Working Paper Series Faculty of Finance. No. 6 Quality of earnings components and the joint issuance of analyst earnings and revenue forecasts

Transcription

1 Working Paper Series Faculty of Finance No. 6 Quality of earnings components and the joint issuance of analyst earnings and revenue forecasts Pawel Bilinski, Michael Eames

2 Quality of earnings components and the joint issuance of analyst earnings and revenue forecasts Pawel Bilinski Cass Business School, City University London EC1Y 8TZ, London Michael Eames Santa Clara University Santa Clara, CA * We thank Asad Kausar, Arthur Kraft, John O'Hanlon, James Ohlson, Ken Peasnell, Peter Pope, Lucio Sarno, Norman Strong, Steve Young, seminar participants at the 2012 European Accounting Association meeting, the 2013 Western Region Meeting, 2013 American Accounting Association meeting, and Cass Business School seminar participants for their comments and suggestions. 1

3 Quality of earnings components and the joint issuance of analyst earnings and revenue forecasts ABSTRACT: This study examines how quality of earnings components, i.e. the quality of revenue and of expenses, affects analysts decisions to supplement their earnings forecasts with revenue forecasts. We show that the accuracy of earnings forecasts is negatively affected by declines in the quality of revenue and expenses. However, revenue forecast accuracy is negatively impacted only by low quality revenue. Consequently, when expense quality is low, revenue forecasts become increasingly useful for valuing firms and investors reliance upon and demand for revenue forecasts increases. Analysts increase their reporting of revenue forecasts and the duration between revenue forecast revisions shortens in response to higher investor demand for these forecasts when the quality of expenses is low. Keywords: joint earnings and revenue forecast issues, earnings quality, earnings and revenue forecast accuracy, price reaction JEL Classification: M41, N20 Data Availability: Data are available from public sources indicated in the text. 2

4 I. INTRODUCTION Analysts routinely report their earnings forecasts to I/B/E/S and often complement these forecasts with revenue forecasts. 1 To date, only Ertimur et al. (2011) and Marks (2008) have examined the determinants of revenue reporting, with both studies investigating whether analyst reputation plays a role in explaining the decision to supply revenue forecasts with earnings-pershare (EPS) estimates. We consider the role of earnings quality components in this decision. We recognize that earnings quality reflects the qualities of both revenue and expenses and that low quality revenue and expenses increase analysts difficulties in forecasting earnings, which decreases the accuracy and value relevance of earnings forecasts (Bradshaw et al. 2001; Hughes et al. 2008). However, the accuracy and value-relevance of revenue forecasts is unaffected by low quality of expenses. Thus, when low quality expenses contribute to low earnings quality, investors relative reliance on and demand for revenue forecasts should increase. 2 Higher investor demand should then lead analysts to increase their revenue forecast reporting when the quality of expenses is low. 3 To examine the associations of the quality of revenue and expenses with analysts decisions to supplement their EPS forecasts with revenue forecasts, we consider I/B/E/S reported annual EPS forecasts and their accompanying revenue forecasts over the fiscal years Our analysis proceeds in two steps. First, we examine the associations of revenue 1 Of the 522,120 individual analyst earnings forecasts in our sample, close to 50% are supplemented by revenue estimates. 2 To illustrate, an investor can attach higher weight to forward sales multiples than earnings multiples in firm valuation. 3 Examining the relation between quality of earnings components and analyst revenue forecast reporting is particularly important given the positive association between earnings quality and analyst propensity to report cash flow forecasts in Bilinski (2013). Bilinski (2013) concludes that cash flow forecasts are not supplied by analysts when earnings quality is low because analysts are unable to accurately forecast accrual components when arriving at cash flow estimates. This conclusion is consistent with the findings in Givoly et al (2009), who argue that cash flow forecasts reflect simple depreciation adjustments to earnings, which makes them of limited value to investors. Eames et al. (2013) also investigate properties of cash flow forecasts and reach similar conclusions. In our study, we show that the observed relations between revenue forecast reporting and quality of earnings components are not driven by cash flow forecast issuance. 4 Consistent with Ertimur et al. (2011) and Marks (2008), we focus on the analyst s decision to provide earnings and revenue forecasts to I/B/E/S. We focus on I/B/E/S as our correspondence with the Thomson Reuters I/B/E/S support team suggests that I/B/E/S forecasts are likely to be timelier than forecasts contained in analyst research reports on Investext as they include forecasts from analyst daily briefs, research notes, and updates issued in addition to standard research reports. This means that investors are likely to pay close attention to the type and 3

5 and expense quality with (1) EPS and revenue forecast errors and (2) price reactions to EPS and revenue forecast revisions. If revenue forecasts accuracy is not affected by low quality expenses, but EPS forecast accuracy is, Bayesian investors will place relatively more reliance on revenue forecasts when expense quality is low and their demand for revenue forecasts will increase. In the second step, we examine the association of the earnings quality components with the likelihood of analysts reporting a revenue forecast with the EPS forecast. We expect a greater supply of revenue forecasts with EPS forecasts when expense quality is low in response to higher investor demand for these forecasts. We initially confirm that revenue and expense quality proxies are inversely associated with EPS forecast error, but expense quality is not associated with revenue forecast error. Further, we document that low quality expenses reduce the price reaction to EPS forecast announcements, but increase the price response to revenue forecasts. These results confirm that when low earnings quality is driven by low quality expenses, the value and weight investors attach to revenue forecasts increase. As the relative usefulness of revenue forecasts increases, so should investor demand for these forecasts, incentivizing analysts to supply more revenue estimates. In the second part of the study, we test if analysts are more likely to report revenue forecasts when expense quality is low. Our tests confirm that analysts propensity to report a revenue forecast with the EPS forecast increases when the quality of expenses is low, which is consistent with analysts responding to higher investor demand for revenue forecasts in this environment. This finding is robust to a battery of sensitivity tests. We confirm that our results are not driven by managements revenue guidance, higher demand for revenue forecasts among institutional investors, or by the lower propensity of more reputable analysts to report revenue quality of forecasts supplied to I/B/E/S. I/B/E/S gives analysts exposure to vast investor groups, including important institutional investors such as pension and mutual funds. Further, Ertimur et al. point out that access to I/B/E/S reduces information search and processing costs for investors by providing standardized forecasts directly comparable across analysts. Ertimur et al. also highlight that analyst ranking services such as StarMine use I/B/E/S to rate analysts and that analyst rankings matter for analyst career prospects and compensation (Hong et al., 2000; Hong and Kubik, 2003; Leone and Wu, 2007). 4

6 forecasts with their EPS estimates (Ertimur et al and Marks 2008). We also show that our conclusions remain unchanged when we use alternative measures of expense quality, and when we remove joint issues of earnings and cash flow forecasts from our sample. Finally, using the Cox (1972) survival model, we document that low expense quality shortens the duration between revenue forecast revisions. Our research contributes to the literature in three major ways. First, we add to the nascent literature that examines the analyst choice to complement an I/B/E/S reported EPS forecast with a revenue forecast. To date, only Ertimur et al. (2011) and Marks (2008) have considered factors contributing to the analyst s decision to supplement the earnings forecast with a revenue forecast, and both studies focus on analyst reputation as a contributing factor to this decision. Our results suggest that analysts rationally respond to investor demand and report revenue forecasts to I/B/E/S when these complementary forecasts are most useful to investors, i.e. when low expense quality reduces the accuracy and value-relevance of EPS forecasts, but not of revenue forecasts. Our results suggest that revenue forecasts can be useful in distinguishing the relative roles of revenue and expense management when earnings quality is low. Jointly reported forecasts of revenue and earnings enable investors to decompose the earnings forecast into expenses and revenue, and pay particular attention to the contribution of the latter, more persistent component to the earnings forecast (Ertimur et al. 2003). Second, we show that the introduction of controls for earnings quality components across analyst groups negates the earlier finding in Ertimur et al. (2011) and Marks (2008) that analyst reputation affects the analyst decision to report a revenue forecast. When we control for the quality of earnings components across high and low analyst reputation groups, we find no evidence that reputation affects revenue reporting. This suggests that the results in Ertimur et al. (2011) and Marks (2008) simply reflect the fact that more reputable analysts tend to follow firms with higher earnings quality. Further, our findings help explain the results in Ertimur et al. (2011) and Keung (2010) that jointly issued EPS and revenue forecasts have a greater price 5

7 impact than stand-alone EPS forecasts. This finding is not surprising since analysts supplement EPS with revenue forecasts to compensate for the lower accuracy and investment relevance of EPS forecasts analysts produce when quality of expenses is low. Third, our study responds to the call in Bradshaw (2011, 24 25), who in his review of the analyst forecasting literature argues that the "furthering [of] our understanding of what analysts do and why they do it requires consideration of their portfolio of activities and that the easiest means for gaining additional understanding of what analysts do is to examine their outputs beyond earnings forecasts. Further, our finding that the quality of expenses plays an important role in the analyst decision to report a revenue forecast adds valuable new insights regarding the capital market effects of firm earnings quality (Healy and Palepu 2001; Dechow et al. 2010), and the dynamic interaction between earnings quality components and information production by analysts (Barth et al. 2001a; Francis et al. 2002; Frankel et al. 2006; Beyer et al. 2010). In particular, the finding that analysts issue revenue forecasts to compensate for the adverse effect of low quality expenses on the accuracy and value-relevance of EPS forecasts adds important new evidence to the literature on the role of financial analysts as information producers in capital markets (Ivkovic and Jegadeesh 2004; Asquith et al. 2005; Ramnath et al. 2008; Chen et al. 2010). II. PREVIOUS LITERATURE AND HYPOTHESIS DEVELOPMENT Earnings is considered a better indicator of firm performance than other accounting numbers, such as revenue (Hopwood and McKeown 1985; Hoskin et al. 1986; Easton et al. 1992; Beyer et al. 2010; Dechow et al. 2010). However, low earnings quality can reduce the reliability and thus usefulness of analysts EPS forecasts. Bradshaw et al. (2001) and Hughes et al. (2008) employ accruals based measures of earnings quality and observe an inverse association between earnings quality and subsequent analyst EPS forecast errors. Bradshaw et al. (2001, 46) conclude that, sell-side analysts lack the necessary sophistication to understand the future implications of high 6

8 levels of accruals, consistent with analysts finding it more challenging to forecast earnings when their quality is low. 5 Revenue is a major component and determinant of earnings. Studies considering the value-relevance of revenue are limited and typically consider revenue s value-relevance simultaneously with that of earnings. 6 Chandra and Rao (2008) find that revenue has incremental information content to that contained in earnings. Further, they report that the value relevance of revenue increases and that of earnings declines for extreme earnings surprises at quarterly earnings announcements. Kama (2009) identifies specific environments, such as high R&D spending and industries with oligopolistic competition, where the market reaction to earnings surprises is not greater than to revenue surprises. Ghosh et al. (2005) find that investors rely more on growth in revenue than in earnings when valuing firms that exhibit continuous increases in both earnings and revenue. Jegadeesh and Livnat (2006) and Ertimur et al. (2003) show that investors value a dollar of revenue surprise more highly than a dollar of expense surprise and that the differential varies with the relative persistence of revenues and expenses. Only two studies have considered factors contributing to sell-side analysts reporting revenue forecasts to I/B/E/S. Marks (2008) and Ertimur et al. (2011) focus on the role of analyst reputation, as captured by analyst forecasting experience, membership on Institutional Investor s All-America Research Team, and the prestige of the analyst s broker, and find that low reputation analysts are more likely than high reputation analysts to report revenue forecasts. They contend that low reputation analysts report revenue forecasts to build their reputations, and that high reputation analysts refrain from reporting revenue forecasts in order to protect 5 The term earnings quality has been broadly used to reflect investors abilities to assess the sources of and probability of recurrence of net income (see Securities and Exchange Commission Accounting Series Release No.159). 6 Accounting research has a long history of studies on the value relevance of earnings (see Kothari 2001; Holthausen et al. 2001; Barth et al. 2001b for reviews of this literature). This research generally finds a positive association between the value relevance of earnings and earnings quality (Bao and Bao 2004; Cahan et al. 2009; Ecker et al. 2006). 7

9 their reputations, since joint revenue and EPS forecast issues can reveal the sources of inaccuracies in analysts EPS forecasts. While previous studies have approached the revenue reporting issue from the standpoint of analysts reputation, we focus on the impacts of revenue and expense quality. We propose that low quality expenses negatively impact analysts earnings, but not revenue, forecast accuracy and value-relevance. Consequently, investors reliance on and demand for revenue forecasts increases when the quality of expenses declines. If analysts respond to this higher investor demand for revenue forecasts, the supply of these forecasts should increase when the quality of expenses is low. Thus, our research hypothesis is: HYPOTHESIS: Analysts are increasingly likely to report revenue forecasts with their EPS forecasts to I/B/E/S when the quality of reported expenses declines. III. DATA We collected annual one-year-ahead EPS and revenue forecasts issued for fiscal years , together with their actual values from the I/B/E/S detail files. We obtained financial statement information from Compustat and stock price information from CRSP. Our sample includes 522,120 EPS forecasts and 261,084 complementary revenue forecasts, by 6,941 analysts and 540 brokerage houses for 2,552 firms. For comparability with EPS estimates, we scale revenue forecasts by the number of shares outstanding at the end of the forecast issue month. Panel A of Table 1 presents sample counts of stand-alone earnings forecasts and of joint earnings and revenue forecasts by fiscal year. [Insert Table 1 around here] The number of EPS forecasts increases from 43,565 in fiscal year 2000 to 66,887 in The number of joint EPS and revenue forecasts increases almost ten-fold, from 5,051 in 2000 to 8

10 46,063 in Thus, the annual percentage of joint revenue and EPS forecasts increases from 11.6% in 2000 to 68.9% in Panel B of Table 1 shows the relative frequencies of stand-alone and revenue accompanied EPS forecasts across 11 industries based on the 2-digit I/B/E/S SIG code. Standalone EPS forecasts are most common in the energy sector and in the public utilities industry. These are mature industries with business models where earnings forecasts are more likely to accurately anticipate firm performance. Revenue forecasts are most common in the technology and healthcare industries. Disaggregating earnings for these industries into revenues and expenses increases the transparency and interpretability of earnings numbers. This is consistent with Bowen et al. (2002), who argue that investors rely on revenue rather than earnings in valuing technology firms. IV. REVENUE AND EXPENSE QUALITY AND THE ACCURACY AND VALUE RELEVANCE OF FORECASTS This section presents our measures of revenue and expense quality and the research methods we use to test the associations of these earnings quality components with (1) EPS and revenue forecast errors and (2) the value relevance of EPS and revenue forecasts. We argue that low quality revenues increase revenue and EPS forecast errors, and lower the value relevance of revenue and earnings forecast revisions. Low quality expenses increase EPS, but not revenue forecast errors, lowering the value relevance of EPS, but not revenue forecasts. Collectively, these results suggest an increase in the relative usefulness and thus demand for revenue forecasts when the quality of expenses is low. Revenue Quality We use the receivables accrual model in Stubben (2010) to obtain our proxy for revenue quality. Stubben (2010) models changes in account receivables ( AR it) as a function of changes in 9

11 revenue ( R it). To control for the impact of firm credit policies on accounts receivable, he includes the interaction terms of revenue change with the firm s financial strength, operational performance relative to industry competitors, and the firm s stage in the business cycle. The log of firm total assets (ln Assets it) proxies for firm financial strength and together with the firm s age reflects the firm s stage in the business cycle. Industry median-adjusted growth rate in revenue (GRR_P it if positive and GRR_N it if negative) and industry median-adjusted gross margin (GRM it) capture the firm s operational performance. The receivables model then takes the form: 2 it 1 it 2 it it 3 it it 4 it it 2 5 Rit GRR _ Pit 6 Rit GRR _ Nit 7 Rit GRMit 8 Rit GRMit it AR R R ln Assets R ln Age R Age (1) where 2 Age it and 2 GRM it control for non-linear associations of age and gross margin with credit policies. Changes in revenue and account receivables are scaled by average total assets. The model s residuals, ν it, measure discretionary receivables. We measure revenue quality as the standard deviation of discretionary receivables for years t 4 to t (RevQ). We estimate equation (1) annually for each 2-digit SIC industry with a minimum of 20 firms. Quality of Expenses We calculate expense quality from the decomposition of earnings quality into quality of revenue and expenses. Following previous studies, we use an earnings quality measure based on discretionary accruals (Johnson et al. 2002; Aboody et al. 2005; Francis et al. 2004, 2005; Biddle and Hilary 2006; Core et al. 2008; Biddle et al. 2009; Elliott et al. 2010). Specifically, discretionary current earnings accruals are the residuals from the McNichols (2002) extension of the Dechow and Dichev (2002) current accruals model. As in Francis et al. (2005), the model takes the form CA CFO CFO CFO R PPE u (2) it 0 1 it 1 2 it 3 it 1 4 it 5 it it where CA it stands for current accruals for firm i in year t, defined as the change in current assets, less change in cash, less change in current liabilities plus the change in debt in current liabilities 10

12 and CFO is cash flow from operations. Both CA it and the three CFO variables are scaled by the average of total assets for the current and previous fiscal year and on their own form the Dechow and Dichev (2002) model. McNichol s (2002) extension adds the gross value of property plant, and equipment, PPE, and changes in firm sales, ΔR, both scaled by average assets, to the Dechow and Dichev s (2002) model. Including PPE and ΔR decreases the measurement error and improves the model s explanatory power (McNichols 2002; Francis et al. 2005). We estimate equation (2) annually for each 2-digit SIC industry with a minimum of 20 firms. The residuals, u it, provide an estimate of discretionary current accounting accruals. To obtain a measure of the quality of expenses, we subtract the discretionary receivables residuals computed in equation 1 from the discretionary current accounting accruals residuals from equation 2, and employ this as a proxy for discretionary expense accruals. We then measure the firm s current quality of expenses as the standard deviation of the discretionary expense accruals for the years t 4 to t (ExpQ). Revenue and expense quality and forecast errors We use the following model to examine the associations of revenue and expense quality with the accuracy of analyst EPS and revenue forecasts: Forecast error ln RevQ ln ExpQ Controls u (3) ijt 0 1 it 2 it ijt where Forecast error ijt is either the EPS or the revenue forecast error by analyst j for firm i issued at time t. The current fiscal year s annual absolute EPS forecast error, EPS FE, is computed as the absolute difference between the forecasted and actual EPS, scaled by the stock price at the end of the previous fiscal year. The corresponding revenue forecast error, REV FE, is computed as the absolute difference between the forecasted and actual revenue, scaled by the product of the end-of-month number of shares outstanding and the stock price at the end of the prior fiscal year. ln indicates the logarithmic transformation of a variable. The coefficient φ 1 captures the association of revenue quality with analyst earnings and revenue forecast errors, and 11

13 the coefficient φ 2 captures the association of the quality of expenses with earnings and revenue forecast errors. Controls include control variables relating to analyst, broker, and firm characteristics. Analyst- and broker-related characteristics We include earnings forecast horizon, Horizon, to reflect the observed increase in EPS forecast accuracy for earnings forecasts issued later in a fiscal year (Sinha et al. 1997). We measure Horizon as the number of days between the EPS forecast announcement and the respective fiscal year-end. We expect analysts with less forecasting experience and analysts with access to fewer resources at the broker to have more difficulties accurately forecasting earnings and revenues. We use analyst general experience (Experience), measured by years in the profession and the size of the analyst s brokerage firm (Broker size), measured as the number of analysts employed by the broker that issued at least one EPS forecast in the previous 12 months to capture analysts forecasting abilities and resources available to the analysts, respectively (Mikhail et al. 1997; Clement 1999). The number of firms an analyst follows (#Firm followed) proxies for task complexity. Actively following and producing research reports on many companies can reduce the time an analyst devotes to producing accurate earnings and revenue forecasts. We measure analyst and broker characteristics at the EPS forecast issue date. Firm characteristics We use market capitalization (MV) and the number of analysts following a company (Analyst following) to capture the quality of the firm s information environment. Higher quality information environments reduce information acquisition costs and should result in more accurate EPS and revenue forecasts. We also include the coefficient of stock price variation (COV), which measures the daily stock price volatility in the last 90 days of the previous fiscal year, and the standard deviation of revenue over the previous four fiscal years (STD REV). 12

14 High values of both measures suggest a more challenging forecasting environment and should negatively correlate with EPS and revenue forecast accuracy. We use the book-to-market ratio (B/M) to proxy for the firm s growth opportunities, which can make forecasting more difficult. We include firm age (Age) as analysts can find forecasting earnings and revenue easier for firms with longer time-series of financial information (Li et al. 2009; Ertimur et al. 2011). We include a dummy variable (Dloss) to indicate loss-making firms, and return on assets (ROA) to capture firm profitability. We expect analysts to produce more accurate forecasts for more profitable and non-loss making firms. We measure net margin (Margin) as the ratio of net income to revenue. Analysts may devote more effort to producing accurate revenue forecasts for firms with high net margin as revenue will have a higher effect on the bottom line net income. To control for solvency and firm distress risk, we include a measure of firm financial leverage (LEV), equal to the ratio of long-term debt to total assets. High distress risk can increase analyst forecasting difficulty and, consequently, revenue and EPS forecast errors. We include a set of year dummies (Year dummies) and industry dummies (Industry dummies) based on 2- digit SIG codes from I/B/E/S to control for year- and industry-effects. 7 All firm characteristics, but analyst following, are measured at the end of the previous fiscal year. Analyst following is measured at the EPS forecast issue date. All variables are winsorized at the 1% level. The full model specification is: Forecast error ijt ln RevQ ln ExpQ ln(1 Horizon ) ln Experience 0 1 it 2 it 3 ijt 4 ijt ln Broker size ln# Firm followed ln MV ln Analyst following 5 ijt 6 ijt 7 it 8 it 9COVit 10STD REVit 11B / Mit 12 ln Ageit 13ROAit Dloss Margin LEV Industry dummies Year dummies u 15 it 16 it 17 k 28 k ijt k 0 k 0 it (4) 7 I/B/E/S SIG code is a six-digit code, representing the sector (2-digits), the industry (2-digits) and the firm s operating group (2-digits). Year dummies are based on the calendar year of the EPS forecast announcement. 13

15 For forecast horizon, we use ln of 1 plus the variable to account for zero values. 8 The coefficients φ 1 and φ 2 capture the associations of revenue and expense quality, respectively, with forecast errors. Regression standard errors are clustered by analyst and firm to control for serial dependence of observations (Petersen 2009). 9 Appendix I provides detailed definitions of our variables. Table 2 presents descriptive statistics for the variables in equation (4), and an indicator variable, DREV, equal to 1 if an earnings forecast is accompanied by a revenue forecast, and 0 otherwise. Panel A shows that half of all EPS forecasts are accompanied by a revenue forecast. Mean RevQ is and mean ExpQ is Panel B reports that the mean absolute EPS and revenue forecast errors are 1.35% and 4.58% of the stock price, respectively. 10 Panel C shows that the average EPS forecast is issued in mid fiscal year. The average analyst is present in our sample for 6.4 years, follows 13.6 firms, and is employed by a brokerage house with 56 analysts. Panel D shows descriptive statistics for our firm-related characteristics. The average firm has a market capitalization close to $3.9 billion, and is followed by 10.5 analysts. The mean coefficient of stock price variation is 0.087, the mean standard deviation of asset scaled revenues is 0.147, and the mean book-to-market ratio is Average firm age is 21 years 11, mean ROA is close to 2% and the median margin is 4.5%. Nearly 18% of firms reported a loss in the previous fiscal year and mean financial leverage is 18.1%. [Insert Table 2 around here] 8 We employ log transformations of variables to ensure more normal distributions of the variables and to account for the possibility of a diminishing impact on the dependent variable. To illustrate, we expect that the increase in forecast accuracy due to an increase in broker size by one analyst should be higher for small than for large brokerage houses. 9 Ertimur et al. s (2011) regressions do not use dual-clustered standard errors to control for cross-sectional dependence of observations, thus the significance of their coefficient statistics is likely to be overstated. 10 Higher average revenue than earnings forecast error reflects that on average (price-scaled) revenue-per-share is15 times larger in magnitude than price-scaled earnings. This means that a one percentage error in a revenue estimate will be on average larger in magnitude than a one percentage error in an earnings forecast. In this study we focus on examining relative accuracy of revenue compared to earnings forecasts and on the relative value-relevance of revenue forecasts compared to earnings estimates with respect to changes in earnings quality. 11 We use CRSP files starting in 1926 to calculate firm age, which includes a few large companies with a long timeseries of stock prices. This explains the high mean firm age. 14

16 As a simple test of the associations of revenue and expense quality with EPS and revenue forecast accuracy, Figure 1 plots mean EPS and revenue forecast errors across decile portfolios of our revenue and expense quality measures, RevQ and ExpQ. Figure 1.a employs RevQ and shows that the mean forecast errors for stand-alone EPS forecasts, revenue accompanied EPS forecasts, and revenue forecasts increase as revenue quality declines. This is consistent with our prediction that low revenue quality has a negative effect on the accuracy of both revenue and earnings forecasts. Figure 1.b replicates the analysis for expense quality deciles. Here we find similar trends in EPS, but not revenue, forecast errors to those observed in Figure 1.a. Mean revenue forecast error is relatively flat or declining in moving from the highest to the lowest decile of expense quality. This supports the prediction that revenue forecast accuracy is not diminished by low expense quality. [Insert Figure 1 around here] Table 3 reports results from regressing EPS and revenue forecast errors on our proxies for the quality of revenues and expenses. To ensure comparability of results, we focus on the sample of 261,084 joint EPS and revenue forecast observations for both the EPS and the revenue forecast error regressions. [Insert Table 3 around here] The results in Table 3 confirm that absolute EPS forecast errors increase when revenue (p=.077) and expense (p=.006) quality decline. 12 We find a significant increase in absolute revenue forecast error when revenue quality falls (p=.000), but no significant association of absolute revenue forecast errors with the quality of expenses (p=.384). Thus, when expense quality is low and negatively impacting earnings forecast errors, revenue forecast errors remain unaffected. Consequently revenue forecasts can be particularly useful and in demand for valuing firms when expense quality is low. 12 In unreported results we find qualitatively similar results for the full sample of 522,120 EPS forecasts. 15

17 Revenue and expense quality and the value relevance of forecasts To test if the relative value relevance of revenue vs. EPS forecasts increases when the quality of expenses is low, we consider how the price reactions to percentage earnings (ΔEPS) and revenue (ΔREV) forecast revisions vary with the quality of revenue and expenses. Our model has the form: CAR EPS REV DREV EPS RevQ EPS ijt 0 1 ijt 2 ijt 3 ijt ijt 4 it ijt RevQ REV ExpQ EPS ExpQ REV Upgrade 5 it ijt 6 it ijt 7 it ijt 8 ijt 9 Downgradeijt uijt. (5) We use a three-day cumulative abnormal return (CAR) centered on the EPS forecast announcement date to measure the price response to EPS and revenue forecast revisions. 13 We expect the coefficients on ΔEPS and ΔREV to be positive if earnings and revenue forecasts have incremental information content to one another. To test if the value relevance of EPS and revenue forecast revisions vary with revenue and expense quality, we introduce the following interaction terms for revenue and expense quality with revenue and earnings forecast revisions: RevQ*ΔEPS, RevQ* ΔREV, ExpQ*ΔEPS and ExpQ *ΔREV. In Table 3 we have shown that revenue quality is negatively associated with both revenue and earnings forecast errors. Consequently, we anticipate a decline in the value relevance of both revenue and earnings forecasts with a decline in the quality of revenues, and thus negative coefficient estimates for RevQ*ΔREV and RevQ*ΔEPS. We have also shown that as expense quality falls we have an increase in earnings, but not revenue forecast errors. Thus, as the quality of expenses falls, we expect a decline in the value relevance of earnings, but not revenue forecast errors (i.e. a negative coefficient estimate for ExpQ *ΔEPS, but not for ExpQ* 13 We use the CRSP value-weighted index as the benchmark to measure abnormal returns. Similarly to Keung (2010), we assume that the revenue forecast revision is zero for stand-alone EPS estimates. We require that the forecasts used to calculate revisions are no more than 300 days apart and that the revisions in EPS and revenue forecasts are for the same fiscal year. The former criterion eliminates infrequently revised forecasts and the later ensures forecast revisions reflect only analyst new information for a fiscal year. These additional selection criteria reduce the sample size to 331,543 observations. 16

18 ΔREV). Together, such results would suggest a relative increase in the value relevance of revenue forecasts when the quality of expenses declines. Our set of controls in equation (5) includes an interaction term between ΔEPS and our indicator variable for a joint revenue and earnings forecast (DREV*ΔEPS). This captures whether the mere presence of a complementary revenue forecast lends credibility to the earnings forecast revision (Keung 2010). 14 To control for the effect of a simultaneous directional revision in the analyst s stock recommendation, we include two dummy variables for directional recommendation revisions, Upgrade, and Downgrade. Regression standard errors are clustered by analysts and firms. 15 Table 4 reports regression results for equation (5). Without controlling for the effect of revenue and expense quality on the price reactions to earnings and revenue forecast announcements (column Without interactions ), we find a positive and significant coefficient on EPS forecast revisions, consistent with previous literature (e.g. Sinha et al. 1997; Francis and Soffer 1997). Revenue forecast revisions have incremental information for pricing, and the market seems to react more strongly to a percentage revision in a revenue than an EPS forecast. A one percent increase in the revenue forecast per share increases the abnormal returns by 18 basis points. A similar increase in the EPS forecast increases the abnormal returns by only three basis points. As in Keung (2010) and Baginski et al. (2004), we find that the presence of a revenue forecast has an incremental effect on the price response to an EPS forecast revision. Specifically, the presence of a revenue forecast increases the return impact of an earnings forecast revision by over 50%. The evidence in Table 4 suggests that revenue forecasts contain incremental information to EPS forecasts and stock recommendations, and that the presence of a revenue forecast adds value to analyst EPS forecasts. 14 For example, the revenue forecast can aid investors in verifying the quality of the earnings forecast by disaggregating it into (more persistent) revenue and (less persistent) expenses. 15 For brevity, we do not report descriptive statistics for model (5). We find that the mean CAR is 0.16%, which is consistent with the mean earnings and revenue forecast revisions of 3.93% and 0.13%, respectively. The negative average values for CARs, and for earnings and revenue forecast revisions are largely due to the inclusion of the recent financial crisis in the sample period. 17

19 [Insert Table 4 around here] The results in Table 4 indicate that low quality expenses are associated with a significantly lower (higher) value relevance for earnings (revenue) forecasts. The coefficient estimates for ExpQ*ΔEPS and ExpQ*ΔREV are significantly negative (p=.002) and positive (p=.032), respectively. We find no significant association of revenue quality with the value relevance of revenue (p=.301) and earnings (p=.794) forecasts. These results confirm a lower value-relevance for earning s forecasts and a higher value-relevance of revenue estimates when earnings quality is low. The results in Table 4 thus far assume that revenue forecast revisions are zero for standalone EPS estimates. This may bias the coefficient estimates on revenue forecast revisions and their interactions with the earnings quality measures towards zero. To address this, the last columns of Table 4 report results for a smaller sample of 174,282 simultaneous EPS and revenue forecast revisions. Here we again find significantly negative and positive coefficient estimates for ExpQ* EPS (p=.007) and ExpQ* REV (p=.027), respectively, which confirms that investors attach a relatively lower (higher) weight to earnings (revenue) forecast revisions when expense quality is low. Again, the coefficient estimates for RevQ *ΔEPS and RevQ* ΔREV are not significant (p=.556 and.172, respectively). Together, the results in Tables 4 confirm that investors value revenue forecasts particularly highly when low earnings quality is due to low expense quality. This provides incentives for investors to increase their demand for the additional information available in revenue forecasts, and should induce analysts to increase the supply of revenue forecasts. V. REVENUE AND EXPENSE QUALITY AND JOINT EPS AND REVENUE FORECASTS Our results thus far indicate declines in EPS forecast accuracy and value relevance in association with decreases in the quality of revenues and expenses, while revenue forecast accuracy does not 18

20 decline with a decrease in the quality of expenses. We also find an increase in the value relevance of revenue forecasts and a decrease in the value relevance of earnings forecasts when expense quality declines. We assert that these factors contribute to an increase in investor demand for revenue forecasts and induce analysts to more frequently accompany their earnings forecasts with a revenue forecast when the quality of expenses is low. We use the following logit model to examine if the quality of revenues and expenses are associated with the likelihood that analysts will report a revenue forecast with their EPS forecasts: 16 Pr DREV ijt ExpQ ln RevQ ln(1 Horizon ) ln Experience 0 1 it 2 it 3 ijt 4 ijt ln Broker size ln # Firm followed ln MV ln Analyst following 5 ijt 6 ijt 7 it 8 it COV STD REV B / M ln Age ROA Dloss 9 it 10 it 11 it 12 it 13 it Margin LEV Industry dummies Year dummies u 15 it 16 it 17 k 28 k ijt k 0 k 0 it. (6) The set of controls in equation (6) is the same as for equation (4). Essentially, the same factors that contribute to analyst difficulties in forecasting earnings should also relate to valuation difficulties for investors and induce them to seek additional information, such as revenue forecasts, for their decision making. To illustrate, we include Horizon as a control variable in equation (4) since it is more difficult for analysts to accurately forecast earnings early in a fiscal year (Sinha et al. 1997). Including Horizon in equation (6) controls for additional investor demand for revenue forecasts early in the fiscal year when EPS forecasts are relatively less useful. We expect joint EPS and revenue forecasts to be relatively more common early in a fiscal year. 16 As we examine the analyst decision to complement an earnings forecast with a revenue forecast, it is necessary to conduct the analysis using individual forecasts. This differs from Ertimur et al. (2011, 39), who model if the analyst issues at least one disaggregated forecast (that is, both a revenue and earnings forecast) during both the first half and second half of the year. Our approach recognizes that revenue forecasts that infrequently complement EPS estimates likely offer limited value to investors. To illustrate, a revenue forecast issued early in a fiscal year provides little information on the quality of analysts consecutive EPS forecast revision during the fiscal year. 19

21 Following Marks (2008) and Ertimur et al. (2011), we include analyst forecasting experience and broker size in equation (6) as we expect analysts with less forecasting experience and analysts from smaller brokers to more commonly issue both revenue and EPS forecasts in order to establish their reputations. Once an analyst has established a reputation, revenue forecasts expose the analyst to additional market scrutiny and the risk of reputation loss, which reduces the analyst s propensity to issue joint EPS and revenue forecasts. We include the number of firms an analyst follows for actively following and producing research reports on many companies is likely to discourage an analyst from devoting the time necessary to produce and report complementary revenue forecasts. The variables MV and Analyst following capture the quality of the firm s information environment. Higher quality information environments should reduce the cost of producing revenue forecasts, increasing the likelihood of reporting a complementary revenue forecast. We include stock price variability and revenue variability because both measures suggest a more challenging forecasting environment, which can reduce the propensity to report a complementary revenue forecast. However, while high price volatility and revenue variability can discourage analysts from reporting revenue forecasts, these factors can also increase investor demand for revenue forecasts. We include the book-to-market ratio because we expect investors to exhibit higher demand for revenue forecasts for growth firms as revenue estimates are more useful for valuing these firms (Ghosh et al. 2005; Ertimur et al. 2003). Firm age reflects that analysts should be more likely to produce revenue forecasts for younger firms that are hard to value on the basis of earnings alone due to their short time-series of financial information (Li et al. 2009; Ertimur et al. 2011). Further, investors demand for revenue forecast can vary on the basis of firm losses and profitability as losses and low profitability can be non-reflective of firm value (Burgstahler and Dichev 1997; Collins et al. 1997). 20

22 We expect increased investor demand for analysts revenue forecasts for firms with high net margin for this increases the impact of revenue on income. As before, we control for solvency and firm distress risk using firm financial leverage. We include year and industry dummies and cluster regression standard errors by analyst and firm to control for the serial dependence of observations (Petersen 2009). Table 5 presents estimation results for equation (6) where we include only revenue and expense quality (columns only revenue and expense quality). We find the coefficient estimate for ln ExpQ is significantly positive (p=.000), while that for ln RevQ is significantly negative (.000), suggesting low expense quality is positively associated with analysts propensity to report revenue forecasts and low revenue quality is negatively associated with the reporting of revenue forecasts. We find that a one standard deviation reduction in our expense quality measure ln ExpQ leads to a 35.6% higher likelihood of jointly issuing EPS and revenue forecasts, which suggests the effect is economically non-trivial. A one standard deviation reduction in revenue quality reduces the likelihood of a revenue forecast issue by 9.41%. [Insert Table 5 around here] The last columns of Table 5 show that our conclusion that analysts increase the supply of revenue forecasts when expense quality is low persists when we use the full specification of equation (6). The coefficient on ln ExpQ is positive and significant (p=.000), while the coefficient on ln RevQ becomes indistinguishable from zero (p=.969) with the introduction of our control variables.. 17 Reviewing the coefficient estimates for our control variables, we find evidence consistent with Marks (2008) and Ertimur et al. (2011) that analysts with more forecasting experience (i.e. more reputable analysts) are less likely to issue revenue forecasts. 18 However, broker size is not 17 In unreported results we find that our conclusions remain unchanged when we use ExpQ and RevQ without log transformations. 18 Our empirical results are qualitatively the same when we use analyst firm-specific and industry-specific experience (results untabulated). The correlation coefficient between analyst general and industry-specific experience is 0.93 and between general and firm-specific experience. 21

23 significantly associated with the issuance of joint EPS and revenue forecasts. Revenue forecasts are more common among smaller, high growth, younger firms, and less profitable firms, and firms with high stock price volatility. These results support the prediction that revenue forecasts dominate among growth firms with high value uncertainty and shorter time-series of financial information. Higher analyst coverage associates with a greater likelihood that an analyst will report a revenue forecast together with the EPS forecast. This is consistent with analysts using revenue forecasts to differentiate themselves when the competition among analysts is high. High variation in firm revenue and low margin reduce the likelihood a revenue forecast will complement the EPS forecast. The former is consistent with the prediction that increased revenue forecasting difficulty reduces analyst propensity to issue revenue forecasts. The latter evidence is consistent with the prediction that revenue forecasts are more informative for firms where revenue has stronger effects on net income. Finally, analysts are less likely to issue joint EPS and revenue forecasts for highly leveraged firms. This suggests that revenue forecasts can be less useful to investors in interpreting firm performance when financial distress risk is high. Overall, the results in Table 5 confirm that analysts exhibit a higher propensity to report an accompanying revenue forecast when expense quality is low, and this result holds when we control for other factors affecting the usefulness of revenue forecasts. VI. SENSITIVITY ANALYSES AND FURTHER TESTS Our sensitivity analyses consider controls for the number of institutional investor shareholders, management revenue guidance, and concurrent cash flow forecasts. We also consider alternative estimates of revenue and expense quality, and additional means of controlling for cross-sectional correlations in our data. Further tests examine the association between earnings quality components and the duration between reported revenue forecast revisions. Higher investor demand for revenue 22

24 forecasts can shorten the duration between reported revenue forecast revisions. Finally, we test whether the result in Ertimur et al. (2011) and in Marks (2008), that more reputable analysts are less likely to produce revenue forecasts, persists once we control for earnings quality differences between high and low analyst reputation groups. Institutional holdings Analysts can increase their revenue reporting to cater to institutional investor demand (Frankel et al. 2006; Ljungqvist et al. 2007). To test if our results are influenced by the number of institutional shareholders, we include this as a control variable in equation (6). Table 6 presents the results. Here we find that a higher number of institutional investor shareholders increases analyst propensity to report revenue forecasts (p=.005), consistent with pressures from institutional investors increasing the supply of reported revenue forecasts. Controlling for the number of institutional investor shareholders, the coefficient on ln ExpQ remains positive and significant (p=.000), consistent with analysts reporting more revenue forecasts when expense quality is low. [Insert Table 6 around here] Revenue forecast guidance Management s revenue guidance reduces the cost of producing revenue forecasts and can stimulate analysts to report more revenue forecasts to I/B/E/S. To control for the possibility that revenue guidance is impacting our results, we include a management guidance dummy (Guidance) equal to one if the firm issued revenue guidance in the 14-day period preceding the analyst s EPS forecast and the guidance is for the same fiscal period, otherwise Guidance is zero. 19 We estimate equation (6) with the inclusion of this guidance variable. Column Guidance in Table 19 We use a 14-day window between management revenue guidance and the analyst EPS forecast issuance because guidance should have a stronger effect on analyst decision to jointly issue EPS and revenue forecasts shortly after release. Consequently, a two-week window will have more power to isolate the effect of management revenue guidance on the analyst revenue forecast reporting decision than longer windows. 23