Effect of Reputation on the Credibility of Management Forecasts*

Effect of Reputation on the Credibility of Management Forecasts* Amy P. Hutton Dartmouth College Phillip C. Stocken Dartmouth College June 30, 2006 Abstract We examine the effect of firm forecasting reputation on investors reaction to management earnings forecasts. We construct a measure of forecasting reputation that reflects prior forecast accuracy and frequency. We hypothesis and find, first, that investors are more responsive to management forecast news when a firm has built a forecasting reputation. Second, we examine forecasts containing extreme earnings news, a setting where investors are expected to find reputation more salient, and find investors are more responsive to these forecasts when the firm has a reputation. Third, when a firm has a forecasting reputation, investors reaction at the management forecast date largely preempts their reaction at the earnings announcement date. Consequently, managers of firms with a reputation have their forward-looking information reflected in stock prices more promptly, yet without a larger overall reaction. Fourth, to establish why all firms do not build a forecasting reputation given the benefits of having one, we demonstrate that it is costly to build a reputation because investors are less responsive to unexpected earnings news when reported earnings fail to reach the management forecast. * We thank Franco Wong, Ken French, Rafael La Porta, and Valentina Zamora for helpful comments. We thank workshop participants at Boston College, MIT, and Tuck School of Business for useful suggestions. Finally, we thank Robert Burnham for excellent research assistance. 1

Effect of Reputation on the Credibility of Management Forecasts You can't build a reputation on what you are going to do. Henry Ford I. INTRODUCTION Traditionally firm managers were reluctant to issue earnings forecasts and other forward-looking statements for fear of litigation. To encourage the release of forecasts, Congress enacted the Private Securities Litigation Reform Act of 1995 (PSLR Act), which contains a safe-harbor provision to shelter managers from litigation arising from unattained projections. Although this legislation was widely welcomed, its opponents contended it would lower the litigation costs associated with unattained forecasts and provide firms with a license to lie (Grundfest and Perino, 1997, ii). In an environment where managers are more likely to offer self-serving forecasts, because the expected threat of litigation is lower, it is more important for investors to evaluate the credibility of management forecasts when impounding this forward looking-information in stock prices. While there are several factors that affect the credibility of management forecasts, the reputation of firm management to forecast in a forthright fashion is a key factor. In this paper, we examine the effect of reputation on investors interpretation and reaction to management earnings forecasts. To examine the effect of reputation, we construct a measure of forecasting reputation for a sample of management earnings forecasts issued after the enactment of the PSLR Act. Our measure of forecasting reputation is a function of both the accuracy and frequency of a firm s prior forecasting behavior. To be classified as having a forecasting reputation, first, a firm must have an accuracy ratio that is better than the average ratio for 2

its industry, where the accuracy ratio is defined as the number of management forecasts that are strictly more accurate than the prevailing consensus analyst forecast relative to the total number of annual forecast issued by the firm. Second, the firm must have offered a sufficient number of forecasts so that investors have had an opportunity to learn about the firm s forecasting ability; the frequency threshold for a particular firm depends on how difficult it is to forecast earnings for that firm. With this measure of forecasting reputation in hand, we separate forecasts of annual earnings issued by firms with a Reputation from those issued by firms with No Reputation and examine whether forecasting reputation affects how investors respond to management earnings forecasts. 1 We find the stock price response to management earnings forecasts is consistent with investors evaluating previously issued management forecasts and responding more to forecasts that are issued by firms that have built a forecasting reputation, irrespective of whether the current forecast contains good or bad news. To further support our primary finding, we examine management forecasts containing extreme news because, in these circumstances, firm reputation is likely to be more important to investors. We find, consistent with our prediction, that when a firm has built a forecasting reputation, investors respond more to both extreme good and bad news. To show that firms that release bad news forecasts are no worse off when they have a strong reputation relative to when they have a weak one, we examine the combined market response to the news in the management forecast and the earnings announcement. 1 Throughout the paper we refer to management s forecasting reputation and the firm s forecasting reputation interchangeably. The presumption is that forecasting reputation stays with the firm even when management changes. If this presumption is incorrect and there are management changes in our sample of Reputation firms, then the power of our tests is weakened. 3

We find that when firms have forecasting reputations, investors reaction at the management forecast date largely preempts their reaction at the earnings announcement date, but without a larger overall reaction. The combined reaction at the management forecast date and earnings announcement date is no different for the two sub-samples. We conclude, therefore, that a forecasting reputation allows a firm to benefit by having its forward-looking information reflected more promptly in its stock price without a larger overall reaction to the earnings news. To examine why all firms do not attempt to build a reputation despite the benefits of having one, we consider whether there are signaling costs associated with developing a reputation. Specifically, we examine whether there is a cost associated with missing a forecast when the earnings realization is revealed and investors observe the management forecast error. Indeed, we find investors capitalize unexpected earnings at a lower rate when the management earnings forecast is not attained. Additionally, we demonstrate that the lower capitalization rate on earnings news is largely driven by missed good news forecasts, indicating that it is more costly to provide good news forecasts. The effect of reputation on firm voluntary disclosure practices has received scant empirical attention. Some empirical research, however, has tangentially considered the role of reputation on how the market responds to firm voluntary disclosure. For instance, Williams [1996], the primary antecedent to our paper, considers how financial analyst responsiveness to a management earnings forecast varies with the usefulness of a previous management forecast. She finds that analyst responsiveness to good news forecasts, which typically are regarded as less credible, varies with the usefulness of a prior good news forecast, whereas analyst responsiveness to bad news forecasts does not 4

vary with the usefulness of the prior forecast. Because management s concern with its reputation can induce voluntary disclosure behavior in a multi-period setting that one does not observe in a single-period setting, an important contribution of our study to the voluntary disclosure literature is to show how in a multi-period setting, management develops a reputation and how this reputation affects the market s response to a firm s forward-looking information. Our study is also reminiscent of Chen, Francis and Jiang [2005] who study how investors learn about analyst forecasting ability. They model analysts as being disinterested information providers. Consistent with this view, they find that investors update their beliefs about analyst ability in a Bayesian fashion in the sense that they weight analyst quarterly earnings forecasts more heavily than their prior beliefs when analyst forecasts are more accurate and frequent. While our study also uses the accuracy and frequency of forecasts as ingredients for determining the market s response to forecasts, we examine the role of management reputation when managers forecast strategically. Accordingly, our measure of forecasting reputation combines these ingredients in a manner that explicitly recognizes that managers might misrepresent their information. This study should be of interest to firm management, policy-makers and regulators. Management may find it of interest because it suggests firms benefit from developing a disclosure reputation. When firms have a forecasting reputation, the market is more responsive to their good and bad forward-looking information, even when the information it is not immediately verifiable or auditable. Because firms with strong reputations have their private information impounded in their stock prices in a more 5

timely fashion, they are more capable of reducing information asymmetries in the market. Consequently, we expect these firms to enjoy a lower cost of capital (e.g., King, Pownall, and Waymire 1990; Healy and Palepu 1993; Coller and Yohn 1997; Verrecchia 2001). Policy-makers and regulators may find this study of interest because it suggests that market forces reduce firms willingness to offer inaccurate forecasts. Policy-makers and regulators have long been concerned with the credibility of management forecasts. Indeed, the Securities and Exchange Commission (SEC) historically prohibited inclusion of forward-looking statements in SEC filings because it argued market forces were insufficient to induce managers to forecast truthfully (Penman 1980; Pownall and Waymire 1989). More recently, policy-makers, securities regulators, and other opponents of the PSLR Act argued that lowering the litigation costs associated with unattained forecasts would lead to misleading forecasting. Our study indicates that market forces encourage forthrightness because firms benefit from having a forecasting reputation. The paper proceeds as follows: Section 2 surveys the existing literature and develops our hypotheses; Section 3 describes our sample; Section 4 defines the variables and specifies the research design; Section 5 discusses the results; and Section 6 summarizes and concludes. Appendix A defines the variables used in our empirical tests. 2. LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT Firm managers forecast for several reasons: The expectations adjustment hypothesis posits that managers issue forecasts to align investors expectations with their own (Ajinkya and Gift 1984; Hassell and Jennings 1986). In addition, managers may forecast to reduce information asymmetry in the market (King et al. 1990), to signal 6

their superior ability to anticipate changes in the economic environment (Trueman 1986), or to reduce expected litigation costs (Skinner 1994; Kasznik and Lev 1995). While managers might typically have incentives to forecast forthrightly, from time to time they have incentives to forecast in a self-serving fashion. Because forward-looking information often is not verifiable or auditable, managers might misrepresent their private information in response to their incentives. For instance, managers may release selfserving forecasts to maximize the value of stock option awards they receive on scheduled option grant dates (Aboody and Kasznik 2000), to reduce the probability of bankruptcy, acquisition, or hostile takeover (Frost 1997), or to discourage entry into their industry (Bamber and Cheon 1998). Consistent with the threat of litigation disciplining management disclosure, several studies examining management earnings forecasts issued before enactment of the PSLR Act find that management forecasts are unbiased on average (e.g., McNichols 1989; Frankel, McNichols, and Wilson 1995; Kasznik 1999). However, following enactment of the PSLR Act, the expected litigation costs associated with unattained forecasts are lower (Johnson, Kasznik, and Nelson 2001). Hence, it is not surprising Rogers and Stocken [2005] find that management earnings forecasts issued after the enactment of the PSLR Act are not only biased, but that managers are more likely to bias their forecasts in response to their incentives when it is more difficult for investors to detect dissembling. Accordingly, in the post-pslr period, it is important for investors to assess the credibility of management forecasts. A key factor affecting investors assessment is management s forecasting reputation. 7

Theoretical studies have examined the effect of reputation on behavior. They show that behavior, which is inconsistent with short-term incentives, can be rationalized in a dynamic equilibrium (see Wilson 1985). Specifically, within the context of a voluntary disclosure game, Stocken [2000] establishes that a manager will forecast truthfully in a multi-period setting even though in a single-period setting the manager dissembles. 2 In his model, the manager is capable of developing a forecasting reputation because investors can use the firm s audited earnings report to assess the credibility of the manager s forecast. If investors assess the manager s forecast to be credible, then they increment the manager s reputation index. In equilibrium, investors impound the manager s forecast in the firm s stock price provided the reputation index is sufficiently high and ignore the manager s forecast otherwise. The ability of the manager to communicate credibly with the market lowers the firm s cost of capital. We posit that investors will be more responsive to management earnings forecasts when the firm has built a forecasting reputation and will be less responsive when the firm has a weak or no reputation. Further, while bad news is typically regarded as being more credible than good news forecasts (e.g., Jennings 1987; Williams 1996; Hutton, Miller and Skinner 2003), in a multi-period setting where managers desire to develop and maintain their forecasting reputation, we expect them to truthfully issue both good and bad news forecasts. Accordingly, when a firm has built a forecasting reputation, we 2 Stocken [2000] establishes that a manager will forecast truthfully, when the manager is evaluated over a sufficiently long time horizon, the manager is sufficiently patient, and the earnings report is sufficiently useful for assessing the truthfulness of the manager s voluntary disclosure. He shows that these conditions are sufficient to guarantee that the manager will truthfully reveal his non-verifiable information even though the earnings report might be a noisy monitor of the manager s forecasting behavior. 8

expect the market to be responsive to both good news and bad news forecasts. Thus, we propose the following directional hypothesis: 3 H1: When a firm has a built a forecasting reputation, investors are more responsive to the news in management earnings forecasts. To support this hypothesis, we examine the investors response to forecasts containing extreme news because these forecasts are typically viewed as being less credible and, hence, firm reputation is likely to be more salient to investors. A well documented feature of the financial disclosure environment is that the returnsearnings news relation is nonlinear, or S-shaped, with the average stock price response declining in the absolute magnitude of the earnings news (e.g., Freeman and Tse 1992; Lipe, Bryant, and Widener 1998). Subramanyam [1996] established that investor uncertainty about the precision of the earnings news is a sufficient condition for this nonlinear pricing function. Further, he showed the extent of the non-linearity is increasing in investor uncertainty about the precision of the earnings report. In this light, we expect that investors will be less uncertain about the precision of a management earnings forecast when the issuing firm has a strong forecasting reputation. Therefore, we predict investors will be more responsive to extreme good and bad earnings forecasts when the firm has a built a forecasting reputation. That is, the non-linear, S-shaped returnsearnings relation is less pronounced, or less non-linear, for firms with forecasting reputations. This argument yields the following directional hypothesis: H2: When a firm has built a forecasting reputation, investors are more responsive to the news in extreme earnings forecasts, and thus the non-linearity in the returnsearnings relation is less pronounced. 3 All hypotheses are stated in the alternative form. 9

A benefit of a forecasting reputation is that the market is more responsive to the information contained in a management forecast and more promptly impounds this information in the stock price. Consequently, when the market learns of a firm s poor performance, the firm s stock price is expected to decline more severely when it has a forecasting reputation than when it does not; this decline, which the firm might interpret as a form of punishment, might discourage it from building a reputation. We argue, however, that the overall response to poor performance is not more sever for firms with a forecasting reputation. Specifically, recognizing that information contained in a firm s reported earnings typically subsumes information in the management earnings forecast, we expect the market s combined response to unexpected earnings at the earnings realization date and to the earnings news at the forecast release date to not depend on whether a firm has a forecasting reputation. This conjecture leads to the following hypothesis: H3: The combined market response to the news in the earnings report and management forecast should not vary with whether a firm has a forecasting reputation. This hypothesis implies that when a firm forecasts, regardless of its forecasting reputation, its stock price is no worse-off and its stock price is no better-off after its earnings report is released. The benefit of a forecasting reputation is in the timeliness of the market s reaction, not in the overall reaction. 4 4 Kasznik and Lev [1995] find firms that provide warnings about negative earnings news are penalized with larger overall stock price reactions than firms that do not provide such warnings. Their findings suggest that it is costly to warn the market. Our analysis differs from theirs since we focus on the stock price reaction to firm voluntary disclosure conditional on firms having offered earnings forecasts. 10

The premise underlying our study is that firms view an increase in investor responsiveness to management earnings forecasts as desirable, even when they release bad news. Indeed, Fuller and Jensen [2002] contend that an overvalued stock can be as damaging to a firm as an undervalued stock because it often leads to dysfunctional firm behavior. As a consequence, they argue management should promptly inform the market when they believe market expectations cannot be met and hence the stock is overvalued. 5 We posit that investors are more responsive to firms that have a forecasting reputation. Since a forecasting reputation is valuable, there must be some costs associated with building a reputation, otherwise every firm would simply choose to build one (see Klein and Leffler, 1981). 6 In our management forecasting setting, there is a signaling cost associated with building a forecasting reputation if investors capitalize unexpected earnings at a lower rate at the earnings realization date when reported earnings fail to reach management s forecasted earnings. When management misses its forecast, investors are more uncertain about the precision of the firm s earnings and accordingly apply a lower capitalization rate to the earnings news. A firm s inability to communicate with the market at the earnings report date raises its cost of capital (e.g., King et al. 1990; Verrecchia 2001). Thus, we expect some firms will choose not to forecast and thereby avoid these costs. 7 This argument leads to the following directional hypothesis: 5 For similar arguments, see Lees [1981] survey and discussion of management forecasting behavior. 6 Klein and Leffler [1981] examine how costly signaling might overcome the moral hazard problem present in settings where firms supply experience goods whose properties are only revealed ex post. They suggest that firms charging higher prices do in fact supply higher quality goods, which are more costly to produce, for fear that if they deviate and supply lower quality goods, they will lose the higher margins. The notion of reputation in Klein and Leffler [1981] and this paper, therefore, differs from that formalized in Kreps and Wilson [1982] and Milgrom and Roberts [1982] where a firm with a fixed type misleads by imitating the behavior of another type until it is advantageous to capitalize on this deception. 7 Indeed, in their survey of factors that affect management performance measurement and voluntary disclosure decisions, Graham, Harvey and Rajgopal [2005] note that firms contemplating whether to offer 11

H4: When a management earnings forecast is not attained, investors are less responsive to unexpected earnings at the earnings realization date. It is commonly argued that managers have incentives to release good news and withhold bad news (e.g., Miller 2002). Further, we expect that missing a forecast after raising investor expectations with a good news forecast sends a more negative signal about a firm s forecasting ability than does missing a forecast after lowering investor expectations with a bad news forecast. Consequently, we expect it to be more costly to develop a reputation for releasing credible good news forecasts. That is, we expect investors to be less responsive to unexpected earnings at the earnings realization date when management released and missed a good news forecast relative to when management released and missed a bad news forecast. 3. SAMPLE DEVELOPMENT To examine how management builds and maintains a reputation, we use the First Call database of Company Issued Guidance (CIG) to identify a sample of 13,877 annual management earnings forecasts released between January 1996 and December 2003. 8 Our sample period begins with the enactment of the PSLR Act and extends through the period for which we have CIG data. Our sample includes only point, range, one-sided directional, and confirming forecasts; it excludes qualitative forecasts not specific enough to provide a numerical earnings per share (EPS) forecast needed to construct our measure earnings guidance recognize that the cost of missing a consensus analyst forecast is greater when the firms has offered guidance than when it has not done so. 8 For the calculation of firm and industry Accuracy Ratios and firm Forecasting Frequency described in section 4, we use all point, range, one-sided directional, and confirming forecasts captured on the CIG data base since its inception in January 1994. 12

of reputation. 9 This sample excludes earnings pre-announcements or earnings warnings, which are released after the end of the fiscal period. Since we use shortwindow stock price reactions to assess investor response to management earnings forecasts, we delete 1,779 observations with more than one forecast made on the same day. We delete these observations because we cannot separately identify investor reaction to each forecast, and it ensures the sample observations are independent. We merged our First Call dataset with financial data and earnings announcement dates from Compustat. We use the Compustat earnings announcement dates to identify and delete 1,851 management forecasts made within 1 to +1 days of an earnings announcement, leaving a sample of 10,247 forecasts. Deleting forecasts made within a three-day window centered on earnings release dates reduces the likelihood that other earnings news explains the observed stock price reactions. We delete observations with missing Compustat data for our control variables (499 observations). We then merged our First Call dataset with stock return data from CRSP, with the First Call database of analyst estimates of annual earnings per share, and with the First Call database of realized or actual annual earnings per share. We delete observations with missing stock return data on CRSP (7 observations) and missing analyst estimates or realized earnings on First Call (360 observations), leaving a sample of 9,381 management forecasts made by 2,138 firms. The development of the sample used to test H1 and H2, labeled as the full sample, is summarized in Table 1. Table 2 describes the distributional properties of the full sample of 9,381 management earnings forecasts. The number of management earnings forecasts in our 9 We recognize that our sample excludes a substantial proportion of non-quantitative management forecasts; see Pownall, Wasley, and Waymire [1993] and Bamber and Cheon [1998]. 13

sample increases from the early to the latter part of the sample period (see Panel A of Table 2). The increase in the early years is likely attributable to the expanding coverage of the First Call CIG database; 10 the more dramatic increase in 2001 through 2003 is most likely due to Regulation Fair Disclosure (Reg FD). Following the introduction of Reg FD on October 23, 2000, firms can no longer privately reveal information to analysts to assist them in the development of their earnings estimates. Thus, managers are more likely to publicly release their forecasts to align investor expectations with their own (see Bailey, Li, Mao, and Zhong 2003; Heflin, Subramanyam, and Zhang 2003; and Straser 2002). First Call provides a classification of the type of management forecast: point, range, one-sided directional, or confirming statements. Approximately 68 percent of the management forecasts included in our sample are range forecasts, 26.6 percent are point forecasts, and only a little over 5 percent are one-sided directional or confirming forecasts (see Panel B of Table 2). To determine a numeric value for each forecast, we use the value of the point and one-sided directional forecasts and midpoint of the range forecasts. 11 For confirming forecasts (that are not point forecasts) we use the median consensus analyst estimate prevailing on the day of the management forecast as the earnings per share value of the management forecast. 10 In an attempt to understand how First Call expanded its coverage over our sample period, we examined the number of firms and the number of management earnings forecasts appearing on the First Call s CIG data base in each of our sample years. We sort observations by size deciles and 2-digit SIC codes. The percentage of the CIG observations falling across size deciles and 2-digit SIC codes does not vary in a systematic way across our sample years. Thus, while it is likely that First Call expanded its coverage during our sample period, First Call s decision rule for expanding its coverage is not obvious to us. 11 Prior research suggests that investors use the mid-point of a range forecast when forming their expectation of earnings (e.g., Baginski, Conrad, and Hassell 1993; Hirst, Koonce, and Miller 1999). Nevertheless, in 144 instances, statements by management (captured in the CIG database) indicated that earnings per share would be at the low (high) end of the range. In these instances, we used the low (high) end of the range as the value for management s earnings per share estimate. 14

To assess whether the management forecast reveals good, bad, or confirming news, we consider the forecast relative to the prevailing median analyst consensus earnings estimate. If the management forecast is higher than the prevailing median analyst consensus estimate, then we classify the forecast as being good news; if it is lower, then we classify the forecast as being bad news; and if it is equal to the prevailing median analyst consensus estimate, then we classify the news as confirming. Approximately 46 percent of the management forecasts convey bad news, approximately 37 percent convey good news, and approximately 17 percent are confirming (see Panel C of Table 2). We define management forecast errors (MFE) as realized earnings per share less the management forecast of earnings scaled by lagged stock price on the third day after the prior period s earnings announcement. For our sample, the mean management forecast error is significantly negative implying that management forecasts are optimistic on average. This observation contrasts earlier work documenting that management forecasts are unbiased on average (McNichols 1989; Frankel et al. 1995; Kasznik 1999). Moreover, bad news forecasts are significantly more optimistic than good news forecasts, which is inconsistent with the view that bad news should be more credible than good news (see Panel D of Table 2). The distribution of management forecast errors by quintiles of forecast news highlights the fact that extreme news is more optimistically biased. Specifically, the mean management forecast error is -0.012 for quintile 1 (extreme bad news) and -0.007 for quintile 5 (extreme good news), but ranges from -0.004 to - 0.002 for the intermediate quintiles (see Panel E of Table 2). The importance of management forecasting reputation is likely to increase with the horizon of a forecast. In particular, investors are less likely to assess the credibility of a 15

forecast when the forecast has a short horizon because the actual earnings release is close at hand. In contrast, investors ought to assess reputation more carefully for forecasts with long horizons. Panel F of Table 2 shows that the mean (median) horizon for our sample of annual forecasts is 225.5 (209) days before the earnings announcement. Hence, focusing on annual as opposed to quarterly forecasts (which generally have horizons of approximately 30 to 45 days) ought to provide a more powerful setting in which to examine the role of reputation. Splitting the forecasts by type of news reveals that bad news is released on average 222.3 days before the earnings announcement and good news is released on average 233.4 days before the earnings announcement. The difference in forecast horizon for good and bad news forecasts is significant, suggesting that bad news forecasts are released later in the fiscal year. The sample used to test H3 and H4 differs from the full sample, described in Tables 1 and 2, used to test H1 and H2. First, we require stock return data around the earnings announcement dates (which reduces the sample from 9,381 forecasts to 9,362 forecasts). Second, we use only post Reg FD forecasts for these tests (which reduces the sample to 6,831 forecasts). 12 Third, we delete observations with no consensus analyst forecast available just before the earnings announcement (which reduces the sample to 6,731 forecasts). Fourth, we use only the last management forecast made for each fiscal period (which reduces the sample to 2,732 forecasts). Finally, data for our additional control variables (earnings-to-price ratios and the standard deviation of analysts forecasts prior to the earnings announcement) is missing for 219 observations, leaving a restricted sample of 2,513 observations for testing H3 and H4. 12 Inclusion of pre Reg FD observations leaves our results qualitatively unaffected. 16

4. VARIABLE DEFINITIONS AND RESEARCH DESIGN Management forecasting reputation can be measured in various ways. We develop a measure of forecasting reputation, similar to the theoretical construct for reputation in Stocken [2000], by assessing the accuracy and frequency of prior forecasting behavior. Chen, et al. [2005] offer an alternative model of investor learning that also requires investors to assess the accuracy and frequency of prior forecasting behavior albeit within the context of analyst quarterly earnings forecasts. Although these two models have similar ingredients, because they represent different environments, they reflect differing views about the credibility of forecasts. In Stocken [2000] the manager strategically forecasts earnings in a multi-period game. Therefore, the precision of each forecast is endogenous. In Chen, et al., [2005], in contrast, investors update their beliefs about an analyst s forecasting ability using signal realizations from a probability distribution with a precision that is exogenous; they assume the analyst is non-strategic and thus the forecast is always credible. These modeling differences imply fundamental differences in the expected behavior of investors. Most importantly, when the precision of each signal is exogenous, Chen, et al., [2005] show that as the number of analyst forecasts increase, investors place strictly more weight on an analyst s forecasting record and strictly less on their prior beliefs when forming posterior beliefs about the analyst s forecasting ability. Alternatively, when the precision of each signal is endogenous, Stocken [2000] characterizes an efficient equilibrium where investors follow a threshold strategy. That is, the manager s forecasting performance is evaluated at the end of a testing period: if the manager s forecasting record is sufficient and exceeds a threshold, then the manager is regarded as having a forecasting reputation and investors 17

impound the forecast information in the firm s stock price; alternatively, if the manager s forecasting record falls short of the threshold, then investors discount the forecast. In this equilibrium, the manager s forecasting reputation depends on the path of state realizations and is a function of the accuracy and frequency of the manager s forecasts. Forecasting accuracy is calculated as the ratio of annual forecasts that are deemed to be relatively accurate to the total number of annual forecasts the firm has issued. 13 Formally, we construct the forecasting Accuracy Ratio as follows: Number of Relatively Accurate Forecasts Total Number of Management Forecasts Issued ; A management forecast is deemed to be relatively accurate when the management forecast is strictly more accurate than the median consensus analyst forecast prevailing on the day the management forecast was released; i.e., Median Analyst Estimate Realized EPS > Management Forecast Realized EPS. The total number of annual forecasts a firm has issued is determined by counting the firm s annual earnings forecasts captured by First Call in its CIG database since the database s inception in January 1994. 14 Forecasting frequency is measured by counting the number of annual forecasts captured by First Call in its CIG database. The number of forecasts investors use to 13 We use a firm s annual earnings forecasting performance only to proxy for the firm s forecasting accuracy. We do not consider quarterly forecast performance when constructing the accuracy ratio. When forecasting quarterly earnings, we expect management to have a strong information advantage over analysts because for forecast with short horizons, the firm-specific information available to management is likely to be more influential than industry and macroeconomic information that analysts specialize in obtaining. This information advantage allows management to forecast quarterly earnings more accurately than analysts. Additionally, it is easier for management to manipulate quarterly earnings to meet their forecasts than annual earnings, which are audited. Thus, we believe that evaluating quarterly forecasting performance is less useful for assessing forecasting reputation. 14 While First Call s CIG database has some observations dated prior to January 1994, its coverage of firm forecasting behavior before January 1994 is generally incomplete. 18

evaluate management forecasting performance ought to increase with the uncertainty and volatility in the firm s information environment. When a firm s information environment is volatile, the earnings report is less useful for assessing the truthfulness of a forecast, and therefore, investors can increase the precision with which they evaluate management forecasting performance by considering more forecasts. Rogers and Stocken [2005] show that the standard deviation of the analysts forecasts captures the usefulness of the earnings report for assessing the truthfulness of a management forecast. In this light, if the standard deviation of the analysts forecasts prevailing at the time the management forecast is released (SD_AF_MF) is at least equal to the sample median standard deviation of the analysts forecasts, then we require a firm to have offered at least eight forecasts for investors to assess whether management forecasted truthfully. Alternatively, if the standard deviation of the analysts forecasts is less than the sample median standard deviation of the analysts forecasts, then the firm has to have made at least five forecasts. Coupling these measures of accuracy and frequency, a firm is defined as having a forecasting reputation if: (i) the firm s accuracy ratio is greater than the mean accuracy ratio for all other firms sharing the same 2-digit SIC; and, (ii) the firm has released at least five or eight annual earnings forecasts depending on the standard deviation of analysts forecasts. 15 15 We use these frequency thresholds because numeric analysis in Stocken [2000] suggests eight forecasts are a sufficiently number for a firm s reputation concerns to induce an efficient outcome. Nevertheless, our results are qualitatively unaltered if we require firms to have released five/three rather than eight/five forecasts. Further, our results are qualitatively unchanged when a firm is defined as having built a reputation for accurate forecasting if it has released at least eight earnings forecasts and its reputation ratio lies above the median for all firms that have released at least eight annual forecasts. Lastly, to calculate SD_AF_MF, we require three or more analyst forecasts. When assigning a firm s forecasting reputation, if we are unable to calculate SD_AF_MF, then the firm is regarded as not having a forecasting reputation. 19

Table 3 reports properties of the Accuracy Ratio for firms and industries in the full sample, the classification of sample firms using our reputation criteria, and the frequency of forecasting by firms with and without a forecasting reputation. Panel A of Table 3 highlights that on average management is more accurate than analysts only about half of the time. This average generally holds across all industries in our sample, as the industry mean ratio ranges from 49 percent to 58 percent from the first to the third quartile. Panel B of Table 3 reports the frequency with which observations in the full sample of 9,381 management forecasts meet each of the reputation criteria and also the classification of these annual forecasts into the Reputation and No Reputation subsamples. We observe that 756 firms issued 2,938 annual forecasts (or about 31 percent of the full sample) when they have an accuracy ratios greater than the mean for their industry, defined by the 2-digit SIC code. 16 Of the over 4,000 firms with a standard deviation of analysts forecasts falling in the upper half of the full sample, only 411 firms issued eight or more forecasts, which accounts for 1,502 observations (or approximately 16 percent of the full sample). For firms with a standard deviation of analysts forecasts falling in the lower half of the full sample, 659 firms issued five or more forecasts, which accounts for 2,470 observations (or approximately 26 percent of the full sample). 17 Coupling the reputation criteria, 414 firms are classified as having a forecasting reputation, and they released 1,630 forecasts after their reputation is built (or 16 The sample firm is excluded from the calculation of the mean accuracy ratio for its 2-digit SIC. 17 For firms issuing eight or more forecasts, nearly three years passed between the first and eighth forecast [mean (median) number of days = 1,054 (923)]. For firms issuing five or more forecasts, a little over two years passed between the first and fifth forecast [mean (median) number of days = 752 (622)]. 20

approximately 17 percent of the full sample); the remaining 7,751 forecasts are classified as having been issued by firms without a forecasting reputation. 18 Panel C of Table 3 presents the frequency tables for the number of annual forecasts issued by the full sample of 2,138 firms, the annual forecasts issued by the sub-sample of 414 Reputation firms, and the annual forecasts issued by the sub-sample of 2,130 No Reputation firms. 19 While this panel indicates forecasting frequency varies widely, it also indicates the frequency distributions are not much different for the sub-samples. Table 4 presents summary statistics of the management forecasts and various firm characteristics. Panel A of Table 4 reports the forecast news, the market response to these forecasts, as well as the management forecast error. Untabulated t-statistics indicate that the mean and median stock price response to good news is significantly positive and to bad news significantly negative; the mean and median stock price response to confirming news is insignificantly different from zero. Panel B of Table 4 partitions the full sample by reputation. This Panel has three noteworthy features: First, firms without a reputation forecast bad news that is significantly worse and good news that is significantly better compared to firms with a reputation. Second, despite the less extreme news, the median stock price reaction to good news is larger for firms with a reputation than for firms without a reputation. Third, management forecasts for no reputation firms are significantly more optimistic across all three news categories. Note that the optimism in these forecasts is not by design, as the Accuracy Ratios used to categorize forecasting 18 We recognize that the cutoffs used in constructing the reputation indicator variable, Reputation, are arbitrary. Accordingly, we conduct sensitivity analyses to ensure that our findings are not sensitive to these cutoffs. The tenor of our results is unaffected. See footnote 15 for details on the alternative cutoffs. 19 Note that 414 firms are classified as having a forecasting reputation and 2,130 firms are classified as having no reputation. The sum of these two numbers is greater than the total number of 2,138 firms in the full sample because forecasts issued before a firm has a forecasting reputation are included in the No Reputation sub-sample. Not knowing exactly when a firm s forecasting reputation is established may weaken the power of our empirical tests. 21

reputations are a function of the absolute forecast errors of prior forecasts only, not the current forecast. Thus, prior forecasting behavior is predictive of future forecasting behavior and our method for classifying forecasting reputations is able to identifying firms that provide less biased forecasts in the future. Panels C and D of Table 4 provide descriptive statistics for the continuous variables and discrete variables considered in the study, respectively. Firms with forecasting reputations issue forecasts later than firms without reputations, and they have a higher median book value to market value of equity ratio, denoted B-to-M t-1, implying that Reputation firms have fewer growth opportunities. 20 Firms with forecasting reputations tend to be larger with more total assets and higher market values. 21 No Reputation firms have larger (in absolute value) special items and smaller (in absolute value) total accruals than the Reputation firms. Additionally, the standard deviation of analyst forecasts, denoted SD_AF_MF, is significantly higher for the No Reputation firms at the management forecast date, but insignificantly different at the earnings announcement date, denoted SD_AF_ER. These differences help to explain why managers of the No Reputation firms are less able to provide accurate earnings forecasts. The descriptive statistics for the discrete variables in Panel D suggest firms with forecasting reputations released fewer point forecasts, fewer forecasts of losses, and more forecasts after the introduction of Reg FD. 20 A higher B-to-M ratio also suggests that these firms are less overvalued (Dechow et al. 2001) and potentially have more informative accounting reports (Tasker 1998). 21 The two groups of firms provide similar returns to their shareholders, as measured by ROE and ROA. 22

Table 5 provides Pearson and Spearman correlation matrices for the control variables. None of the correlations between our control variables are sufficiently high (greater than 30%) to raise concerns about multicollinearity. 22 4.1 Research design for Hypothesis 1 Hypothesis H1 posits that investors are more responsive to management forecasts when the issuing firm has a forecasting reputation. To examine how forecasting reputation affects the investor response to management forecasts, we estimate the following model for the full sample using a cross-sectional OLS regression (firm and time subscripts have been suppressed): MF_CAR -1,+1 = α 0 + α 1 Reputation + α 2 News Good + α 3 News Bad + α 4 Reputation News Good + α 5 Reputation News Bad + Control variables + ε. (1) The model s variables are defined as follows: Event Return: The market response to a management earnings forecast, denoted MF_CAR -1,+1, is the three-day event window return centered on the day of the management forecast. We report results using size-adjusted returns, where for stock listed on the NYSE or AMEX (Nasdaq), we calculate the size-decile return based on the CRSP NYSE/AMEX (Nasdaq) Capitalization Deciles. Further, MF_CAR -1,+1 is winsorized at the 1 and 99 percent levels. Untabulated results using raw returns and market-adjusted returns are similar. Forecast news: Forecast news, denoted News, equals the management earnings per share forecast less the median consensus analyst estimate prevailing on the day of the 22 Although the correlations between the two measures of the standard deviation of analyst forecasts are greater than.3, these two measures are not used in the same regression. 23

management forecast scaled by the stock price on the third day after the prior period s earnings announcement; News is winsorized at the 1 and 99 percent levels. The good news indicator variable, Good, equals one when News is positive and zero otherwise; the bad news indicator variable, Bad, equals one when News is negative and zero otherwise. Reputation: We proxy for a firm s forecasting reputation using the reputation indicator variable, Reputation, which is assigned a value of one for a firm that has a forecasting reputation and a value of zero otherwise. When determining forecasting reputation, we lag the accuracy ratio to ensure that it is observable to investors on the day of management s current forecast. In particular, the announcement date of the management earnings forecast occurs after the earnings release date for all of the prior forecasts used to calculate the accuracy ratio. Previous studies examining firms voluntary disclosure find good and bad news induce differing market responses; in addition, the extant literature typically regards good news forecasts as being less credible than bad news forecasts (e.g., Jennings 1987; Skinner 1994; Williams 1996; Hutton et al. 2003). Consequently, we allow the response coefficients to vary with whether a forecast contains good or bad news. Based on Hypothesis H1, we predict the coefficients on Reputation News Good and Reputation News Bad to be positive; i.e., α 4 > 0 and α 5 > 0. In contrast, if the market does not consider a firm s reputation when responding to a forecast or if the reputation indicator, Reputation, is an inappropriate proxy for a firm s reputation, then the coefficients on Reputation News Good and Reputation News Bad should be zero. 24

Control Variables: Several variables identified in previous studies as affecting forecasting behavior or response coefficients are introduced to control for cross-sectional differences in response coefficients. The variables, defined in Appendix A, control for forecast specificity (Baginski et al. 1993), forecast horizon (Johnson et al. 2001), growth (Bamber and Cheon 1998), predicted losses (Hayn 1995; Basu 1997), the effects of Reg FD (Heflin et al. 2003), size (Baginski and Hassell 1997; Bamber and Cheon 1998), management s reporting discretion (Kasznik 1999), and special items (Bradshaw and Sloan 2002). Lastly, we include fixed effects for each year. 23 We do not predict a sign on the coefficients of the control variables for forecast specificity, Point, and forecast horizon, Horizon, because these variables are endogenous to the decision to forecast. First, consider forecast specificity. Baginski et al. [1993] argue that managers reveal their uncertainty about earnings by issuing forecasts with wider ranges. Because Point equals one when a firm issues a point forecast and zero otherwise, one might expect the coefficient on Point News to be positive. However, in an environment where forecasting reputation is a concern, management might only issue point forecasts when they have the reporting flexibility to manipulate earnings and thereby meet the forecast. If investors hold these expectations about the firm, then they will be less responsive to point forecasts, and accordingly, the coefficient on Point News will be negative. Second, consider forecast Horizon, which equals the number of days between the firm s forecast and the earnings announcement (realization). Forecasts that are issued closer to the earnings announcement date are expected to be more precise (Johnson et al. 2001). Hence, one might expect a negative coefficient on Horizon 23 The results reported throughout the paper are not qualitatively different if we included fixed effects for industry, where industry is defined as all firms reported on Compustat sharing the same two-digit SIC code. 25