1997) find analysts do not efficiently recognize implications of fundamental signals for future earnings; and Bradshaw, Richardson and Sloan (2001) show analysts do not fully adjust their earnings forecasts for past accruals. While management forecasts and other voluntary disclosures help the market incorporate information (Kimbrough 2005), similar to analysts, managers do not efficiently incorporate all publicly available information into their forecasts. McNichols (1989) shows management forecasts underutilize information in past stock returns; Gong et al. (2009) and Xu (2010) demonstrate management forecasts overestimate the persistence of accruals; and Gong et al. (2010) find management forecasts underutilize information in past earnings changes. The intersection of these two literatures points to three types of information that both analysts and managers inefficiently impound into their forecasts: past earnings changes, past accruals, and past stock returns. We use past earnings changes and past accruals as our proxies for publicly 4
available financial statement information, and past stock returns as our proxy for publicly available non-financial statement information. 2.2 Determinants of the issuance of management earnings forecasts As an important voluntary disclosure mechanism, management earnings forecasts have been extensively researched. Numerous studies have examined factors influencing management decisions to issue earnings forecasts, such as reducing information asymmetry (Verrecchia 2001), mitigating litigation risk (Skinner 1994, 1997), and establishing reputations for reporting transparency (Graham et al. 2005). We focus our discussion on the association between analyst forecasts made prior to management forecasts and the decision of managers to make their own earnings forecasts. Ajinkya and Gift (1984) and Ruland et al. (1990) show managers are more likely to issue forecasts when analyst forecasts contain large errors. Cotter et al. (2006) suggest managers are more likely to issue forecasts when analysts are optimistic. Larocque (2012) finds managers propensity to issue forecasts increases with unpredictable analyst forecast errors, but she did not examine how different sources of analyst errors affect managers decisions to issue forecasts. In contrast, we examine the relation between three key ex ante determinants of analyst forecast errors on management forecast issuance. 2.3 Determinants of improvement in analyst forecasts after management forecasts Waymire (1986) finds management forecasts are more accurate than contemporaneous analyst forecasts, and analyst earnings forecast accuracy improves after management forecasts are released. Cotter et al. (2006) show analysts react quickly to management guidance, making it more likely that their forecasts can be met or beaten. Jennings (1984) finds analyst forecast revisions in response to management earnings forecasts depend on management forecast ex post accuracy. Williams (1996) shows analyst response to management forecasts depends on the 5
usefulness of managers prior forecasts. Bagnoli et al. (2009) find analysts are more likely to mimic management guidance for managers with a history of unbiased or accurate guidance. Feng and McVay (2010) argue analysts overweight management guidance when revising their shortterm forecasts to curry favor with managers. Francis et al. (1997) show analysts forecasting activity increases after management presentations, but do not find improved analyst forecast accuracy, bias or reduction in dispersion. Bowen et al. (2002) show analyst forecasts become more accurate after conference calls. While the literature has established a link between managers forecast accuracy and analysts post-management forecast accuracy, we are the first to examine determinants of analysts improved forecast accuracy post-management forecasts. 3. Hypothesis Development Managers have superior information regarding their own firms performance due to their involvement in their firms daily operations which include making their firms business and reporting decisions. We study the comparative advantage of managers relative to analysts with respect to incorporating financial and non-financial information into their earnings forecasts. We choose past earnings changes and past accruals as proxies for financial information, and past returns as a proxy for non-financial information because they have important implications for valuation as widely documented in the literature. 1 Besides the above mentioned implications for valuation, another reason to use stock returns as the proxy for non-financial information is that returns aggregate various information available to the market. Stock prices incorporate the discounted value of the vector of expected future abnormal earnings (Ohlson 1995); hence they contain information relevant for predicting 1 The post-earnings announcement drift phenomenon has been interpreted as the capital market s underreaction to past earnings changes (Bernard and Thomas 1989). The accruals anomaly (Sloan 1996) suggests capital markets overestimate persistence of past accruals. The momentum anomaly (Jegadeesh and Titman 1993, 2001) has been attributed to capital markets underreacting to information in past returns. 6
future earnings. However, returns also reflect considerable information unrelated to future earnings, such as risks or noises. Managers, with their close involvement in their firms operations and decision making, are likely to have an information advantage over others (including analysts) in differentiating information contained in returns that is relevant to predicting future earnings or that which reflects risk or noise. Managers also have superior knowledge of which period s future earnings current stock price impounds. In other words, relative to analysts, managers have superior information regarding which scalar on the future earnings vector current returns reflects. On the other hand, financial statements are widely distributed and perused by analysts so it is likely managers superior information advantage pertains more to past returns than it does to past earnings changes or past accruals. This reasoning gives rise to our first hypothesis. H1. Managers have a greater information advantage versus analysts in understanding how past stock returns impact a particular future period s earnings than at how past earnings changes or past accruals impact a particular future period s earnings. We make two assumptions in order to develop our second hypothesis: (1) managers recognize they have a superior information advantage versus analysts in understanding how past returns relate to each period s expected future earnings; (2) managers are most likely to make earnings forecasts when they believe that analysts forecasts are especially imprecise. Because managers understand that the comparative precision of their forecasts versus analysts forecasts is greatest with respect to incorporating stock returns (rather than financial information), they are most likely to make earnings forecasts when analysts forecasts of future earnings are especially imprecise at incorporating past stock returns. These assumptions lead to our second hypothesis. 7
H2. Managers are most likely to issue earnings forecasts when they perceive that analysts forecasts of future earnings are especially imprecise at impounding past stock returns. We make three assumptions in order to develop our third and final hypotheses: (1) analysts are aware of managers comparative information advantage regarding the relation of past stock returns to expected future earnings; and (2) analysts cannot determine the nature of managers superior information advantage without perusing managers earnings forecasts (e.g., they do not know which particular scalar of the vector of expected future earnings past returns impound without observing the manager s forecast for a particular future period). These assumptions lead to our third hypothesis: H3. Analysts improve their forecasts after versus before management forecasts more by better incorporating past stock returns into earnings forecasts than at incorporating financial information into earnings forecasts. 4. Data and Sample Description We obtain analyst forecasts and actual earnings from the First Call Historical Database; management forecasts from the First Call Company Issued Guidance Database; and data for calculating all other variables from Compustat and CRSP. Our sample covers 14 fiscal years from 1996 to 2009 2. We examine the association between either signed analysts or managers forecast errors for quarter t (quarter t s actual EPS minus forecast EPS, scaled by stock price) and three pieces of publicly available information as of quarter t-1 s earnings announcement: 2 We start from 1996 because First Call s coverage on management earnings forecasts is sparse before 1996. We end in 2009 because we require other data (e.g., actual earnings) to be available in 2010, the last year with complete data coverage when we started our project. 8
(1) Seasonal earnings change of quarter t-1 (quarter t-1 s earnings minus the earnings of the same quarter of last year, scaled by stock price); (2) Accruals of quarter t-1 (quarter t-1 s earnings minus quarter t-1 s cash flow from operations, scaled by stock price); and (3) Market-adjusted buy-and-hold stock returns 3 from one day after quarter t-2 s earnings announcement to one day before quarter t-1 s earnings announcement. 4 Figure 1 shows the chronological timeline of the six earnings-related events we study: (1) earnings announcement for quarter t-1; (2) first analyst forecast for quarter t made after (1); (3) last analyst forecast for quarter t made before the first management forecast for quarter t made after (2); (4) first management forecast for quarter t made after (2); (5) first analyst forecast for quarter t made after (4); and (6) earnings announcement for quarter t. For analyses of managers comparative information advantage versus analysts (H1), we obtain the first point or range management forecast 5 for quarter t issued after at least one analyst forecast was made for quarter t conditioned on the quarter t-1 earnings announcement (4 above). We then compare the management forecast with the analyst forecast made closest to but before the management forecast (3 above). To investigate whether managers issue guidance in response to specific inefficiencies in analyst forecasts (H2), we obtain the first analyst forecast for quarter t made after the quarter t-1 earnings announcement (2 above), regardless whether management issues a forecast for that quarter. If multiple analysts make forecasts on the same day, we average them. Consistent with Stickel (1989), analysts respond quickly to quarterly earnings announcements; the mean 3 Our results are qualitatively similar if we use raw returns instead of market-adjusted returns. 4 To eliminate the effects of outliers, we delete observations in the top and bottom 1% of the distributions. 5 The Appendix provides detailed variable definitions. Similar to past research, we use the midpoint of the range. 9
(median) number of days between the first analyst earnings forecast and the quarter t-1 earnings announcement is 1.74 (1) days. To examine the improvements in analyst forecasts after management forecasts (H3), we compare analyst forecasts made immediately before management forecasts (3 above) and analyst forecasts made immediately after management forecasts (5 above). The sample used to test H2 is the most comprehensive one as it includes both firms with and without management forecasts. Therefore we provide summary statistics for this sample in Table 1. Table 1 Panel A shows the sample distribution by fiscal year. The entire sample consists of 108,188 firm-quarters with at least one analyst forecast for quarter t made after quarter t-1 s earnings announcement. Only 9.26% of these analyst forecasts are followed by one or more management forecasts. 6 The percent of analyst forecasts followed by management forecasts does not have a linear trend. It starts out (1996-1999) and ends up below the average (2006-2009), and it is above the average in the middle of the period (2000-2004). Panel B presents descriptive statistics of the sample. Consistent with the Abarbanell and Lehavy (2003) evidence that the mean (median) analyst forecast error is optimistic (pessimistic), the mean (median) analyst forecast error (AFE) is -0.0017 (0.0001). Consistent with Waymire (1986) that management forecasts have less error than analyst forecasts, the mean management forecast error (MFE) is -0.0012. Consistent with Foster (1977) and Sloan (1996) respectively, the mean seasonal change in earnings (Chgearn) is positive (0.0014) and mean accruals (Accruals) are negative (-0.0137). Consistent with Fama (1970), the mean return between two adjacent quarters earnings announcements (Ret) is positive (0.0024). 6 This percentage is small compared with some prior studies because we require that a management forecast be made for quarter t after analysts update their forecasts based on the earnings announcements of quarter t-1. For example, Gong et al. (2012) find that about 40-50% of the firm-quarters in their sample have management forecasts but they include all management forecasts made for all subsequent years or quarters, many of which are issued on the same day of earnings announcements; thus before analysts update their forecasts. 10
Panel C shows sample correlation coefficients. 7 Analyst forecast errors (AFE) are negatively correlated with management forecast issuance (D_MF), consistent with evidence that managers walk down analyst forecasts to meet or beat them (Richardson et al. 2004). Consistent with Elgers and Lo (1994), Abarbanell (1991) and Bradshaw et al. (2001) respectively, AFE is positively correlated with past earnings changes (Chgearn) and returns (Ret), and negatively correlated with accruals (Accruals). Similar to analysts forecast errors, managers forecast errors (MFE) are negatively correlated with Accruals (Gong et al. 2009). In contrast to analysts forecast errors, managers forecast errors are negatively correlated with Chgearn and they are unrelated to Ret. The latter suggests managers, unlike analysts, fully incorporate information in past returns into their earnings forecasts. 8 5. Empirical Results Section 5.1 examines the comparative advantages of the management forecast relative to the immediately preceding analyst forecast with respect to our financial information variables: past earnings changes and past accruals, and our nonfinancial information variable: past returns. Section 5.2 examines whether the findings in Section 5.1 are due to management forecasts being 7 The correlation coefficients are generally modest so multicollinearity is not a concern. 8 Our findings that management forecasts overreact to information in past earnings changes differ from Gong et al. (2010) who show an underreaction. The discrepancy may be due to different sample selection criteria. Gong et al. (2010) select the first management forecast issued after the prior quarter s earnings announcement, while we select the management forecast issued after analysts update their forecasts subsequent to the prior quarter s earnings announcement. Given that a substantial proportion of management forecasts are made concurrently with prior quarters earnings announcements, our sample is very different from theirs, our management forecasts are, on average, made later than theirs. More importantly, managers in our sample have already observed analyst forecasts updated after the prior quarter s earnings announcement, so characteristics of these managers forecasts are able to respond to characteristics of analysts forecasts. Our lack of significant results on returns differs from McNichols (1989) who finds management forecasts underreact to information in returns. Her sample differs from ours in two ways. First, similar to Gong et al. (2010), she selects the first management forecasts made for a fiscal period so her sample contains management forecasts made concurrently with earnings announcements. Second, her study uses management forecasts for annual earnings and the stock returns 120 days before and after management forecasts, while we use management forecasts for quarterly earnings and returns between adjacent earnings announcements. Consistent with our results, Gong et al. (2010) find no significant correlation between quarterly returns and management forecasts for quarterly earnings. 11
issued later than the preceding analyst forecasts by comparing them to those of a control sample of analyst forecasts that were not followed by management forecasts. Section 5.3 explores how these financial and nonfinancial information variables impact the likelihood that managers choose to make their forecasts subsequent to analyst forecasts. Section 5.4 examines how analysts improve their forecasts after observing management forecasts with respect to the financial and nonfinancial information variables above. 9 Section 5.5 inspects whether the results in Section 5.4 are attributable to post-management forecast analyst forecasts enjoying a time advantage over pre-management forecast analyst forecasts by comparing the results with a control sample without management forecasts. In sum, sections 5.1 and 5.2, section 5.3, and sections 5.4 and 5.5 respectively provide evidence regarding hypotheses 1 to 3. 5.1 Comparative advantage of management forecasts over analyst forecasts with respect to information in past earnings changes, past accruals and past stock returns In order to test our first hypothesis that managers information advantage relative to analysts is greater with respect to non-financial versus financial information, we first regress the signed error of the last analyst forecast for quarter t made before the first management forecast (AFE_pre) on past earnings changes, past accruals and past stock returns in the multivariate regression below: 10 AFE_pre = β 0 + β 1 Chgearn + β 2 Accruals + β 3 Ret + quarter dummies (1) We report the results in Table 2 Panel A Column 2. Our results are consistent with the extant literature: the coefficients on past earnings changes (Chgearn) and past returns (Ret) are positive, suggesting analyst underreaction to information in past earnings changes and past 9 We assume that the three publicly available sources of information we consider are uncorrelated with analyst private information incorporated in their forecasts, lest our regression specifications suffer from correlated omitted variable problems. Prior studies exclude analyst private information so they also make this assumption. 10 For simplicity, we omit time and firm subscripts in all our equations. We estimate all regressions with robust estimators of variance clustered at the firm level. The dependent variable is based on the third chronological event in Figure 1. 12
returns; accruals (Accruals) has a negative coefficient, suggesting analyst overreaction to information in past accruals. More specifically, the coefficients on Chgearn, Accruals, and Ret are 0.0301 (t-statistic = 3.95), -0.0087 (t-statistic = -2.66) and 0.0063 (t-statistic = 10.92), respectively. Prior research shows managers do not efficiently incorporate all publicly available information into their forecasts (e.g., Gong et al. 2009). We examine this issue using the same information variable that we used to model analyst forecast errors. More specifically, we regress management forecast error (MFE, defined as actual EPS minus management forecasts scaled by price) 11 on Chgearn, Accruals and Ret. The multivariate equation is: MFE = β 0 + β 1 Chgearn + β 2 Accruals + β 3 Ret + quarter dummies (2) Table 2 Panel A Column 3 reports that MFE in equation (2) is negatively associated with Chgearn (coefficient = -0.0107, t-statistic = -2.25). The coefficients on Accruals and Ret are also negative, but, unlike Chgearn, they are insignificant (coefficients = -0.0034 and -0.0003, with t- statistics = -0.77 and -1.51). 12 To ascertain the comparative efficiency of managers versus analysts at incorporating publicly available information into their forecasts, we compare the coefficients for the regression of management forecast errors on our information variables with the coefficients for the regression of analyst forecast errors on our information variables. The results appear in Table 2 Panel A Column 4. The reductions in absolute values of coefficients for Chgearn are 0.0194 (Chi2-statistic = 4.25), for Accruals are 0.0053 (Chi2-statistic = 2.28), and for Ret are 0.0060 11 MFE, the dependent variable, is based on the fourth chronological event in Figure 1. 12 Please refer to footnote 9 for comparison with prior literature. 13
(Chi2-statistic = 54.80), suggesting managers are relatively more efficient than analysts at incorporating information in past earnings change and returns. 13 To test H1, we need to assess the magnitude of relative efficiency of managers versus analysts regarding the three pieces of information into their earnings forecasts. We compare the reduction in absolute values of standard-deviation-adjusted coefficients on Chgearn, Accruals and Ret, calculated as reduction in absolute values of coefficients multiplied by the standard deviation of corresponding variables. We interpret this as the reduction in management forecast errors compared to analyst forecast errors when the same variable is changed by one standard deviation. The results appear in Table 2 Panel A Column 5. The reduction in the absolute values of standard-deviation-adjusted coefficients (multiplied by 1,000) on Chgearn, Accruals and Ret are 0.4332, 0.1982, and 1.2502, respectively 14. Panel B provides tests on the significance of contrasts between differences in absolute values of standard-deviation-adjusted coefficients. We find that the contrasts between Ret and Chgearn (Chi2-stat = 8.57, p-value = 0.0034) and between Ret and Accruals (Chi2-stat = 24.80, p-value = 0.0000) are significant, revealing that managers comparative advantage versus analysts in impounding past returns into their earnings forecasts is greater than for impounding either past accruals or past earnings changes. 15 Consistent with our first hypothesis, the 13 We do not obtain a significant result for past accruals. We use t-tests to determine the significance of a coefficient within an equation and Chow tests to determine the significance of the difference of coefficients between equations. Managers do not issue forecasts randomly so our sample suffers from the self-selection of management forecasts. However, we focus on the differences among past earnings changes, past accruals and past returns so it is unlikely that self-selection issues affect these three pieces of information variables in such a way that self-selection issues provide a plausible alternative explanation for our results. 14 Adjusting for standard deviations does not affect the test statistics or the significance levels so we omit the test statistics for the standard-deviation-adjusted coefficients. 15 We find that the differences in the standard-deviation-adjusted coefficients between the AFE_pre and MFE regressions (instead of the absolute values of these adjusted coefficients) demonstrate a similar pattern (untabulated). While not pertinent to our hypothesis test, we show that the contrast between Chgearn and Accruals and find that it is insignificant (Chi2-stat = 0.92, p-value = 0.3376). 14
comparative efficiency of analysts versus managers is greater at incorporating nonfinancial information (returns) than at incorporating financial information. 5.2 Sensitivity analysis on comparative advantage of management forecasts over analyst forecasts with a control sample In the above analysis, managers issue forecasts after analysts so they possess a timing advantage over analysts. To address the concern that our results may be attributable to information that becomes available after analyst forecasts (Fried and Givoly 1982; Brown et al. 1987) rather than managers comparative advantage, we construct a control sample with no management forecasts, and we test whether the latter analyst forecasts demonstrate an advantage over the earlier analyst forecast similar to what we showed in Table 2. Our control sample consists of analyst forecast pairs for firms in the same two-digit SIC industry as the firm making a management forecast but whose two analyst forecasts are made closest in time to the two forecasts for the Table 2 analysis. We require the number of days between the matched pairs of forecasts to be no more than 10 days. We can match 9,097 out of the 10,013 pairs of analyst forecasts-management forecasts used in Table 2. After deleting cases with extreme 1% values we have 8,397 analyst forecast pairs. We denote the analyst forecasts in the control sample made closest to the premanagement-forecast made by analysts as prior forecasts, and their errors as AFE_prior; and we denote the analyst forecasts in the control sample made close to the management forecasts as subsequent forecasts, and their errors as AFE_subseq. We replicate the same analysis whose results are shown in Table 2 with AFE_prior and AFE_subseq. We report these results in Table 3. To facilitate comparison with the test sample, we also present regression results with AFE_pre and MFE from the test sample. 15
Table 3 Panel A Column 7 shows that for the control sample, analyst forecasts issued later (AFE_subseq) have significantly smaller coefficients for all three sources of information than do analyst forecasts issued earlier (AFE_prior), suggesting that the timing advantage enables analysts to better incorporate all three pieces of public information into their forecasts. More importantly, column 8 shows that the reduction in absolute values of coefficients between the AFE_pre and MFE regressions is significantly greater than the reduction in absolute values of coefficients between the AFE_prior and AFE_subseq regressions for Ret (difference = 0.0050, Chi2-statistic = 28.01), suggesting management forecasts incorporate information in returns more efficiently than analyst forecasts even after considering the time advantage enjoyed by the former. In contrast, the reduction in absolute values of coefficients on Chgearn (difference = 0.0167, Chi2-statistic = 2.12) and Accruals (difference = 0.0025, Chi2-statistic = 0.34) between the AFE_pre and MFE regressions is not significantly different from the reduction between the AFE_prior and AFE_subseq regressions, indicating managers comparative advantage over analysts regarding financial statement information may simply be their timing advantage. To test H1, we evaluate the relative magnitude of the incremental advantage of management forecasts over analyst forecasts (the difference between the reductions in the absolute values of coefficients from AFE_pre to MFE and the reductions from AFE_prior to AFE_subseq) with respect to Chgearn, Accruals and Ret. We compare differences in reductions in absolute values of coefficients adjusted for each variable s standard deviation, i.e., the differences in improvements between the test sample and the control sample when a variable changes by one standard deviation. The differences in improvement in analyst forecasts between the test sample and the control sample are 0.3583, 0.0884 and 1.0075 for one standard deviation of change in Chgearn, Accruals and Ret, respectively. Table 3 Panel B reports that the difference 16
in improvement in analyst forecasts between the test and control samples is greater for Ret than for both Chgearn (Chi2-stat= 3.99, p-value = 0.0458) and Accruals (Chi2-stat = 14.29, p-value = 0.0002), providing further evidence consistent with H1. 5.3 The relation between financial and nonfinancial information sources of analyst forecast error and management forecast issuance We now turn to testing our second hypothesis that managers are most likely to issue forecasts when analysts are inefficient at incorporating past returns into their forecasts. We define D_MF as an indicator variable equal to one if an analyst forecast is followed by a point or range management forecast for the same firm-quarter, zero otherwise. We regress AFE on Chgearn, Accruals and Ret, and their interactions with D_MF and D_MF itself as shown in Equation (3). 16 AFE = β 0 + β 1 Chgearn + β 2 Accruals + β 3 Ret + γ 0 D_MF + γ 1 D_MF Chgearn + γ 2 D_MF Accruals + γ 3 D_MF Ret + quarter dummies (3) If managers are more likely to issue forecasts when analyst forecasts contain errors more closely related to a certain variable (i.e., when analysts are particularly inefficient at incorporating that information), we expect the coefficient on the interaction between D_MF and that variable to be of the same sign as the coefficient on the stand alone variable so that the absolute value of β i + γ i is greater than the absolute value of β i (i = 1-3). If managers are more likely to issue their own earnings forecasts when analyst earnings forecasts contain errors that are more closely related to Ret than Chgearn or Accruals, we expect γ 3 to be more likely to be significant than either γ 1 to γ 2. Given that Table 3 shows that after controlling for their timing advantage, managers possess no comparative advantage over analysts with respect to information in past earnings changes and past accruals, it is plausible that only γ 3 will be significant. 16 The dependent variable, AFE, is based on the second chronological event in Figure 1. 17
Notice that while managers observe analyst forecasts and may decide that analyst forecasts contain errors because they deviate from managers own forecasts, they do not observe what are driving these errors. We argue that managers are most confident that analysts deviations from their own projections are errors when managers own projections are based on the set of information used by investors to establish stock prices. For instance, a manager could have a very positive projection of future earnings because of a big contract to be signed soon (suppose the stock returns have at least partially reflected the forthcoming contract). If the manager believes that she has great comparative advantage over analysts with respect to nonfinancial information (in this case, the forthcoming contract), then when she observes low analyst forecasts, she is likely convinced that her forecast is better than analysts and is likely to issue guidance. Alternatively, the manager could predict high future earnings because she expects past earning increases to persist. If the manager believes that she does not have great comparative advantage over analysts with respect to financial information (in this case, past earnings increases), she is less convinced that analysts lower forecasts are wrong and is less likely to issue guidance. This reasoning indicates that managers are more likely to be propelled to issue guidance when analyst forecast errors are more strongly correlated with returns than when analyst forecast errors are more strongly correlated with financial information. Table 4 provides the results which illustrate three main points. First, consistent with past research and Table 2 Panel A Column 2, Chgearn and Ret have significant and positive coefficients while Accruals has a significantly negative coefficient. Second, consistent with evidence that managers guide analyst forecasts downwards so they can meet or beat them (Richardson et al. 2004), D_MF has a significantly negative coefficient (-0.0014, t-statistic = - 9.08). Third, of the three interactive coefficients we consider, only D_MF Ret is significant (γ 3 = 18
0.0020; t-statistic = 2.80), suggesting that analyst forecasts which are especially inefficient at incorporating the implications of past stock returns rather than past earnings or past accruals for future earnings are more likely to lead managers to issue their own forecasts. This result supports our second hypothesis, which itself builds on our first hypothesis. Our evidence is consistent with: (1) managers have a comparative information advantage versus analysts in understanding the relation of past returns with future earnings; (2) managers are aware of their information advantage; (3) managers seek to alter analysts forecasts to conform to their own forecasts when they are most confident that their forecasts are more precise than analysts forecasts; and (4) managers are most confident that their forecasts are more precise than analysts forecasts when analysts are especially inefficient at impounding past stock prices into their forecasts. 5.4 Sources of Improved Efficiency of Analyst Forecasts before versus after Management Forecasts Having shown that: (1) managers relative information advantage over analysts is greatest in incorporating past returns (rather than financial information such as past earnings changes or past accruals) into their forecasts; and (2) managers are more likely to issue forecasts when analysts are especially inefficient at incorporating past returns into their forecasts, we are ready to test our third hypothesis. We assume that analysts are aware of managers comparative information advantage regarding incorporating past returns into their forecasts, and that they are able to extract (at least part of) managers comparative information advantage by perusing managers forecasts. 17 To test H3, we regress the error of the first analyst forecast made after the 17 The same reasoning could lead to the prediction that analyst forecast revisions around management forecasts should be more strongly correlated with Ret, which is confirmed by our empirical results. Given that analyst forecast revisions around management forecasts are equal to AFE_post minus AFE_pre, the results with analyst forecast 19
management forecast (AFE_post) on Chgearn, Accruals and Ret and we compare the results with those of the AFE_pre regression. 18 Table 5 presents the results. It is evident from Table 5 Panel A Column 3 that after perusing the management forecast, analysts are still inefficient at incorporating Accruals and Ret into their forecasts, but Column 4 shows they are more efficient at incorporating all three pieces of information than they were earlier. Specifically, reductions in the absolute values of coefficients from the pre- to the postmanagement forecast regressions for the three pieces of information are all significant: 0.0265 (Chi2-stat = 10.81) for Chgearn, 0.0083 (Chi2-stat = 6.04) for Accruals 19, and 0.0062 (Chi2-stat = 115.59) for Ret respectively. 20 To test H3, we need to evaluate the relative improvement in analyst efficiency of AFE_post versus AFE_pre across the three pieces of information. We compare reductions in the absolute values of standard-deviation-adjusted coefficients on Chgearn, Accruals and Ret, calculated as the reductions in the absolute values of coefficients multiplied by the standard deviations of corresponding variables. The reduction in the absolute value of a standarddeviation-adjusted coefficient can be interpreted as the reduction in AFE_post as compared to AFE_pre when the corresponding variable is changed by one standard deviation. The reduction in the absolute values of standard-deviation-adjusted coefficients (multiplied by 1000) on revisions can be inferred from the AFE_pre and AFE_post regression results reported in Table 5, hence they are not tabulated. 18 The Table 5 sample is smaller than the Table 2 sample because we now add the requirement that management forecasts are followed by analyst forecasts. However, the results with AFE_pre in Table 5 are qualitatively similar to those of AFE_pre in Table 2. 19 Table 3 shows management forecasts are not more efficient than analyst forecasts in incorporating information in Accruals. However, analyst forecasts still improve after management forecasts with respect to Accruals. This is possible as long as management forecasts are not perfectly correlated with analyst forecasts, hence provide incremental information. 20 Managers do not issue forecasts randomly so our AFE_pre and AFE_post sample suffers from self-selection. However, since our focus is on the differences across past earnings changes, past accruals and past returns, selfselection is only a concern if it is a plausible alternative explanation for our results. We doubt that this is the case, but we must admit we may be wrong. 20
Chgearn, Accruals and Ret are 0.5858, 0.3089 and 1.2993, respectively. Table 5 Panel B shows the reduction in the absolute value of standard-deviation-adjusted coefficient on Ret is significantly greater than that on Chgearn (Chi2-stat = 9.03, p-value = 0.0027) or Accruals (Chi2-stat = 30.40, p-value = 0.0000), indicating that, consistent with H3, the greatest improvement in analyst forecast efficiency is due to analysts incorporating past returns more efficiently into their forecasts. Our findings are consistent with: (1) analysts awareness of managers comparative advantage at understanding the relation between past returns and particular period s future earnings; and (2) analysts ability to improve their earnings forecasts by extracting managers comparative advantage from managers earnings forecasts. Our results support H3, which builds on both H1 and H2. 5.5 Sensitivity analysis on comparative efficiency of analyst forecasts made before versus after management forecasts with a control sample Analyst forecasts made closer in time to the next quarterly report may benefit from information unavailable to analysts who made their forecasts earlier in the time period (Fried and Givoly 1982; Brown et al. 1987). To mitigate the validity threat that analyst improvement is due to a timing advantage rather than their ability to extract information from managers forecasts, we replicate the Table 5 analysis on a control sample with a similar timing advantage but without intervening management forecasts. Specifically, our control sample consists of analyst forecast pairs for firms in the same two-digit SIC industry made closest in time to the pre- and postmanagement forecast analyst forecasts, but without intervening management forecasts. We require the number of days between analyst forecasts for the control firm and the test sample to be no more than 10 days. We can match 7,823 out of the 8,974 pairs of analyst forecasts used in Table 5. After deleting cases with extreme 1% values we have 7,216 analyst forecast pairs. 21
We denote the analyst forecasts in the control sample made close to the pre- (post-) management-forecasts analyst forecasts as early (late) forecasts, and their errors as AFE_early (AFE_late). We replicate the same analysis whose results are shown in Table 5 with AFE_early and AFE_late. We report these results in Table 6. To facilitate comparison with the test sample, we also present regression results with AFE_pre and AFE_post from the test sample. Table 6 Panel A Columns 7 shows that AFE_late has significantly smaller coefficients for Ret than AFE_early for the control group of firms, but not for Chgearn or Accruals. 21 Reductions in absolute values of coefficients between the AFE_pre and AFE_post regressions are significantly greater than reductions in absolute values of coefficients between the AFE_early and AFE_late regressions for Chgearn (difference = 0.0119, Chi2-statistic = 2.83), Accruals (difference = 0.0093, Chi2-statistic = 9.65), and Ret (difference = 0.0031, Chi2- statistic = 30.68), suggesting management forecasts aid analysts to incorporate information in past earnings changes, accruals and returns more efficiently. To test H3, we need to evaluate the relative magnitude of the incremental effects of management forecasts on improvement in analyst forecasts (the difference between the reductions in the absolute values of coefficients from AFE_pre to AFE_post and the reductions from AFE_early and AFE_late) with respect to Chgearn, Accruals and Ret. We compare differences in reductions in absolute values of coefficients adjusted for each variable s standard deviation 22. These can be interpreted as differences in improvement in analyst forecasts between the test sample and the control sample when a variable changes by one standard deviation. The 21 The results with AFE_early and AFE_late differ from the results with AFE_prior and AFE_subseq reported in Table 3 with respect to Chgearn and Accruals. We have no explanation for these result but we note that our focus is on the difference between Ret and Chgearn/Accruals, and all results regarding Ret and its comparison with Chgearn/Accruals and tests of our hypotheses are robust across all analyses. 22 The standard deviations of Chgearn, Accruals and Ret are similar for the test sample and the control sample. We use the average standard deviations of the variables of the two samples to calculate the standard-deviation-adjusteddifference in coefficients. 22
differences in improvement in analyst forecasts between the test sample and the control sample are 0.2343, 0.3261 and 0.6104 for one standard deviation of change in Chgearn, Accruals and Ret, respectively. Table 6 Panel B reports that the difference in improvement in analyst forecasts between the test and control samples is greater for Ret than for Chgearn (Chi2-stat = 3.86, p- value = 0.0494) or Accruals (Chi2-stat = 3.22, p-value = 0.0730), suggesting management forecasts help analysts more with respect to information in past returns than financial information. Our results with a control sample provide further evidence consistent with H3. 6. Conclusions We examine and provide evidence regarding three hypotheses: (1) Is managers relative information advantage versus analysts at predicting future earnings greater with respect to nonfinancial information such as past stock prices than it is with respect to financial information such as past earnings change or past accruals?; (2) Are managers most likely to make earnings forecasts when analysts are especially inefficient at incorporating past returns into their earnings forecasts rather than when they are inefficient at impounding past earnings changes and past accruals into their earnings forecasts?; and (3) Do analysts act as if they are aware of managers comparative information advantage regarding impounding past returns into earnings forecasts and they extract this comparative information advantage, enabling them to improve their earnings forecasts? Our evidence is consistent with the following: (1) managers have a comparative information advantage versus analysts in understanding the relation of past returns with future earnings; (2) managers are aware of their information advantage; (3) managers seek to alter analysts forecasts to conform to their own forecasts when they are most confident that their 23
forecasts are more precise than analysts forecasts; (4) managers are most confident that their forecasts are more precise than analysts forecasts when analysts are especially inefficient at impounding past stock prices into their forecasts; (5) analysts are aware of managers comparative advantage at understanding the relation between past returns and future earnings; and (6) analysts are able to improve their own forecasts by extracting managers comparative advantage from managers forecasts. Our results enable a better understanding of the comparative efficiency and interaction of managers and analysts forecasts. We identify circumstances in which analyst forecasts prompt managers to issue their own forecasts, and how analysts improve their forecasts after perusing managers forecasts. Future research should recognize that analysts and managers forecasts are interdependent and should be considered jointly rather than separately. 24