Earnings Management and Earnings Surprises: Stock Price Reactions to Earnings Components * Larry L. DuCharme Yang Liu Paul H. Malatesta University of Washington School of Business Box 353200 Seattle, WA 98195 First Draft: February 29, 2004 This Draft: September 7, 2004 Abstract: We examine the stock price reactions to earnings announcements. We use a database that contains analysts' forecasts of earnings and revenues. This allows us to decompose earnings surprises into three components: innovations to expected cash flow and expected normal accruals, and an abnormal accrual component. We find that abnormal stock returns are contemporaneously positively related to all three components of the earnings surprise. The impact on stock prices varies, however, across the components. The marginal value of innovations to expected normal accruals much exceeds the marginal value of innovations to expected cash flow, which exceeds the marginal value of abnormal accruals. We also examine the relation between the earnings components and future cash flow and stock returns. Innovations to the earnings components are positively related to future cash flow. * We appreciate helpful comments from Shuping Chen, Jon Karpoff, Terry Shevlin, and from seminar participants at Seattle University. We thank Lew Thorsen for assistance with programming and data collection. Malatesta gratefully acknowledges research grants provided by the University of Washington.
Earnings Management and Earnings Surprises: Stock Price Reactions to Earnings Components Abstract We examine the stock price reactions to earnings announcements. We use a database that contains analysts' forecasts of earnings and revenues. This allows us to decompose earnings surprises into three components: innovations to expected cash flow and expected normal accruals, and an abnormal accrual component. We find that abnormal stock returns are contemporaneously positively related to all three components of the earnings surprise. The impact on stock prices varies, however, across the components. The marginal value of innovations to expected normal accruals much exceeds the marginal value of innovations to expected cash flow, which exceeds the marginal value of abnormal accruals. We also examine the relation between the earnings components and future cash flow and stock returns. Innovations to the earnings components are positively related to future cash flow.
Earnings Management and Earnings Surprises: Stock Price Reactions to Earnings Components 1. Introduction The relationship between stock prices and company performance measured by accounting numbers has attracted much attention among accountants and financial economists. Many studies have examined stock price reactions to earnings announcements. Ball and Brown (1968), Beaver (1968), and Rendleman, Jones, and Latané (1982) report that stock returns are positively related to contemporaneous earnings surprises. This is a very robust result and subsequent work has confirmed the findings of the early studies in this area. Other work has sought to determine if the composition of earnings contains information beyond that conveyed by the level of earnings alone. For example, Bowen, Burgstahler, and Daley (1987) and Wilson (1986, 1987) break earnings into accrual and cash (or funds) flow components. They show that innovations in both components are statistically significantly related to the abnormal stock returns of reporting firms. 1 Another line of research has focused on how firms may influence, or manage, reported earnings through their choices of accounting policies. Residual accruals are widely used to measure the thrust and extent of earnings management. Earnings are divided into cash flows and accruals. Most often only the current accruals, those related to working capital, are analyzed. The accruals are further decomposed into normal and abnormal components using Jones' (1991) time series regression model or, more commonly, some version of DeFond and Jiambalvo's (1994) cross-sectional regression model. The residuals from the regression, the abnormal accruals, are attributed to discretionary earnings management. The remainders of the accruals are the normal, nondiscretionary components. There is a body of evidence that suggests that the information contained in accruals is not efficiently impounded in stock prices when it enters the public domain. Teoh, Welch, and Wong (1998a, 1998b), and DuCharme, Malatesta, and Sefcik (2001, 1 See also Rayburn (1986). 1 1
2004) find that abnormal working capital accruals around the time of stock issues, both IPOs and SEOs, are significantly negatively related to subsequent abnormal stock returns. Eckbo, Masulis, and Norli (2000) argue that these results are spurious, arising from inadequate controls for differences in risk between firms and the resulting mismeasurement of abnormal returns. DuCharme, Malatesta, and Sefcik (2001) show, however, that their results for IPOs hold even for risk-adjusted returns measured using the multi-factor CAPM of Eckbo, Masulis, and Norli (2000). In addition, Sloan (1996) and Xie (2001) report that accruals have significant power to predict subsequent stock returns in general. Buying stock in firms with low accruals and selling stock in firms with high accruals generates significantly positive abnormal returns relative to the Sharpe (1964) CAPM and the three-factor model of Fama and French (1993). How to interpret these provocative results remains debatable. Despite the results of DuCharme, et al. (2001), it is not possible to dismiss the central point raised by Eckbo, et al. (2000). We do not know the correct way to control for differences in risk and this is a crucial issue in the studies cited above, which attempt to measure abnormal returns over long holding periods. In the current research we provide further evidence on the relation between stock returns and accounting information. Our study is closely related to those of Callen and Segal (2004), Subramanyam (1996), and DeFond and Park (2001). Callen and Segal (2004) examine the relation between stock returns, and news about cash flows, accruals, and expected returns within a variance decomposition framework similar to that developed by Vuolteenaho (2002). They conclude that news about accruals significantly affects stock returns. The measures of news used in the study, however are indirect, derived from estimated VAR models of the accounting numbers. Moreover, Callen and Segal (2004) do not distinguish between news about normal accruals and abnormal accruals. Subramanyam (1996) decomposes earnings into cash flow, normal accrual, and abnormal accrual components using several different methods. He then estimates a mixed cross-sectional time series regression of annual stock returns on these earnings components and concludes that all three are priced. These results raise several questions. Variation in realized stock returns arises in part from variation in ex ante expected 2 2
returns. Thus, in the regression of raw returns on earnings components we do not know whether the results mean that the components are correlated with priced risk(s), or if news about the components is itself driving unanticipated returns. The experimental design does not extract the expected components of returns. As Fama (1991, 1998) notes, this is a serious issue when low frequency (e.g., annual) return data are used. Another point is that the formal hypothesis tests make no allowance for the crosssectional correlation of contemporaneous returns. In effect, these correlations are assumed to be zero. This is not very plausible. For this reason, the test significance levels are likely to be overstated. Finally, the regressors are the realized earnings components. These also contain expected as well as unexpected parts. We cannot tell from these results if, for example, expected stock returns are positively related to expected earnings components or if unexpected returns are positively related to news about earnings. DeFond and Park (2001) also examine the relation between stock returns and earnings components, but within an event study framework using daily returns data. This avoids or mitigates the problems confronted by Subramanyam (1996). They present evidence that stock prices react less to earnings surprises when abnormal accruals increase the magnitude of the surprise, and react more when the opposite is true. This suggests that investors do distinguish among earnings components. However, DeFond and Park do not actually decompose earnings surprises themselves and they do not measure the marginal values of the earnings components. Moreover, their results are not robust to alternative measures of abnormal accruals. When they use a time series approach similar to that of Jones (1991) to measure abnormal accruals, their evidence provides no support for the hypothesis that investors distinguish between different earnings components. Regardless of how the abnormal accruals are measured, DeFond and Park assume that all of the relevant accounting information becomes publicly available on the earnings announcement date. Investors, however, must often wait until a later date to obtain information from balance sheets and thus, by inference, about accruals. Therefore, the stock price reaction to abnormal accruals on the earnings announcement date is surely an attenuated measure of the marginal value of accruals. 3 3
In this paper we extend previous work by examining the relation between stock returns, cash flows, and accruals around earnings release dates. We postulate a simple model where abnormal stock returns are a linear function of contemporaneous shocks to expected cash flow, expected normal accruals, and expected abnormal accruals. Following Wilson (1986, 1987) we recognize that information about the various components of earnings may arrive at two dates. As noted above, earnings and revenues typically are released before balance sheet information becomes available to investors. In our model new earnings and revenue information cause investors to update their beliefs about contemporaneous cash flow and normal accruals. The release of balance sheet information in conjunction with earnings and revenues reveals the abnormal accrual and causes an additional change in expected cash flow. We show explicitly how earnings and revenue surprises, together with abnormal accruals, are related to cash flow and accrual surprises. We use analysts' forecasts of firm earnings and revenues compared to actual earnings and revenues to estimate earnings and revenue surprises. With the surprise estimates, we decompose the earnings surprise into three parts: unexpected cash flow, unexpected normal accruals driven by the revenue surprise, and the abnormal accrual attributable to discretionary earnings management. We then measure stock price reactions to the three components of the earnings surprises using a variety of methods, including multiple regression. To our knowledge, this three-part decomposition is unique to this study. Our analysis allows us to determine if investors discriminate between the various components of earnings. In particular, we test to see if investors value abnormal accruals differently than the other components of earnings. By examining stock price reactions at both the earnings announcement date and when balance sheet information is released we provide direct evidence on the extent to which investors distinguish among different earnings components. This has potentially important implications regarding market efficiency and the way that we interpret the results of, e.g., Xie (2001). Suppose that stock prices react differently to news about the various components of earnings. Specifically, suppose that the marginal value of abnormal accruals differs from the marginal values of cash flow and of normal accruals. This would be consistent with the notion that investors are sophisticated consumers of accounting information and 4 4
that they price securities efficiently with respect to that information. In this case, one would incline towards the viewpoint of Eckbo, et al. (2000) who contend that the apparent power of abnormal accruals to predict subsequent abnormal returns over long horizons arises spuriously from omitted risk factors. Assume, on the other hand, that stock prices do not react differently to news about the various components of earnings. This would suggest that investors process accounting information incompletely and therefore price securities inefficiently. We would be led to interpret negative correlation between abnormal accruals and measures of subsequent abnormal stock returns as evidence that the stock market is inefficient in processing publicly available information contained in financial statements. In our empirical analysis we find, like others before us, that abnormal stock returns are significantly positively related to contemporaneous earnings surprises. In addition, we report strong evidence that normal accruals contain priced information beyond that contained in earnings alone. Moreover, in our sample the valuation impact of shocks to normal accruals significantly exceeds that of unexpected cash flow. Finally, the evidence supports the hypothesis that investors value earnings supported by cash flow more highly than earnings supported by abnormal accruals. The estimated marginal value of abnormal accruals is positive and statistically significant, but it is smaller than that of unexpected cash flow. We also examine the relation between shocks to the earnings components and subsequent cash flow using several different regression models. Our results indicate that cash flow and normal accrual surprises are both positively related to subsequent cash flow, though some of the estimates are imprecise. Abnormal accruals, however, are significantly positively related to subsequent cash flow in all of our models. Thus, it appears that positive abnormal accruals convey good news about future cash flow. This helps to explain the positive contemporaneous relation between abnormal accruals and abnormal stock returns. The remainder of the paper is organized as follows. Section 2 sets forth our model and discusses our assumptions about the information arrival process pertaining to 5 5
earnings, cash flows, and accruals. Section 3 describes our data sources and empirical methods. Our results are presented in Section 4. Section 5 concludes. 2. A Model of Stock Price Reactions to Earnings Components 2.1 Accounting Relationships We begin with the accounting identity relating earnings to cash flows and accruals. For any firm j and performance measurement interval p, we have Y jp = C jp + A jp [1] where Y, C, and A denote earnings, cash flow, and accruals per share, respectively. Accruals may be thought of as the sum of two components, normal accruals, A n, and abnormal (residual) accruals, A r, that arise from discretionary earnings management. Hence, we can represent earnings as the sum of three components. Y jp = C jp + A n jp + A r jp [2] We assume that investors formulate expectations regarding the levels of earnings and its components and that stock prices may depend on these expectations. If investors are rational, then expected earnings must equal the sum of the expected components of earnings: E t (Y jp ) = E t (C jp ) + E t (A n jp) + E t (A r jp) [3] where E t denotes the expectation operator conditional on information known at date t. It follows that the unexpected part of realized earnings is the sum of unexpected cash flows and unexpected normal and abnormal accruals. [Y jp - E t (Y jp )] = [C jp - E t (C jp )] + [A n jp - E t (A n jp)] + [A r jp - E t (A r jp)] [4] In general, rationality requires that the innovation to expected earnings at time t equals the sum of contemporaneous innovations to expected cash flow, expected normal accruals, and expected abnormal accruals. So, 6 6
[E t (Y jp ) - E t-1 (Y jp )] = [E t (C jp ) - E t-1 (C jp )] + [E t (A n jp) - E t-1 (A n jp)] + [E t (A r jp) - E t-1 (A r jp)] [5] We assume that normal accruals per share are a linear function of the change in revenues per share from the previous period. We write A n jp = α j + φ j (S jp - S j p-1 ) = α j + φ j S jp [6] where S jp and S j p-1 denote current period and lagged revenues per share for firm j, respectively. The model in [6] is similar to the working capital accruals model of Teoh, et al. (1998a). 2 Substituting [6] into [5] and simplifying yields [E t (Y jp ) - E t-1 (Y jp )] = [E t (C jp ) - E t-1 (C jp )] + φ j [E t (S jp ) - E t-1 (S jp )] + [E t (A r jp) - E t-1 (A r jp)] [7] Note that the innovation to the normal accrual component is proportional to the change in expected revenues per share. 2.2 Expectations and Information Arrival In the model, investors formulate prior expectations about earnings and revenues. Sometime after the end of interval p, information is revealed regarding the actual levels of earnings and revenues. At this time, investors update their beliefs regarding earnings, cash flows, and normal accruals. Often, however, the end-of-period balance sheet is not reported until a subsequent date. At that subsequent date, investors observe actual accruals and, therefore, the abnormal accrual component. This information resolves the remaining uncertainty regarding realized cash flow. 3 2 Teoh, et al. (1998a) estimate a model like [6] using cross-sectional data. When they employ the fitted model to forecast normal accruals, however, they adjust the regressor by subtracting the change in receivables over the reporting interval. Their approach, while commonplace in the accounting literature, is rather unusual from an econometric viewpoint. Also, it is generally true that the receivables adjustment has little practical effect, on average. For these reasons we ignore the adjustment here. 3 Of course, income statement and balance sheet items may later be restated. Hence, some uncertainty remains regarding all of the earning components even after the reports are filed with the Securities and Exchange Commission. We assume that this uncertainty is negligible. 7 7
Suppose that on date t = a earnings and revenues are announced and that this occurs before the balance sheet is reported. We assume that the expected abnormal accrual is zero until its realization is revealed. From [7] we obtain [Y jp - E a-1 (Y jp )] = [E a (C jp ) - E a-1 (C jp )] + φ j [S jp - E a-1 (S jp )] [8] Hence, at date a the innovations to expected cash flow, expected normal accruals, and expected abnormal accruals are, δ ja c = [Y jp - E a-1 (Y jp )] - φ j [S jp - E a-1 (S jp )] δ ja n = φ j [S jp - E a-1 (S jp )] [9] δ ja r = 0 Later, on date t = f, the balance sheet is reported. This reveals the realized accrual, its abnormal component, and actual cash flow. The innovations at date f, therefore, are δ jf c = -{A jp - [α j + φ j S jp ]} δ jf n = 0 [10] δ jf r = A jp - [α j + φ j S jp ] The other case of interest is when dates a and f coincide, when all information is released simultaneously. In this case, we have δ jf c = [Y jp - E f-1 (Y jp )] - φ j [S jp - E f-1 (S jp )] - {A jp - [α j + φ j S jp ]} δ jf n = φ j [S jp - E f-1 (S jp )] [11] δ jf r = A jp - [α j + φ j S jp ] 2.3 Stock Returns Let R jt denote the stock return for firm j at time t and suppose that 8 8
R jt = ρ j + β j Z t + ε jt [12] where ρ is a constant, β measures systematic risk(s) relative to common risk factor(s) Z, ε is the abnormal return, and E(Z) = E(ε) = 0. We assume that the abnormal return reflects new firm-specific information arriving at time t, including any information pertaining to earnings. Our specification is: ε jt = γ 0 + γ c (δ c jt /T jp-1 ) + γ n (δ n jt /T jp-1 ) +γ r (δ r jt /T jp-1 ) + u jt [13] In [13] T jp-1 denotes total assets per share for firm j at the end of performance measurement interval p-1. We scale the innovations δ by dividing them by assets per share. 4 The regression coefficients γ measure the marginal value of scaled cash flow, normal accruals, and abnormal accruals per share, respectively. The residual term u jt is the part of stock return arising from firm-specific information that is unrelated to information about earnings. We may rewrite [13] in terms of the scaled innovations to expected earnings, revenues, and residual accruals. ε jt = γ 0 + γ c [(E t (Y jp ) - E t-1 (Y jp )) - φ j (E t (S jp ) - E t-1 (S jp )) - (E t (A r jp) - E t-1 (A r jp))]/t jp-1 + γ n [φ j (E t (S jp ) - E t-1 (S jp ))]/T jt-1 + γ r [E t (A r jp) - E t-1 (A r jp)]/t jp-1 + u jt [14] Collecting terms, we have ε jt = γ 0 + γ c (E t (Y jp ) - E t-1 (Y jp ))/T jt-1 + (γ n - γ c ) φ j (E t (S jp ) - E t-1 (S jp ))/T jp-1 + (γ r - γ c ) (E t (A r jp) - E t-1 (A r jp))/t jp-1 + u jt [15] This formulation is more useful in the empirical analysis that follows. 4 The innovations are themselves measured on a per share basis. Thus, when we divide them by assets per share the number of shares in the numerators and denominators of the ratios cancel and we are left with the total dollar innovations as fractions of total assets. Hence, the decomposition actually breaks surprises in the return-on-assets into parts related to cash flow, normal accruals, and abnormal accruals. 9 9
3. Data Sources and Empirical Methods In our empirical analysis we require proxies for innovations to expected earnings and revenues. We use stock analysts forecasts of annual earnings and revenues, and their forecast errors. We obtained the forecasts from the First Call Historical Database. 5 These continue through 2001. Revenue forecasts, however, are not available for years prior to 1998 and there are very few of them in the database for 1998 and 1999. As a consequence our sample contains forecasts primarily for 2000 and 2001. Realized earnings were also taken from First Call. The First Call Database does not include realized revenues, so we drew these from the Compustat Research File. 6 We focus on the days when firms announce their results for annual earnings and revenues, and when they file their year-end financial statements with the SEC. Announcement dates are taken from Compustat and the filing dates from the SEC website (www.sec.gov). To calculate an earnings forecast error we subtract from actual earnings the forecast made most recently prior to the date when realized earnings first become publicly known. This date is most often, but not always, before the SEC filing date. Revenue forecast errors are calculated in the same way as earnings forecast errors. We use the analyst forecast errors to proxy for the innovations to expected earnings and revenues on the information release date, either the filing date or an earlier announcement date depending on the specific case. To measure accruals we must have information from firm balance sheets. We obtained this information from Compustat. In our empirical analysis we restrict our attention to current accruals only. We assume that unexpected long-term accruals are negligible. Abnormal accruals are estimated using the cross-sectional method of Teoh, et al. (1998a). Each sample firm j is pooled with all other Compustat firms having the same two-digit SIC code. Current accruals for the year are then regressed on the change in 5 First Call is a Thomson Financial Company. We obtained access to the First Call Historical Database through special permission. First Call compiled the data on revenue forecasts at our request. 6 We used Compustat item 12, net sales. 10 10
revenues from the previous year. Consistent with standard practice, we employ a weighted least squares regression, with weights equal to the inverse of total assets as of the end of the previous year. The residual from this regression for firm j is our proxy for the innovation to the expected abnormal accrual for the firm on the SEC filing date. The estimated slope coefficient, φ j, multiplied by the revenue forecast error is our proxy for the innovation to expected normal accruals for the firm on the date when its revenues are first publicly known. We refer to this as the adjusted revenue error. Stock return data are obtained from the CRSP files 7. We use two different measures of abnormal stock returns, ε jt. One is a simple market-adjusted measure. To obtain the abnormal return estimate we subtract the daily return to the CRSP equally weighted market index from the daily return to the stock of each sample firm. The other measure is a market model prediction error. We estimate the intercept and the slope of the model for each firm by regressing its stock returns on the return to the equally weighted index. For this regression we use daily returns over the period from 180 days through 11 days before the announcement date. We then use the estimated intercept and slope to predict firm stock returns, conditional on the market index return. The estimated abnormal return equals the actual stock return less the predicted return. There are a total of 3,049 firm-year observations in the First Call Database for the years from 1998 through 2001 where annual earnings per share forecasts, revenue per share forecasts, and realized earnings per share are all available. 149 observations are eliminated because missing information on Compustat prevents us from calculating current accruals and therefore we cannot implement the Teoh, et al (1998a) method to estimate abnormal accruals. An additional 233 observations are eliminated because we are unable to determine the SEC filing date. We also lose 209 firm-years because of missing event date stock returns on the CRSP file. This leaves us with 2,458 firm-year observations for which we can calculate earnings and revenue forecast errors, and also to estimate the abnormal accrual and φ j, the slope coefficient in the accruals model. (See equation [6].) To moderate the influence of outliers in the data we trimmed the sample by excluding observations in the largest and the smallest one percent of the forecast 7 Center for Research in Security Prices at the University of Chicago. 11 11
errors, abnormal accruals, and φ j. After removing the outliers, there are 2,356 observations remaining in the trimmed sample. 4. Empirical Results 4.1 Stock Price Reactions to Earnings Components Table 1 presents summary statistics describing the distributions of the earnings and revenue forecast errors and of the abnormal accruals. All of the observations are expressed as fractions of total assets, consistent with the discussion in section 2.3. The median forecast errors and abnormal accruals are very close to zero. The mean revenue forecast error, however, differs significantly from zero, with a p-value less than 0.10. The mean earnings and revenue forecast errors are both negative. This indicates that stock analysts' forecasts were overly optimistic, on average, about our sample firms. Others have reported previously a similar result regarding earnings. 8 The results in table 1 show that analysts' revenue forecasts also tend to display a positive bias. Table 1 also shows that the mean abnormal accrual is positive and statistically significant at the one percent level of confidence. This is a somewhat surprising result. The estimated abnormal accruals are residuals from cross-sectional weighted least squares regressions. The weights in these regressions are the inverses of total assets. By construction, the weighted average residual for these regressions equals zero, but our sample firms display a positive weighted average. The firms in our sample all share one characteristic. They are followed by stock analysts. It is possible that close and continuing scrutiny by analysts creates pressure on firms to increase earnings and that firms respond, at least in part, by managing accruals upward. Table 2 summarizes our analysis of market-adjusted stock returns around earnings announcement and filing dates. Panel A contains the results for those cases where the two dates are distinct, and Panel B contains the results for those cases where the dates 8 See Abarbanell (1991), Brown, Foster, and Noreen (1985), Chopra (1998), Dremen and Berry (1995), and Stickel (1990) for examples. 12 12
coincide. Table 3 reports a parallel set of results for the market-model stock return prediction errors. It is apparent from either of tables 2 or 3 that the earnings reports of our sample firms disappointed investors, on average. The mean cumulative market-adjusted return for the three days centered on the announcement date is negative and highly significant. So is the mean cumulative market-model prediction error. For the cases with distinct announcement and filing dates, the mean cumulative abnormal returns around the latter date are also negative, and significant at the.01 confidence level. For cases where the announcement and filing dates coincide, the mean cumulative abnormal returns are negative, and larger in absolute value than for the cases with distinct dates, but they differ insignificantly from zero. Of course, neither do they differ significantly from the mean cumulative abnormal returns around the filing date for the cases with distinct announcement and filing dates. The group with coincident dates contains only 84 observations. It is not surprising that earnings reports tended to be bad news during the period spanned by our sample. Recall that the First Call Database contains few revenue forecasts for periods before the year 2000. All but one of our sample forecasts are for the years between 1999 and 2001, which would be related to earnings reports announced and filed between about March 2000 and March 2002. Over that period, the S & P 500 stock index fell by 24 percent. The NASDAQ index fell by 59 percent. Table 4 provides information on the joint distribution of the analysts' forecast errors and cumulative abnormal stock returns around the earnings announcement dates. The number of observations here is less than in the preceding tables because we exclude cases where earnings or revenue forecast errors equal zero. All of the observations underlying the figures reported in this table pertain to cases with distinct announcement and filing dates. The figures in table 4 reveal that investor reactions to news about earnings and revenues were asymmetric during our sample period. If either realized earnings or revenues were less than analysts forecast, stock prices tended to fall on the news. Over the entire sample period there are 444 cases where both forecast errors are negative. For those cases, the mean cumulative abnormal returns around the earnings 13 13
announcement date are less than -0.03. There are 842 cases where only one of the forecast errors is negative. For those cases, the mean cumulative abnormal returns range between -0.014 and 0.000. Most of these estimates differ significantly from zero at conventional confidence levels. Stock prices do not, however, respond to favorable news about earnings and revenues. When both forecast errors are positive the mean cumulative abnormal returns around the announcement date are close to zero and statistically insignificant. When firms release information about their most recent performance they often make concurrent announcements regarding management's expectations about future performance. Given our sample period, it is plausible that favorable news about recent earnings and revenues would arrive together in many cases with pessimistic assessments by management regarding near-term future performance. There is no reason to expect that stock prices would react positively to such a mixed message. Table 5 shows how cumulative abnormal stock returns around the SEC filing date are related to abnormal accruals. As in table 4, all of the observations underlying the figures reported in table 5 pertain to cases with distinct announcement and filing dates. The table shows clearly that investors distinguish between cases where the abnormal accrual is positive and where it is negative. Over the full sample period, there are 1024 cases where the abnormal accrual is positive. Around the filing date the mean cumulative market-adjusted return is -0.016 and the mean cumulative market-model prediction error is -0.011. Both estimates differ significantly from zero at the 0.01 level of confidence. There are 944 cases where the abnormal accrual is negative. For these cases, the cumulative abnormal returns around the filing date are negative, but closer to zero. Moreover, the two sets of estimates based on market-adjusted returns differ significantly from each other. The mean cumulative market-adjusted return for the positive abnormal accruals group is significantly less than that for the negative abnormal accruals group, at the 0.05 level. The mean cumulative market-model prediction error for the negative abnormal returns group is also less than that for the negative abnormal accruals group. The difference is not significant at conventional confidence levels. The one-sided p- value, however, is 0.07. Finally, the results are consistent over the sample period. The mean cumulative abnormal returns around the filing date are smaller for the positive 14 14
abnormal accruals group than for the negative abnormal accruals group in every year for which a comparison can be made. The only exception is for the year 1998, for which we have but a single observation. Recall from equation [10] that when the earnings announcement and SEC filing dates are distinct that the abnormal accrual corresponds to an innovation in expected cash flow. The implied shock to expected cash flow is simply the opposite of the abnormal accrual. Hence, the results in table 5 also indicate that stock prices respond negatively to negative innovations in expected cash flow. We performed several regressions to estimate the marginal values of innovations to cash flow, normal accruals, and abnormal accruals. Table 6 reports the results of some of these regressions. These results are based upon a clean sample, where we have available complete data on the forecast errors and the abnormal accruals. Also, the cases used in these regressions all entail distinct announcement and filing dates. Cumulative abnormal returns around the announcement dates are regressed on earnings and adjusted revenue forecast errors. Cumulative abnormal returns around the filing dates are regressed on the abnormal accruals. All of the estimated slope coefficients in table 6 differ significantly from zero at conventional confidence levels. The results for market-adjusted returns are very similar to those for market-model prediction errors. We confirm the well-known finding that stock returns are positively related to earnings surprises. It is more interesting, however, to examine the impact of adjusted revenue surprises. These results are reported in columns 2 and 5 of the table, which show that the estimated marginal value of an adjusted revenue surprise is positive and significant at the 0.05 level. We can interpret the results in table 6 within the framework of our model. (See equation [15].) In the model, the true coefficient of the earnings surprise is γ c, the marginal value of an innovation to expected cash flow. The true coefficient of the adjusted revenue surprise is the difference (γ n - γ c ), where γ n denotes the marginal value of an innovation to expected normal accruals. Hence, the results in columns 2 and 5 of table 6 indicate that the marginal value of the expected cash flow innovation is positive and that the marginal value of the expected normal accrual innovation significantly 15 15
exceeds that of the expected cash flow innovation. The last line in the table reports t- statistics for tests of the hypothesis that the sum of the coefficients on the earnings and adjusted revenue forecast errors equals zero. This is equivalent to a test of the hypothesis that γ n equals zero. We reject this hypothesis at the 0.01 level for market-adjusted returns and at the 0.05 level for market-model prediction errors. Columns 3 and 6 of table 6 show the results for regressions of abnormal returns around the SEC filing date on the estimated abnormal accruals. The estimated coefficients are negative, and significant at the 0.05 and 0.10 levels for market-adjusted returns and market-model prediction errors, respectively. These coefficients correspond to the term (γ r - γ c ) in our model, where γ r denotes the marginal value of the innovation to expected abnormal accruals. Hence, the significant negative coefficient estimates indicate that the marginal value of the abnormal accruals is less than that of cash flow. Stock prices respond more to earnings supported by cash flow than to earnings supported by abnormal accruals. Regressions reported in table 7 pool the time series and cross-sectional data. The dependant variable equals cumulative abnormal returns around either the announcement or filing dates. The regressors equal earnings and adjusted revenue surprises, and the abnormal accruals. The design matrix reflects the information arrival sequence, as discussed in section 2. When announcement and filing dates are distinct, the case generates an observation for each. The announcement date abnormal return corresponds to the earnings and adjusted revenue errors, with the abnormal accrual set to zero. For these cases, the filing date abnormal return corresponds to the revealed abnormal accrual, and the earnings and adjusted revenue errors are set to zero. When the announcement and filing dates coincide, the case generates a single observation. The cumulative abnormal returns around the filing date then correspond to the (simultaneously) revealed earnings and adjusted revenue errors, and the abnormal accrual. As in table 6, the results in table 7 are based upon a clean sample, where we have available complete data on the forecast errors and the abnormal accruals. The estimated coefficients for the pooled regressions are quite similar to those reported in table 6. Again, all of the estimated coefficients differ significantly from zero. 16 16
The coefficients on the earnings and adjusted revenue forecast errors are both positive. We reject the hypothesis that the sum of the coefficients of the earnings and adjusted forecast errors equals zero. This supports the conclusion that the marginal value of innovations to expected normal accruals is positive, as is the marginal value of innovations to expected cash flow. In table 7, the estimated coefficient of abnormal accruals is significantly negative. As we noted earlier, this is an estimate of the term (γ r - γ c ) in our model, where γ r denotes the marginal value of the innovation to expected abnormal accruals and γ c denotes the marginal value of the innovation to expected cash flow. We can recover an estimate of γ r itself by adding together the estimated coefficients of the earnings forecast error and of the abnormal accruals. This results in estimates of 0.223 and 0.212 for the marketadjusted return and market-model prediction error regressions, respectively. These point estimates are positive, which supports the view that investors value abnormal accruals, albeit somewhat less than cash flow, and much less than normal accruals. Moreover, we perform a t-test of the hypothesis that the sum of the two estimated coefficients equals zero and we reject this hypothesis at the 0.05 confidence level. Table 8 reports the results for regressions like those reported in table 7. For these regressions, though, we relax the requirement that all data be available for every case. In some cases data on earnings or revenue forecast errors, or on the abnormal accruals are not available. Where data are missing, we assume that the corresponding innovation is zero. The impressions gained from the results in table 8 differ little from those gained from tables 6 and 7. The implicit estimates for the marginal values of cash flow and normal accrual innovations are positive and significant. The implicit estimates of the marginal value of abnormal accruals are somewhat less than those from table 7. We again reject at the 0.05 level the hypothesis that the sum of the coefficients on the earnings forecast errors and on the abnormal accruals is zero. This provides strong evidence that the abnormal accrual component of earnings moves stock prices. 17 17
4.2 Predictive Power of Earnings Components Accrual accounting creates natural links between the components of current and future earnings. We expect to observe some persistence in cash flow from one period to the next and for current accruals to contain information about future cash flow. For example, under ordinary circumstances, this period's receivable becomes next period's cash receipt. Thus, unexpected cash flow and unexpected normal accruals should be positively related to subsequent cash flow. For abnormal accruals, however, the case is not as clear. Suppose that a firm's managers have an incentive, such as might arise when a stock offer is planned, to bolster its stock price temporarily. Under these circumstances managers could make discretionary accounting choices affecting accruals that would boost current earnings without affecting current or future cash flows. For example, underestimating inventory losses would abnormally inflate current accruals and increase current earnings. Eventually the losses would be realized and the abnormal accruals would reverse, as would the transitory increase in earnings. If this type of scenario occurs pervasively, then abnormal accruals generally would be unrelated to subsequent cash flows. It is also possible, though, that managers make discretionary accounting choices to signal their beliefs regarding their firm's future prospects and intrinsic value. Under this hypothesis, abnormally high accruals this period would signal that managers expected higher cash flows in the future. Abnormal accruals would be positively related to future cash flows. We investigate the relation between current earnings surprises and subsequent cash flows by regressing cash flow on lagged earnings and adjusted revenue surprises, and lagged abnormal accruals. The relation between cash flow, earnings, and accruals in equation [2] implies that C jp+1 = Y jp+1 - A n jp+1 - A r jp+1 [16] We postulate a linear relation between scaled cash flow and lagged scaled innovations to the earnings components. 18 18
C jp+1 /T jp-1 = λ 0 + λ c (δ jt c /T jp-1 ) + λ n (δ jt n /T jp-1 ) + λ r (δ jt r /T jp-1 ) + w jp+1 [17] Using earlier results from section 2, we rewrite [17] in terms of the lagged scaled earnings and adjusted revenue surprises, and the abnormal accruals to obtain C jp+1 /T jp-1 = λ 0 + λ c (E t (Y jp ) - E t-1 (Y jp ))/T jp-1 + (λ n - λ c ) φ j (E t (S jp ) - E t-1 (S jp ))/T jp-1 + (λ r - λ c ) (E t (A r jp) - E t-1 (A r jp))/t jp-1 + w jp+1 [18] where w jp+1 is the regression residual, assumed normally distributed with mean zero. Note that in [17] and [18] the slope coefficients λ c, λ n, and λ r represent the marginal effects of lagged shocks to expected cash flow and normal accruals, and abnormal accruals, respectively, on cash flow. We estimate the marginal effects of the lagged shocks on cash flow by ordinary linear regression, pooling the sample across firms and time. Each observation requires that a firm be in our sample for two consecutive years. Our time series is short and all but 8 of our 728 observations pertain to cash flows of 2001 regressed on earnings surprises of 2000. The results are reported in table 9. The first column of table 9 contains estimates for the model set forth in [18]. The second column contains estimates for a slightly different specification that includes lagged scaled cash flow among the regressors. The third column reports results where annual fixed effects dummies are included. The results for the three model specifications in table 9 are broadly similar. The estimated coefficient on the lagged earnings forecast error is positive for all three models and statistically significant at the 5% level of confidence in two of the three. The coefficients on the lagged adjusted revenue forecast error are all positive, also, but statistically insignificant. Recall that these are estimates of the difference between the marginal effects on cash flow of lagged normal accruals surprises and lagged cash flow surprises. Therefore, we must add these estimated coefficients to the corresponding estimated coefficients on the lagged earnings forecast error to obtain estimates of λ n, the marginal effect on cash flow of lagged normal accrual surprises. The relevant sums are 19 19
all positive. For two of the three specifications they are statistically significant at the 5% level. Taken as a whole, this evidence is persuasive. The results suggest that current cash flow and normal accruals surprises are positively related to subsequent cash flow. The strongest evidence in table 9 concerns the abnormal accruals. In each specification the estimated coefficient on the lagged abnormal accrual is positive and significant at the 1% level. In this case, these are estimates of the difference between the marginal effects on cash flow of lagged abnormal accruals and lagged cash flow surprises. The results indicate that the marginal effect of lagged abnormal accruals exceeds that of cash flow surprises. We recover estimates of λ r, the marginal effect on cash flow of lagged abnormal accruals, by summing the estimated coefficients on the lagged earnings forecast error and the abnormal accruals. This sum is positive for each of the three models. The t-statistics for testing the hypothesis that this sum is zero are reported in the table. We reject this hypothesis for all three of the model specifications, twice at the 1% level and once at the 5% level. Hence, the evidence tends to confirm the view that abnormal accrual surprises are positively related to future cash flow. 5. Conclusions Using analysts' forecasts of earnings and revenues, we decompose innovations to expected earnings into three components: cash flow shocks, normal accrual shocks, and abnormal accruals. We find that all three components are positively related to abnormal stock returns measured around the times when earnings, revenue, and financial statement information becomes publicly available. Stock prices respond differently, however, to the different components. The marginal value of innovations to expected normal accruals is significantly greater than the marginal value of innovations to expected cash flow. In turn, the marginal value of expected cash flow innovations is significantly greater than the marginal value of abnormal accruals. Our results extend the work of Subramanyam (1996) and DeFond and Park (2001). Stock price reactions to earnings announcements are consistent with the notion that investors subtly distinguish between earnings components. Apparently, investors perceive differences in the quality of reported earnings, depending on their composition. 20 20
It is particularly interesting that investors seem to distinguish between earnings that are supported by cash flows and those that are generated by abnormal accruals, valuing the former more highly than the latter. Yet, investors do not entirely disregard the abnormal accruals component of earnings. We also analyze the relation between cash flow and lagged values of earnings surprise components. This analysis provides additional support for the conclusion drawn by Subramanyam (1996). Our results yield strong evidence that abnormal accruals are positively related to future cash flows. This helps to explain why abnormal accruals have positive marginal value. Overall, our evidence suggests that investors are sophisticated consumers of accounting information. They value abnormal accruals because these provide signals about future cash flows and these signals are often valid. Our results do not, of course, fully inform us how to interpret correctly the results of Xie (2001) and others who find that abnormal accruals are negatively related to measures of subsequent abnormal returns. We are stuck with the joint hypothesis problem so famously described by Fama (1970). We do not know if this negative relation means that the stock market is inefficient, or if it means that the models of risk-adjusted normal returns are all bad. Our event study results using high frequency return data substantially mitigate this joint hypothesis problem because cumulative errors in risk adjustments made over short holding periods are likely to be negligible. We find that investors apparently distinguish in subtle ways between the components of earnings and that there is a rational basis for abnormal accruals to have positive marginal value. These findings make us leery of too quickly embracing the view that the stock market, in general, inefficiently impounds the information contained in abnormal accruals. 21 21
References Abarbanell, J. S., 1991. Do analysts' earnings forecasts incorporate information in prior stock price changes?, Journal of Accounting and Economics 14, 147-165. Ball, R. J. and P. Brown, 1968. An empirical evaluation of accounting income numbers. Journal of Accounting Research 6, 159-178. Beaver, W.H., 1968. The information content of annual earnings announcements. Journal of Accounting Research supplement to 6, 67-92. Bowen, R., Burgstahler, D., and L. Daley 1987. The incremental content of accrual versus cash flow data, The Accounting Review 62, 723-747. Brown, L. D., G. Foster, and E. Noreen, 1985. Security analyst multi-year earnings forecasts and the capital market, American Accounting Association: Sarasota, FL. Callen, J. L. and D. Segal, 2004. Do accruals drive firm-level stock returns? A variance decomposition analysis, Journal of Accounting Research 42, 527-560. Chopra, V. K., 1998. Why so much error in analysts' earnings forecasts?, Financial Analysts Journal 54, 30-37. DeFond, M.L. and J. J. Jiambalvo, 1994. Debt covenant violation and manipulation of accruals, Journal of Accounting and Economics 17, 145-1476. DeFond, M. L. and C. W. Park, 2001. The reversal of abnormal accruals and the market valuation of earnings surprises, The Accounting Review 76, 375-404. Dremen, D. and M. Berry, 1995. Analyst forecasting errors and their implications for security analysis, Financial Analysts Journal 51, 30-42. DuCharme, L.L., Malatesta, P.H., Sefcik, S.E., 2001. Earnings management: IPO valuation and subsequent performance, Journal of Accounting, Auditing & Finance 16, 369-396. 22 22