The Use of Revenue Disclosures. to Inform and Influence the Market

Transcription

1 The Use of Revenue Disclosures to Inform and Influence the Market April 2017 Lorien Stice-Lawrence University of North Carolina at Chapel Hill Stephen R. Stubben University of Utah We thank workshop participants at Nanyang Technological University, Singapore Management University, London Business School, Massachusetts Institute of Technology, and the 2015 Accounting and Audit Convergence Conference at Babes-Bolyai University. We also thank Elia Ferracuti, Won Kim,and Arthur Morris for excellent research assistance.

2 The Use of Revenue Disclosures to Inform and Influence the Market Abstract: This study addresses whether firms use revenue disclosures to inform and/or influence the market. We find that when firms provide a more thorough discussion of revenues in their 10- K filings, (1) analysts are more likely to issue revenue forecasts, (2) the revenue forecasts tend to be more accurate, and (3) investors place a greater weight on revenues in valuation. These findings are consistent with revenue disclosures providing useful information to the market about reported revenues. However, we also address the possibility that some firms use revenue disclosures to influence or mislead stakeholders. We find that although revenue disclosure is associated with a greater weight on revenues in valuation, it is not associated with more persistent revenues, suggesting the greater weight on revenues may be unwarranted. We also find that firms tend to discuss revenues less when they are managing revenues to meet reporting targets, consistent with an effort to misdirect investors and regulators.

3 1. Introduction This study addresses whether firms use revenue disclosures to inform and/or influence the market. A large literature in accounting has examined corporate disclosure and shows that disclosure can have significant effects on such outcomes as liquidity (Diamond and Verrecchia 1991, Leuz and Verrecchia 2000) and cost of capital (Botosan 1997, Lambert et al. 2012), among others. The combined body of evidence suggests that an increase in information disclosed by firms will benefit market participants. However, the prior literature has largely overlooked specific disclosure topics, instead focusing on more general measures, for example aggregate disclosure indices, management earnings guidance, or accrual quality measures. While useful, these measures ignore the fact that specific types of disclosure convey different types of information and may have different implications for financial statement users. In addition, the prior literature has generally focused more on whether disclosures are informative and less on whether they might be used to strategically influence investor behavior. In this study, we measure dimensions of one specific disclosure type revenue-related disclosure and demonstrate that while revenue disclosures are informative to financial statement users on average they may also be used strategically to influence stakeholders. As a proxy for revenue-related disclosure, we introduce a new measure that captures the amount of revenue-related information provided in 10-K filings. This revenue disclosure index counts the number of commonly discussed revenue-related topics included in each firm s 10-K filing, reflecting the amount of revenue-related information provided by the firm. As a second measure of revenue disclosure, we capture firms emphasis on revenues by counting the percentage of words in the 10-K that relate to revenues. Whereas the first measure proxies for the 1

4 extent and completeness of revenue disclosure, the second measure may simply pick up repetitive discussion that is not necessarily informative. We use these revenue disclosure measures to address both the informational role of revenue disclosures themselves and their role in influencing how financial statement users observe and interpret financial information. First, we assess whether revenue disclosures are informative to analysts by testing whether greater revenue disclosure is associated with a higher likelihood that analysts will forecast revenues and whether these revenue forecasts are more accurate. We find that greater revenue disclosure is associated with a higher likelihood of analyst revenue forecasts subsequent to the 10-K filing, even when controlling for a broad set of control variables and other general measures of corporate disclosure. Because we restrict our analysis to firms with earnings forecasts, variation in the existence of revenue forecasts is unlikely to be driven by more general factors in the firm s information environment that affect analyst coverage. We also find that an increase in unique revenue disclosures, though not an overall emphasis on revenues, is associated with more accurate subsequent revenue forecasts. Because we control for earnings forecast accuracy, this finding is more likely attributable to revenue disclosure than to more general earnings-related news. This finding for the revenue disclosure index is incremental to two measures of general disclosure, the length of the 10-K and the Fog Index (Li 2008), which supports the validity and usefulness of the revenue disclosure index as a measure of corporate disclosure. Further, as we include firm and year fixed effects in our regression analyses, the relation is unlikely to be driven by omitted factors that vary across firms or over time. 2

5 While these two findings are consistent with prior general findings on the association between disclosure and earnings forecast attributes (Lang and Lundholm 1996, Hope 2003), our focus on revenue disclosure and revenue forecasts (controlling for earnings forecast attributes) allows us to more directly link the information provided by firms and their capital market effects. Together, these findings are consistent with analysts being more willing to issue revenue forecasts for firms with greater available revenue information and with that information aiding analysts in providing accurate forecasts of revenues. Next, we test whether and how revenue disclosures influence investors. We find that greater revenue disclosure is associated with higher revenue response coefficients, incremental to earnings response coefficients. While a higher revenue response coefficient may be justified when revenue disclosure is high, either because the revenue figure is more credible or because the firm appropriately leads investors to place more weight on revenues, we provide further evidence that this higher revenue response coefficient is unwarranted. In particular, we find that the persistence of revenues does not vary with revenue disclosure levels, which suggests that the higher weight placed by investors on revenues when revenue disclosures are more detailed may not be justified. Finally, we find that firms discuss revenues less when they are managing revenues to meet reporting targets, consistent with these firms attempting to shift investors and regulators focus away from manipulated accounts. Our study makes several contributions to the disclosure literature. First, we show that specific disclosures (e.g., those related to revenues) provide specific information to the market that is not captured in more general disclosure measures such as 10-K length. These revenue disclosures also influence how the market prices the firm s financial information. Second, our results provide additional evidence supporting prior findings on the consequences of corporate 3

6 disclosure by documenting effects on analyst coverage and forecast accuracy; this supports the notion that revenue disclosures are informative to financial statement users and not just boilerplate. Third, we present a new measure of disclosure that overcomes some of the limitations of commonly used disclosure measures. Our measure is able to identify specific dimensions of disclosure, similar to studies that use hand-collected metrics (Botosan 1997, Francis et al. 2008), but it overcomes the issue of selection bias present in these studies because it is available for a broad set of firms (all those with 10-Ks posted on Edgar). It is also based on objective criteria and available each year. Finally, we add to the literature that addresses how firms use voluntary disclosures to influence market perceptions. While the majority of disclosure research considers the informational role of corporate disclosure, we also provide evidence consistent with disclosures being used to sway investors away from optimal decision making. Our study proceeds as follows. Section 2 summarizes related research and our contributions to the literature and motivates and states our primary predictions. Section 3 describes our measures of revenue disclosure, the models we use in our analysis, and the sample. Section 4 discusses the results, and Section 5 concludes. 2. Related Research and Predictions 2.1 Roles of Disclosure As discussed in Merkl-Davies and Brennan (2007), discretionary disclosures either (a) contribute to useful decision making by overcoming information asymmetries between managers and firm outsiders; or (b) constitute opportunistic behavior whereby managers exploit information asymmetries between them and firm outsiders through engaging in biased reporting, i.e., impression management (page 3). Regarding the latter case, Yuthas et al. (2002) argue that 4

7 managers use corporate reports to strategically manipulate the perceptions and decisions of stakeholders. Merkl-Davies and Brennan (2007) note that the majority of large-scale empirical studies on corporate disclosure have focused on the informational role of corporate disclosures. Much of the work to date on strategic disclosure, or impressions management, has been conducted outside the U.S., often using qualitative rather than quantitative methods. Our study uses a large-scale empirical approach that examines both informational and strategic roles of corporate disclosure. We start by examining the informational role of disclosure, which has been addressed by a variety of previous studies. For example, theoretical studies suggest that greater voluntary disclosure can affect information asymmetry, liquidity, and the cost of capital (Glosten and Milgrom 1985, Diamond and Verrecchia 1991, Lambert et al and 2012, Kim and Verrecchia 1994, Kondor 2012). Another strand of mostly empirical research on the consequences of disclosure examines more directly the information provided to the market, or the means by which information asymmetry is reduced. For example, Lang and Lundholm (1996) find that firms with more informative disclosure policies have larger analyst following, and more accurate and less disperse analyst earnings forecasts. Our study adds to this prior literature on the informational role of disclosure by documenting additional evidence of reductions in information asymmetry (i.e., improved analyst forecast accuracy) to firms with greater disclosure. Further, while much of the prior literature focuses on variation in disclosure in the cross section, our within-firm analysis provides more credible evidence on the effect of disclosure. Unlike most of the previous empirical research on corporate disclosure, our study also addresses the role of disclosure in affecting how financial statement users interpret financial 5

8 information. On the one hand, disclosure could help financial statement readers to more correctly use and interpret accounting numbers. On the other hand, disclosure could have the effect of drawing a disproportionate amount of attention to certain accounting numbers, potentially resulting in mispricing. Supporting the first theory, Drake et al. (2009) and Hirst and Hopkins (1998) find that better disclosure can lead to stock prices that more accurately reflect the relative persistence of accruals and cash flows and help financial statement users identify earnings management, respectively. Research on the strategic use of disclosure is limited. Several studies have examined the strategic use of a quantitative voluntary disclosure pro form earnings (e.g., Frederickson and Miller 2004, Bowen et al. 2005, Elliott 2006). Few studies have examined the strategic use of qualitative disclosures. One notable exception is Lang and Lundholm (2000), who find evidence that some firms opportunistically increase disclosure around seasoned equity offerings, potentially in an attempt to hype the stock. We attempt to add to this area of research using a larger set of firms in a more general setting than seasoned equity offerings. Overall, it is unclear whether increased disclosure of a specific type will increase overall informativeness or cause investors to fixate unnecessarily on certain pieces of financial information. We address this question in the context of the pricing of revenues of firms with high versus low revenue disclosure levels. 2.2 Empirical Measures of Disclosure One contribution of this study is the introduction of a new measure of corporate disclosure that has certain advantages over commonly used measures of disclosure such as selfconstructed measures, AIMR ratings (Lang and Lundholm, 1993, 1996), management forecasts (Baginski and Rakow 2012), length of the 10-K, and 10-K readability (Li 2008). As Healy and 6

9 Palepu (2001) and Beyer et al. (2010) discuss, these measures suffer from a variety of problems, including selection bias and generalizability, which limit their ability to reliably empirically document relations between disclosure and other determinants and outcomes. For example, hand-collected, self-constructed measures, such as those used by Botosan (1997) and Miller (2002), can be labor intensive to assemble. Given the high cost, these measures are usually used on small samples and are difficult to replicate. Similarly, AIMR ratings represent analysts subjective perceptions of disclosure quality, and are available only for large firms in the 1980s and 1990s. Finally, management earnings forecasts capture only disclosure relating to management s forecast of future earnings and are used by a small proportion of firms. Recent improvements in technology have opened the possibility of various machineconstructed disclosure measures. One of the most basic of these measures of disclosure is the length of the 10-K. However, this disclosure measure is ambiguous because it is not clear from the length alone whether a longer 10-K reflects greater disclosure or instead more complexity. The Fog Index (Li 2008) examines another aspect of the 10-K by examining complexity of disclosures, specifically the complexity of sentences (i.e., number of words per sentence) and the complexity of words (i.e., number of syllables per word). This measure has been used in a number of studies (e.g., Lehavy et al. 2011). However, Loughran and McDonald (2014) discuss issues of misspecification and measurement error in the Fog Index. They conclude that even a rough measure of 10-K filing file size is a proxy for readability that outperforms the Fog Index. Our revenue disclosure measures contribute to the literature along two dimensions. First, our revenue disclosure index overcomes many of the empirical difficulties encountered by disclosure measures in the prior literature. For example, this measure isn t as affected by 7

10 selection bias because it is available for every firm with an electronic 10-K filed on Edgar. Also, because the index is the sum of separate disclosure terms, researchers can easily interpret what drives the level of and change in disclosure. Second, because our measures are specifically associated with revenues, they allow us to examine the particular effects of revenue-related disclosure on outcomes and give us more confidence in the empirical links we document because they are directly tied with the nature of the disclosure we measure. Along these lines, Merkley (2014) finds that his measure of R&D disclosure is positively related to analyst-related outcomes such as earnings forecast accuracy but negatively associated with earnings disclosure, while Kravet and Muslu (2013) find that increased MD&A risk disclosures actually increase uncertainty and perceived risk. Both studies highlight the importance of considering the fact that not all disclosures have the same purpose or effects and suggest that more research is needed on the determinants and outcomes of specific disclosures, as they can potentially have different implications than more general measures. 2.3 Predictions We test whether revenue disclosures provide useful information to the market. Using analysts as a proxy for the market, we expect that if revenue disclosures are informative, then analysts would be more likely to issue revenue forecasts and those forecasts would be more accurate. These tests are similar to those in prior research that have examined associations between disclosure and properties of earnings forecasts (Lang and Lundholm 1996, Hope 2003), except that in these revenue tests we attempt to control for the general effects of earnings information and explore the incremental effect of revenue disclosure on revenue information. 8

11 First, we test whether analysts are more likely to issue revenue forecasts when revenue disclosure is higher. The decision made by an analyst to cover a firm is influenced by many factors (Bhushan 1989). Rather than attempt to identify and control for all possible determinants of an analyst s decision to cover a firm, we focus our revenue forecast analysis on the subset of firms that also have earnings forecasts. The decision to supplement an earnings forecast with a revenue forecast is a choice made by the analyst, and prior research indicates that analysts are more likely to issue an accompanying revenue forecast when they have better information (Keung 2010). We test whether this decision is associated with the firm s level of revenue disclosure and thus whether this disclosure is informative to analysts. H1: Revenue disclosure is associated with a greater incidence of revenue forecasts Next, we test whether revenue forecasts are more accurate for firms that have greater revenue disclosure. If revenue disclosures provide information to analysts that is useful in predicting future revenues, we expect to see a negative association between revenue disclosure and absolute revenue forecast errors. Because our disclosure measure may capture information in general, we isolate the revenue-related information by controlling for absolute earnings forecast errors in the test. H2: Revenue disclosure is associated with smaller revenue forecast errors Finally, we test whether revenue disclosure affects the extent to which the market uses revenue information in valuing the company. If revenue disclosures are informative, investors may have a stronger response to revenues because they are viewed as more credible. Or, more discussion of revenues in the 10-K could lead investors to fixate on revenues when valuing the firm. Either way, we expect to see higher revenue response coefficients for firms with greater revenue disclosure. 9

12 H3a: Revenue disclosure is associated with higher revenue response coefficients in valuation We utilize an additional test of revenue persistence to distinguish these two possible explanations. Because more persistent revenues are justifiably valued more highly, if revenue disclosure simply reflects that higher persistence, a higher weight on revenues in valuation may be justified. Otherwise, the higher weight on revenues in valuation may reflect investors being influenced by extensive revenue disclosures. H3b: Revenue disclosure is associated with higher revenue persistence Finally, we examine whether firms discuss revenues less when they are managing revenues to meet reporting targets. If firms wish to avoid unnecessary attention on their manipulated revenues by investors, regulators, or other stakeholders, they are likely to discuss revenues less in their 10-K. H4: Revenue disclosure is lower for firms suspected of managing revenues 3. Research Design and Data 3.1 Research Design Revenue Disclosure Measures Our first measure of revenue disclosure, REV_DISCL, is constructed by counting the number of revenue-related topics discussed in a firm s annual Form 10-K filed with the SEC. In order to identify these disclosure topics, we first identified a set of common revenue disclosure topics and the common phrases and vocabulary associated with them by having an RA manually read the 10-K disclosures of a sample of 25 training firms across several industries. This process allowed us to identify seven common revenue disclosure categories and a list of key phrases and 10

13 terminology that were associated with each disclosure topic (see Appendix B for a list of the specific disclosure categories and some example phrases). These are the revenue disclosures that we track in order to calculate REV_DISCL. 1 Although our measure is limited in that we will not identify revenue disclosure topics that did not appear in the 10-Ks of our training firms or which used substantially different terminology than our hand-collected sample, we believe that the sample of firms we chose in this initial exercise was sufficiently broad to capture common disclosure across a variety of industries. Even within our diverse sample of training firms, we found that revenue disclosure phrases and terminology were highly standardized. Additionally, we generalized our phrases in order to make them flexible enough to accommodate variation in disclosure across firms. Whether our final measure is a useful measure of revenue disclosure is an empirical matter, and any noise caused by the issues above is likely to prevent us from finding significant results. After compiling the list of phrases associated with each revenue disclosure type, we wrote a program in Perl that generalized these phrases and parsed Form 10-Ks to identify whether each disclosure type appeared in firms disclosure. For some disclosure categories, this was a relatively simple operation; for example, order backlog disclosure was easily identified by searching for whether the phrase backlog occurred in the 10-K. However, other categories were more difficult to identify, for example disclosures of revenue risk. In the end, we categorized disclosures where revenues or sales were discussed in close proximity to such words as risk, sustainability, or competition as revenue risk disclosures, but this ignores revenue risk disclosures related to specific other topics that do not use any of these identifying words. 1 We excluded a number of revenue phrases that pertain to certain activities not common across firms (e.g., deferred revenues, percentage-of-completion accounting, the effect of foreign exchange rates on revenues, revenue growth from acquisition activity, etc.) because non-disclosure of these items could either indicate poor disclosure or a lack of involvement in a particular activity. Instead, we focus only on items that could reasonably be disclosed by most of the firms in our sample. 11

14 Although our program had a high in-sample success rate for correctly identifying the specific phrases that we collected from annual reports without incorrectly classifying phrases into the wrong revenue disclosure category, there are undoubtedly limitations in the ability of our program to correctly identify the revenue categories we would like to track, especially when it encounters relevant phrases that were not in our training set. Again, this will decrease the ability of our measure to accurately capture the effects that we are trying to document. After parsing each 10-K and identifying which revenue disclosure categories are present, we calculate our final revenue disclosure measure, REV_DISCL, by counting the total number of revenue disclosure categories that are present in the firm s 10-K in that year. Because we identify seven unique revenue disclosure categories, the maximum possible value for REV_DISCL is 7 and the minimum is 0. Although some sections of the 10-K are specifically devoted to revenue disclosure (e.g. the accounting policy footnote has a section devoted to revenue recognition), we identify revenue disclosure present in any part of the 10-K because we are interested in all disclosure relating to revenues which is available to financial statement users. 2 If we assume that disclosure present in all sections of the 10-K will be reviewed by at least some financial statement users, then not including revenue disclosure which is found in non-revenue-specific sections would ignore information that is available to investors. We also measure a second dimension of revenue disclosure: the percent of words in the 10-K that relate to revenues, which we denote as PCT_REV. Specifically, we count the number of times the words revenue, revenues, or sales appear in the 10-K and divide that sum by the total number of words in the 10-K. Whereas REV_DISCL aims to capture the total number of 2 Peterson (2011) examines the disclosure about revenue recognition policies in the 10-K accounting policy footnote but focuses on the length of disclosure in that footnote only as a measure of the complexity of the firm s revenue recognition policies to predict future revenue restatements 12

15 a few key revenue-related topics covered in revenue disclosures (without being inflated by repetitive disclosures), PCT_REV captures the emphasis a firm places on revenue relative to other topics in the 10-K, including, potentially, repetition of a few items Regressions In order to establish the validity of our measures and demonstrate that they generally behave consistent with prior theoretical and empirical findings, we first examine the determinants of revenue disclosure. To do so, we estimate Equation (1): 3 REV_DISCL t or PCT_REV t = b0 + b1 SIZE t + b2 REV t + b3 INC t + b4 LOSS t + b5 BM t (1) + b6 LEV t + b7 R&D t + b8 MERGER t + b9 SPI t + b10 ISS t + b11 VOLE t + b12 LITIG t + Firm FE + Year FE + e t This specification investigates the effects of various determinants. First, we include a measure of firm revenues, REV, as firms with relatively higher revenues are naturally expected to discuss them more frequently. Lang and Lundholm (1993) find that disclosure increases with firm size and profitability, so we include several size and profitability measures: SIZE is measured as the natural log of the firm s year-end market value of equity, INC is net income before extraordinary items divided by beginning-of-year market value of equity, and LOSS is an indicator variable that equals one when INC is negative. The book-to-market ratio, BM, which is often used as a proxy for growth opportunities, and leverage, LEV, are also both likely to be associated with revenue disclosure. Research and development expense (R&D), calculated as annual R&D expense divided by year-end total assets, is argued to be positively related to proprietary costs, which have commonly been linked with the costs of disclosure (Verrecchia 1983, Hayes and Lundholm 1996, Wang 2007). MERGER is an indicator that equals one if the 3 With the exception of Equation (2), each equation is estimated using OLS with firm and year fixed effects and standard errors clustered by firm and year. As explained later, Equation (2) is estimated using a logistic regression. 13

16 firm completed an acquisition during the fiscal year (Botosan and Stanford 2000). SPI equals special items divided by total assets at the end of the year. ISS is the ratio of new debt and new equity issued to year-end total assets; Lang and Lundholm (2000) find that disclosure is higher when firms issue securities. VOLE is equity volatility, which captures the precision of information available to management; it is measured as the standard deviation of monthly logged stock returns during the fiscal year. In addition, as Skinner (1994) and Field et al. (2005) find that managers may provide earnings forecasts to avoid costly litigation, we include LITIG as a measure of litigation risk; it is an indicator variable that equals one if the firm operates in a litigious industry (following Francis et al. 1994). Finally, we include year and firm fixed effects to capture trends in revenue disclosure over time and unmodeled, time-invariant firm characteristics that may explain firms disclosure decisions, respectively. After examining the determinants of revenue disclosure, we test the effects of revenue disclosure on security analysts. Lang and Lundholm (1996) and Hope (2003) find that disclosure is associated with more accurate earnings forecasts. To our knowledge, no study has examined the association between revenue disclosure and revenue forecast likelihood and accuracy. Our first test addresses whether analysts are more likely to issue revenue forecasts for firms with greater revenue disclosures. I_RF is an indicator variable that equals one if at least one analyst issues a revenue forecast for the subsequent year during the seven days following the current year s 10-K filing date. To isolate the effect of specific incremental information provided about revenues, we estimate this logistic regression on the subset of firm-year observations where an earnings forecast is issued. That is, each firm in the subsample has an earnings forecast, and we assess the likelihood that a revenue forecast is also issued in connection with the earnings forecast. 14

17 I_RF t+1 = b0 + b1 [REV_DISCL t or PCT_REV t ] + b2 LENGTH t + b3 READABLE t (2) + b4 SIZE t + b5 REV t + b6 INC t + b7 LOSS t + b8 BM t + b9 LEV t + b10 R&D t + b11 MERGER t + b12 SPI t + b13 ISS t + b14 VOLE t + b15 LITIG t + Industry FE + Year FE + e t We include in Equation (2) the explanatory variables from Equation (1) above, which make the incremental effect of revenue disclosure on forecast likelihood, b1, independent of those firm characteristics that are associated with disclosure. 4 We also include two additional text-based measures of general disclosure to assess whether REV_DISCL and PCT_REV have incremental explanatory power over commonly used disclosure measures. LENGTH is the natural log of the number of words in the 10-K filing, and READABLE is financial statement readability, or (-1) x the Fog Index (Li 2008). 5 Finally, we include industry and year fixed effects to address the natural variation in revenue forecast propensity across industries and the increasing likelihood of revenue forecasts over time. 6 Our next test addresses whether revenue disclosure aids analysts in issuing these revenue forecasts by helping to improve their forecast accuracy. We measure revenue forecast accuracy using revenue forecasts for the subsequent fiscal year that are issued during the seven days 4 Healy and Palepu (2001) note that endogeneity is an important limitation of findings related to disclosure. For example, firms with high disclosure ratings tend to also have high contemporaneous earnings performance (Lang and Lundholm 1993). This may be caused by a self-selection bias firms may increase disclosure when they are performing well. As a result, the association between capital market variables and disclosure may be driven by firm performance rather than disclosure per se. In order to mitigate the effect of endogeneity on our results, our tests control for performance and other factors that influence disclosure, and also include firm fixed effects when appropriate. 5 In untabulated analyses, we include earnings quality as an additional control variable (Francis et al. 2008) and none of our inferences relating to revenue disclosure are qualitatively changed. We do not include earnings quality in our tabulated analyses for two reasons: (1) the five-year measurement period reduces the sample size and introduces survivorship bias and (2) VOLE has a strong negative correlation with earnings quality and captures a similar notion of information uncertainty. 6 Whereas we include firm fixed effects in all other (OLS) regressions, we use industry fixed effects in the estimation of Equation (2), which uses a logitistic estimation. Greene (2004) cites prior studies that have analyzed fixed effects in binary choice models and concludes that the fixed effects estimator is inconsistent and substantially biased away from zero when group sizes are small (e.g., firms). The bias diminishes as group size increases (e.g., industries). 15

18 following the current year s 10-K filing date. Using this short timeframe helps us to isolate the effect of information provided by 10-K disclosures and avoid confounding subsequently revealed information. Revenue forecast accuracy, RFE, is the absolute difference between the mean of revenue forecasts issued in the seven days following the 10-K filing and actual sales revenue, scaled by actual sales revenue. We include earnings forecast accuracy to the control variables to allow the assessment of revenue forecast accuracy independent of earnings forecast accuracy. Thus, we focus on forecasting accuracy as it specifically relates to revenues. RFE t+1 = b0 + b1 [REV_DISCL t or PCT_REV t ] + b2 LENGTH t + b3 READABLE t (3) + b4 EFE t+1 + b5 SIZE t + b6 REV t + b7 INC t + b8 LOSS t + b9 BM t + b10 LEV t + b11 R&D t + b12 MERGER t + b13 SPI t + b14 ISS t + b15 VOLE t + b16 LITIG t + Firm FE + Year FE + e t Again, because we control for determinants of disclosure, the coefficients on REV_DISCL and PCT_REV represent the incremental effect of disclosure on revenue forecast accuracy. In addition, with firm fixed effects, our coefficient estimates capture the effect of revenue disclosure on forecast accuracy using within-firm variation in disclosure; it does not capture cross-sectional variation in revenue disclosure, which is more likely to be correlated with omitted firm-specific factors. Next, we examine the pricing of revenues and income conditional on the level of revenue disclosure. If revenue disclosure leads the market to place relatively more valuation emphasis on revenues, then we expect to see a higher revenue response coefficient when disclosure is high. We use a simple model of returns as a function of income, with an incremental coefficient for revenues because revenues are likely to be valued differently than expenses 16

19 (Ertimur et al. 2003). The annual stock return, RET, is the monthly compounded stock return, beginning three months into the fiscal year and ending three months after fiscal year end. The three-month lag in computing the annual stock return helps ensure that the price response to 10- K disclosures is included in the return. Annual revenue, (REV) and annual income (INC) are deflated by beginning-of-year market value of equity, as above. An indicator variable for high revenue disclosure, DISCL_HI, is measured in one of two ways. First, it is set to one if disclosure is at or above the sample median of REV_DISCL. In a second analysis, DISCL_HI is set to one if PCT_REV is above the sample median. Because we include both revenues and income in the model, the coefficient on REV can be interpreted as the effect of revenues on returns, incremental to the coefficient on INC that applies to both revenues and expenses. A significantly positive coefficient on the interaction of REV and the high disclosure indicator variable, b3, indicates that investors place more weight on valuing revenues when revenue disclosure is high. RET t = b0 + b1 REV t + b2 INC t + b3 REV t x DISCL_HI t + b4 INC t x DISCL_HI t (4) + b5 DISCL_HI t + Firm FE + Year FE + e t Equation (5) replaces RET in Equation (4) with earnings in the subsequent year, scaled by market value of equity. This equation allows us to assess whether a difference in persistence of revenues in earnings prediction exists for firms with high versus low disclosure. 7 For example, if firms with high revenue disclosure also tend to have more persistent revenues, then a larger revenue response coefficient in Equation (4) is warranted and investors may be responding rationally to the signal of revenue persistence provided by revenue disclosures. Otherwise, a 7 Because of the autoregressive nature of Equation (5), our coefficient estimates may be biased due to the inclusion of firm fixed effects. Untabulated regressions using industry and year fixed effects or only year fixed effects lead to the same conclusions revenue disclosure is not significantly associated with higher persistence. 17

20 larger revenue response coefficient in Equation (4) could represent the market being unduly influenced by a focus on revenues in 10-K disclosures. INC t+1 = b0 + b1 REV t + b2 INC t + b3 REV t x DISCL_HI t + b4 INC t x DISCL_HI t (5) + b5 DISCL_HI t + Firm FE + Year FE + e t Finally, Equation (6) builds on Equation (1) by adding two variables and an interaction term intended to capture firms that are likely to have managed revenues to meet a reporting target. It includes a measure of discretionary revenues, DREV, and an indicator variable for firms more likely to have managed earnings and/or revenues to meet a financial reporting target, SUSP. The interaction of SUSP and DREV captures revenue disclosure by firms suspected to have manipulated revenues; it is an estimate of the association between discretionary revenues for firms that just met an earnings target. REV_DISCL t or PCT_REV t = b0 + b1 SUSP t + b2 DREV t + b3 SUSP t x DREV t (6) + b4 SIZE t + b5 REV t + b6 INC t + b7 LOSS t + b8 BM t + b9 LEV t + b10 R&D t + b11 MERGER t + b12 SPI t + b13 ISS t + b14 VOLE t + b15 LITIG t + Firm FE + Year FE + e t We calculate discretionary revenues, DREV, following the approach described in Stubben (2010). SUSP is an indicator variable that equals one if the firm reported earnings per share equal to or one cent greater than the consensus analyst forecast. We use the same set of control variables as Equation (1). 3.2 Sample Our sample spans from 1997 to 2013, and excludes regulated industries (Financial, Insurance, and Utilities). We use accounting data from Compustat, stock market data from CRSP, and analysts forecasts from I/B/E/S. We use 10-K filings on Edgar to calculate the 18

21 revenue disclosure index, 10-K length, and 10-K readability. After requiring data for all variables used in the analysis except revenue forecast errors, our sample contains 50,310 observations. 4. Results 4.1 Descriptive Statistics Table 1, Panel A, presents the level of and changes in the revenue disclosure index by disclosure topic. The first column shows the fraction of firm-year observations that discuss each of the 7 individual search terms. The most commonly discussed revenue items are revenue growth (98% of observations) and the source of that growth (95%). The least commonly discussed items are order backlog (42%) and revenue risk (48%). After summing the revenue terms, the average firm discloses 5.22 of the 7 items. Annual variation in disclosure exists across the component terms. The percentage of observations with changes in the presence of disclosure range from 2% (revenue growth) to 14% (revenue risk). The aggregate revenue disclosure index changes in 34% of firm-year observations. The relatively high frequency of annual changes supports the notion that the revenue disclosures are not merely boilerplate and copied over from the prior year s 10-K. Table 1, Panel B, presents annual changes in the revenue disclosure index level. For example, the median revenue disclosure index is 5 out of 7 items. Of the 13,198 observations with a revenue disclosure index of 5 in year t-1, 8,262 (63%) also have a revenue disclosure index of 5 in the following year and 12,658 (96%) have a revenue disclosure index that changes by no more than 1 in the following year. The small frequency of extreme changes in the index suggests that noise in the measure is relatively small. 19

22 Summary statistics for the variables used in subsequent analyses are presented in Table 2, Panel A. As discussed previously, the average firm discloses 5.22 out of 7 revenue items. The range of REV_DISCL from the first to the third quartile is 4 to 6 items. For the average firm, revenues represent 0.42% of the words included in the 10-K filing. Financial statement readability is multiplied by -1 (i.e., it is the inverse of the Li (2008) Fog Index) so that higher values represent more readable disclosure. Table 2, Panels B and C present means of select variables by year and by industry, respectively. Panel B reveals that revenue disclosure increases over time. REV_DISCL increases from 4.78 in 1997 to 5.52 in 2013, and PCT_REV increases from 0.38% to 0.44%. In contrast, financial statement readability (READABLE) does not exhibit such a strong trend. It increases slightly from in 1997 to by 2001 but then subsequently shows a gradual decrease. Overall, the different trends among disclosure measures suggest that they are not perfect substitutes and each is affected by different factors. Panel B also shows a steady increase in the frequency of revenue forecasts over time, consistent with Ertimur et al. (2011). This temporal trend underscores the importance of our inclusion of year fixed effects in our multivariate analyses. Likewise, the industry variation evident in Panel D supports our inclusion of industry and/or firm fixed effects. 8 The Pearson correlations presented in Panel D reveal some differences among the disclosure measures. REV_DISCL is positively correlated with PCT_REV (0.33), 10-K length (0.13) and readability (0.02), while 10-K length is negatively correlated with PCT_REV (-0.46) and readability (-0.40). The negative correlation between length and PCT_REV is likely due to the fact that PCT_REV is scaled by document length. Panel D also shows that REV_DISCL is positively correlated with the issuance of revenue forecasts (0.04) and negatively correlated with 8 We use industry classifications as defined by Barth et al. (2005). 20

23 absolute revenue forecast errors (-0.04). However, we base our inferences on the multivariate tests that follow. 4.2 Determinants of Revenue Disclosure Table 3, Panel A, presents the determinants of revenue disclosure. Determinants of the revenue disclosure index (REV_DISCL) appear in the first column of results. Consistent with Lang and Lundholm (1993) disclosure is positively associated with firm size (SIZE coefficient = 0.08, t-statistic = 5.73). However, contrary to Lang and Lundholm (1993), revenue disclosure is not positively associated with profitability (INC coefficient = -0.02, t = -0.91). In fact, revenue disclosure is slightly higher for loss firms (LOSS coefficient = 0.02, t = 1.80), suggesting that firms tend to discuss revenues more when profits are low. Revenue disclosure is positively associated with the extent of special items recognized (SPI coefficient = -0.07, t = -1.85, which is consistent with the argument that firms with financial difficulties tend to disclose less. Finally, disclosure is positively associated with leverage (LEV coefficient = 0.12, t = 2.45) and equity volatility (VOLE coefficient = 1.21, t = 6.27) but not significantly associated with the book-tomarket ratio, involvement in acquisitions, debt or equity issuance, or litigation risk. The second column of results relates to determinants of a firm s emphasis on revenues in the 10-K, PCT_REV. Larger firms tend to emphasize revenues less (SIZE coefficient = -0.01, t = -2.16), along with loss firms (LOSS coefficient = -0.02, t = -5.21), firms with high proprietary costs (R&D coefficient = -0.03, t = -3.12), firms involved in mergers and acquisitions (MERGER coefficient = -0.01, t = -2.34), firms issuing debt or equity (ISS coefficient = -0.01, t = -5.06), and firms with high volatility (VOLE coefficient = -0.05, t = -3.07). Firms emphasize revenues more when leverage is lower (LEV coefficient = -0.02, t = -2.74), and when special items are higher 21

24 (SPI coefficient = 0.02, t = 2.47). PCT_REV is not significantly associated with the book-tomarket ratio or litigation risk. Differences in results between the regression of REV_DISCL and PCT_REV are likely due to a denominator effect in PCT_REV. Whereas REV_DISCL is a raw index of disclosure related to revenues, PCT_REV captures the amount of revenue disclosure relative to total disclosure. So while events like mergers and acquisitions may not directly affect the amount of disclosure related to revenues, they may result in an increase in total disclosure that lowers PCT_REV. 4.3 Results Pertaining to the Informational and Influential Roles of Revenue Disclosure The next series of tests examines the outcomes associated with the two roles of revenue disclosure: the informational role, and its role in affecting financial statement users interpretation of financial information. Table 4 addresses the information provided about revenues to security analysts, Table 5 addresses the market s pricing of revenues, and Table 6 addresses the possibility that firms attempt to use revenue disclosures to shift investors and regulators focus away from the revenue account when it is being manipulated for financial reporting purposes. Table 4, Panel A, reveals that revenue disclosure is associated with a greater likelihood of revenue forecasts (REV_DISCL coefficient = 0.09, t = 2.44; PCT_REV coefficient = 0.97, t = 5.58). Because this analysis is conducted on the subsample of firms with earnings forecasts, this association is unlikely to be driven by the general conditions that lead an analyst to issue forecasts. Also, because the analysis includes both year and industry fixed effects, the association is unlikely to be driven by the general trends of increasing revenue disclosure and increasing frequency of revenue forecasts over time, or by variation across industries. 22

25 Column (1) indicates that revenue forecasts are not significantly associated with the length or readability of the 10-K, which is not surprising because these are measures of general disclosure not specifically related to revenues. The incidence of revenue forecasts is significantly positively associated with SIZE (coefficient = 0.06, t = 1.91), INC (coefficient = 0.24, 2.69), and VOLE (coefficient = 1.66, t = 2.43). It is marginally significantly negatively associated with R&D (coefficient = -0.43, t = -1.75). The coefficients on control variables are similar in Column (2). Results on the association between revenue disclosure and revenue forecast accuracy are provided in Table 4, Panel B. Whereas REV_DISCL has a negative and statistically significant association with absolute revenue forecast errors (REV_DISCL coefficient = -0.01, t = -3.11), PCT_REV is not statistically significantly associated with absolute revenue forecast errors (PCT_REV coefficient = -0.04, t = -1.50). Thus, while revenue forecast accuracy is higher when revenue disclosure is greater, as measured by the disclosure index, a simple emphasis and possibly repetitive discussion of revenues in the 10-K is not associated with greater forecast accuracy. In addition, Column (1) reveals that absolute revenue forecast error is significantly positively associated with EFE (coefficient = 0.09, t = 2.35) and ISS (coefficient = 0.03, t = 1.98). In both Column (1) and Column (2), the revenue disclosure variables appear to subsume any association between 10-K length or readability and revenue forecast accuracy. This is consistent with these revenue variables being less noisy proxies for corporate disclosure than two commonly used measures in the academic literature. Taken together, the results of Table 4 indicate that 10-K revenue disclosure is informative to the market, in particular securities analysts. Analysts are more likely to issue revenue forecasts 23

26 for high revenue disclosure firms, consistent with the finding from Keung (2010) that analysts are more likely to issue revenue forecasts when they have better information. We also find that revenue forecast errors are lower for high revenue disclosure firms, further indicating that analysts have more precise information for these firms. However, this result is obtained only for the revenue disclosure index, which measures disclosures about different topics, but not the emphasis on revenues in the 10-K that could simply be repetitive and non-relevant information. The result for the revenue disclosure index is particularly interesting given that it holds even when controlling for earnings forecast errors and thus, to some extent, the general information environment of the firm. It supports our intuition that specific firm disclosures have a specific informational role, instead of the vague notion that any general increase in information will lead to benefits in general outcomes. Table 5, Panel A, presents revenue response coefficients as they relate to revenue disclosure. The regression is based on a simple model of annual stock return on revenues and income an extension of the basic returns-earnings model that allows for incremental coefficients on revenues and then allows these coefficients to vary with an indicator for revenue disclosure that is at or above the sample median. In Column (1), which focuses on revenue disclosure measured as REV_DISCL, the coefficient on REV (coefficient = 0.04, t = 4.91) indicates the incremental revenue response coefficient when disclosure is below the median. This larger coefficient on revenues as compared to earnings is consistent with prior findings (Ertimur et al. 2003). When disclosure is above the median, the incremental revenue response coefficient is higher by 0.03 (t = 4.28). 24

27 The results presented in Column (2) indicate a similar pattern in revenue response coefficients when PCT_REV is above the sample median: the coefficient on revenues is higher (coefficient = 0.01, t = 2.00) when there is a greater emphasis on revenues in the 10-K. Given the results in Table 5 Panel A alone, it is unclear whether revenue disclosure is informing or biasing the market. If revenue information is relatively more informative for high disclosure firms, then the results we document in Panel A are a rational response by the market. However, if revenues are not an incrementally better predictor of future performance for high revenue disclosure firms, then this is evidence that revenue disclosures influence investors to put an undue weight on revenue information when valuing the firm. To distinguish between these two possibilities, Panel B shows that the actual persistence of revenues in earnings prediction does not vary with the level of revenue disclosure. The incremental coefficients on revenues when disclosure is high range are 0.00 in both regressions, with t-statistics ranging from 0.77 to This suggests that revenue disclosure itself, not just the underlying characteristics of the revenues of firms with high disclosure, is associated with greater market emphasis on revenues in valuation. Our final analysis examines a separate setting wherein firms may use revenue disclosures to shift attention away from attempts to manage revenues to meet financial reporting targets. Our focus is the coefficient on the interaction of SUSP and DREV, which captures revenue disclosure by firms suspected to have manipulated revenues. It is an estimate of the association between discretionary revenues for firms that just met an earnings target. In Column (1), the coefficient on SUSP and DREV is not statistically significantly associated with the revenue disclosure index, REV_DISCL (coefficient = 0.01, t = 0.02). However, in Column (2), the coefficient on SUSP and DREV is significantly negative 25