The Role of Financial Analysts in Stock Market Efficiency with Respect to Annual Earnings and its Cash and Accrual Components

Transcription

1 The Role of Financial Analysts in Stock Market Efficiency with Respect to Annual Earnings and its Cash and Accrual Components Dana Hollie a, Phil Shane b, Qiuhong Zhao c a Louisiana State University b University of Virginia and University of Auckland c University of Missouri at Columbia August 2012 Abstract This paper examines biases in stock prices and financial analysts earnings forecasts with respect to persistence characteristics of net operating income and its free cash flow and accrual components. These biases take the form of systematic over-reaction or underreaction. We find that stock prices have a general tendency to over-react to the persistence of net operating income and its components; whereas, financial analysts have a tendency to under-react to the same information. Analysts forecasting bias mitigates what would otherwise appear as more pronounced stock price over-reaction. On the other hand, we find little evidence that the bias in stock prices mitigates analyst under-reaction. Overall, we find that, left to their own devices, analysts earnings forecasts under-react to the persistence characteristics of annual earnings, and stock prices would over-react, but biased analysts forecasts largely mitigate this over-reaction tendency. This paper brings a new perspective to the literature regarding the disciplining role of financial analysts in capital markets. JEL classification: G14, M41 Keywords: Analyst Forecast, Accruals, Cash Flows, Market Efficiency We thank I/B/E/S for providing the analysts forecast data for this study. For their helpful comments, we would also like to thank Sudipta Basu, Katherine Gunny, Steve Rock and workshop participants at Bond University, Colorado State University, Louisiana State University, Monash University, the Universities of Auckland, Colorado, and Queensland, the 2012 AAA Mid-Atlantic conference, and the 2008 Annual AAA conference. * addresses: danahollie@lsu.edu, Phil.Shane@Virginia.edu, and zhaoq@missouri.edu. 0

2 The Role of Financial Analysts in Stock Market Efficiency with Respect to Annual Earnings and its Cash and Accrual Components I. Introduction This paper investigates the role of financial analysts in stock market efficiency with respect to earnings information derived from corporate annual reports to shareholders. 1 Sloan (1996) finds that the market over-reacts (under-reacts) to the persistence characteristics of accrual (cash flow) components of earnings, where persistence is estimated with reference to the coefficient relating each current year earnings component to the whole of next year s net operating income. 2 Mashruwala, Rajgopal and Shevlin (2006) report that the cash flow and accrual anomalies are concentrated in stocks with low trading volume and other characteristics associated with high arbitrage costs. Since stocks covered by financial analysts tend to have relatively high trading volume [Alford and Berger (1999); Frankel, Kothari and Weber (2006)], we consider it unlikely that financial analyst activity drives the anomalous stock price behavior discovered by Sloan. Our paper focuses on whether financial analysts play a role in mitigating the cash flow and accrual anomalies in the stocks they follow. Prior research provides evidence that stock prices respond, at least partially, to analysts earnings forecast revisions [Gleason and Lee (2003)], and analysts respond, at least partially, to information in stock returns [Clement, Hales and Xue (2011)]. As a result, returns on equity securities exhibit a strong contemporaneous relation with both 1 See Schipper (1991), Brown (1993) and Ramnath et al. (2008a and 2008b) for extensive reviews of the literature investigating the role of financial analysts in capital markets. 2 See Collins and Hribar (2000) for a similar result in the context of quarterly accrual and cash earnings. 1

3 analysts earnings forecast errors and analysts earnings forecast revisions [Liu and Thomas (2000)]. Our study is the first to investigate the bias (if any) in financial analysts' response to firms annual net operating income and its detailed accrual and cash flow components. We focus on how, without the influence of the bias in stock prices (analysts forecasts), analysts forecasts (stock prices) would respond to the free cash flow and accrual earnings components derived from corporate annual reports. Our disaggregation of annual net operating income into free cash flow and accrual components follows Dechow, Richardson and Sloan (DRS 2008). We also apply the DRS technology in further disaggregating free cash flow into components that reflect the choices firms make in distributing (funding) the year s free cash flow surplus (deficit). The firm can choose to hold surplus free cash flow in the form of financial assets (i.e., cash) or distribute the surplus to investors (i.e., debt holders and equity holders). 3 On the other hand, the firm can choose to fund a free cash flow deficit by reducing the firm s cash balance or by obtaining additional capital from debt holders and equity holders. Thus, we disaggregate free cash flow into its change in cash, net distributions to equity holders and net distribution to debt holders components. We follow Richardson, Sloan, Soliman and Tuna (2001, 2006) in disaggregating accruals into sales growth and change in asset turnover (i.e., asset efficiency) components. Accruals (defined as changes in net operating assets) increase with increases in sales and, for a given level of sales, with decreases in asset efficiency, where asset efficiency equals sales divided by net operating assets. We assess analyst and market efficiency with respect to the persistence characteristics of net operating income as a whole and then disaggregated into: (i) free 3 Throughout the paper, we simply refer to financial assets as cash. 2

4 cash flow and accruals; (ii) free cash flow, sales growth and changes in asset efficiency; and (iii) accruals, changes in cash, net distributions to equity holders, and net distributions to debt holders. We assess analyst efficiency with and without controls for the influence of any bias in stock returns, and we assess market efficiency with and without controls for the influence of any bias in analysts earnings forecasts. In this manner, we assess how the market (analysts) would respond to earnings information if extant analysts forecasts (stock prices) had responded efficiently with respect to the same earnings information. In our sample of relatively large firms followed by analysts, we begin by estimating the persistence of annual net operating income and its components. Consistent with prior literature, we estimate that the persistence of the free cash flow component of net operating income exceeds the persistence of the accrual component by 26%, a difference that is economically and statistically significant. As expected, the sales growth component of accruals has significantly greater persistence than the change in efficiency component. All three components of free cash flow are highly persistent and, interestingly, the change in cash variable has greater persistence than the net distributions to debt holder and equity holder variables. To assess the bias in the stock price response to net operating income and its components, we regress returns for year t+1 on the year t income variables described above. The returns accumulation period extends for a full year beginning with the fifth month following the end of fiscal year t. To assess the bias in analysts earnings forecasts, we estimate two regressions. The first (second) evaluates the bias in analsyts year t+1 (t+2) earnings expectations as of the beginning of the return accumulation 3

5 period. The first (second) model regresses the year t+1 (t+2) earnings forecast error (revision) on the year t income variables. The key results are summarized as follows. First, removing the influence of the bias in analysts earnings forecasts, we find substantial evidence of market over-reaction to year t operating income and its components. The net distributions to debt holders variable is the only component with respect to which we have no evidence of market inefficiency. Otherwise, the evidence suggests that, without the influence of analysts biased response to year t earnings information, the stock market would over-react to the persistence characteristics of: net operating income as a whole; the free cash flow and accrual components of net operating income; the sales growth and changes in asset efficiency components of accruals; and the change in cash and net distributions to equity holders components of free cash flow. Second, without removing the influence of the bias in analysts earnings forecasts, we find no evidence of market inefficiency with respect to the persistence characteristics of: net operating income as a whole; the free cash flow component of net operating income; the sales growth component of accruals; and the change in cash and distributions to equity holders components of free cash flow. Without removing the influence of the bias in analsyts earnings forecasts, we only find market over-reaction to accruals and its change in asset efficiency component. 4 Thus, biased analsyts forecasts offsets what otherwise would be market over-reaction to net operating income as a whole and most of 4 Our evidence that efficiency changes drive the market over-reaction to accrual earnings contrasts with Hribar and Yehuda (2008), who provide evidence suggesting that accruals mispricing arises from misunderstanding of returns to the firm s growth opportunities, particularly during the firm s growth stage. Subsequent research might explore whether the sales growth versus efficiency change explanation for accruals mispricing depends on the stage of the firm in its life cycle. 4

6 its components. Biased analsyts earnings forecasts have the effect of mitigating market inefficiency. Third, removing any influence of biased stock price response to year t earnings information, we find that analsyts year t+1 earnings forecasts exhibit significant underreaction to net operating income and each of its individual components. Without removing any influence of biased stock price response, the only thing that changes is that we find no evidence of analyst under-reaction to the change in efficiency component of accruals. Thus, biased stock prices generally do not mitigate analyst under-reaction to the implications of current earnings and its components for predictions of next year s earnings. Under-reaction in analysts forecasts of year t+2 earnings is similarly unaffected by stock price over-reaction to year t earnings information. Only analyst under-reaction to the total accruals component of net operating income disappears when we remove controls for the over-reaction bias in stock prices. Overall, we conclude that, left to their own devices, analsyts would generally under-react and stock prices would generally over-react to the persistence characteristics of annual net operating income and its components. The general under-reaction in analysts earnings forecasts serves to mitigate the natural tendency of investors to overreact, and stock prices become much more efficient. Mitigation of biased reaction to earnings information is generally a one-way street, with analyst undereraction mitigating investor over-reaction, and with market over-reaction having virtually no effect on analyst under-reaction. This paper makes important contributions to the literature. First, we bring a new perspective to the literature regarding the role of financial analysts in capital markets. 5

7 Beginning with Abarbanell and Bernard (1992), many papers examine the role of financial analysts in relation to anomalous stock market behavior. However, these papers invariably look to analysts for root causes of market inefficiency. Instead, we find that analysts play a disciplining role that reins in the tendency for investors and stock prices to over-react to information about future earnings. Thus, the economic incentives leading to analyst under-reaction could include an incentive to put the brakes on what otherwise would appear as generalized market over-reaction to information about future earnings. We leave investigation of such incentives for future research. Second, we draw attention to a technique that others can use to investigate the interaction of analyst and market inefficiency in other contexts. Finally, to our knowledge, this paper is the first to investigate analyst forecasting bias with respect to the persistence characteristics of detailed components of both the free cash flow and accrual portions of annual net operating income. The rest of this paper is organized as follows. Section II reviews the most relevant literature. Section III provides a detailed description of our research design. Section IV describes our sample, section V presents our results, and section VI summarizes and concludes the paper. II. Prior Research Beginning with Sloan (1996), researchers have investigated stock market efficiency with respect to information in accrual and cash flow components of earnings. While Sloan examined the market s response to operating cash flows and working capital accruals minus depreciation, the research has evolved to consider the market response to 6

8 total accruals, accruals disaggregated into various components, and free cash flows disaggregated into retained and distributed components. Richardson et al. (2005) find that the market over-reacts to the persistence of both current (non-cash working capital) and non-current operating accruals and that trading strategies based on total operating accruals generate more profits than strategies based on current operating accruals. Richardson et al. (2001, 2006) disaggregate total operating accruals into sales growth and asset efficiency (i.e., turnover) components. Richardson et al. (2001) find that the market over-reacts to the persistence of both the sales growth and change in asset efficiency components of total accruals. The authors interpret this result to support both: (i) the perspective of Fairfield et al. (2003), Cooper et al. (2005) and others that the accrual anomaly emerges from diminishing returns on firms' growth opportunities, and (ii) the perspective of Sloan (1996), Xie (2001) and others that the accrual anomaly emerges from accounting distortions. We extend this research by examining the role of financial analysts in the interpretation of information in components of free cash flows and accruals. DRS demonstrate that the persistence of the free cash flow component of accounting earnings depends on how the firm chooses to distribute (fund) free cash flow surpluses (deficits). DRS find no evidence of market inefficiency with respect to the persistence of the portion of free cash flow that the firm distributes to (or obtains from) investors; whereas, the market over-reacts to the persistence of the portion of cash earnings that the firm retains (uses). This contradicts prior research evidence that the market under-reacts to total cash earnings [e.g., Sloan (1996); Desai et al. (2004); Ahmed et al. (2006)]. In a sample of larger firms followed by analysts during a more recent time 7

9 period, we re-examine the market's response to cash components of earnings, and extend the literature by examining the role of analysts forecasting behavior in market efficiency. Many prior research papers investigate the role of financial analysts in making the market inefficient [e.g., Shane and Brous (2001) in the context of post-earnings and postforecast revision drift; La Porta (1996) and Doukas, Kim and Pantzalis (2002) in the context of the value-glamour stocks anomaly; Dechow and Sloan (1997) in the context of market optimism with respect to firms analysts long-term growth forecasts; Rajan and Servaes (1997), Dechow, Hutton and Sloan (1999), Teoh and Wong (2002) and Purnanandam and Swaminathan (2004) in the context of long-run underperformance of IPOs and SEOs; Billings and Morton (2001) in the context of the book-to-market anomaly; Bradshaw and Sloan (2002) in the context of emphasis on street earnings ; Ikenberry and Ramnath (2002) in the context of under-reaction to the information content of stock splits; Elgers, Lo and Pfeiffer (2003) in the context of over-reaction to the persistence of working capital accruals; Kadiyala and Rau (2004) in the context of corporate events including mergers and stock repurchases; and Jackson and Johnson (2006) in the context of returns momentum]. Some studies refer to market under-reaction to information in news and events and investigate the role of financial analysts in speeding up the market s assimilation of that information [e.g., Gurun, Johnston and Markov (2011) in the context of debt markets; and Hong, Lim and Stein (2000) in the context of equity markets]. However, our study is the first to investigate whether the well-known under-reaction in analsyts earnings forecasts reins in what otherwise would be market over-reaction to the 8

10 implications of current earnings (and its components) for predictions of future earnings, and we have a unique approach for doing so (as described in section III below). III. Research design Our investigation of biased analyst earnings forecast and stock price response to cash and accrual earnings components builds on earnings disaggregations developed by Richardson et al. (2001, 2006) and Dechow et al. (2008). Consistent with Dechow et al. (2008), we disaggregate earnings into free cash flow (cash) and accrual components as follows (firm subscripts suppressed throughout all models). The balance sheet equation implies: ΔA t = ΔL t + ΔSE t where t identifies the firm s fiscal year, A t = total assets, L t = total liabilities, and SE t = total equity holders equity. 5 Then, separating financing from operating activities yields: ΔOA t + ΔCASH t = ΔOL t + ΔFL t + ΔPSE t + ΔCSE t where OA t = operating assets, CASH t = all financial assets, OL t = operating liabilities, FL t = financial liabilities, PSE t = preferred equity holders equity, and CSE t = common equity holders equity. Rearranging terms: ΔNOA t = ΔDEBT t ΔCASH t + ΔCSE t, (1) where NOA t = OA t OL t = net operating assets, and DEBT t = all non-owner financing, including financial liabilities and preferred equity holders equity. Assume clean surplus: ΔCSE t = NI t DIST_EQ t, (2) 5 The appendix provides detailed definitions of each variable used in the study, along with Compustat labels. 9

11 where NI t = net income and DIST_EQ t = net distributions to common equity holders. DIST_EQ t represents distributions to common equity holders by way of dividends and stock repurchases net of new owner investment. Substituting the right-hand side of (2) for ΔCSE t in (1), decomposing NI t into net operating income (NOI t ) and net financing expense (NFE t ), defining DIST_D t (net distributions to non-owner providers of financing) as NFE t ΔDEBT t and rearranging terms yields NOI t ΔNOA t = ΔCASH t + DIST_D t + DIST_EQ t (3) Thus, free cash flow generated by operations is represented on the left-hand side of (3) as net operating income less the change in net operating assets, and uses of free cash flow are represented on the right-hand side of (3) as the change in cash (build-up of the investment in financial assets) plus distributions to non-owner contributors of financing plus distributions to owners. All terms in (3) can be negative or positive; e.g., when NOI t ΔNOA t < 0, free cash flow is used in operations and the right-hand side of (3) must also be negative indicating the sources of the free cash flow used in operations (net decreases in the cash balance and/or net new investment by non-owner and owner providers of capital). Defining ΔNOA t as accruals (ACC t ) and ΔCASH t + DIST_D t + DIST_EQ t as free cash flow (FCF t ) yields NOI t = FCF t + ACC t. Thus, net operating income equals free cash flow (i.e., cash earnings) plus accruals (i.e., accrual earnings). Finally, deflate all terms by NOA t-1 and define NOI t /NOA t-1 as RNOA t. RNOA t = FCF t + ACC t (4) 10

12 where RNOA t = NOI t / NOA t-1 = return on net operating assets, FCF t = free cash flow deflated by lagged net operating assets, and ACC t = ΔNOA t deflated by lagged net operating assets. 6,7 Following Richardson et al. (2001, 2006), we define and disaggregate total operating accruals as follows: ACC t = SG t ΔEFF t (SG t * ΔEFF t ) (5) where: SG t = ΔSales t /Sales t-1 and represents sales growth; ΔEFF t = ΔAT t /AT t and represents the change in asset efficiency measured as (Sales t /NOA t Sales t-1 /NOA t-1 ) / (Sales t /NOA t ) (i.e., Sales t /NOA t is a measure of asset turnover or asset efficiency). Richardson et al. (2006) provide an algebraic proof showing that the three terms on the RHS of (5) sum to total accruals (ACC t ). The authors argue that this decomposition allows tests with the potential to distinguish between two competing explanations for the accrual anomaly: market failure to fully impound mean reversion in sales growth (i.e., over-reaction to sales growth) versus market failure to fully impound information in accruals about temporary accounting distortions (i.e., holding sales constant, ΔEFF t increases with decreases in net operating assets, and the market might fail to impound the temporary nature of the portion of changes in NOA due to accounting distortions). If diminishing marginal returns on investment drive lower persistence of accruals, this should be reflected in the growth component of ACC t [i.e., SG t in equation (5)]. In contrast, if accounting distortion or declining operating asset efficiency drives lower 6 Theoretically, RNOA t in (1) should be NOI t (i.e., net operating income after tax). Following Richardson et al. (2005, 2006), we simply represent RNOA t as income after depreciation (Compustat item # 178). 7 Theoretically, distributions to preferred equity holders should be included in DIST_D. Following Dechow et al., we include distributions to/from preferred equity holders in DIST_EQ. 11

13 persistence of accruals, then this should be reflected in the efficiency component of ACC t [(i.e., ΔEFF t in equation (5)]. 8 Summarizing (3), (4) and (5) above, our fully disaggregated model of net operating income becomes: RNOA t = (ΔCASH t + DIST_D t + DIST_EQ t ) + [SG t ΔEFF t (SG t * ΔEFF t )] (6) where the first bracketed term on the RHS of (6) disaggregates free cash flow (FCF t ), and the second bracketed term disaggregates accruals (ACC t ). 9 Earnings Persistence Replacing RNOA t in model (6) with RNOA t+1, regression models (7) through (10) examine the persistence of net operating income and the various components described above. RNOA t+1 = ψ 0 + ψ 1 RNOA t + e t+1 (7) RNOA t+1 = θ 0 + θ 1 FCF t + θ 2 ACC t + e t+1 (8) RNOA t+1 = β 0 + β 1 ACC t + β 2 CASH t + β 3 DIST_EQ t + β 4 DIST_D t + e t+1 (9) 8 Results in Hribar and Yehuda (2008) suggest that market over-reaction to the persistence of firms growth opportunities, rather than over-reaction to the persistence of accounting accruals, drives the accruals anomaly, particularly for growth firms early in their life cycle. If this is the case, then we expect that the sales growth component of equation (5) should drive any market over-reaction to accruals. 9 Since Compustat does not aim to produce the theoretically sound summary variables represented in (6), we are left with the following issues, which we do not believe bias our results: (a) FCF is in pretax dollars, because NOI is represented by the Compustat variable, OIADP, which is operating income after depreciation but before tax. Thus, FCF and DIST_EQ are over- or under-stated by the same amounts. (b) Net financing expense is omitted from the calculations, causing misstatement of FCF and DIST_D by the same amounts. (c) Preferred stock and minority interest are included with equity holder investment. This causes misstatement of DIST_D and DIST_EQ by opposite amounts. (d) Variable measurement ignores other comprehensive income and any other changes in equity holders equity that do not pass through operating income and are not due to owner investment or disinvestment. This causes a misstatement of FCF and DIST_EQ by the same amounts. 12

14 RNOA t+1 = λ 0 + λ 1 FCF t + λ 2 SG t + λ 3 ΔEFF t + λ 4 (SG t * ΔEFF t ) + e t+1 (10) Model (7) represents basic earnings persistence where the return on net operating assets is a function of the lagged value of the same variable, and the coefficient ψ 1 estimates earnings persistence. Models (8) through (10) allow components of RNOA t to have different persistence parameters. Positive θ 1 and θ 2 in model (8), and positive β 1 and λ 1 in models (9) and (10), indicate that the summary earnings components, FCF t and ACC t, respectively, are positively related to future earnings; i.e., they persist, but not necessarily to the same degree (θ 1 might differ from θ 2, and both could differ from the summary persistence parameter, ψ 1 ). In models (9) and (10), positive β 2, β 3, β 4 and λ 2, and negative λ 3 and λ 4 indicate that the detailed earnings components, CASH t, DIST_EQ t, DIST_D t, SG t, and ΔEFF t, respectively, are positively related to future earnings (i.e., they persist but, again, not necessarily to the same degree). Market Efficiency Next, we replace RNOA t+1 in (7)-(10) with RET t+1 and evaluate market efficiency with respect to the various components of cash and accrual earnings using the following models. RET t+1 = φ 0 + φ 1 RNOA t + e t+1 (11) RET t+1 = ρ 0 + ρ 1 FCF t + ρ 2 ACC t + e t+1 (12) RET t+1 = κ 0 + κ 1 (FCF t ) + κ 2 (SG t ) + κ 3 (ΔEFF t ) + κ 4 (SG t *ΔEFF t ) + e t+1 (13) RET t+1 = γ 0 + γ 1 (ΔCASH t ) + γ 2 (DIST_EQ t ) + γ 3 (DIST_D t ) + γ 4 ACC t + e t+1 (14) 13

15 where RET t+1 = firm i s raw returns, accumulated from the fifth month following the end of fiscal year t through the fourth month following the end of fiscal year t+1, minus the similarly accumulated mean return of all firms in the same size-decile as firm i (with size deciles formed as of the end of fiscal year t). 10 If the market efficiently impounds the information in year t operating income and its components, then the coefficients on all variables in models (11)-(14) are zero. On the other hand, if, as described in Richardson et al. (2001), stock prices over-react to the persistence of sales growth and efficiency components of total accruals, then we expect κ 2 < 0 and κ 3 >0 in (13) above. 11 If, as described in Dechow et al. (2008), stock prices over-react to the persistence of the retained component of cash earnings, then we expect γ 1 < 0 in (14) above. 12 Next we use models (15)-(18) below to assess how biased analyst earnings forecasts affect any market inefficiency estimated in models (11)-(14) above. RET t+1 = α 0 + α 1 RNOA t + α 2 FE t+1 + α 3 FREV t+2 + e t+1 (15) RET t+1 = η 0 + η 1 FCF t +η 2 ACC t + η 3 FE t+1 + η 4 FREV t+2 + e t+1 (16) 10 Accumulating returns from the beginning of the fourth month following the end of fiscal year t does not qualitatively change our results. 11 Controlling for sales growth, increases in asset efficiency (ΔEFF t ) correspond to decreases in NOA t and decreases in accruals. Thus, if investors over-react to the persistence of accruals components of earnings, then ΔEFF t > 0 (corresponding to decreases in accruals) should lower stock prices, manifest as higher subsequent returns (as correction occurs), and appear as κ 3 > 0 in model (13). 12 Rather than using the framework developed by Mishkin (1983), we use a conventional OLS model to test the rational expectations hypotheses. Kraft et al. (2007) show that OLS and the Mishkin test generate identical inferences in accounting settings when samples are large. As described in Kraft et al. (2007), to obtain estimates of the difference between actual and implied market estimates of the persistence of earnings components, the residual terms from models (9) and (10) are added as independent variables in models (13) and (14). Then the difference between actual and implied market persistence estimates are derived by dividing each component variable coefficient by the residual variable coefficient in models (13) and (14). 14

16 RET t+1 = π 0 + π 1 (FCF t ) + π 2 (SG t ) + π 3 (ΔEFF t ) + π 4 (SG t )*(ΔEFF t ) + π 5 FE t+1 + π 6 FREV t+2 + e t+1 (17) RET t+1 = δ 0 + δ 1 (ΔCASH t ) + δ 2 (DIST_EQ t ) + δ 3 (DIST_D t ) + δ 4 (ACC t ) + δ 5 FE t+1 + δ 6 FREV t+2 + e t+1 (18) where FREV t+2 = ( t+1 F t+2 t F t+2 ) / P t ; t+1 F t+2 is the first individual analyst forecast of t+2 earnings issued after the end of the return accumulation period; t F t+2 is the most recent forecast of t+2 earnings issued prior to the beginning of the return accumulation period; FE t+1 = A t+1 F t+1 ; A t+1 = actual year t+1 earnings (per I/B/E/S); and F t+1 is the most recent forecast of t+1 earnings issued prior to the beginning of the return accumulation period. 13 Models (15)-(18) describe the returns models that control for bias in analyst response to RNOA t and its components. Each of these models includes forecast error and forecast revision variables that remove the variability in RET t+1 due to the influence of any biased forecast response to RNOA t and its components. If analyst forecasting behavior mitigates (drives) market over-reaction to the persistence characteristics of RNOA t and its components, then we expect the under/over-reaction coefficients to become increasingly negative and significant (approach zero) due to the inclusion of control variables, FE t+1 and FREV t Including FE t+1 and FREV t+2 in the returns models above removes the portion (if any) of the estimated under/over-reaction coefficients on RNOA t and its components due 13 See Figure 1 for a pictorial representation and definitions of the market and analyst reaction variables: RET t, FE t+1 and FREV t See Shane and Brous (2001) for a detailed explanation of this approach, which those authors use to gauge the degree to which analyst forecasting behavior drives drifts in returns following quarterly earnings news. 15

17 to biased response of FE t+1 and FREV t+2 to RNOA t and its components. We are then left with under/over-reaction coefficients reflecting the bias, if any, in how the market would have responded to RNOA t and its components without the influence of any analyst forecasting bias. Comparing the under/over-reaction coefficients in models (15) through (18) to the under/over-reaction coefficients in models (11) through (14) provides insight into the portion of the bias in the stock price response to RNOA t and its components due to bias in analysts forecasting response to the same information. As described in Shane and Brous (2001), if analyst forecasting behavior completely drives market inefficiency with respect to information about future earnings, then adding the forecast error and forecast revision variables to the returns regressions should make the coefficients on the information variables go to zero. On the other hand, if the market would have under/over-reacted to RNOA t and its components without the influence of any bias in analysts forecasts, then the under- (over-) reaction coefficients should be significantly positive (negative) in models (15) through (18). Financial Analyst Efficiency Next we assess whether any bias in the market s response to RNOA t and its components influences any bias in analysts response to the same information. Models (19) through (22) below evaluate analyst two-year-ahead forecasting efficiency before controlling for effects of investor behavior. FREV t+2 = φ 0 + φ 1 RNOA t + e t+2 (19) FREV t+2 = ρ 0 + ρ 1 FCF t + ρ 2 ACC t + e t+2 (20) 16

18 FREV t+2 = κ 0 + κ 1 (FCF t ) + κ 2 (SG t ) + κ 3 (ΔEFF t ) + κ 4 (SG t *ΔEFF t ) + e t+2 (21) FREV t+2 = γ 0 + γ 1 (ΔCASH t ) + γ 2 (DIST_EQ t ) + γ 3 (DIST_D t ) + γ 4 ACC t + e t+2 (22) If analysts issue efficient forecasts following firms release of their financial statements containing cash and accrual components of earnings, then the information in those financial statements should not predict analysts forecast revisions (FREV t+2 ), and the coefficients on the earnings variables should equal zero. The intercept terms capture analysts general optimism/pessimism. For example, significantly negative intercepts in models (19)-(22) indicate systematically optimistic t F t+2. To remove any influence of biased stock price response to RNOA t and its components, models (23)-(26) add year t+1 returns, accumulated over the same time as FREV t+2, to the FREV t+2 regressions. FREV t+2 = α 0 + α 1 RNOA t + α 2 RET t+1 + e t+2 (23) FREV t+2 = η 0 + η 1 FCF t + η 2 ACC t + η 3 RET t+1 + e t+2 (24) FREV t+2 = π 0 + π 1 (FCF t ) + π 2 (SG t ) + π 3 (ΔEFF t ) + π 4 (SG t *ΔEFF t ) + π 5 RET t+1 + e t+2 (25) FREV t+2 = δ 0 + δ 1 (ΔCASH t ) + δ 2 (DIST_EQ t ) + δ 3 (DIST_D t ) + δ 4 ACC t + δ 5 RET t+1 + e t+2 (26) Including RET t+1 in models (23)-(26) removes any impact of bias in the price reaction to the persistence characteristics of RET t+1 and its components on the under/over-reaction coefficients, which now represent analyst inefficiency with respect to earnings information absent the influence of any bias in stock prices. If analyst forecasting behavior simply mimics the behavior of the marginal investor, then adding RET t+1 to the FREV t+2 regressions should make the under/over-reaction coefficients on 17

19 RNOA t and its components go to zero. On the other hand, if investor over-reaction mitigates analyst under-reaction to RNOA t and its components, then the coefficients in models (23) through (26) should become significantly positive and more so than in models (19) through (22). Finally, we use models (27) through (34) below to evaluate analyst one-yearahead forecasting efficiency by estimating the relation between year t earnings information and the year t+1 analyst forecast error. FE t+1 = φ 0 + φ 1 RNOA t + e t+1 (27) FE t+1 = ρ 0 + ρ 1 FCF t + ρ 2 ACC t + e t+1 (28) FE t+1 = κ 0 + κ 1 (FCF t ) + κ 2 (SG t ) + κ 3 (ΔEFF t ) + κ 4 (SG t *ΔEFF t ) + e t+1 (29) FE t+1 = γ 0 + γ 1 (ΔCASH t ) + γ 2 (DIST_EQ t ) + γ 3 (DIST_D t ) + γ 4 ACC t + e t+1 (30) FE t+1 = α 0 + α 1 RNOA t + RET t+1 + e t+1 (31) FE t+1 = η 0 + η 1 FCF t + η 2 ACC t + η 3 RET t+1 + e t+1 (32) FE t+1 = π 0 + π 1 (FCF t ) + π 2 (SG t ) + π 3 (ΔEFF t ) + π 4 (SG t *ΔEFF t ) + π 5 RET t+1 + e t+1 (33) FE t+1 = δ 0 + δ 1 (ΔCASH t ) + δ 2 (DIST_EQ t ) + δ 3 (DIST_D t ) + δ 4 ACC t + δ 5 RET t+1 + e t+1 (34) Models (27) through (30) estimate under/over-reaction coefficients without controlling for biased stock prices, and models (31)-(34) add RET t+1 to the right-hand side in order to negate any analyst under/over-reaction coefficient effects due to the influence (if any) of biased stock prices. These models work exactly like models (19)- (26) described above, except we replace FREV t+2 with FE t+1 as the dependent variable. IV. Sample 18

20 Since we rely on I/B/E/S to measure the forecast error (FE t+1 ) and forecast revision (FREV t+2 ) variables in various models throughout this study, our sample represents larger firms and a more recent time period relative to the sample and time periods in Dechow et al. (2008) and Richardson et al. (2001, 2006). Our time period spans the years ; whereas, Richardson et al. (2001) spans , Richardson et al. (2006) spans , and Dechow et al. (2008) spans To estimate the variables in our models, we obtain: financial statement data from Compustat; returns data from CRSP; and earnings forecasts, actual earnings, and stock prices from I/B/E/S. Table 1 describes our sample selection procedure. Our initial sample finds 174,907 firm-year observations on Compustat s Annual Industrial, Research, and Full Coverage files between the years, 1988 and Following Richardson et al. (2005) and Richardson et al. (2006), we exclude firms in the financial services industry (32,860 firm-years with SIC codes in the range ). We omit 22,819 observations without Compustat data needed to compute our cash flow and accrual earnings component variables. We exclude 8,197 observations with net operating assets less than zero. We lose 32,510 observations missing the CRSP data needed to compute our returns variable. We lose another 42,751 firm-year observations without I/B/E/S data needed to compute our forecast error variable (FE t+1 ), and we lose 14,139 observations without I/B/E/S data needed to compute our forecast revision variable (FREV t+2 ). We exclude 3,916 observation outliers, with any earnings component or analyst forecast variable greater than 1 or less than -1, resulting in a final sample of 27,686 firm-years spanning {Insert Table 1 about here} 19

21 V. Results Descriptive statistics Table 2 presents descriptive statistics, including correlations between variables used in the study. Panel A shows that, overall, the market was relatively flat during our sample period (mean RET = 0.042, median = ). The forecast error and forecast revision variables are deflated by the stock price reported in the most recent I/B/E/S report prior to the return accumulation period. Analysts provide slightly optimistic current year forecasts for our sample firms (FE t+1 < 0), and, consistent with Raedy et al. (2006), we see evidence of horizon-dependent optimism whereby, as of the beginning of the return accumulation period, forecasts of year t+2 earnings are more optimistic (and result in a more negative expectation adjustment) than forecasts of year t+1 earnings (i.e., FREV t+2 < FE t+1 < 0). {Insert Table 2 about here} RNOA t is naturally deflated by beginning of year net operating assets, and the accrual and free cash flow components of net operating income are similarly deflated. Panel A shows that, on average, one-third (two-thirds) of the 20.5% mean return on net operating assets comes from the free cash flow (accruals) component of net operating income. Specifically, the mean FCF t = 7.5% and the mean ACC t = 13%; however, the median FCF t is 9.8% and the median ACC t is 7.3%. ACC t appears to be skewed towards the right-hand side of the distribution. At 9.6% of net operating assets, on average, 20

22 distributions to equity holders drives the free cash flow contribution to RNOA t ; and at 13.3%, sales growth drives the contribution of accruals to RNOA t. Panel B of Table 2 provides pair-wise Pearson and Spearman correlations among the variables used in this study. The significant positive contemporaneous correlation between RET t+1 and FE t+1, and between RET t+1 and FREV t+2 (Pearson correlation coefficients = 0.19 and 0.40, respectively, with p-values < ) reflects strong common elements between the behavior of analysts and investors. We attribute the stronger correlation with FREV t+2 to a better match between the forecast revision and return accumulation windows and to the relatively long term perspective of both variables.. The FREV t+2 window ends at (or shortly after) the end of the return window (4 months after the end of fiscal year t+1), whereas, the FE t+1 window ends with the year t+1 earnings announcement (about 9 or 10 months into the 12-month return accumulation period). In any case, the return and analyst variables are not perfectly correlated, which creates tension and makes it is an empirical question as to whether bias in the analyst variables creates bias in the returns variable and vice versa. On a univariate basis, consistent with Sloan (1996), it appears that investors overreact to accruals (significant negative correlation between ACC t and RET t+1 ) and underreact to the free cash flow component of earnings (significant positive correlation between FCF t and RET t+1 ). Investors appear to over-react to both components of accruals, as we find significant negative correlation between SG t and RET t+1 and significant positive correlation between ΔEFF t and RET t+1. Consistent with Dechow, et al. (2008), the significant positive correlation between returns and distributions to both equity holders and debt holders appears to drive the significant positive correlation 21

23 between free cash flow and next year s returns. Finally, the significant negative Pearson correlation coefficient relating ΔCASH t to RET t+1 supports Dechow, et al. s inference that investors over-react to the change in cash component of free cash flow; although the Spearman correlation between ΔCASH t and RET t+1 is not significant. Turning to analyst forecasting behavior, consistent with Bradshaw et al. (2001), the univariate correlations in Table 2 Panel B suggest that analysts over-react to the persistence of accruals (significant negative correlations between ACC t and both FREV t+2 and FE t+1 ). Also, it appears that analysts under-react to free cash flow (significant positive correlation between FCF t and both FREV t+2 and FE t+1 ). Univariate statistics suggest that analyst under-reaction to free cash flow permeates all three free cash flow components: ΔCASH t, DIST_D t and DIST_EQ t. The correlation matrix suggests that analysts, at least with reference to their longer horizon forecasts, over-react to both components of accruals. Table 2 Panel B reports significant negative correlation between SG t and FREV t+2 (Pearson correlation = and Spearman correlation = ), and significant positive Pearson correlation between SG t and FE t+1. Panel B reports significant positive correlations between ΔEFF t and both FE t+1 and FREV t+2, suggesting that analysts over-react to the change in asset efficiency component of accruals. Overall, the univariate tests suggest that analysts tend to over-react to accounting accruals and its components, at least over longer horizons. The significant negative correlation between FCF t and ACC t (-0.60 Pearson and Spearman correlation) means that only a multivariate test can sort out each variable s unique relation with future returns (RET t+1 ), future forecast errors (FE t+1 ) and future forecast revisions (FREV t+2 ). Thus, conclusions, drawn from univariate statistics in Table 22

24 2, about analyst/market over/under-reaction to free cash flow and accrual earnings components are premature. Earnings Persistence Table 3 shows that the cash and accrual earnings components, derived from our sample, exhibit persistence characteristics consistent with prior literature. As described by the regression in Panel A, return on net operating assets persists strongly from one year to the next. The regression of RNOA t+1 on RNOA t has a highly significant slope coefficient of 0.68 and the adjusted R-square statistic indicates that RNOA t explains 42.2 per cent of the variation in RNOA t+1. Panel B shows that coefficients relating cash and accrual earnings to next year earnings are and 0.589, respectively, and while both coefficients are significantly greater than zero, they are also significantly different from each other (p-value < ). 15 Consistent with Sloan (1996) cash flows are significantly more persistent than accruals. Table 3 Panels C and D provide evidence that every component of cash and accrual earnings is highly persistent. Each disaggregated cash and accrual earnings component is significantly related to RNOA t+1 with the expected sign. The persistence parameters range from a low of on the efficiency change component of accruals to a high of on the retained cash component of free cash flow. Untabulated F-tests show that each coefficient estimate is significantly different from each other, at a one percent level. Thus, disaggregating earnings into components that have different persistence characteristics provides useful information for purposes of forecasting earnings and valuing securities. 15 All p-values refer to two-tailed significance levels. 23

25 {Insert Table 3 about here} Market Efficiency Table 4 describes the behavior of stock returns following publication of financial statements with enough information to disaggregate earnings into the accrual and cash components described in table 3 and models 9 and 10. In Panel A estimates of model (15) coefficients indicate that, controlling for the bias in analysts forecasts, the market s overreaction coefficient equals (p-value < 0.01) with reference to net operating income as a whole. Removing the control variables, the ostensible market over-reaction coefficient is statistically insignificant at , and the p-value associated with the 30% decline in absolute value from to is less than Apparently, with reference to RNOA t, under-reaction bias in analysts earnings forecasts mitigates the over-reaction in stock market prices to the point where it registers as statistically insignificant. {Insert Table 4 about here} We attribute the stronger relation between RET t+1 and FREV t+2 (as opposed to FE t+1 ) to three factors. First, we have a closer match between the information acquisition windows for FREV t+2 and RET t+1 (i.e., both are measured over the entire 12-month return accumulation period). Second, like RET t+1, FREV t+2 is more forward looking (as compared to FE t+1 ), and Liu and Thomas (2000) find that earnings response coefficients 24

26 increase as the earnings variable contains more forward information. Finally, FREV t+2 provides analysts with enough time to analyze the full set of year t+1 financial statements (including the ability to disaggregate operating income into its cash and accrual components), as opposed to just the earnings number that represents the earnings expectation adjustment associated with FE t+1. We conjecture that these three reasons lead to the much stronger relation between RET t+1 and FREV t+2, as compared to the relation between RET t+1 and FE t+1, identified in Table 4 Panel A. In Table 4 Panel B, like Richardson et al. (2001, 2005), we find a significantly negative relation between total accruals and following year returns, suggesting market over-reaction to the persistence of accrual earnings. Also, like Richardson et al., we do not find a significant relation between cash earnings and subsequent returns. Compared to Richardson et al., the coefficient on the total accruals variable is smaller in our sample. We attribute this difference to mitigation of the accrual anomaly (i.e., the negative relation between accruals and subsequent returns) during the time period. 16 As shown in panel B, without FE t+1 and FREV t+2 in model (12), the coefficient relating ACC t to RET t+1 equals and is significantly less than zero (two tailed p-value < 0.05). Adding FE t+1 and FREV t+2 to model (12) does not enhance the estimated market over-reaction coefficient. In fact, panel B shows that the coefficient on ACC t, shrinks slightly to , but becomes more significant (p-value < 0.01). Apparently, any analyst inefficiency with respect to the persistence of accrual earnings does not explain/mitigate the market over-reaction. 16 Our results do not change when we remove our forecast error and forecast revision data constraint. However, when we maintain that constraint but eliminate the years , the estimated coefficient on the accruals variable becomes similar in magnitude to the coefficient in the Richardson et al. study. 25

27 With respect to the free cash flow component of net operating income, it appears that analysts under-react, because the coefficient relating year t+2 returns to year t free cash flow becomes significantly negative when we control for analyst forecasting behavior. Adding FREV t+2 and FE t+1 to model (12) causes the coefficient relating FCF t with RET t+1 to drop from a statistically insignificant to a highly significant , and the p-value associated with the effect of controlling for analyst forecasting behavior is less than Thus, it appears that analyst under-reaction mitigates market overreaction to free cash flow. A picture of analyst under-reaction and investor over-reaction to annual earnings and its components is beginning to emerge. Without controlling for analyst underreaction, the strong positive relation between RET t+1 and FREV t+2 obscures the influence of analyst forecasting bias on the stock price over-reaction to FCF t. Controlling for the strong positive relation between investor and analyst behavior, we find evidence of general market over-reaction to both cash and accrual earnings components. The market over-reaction to cash earnings that would occur without the influence of the bias in analysts forecasts is a new result, inconsistent with inferences drawn from Sloan (1996) and the many papers that followed. Panel C of Table 4 replicates the Richardson et al. (2001) analysis of the relation of cash flow and disaggregated accrual earnings components with following year returns. In both our replication and the Richardson et al. (2001) study, before controlling for analyst behavior, the cash flow component of earnings is not significantly related to future returns, and the change in asset efficiency component of accruals is positively related to future returns (indicating over-reaction). The estimated coefficient on the 26

28 change in asset efficiency component of accruals is about 40% smaller than the estimated coefficient in the Richardson et al. (2001) study. 17 Nonetheless, using our smaller sample of larger firms in a more recent time period, we are able to replicate the Richardson et al. (2001) finding that the market apparently over-reacts to the persistence of the change in the efficiency component of total accruals. 18 Interestingly, the coefficient on the sales growth component of total accruals becomes significantly negative only when we control for analyst forecasting behavior in model (17). Omitting FREV t+2 and FE t+1 from model (13) obscures the mitigating influence of biased analyst forecasts, which enhances the market s over-reaction coefficient on SG t from to a statistically significant (p-value< 0.05). Controlling for analyst forecasting behavior, estimation of model (17) in Panel C reveals generalized market over-reaction to free cash flow and both components of accruals. Apparently, analyst under-reaction mitigates investors tendency to over-react and leaves the market with a relatively efficient price with respect to persistence characteristics of FCF t and SG t. We will explore the efficiency of the analyst reaction more directly in tables 5 and 6, but the evidence in table 4 suggests that analysts under-react at least to the free cash flow and sales growth components of net operating income. Finally, Panel D of Table 4 shows that, without controlling for analyst forecasting behavior, estimates of coefficients in model (14) show no evidence of market inefficiency with respect to the components of free cash flow. However, controlling for FE t+1 and 17 The analysis of the relation of future returns with current accruals disaggregated into sales growth and change in asset efficiency components appears only in the Richardson et al. (2001) working paper and not in any subsequent published papers. 18 Recall that, holding sales growth constant, an increase in asset turnover (efficiency) corresponds to a decrease in accruals, so a positive coefficient on the efficiency change variable indicates that the market has over-reacted to the persistence of this component of accruals. 27