Do Publicly Signalled Earnings Management Incentives Affect Analyst Forecast Accuracy?
The authors would like to acknowledge the valuable feedback received regarding this paper at the Amsterdam Business School, University of Bristol and University of New South Wales Seminar Series, and the British Accounting Association Conference 2009. We would particularly like to acknowledge the constructive comments of Nikola Petrovic, Sanjay Bissessur, David Veenman, Neil Fargher, Greg Shailer and Jeff Coulton, and two anonymous reviewers.
Abstract
Using a panel of listed Australian firms for the years 1999–2007, this paper investigates whether analysts' forecast efficiency is improved by the occurrence of a publicly observable event, such as a CEO appointment, which signals a firm's earnings management incentives. Two supporting hypotheses are also tested: first, that CEO appointments are associated with income-decreasing earnings management; and second, that analyst forecast errors increase with the level of earnings management present in current period financial statements. Consistent with prior literature, we find income-decreasing earnings management in the year of CEO appointment. Earnings management, as a general phenomenon, is found to be significantly related to analyst forecast errors in the period in which the earnings management occurs. However, we present evidence that analyst forecasts for current year earnings are significantly more accurate with respect to earnings management in cases where a CEO is appointed during the current financial period.
The economic consequences of one party's actions on third parties have, for some time, motivated a significant body of accounting research (Zeff, 1978). While accounting standards allow considerable discretion in the measurement and reporting of accruals, with the intention of allowing published financial reports to reflect the private information of well-informed insiders, this discretion also creates potential for insiders to abuse their private information when estimating accruals, a phenomenon described in the literature as ‘earnings management’. Earnings management can result in biased measures of reported profit. Financial analysts assist investors to value firms by making well-informed forecasts of the determinants of firm value, including earnings, net operating cash flow and dividends, and may potentially reduce the undesirable market effects of earnings management. Teoh and Wong (2002) report that many investors do not ‘see through’ extant systematic forecast inaccuracy. This paper seeks to examine analysts' ability to anticipate earnings management, evidence of which is of relevance to both investors and those charged with the responsibility of regulating financial reporting.
The manipulation of accruals affecting forthcoming annual earnings reports impacts the accuracy of analysts' current year earnings forecasts. In the absence of publicly observable signals of the likelihood of earnings management, signed short-horizon analyst forecast errors are likely to be associated with directional measures of earnings management. If this hypothesized association is supported empirically, this would suggest that contemporary earnings management models generate estimates of ‘unexpected accruals’ that sophisticated market participants are unable to anticipate perfectly. Where, however, publicly observable signals of likely earnings management, such as the appointment of a new CEO, occur during the current financial period, analysts may incorporate this information in their forecasts, thereby reducing earnings management-related forecast errors. A significant reduction in the sensitivity of forecast errors to earnings management in these would suggest that analysts are able to anticipate, at least partially, firms' earnings management behaviour when publicly observable signals of its likelihood exist. Using an unbalanced panel of listed Australian firms between 1999 and 2007, we investigate whether a publicly observable and well-recognized signal of the likelihood and direction of earnings management, the appointment of a new CEO, is associated with more accurate short-horizon analyst forecasts.1
We focus on the CEO appointments as our publicly observable signal of earnings management incentives for two main reasons. First, CEO appointments represent a discrete publicly observable signal of a phenomenon widely accepted as inducing earnings management in the current reporting period (Moore, 1973; Strong and Meyer, 1987; Godfrey et al., 2003). As such, CEO appointments signal both an earnings management incentive, and the particular future earnings reports argued to be influenced by that earnings management. Other events commonly associated with earnings management incentives, such as seasoned equity offerings, are most strongly associated with earnings management occurring prior to public announcement of the event (Teoh and Wong, 2002). Second, CEO appointments provide a clear directional prediction: newly appointed managers are predicted to manage earnings downwards in the appointment year. This allows a statistical examination of the relationship between predicted directional earnings management and the direction of analyst forecast errors.
LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT
Earnings Management Following CEO Appointments
Central to this paper is whether the occurrence of a publicly observable event generally recognized to be associated with earnings management, in this case the appointment of a new CEO, moderates the relationship between earnings management and forecast errors. The relevance of this question is contingent on the existence of a significant relationship between CEO appointments and firms' subsequent earnings management behaviour. Consistent with prior literature (e.g., Moore, 1973; Pourciau, 1993; Wells, 2002; Wilson and Wang, 2010), we argue that newly appointed CEOs have an incentive to manage earnings downwards in the initial reporting year of their tenure. Such earnings management may serve to attribute responsibility for a firm's poor performance to the outgoing CEO, and lower the benchmark against which the new CEO's performance is evaluated (Wilson and Wang, 2010).
Numerous studies examine the relationship between discretionary accounting decisions and CEO appointments, and report diverse results. Many studies of either specific accounting accruals, or total accruals as a proxy for earnings management, report that discretionary accounting choices reduce reported earnings in new CEOs' initial year of incumbency (Moore, 1973; Strong and Meyer, 1987; Elliot and Shaw, 1988; Godfrey et al., 2003). Others find similar evidence, but only for CEO appointments classified as ‘non-routine’ (e.g., Pourciau, 1993; Murphy and Zimmerman, 1993). In a study of the top 100 Australian firms, Wells (2002) finds only a weak relationship between ‘non-routine’ CEO changes and earnings management proxied by the behaviour of abnormal charges and extraordinary items. More recently, Wilson and Wang (2010) report significant downward earnings management in the CEO's initial year of incumbency, a phenomenon which is amplified in cases where there is also a change in company chairperson.2
Given the inconsistent results reported in the existing literature, we retest the relationship between discretionary accruals (our proxy for earnings management) and CEO appointments for listed Australian firms. Adopting Wells' (2002) formulation, H1 is stated formally below:
-
H1: In the year of a CEO appointment, income-decreasing discretionary accruals are used to decrease the reported earnings.
Analyst Forecast Accuracy and Earnings Management
We define analyst forecast accuracy as the signed forecast error, which is equal to the difference between actual earnings per share and forecast earnings per share, deflated by the stock price at the beginning of the forecast month. If analysts are rational and sufficiently competent, and all information relevant to earnings management incentives is freely available, forecast errors should not be related to discretionary accruals. In such a stylized world, analysts' forecasts would impound without bias expected earnings management, yielding no systematic relationship between actual earnings management and forecast errors. If, however, analysts are irrational, or lack necessary ability, or information regarding earnings management incentives is incomplete, it is plausible that forecast errors will be systematically related to current period earnings management, because earnings management may bias actual earnings away from forecast earnings.
We cannot identify any studies concerning the extent to which analysts fail to anticipate discretionary accruals in current year annual earnings, but some studies have considered this relationship with respect to quarterly earnings reports. Abarbanell and Lehavy (2003) identify two asymmetries in the distribution of analyst forecast errors: the tail asymmetry, describing the greater frequency of large optimistic forecast errors relative to pessimistic forecast errors; and the middle asymmetry, reflecting the greater frequency of small or zero pessimistic forecast errors relative to small optimistic forecast errors. They identify a significant relationship between each of these asymmetries and current quarter earnings management, suggesting that analysts are unable to anticipate firms' earnings management activities in relation to quarterly earnings. However, they do not control for performance, which biases estimated discretionary accruals (Kothari et al., 2005). Liu (2005) provides evidence that analysts anticipate earnings management to some extent in circumstances where capacity or incentives to manage earnings are evident, but does not test the relationship between this earnings management and forecast errors.
If Abarbanell and Lehavy's results are generalizable, both to annual earnings and to the entire distribution of discretionary accruals and forecast errors, then it is likely that analysts will fail to anticipate current discretionary accruals, resulting in biased forecasts. Given the specification of our forecast error calculation, earnings management which is not fully anticipated by analysts will result in an error of the same sign as that of the observed earnings management. Thus, Hypothesis 2 is:
-
H2: There is a positive relationship between forecast errors and current period discretionary accruals.
Thus, more positive (negative) discretionary accruals will inflate (deflate) the current year's earnings relative to cases where there is no earnings management. If analysts fail to predict perfectly earnings management, actual earnings will be greater (smaller) than forecast earnings, leading to positive (negative) forecast errors.
We test this relationship here primarily to enhance the interpretation and understanding of our main research question, developed further in the following section, and which asserts that analysts' ability to anticipate earnings management is contingent on the existence of public signals of its likelihood. Corroboration of H2 would, in and of itself, be unsurprising. Confirmation of H2 may simply imply that contemporary statistical models of unexpected accruals generate estimates that are, in fact, unexpected by sophisticated market participants.
Forecast Accuracy, Earnings Management and the Moderating Effect of Signals of Earnings Management Incentives
There is very little prior literature that considers the moderating role of public signals of earnings management incentives with regard to the relationship between forecast errors and earnings management. Kim and Schroeder (1990) show that the distribution of analyst forecast errors is consistent with the forecasts reflecting information regarding earnings management incentives entailed in managerial bonus plans. Coles et al. (2006) regress short-horizon forecast errors on discretionary accruals for a sample of earnings reports released between cancelation and re-issue of employee stock options and find no significant relationship, suggesting that analyst forecasts efficiently impound information regarding this particular earnings management incentive. The findings of this paper should be treated with caution, however, as neither the proxy for discretionary accruals nor the main regression model used controlled for the influence of performance, which is likely to be correlated with the incidence of option reissues. Nevertheless, there is a sufficient empirical and logical basis to predict that the occurrence of a transparent event which is likely to induce earnings management may make analysts better aware of the likelihood of forthcoming earnings management. Applying this general rationale to the case of CEO appointments, it is plausible that financial analysts may better predict the magnitude and direction of earnings management in the period of CEO appointment, relative to cases of earnings management behaviour where there is no public signal of its likelihood.3 Therefore, with respect to the first earnings report for which the new CEO is responsible, analysts' forecasts may be less sensitive to earnings management than in other cases. Hypothesis 3 is:
-
H3: The sensitivity of forecast errors to current period discretionary accruals is decreased by the occurrence of a CEO appointment in the current period.
To illustrate, assume that there are two firms in an industry (A and B), of which only firm A experiences a CEO change in year t. Firm A and firm B both report earnings for year t, which contain income-decreasing discretionary accruals of similar magnitude. Assuming that analysts recognize the new CEO's likely incentives, the appointment serves as an ex ante signal of earnings management, and analysts may adjust their earnings forecasts accordingly. As no ex ante signals of likely earnings management by firm B are received by the market, earnings forecasts for this firm are less likely to anticipate the realized earnings management. Thus, for a given level of discretionary accruals, forecast errors should be lower for firm A than firm B.
Another branch of the literature which is relevant to the interpretation of the results of tests of H3 concerns whether the existence of analyst forecasts of current earnings encourages income-increasing earnings management intended to cause reported earnings to meet or just beat the consensus forecast. Such earnings management would induce a positive relationship between discretionary accruals and signed forecast errors, as the income-increasing accruals cause the forecast error to be more positive (or less negative). Because a recently appointed CEO is arguably less likely to experience incentives to meet or beat analyst forecasts than the CEOs of other firms, empirical support for H3 could also be explained by this difference in incentives.
Degeorge et al. (1999) argued that the lower than expected number of small negative forecast errors (forecast earnings exceed actual earnings) observed, and the corresponding greater incidence of small positive forecast errors, is attributable to earnings management. Burgstahler and Eames (2006) find evidence that small positive forecast errors are associated with income-increasing earnings management for firms in the upper quartile of discretionary accruals only. Recent studies in the U.K. (Athanaskou et al., 2009) and Australia (Habib and Hossain, 2008), however, find no evidence of a significant difference in earnings management behaviour between small negative and zero or small positive forecast error firms. Given the weakness of the published evidence in support of this purported relationship between earnings management and forecast errors, we do not argue for this potential explanation of our results. However, in our robustness tests we re-estimate our regressions based on a sample restricted to exclude firms whose earnings are exactly equal to, or just exceed, forecast earnings, and which are thus most likely to have been affected by this ‘forecast chasing’ motivation.
DATA AND MODELS
Data and Method for the Estimation of Discretionary Accruals
(1)
(2)
(3)Requisite annual financial statement information was drawn from the Aspect Financial Analysis database for the years 1999–2007. Financial sector firms are excluded due to the different reporting requirements affecting this sector.
The determination of the final samples used in this study is described in Table 1. Of the 9,715 firm-year observations available, 3,509 observations were eliminated from the sample employed to estimate discretionary accruals for the following reasons. Firms in industries with fewer than fifteen firms (625 firms) were eliminated, to allow sufficient degrees of freedom in our discretionary accruals regressions. We also eliminate firms with revenues less than 1% of total assets (1,548 firms), many of which are exploration companies, and insolvent firms (233 firms), as an analysis of a preliminary estimation of discretionary accruals indicated that such firms were likely to be associated with discretionary accrual estimates of unfeasible magnitude, and that their inclusion in the estimation process was likely to bias the estimated discretionary accruals of other firms. We further eliminate 278 observations where total accruals exceed 50% of the value of total assets. Finally, after estimating discretionary accruals, observations for which the Cook's distance statistic exceeded 3 were eliminated and the regressions re-estimated. Discretionary accruals were thus estimated for 6,206 firm-years. The final sample employed in our main tests of H1 employs only 5,095 of these observations, due to missing data relating to control (245 cases) or managerial appointment variables (212 cases), or because the firms experience extreme performance (654 cases), which we define as cases where either the level of, or changes in, ROA or the ratio of cash from operations to assets exceeds 50% of the value of total assets. These cases of extreme performance were eliminated from our reported results because prior studies have shown that conventional discretionary accruals models may be unreliable in such cases.
| Panel A: Determination of discretionary accruals sample | ||
|---|---|---|
| Firms eliminated | Cum. total | |
| Total number of firm-years for which current year financial data was available | 9,715 | |
| Industry population <15 firms | 625 | 9,090 |
| Revenue/Total assets <1% | 1,548 | 7,542 |
| Technically insolvent firms | 223 | 7,319 |
| Lagged data unavailable | 733 | 6,586 |
| ABS[Total Accruals] >50% of total assets | 278 | 6,308 |
| Cook's distance >3 | 102 | 6,206 |
| Discretionary accruals estimated | 6,206 | |
| Eliminated from main tests after estimation of discretionary accruals | ||
| Unavailable control variables (GROWTH) | 245 | 5,961 |
| Unavailable managerial change data | 212 | 5,749 |
| Levels of, or changes in, ROA or CFO/TA exceeded 50% of the value of opening total assets | 654 | 5,095 |
| Total firm-years included in final sample used to test H1 | 5,095 | |
| Panel B: Determination of forecast error sample | ||
|---|---|---|
| Total number of firm-year observations (1998–2007) for which DAs and final forecast errors are available | 2,233 | |
| Eliminate | ||
| Final Forecasts Issued Prior to end of financial year | 201 | 2,032 |
| Observations where forecast error exceeds 30% of share price | 13 | 2,019 |
Earnings Forecast Data
Analysts' consensus earnings per share (EPS) forecasts and firms' share prices (1998–2007) were obtained from the Institutional Brokers Estimate Service (IBES).5 Following prior research (Ali et al., 1992; Bradshaw et al., 2001), we employ the median consensus forecasts, as this measure is less sensitive to outliers than mean forecasts. To maintain data consistency, both the forecast earnings and actual earnings used to calculate forecast errors were collected from IBES.6
Following the relevant literature (Abarbanell and Lehavy, 2003; Coles et al., 2006), we focus on the accuracy of the final forecasts made for a given year's earnings. These forecasts (commonly referred to as short-horizon forecasts) are used in tests of H2 and H3, which focus on analysts' ability to anticipate forthcoming earnings management. Short-horizon forecasts are typically issued between fifteen and 30 days prior to the release of earnings data, and best reflect the totality of information relevant to the prediction of year t earnings, limiting the potential for measurement noise. Additionally, by employing short-horizon forecasts, we ensure that all CEO appointments occurring within a reporting year were known publicly prior to the issue of the forecasts whose accuracy we test. Our focus on final forecasts also has the advantage of restricting the variation in analyst forecast horizon, the effect of which is very difficult to control robustly when forecast horizons differ by several months. The typical timing of the ‘final’ forecasts, and their relationship to the timing of CEO appointments, assuming a 30 June balance date, is illustrated in Figure 1.
ILLUSTRATION OF FORECAST TIMING
(4)Of the 5,095 firm-years employed in our discretionary accruals tests, complete final forecast data were available for 2,233 cases. We eliminate cases where the final available forecast for a firm occurs prior to balance date, as these forecasts are stale, and may fail to capture information from CEO appointments that occur subsequent to the forecast date. To limit the effect of outliers, observations where forecast errors exceed 30% of the firm's share price are eliminated from the sample.
CEO Change Data
Table 2 describes the determination of the final CEO change variables to be examined in this study. CEO changes between 1999 and 2004 were identified by scanning the Australian Stock Exchange (ASX) online announcements for news of CEO appointments. CEO appointments between 2005 and 2007 were identified from Connect 4 Boardroom Review's ‘Position Changes’ file.
| CEO appointment data | |
| Observations for which CEO appointment data is available | 10,268 |
| CEO appointments | 1,476 |
| Proportion of total | 14.37% |
| Observations in discretionary accrual sample | 5,095 |
| CEO appointments | 637 |
| Proportion of total | 12.50% |
| Observations in forecast errors sample | 2,014 |
| CEO appointments | 249 |
| Proportion of total | 12.36% |
For the final sample employed to test the relationship between discretionary accruals and CEO appointments (5,095 firm-year observations), 637 CEO appointments are identified. When the sample is restricted further, to include observations for which forecast errors are available (2,014 firm-years), 249 CEO appointments are identified.
Specification of Regression Models
This section specifies the regression models used to test our hypotheses. To estimate the association between earnings management and current period CEO appointments (H1), this study adopts the approach employed by Pourciau (1993) and Geiger and North (2006), in which discretionary accruals are regressed on a dummy variable indicating the appointment of a new CEO and control variables. Two alternative models are employed, differing only in regard to the control employed for growth.
((Models 1a/b))Hypothesis 1 predicts income-decreasing earnings management following CEO appointments. A number of control variables are included in the regressions. One-year sales growth (GROWTH) has been identified as bearing a positive relationship with measured discretionary accruals (Menon and Williams, 2004; Geiger and North, 2006). As earnings growth is endogenous to firms' earnings management behaviour, we test an alternative specification of the model, in which GROWTH is replaced by the one-year change in the ratio of cash from operating activities to total assets (ChgCFO), a measure which is theoretically independent of accrual-based earnings management. We include the natural log of market value of equity (SIZE), because firm size has been found to be significantly related to discretionary accruals (Frankel et al., 2002; Davidson et al., 2005). Following Geiger and North (2006), financial distress as measured by Zmijiewski's Z Score is included to control for the likelihood that Jones-type models may overestimate discretionary accruals for firms under extreme financial stress.7 To control for the impact of long-term growth opportunities on measured discretionary accruals, firms' book to market ratios (BM) are included. This ratio has previously been found to bear a negative relation to discretionary accruals (Ashbaugh et al., 2003; Geiger and North, 2006). Lagged ROA is included to control for the fact that the performance-adjusted modified Jones model may model the impact of firm performance imperfectly. Although the models of normal and discretionary accruals employed include the effects of lagged firm performance (LAGROA), this variable is also included in the second-stage regressions to capture non-linearities in the relationship between discretionary accruals and performance. We include a dummy variable indicating firm-years in which a financial loss is reported (LOSS) to control for possible differences in earnings management incentives or, in the estimation of discretionary accruals, for loss firms. Finally, year dummies are included to control for changes in the mean level of discretionary accruals across time.
((Model 2))Hypothesis 2 predicts that unanticipated income-increasing (income-decreasing) discretionary accruals inflate (deflate) earnings, causing earnings to be greater (less) than forecast, implying a positive (negative) forecast error. Thus the predicted sign of DA is positive.
Most of the control variables included in Model 2 are identical to those employed in Model 1b. ZSCORE, which is highly positively correlated with leverage, increases the sensitivity of profits to operating risk (Cormier and Martinez, 2006), and thus is controlled here. There is evidence that analyst forecasts for small firms are more optimistic than those for large firms (Coles et al., 2006; Gu and Wu, 2003), and consequently we control for SIZE in our regressions. Firms' book-to-market (BM) ratio which is inversely correlated with a firm's future growth options is included because signed forecast errors increase with a firm's growth options. This implies a negative relationship between BM and forecast errors. The natural log of the time elapsing between the IBES forecasting date and the earnings release date (TIME) is also included because forecast accuracy decreases as the forecast horizon expands, although as our forecast errors are signed this variable will only be significant if early (late) forecasts are biased up or down. Gu and Wu (2003) report that systematic forecast optimism is likely to be associated with firms that report losses, hence a dummy variable indicating actual earnings losses (LOSS) is included to control for this effect. A negative relationship between this variable and forecast errors is expected. LAGROA is included to control for possible over or under-reaction to prior performance (Debondt and Thaler, 1990; Abarbanell and Bernard, 1992). Current year cash from operating activities (CFO) is included to control for performance-level effects not captured by LOSS. The one-year change in CFO is included to control for the likelihood that the magnitude of the change in firm performance is likely to be negatively associated with forecast errors.
((Model 3))If analysts are more efficient in anticipating, or incorporating the impact of, discretionary accruals that are associated with a firm's CEO change, the sensitivity of forecast errors to either current or past discretionary accruals should be decreased by the occurrence of such an event. For H3 to be supported, the coefficient for the interaction term (CEOAppt*DA) should be negative and of the opposite sign to that of the main effect discretionary accruals term (DA). As forecast errors are expected to be positively associated with discretionary accruals (H2), a positive coefficient sign is predicted for DA in Model 3. Consequently, a negative coefficient is expected for CEOAppt*DA. We make no prediction in regard to the sign of the CEOAppt main effect variable.
In addition to testing for the significance of the interaction term (CEOAppt*DA), which tests for a significant difference in the sensitivity of forecast errors to discretionary accruals between CEO appointing firms and others, we also test the significance of the sum of the main effect (DA) and interaction terms (CEOAppt*DA). This tests whether there exists a significant association between discretionary accruals and forecast errors for the sub-sample of CEO change firms, and potentially provides evidence on the semi-strong form efficiency of analyst forecasts with respect to the information contained in the CEO appointment signal. Should the sum of these coefficients be insignificantly different from zero, one cannot reject the null hypothesis that analysts efficiently impound information regarding future accrual behaviour suggested by CEO appointments.
Statistical Method
Diagnostic tests conducted on all models and sub-samples confirmed the presence of firm fixed effects (F-test), and subject specific random effects (Breusch-Pagan Lagrange Multiplier test), suggesting that OLS regression estimators are inconsistent. A Hausmann test, however, rejected the null hypothesis that random effects coefficients were strictly uncorrelated with firm fixed effects, rendering these estimators inconsistent (Greene, 2003). Consequently, this study employs two-way fixed effects regressions in all tests conducted. Fixed effects panel regressions are essentially OLS regressions that include dummy variables representing each subject (firm) that occurs more than once in the sample, and thus test the extent to which within-subject variation in the dependent is explained by the within-subject variation in the regressors.8 The use of fixed effects models is particularly desirable in our research setting, as the firm fixed effect variable potentially reduces the impact of noisy discretionary accrual estimates, if such noise is a function of firm attributes that vary little over time. Time dummy variables (not reported in our output tables) are also included in all models to capture potential differences in the mean level of the dependent variables across time.
Table 3 reports descriptive statistics for variables included in Models 1a and 1b. All continuous variables are winsorized at the 1st and 99th percentiles. Both mean (−0.005) and median (−0.004) discretionary accruals for this sample are negative, with a standard deviation of 0.125. This distribution is comparable to those reported in the U.S. literature (Menon and Williams, 2004; Geiger and North, 2006). There are some significant outliers, which reflect the limitations of discretionary accrual models in estimating earnings management for very small firms and those exhibiting abnormal growth. Predictably, mean LAGROA of firms appointing a new CEO (−0.033) is considerably lower than that of other firms (−0.012). With the exception of GROWTH, none of the control variables exhibits severe breaches of normality. Despite winsorization, GROWTH includes a small number of extreme positive outliers. All tests reported in this paper were replicated on a sample truncated at the GROWTH= 200% level, with no significant impact on results.
| N | M | 1% | 25% | Mdn | 75% | 99% | SD | |
|---|---|---|---|---|---|---|---|---|
| Firm-years without CEO appointment | ||||||||
| DA | 4,458 | −0.003 | −0.341 | −0.060 | −0.003 | 0.052 | 0.399 | 0.125 |
| CEOAppt | 4,458 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| GROWTH | 4,458 | 3.163 | −0.882 | −0.046 | 0.120 | 0.437 | 92.670 | 21.119 |
| ChgCFO | 4,458 | 0.004 | −0.351 | −0.045 | 0.003 | 0.050 | 0.360 | 0.117 |
| SIZE | 4,458 | 18.041 | 14.136 | 16.359 | 17.752 | 19.502 | 23.130 | 2.179 |
| ZSCORE | 4,458 | −2.073 | −4.718 | −3.003 | −2.088 | −1.268 | 1.499 | 1.289 |
| BM | 4,458 | 0.836 | 0.055 | 0.360 | 0.625 | 1.087 | 3.125 | 0.693 |
| LOSS | 4,458 | 0.334 | 0.000 | 0.000 | 0.000 | 1.000 | 1.000 | 0.472 |
| LAGROA | 4,458 | −0.012 | −0.743 | −0.056 | 0.043 | 0.082 | 0.298 | 0.190 |
| CFO | 4,458 | 0.032 | −0.393 | −0.036 | 0.053 | 0.114 | 0.359 | 0.143 |
| Firm-years with CEO appointment | ||||||||
| DA | 637 | −0.017 | −0.338 | −0.075 | −0.019 | 0.050 | 0.316 | 0.125 |
| CEOAppt | 637 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.000 |
| GROWTH | 637 | 2.700 | −0.863 | −0.075 | 0.087 | 0.386 | 96.652 | 18.287 |
| ChgCFO | 637 | 0.001 | −0.380 | −0.056 | 0.000 | 0.053 | 0.413 | 0.129 |
| SIZE | 637 | 18.135 | 14.433 | 16.571 | 17.776 | 19.519 | 23.130 | 2.129 |
| ZSCORE | 637 | −1.806 | −4.439 | −2.706 | −1.890 | −1.085 | 2.231 | 1.365 |
| BM | 637 | 0.846 | 0.041 | 0.372 | 0.641 | 1.111 | 3.125 | 0.668 |
| LOSS | 637 | 0.421 | 0.000 | 0.000 | 0.000 | 1.000 | 1.000 | 0.494 |
| LAGROA | 637 | −0.033 | −0.708 | −0.078 | 0.029 | 0.072 | 0.331 | 0.209 |
| CFO | 637 | 0.009 | −0.405 | −0.066 | 0.033 | 0.094 | 0.363 | 0.151 |
| All firm-years | ||||||||
| DA | 5,095 | −0.005 | −0.341 | −0.061 | −0.004 | 0.052 | 0.391 | 0.125 |
| CEOAppt | 5,095 | 0.125 | 0.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.331 |
| GROWTH | 5,095 | 3.105 | −0.881 | −0.049 | 0.113 | 0.434 | 92.917 | 20.785 |
| ChgCFO | 5,095 | 0.004 | −0.354 | −0.046 | 0.003 | 0.050 | 0.368 | 0.118 |
| SIZE | 5,095 | 18.053 | 14.161 | 16.387 | 17.755 | 19.506 | 23.130 | 2.173 |
| ZSCORE | 5,095 | −2.040 | −4.710 | −2.968 | −2.057 | −1.249 | 1.602 | 1.302 |
| BM | 5,095 | 0.837 | 0.055 | 0.361 | 0.629 | 1.087 | 3.125 | 0.690 |
| LOSS | 5,095 | 0.345 | 0.000 | 0.000 | 0.000 | 1.000 | 1.000 | 0.475 |
| LAGROA | 5,095 | −0.014 | −0.735 | −0.058 | 0.042 | 0.080 | 0.305 | 0.193 |
| wSclCFO | 5,095 | 0.029 | −0.396 | −0.040 | 0.051 | 0.112 | 0.359 | 0.144 |
Variance inflation factors (VIFs) were calculated for all models. With the exception of firm size, no VIFs exceeded 4, suggesting no significant collinearity problems in our models.9
Table 4 provides descriptive statistics for variables that are included in Models 2 and 3, in which forecast errors are the dependent variable. Consistent with prior research, forecast errors are skewed about an approximately zero median, with negative (optimistic) forecast errors being typically larger than positive (pessimistic) forecast errors. There is no significant difference in mean forecast errors for firms with and without CEO appointments. Mean LAGROA is significantly higher (0.057) than in the sample used to test H1, reflecting the fact that analysts are more likely to study larger, more consistently profitable entities. Once more, with the exception of the SIZE variable, no VIFs exceed 4, suggesting that multicolinearity presents no significant challenges to the interpretation of our results.
| N | M | 1% | 25% | Mdn | 75% | 99% | SD | |
|---|---|---|---|---|---|---|---|---|
| Firm-years without CEO appointment | ||||||||
| FE | 1,765 | −0.004 | −0.177 | −0.005 | 0.001 | 0.005 | 0.103 | 0.038 |
| DA | 1,765 | −0.004 | −0.283 | −0.046 | −0.005 | 0.039 | 0.305 | 0.096 |
| CEOAppt*DA | 1,765 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| ZSCORE | 1,765 | −2.049 | −4.794 | −2.756 | −1.996 | −1.331 | 0.557 | 1.074 |
| BM | 1,765 | 0.608 | 0.054 | 0.280 | 0.493 | 0.787 | 2.632 | 0.482 |
| SIZE | 1,765 | 19.718 | 16.482 | 18.552 | 19.590 | 20.828 | 23.130 | 1.620 |
| TIME | 1,765 | 2.567 | 0.000 | 1.946 | 2.639 | 3.219 | 4.787 | 0.791 |
| LOSS | 1,765 | 0.118 | 0.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.323 |
| LAGROA | 1,765 | 0.060 | −0.404 | 0.039 | 0.067 | 0.102 | 0.331 | 0.118 |
| CFO | 1,765 | 0.087 | −0.313 | 0.040 | 0.085 | 0.139 | 0.439 | 0.120 |
| ChgCFO | 1,765 | 0.002 | −0.340 | −0.038 | 0.002 | 0.038 | 0.336 | 0.115 |
| Firm-years with CEO appointment | ||||||||
| FE | 249 | −0.018 | −0.271 | −0.017 | 0.000 | 0.005 | 0.115 | 0.059 |
| DA | 249 | −0.015 | −0.347 | −0.057 | −0.017 | 0.044 | 0.318 | 0.112 |
| CEOAppt*DA | 249 | −0.015 | −0.347 | −0.057 | −0.017 | 0.044 | 0.318 | 0.112 |
| ZSCORE | 249 | −1.970 | −4.794 | −2.615 | −1.890 | −1.315 | 0.999 | 1.072 |
| BM | 249 | 0.740 | 0.030 | 0.340 | 0.588 | 0.917 | 2.857 | 0.596 |
| SIZE | 249 | 19.620 | 15.930 | 18.382 | 19.418 | 20.897 | 23.130 | 1.761 |
| TIME | 249 | 2.615 | 0.000 | 2.398 | 2.639 | 3.091 | 3.738 | 0.713 |
| LOSS | 249 | 0.193 | 0.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.395 |
| LAGROA | 249 | 0.030 | −0.576 | 0.023 | 0.057 | 0.091 | 0.301 | 0.224 |
| CFO | 249 | 0.080 | −0.378 | 0.028 | 0.077 | 0.132 | 0.439 | 0.124 |
| ChgCFO | 249 | 0.022 | −0.331 | −0.027 | 0.008 | 0.054 | 0.562 | 0.137 |
| All firm-years | ||||||||
| FE | 2,014 | −0.005 | −0.204 | −0.005 | 0.000 | 0.005 | 0.103 | 0.042 |
| CEOApp | 2,014 | 0.124 | 0.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.329 |
| DA | 2,014 | −0.006 | −0.290 | −0.048 | −0.005 | 0.040 | 0.305 | 0.099 |
| CEOAppt*DA | 2,014 | −0.002 | −0.166 | 0.000 | 0.000 | 0.000 | 0.108 | 0.040 |
| ZSCORE | 2,014 | −2.040 | −4.794 | −2.750 | −1.977 | −1.330 | 0.574 | 1.074 |
| BM | 2,014 | 0.624 | 0.052 | 0.287 | 0.505 | 0.800 | 2.632 | 0.499 |
| SIZE | 2,014 | 19.706 | 16.482 | 18.525 | 19.584 | 20.845 | 23.130 | 1.638 |
| TIME | 2,014 | 2.573 | 0.000 | 1.946 | 2.639 | 3.219 | 4.700 | 0.782 |
| LOSS | 2,014 | 0.127 | 0.000 | 0.000 | 0.000 | 0.000 | 1.000 | 0.333 |
| LAGROA | 2,014 | 0.057 | −0.406 | 0.037 | 0.066 | 0.100 | 0.331 | 0.136 |
| CFO | 2,014 | 0.086 | −0.313 | 0.038 | 0.084 | 0.138 | 0.439 | 0.121 |
| ChgCFO | 2,014 | 0.005 | −0.334 | −0.037 | 0.003 | 0.040 | 0.354 | 0.118 |
RESULTS
Tests of the Association Between Discretionary Accruals and CEO Appointment (H1)
The results of tests of Hypothesis 1, which predicts income-decreasing discretionary accruals in the year of a CEO appointment, are provided in Table 5. Both variants of the regression model exhibit R-squared statistics of approximately 32%. Consistent with H1 and the findings of much of the prior literature, in each model CEO appointments are significantly negatively associated with discretionary accruals at the 95% confidence level (p= 0.0327, p= 0.0329). The estimated relationships between discretionary accruals and CEO appointments also appear to have economic significance, with the beta coefficients indicating an average impact on discretionary accruals equal to 0.91% of total assets. Applying the cross-sample median relationships between revenues and total assets, and absolute earnings and total assets, the decrease in signed discretionary accruals in cases of CEO appointments equates to approximately 1.25% of revenues and 11.5% of the absolute value of earnings. Neither of the growth proxies, nor SIZE, is significantly associated with discretionary accruals, but the remainder of the control variables are significant in the predicted direction.
| Dependent variable: Discretionary accruals | (1) | (2) | |
|---|---|---|---|
| Independent variables | Pred. sign | Model 1a | Model 1b |
| CEOAppt | − | −0.0091* | −0.0091* |
| (0.0327) | (0.0329) | ||
| GROWTH | + | 0.0001 | |
| (0.1373) | |||
| ChgCFO | + | −0.0113 | |
| (0.6248) | |||
| SIZE | − | −0.0009 | −0.0008 |
| (0.4202) | (0.4242) | ||
| ZSCORE | − | −0.0454*** | −0.0454*** |
| (0.0000) | (0.0000) | ||
| BM | − | −0.0091* | −0.0091* |
| (0.0395) | (0.0392) | ||
| LOSS | − | −0.0899*** | −0.0895*** |
| (0.0000) | (0.0000) | ||
| LAGROA | ? | −0.0918*** | −0.0964*** |
| (0.0000) | (0.0000) | ||
| CFO | − | −0.6906*** | −0.6796*** |
| (0.0000) | (0.0000) | ||
| Constant | −0.0284 | −0.0303 | |
| (0.7106) | (0.6905) | ||
| Observations | 5,095 | 5,095 | |
| R 2 | 0.3227 | 0.3224 | |
- *** p < 0.01, ** p < 0.05, * p < 0.1. p-values are one-sided if estimated coefficient is of predicted sign.
- CEOAppt= 1 if a new CEO appointed in current financial year, 0 otherwise; GROWTH= rate of revenue growth proportional previous year revenue; ChgCFO= current year cash from operating activities / average total assets – previous year cash from operating activities / previous year average total assets; SIZE= natural log of opening market value of equity; ZSCORE= Zmijiewski's Z Score; BM= ratio of opening book value per share to market value per share; LOSS= 1 if current year earnings are negative, 0 otherwise; LAGROA= previous year return on assets; CFO= ratio of cash from operating activities to average total assets.
Tests of the Relationship Between Discretionary Accruals and Forecast Errors
Table 6 reports the results of tests of Hypothesis 2, which predicts that final forecast errors are positively associated with discretionary accruals detected in actual earnings. In addition to the main test, conducted on the full sample as described previously, the regression is also estimated on samples restricted according to firms covered by at least three analysts (column 2) and to forecast errors of absolute value less than 10% (column 3).
| Dependent variable: Forecast error | (1) | (2) | (3) | |
|---|---|---|---|---|
| Independent variables | Pred. sign | All firms | Coverage ≥ 3 analysts | |FE| < 10% |
| DA | + | 0.0726*** | 0.0800*** | 0.0389*** |
| (0.0004) | (0.0009) | (0.0003) | ||
| ZSCORE | ? | −0.0024 | −0.0023 | 0.0002 |
| (0.2531) | (0.3451) | (0.8428) | ||
| BM | − | −0.0208*** | −0.0229*** | −0.0037 |
| (0.0001) | (0.0004) | (0.1095) | ||
| SIZE | − | −0.0008 | −0.0022 | 0.0007 |
| (0.3698) | (0.2461) | (0.3126) | ||
| TIME | ? | −0.0035** | −0.0041*** | −0.0026*** |
| (0.0102) | (0.0027) | (0.0005) | ||
| LOSS | − | −0.0156** | 0.0005 | −0.0023 |
| (0.0238) | (0.9701) | (0.2743) | ||
| LAGROA | − | −0.0705*** | −0.0804* | −0.0157 |
| (0.0044) | (0.0286) | (0.1341) | ||
| CFO | + | 0.0744*** | 0.0402 | 0.0258* |
| (0.0054) | (0.1664) | (0.0805) | ||
| ChgCFO | − | −0.0131 | −0.0017 | 0.0038 |
| (0.2408) | (0.4662) | (0.3658) | ||
| Constant | 0.0347 | 0.0667 | −0.0036 | |
| (0.4624) | (0.3459) | (0.8969) | ||
| Observations | 2,014 | 1,264 | 1,928 | |
| R 2 | 0.1180 | 0.1343 | 0.0625 | |
- *** p < 0.01, ** p < 0.05, * p < 0.1. Robust p values in parentheses, p-values are one-sided if estimated coefficient is of predicted sign, otherwise two-sided. DA= discretionary accruals estimated from performance-adjusted modified Jones model; ZSCORE= Zmijiewski's Z Score; BM= ratio of opening book value per share to market value per share; SIZE= natural log of opening market value of equity; TIME= natural log of the number of days between forecast date and reporting date; LOSS= 1 if current year earnings are negative, 0 otherwise; LAGROA= previous year return on assets; CFO= ratio of cash from operating activities to average total assets; ChgCFO= current year cash from operating activities / average total assets – previous year cash from operating activities / previous year average total assets.
The coefficient for discretionary accruals is significantly positive at an alpha confidence level of 1% in the main test (B= 0.072, p= 0.0004) and both additional tests. Given that forecast errors are deflated by end-of-year stock price (market value), discretionary accruals are deflated by beginning-of-year total assets (book value), and the median ratio of these deflators is approximately 2.36, this coefficient implies that analysts fail to anticipate 17.1 cents (0.072 × 2.36) in each dollar of discretionary accruals recorded by a representative firm. This appears to represent strong evidence that analysts are unable to anticipate perfectly firms' earnings management behaviour as measured by discretionary accruals. This is not a surprising result, and provides support for the contention that the measures commonly described by researchers as ‘unexpected accruals’ are to some extent ‘unexpected’ by sophisticated market participants. Of course, our results also suggest that a significant proportion of our estimated discretionary accruals are impounded in analysts' forecasts, either because analysts are able to predict firms' discretionary reporting behaviour, or because contemporary discretionary accrual measures are imperfect and are correlated with non-discretionary reporting behaviour. The tests reported in the following section provide evidence on the extent to which analysts' forecasts reflect predicted discretionary accounting behaviour.
Tests of the Moderating Effect of CEO Changes on the Relationship Between Forecast Errors and Discretionary Accruals
The central research question considered in the paper concerns the extent to which publicly observable signals of likely earnings management moderate the relationship between this phenomenon and analyst forecast errors. Results for tests of Hypothesis 3, which predicts that forecast errors are less sensitive to earnings management in cases where a new CEO has been appointed than otherwise, are reported in Table 7. A series of additional tests (discussed below) are also reported.
| Dependent variable: Forecast error | (1) | (2) | (3) | (4) | (5) | (6) | |
|---|---|---|---|---|---|---|---|
| Variables | Pred. sign | Sample employed | |||||
| All firms | Coverage ≥ 3 analysts | |FE| < 10% | Excl small positive FEs (FE < 0.0002) | Firms with no ABNMLS | Alternative measure of DA | ||
| CEOAppt | ? | −0.0056 | −0.0070** | −0.0021 | −0.0055 | −0.0021 | −0.0060 |
| (0.1217) | (0.0331) | (0.2087) | (0.1729) | (0.4560) | (0.1055) | ||
| DA | + | 0.0808*** | 0.0897*** | 0.0442*** | 0.0944*** | 0.0483*** | 0.0421** |
| (0.0002) | (0.0008) | (0.0001) | (0.0001) | (0.0017) | (0.0138) | ||
| CEOAppt*DA | − | −0.0625** | −0.0667** | −0.0377** | −0.0803** | −0.0598* | −0.0842*** |
| (0.0307) | (0.0426) | (0.0278) | (0.0167) | (0.0358) | (0.0245) | ||
| ZSCORE | ? | −0.0020 | −0.0017 | 0.0004 | −0.0033 | −0.0017 | −0.0035* |
| (0.3452) | (0.4793) | (0.7478) | (0.2078) | (0.3791) | (0.0916) | ||
| BM | − | −0.0196*** | −0.0218*** | −0.0031 | −0.0161*** | −0.0094** | −0.0200*** |
| (0.0001) | (0.0055) | (0.1510) | (0.0033) | (0.0355) | (0.0001) | ||
| SIZE | − | −0.0001 | −0.0018 | 0.0010 | 0.0019 | −0.0031* | 0.0003 |
| (0.4760) | (0.2918) | (0.4622) | (0.4946) | (0.0861) | (0.9206) | ||
| TIME | ? | −0.0034** | −0.0041*** | −0.0026*** | −0.0039** | −0.0025* | −0.0035** |
| (0.0150) | (0.0030) | (0.0006) | (0.0193) | (0.0879) | (0.0149) | ||
| LOSS | − | −0.0158** | −0.0015 | −0.0023 | −0.0158* | 0.0138 | −0.0145* |
| (0.0380) | (0.9068) | (0.5313) | (0.0716) | (0.5370) | (0.0814) | ||
| LAGROA | − | −0.0753*** | −0.0773** | −0.0191* | −0.0775*** | −0.0796** | −0.0957*** |
| (0.0021) | (0.0336) | (0.0897) | (0.0009) | (0.0238) | (0.0007) | ||
| CFO | + | 0.0794*** | 0.0424 | 0.0283* | 0.1034*** | 0.0262 | 0.0744*** |
| (0.0029) | (0.1528) | (0.0569) | (0.0011) | (0.1515) | (0.0038) | ||
| ChgCFO | − | −0.0154 | −0.0019 | 0.0030 | −0.0247 | 0.0070 | −0.0335* |
| (0.1999) | (0.4609) | (0.3915) | (0.2090) | (0.3471) | (0.0284) | ||
| Constant | 0.0223 | 0.0581 | −0.0103 | −0.0287 | 0.0783 | 0.0128 | |
| (0.6371) | (0.4136) | (0.7118) | (0.6278) | (0.1090) | (0.7982) | ||
| Observations | 2,014 | 1,264 | 1,928 | 1,766 | 935 | 1,909 | |
| R 2 | 0.1224 | 0.1408 | 0.0670 | 0.1281 | 0.1602 | 0.1037 | |
| F test: DA+CEOAppt*DA= 0 | 0.3109 | 0.4200 | 0.1172 | 0.1472 | 0.1635 | 2.0911 | |
| p-value: DA+CEOAppt*DA= 0 | 0.577 | 0.517 | 0.732 | 0.701 | 0.686 | 0.149 | |
- DA= discretionary accruals estimated from performance-adjusted modified Jones model; ZSCORE= Zmijiewski's Z Score; BM= ratio of opening book value per share to market value per share; SIZE= natural log of opening market value of equity; TIME= natural log of the number of days between forecast date and reporting date; LOSS= 1 if current year earnings are negative, 0 otherwise; LAGROA= previous year return on assets; CFO= ratio of cash from operating activities to average total assets; ChgCFO= current year cash from operating activities / average total assets – previous year cash from operating activities / previous year average total assets.
The interaction variable (CEOAppt*DA) is the focus of these tests. In the main test (column 1), the predicted positive relationship between forecast errors and discretionary accruals for firms that did not appoint a new CEO (DA) is confirmed (B2= 0.0808, p= 0.0002). After using the median ratio of end-of-year market value to beginning-of-year book value (2.36 to 1) to adjust for the difference in the deflators implicit in our forecast error and discretionary accrual measures, the coefficient for DA implies that a $1.00 increase in discretionary accruals is associated with a $0.192 increase in forecast errors for firms that did not appoint a CEO during the period. The negative coefficient (B3=−0.0625, p= 0.0307) for the interaction term indicates a significantly lower sensitivity of forecast errors to current period earnings management in cases where a CEO change has occurred during the financial year. This evidence is consistent with analysts using their knowledge of the earnings management implications of CEO turnover to revise their expectations and to improve the accuracy of their forecasts. The results reported in column 1 suggest that the occurrence of CEO appointment reduces the sensitivity of forecast errors to discretionary accruals by approximately 77%.
In addition to the significant reduction in the sensitivity of forecast errors to discretionary accruals when CEO appointments occur, we test the sum of the coefficients for the main effect (DA) and the interaction term (CEOAppt*DA) to determine whether there is any significant relationship between earnings management and discretionary accruals for these CEO change firms. The sum of the main effect and interaction terms (B2+B3= 0.018, p= 0.577) is statistically indistinguishable from zero. Thus, we show that analysts are not only able to use signals of earnings management incentives to reduce forecast errors associated with earnings management, but that where CEO appointments occur the improvement in forecast accuracy is so great that forecast errors bear no significant relation to our estimate of discretionary accruals.
Sensitivity Analysis
Columns 2 through 6 of Table 7 report the results of various robustness tests. In column 2, the model is re-estimated based on a sample of firm-years for which at least three analysts' forecasts were used to calculate the consensus forecast. In column 3, the sample is restricted to cases where the absolute value of the forecast error is less than 10%. In each of these cases the main test results are confirmed.
An alternative explanation for our main test results is that the existence of analyst forecasts motivates firms whose ‘unmanaged’ earnings are below the consensus forecast to engage in income-increasing earnings management to meet or just beat this benchmark. If firms that did not appoint a new CEO in a given year experience an incentive to manage earnings upwards to meet consensus forecasts, and firms with new CEOs do not experience this incentive (as the new CEO is unlikely to be blamed for missing this benchmark), then this may at least partially explain why forecast errors are less sensitive to earnings management when practised by firms that have recently appointed a new CEO.10 To investigate this possibility, we re-estimate Model 3 using a sample which excludes firms with forecast errors between 0 and 0.0002% inclusive, which we interpret as representing firms that meet or just beat consensus forecasts. Our key result remains unaffected by this sample restriction.11
In columns 5 and 6 we investigate whether our results could be explained by the fact that analysts' earnings forecasts and IBES actual earnings are calculated on a continuing operations basis, while our discretionary accruals measure is based on GAAP profits. If this articulation difference has any effect, it should logically be concentrated in cases where significant non-recurring items affect GAAP earnings. In column 5 we re-estimate Model 3 based on a sample which excludes all firm-years in which the Aspect-Huntley database identifies abnormal items within profit.12 While the magnitude of the coefficients relating to DA and CEOAppt*DA are smaller, their significance remains similar to the main test results. In column 6, we use an alternative measure of discretionary accruals in our regression. In estimating discretionary accruals, we define the dependent variable, total accruals, as the difference between net profit before abnormal items and cash from operations. In doing so, large one-off write-downs such as restructuring charges are excluded from the total accruals that the performance-adjusted modified Jones model attempts to explain.13 Such write-downs are more likely in the case of CEO turnover, but arguably do not constitute opaque earnings management. When this alternative measure of discretionary accruals is employed our results remain substantively similar to our main results.
To investigate whether our results are sensitive to the choice of discretionary accrual model employed, we repeated our main tests of Hypotheses 1 through 3 using two well-known alternative models of accrual behaviour: the Dechow–Dichev (2002) model (hereafter DD), and the McNicholls (2002) modification (MM) to the DD model. The DD model is summarized in equation (5).
(5)
(6)The distributions of the DD and MM discretionary accruals (not tabulated) were similar to those of the performance-adjusted modified Jones discretionary accruals, with mean values of 0.0008 (DD) and 0.0003 (MM) and standard deviations of 0.1241 (DD) and 0.1136 (MM). The results of tests of Hypotheses 1 through 3 using the DD and MM discretionary accruals estimates are reported in Table 8. For the sake of brevity, coefficients for control variables are not tabulated, but are similar to those in our main tests. Tests of H1, which posits a negative association between discretionary accruals and current period CEO appointments, generate significant negative coefficients with p-values less than 0.05 for the relevant variable (CEOAppt) in both the DD model (column 1) and the MM model (column 4). Thus our main results are supported. Both the DD model (column 2) and McNicholls model (column 5) regressions report strong positive associations (p-values ≤ 0.001) between signed forecast errors and current period discretionary accruals, providing further support for H2. Finally, both alternative models (columns 3 and 6) report significant negative coefficients for the interaction between CEO appointments and discretionary accruals, supporting H3 which contends that the sensitivity of forecast errors to discretionary accruals is lower in cases where a CEO is newly appointed in the current year. While the coefficient for CEOAppt*DA is only barely significant under the DD model, it must be remembered that this proxy for discretionary accruals ignores any manipulation of long-term accruals. Further, in both the DD and MM models the sum of the main discretionary accrual effect (DA) and the interaction term (CEOAppt*DA) is not significantly different from zero, consistent with the proposition that analyst forecasts are efficient with respect to discretionary accruals where CEO appointments occur in the current year.
| Variables and predicted signs for H1/H2/H3 | Dechow-Dichev (2002) discretionary accruals | McNicholls (2002) discretionary accruals | ||||
|---|---|---|---|---|---|---|
| (1) | (2) | (3) | (4) | (5) | (6) | |
| H1 | H2 | H3 | H1 | H2 | H3 | |
| Dependent variable: DA | Dependent variable: FE | Dependent variable: FE | Dependent variable: DA | Dependent variable: FE | Dependent variable: FE | |
| CEOAppt (−/NA/?) | −0.0092** | −0.0051 | −0.0072* | −0.0040 | ||
| (0.0237) | (0.1012) | (0.0493) | (0.1525) | |||
| DA (NA/+/+) | 0.0683*** | 0.0713*** | 0.0692*** | 0.0808*** | ||
| (0.0000) | (0.0000) | (0.0004) | (0.0005) | |||
| CEOAppt*DA (NA/NA/−) | −0.0434 | −0.0847* | ||||
| (0.0908) | (0.0496) | |||||
| N | 4,936 | 1,749 | 1,690 | 5,048 | 1,820 | 1,711 |
| R 2 | 0.260 | 0.100 | 0.106 | 0.231 | 0.099 | 0.109 |
| p-value from F-test of DA+CEO*DA= 0 | 0.3931 | 0.9322 | ||||
- *** p < 0.01, ** p < 0.05, * p < 0.1. Coefficients for control variables defined in equations (5) and (6) not reported.
As our descriptive statistics indicated the presence of significant outliers in the GROWTH variable, we conducted a series of untabulated tests to examine the possible impact of abnormal revenue growth on the estimation of discretionary accruals and thus on each of our main regression models. To this end we replicated our regressions, excluding from the samples firms whose revenue growth exceeded the mean industry-year revenue growth by more than one standard deviation. In each of our main tests of Hypotheses 1 through 3, and in our additional tests of these hypotheses using DD and MM discretionary accrual methods, our results remained substantively unaffected.
Finally, to test the sensitivity of our results to outliers in estimation of discretionary accruals, we re-estimated our main tests based on samples restricted to discretionary accruals of absolute value less than 10% of opening total assets. Our key results held in these tests and also when we varied the outlier threshold to 5%, 15% and 20% of total assets.
DISCUSSION AND CONCLUSION
Discretionary Accruals Following CEO Changes
The above section presented results generally consistent with the proposition that incoming CEOs take an ‘earnings bath’ in their initial year of incumbency, in an attempt to signal their predecessor's liability for the firm's poor performance and to lower the benchmark on which the incoming CEO's future compensation will be based (H1). This is consistent with prior findings, such as Moore (1973), Elliot and Shaw (1988), Pourciau (1993), Wells (2002) and Godfrey et al. (2003).
Forecast Efficiency and Discretionary Accruals
While prior studies have examined the extent to which analysts are able to anticipate earnings management present in quarterly earnings, and the extent to which firms manage earnings in an attempt to meet or beat consensus earnings forecasts, this study is the first to examine analysts' ability to anticipate discretionary accruals in annual earnings. Findings are consistent with H2, which predicts that on average, analysts are unable to predict perfectly the magnitude of as yet unreported earnings management. Forecast errors are positively associated with discretionary accruals realized in the current period, suggesting that more positive discretionary accruals induce more positive forecast errors (larger ‘good news’ earnings surprises) and more negative discretionary accruals lead to more negative forecast errors (larger ‘bad news’ earnings surprises). This result is consistent with Abarbanell and Lehavy's (2003) findings with respect to earnings management in U.S. quarterly reports. Whether this systematic bias represents market inefficiency, however, depends on the assumptions one makes about the availability of information from which analysts should be expected to infer likely earnings management behaviour. Beyond this question of market efficiency, the implications of our results are reasonably clear. Measured discretionary accruals are, to some extent at least, genuinely ‘unexpected accruals’. The likelihood that analysts will not perfectly anticipate future earnings management renders this behaviour an effective (if socially undesirable or at times fraudulent) tool with which to attempt to meet or beat analyst forecasts.
The Moderating Effect of CEO Changes on the Relationship Between Forecast Efficiency and Discretionary Accruals
The main focus of this study is to test whether the occurrence of a CEO appointment improves analysts' abilities to anticipate earnings management affecting the accuracy of short-horizon forecasts. The incentives for a new CEO to manage earnings are widely recognized both in the literature and in the market, thus it is reasonable to expect sophisticated analysts to be aware of likely earnings management surrounding CEO appointments and adjust their forecasts accordingly.
We report strong support for H3, which predicts a significant moderating effect of CEO changes on the relation between discretionary accruals and forecast errors in the year of CEO change. In our main tests and numerous robustness tests, forecast errors are less sensitive to discretionary accruals where a new CEO has been appointed during the financial year. Further, we report evidence that there is no significant relationship between forecast errors and discretionary accruals for CEO appointing firms. This evidence is consistent with the conjecture that analysts are able to impound relatively clear earnings management signals into their forecasts.
Limitations
Our study is subject to a number of limitations. First, while we employ a method common in the literature, it is widely accepted that discretionary accrual models are imperfect. While contamination of the discretionary accrual measure by non-discretionary items is almost inevitable, we have taken considerable steps in the estimation of discretionary accruals to reduce the effect of significant outliers and to report robustness tests employing alternative discretionary accrual measures.
A further limitation is that this paper does not attempt to differentiate between the routine and non-routine CEO changes, which would necessarily introduce additional subjectivity to the measurement process. Some previous studies, however, show that non-routine CEO changes have a stronger relationship with earnings management relative to routine CEO changes in the year of the changes (Pourciau, 1993; Wells, 2002). In a similar light, Wilson and Wang (2010) demonstrate that concurrent changes in chairperson and CEO are associated with significantly greater earnings management behaviour than cases where the CEO change occurs alone. A more detailed classification of CEO appointments would, however, reduce the number of events per dichotomous variable to a level where a small number of observations could exert potentially great leverage on our results.
While the sample used to test our main hypothesis includes analyst forecasts for 2,014 firm-years, the sub-sample of CEO appointing firms is only 249. This represents quite a small proportion of all analyst forecasts, and so caution should be exercised in generalizing from these results. Nevertheless, this sub-sample is of similar size to that employed in the only directly comparable study (Coles et al., 2006).
Our results suggest a number of areas for future research. First, the study of analyst forecast revisions immediately following either the appointment of a new CEO, or the end of an outgoing CEO's tenure, may provide further evidence into the extent to which analysts utilize information from such events. Further, while we have selected CEO changes as an example of a public signal of likely earnings management, analysts' efficiency with respect to other signals, such as regulatory events and the disclosed terms of executive option schemes, may provide fruitful avenues for future research.
