The Value of Long-Term Accrual Management *

Transcription

1 The Value of Long-Term Accrual Management * Satiprasad P. Bandyopadhyay, Alan G. Huang and Tony S. Wirjanto This Version: February, 2010 Abstract In this paper, we examine both value-enhancing and value-detrimental effects of smooth earnings on forward-looking value proxies of future returns from one-month to five-year horizons. Some prior studies, on one hand, find that smoothed earnings signal superior performance, enhancing thereby firm value; some others, on the other hand, argue that smooth earnings achieved through managerial intervention are noisy. We capture the potentially harmful long-term managerial intervention aspects of earnings smoothing by multi-period accruals volatility that increases with smooth earnings. Consistent with previous literature, we treat the correlation ( correlation ) between cash flows and accrual as the value enhancing aspect of earnings smoothing. In multivariate regressions that control for a battery of risk factors, future returns are positively related to earnings smoothing of the correlation variable but negatively associated with accrual volatility. Moreover, economic significance tests performed in this paper show that the value detrimental effects of accrual volatility completely wipe out future returns increases associated with any increase in correlations. These results are particularly pronounced in discretionary accruals and in firms with high information asymmetry and underlying operating uncertainty. The evidence suggests that due to its large accrual volatility effect, excessive long-term accrual management, whether voluntary or involuntary, is valuedetrimental. * We thank Asher Curtis, Si Li, and seminar participants at the 2010 FARS Mid-Year Meeting. All errors are ours. School of Accounting and Finance, University of Waterloo, Waterloo, ON N2L 3G1 Canada, bandy@uwaterloo.ca, tel: ext School of Accounting and Finance, University of Waterloo, Waterloo, ON N2L 3G1 Canada, aghuang@uwaterloo.ca, tel: ext School of Accounting and Finance, and Department of Statistics and Actuarial Science, University of Waterloo, Waterloo, ON N2L 3G1 Canada, twirjant@uwaterloo.ca, tel: ext

2 The Value of Long-Term Accrual Management Abstract In this paper, we examine both value-enhancing and value-detrimental effects of smooth earnings on forward-looking value proxies of future returns from one-month to five-year horizons. Some prior studies, on one hand, find that smoothed earnings signal superior performance, enhancing thereby firm value; some others, on the other hand, argue that smooth earnings achieved through managerial intervention are noisy. We capture the potentially harmful long-term managerial intervention aspects of earnings smoothing by multi-period accruals volatility that increases with smooth earnings. Consistent with previous literature, we treat the correlation ( correlation ) between cash flows and accrual as the value enhancing aspect of earnings smoothing. In multivariate regressions that control for a battery of risk factors, future returns are positively related to earnings smoothing of the correlation variable but negatively associated with accrual volatility. Moreover, economic significance tests performed in this paper show that the value detrimental effects of accrual volatility completely wipe out future returns increases associated with any increase in correlations. These results are particularly pronounced in discretionary accruals and in firms with high information asymmetry and underlying operating uncertainty. The evidence suggests that due to its large accrual volatility effect, excessive long-term accrual management, whether voluntary or involuntary, is valuedetrimental. 1. Introduction The practice of earnings smoothing is pervasive among companies in North America. 1 A number of recent studies provide evidence to suggest that earnings smoothing creates value. Specifically these studies find that earnings smoothing leads to higher contemporaneous valuation metrics such as market value (Allayannis and Simko, 2009), price to earnings ratio (Barth, Elliott and Finn, 1999; Hunt, Moyer and Shevlin, 2003; and Francis et al., 2003), or lower cost of debt or equity capital (Francis et al., 2005; and Li, Richie and Tucker, 2009). In addition, earnings smoothing is also shown to improve the informativeness of current prices for future earnings or firms information environment (Tucker and Zarowin, 2006). In the aforementioned studies, earnings smoothing is measured in a number of ways, such as (low levels of) volatility of reported earnings as a 1 See, for example, Smith et al. (1994), DeFond and Park (1997), and more recently, Graham et al. (2005). 1

3 proportion of cash flow volatility, and large negative correlation between discretionary accruals and pre-managed earnings (or between cash flows and accruals). In summary, the empirical evidence assembled by these prior studies links higher current firm value to historical efforts in managing accruals to smooth earnings over time. 2 Despite the findings that income smoothing creates value, some studies point to their potential harmful effects. For example, Leuz et al. (2003) contend that management uses earnings management to conceal firm performance from outsiders in order to protect their private control benefits; as such, earnings are less informative about future firm performance. Schipper and Vincent (2003) raise the point that the process of converting a volatile cash flow stream into a smooth net income stream by means of accruals manipulation introduces transitory elements or measurement error in the accruals and, in the process, lowers earnings quality. However, to the best of our knowledge, there has been little systematic effort in the literature to examine both the benefits and costs (harmful effects) of earnings smoothing at the same time. This paper provides such an attempt. We begin our analysis by noting that while income smoothing reduces the variability of earnings, it also increases the variability of accruals for a given level of cash flow volatility. The literature typically uses as the earnings smoothing measure either the ratio of earnings volatility to cash flow volatility, EV / CFV, or the correlation between cash flows and accruals, ρ (see, for example, Francis et al. 2003; Hunt, Moyer, Shevlin 2003; and Tucker and Zarowin 2006). These are multi-period measures benchmarked against cash flow volatility. Managers of earnings-smoothing firms (with either a low EV / CFV or a high negative ρ ) manage accruals to ensure that changes in cash flows over time are offset by the opposite changes in accruals over multiple periods. 2 In general, this empirical finding supports the signaling hypothesis as opposed to the garbling hypothesis. A number of authors view smooth earnings as signals of firms superior performance (Ronen and Sadan, 1981; Chaney and Lewis, 1995; Fudenberg and Tirole, 1995; and Demski, 1998). Truman and Titman (1998) suggest the smooth earnings reduce the probability of bankruptcy and thus (indirectly) reduce smoothed firms cost of capital. Goel and Thakor (2003) argue that smooth earnings reduce information asymmetry and are valued. Kirchensheiter and Melamud (2002) suggest that the market favours smooth earnings because they enhances the precision of earnings numbers. 2

4 Given that accruals are mean-reverting over the long run and also that accruals absorb shocks in cash flows over time, for a given level of cash flow volatility, the greater the extent of earnings smoothing, the larger the magnitude of accrual volatility. Therefore, there are two sides to multi-period earnings smoothing: ρ and accrual volatility. The former measures the degree of earnings smoothing relative to cash flows, and the latter measures the absolute degree of earnings smoothing (Gu, Lee and Rosett, 2005). While the former is value-enhancing, as discussed in the literature, the latter appears to be value-detrimental, as implied in Leuz et al. (2003) and Schipper and Vincent (2003). The value-creating properties of smoothed earnings discussed earlier also appear to be in conflict with the well known accrual anomaly. First reported by Sloan (1996), the accrual anomaly literature shows that the level of accruals is negatively associated with future stock returns. 3 To see why the value-relevance evidence of earnings smoothing is seemingly at odds with the accrual anomaly, consider the role of accruals in earnings smoothing. The identity that earnings equal cash flows plus accruals implies that if a firm smoothes earnings in order to reduce its earnings volatility, it must be the case that either cash flows are not volatile enough, or that cash flows are volatile but are offset by large changes in accruals to reverse cash flow shocks. In other words, for a firm with a given high level of volatile cash flow stream but a low level of earnings volatility, earnings smoothing by inducing lower earnings volatility creates large (absolute) values of accruals. While smooth earnings reflect a higher current price, as suggested by the value-enhancing literature of earnings smoothing, large values of accruals indicate a lower future return by the accrual anomaly. Taken together, this implies that earnings smoothing leads to higher current prices, but only to be followed by lower (or even negative) future returns. This argument casts doubt on the ability of earnings smoothing activities to continue to create value over a longer time period. 3 Sloan (1996) shows that accruals are negatively correlated with future size-adjusted returns. This finding is robust to different country settings (Pincus et al., 2007), different sample periods (Lev and Nissim, 2006; Mashruwala et al., 2006), the inclusion of Nasdaq firms (Lev and Nissim, 2006; Mashruwala et al., 2006), alternative definitions of accruals (Xie, 2001; and Hribar and Collins, 2002), and considerations of additional risk/mispricing factors (Collins and Hribar, 2000; Mashruwala et al., 2006; and Hirshleifer et al., 2009). 3

5 The link between the value-relevance of income smoothing and the accrual anomaly literature calls for a common metric for value when the costs and benefits of earnings smoothing are examined. Based on the accrual anomaly literature and also on the portfolio performance evaluation literature, 4 we use forward-looking future returns to compare the value relevance of the two accruals-related effects, namely, relative earnings smoothing and its counterpart, high accrual volatility. Thus, in contrast to the earnings smoothing literature which typically uses current prices to assess the value of earnings smoothing, we ask whether earnings smoothing creates a value for shareholders over time. We confirm that earnings smoothing, as measured by EV / CFV or ρ over the past sixteen quarters, is associated with higher future returns for the NYSE/Nasdaq/AMEX firms during the period of In particular, in multivariate cross-sectional regressions that control for the Fama-French three factor variables of stock return beta, size and book-to-market-equity and Carhart s (1997) price momentum (hereafter FF4- factor variables ), earnings smoothing is positively related to returns over one-monthahead to five-year-ahead horizons. When we add additional return-informative variables of earnings momentum (Chan, Jegadeesh and Lakonishok, 1996), earnings yield (Haugen and Baker,1996), illiquidity (Amihud, 2002), and idiosyncratic return volatility (Ang et al., 2006) to the empirical models, the predictive powers of earnings smoothing are weakened over the horizons that extends from one year to five years. Since we show that the earnings-smoothing results using ρ and EV / CFV as the smoothing metrics are broadly in conformity with extant findings in the literature, we further examine whether accrual volatility has an adverse effect on firm value as implied by studies such as Leuz et al. (2003) and Schipper and Vincent (2003). We achieve this by including accrual volatility as an additional control variable in our multivariate regression models. We find that the smoothing effect (as measured by ρ 4 Future returns or risk-adjusted future returns (Jensen s alpha) are widely used for portfolio performance evaluation in the literature and in practice. For example, the question of whether a mutual fund manager creates value is often addressed by examining whether the fund under his/her management has a positive alpha. 4

6 or EV / CFV ) is significantly positive for most future return horizons even in the presence of accrual volatility as an additional control variable. In contrast, in specifications that include (i) the FF4-factor variables, (ii) the additional return-informative variables as mentioned above, and/or (iii) the earning-smoothing variable, accrual volatility is found to be significantly negatively related to future returns from one-month to five-year-ahead. Given the findings of a robust association between accrual volatility and future returns, we proceed to compare the economic significance of earnings smoothing and accrual volatility by examining the magnitude of change in return resulting from a standard deviation change in both variables. 5 In all of the cases considered in this paper, the economic significance of accrual volatility is always found to be much greater than that of earnings smoothing. That is, any higher future returns generated from smooth multiperiod earnings are found to be more than offset by the negative return induced by volatile accruals. For example, in the regressions that use all of the above returninformative variables to explain one-year-ahead return, the economic significance of ρ CF,ACC is 0.39% (with positive future returns observed 68% of times), but the economic significance of accrual volatility significance is -1.31% (with negative future returns observed 75% of times). Thus, if one manages earnings to cause one standard deviation decrease in ρ but one standard deviation increase in accrual volatility at the same time, the net decrease in annual return is 0.92% (t-statistic = 5.23, 57% of times observed to be negative). We therefore conclude that the combined effect of earnings smoothing and accrual volatility, or what we coin in this paper as the long-term accrual management effect, is negative. Next we provide an economic interpretation for why earnings smoothing is value enhancing when measured as ρ or EV/CFV, but value detrimental when accounting for the effect of accrual volatility. A standard rationale in the literature for the value of 5 The economic significance of a variable is defined as the coefficient estimate on the variable times the standard deviation of the variable. The use of standard deviation addresses the problem that earningssmoothing of ρ or EV/CFV and accrual volatility are not measured in the same unit (and therefore a comparison of their marginal effects is misleading). 5

7 earnings management is that earnings management serves to signal private knowledge of firm value to external shareholders (e.g. Demski and Sappington, 1990), or to induce efficient compensation contracting when there is information asymmetry between the principal (the owner of the firm) and the agent (the manager of the firm) (e.g. Dye, 1988 and Evans and Sridhar, 1996). We show that firms with more negative values of ρ have a lower degree of information asymmetry based on information asymmetry measures such as the (absolute) difference between the actual EPS and the analyst consensus forecast, firm size, the ratio of expense used to generate sales, and the asset utilization ratio. If smaller information asymmetry signals a higher firm value, as is usually the case in the literature, then ρ can be value-enhancing in this regard. On the other hand, accrual volatility, ACCV, provides a direct measure for the consistency (or lack of) of earnings disclosure of the firm, as accruals add noise to the earnings stream. Easley and O Hara (2004) argue that a firms treatment of accounting accruals impacts the information allocation process. Hence, ACCV can be regarded as a measure of information uncertainty. Recent literature documents a negative cross sectional association between future returns and historical information uncertainty proxies (Zhang, 2006), such as idiosyncratic return volatility (Ang et al., 2006), analyst forecast dispersion (Diether et al., 2002), and cash flow volatility (Huang, 2009). ACCV is correlated with these information uncertainty measures. Thus, the fact that ACCV leads to a lower return can be treated as an information uncertainty phenomenon. We also identify a subset of accruals and a subset of firms that the dominance of accrual volatility effect is particularly acute. For the former, we find that the dominance of accrual volatility over smoothing measures (namely, ρ and EV/CFV) persists regardless of whether total accruals, discretionary accruals, or changes in working capital are used as the definition for accruals. Consistent with Subramanyam (1996) and Xie (2001) who emphasize the importance of discretionary accruals in generating the accrual anomaly, the economic significance of discretionary accrual volatility is found to be the highest among the three definitions of the accruals analyzed. It follows that discretionary 6

8 accruals can be identified as a subset of accruals whose volatility is particularly valuedetrimental. The subset of firms with the most pronounced dominance of the accrual volatility effect is those with high information asymmetry and underlying operating uncertainty. We partition the sample firms into four quadrants based on the median values of ρ and accrual volatility. We find that (1) cash flow volatility is almost monotonically increasing across these quadrants, (2) accrual volatility dominates ρ in all four quadrants, and (3) the dominance is the most pronounced in the quadrant with a less negative value of ρ CF,ACC (more information asymmetry) and high accrual volatility. These results suggest that for firms with greater operating uncertainty and information asymmetry, the resulting high accrual volatility is extremely value detrimental even though a smoothing measure, such as ρ, still provides a channel for value signaling. Our final robustness check focuses on the relation between our findings and the accrual anomaly phenomenon. To the extent that accruals are mean reverting over time, high accruals may be taken to imply high accrual volatility and, therefore, the accrual volatility effect may simply be interpreted as a rediscovery of the accrual anomaly. However we show in this paper that this is not the case, at both the firm and portfolio levels. At the firm-level, the accrual volatility effect is shown to remain strong in multivariate regressions that control for the level of accruals. In addition, at the portfoliolevel, returns on the portfolios first sorted by the level of accruals (and thus controlling for accruals) and then by accrual volatility are shown to retain a strong decreasing pattern. This paper contributes to the literature in a number of ways. First, we use the notion of future return to assess the value of earnings smoothing, thus emphasizing future values for earnings smoothing activities. While the use of future returns to examine the effectiveness of specific trading strategies is widely used in the portfolio management literature as well as in the money management industry, it has not been widely adopted in the pricing of smooth-earnings literature. Secondly, we argue that there are two different 7

9 channels through which earnings smoothing can affect accrual management, namely, (i) increased correlation between cash flow and accruals, and (ii) increased accrual volatility. So far the literature has focused on the first channel only but not on both channels simultaneously. We show that although earnings-smoothing activities via the correlation is value enhancing, as is claimed in the literature, earnings-smoothing activities via accrual volatility is not. Moreover, the loss in value arising from accrual volatility dominates the value created via ρ. This dominance of accrual volatility persists for total accruals, discretionary accruals, and change in working capital. Thus, our evidence shows that an excessive long-term accrual management, whether voluntary or involuntary, is value-detrimental due to its large negative accrual volatility effect. Thirdly, we provide a resolution between two seemingly conflicting results in the literature, namely, the value-enhancing effects of smooth earnings versus value-detrimental effects of high levels of accruals arising from earnings smoothing in the accrual anomaly literature. We demonstrate that the former can be due to value signaling and the latter can be attributed to to the noisiness of accruals in the earnings management process and hence the information uncertainty effect. The rest of the paper is organized as follows. Section 2 describes the sample and variable definition used in this study. Section 3 presents the main results of the paper and Section 4 provides the economic interpretation of ρ and ACCV. Section 5 offers a discussion of the results and presents robustness checks on the results. Section 6 concludes the paper. 2. Sample and Variable Measurements 2.1 Sample Our study requires the estimation of accrual volatility and the measure of the correlation between accruals and cash flows, for which the sample should ideally contain as many time-series observations as possible. Therefore, unlike many previous studies that use annual data (e.g. Sloan, 1996; Xie, 2001), we use quarterly data to increase the number of 8

10 observations. We select all NYSE/NASDAQ/AMEX-listed firms on the merged CRSP/Compustat database and extract monthly stock returns and quarterly accounting items for the 1976 to 2007 period. 6 To ensure that accounting information is known prior to trading, we match monthly stock returns to accounting numbers reported for the prior fiscal quarter on the calendar basis. Finally, we eliminate financial services companies (SIC code between 6000 and 6999) and observations with negative assets or negative sales. 2.2 Measures for Accruals, Accrual Volatility, and Earnings Smoothing We follow the accrual anomaly literature in defining total accruals (e.g. Sloan, 1996), operating accruals (e.g. Fama and French, 2008), and discretionary accruals (e.g. Xie, 2001). Our primary measure for accruals is total accruals, derived from the cash flow statement. Following Richardson et al. (2005), we define operating accruals as changes in net working capital with respect to a previous quarter, where net working capital is defined as: (Current assets Cash and short-term investment) (Current liabilities Debt in current liabilities). 7 We define total accruals as operating accruals minus depreciation. Discretionary accruals, or the so-called abnormal accruals (Xie, 2001), are estimated using the following model due to Jones (1991) for each industry: ACC Sales ΔSalesi, j, t PPEi, j, t = β + β1, j + β 2, j + ε i, j, t (1) Sales Sales Sales i, j, t 1 0, j i, j, t i, j, t i, j, t i, j, t where ACC is total accruals; PPE is property, plant and equipment; i indexes firm; j indexes industry (defined as two-digit SIC code); t indexes quarter; and Δ Sales i, j, t is the 6 We select 1976 as the beginning of our sample because quarterly data items that are necessary for computing accruals exist in a large scale from 1976 onwards. 7 This definition slightly differs from that used in Sloan (1996) in that we drop income taxes payable from the second difference term. The purpose for this is to retain the largest possible sample size, as there are many missing observations in income taxes payable in the quarterly data tape. 9

11 change in sales with respect to the quarter of a year ago. The estimate of the error term, ε i, j, t, from (1) is taken to be a measure of discretionary accruals. We estimate Equation (1) with past four years data and by rolling forward every month. The rolling estimation is intended to ensure that when we use discretionary accruals-based measures to predict returns, the conditioning information (i.e., the discretionary accruals) is known to us. We use the same quarterly accounting numbers for each month within each quarter. We choose sales as a deflator in Equation (1) for the following reasons. First, the sales variable is less affected by accruals than total assets. Accrual adjustments to operating earnings directly affect the level of assets and book equity, making it more difficult to disentangle the net effect of accruals. The sales variable, on the other hand, is less likely to be affected by the choice of accruals. Second, the sales variable has been used in the prior literature as a measure of firm size (e.g. Berk, 1997). Third, using the sales variable as the deflator greatly reduces the seasonality effect in the quarterly accruals. There is substantial evidence that accounting variables exhibit significant seasonality (e.g. Brown, 1993). It is well known that in the presence of seasonality, using the raw data to estimate the second moments (e.g. correlation and standard deviation) would impart upward biases in the coefficient estimates in multiple regressions. For example, during our sample period, the full-sample lag-four autocorrelation between the current quarter total accruals and the total accruals of a year ago is In contrast, the lag-four autocorrelation of accruals to sales is only We take this as some evidence that scaling accruals by sales is useful in addressing the seasonality in accruals. Based on the above accrual proxies, we compute accrual volatility as the rolling standard deviation of the standardized accruals (accruals scaled by sales) over the past sixteen quarters (four years). Within the same estimation window, we also compute earnings volatility, cash flow volatility and ρ analogously, where cash flow is defined as earnings minus total accruals, again scaled by sales. We require at least eight non-missing observations of accruals within this estimation window. Although the choice of the estimation window of four years is somewhat arbitrary, we find that virtually all our results are robust to estimation windows of three years (twelve quarters) and five years 10

12 (twenty quarters). When we adjust for the first four years needed for accrual volatility calculation, our usable sample period reduces to the period of We label earnings volatility as EV, cash flow volatility as CFV, and accrual volatility based on total accruals, operating accruals and discretionary accruals as ACCV, CWCV, and DAV, respectively. EV/CFV and ρ are the standard measures of earnings smoothing (e.g., Leuz et al., 2003; Myers et al., 2003; Rountree et al. 2008; Allayannis and Simko, 2009). 8 For brevity, we present most of our results using ACCV, and report results using CWCV and DAV only as a robustness check. 3. The Value of Long-Term Accrual Management 3.1 Research Method Regression Specification We test the value of long-term accrual management (earnings smoothing versus accrual volatility) at the firm level using Fama and MacBeth s (1973) cross-sectional methodology. In the two-step procedure, the regression model is first run every month, and then the monthly coefficient estimates are averaged to obtain a full-sample estimates and t-statistics. We follow Fama and French (1992) to use beta, size and book-to-market equity as our first set of control variables since they represent the factors of market, SMB and HML in the well-known Fama-French three-factor model. In addition, we also include other return-informative measures as the control variables, namely, price momentum (Jegadeesh and Titman, 1993), earnings momentum (Chan, Jegadeesh and Lakonishok,1996), earnings yield (Haugen and Baker, 1996), illiquidity (Amihud, 2002), and idiosyncratic return volatility (Ang et al., 2006). To test the value-relevance of longterm accrual management, we augment the above set of variables with two additional 8 We also use the correlation between the change in (scaled) accruals and the change in (scaled) cash flows as an alternative to ρ for the measure of earnings smoothing (e.g. Leuz et al., 2003). These two measures are highly correlated in our sample, with a correlation coefficient of Therefore, none of our results is expected to change substantively with this alternative correlation measure. 11

13 variables, ρ and ACCV, in the cross-sectional multiple regression for every month t. The final estimating equation is given below: R i, t+ 1 = i, t + γ 1, t β i, t + γ 2, t ln( ME) i, t + γ 3, t ln( BEME) i, t α + γ PMOM + γ + 6, t i, t + γ 7, t ILLIQi, t + γ 8, t IRVi, t + γ 9, t ( ρcf, ACC ) i, t + γ 10, t ACCVi, t 4, t i, t it 5, t SUE γ EY + v (2) i, t where subscript i indexes stock, subscript t indexes month, R is raw return, ME is lagged market equity, BEME is book-to-market equity, PMOM is price momentum, SUE is standardized unexpected earnings, EY is earnings yield, ILLIQ is illiquidity, and IRV is idiosyncratic return volatility. Note that, in principle, the measure ρ or ACCV can be replaced by other measures of earnings smoothing or accrual volatility. We also note that although we use t+1 to label future returns, we use one-month- to five-year-ahead returns as discussed in the previous sections. The measurement of the variables in regression equation in (2) follows prior literature. Specifically, beta is measured as the CAPM beta estimated from the past five years monthly returns, ME is the market equity at the beginning of the month, BEME is book value equity of the month to ME, PMOM is the cumulative return of past 12 months, and EY is earnings to ME. Following Chan, Jegadeesh and Lakonishok (1996), earnings momentum (SUE) at month t is defined as: SUE i, t e = i, q e σ i, t i, q 4 where e i, q is the most recent quarterly earnings, ei, q 4 is earnings four quarters ago, and e σ i,t is the standard deviation of unexpected earnings ( i, q i, q 4 e ) over the preceding sixteen quarters. Following Amihud (2002) and using daily returns, illiquidity (ILLIQ) for month-t is defined as: ILLIQ i, t daily returni previous within month mean of daily volumei = cross sectional average of the numerator 12

14 The ILLIQ ratio gives the absolute percentage price change per dollar of daily trading volume, or the daily price impact of the order flow. The higher the ratio, the more illiquid is the stock. Following Ang et al. (2006), we define idiosyncratic return volatility (IRV) relative to the Fama-French three-factor model. Specifically, every month we run the following regressions of daily excess stock for firm i using daily returns within month t: R i, t R f, t = ai + bmkt, imktt + bsmb, ismbt + bhml, ihmlt + ui, t (3) where R, is the day-t stock return during the month, i t R, is the riskfree rate, MKT, SMB f t and HML are the daily Fama-French three factors of market, size and value, respectively, and u i, t is the error term. The standard deviation of the estimate of i t u, from Equation (3) within the month is used to proxy IRV for the next month. Except the logarithm of market equity and the logarithm of book to market, all of the variables in Equation (2) are winsorized at the 1 st and 99 th percentiles over the full sample to reduce the effect of outliers Comparison of Economic Significance between Earnings Smoothing and Accrual Volatility The coefficient estimates of γ 9, t and γ 10, t in the regressions in (2) represent, respectively, the marginal effect of ρ and ACCV on returns. If ρ and ACCV are measured on the same unit, their economic significance equals the marginal significance. Note that, however, ρ CF, ACC is measured unit free but ACCV is not. Thus in order to compare the economic significance of ρ and ACCV, we need to standardize both variables. The standardization is achieved through the following transformation, similar to standardizing a normally distributed random variable: ( ρ CF, ACC ) ( ρ i, t σ ρ, t ) t, and 13

15 ACCV i ACCV σ, t ACCV, t t, where ρ and σ ρ, t are the cross-sectional mean and standard deviation of ( ) t (ρ ) i, t at time t, respectively. ACCVt and σ ACCV, t have an analogous interpretation. Replacing (ρ ) i, t and ACCV i, t with their standardized counterparts in the regressions in (2) results in new coefficient estimates that equal γ 9,t σ ρ, t and γ 10,t σ ACCV, t without affecting the statistical significance of ρ CF, ACC and ACCV. Therefore, the quantities γ σ and γ 10,t σ ACCV, t can be used to compare the economic significance of ρ CF, ACC and 9,t ρ,t ACCV. In other words, we compare the economic significance of earnings smoothing and accrual volatility by examining the magnitude of change in return resulting from a standard deviation change in both variables. We are also interested in the aggregate effect of the long-term accrual management. Firms that manage accruals at both the relative and absolute levels over the long term will likely see a more negative ρ and a larger ACCV. Therefore, the standardized effect of long-term accrual management on return due to ρ is given by 9,t ρ, t γ σ (because ρ CF, ACC becomes more negative, or smaller), and due to ACCV is given by γ 10,t σ ACCV, t (because ACCV becomes larger). The aggregate effect of long-term accrual management is thus given by the following expression: γ ρ + 9, t σ, t γ 10, tσ ACCV, t. Corresponding to the cross-sectional nature of the regressions specified in (2), we can derive a time-series counterpart of these effects. Using these time series, we then can compute the estimate and significance in a manner analogous to the derivation of estimate and significance of regression coefficients in the Fama-MacBeth regressions. 3.2 Univariate Characteristics of the Sample and the Relationship between Earnings Smoothing and Accrual Volatility 14

16 Table I provides the descriptive statistics of the major variables for the sample. We make several brief observations. First, our sample is relatively large. In Table I, ρ has the smallest number of observations of 336,512 firm-quarters in variables in Equation (2). This means that the average number of firms used in the regression model is about 3,000 ( 336,512 / (28 years 4 quarters) ). Second, firms appear to engage in earnings smoothing quite extensively. The mean (median) value of ρ is (-0.93). Third, the standard deviations of ACCV and CWCV are also relatively large. Since both variables are positive, this magnitude of standard deviation implies that the right tail of the accrual volatility distribution is very thick. This perhaps is not surprising given that our sample consists of many small firms, as illustrated by the distribution of market value of equity (note that the mean is much larger than the median). 9 [Table I about here.] Panel A, Table II provides the correlation matrix of the variables. Corresponding to our cross-sectional regressions, we provide the time-series means of cross-sectional correlations. Since our variables for earnings smoothing and accrual volatility are estimated using a rolling window, estimating the full sample correlations this way also helps us to address the autocorrelation issue common to these variables. Some brief comments on these correlations are in order here. First, earnings smoothing and accrual volatility measures are all significantly correlated with future returns, with magnitudes comparable with other return-informative variables such as book-to-market and price momentum. The correlation between earnings smoothing and returns is negative, implying that higher levels of earnings smoothing, as evidenced by smaller (or more negative values) of ρ CF, ACC or EV/CFV,) are associated with higher returns. On the other hand, the correlation between accrual volatility and returns is negative, implying that larger values of accrual volatility lead to lower returns. Second, the three accrual volatility measures (ACCV, CWCV, and DAV) are highly correlated with one another (e.g. 9 The dominance of small firms also explains why earnings and cash flow are on average negative in our sample. 15

17 the correlation between ACCV and CWCV is 0.95, and between ACCV and DAV is 0.71). The two earnings smoothing measures ( ρ and EV/CFV) are also highly correlated (with a correlation coefficient of 0.84). This high degree of correlation motivates our use of ACCV as the primary accrual volatility measure. [Table II about here.] The results reported in Panel A of table II shows that the correlation between earnings smoothing measures and accrual volatility measures is positive. This indicates that unconditionally, firms with small values of ρ, reflecting high levels of relative earnings management, exhibit smaller absolute amounts of accrual volatility. While this might appear somewhat puzzling at first sight, it is important to note that the above positive unconditional correlation between ρ and ACCV does not take into account firms differing underlying business volatility. For example, if the underlying business of a firm becomes more volatile, in order to engage in earnings smoothing, the firm naturally will exercise a large absolute amount of accruals (resulting in large ACCV). However, it might be the case that the amount of accrual manipulation is not enough to keep up with the growing volatility in the underlying business, resulting in a less negative value of ρ and, thus, a positive association between ACCV and ρ. Therefore, our earlier claim that a higher degree of earnings smoothing, as measured by a negative value of ρ, leads to larger accrual volatility and thus a negative association between ACCV and ρ CF, ACC, is only meaningful for a given level of cash flow volatility. We find that this is indeed the case, as Panel B, Table II shows a negative partial correlation between ρ and accrual volatility after we control for cash flow volatility. Using cash flow volatility and firms industry group (as defined by 2-digit SIC code) to capture the underlying business volatility, we recalculate the correlation between earnings smoothing and accrual volatility after controlling for business volatility for each industry group. Panel B, Table II reports the time-series means of industry-wise, crosssectional partial correlations between earnings smoothing ( ρ or EV) and accrual 16

18 volatility (ACCV, CWCV, and DAV). Note that for the second earnings smoothing measure, we use EV in lieu of EV/CFV, this is because cash flow volatility is already controlled for and hence CFV is no longer required as a deflator. As expected, the partial correlations between all these earnings smoothing measures and accrual volatility measures are significantly negative. For example, the partial correlation between ACCV and ρ (EV) is highly significant at ( ). 3.3 Regression Results Table III presents the regression results of Equation (2), using ACCV and ρ as the two long-term accrual management variables. In Panel A, we present various model specifications for one-month-ahead returns, and in Panels B to E, we report the full model results for 6-month to 5-year-ahead returns. [Table III about here.] Let us examine the value of earnings smoothing first. From Panel A, the predictive power of ρ on one-month return depends on the model specification. In the univariate case (Model 11) and the case with the Fama-French four-factors variables of beta, size, book to market and price momentum (Model 21), the coefficient estimate of ρ is significantly negative, as expected. However, the significance of ρ vanishes when we add other return-informative variables of SUE, EY, ILLIQ, and IRV (Model 31). Furthermore, adding ACCV does not change the sign and significance of materially (Models 13, 23, and 33). ρ CF, ACC For longer-term returns in Panels B to E, the value of earnings smoothing is again somewhat mixed. The estimated ρ is not significant in explaining six-month returns. It is, however, significant in explaining one-year to five-year returns. Although the results are not tabulated here, we can report that after we control for the Fama-French four-factor variables, the estimate of the coefficient on ρ becomes significantly 17

19 negative in explaining the six-month-ahead returns. Overall, although the evidence is not uniformly affirmative, the results presented in Panels A-E can be interpreted as being supportive to the value-relevance argument of earnings smoothing in general. Let us now turn to accrual volatility. The most telling result from Panel A is that accrual volatility negatively predicts one-month-ahead returns, consistent with the previous portfolio-sorting results. In Model 12 when ACCV is the only explanatory variable in the regression models, its coefficient estimate is (with a t-statistic of -3.40). In Model 22, when combined with the Fama-French four factors, the coefficient estimate of ACCV drops to (with a t-statistic of -2.79). In Model 32, a further addition of the remaining control variables in (2) reduces the ACCV coefficient estimate slightly to (with a t-statistic of -2.32). In these models, a further inclusion of ρ in the regressions barely changes the magnitude and significance of ACCV (Models 13, 23, and 33). Panels B to E show that the accrual volatility effect identified in Panels A remains strong for six-month to five-year ahead returns. The coefficient estimate of ACCV in the regression models remains significantly negative for all of these horizons, and actually with a larger magnitude of t-statistics than the case of one-month-ahead return. In terms of the magnitude of the ACCV effect over these longer horizons, we note that it remains roughly at the same level at an annualized loading of -.80 to The benchmark annualized loading of ACCV using a one-month-ahead return is -.84 (= ). The annualized loading using a six-month-ahead return is -.98 (= ), and using a oneyear-ahead return the annualized loading is Using a two-year-ahead return the annualized loading reduces to -.74 (= -1.48/2), but it increases to (= -7.40/5) when using a five-year-ahead return. The signs on the control variables in Panel A are by and large unremarkable. We use onemonth return as the benchmark base as this is the most widely used return measure in the literature. In Panel A, beta is generally not significant, as is often shown in the literature (e.g. Fama and French, 1992). Consistent with the finding of disappearing size effect 18

20 after the 1980s, the coefficient on size (ln(me)) is insignificant in the four-factor specifications. However, when the other return-informative variables are introduced into the regression models, the sign on size becomes significantly negative, as expected. Finally, the coefficients on book to market, price momentum, earnings momentum, earnings yield, and illiquidity are all significantly positive, and the coefficient on IRV is significantly negative, as expected. Most of these results also hold for longer-horizon returns, except that the signs on beta and earnings yield are negative for two- and fiveyear returns, price momentum effect disappears at a five-year horizon. The negative sign of beta to explain stock returns is also noted in the literature (e.g. Petkova, 2006), and it is known that the price momentum effect does not last over two years (e.g. Jegadeesh and Titman, 1993). Table IV repeats the exercises of Table III, except that EV/CFV is used as the earnings smoothing measure. In terms of the significance, the only difference between Table IV and Table III is that in Table IV EV/CFV is marginally significant for a one-year return (Panel C). The similarity of the results reported in Table III and IV is probably not surprising given a high degree of correlation between ACCV and ρ. Again, Table IV reinforces the findings that earnings smoothing (as proxied by ρ or EV/CFV) is generally value enhancing, but accrual volatility is universally value detrimental at least for the cases presented so far. [Table IV about here] 3.4 Dominance of Accrual Volatility over Earnings Smoothing Table V shows the economic significance of earnings smoothing and accrual volatility. In Panel A, we present the economic significance of ρ CF, ACC and ACCV based on the regression results in Tables III, and in Panel B, we present the economic significance of EV/CFV and ACCV based on the regression results in Tables IV. To save space, we only report the results using the full-model specification (e.g., Model 33 that includes both earnings smoothing and accrual volatility in both Tables). The results continue to hold in other specifications as well. 19

21 [Table V about here.] As the results reported in Panels A and B are almost identical, we focus our ensuing discussion on Panel A only (which uses ρ as the proxy for earnings smoothing). The measure ρ is not economically significant for one-month to one-year returns but is economically significant for two- to five-year returns. This is largely consistent with the results reported in Table III. For example, for a two-year-ahead return, the economic significance of ρ, γ 9,t σ ρ, t, is found to be, on average, 0.69% (with a t- statistic of 2.88), indicating that smoothing earnings by a standard deviation change in ρ will increase a two-year-ahead return by 69 basis points. In contrast, the economic significance of ACCV is unambiguously negative across all of the horizons considered in the study. In addition, we find that the significance of ACCV is much stronger than that of ρ. For example, the economic significance of ACCV, γ, for the two-year-ahead return is found to be, on average, -2.58% (with a t- 10,t σ ACCV, t statistic of ). As a result, the aggregate long-term accrual management effect, defined as the sum of the economic significance of ρ and ACCV, is negative. Using a two-year return as an example, the aggregate economic significance of long-term accrual management is given by -1.89% (t-statistic of -7.47). These results suggest that although earnings smoothing is value relevant, the value detriment of accrual volatility more than offsets the positive return, if there is any, generated by earnings smoothing. In other words, accrual volatility dominates earnings smoothing and renders long-term accrual management to be harmful in value. We also show that the dominance of accrual volatility over earnings smoothing is not clustered over certain periods of time. We illustrate this in two ways. In the first instance, we examine the percentage of times that a certain effect is above zero. The results are presented in Table V as well. Again, we focus on ρ but there we observe that the economic significance of ρ is above zero more than 50% of the time for most 20

22 return horizons. Consistent with the finding that the earnings-smoothing effect seems to be stronger at longer horizons, this percentage is generally increasing in return horizon (from 50% in the one-month return to 69% in the five-year return). In contrast, the percentage of times that the ACCV s economic significance is greater than zero is well below 50% all the time, and is decreasing in the return horizon (from 43% in the onemonth return to only 13% in the five-year return). As a result of the previously documented dominance of ACCV, the aggregate economic significance of the long-term accrual management is above zero for only much less than half of the time across all the horizons presented (from 41% in the one-month return to 24% in the five-year return). To further illustrate that the effect obtained above is not clustered, we plot these effects over time. Panels (a) and (b) of Figure 1 depict, respectively, the monthly time-series and the trailing 12-month average of the value effects of ρ and ACCV based on onemonth-ahead returns. Panel (c) shows the aggregate value effect of ρ and ACCV based on Panels (a) and (b). Figure 2 shows the monthly time-series and the trailing twelve-month average of the aggregate long-term accrual management effect for the sixmonth to five-year horizons. We observe that although there is a negative spike for the aggregate effect in the late 1999 and the early 2000 (which coincides with the period in which the Internet bubble bursted), the frequency of the aggregate effect being negative is visibly more than the frequency of the aggregate effect being positive. [Figures 1 and 2 about here.] 4. Economic Interpretation for the Value Effects of ρ versus ACCV As previously discussed, the earnings smoothing measures of ρ and EV/CFV, which are benchmarked against cash flow volatility, reflects a relative degree of earnings management. In contrast, the accrual volatility measure, ACCV, is derived without having to consider cash flow volatility; as such, it reflects an absolute level of earnings management. In the previous sections, we have shown that the former is value increasing 21

23 while the latter is value detrimental. In this section we explore the economics of this value divergence. In particular, we show that both ρ and ACCV reveal information contents that are consistent with their respective value implications. A standard rationale in the literature is that earnings management serves to signal private knowledge of firm value to external shareholders and to induce efficient compensation contracting when there is information asymmetry between the principal (the owner of the firm) and the agent (the manager of the firm). An example of signaling role of earnings management can be found in Demski and Sappington (1990), where the authors argue that an unmanaged performance measure (such as operating cash flows) contains some firm specific information but managers typically possess more information than what is contained in cash flows. Also managers try to communicate this private knowledge to investors through discretionary accruals. For efficient contracting, for example, Dye (1988) and Evans and Sridhar (1996) both show that when the manager has private information about the firm s economic profit, an efficient compensation contract that aims to induce the manager to exercise optimal work efforts may necessarily lead to some earnings management. In these papers, the manager s compensation is tied not only to the manager s efforts, but also to reported earnings. The demand for income smoothing emerges in equilibrium because smooth earnings elevate the manager s utility level. To see this, imagine a compensation contract that is monotonically increasing in reported earnings. If total economic earnings over multiple periods are constant, then for a manager with a concave utility function, the manager s optimal choice, ceteris paribus, is to report smooth earnings over time, since this will give him/her the smoothest stream of consumption and, hence, the highest level of utility. As to ACCV, aside from being a measure of earnings management, it also provides a direct measure for the consistency (or lack of) of earnings disclosure of the firm. This is because accruals add noise to the earnings stream by imparting transitory components in the process of converting cash flows to earnings (e.g., Schipper and Vincent, 2003; Leuz 22