Journal of accounting, auditing finance tập 27, số 02, 2012 4

The Predictive Value of Accruals and Consequences for Market Anomalies Journal of Accounting, Auditing & Finance 27(2) 151–176 Ó The Author(s) 2012 Reprints and permission: sagepub.com/journalsPermissions.nav DOI: 10.1177/0148558X11409149 http://jaaf.sagepub.com Seunghan Nam1, Francois Brochet2, and Joshua Ronen3 Abstract In this article, the authors revisit the role of the cash and accrual components of accounting earnings in predicting future cash flows using out-of-sample predictions and market value of equity as a proxy for all future cash flows They find that, on average, accruals improve upon current cash flow from operations (CFO) in predicting future cash flows In the crosssection, accruals’ contribution is positively associated with proxies for quality of accruals and governance Next, the authors investigate the implications of accruals’ predictive value for accrual-based market anomalies They find that portfolios formed on stock return predictions using information from current CFO and accruals yield significantly positive returns on average, as opposed to CFO alone They also find that Sloan’s accrual anomaly is related to our accrual contribution anomaly Indeed, when accruals’ contribution to future cash flow prediction is the highest, the accrual anomaly vanishes Collectively, the results suggest that the predictive value of accruals and market participants’ ability to process it are a significant driver of accrual-based anomalies Keywords accruals, cash flows, cash flow predictions, anomalies The amount of aggregate future cash flows is key to the valuation of a firm’s securities Alternative valuation models by both academics and financial analysts have focused on the prediction of free cash flows (FCFs; Copeland, Koller, & Murrin, 1994) or residual income (Edwards & Bell, 1961; Ohlson, 1995; Preinreich, 1938) The prediction of cash flows is invariably based on past accounting numbers One question that has occupied much of the researchers’ attention is the extent to which the accrual component of past earnings contributes to the prediction of future realizations of cash flows and market participants’ expectations of future cash flows Rensselaer Polytechnic Institute, Troy, NY, USA Harvard Business School, Boston, MA, USA New York University, Stern School of Business, USA Corresponding Author: Seunghan Nam, Lally School of Management & Technology, Rensselaer Polytechnic Institute, 110 8th Street, Room 1120, Pittsburgh Bldg., Troy, NY 12180, USA Email: nams2@rpi.edu Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 152 Journal of Accounting, Auditing & Finance Promoting the accrual basis of accounting, the Financial Accounting Standards Board (FASB) asserts that earnings and their components are better predictors of future cash flows than current cash flow (FASB, 1978) In spite of the FASB argument, scholars and practitioners argue that the subjectivity inherent in estimates embedded in accruals introduces noise that can have a negative impact on their informational value (Dechow & Dichev, 2002) Firm managers may engage in self-serving earnings manipulation by reporting numbers based on distorted estimates, which has been shown to decrease the value relevance of earnings (Marquardt & Wiedman, 2004) Hence, whether they are made in good faith or with manipulative intent, accruals can be misleading and not representative of firm’s future performance We first revisit the findings on cash flow predictability by testing the Dechow, Kothari, and Watts’s (1998) theoretical predictions with a methodology that simultaneously addresses the following three dimensions: (a) judgment of the superiority of the predictor being based on out-of-sample forecasts rather than in-sample properties such as R2, (b) the estimation of firm-specific versus cross-sectional coefficients, and (c) the level of aggregation of future cash flows as the predicted variable Our evidence based on these methodological choices supports the view that accruals contribute to the prediction of future cash flows and provides detailed information on cross-sectional differences in the predictive value of accruals.1 In addition, we compare out-of-sample forecasts of future market capitalizations using firm-specific regressions with and without accruals as a predictor We consider market values of equity as the best available proxy for the present value of all future cash flows, that is, the highest level of aggregation of future cash flows After obtaining these forecasts, we compute predicted returns derived from the forecasts and form portfolios on the basis of the sorted predicted returns We are thus able to assess whether investors properly use the predictive ability of current accounting data for future cash flows in forming their expectations, in which case our sorting procedure should not predict actual stock returns However, we find that investors fail to fully understand the predictive ability of accruals in their investment decisions We also establish that the more predictive content accruals have, the more accurately investors are able to use them in investment decisions More specifically, we show that accrual anomalies (e.g., Sloan, 1996) are more a consequence of current accruals’ ability to forecast future cash flows than other cross-sectional differences like sign or size Our sample utilizes post-SFAS 95 quarterly data from Compustat We define cash flow as cash flow from operations (CFO) and accruals as the difference between net income and CFO, consistent with Hribar and Collins (2002) In our main analysis, we require 56 timeseries observations to develop firm-specific regression estimates As a result, our holdout sample period is from the third quarter of 2002 to the fourth quarter of 2006 To account for seasonal variations in quarterly cash flows, we deseasonalize our data using the X11 method developed by the U.S Bureau of Census.2 The economic significance of accruals’ predictive ability in our sample is most pronounced when the predicted variable is current or one-quarter-ahead market value of equity: The model including accruals as a predictor along with CFO exhibits significantly smaller mean and median absolute prediction errors than the model using current CFO alone, by about 5% of total assets In our portfolio tests, the average hedge return adjusted for the three Fama–French factors and momentum for a 90-day holding period when going long (short) on the highest (lowest) quintile of the quarterly predicted return distributions is insignificantly different from zero, with or without accruals as a predictor However, as the holding period Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Nam et al 153 increases, the returns earned on the portfolio using CFO and accruals become significantly higher than those using CFO only as a predictor For instance, 270- and 365-day incremental returns when accruals are added as a predictor are about 2% per quarter on average As for our tests related to the accrual anomaly, we replicate the results first documented by Sloan (1996) using quarterly data and find that the accrual anomaly is nonexistent for stocks in the top quintile of accruals’ contribution to future cash flow predictions This result supports our view that the current accruals’ ability to forecast future cash flows—rather than properties of current accruals per se, such as their sign and size—is the primary driver of accrual-based anomalies.3 Our contribution to the literature is twofold First, our study demonstrates that accruals’ contribution to future cash flow predictions is most significant when predicting future market capitalizations Assuming that market capitalization is a good proxy for all future cash flows, this implies that accruals contribute to the prediction of all future cash flows Many studies show that cash flow and accruals exhibit higher associations with future cash flows and/or stock returns than current cash flow alone (e.g., Barth, Cram, & Nelson, 2001; Dechow, 1994), but none provides such evidence in terms of out-of-sample predictions Second, our results add to the literature on accounting-based stock anomalies By documenting predictable abnormal returns based on hedge portfolios that use current accounting data as a sorting criterion, we show that market participants not fully understand the implications of current CFO and accruals for the present value of future cash flows In particular, the contribution of accruals to future cash flow predictions does not appear to be fully taken into account by investors, as accruals help improve upon CFO alone in earning abnormal returns over horizons of months and more Finally, we show that our contribution anomaly is related to the accrual anomaly documented by Sloan (1996) In addition, our methodological considerations have practical implications because they address issues of relevance to investors who use current accounting data for equity valuation purposes With respect to finite cash flow predictions, finite horizon predictions are of particular relevance to equity valuation techniques that consist of forecasting earnings, cash flows, or dividends over a finite period and computing a terminal value (Penman & Sougiannis, 1998) Our study is subject to caveats that apply to most studies in this field First, by using firm-specific regressions, we not only require time-series data that unavoidably reduce sample size but also introduce potential survivorship bias.4 Second, some accruals and deferrals are estimates subject to moral hazard between managers who report them and shareholders Our attempt to separate accruals based on their discretionary or unverifiable components using the Jones (1991) model is subject to the usual criticism regarding discretionary accruals estimation error The rest of the article is organized as follows: Section titled ‘‘Prior Literature and Empirical Predictions’’ reviews the relevant literature Section titled ‘‘Research Design’’ specifies the empirical tests, and the next section titled ‘‘Research Design’’ describes the sample selection process and presents the main results The final section titled ‘‘Conclusion’’ summarizes and concludes Prior Literature and Empirical Predictions Prior Literature Our article relates to an extensive literature that investigates the valuation implications of components of accounting earnings, either indirectly through their association with future Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 154 Journal of Accounting, Auditing & Finance accounting measures or directly through their association with market values of equity Wilson (1986), who uses stock returns around earnings announcements and Form 10-K filings to show that the accrual component of earnings has incremental information content over cash flow, is one of the earliest studies in this literature strand Because this question has subsequently generated a vast number of studies, which generally differ by methodological choices, we provide a matrix (see Appendix A and all other appendices online at http://jaaf.sagepub.com/supplemental) that highlights the key findings of prior research based on the three dimensions along which we position our study Determinants of Accruals’ Contribution The extent to which current accruals contribute to more accurate predictions of future cash flows is expected to vary across firms and time periods First, we expect the accruals’ contribution to vary with specifics of the economics of a firm, as manifested in properties of past or current accounting numbers For example, if firms operate in an uncertain environment, their stream of cash flows is more likely to exhibit greater volatility As a result, past realizations of cash flows are likely to be noisy and to be a less useful predictor of future cash flows Financial statement users are more likely to draw inferences about the timing and amount of future cash flows by using accruals Indeed, accruals tend to smooth out some of the variability in cash flow patterns by mitigating issues arising from discrepancies between cash flows and the underlying economics in terms of timing of recognition In addition, current cash flow in firms with greater growth options could be of relatively limited use in predicting future streams of cash flows Indeed, growth firms are more likely—ceteris paribus—to be in a transitory stage where past realizations of cash flows bear little association with future cash flows Although short-term accruals may also be uninformative, long-term accruals are likely to provide incremental information For instance, amortization policies for recent investments can provide useful insight about the economic life of the type of projects that the firm can undertake in the near future However, this benefit would not arise if a firm invests in R&D and other unrecognized internally developed intangibles At any rate, rather than the volatility of past cash flow, it is the expected volatility of future cash flows that provides a role for current accruals in terms of predictive ability for growth firms Our predictions so far rely on the assumption that management uses accruals in a manner that is not self-serving However, agency conflicts between managers and shareholders can induce management to deviate from truthful reporting to maximize their own wealth Indeed, prior studies have shown that executives report income-increasing discretionary accruals in years where they sell their stock so as to increase the proceeds from those transactions (Bartov & Mohanram, 2004; Cheng & Warfield, 2005) If reported accruals are distorted by measurement bias, their informativeness vis-a`-vis future cash flows may be impaired to a point where they no longer provide incremental prediction value or even worsen predictions compared with those based on current cash flows only Consistent with this idea, the results documented by Xie (2001) suggest that investors’ mispricing of accruals is driven by discretionary accruals However, managers can also report income-decreasing accruals for their own benefit In particular, when truthful reporting falls short of expectations by a large margin, they are better off taking a ‘‘big bath,’’ that is, reporting large income-decreasing accruals Hence, we expect that the magnitude of discretionary accruals (to the extent they are driven by measurement bias) Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Nam et al 155 exhibits a negative association with the contribution of total accruals to the prediction of future cash flows Accruals’ Contribution and Their Mispricing Starting with Sloan (1996), prior research has shown that investors not fully understand the implications of accruals for future earnings Sloan posits that investors fixate on earnings and fail to recognize the lower persistence of accruals compared with cash flows As a result, one can generate a profitable trading strategy by buying low-accrual stocks and selling high-accrual stocks We further explore the role of accruals in explaining future stock returns through their predictive ability for future cash flows To the extent that accruals contribute to more accurate predictions of future cash flows and market values of equity, and if investors fail to act such as to cause stock prices to fully reflect the predictive ability of accruals at the time accounting information becomes publicly available, then one may observe predictable stock returns subsequent to the release of that information For instance, the predicted market capitalization conditional on current accounting data can be viewed as a proxy for fundamental value, and if current market prices gravitate toward fundamental values, one can sort stocks based on the degree to which their prices deviate from fundamental value so as to predict stock returns (Frankel & Lee, 1998) The role of accruals in explaining such anomaly can be judged by comparing predicted values of future cash flows (proxies for fundamental value) with and without accruals as a predictor We expect that if, indeed, investors not fully understand the implications of accruals for future cash flows, then a trading strategy going long (short) on high (low) ratios of predicted to actual market values of equity should yield higher riskadjusted returns with accruals as a predictor than without accruals Finally, we investigate whether accruals’ predictive value is related to the accrual anomaly first documented by Sloan (1996) We posit that accounting-based anomalies should be primarily driven by investors’ incorrect expectations of future cash flows, rather than properties of current accounting data per se Hence, we expect that sorting stocks on accrual size need not be associated with predictable stock returns when current accruals are an accurate predictor of future cash flows.5 Research Design Prediction Models We use regression models to predict various measures of future cash flows out of sample In all models, we use the generic term Predictedt11 to designate the dependent variable, which can be either CFO or FCF,6 both measured over one to eight quarters ahead, or market value of equity (MKTCAP), at the beginning or at the end of the fiscal quarter, as a proxy for the present value of all future cash flows.7 All variables, whether they are being predicted or used as predictors, are scaled by total assets at the end of the previous fiscal quarter Our main analysis is based on firm-specific estimations using time-series data Our benchmark ‘‘cash-flow only’’ model is the following: CFOt11 5g0 1g1 CFOt 1e: ð1Þ Our accounting variables are subject to seasonality This is particularly the case for firm-level quarterly cash flows time series, which exhibit purely seasonal characteristics, as Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 156 Journal of Accounting, Auditing & Finance documented by Lorek and Willinger (1996) As we use adjacent quarters to make our predictions, we need to adjust for seasonality in our cash flow series To so, we use the X11 method as described in Appendix B In brief, the X11 procedure, developed by the Bureau of Census, decomposes monthly or quarterly data into trend, seasonal, and irregular components using moving averages One can subsequently subtract the estimated seasonal component to come up with a deseasonalized series To test whether accruals contribute to reducing prediction errors, we compare Model to models wherein aggregate accruals are included as an independent variable, either aggregated with cash flows or as a separate predictor: CFOt 5h0 1h1 CFOtÀ1 1h2 ACCtÀ1 1e: ð2Þ CFOt 5j0 1j1 EARNtÀ1 1e: ð3Þ ACC stands for total accruals, defined as the difference between net income before extraordinary items EARN (Compustat Quarterly Data Item 8) and CFO (Compustat Quarterly Data Item 108) net of extraordinary items/discontinued operations that affect cash flows (Compustat Quarterly Data Item 78).8 In Model 3, the coefficients on the cash flow and accrual components of earnings are equal, whereas they are allowed to differ in Model We include Model to assess whether aggregate earnings improve upon current cash flow alone in predicting cash flows We further proceed to disaggregate total accruals into their components, based on the premise that different subsets of accruals carry different implications for future cash flows (Barth et al., 2001), such as stemming from the horizon over which cash collectability is expected or from differing degrees of subjectivity inherent in different subsets of accruals: CFOt 5b0 1b1 CFOtÀ1 1b2 DARtÀ1 1b3 DINVtÀ1 1b4 DAPtÀ1 1b5 DEPAMORtÀ1 1b6 OTHERtÀ1 1e: ð4Þ Model is similar to the cross-sectional regression that Barth et al (2001) run to test the incremental explanatory power of disaggregated earnings This model presents the highest level of accrual disaggregation that we consider DAR, DINV, and DAP are changes in working capital accounts: accounts receivable, inventories, and accounts payable, respectively DEPAMOR is depreciation and amortization.OTHER is simply the difference between total accruals ACC and (DAR DINV DAP DEPAMOR When it is available, we use data from the statement of cash flow for our individual accrual components; otherwise, we use changes in balance sheet accounts That is, we use changes in accounts receivable, inventory, and accounts payable (Compustat Quarterly Data Items 103, 104, and 105, respectively) if they are available; otherwise, we use changes in Data Items 37, 38, and 46 from the previous fiscal quarter Depreciation and amortization expense is Compustat Quarterly Data Item 77 Market capitalization is the product of Compustat Quarterly Data Items 14 and 61 Finally, our deflator is total assets (Compustat Quarterly Data Item 44) as of the beginning of the quarter One major distinction among accrual components is the timing of their conversion into cash in- or outflows The changes in working capital variables are expected to affect future cash flows in the near term (within a year) By contrast, DEPAMOR should exhibit a greater association with cash flows in the longer run Indeed, depreciation and amortization expenses are intended to match costs of Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Nam et al 157 investments with their benefits over the expected life of the asset that is being depreciated/ amortized, typically several years Overall, although the use of individual accrual components may help improve prediction accuracy, the decrease in the number of degrees of freedom may offset such a benefit for firm-specific estimations, for which the number of observations is limited Each firm-specific model is estimated using 56 consecutive quarterly observations.9 We use rolling windows so that coefficients are ‘‘updated’’ every quarter The required number of observations represents a trade-off between sample size and the reliability and stability of time-series estimates Alternatively, we estimate coefficients cross-sectionally, separately for each fiscal quarter Once we run a regression, we use the coefficient estimates to comd CFOt t11 pute predicted values For example, based on Model 1, CFO g0 1c g1 Assets , Assetst is equal to c tÀ1 where c g0 ; c g1 are estimated from the regression The predicted value is then compared with the actual value We compute our absolute prediction errors as follows: ABSEj Huang 273 DVt11 1Dt11 DBt11 Bt D 5v ½ Š1v0 ½ Dqt11 Š1 t11 : Vt Vt Vt Vt DBt11 B D t t11 1v0 ½ Dqt11 Š1 : Bt Vt Vt Rt11 ð8Þ Using the clean surplus relation, we replace Dt11 in Equation (8) with Xt11 and DBt11 and rearrange to obtain the following expression for the period t 1 return: Rt11 5½ Xt11 Bt Vt DBt11 Š1v0 ½ Dqt11 Š1½( À 1) Š: Vt Vt Bt Vt ð9Þ Equation (9) shows that the return in period t 1 is a function of the profitability change, Dqt11, and contemporaneous capital investment, DBt11, in addition to earnings, Xt11 Changes in profitability affect returns because they revise expectations about a firm’s ability to generate value from invested capital Investment results in a change in the capital base used to generate value, and so it also affects returns Both variables revise expectations about future cash flows In this model, value generation hinges on two basic attributes of operations as viewed from equity holders’ standpoint: the amount of capital invested (equity book value) and the efficiency in utilizing capital to generate profit (profitability) Furthermore, as a firm’s operations move forward, the scale of operation is adjusted in accordance with changes in profitability, thus, giving rise to real options With equity value depending on equity book value and profitability, returns as changes in equity value naturally depend on contemporaneous equity investment, DBt11, and changes in profitability, Dqt11, both of which require balance-sheet data Although the profitability variable is constructed jointly with balancesheet and income-statement data, we classify it as a balance-sheet variable because when we have already controlled for earnings and the earnings change, any IEP of profitability changes in explaining return comes from balance-sheet information The two balance sheet–related factors, DBt11 and Dqt11, are linked to real options through their coefficients as in Equation (9) The coefficient on Dqt11 contains v0 5dv=dqt P (qt )11=r1gC (qt ), which is positive and increasing in qt, given that v itself is increasing and convex in qt In our empirical analysis below, we exploit this nonlinearity feature caused by real options to allow the coefficient on Dqt11 to vary with the level of profitability The coefficient on DBt11 is (Vt / Bt 1) = P(qt )1qt =r1gC(qt )21, representing the net present value per unit of incremental investment, which incorporates the effect of real options Empirically, this coefficient can be either positive or negative, depending on whether the additional investment is profitable, that is, whether P(qt )1qt =r1gC(qt )21 0.13 To the extent that firms on average make profitable (positive NPV) investments, we expect contemporaneous capital investment (DBt11) to have a positive coefficient in the return model Thus, the real options–based model of Zhang (2000) provides an economically meaningful interpretation of the coefficient of DBt11 Finally, consistent with Equation (5), the coefficient on Xt11 is predicted to be one Empirical Research Design and Sample Research Design From Ohlson’s (1995) model, we show that the previous period’s capital investment, along with the earnings level and the earnings change, explains returns Moreover, from Zhang’s Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 274 Journal of Accounting, Auditing & Finance (2000) model, we derive two additional balance sheet–related variables, contemporaneous capital investment and the profitability change, together with earnings The two models are developed from two distinct economic settings, which are complementary in certain aspects Although Zhang considers contingent capital investment decisions, which give rise to real options, his model is embedded with an assumption that a firm’s scale of operations is kept constant in the period preceding the date of valuation (so that lagged capital investment is zero; Zhang, 2000); however, this condition is not imposed in Ohlson’s linear model Due to the complementarity between the two models, the factors from one model will not completely subsume those from the other in explaining returns Thus, in our empirical analysis, we combine the factors from Ohlson and Zhang, and use them to jointly explain returns The main return model for our empirical analysis is the following linear specification: Rit 5a1b xit 1g Dxit 1h Dqit 1u Dbit 1d DbitÀ1 1eit : ð10Þ In Equation (10), Dxit and Dbit21 arise from Ohlson (1995), Dqit and Dbit from Zhang (2000), and xit from both settings Although it is true that Dqit already contains the information in Dxit and Dbit21 (which is combined in a particular fashion), the three factors originate from two distinct economic settings (explained above), and empirically, whether Dqit is sufficient for summarizing the information in Dxit and Dbit-1 to explain returns is unclear For this reason, we keep all three factors (Dqit, Dxit, and Dbit21) in return Model (10) Model (10) also enables us to conveniently evaluate the incremental usefulness of balancesheet information beyond an earnings only–based model (that uses only earnings and the earnings change) The dependent variable in (10), Rit, is the annual stock return, which is calculated from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end The independent variables are specified as follows: xit = Xit / Vit21 is the earnings in year t (Xit) scaled by the market value of equity at the beginning of year t (Vit21); Dxit = (Xit Xit21) / Vit21 is the earnings change in year t relative to year t scaled by Vit21; Dqit = qit qit21 is the profitability change in year t relative to year t 1, with qit = Xit / Bit21; Dbit = (Bit Bit21) / Vit21 is capital investment (the change in equity book value) in year t scaled by Vit21; and Dbit21 = (Bit21 Bit22) / Vit21 is the lagged capital investment (change in equity book value in year t 1) scaled by Vit21 According to Model (9), the coefficient on Dqit involves the first-order derivative of the growth option (included in v’) Due to the convex behavior of real options, this coefficient is an increasing function of profitability To capture this property, we distinguish the coefficient on Dqit between high- and low-profitability firms in the extended specification below, Rit 5a1b xit 1g Dxit 1h Dqit 1hH HDqit 1u Dbit 1d DbitÀ1 1eit ; ð10aÞ where H is a dummy variable equal to for firms with profitability above the median in a year and otherwise Based on the above theoretical analysis, we expect b = 1, g 0, h 0, u 0, and d \ in Model (10a) In addition, following the prediction that the return impact associated with one unit of profitability change is greater for more profitable firms, we expect hH To evaluate the incremental usefulness of the balance-sheet information as incorporated in (10a), we compare the performance of (10a) with that of the following benchmark Model (11) (Easton & Harris, 1991), which relies only on earnings variables: Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 275 Rit 5a1b xit 1g Dxit 1eit : ð11Þ The factors in Model (11) are a subset of those in Model (10a) Observe that (11) can be viewed as a ‘‘cut-down’’ version of the return Model (5), which is derived from Ohlson (1995), without the lagged capital investment term We assess the incremental usefulness of balance-sheet information in two aspects First, for individual balance sheet–related variables, we test whether the coefficients are consistent with the theoretical predictions We examine the significance of these variables within our (more comprehensive) return Models (10) and (10a), controlling for the earnings variables Second, we examine whether there is a significant improvement in model performance after introducing our balance sheet–related variables, as measured by the IEP, which is calculated as the R2 of Model (10a) minus that of Model (11) (Biddle, Seow, & Siegel, 1995; Brown, Lo, & Lys, 1999).14 That is, we attribute the IEP of Dbit21, Dbit, and Dqit, beyond Dxit and xit, to balance-sheet information As already explained above, although Dqit combines both earnings and book value, its IEP over and above that of the earnings variables is attributable to balance-sheet information The difference in the explanatory power between (10a) and (11) represents the IEP of the three balance-sheet variables as a group In addition, we also estimate the IEPs of the balance-sheet variables individually This is computed as the R2 of Model (10a) minus that of Model (10a), excluding the concerned balance-sheet variable Beyond examining the IEP on broad cross-sectional samples, we evaluate whether the IEP varies over time and across subsets of firms We describe how the IEP of balancesheet information fluctuates from year to year in relation to the explanatory power of earnings numbers We also compare the IEPs of firms that differ in earnings informativeness (negative vs positive earnings firms and young vs mature firms) or in the predictability of future earnings (firms with low vs high analyst forecast accuracies and firms with large vs small forecast dispersions) The Sample and Descriptive Statistics We extract the data on earnings before extraordinary items and discontinued operations (Xit, No 18) and equity book value (Bit, No 60) from the Compustat annual file We extract the stock returns and beginning market values of common equity from the CRSP monthly files Annual returns with dividends (Rit) are compounded from monthly returns starting from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end We exclude observations with an equity book value less than US$0.5 millions or total assets less than US$1.5 millions To reduce the impact of outliers and extremely illiquid stocks, we require the stock price at the beginning of a fiscal year to be higher than US$3 We exclude firms in financial industries (whose balance sheets have distinctively different features) and utility firms (whose profitability is subject to regulations) The resulting sample consists of 87,439 firm-year observations for the period 1968 to 2007 In some parts of the analysis, the sample size varies where we also use analyst earnings forecasts from the Institutional Brokers’ Estimate System (I/B/E/S) detailed file.15 We winsorize the continuous variables at the top and bottom 1% of the distribution Panel A of Table presents the descriptive statistics of the main variables for the pooled sample The annual stock return (Rit) has a mean (median) of 0.17 (0.09) The scaled earnings (xit) have a mean (median) of 0.07 (0.07), and the scaled earnings change (Dxit) has a mean and median of 0.01 The profitability change (Dqit) has a mean of 20.01 Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 276 Journal of Accounting, Auditing & Finance Table Summary Statistics Panel A: Descriptive Statistics Rit xit Dxit Dbit Dqit Dbit21 M SD 1% 25% 50% 75% 99% 0.17 0.07 0.01 0.06 20.01 0.06 0.46 0.07 0.09 0.13 0.15 0.12 20.57 20.10 20.33 20.41 20.49 20.41 20.13 0.04 20.02 0.02 20.06 0.02 0.09 0.07 0.01 0.06 0.00 0.05 0.36 0.11 0.03 0.10 0.04 0.10 1.82 0.21 0.33 0.53 0.43 0.44 Panel B: Correlation Matrix Rit Rit xit Dxit Dbit Dqit Dbit21 xit 0.34 0.28 0.27 0.21 0.27 20.05 0.52 0.56 0.39 0.20 Dxit Dbit 0.35 0.50 0.25 0.60 0.48 0.46 0.72 20.30 0.30 0.17 Dqit 0.32 0.37 0.84 0.32 Dbit-1 20.04 0.27 20.18 0.27 20.38 20.37 Note: This table reports the summary statistics for our sample for the period 1968 to 2007 There are 87,439 firm-year observations The variables are defined as follow: Rit is the annual stock return from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end; xit = Xit / Vit21 is the earnings in year t (Xit) scaled by the beginning market value of equity (Vit21); Dxit = (Xit Xit21) / Vit21 is the earnings change in year t relative to year t scaled by Vit21; Dqit = (qit qit21) is the profitability change of year t relative to year t 1, with qit = Xit / Bit21; Dbit21 = (Bit21 2Bit22) / Vit21 is the capital investment in year t scaled by Vit21; and Dbit = (Bit Bit21) / Vit-1 is the current year’s capital investment scaled by Vit21.In Panel B, the Spearman correlation coefficients are above the diagonal, and Pearson correlation coefficients are below the diagonal All of the coefficients are significant at the 01 level and median around 0, suggesting that the profitability tends to decline over time Scaled contemporaneous capital investment (Dbit) has a mean (median) of 0.06 (0.06), and scaled lagged capital investment (Dbit21) has a mean (median) of 0.06 (0.05) Panel B reports the pairwise correlations among the variables We find that all of the correlations in the panel are significant at the 01 level The annual stock return is positively correlated with earnings (with a Pearson correlation equal to 28 and a Spearman correlation equal to 34) and with the earnings change (.27 and 35) More importantly, the return is also positively correlated with the profitability change (.27 and 32), contemporaneous capital investment (.21 and 25), and is negatively correlated with lagged capital investment (2.05 and 2.04), all having the predicted signs We find strong correlations among the accounting variables The earnings level is positively correlated with the earnings change, which is expected (shocks causing earnings to increase in year t tend to cause earnings to be higher than in year t 1) Current capital investment is positively correlated with earnings, earnings change, and profitability change, and is grossly consistent with the notion of ‘‘capital following profitability’’ (Biddle, Chen, & Zhang, 2001) Current capital investment is also positively correlated with lagged capital investment, suggesting that in general firms not alter their investment activities drastically from one year to the next The correlation between the earnings change and the profitability change is particularly high (.72 and 84) Earnings are a product of equity book value and profitability In Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 277 principle, the earnings change in a year relative to the prior year can arise from two sources: a change in book value (due to incremental investment or divestment) and a change in profitability (due to improvement or deterioration in operational efficiency) The high correlation between the earnings change and the profitability change suggests that, empirically, the earnings change between two consecutive years is mostly driven by the profitability change, rather than capital investment Results From Broad-Based Samples This section examines the empirical importance of balance-sheet information in broad cross-sectional samples The objective here is to gain an overall view of how useful the balance-sheet variables are beyond earnings variables in explaining stock return Results From the Pooled Sample Using pooled samples, Table reports the performance of our return Models (10) and (10a), relative to the performance of several variants of these two models and earningsonly Model (11) In running these pooled regressions, we adjust for cross and serial correlations with two-way (firm and year) clustering Panel A shows the regression results of Models (10), (10a), and (11) Controlling for earnings and the earnings change, the three balance sheet–related variables, Dqit, Dbit, and Dbit-1, all have a significant effect on returns and the directions of the effects are as predicted The profitability change (Dqit) has a consistently positive coefficient in all the specifications in the table In row (ii), without introducing the piecewise linear structure, the coefficient on Dqit is 0.37 (t = 13.31), significant at the 01 level For the piecewise linear model (row i), the coefficient on Dqit is 0.15 (t = 4.02) for the low-profitability range, significant at the 01 level, and increases to 0.53 (=0.15 0.38) for the high-profitability range The coefficient increase from the low-profitability to the high-profitability range is significant at the 01 level (t = 7.19) These results indicate that the effect of a change in profitability on returns is positive and is greater for firms with higher profitability The magnitude of the coefficient demonstrates that the effect of the profitability change is economically important Ceteris paribus, an increase in profitability by one standard deviation within the pooled sample (=0.15) is, on average, associated with a return increase of 0.02 for low-profitability firms and 0.08 for high-profitability firms, which amounts to 13.2% and 46.8% of the average annual return (0.17), respectively In row (i), lagged capital investment (Dbit21) has a negative coefficient of 20.13 (t = 2.56), significant at the 05 level, consistent with the prediction This suggests that a change of lagged capital investment by one standard deviation (0.12) is associated with an average return change of 20.02 Contemporaneous capital investment (Dbit) has a coefficient of 0.28 (row i), significant at the 01 level (t = 4.89) In absolute terms, the coefficient on Dbit is almost twice of that on Dbit21 and has a much higher t value An increase in capital investment by one standard deviation (0.13) is associated with an average return increase of 0.04, which is economically significant The coefficient on xit is highly significant; the coefficient is 1.02 (t = 3.56) in Model (10a), row (i), and is 0.97 (t = 3.40) in Model (10), row (ii) These values are not significantly different from the theoretical value of one at the level The coefficient on Dxit is sensitive to whether balance-sheet information is present In the benchmark Model (11), row (iii), where only the two earnings variables are used, this Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 278 Journal of Accounting, Auditing & Finance Table Pooled Sample Regressions Intercept xit Dxit Dqit Dbit Dbit21 H Dqit R2 Panel A: Comparison of our return models with earnings-only-based return models (i) Model (10a) (ii) Model (10) (iii) Model (11) 0.09a (3.23) 0.10a (3.55) 0.10a (3.27) 1.02a,d (3.56) 0.97a,d (3.40) 1.11a,d (4.27) 0.30b (2.56) 0.19c (1.73) 0.84a (7.08) 0.15a (4.02) 0.37a (13.31) 0.28a (4.89) 0.30a (5.15) 20.13b (22.56) 20.14a (22.69) 0.38a (7.19) 090*** 087*** 076*** Panel B: The effect of individual balance-sheet variables (iv) (v) (vi) 0.09a (3.22) 0.10a (3.26) 0.10a (3.30) 1.14a,d (4.42) 0.91a,d (3.19) 1.25a,d (4.65) 0.50a (4.02) 0.75a (6.95) 0.72a (5.35) 0.14a (3.41) 0.42a (7.51) 0.25a (4.08) 087*** 079*** 20.17a (23.04) 077*** 20.24a (24.59) 20.07 (21.26) 081*** Panel C: IEP of individual balance-sheet variables (vii) (viii) (ix) 0.10a (3.30) 0.09a (3.21) 0.09a (3.23) 1.06a,d (3.69) 1.20a,d (4.48) 0.93a,d (3.30) 0.56a (4.71) 0.46a (3.46) 0.38a (3.43) 0.30a (5.09) 0.13a (3.40) 0.17a (4.39) 0.26a (4.32) 0.41a (7.44) 0.39a (7.36) 087*** 090*** Note: IEP = incremental explanatory power This table reports the pooled regression results for Models (10), (10a), and (11): Rit 5a1b xit 1g Dxit 1h Dqit 1u Dbit 1d DbitÀ1 1eit ; ð10Þ Rit 5a1b xit 1g Dxit 1h Dqit 1hH HDqit 1u Dbit 1d DbitÀ1 1eit ; ð10aÞ Rit 5a1b xit 1g Dxit 1eit : ð11Þ Rit is the annual stock return from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end; xit = Xit / Vit21 is the earnings in year t (Xit) scaled by the beginning market value of equity (Vit21); Dxit = (Xit Xit21) / Vit21 is the earnings change in year t relative to year t scaled by Vit21; Dqit = (qit qit-1) is the profitability change of year t relative to year t 1, with qit = Xit / Bit21; Dbit21 = (Bit-1 Bit22) / Vit-1 is the capital investment in year t scaled by Vit21; and Dbit = (Bit Bit-1) / Vit21 is the current year’s capital investment scaled by Vit21 H is a dummy variable equal to for firms whose profitability is larger than the annual median level The t statistics in the parentheses are adjusted for firm-year two-way clustering a,b , and c indicate the coefficient being significantly different from at the 01, 05, and levels, respectively d indicates the coefficient is not significantly different from the predicted value of one at the level ***indicates the Vuong’s Z statistics for comparing balance-sheet-information-integrated model (Model [10a] and its variants) with earnings-only model (Model [11]) being significant at the 01 level The Z statistics are 26.36 (row [i]), 23.35 (row [(ii]), 26.77 (row [iv]), 11.47 (row [v]), 7.70 (row [vi]), 15.60 (row [vii]), 22.98 (row [viii]), and 25.70 (row [ix]), respectively, in favor of balance-sheet-information-integrated models coefficient is 0.84 (t = 7.08), but it decreases by more than half to 0.30 (t = 2.56) in the comprehensive Model (10a), row (i), which incorporates the balance sheet–related variables The explanatory power of Model (10a) is 9%, compared with that of 7.6% for benchmark Model (11) (row [i] vs row [iii]) Thus, collectively, the balance-sheet variables have Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 279 an IEP of 1.4% Vuong’s Z statistics for comparing Models (10a) and (11) is 26.36, significant at the 01 level, in favor of Model (10a) Similarly, Vuong’s Z statistics for comparing Models (10) and (11) is 23.35, also significant at the 01 level, in favor of 10 over 11 Panel B details how the individual balance-sheet variables impact the role of Dxit in the return model Inclusion of Dqit has the greatest impact, which reduces the coefficient on Dxit from 0.84 in row (iii) to 0.50 in row (iv), whereas the inclusion of Dbit and Dbit21, row (v) and (vi), has a smaller impact The results demonstrate that Dqit has a more robust relationship with stock returns than does Dxit, reaffirming the usefulness of balance-sheet information for enhancing the performance of return models The IEP of individual balance-sheet variables is provided in Panel C Among the individual factors, the profitability change has the largest IEP of 9% The IEP of contemporaneous capital investment is 3% The IEP of lagged capital investment is the smallest at 01% In Panels B and C, we conduct Vuong’s tests to examine the performance of various balance-sheet-information-integrated models relative to the earnings-only model The results show that the models incorporating various subsets of our balance-sheet variables all perform significantly better at the 01 level than the earnings-only model Although the results in Table are from regressions using raw returns as the dependent variable (which is originally derived from the underlying valuation models), we also perform regressions using market-adjusted returns as the dependent variable, which aim to mitigate potential concerns caused by the differences in the general level of returns across years The results, presented in Table 3, are similar to those reported in Table Therefore, our conclusion about the usefulness of the balance-sheet variables that we have identified (Dqit, Dbit, and Dbit21), both individually and as a whole, remains unchanged.16 Results From the Annual Samples Panel A of Table presents the results of Model (10a) from the annual samples The top part of the panel shows the mean coefficients from the annual regressions across the sample years and the Fama–MacBeth t values adjusted with Newey–West approach The average coefficient on xit is 1.26 (t = 7.76), which is not significantly different from the theoretical value of one at the level In annual regressions, the coefficient on xit is significantly different from for 27 years at the 0.1 level or better and is not significantly different from for 13 years The average coefficient on Dqit is 0.21 (t = 4.61) for low-profitability firms, and the incremental coefficient on Dqit for high-profitability firms is 0.43 (t = 7.68), showing a relationship between returns and Dqit that is dependent on the level of profitability The coefficient on Dqit is significantly positive for low-profitability firms in 21 of the 40 sample years and the incremental coefficient for high-profitability firms is significantly positive in 30 years at the level or better, and generally insignificant for the remaining years, conditional on the earnings variables The average coefficient on Dbit is 0.16 (t = 3.78) and that on Dbit21 is 20.09 (t = 22.54) The coefficient on contemporaneous capital investment is significantly positive in 17 years (at the 0.1 level or better), insignificant in 21 years, and significantly negative in years The coefficient on Dbit21 is significantly negative (at the level or better) in 17 years, insignificant in 16 years, and significantly positive in years In Panel B of Table 4, we compare the annual R2s of Model (10a) with those of Model (11) In all the 40 years in our tests, introducing the balance-sheet information significantly improves the explanatory power of the return model (at the level or better), as indicated Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 280 Journal of Accounting, Auditing & Finance Table Pooled Sample Regressions With Market-Adjusted Returns Intercept xit Dxit Dqit Dbit Dbit21 H Dqit R2 Panel A: Comparison of our return models with earnings-only-based return models (i) Model (10a) (ii) Model (10) (iii) Model (11) 20.04 (21.57) 20.02 (21.09) 20.032 (21.42) 1.06a,d (5.40) 1.02a,d (5.16) 1.10a,d (6.40) 0.39a (4.33) 0.28a (3.27) 0.93a (11.41) 0.15a (5.08) 0.37a (15.08) 0.25a (5.00) 0.27a (5.33) 20.16a (23.56) 20.16a (23.65) 0.37a (8.04) 109*** 106*** 094*** Panel B: The effect of individual balance-sheet variables (iv) (v) (vi) 20.03 (21.57) 20.03 (21.46) 20.03 (21.39) 1.13a,d (6.63) 0.93a,d (4.71) 1.27a,d (7.13) 0.58a (6.49) 0.86a (11.88) 0.79a (8.01) 0.16a (4.88) 0.40a (8.44) 0.21a (4.11) 106*** 096*** 20.20a (24.48) 096*** 20.26a (26.10) 20.10b (22.20) 099*** Panel C: IEP of individual balance-sheet variables (vii) (viii) (ix) 20.03 (21.43) 20.04 (21.57) 20.04 (21.56) 1.10a,d (5.57) 1.22a,d (6.90) 0.95a,d (4.87) 0.65a (7.52) 0.53a (5.32) 0.48a (5.99) 0.26a (5.28) 0.13a (4.45) 0.18a (5.87) 0.22a (4.32) 0.40a (8.31) 0.38a (8.29) 106*** 108*** Note: IEP = Incremental explanatory power This table reports the pooled regression results for models (10), (10a), and (11): ExRit 5a1b xit 1g Dxit 1h Dqit 1u Dbit 1d DbitÀ1 1eit ; ð10Þ ExRit 5a1b xit 1g Dxit 1h Dqit 1hH HDqit 1u Dbit 1d DbitÀ1 1eit ; ð10aÞ ExRit 5a1b xit 1g Dxit 1eit : ð11Þ ExRit is the annual market-adjusted stock return from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end; xit = Xit / Vit21 is the earnings in year t (Xit) scaled by the beginning market value of equity (Vit21); Dxit = (Xit Xit21) / Vit21 is the earnings change in year t relative to year t scaled by Vit21; Dqit = (qit2qit21) is the profitability change of year t relative to year t 1, with qit = Xit / Bit21; Dbit21 = (Bit21 Bit22) / Vit21 is the capital investment in year t scaled by Vit21; and Dbit = (Bit Bit21) / Vit21 is the current year’s capital investment scaled by Vit21 H is a dummy variable equal to for firms whose profitability is larger than the annual median level The t statistics in the parentheses are adjusted for firm-year two-way clustering ***indicates the Vuong’s Z statistics for comparing the balance-sheet information integrated model (Model [10a] and its variants) with earnings-only model (Model [11]) being significant at the 01 level The Z statistics for these models are 27.63 (row [i]), 24.57 (row [ii]), 24.30 (row [iv]), 10.30 (row [v]), 9.80 (row [vi]), 16.16 (row [vii]), 24.78 (row [viii]), and 26.60 (row [ix]), respectively, in favor of balance-sheet-information-integrated models a,b and c indicate the coefficient being significantly different from at the 01, 05, and levels, respectively d indicates the coefficient is not significantly different from the predicted value of one at the 01 level by the significant Vuong’s Z statistics The IEP of the balance-sheet information ranges from 6% (year 1984) to 8% (year 1969), with a mean of 2.4% (t = 7.31) Moreover, the average annual R2 is 13.9% for Model (10a), compared with that of 11.5% for Model (11) Following Ball, Kothari, and Robin (2000), we perform a t test and find that the time series Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 281 Table Annual Regression Results and the IEP of Balance-Sheet Information Year Intercept xit Dxit Dqit Dbit Dbit21 H*Dqit R2 Mean 0.07 (2.66) 0.07a (2.86) 20.27a (217.31) 0.01 (0.57) 0.04b (2.53) 20.18a (211.89) 20.27a (220.79) 20.21a (211.08) 0.23a (9.71) 20.07a (23.71) 0.02 (1.07) 0.16a (8.02) 0.15a (6.77) 0.47a (18.35) 20.15a (210.05) 0.38a (16.61) 0.08a (5.12) 20.10a (29.32) 0.23a (16.99) 0.11a (10.09) 20.06a (26.25) 0.04a (3.96) 0.05a (4.35) 0.01 (0.75) 0.23a 1.26a (7.76) 2.45a,*** (5.72) 1.59a,** (5.60) 3.30a,*** (11.62) 1.99a,*** (7.86) 1.08a (4.85) 0.94a (7.03) 1.23a (8.18) 1.44a,** (7.03) 1.90a,*** (11.16) 1.18a (6.91) 0.28*** (1.56) 20.47b,*** (22.46) 20.36c,*** (21.65) 1.70a,*** (11.35) 1.33a (6.03) 3.21a,*** (15.06) 2.15a,*** (16.95) 1.56a,*** (8.66) 2.39a,*** (13.04) 1.28a,* (7.88) 1.61a,*** (11.16) 0.81a (4.78) 0.90a (4.55) 1.11a 0.41a (4.71) 0.91c (1.76) 20.44 (21.41) 20.78a (23.03) 0.61b (2.26) 0.06 (0.24) 0.59a (3.36) 0.66a (4.37) 0.43b (2.49) 0.65a (4.14) 1.29a (8.25) 1.03a (5.74) 1.80a (9.79) 1.42a (7.18) 20.09 (20.62) 0.10 (0.53) 0.17 (0.91) 20.10 (20.82) 20.10 (20.67) 20.14 (20.91) 0.31b (2.20) 0.15 (1.08) 0.44a (2.72) 0.43b (2.23) 0.21 0.21a (4.61) 0.40b (2.49) 20.24b (22.30) 20.55a (24.21) 0.09 (0.81) 20.09 (20.83) 0.06 (0.67) 20.14 (21.49) 0.20c (1.67) 20.11 (21.16) 20.03 (20.34) 0.08 (0.92) 0.13 (1.37) 0.57a (5.39) 0.05 (0.69) 0.11 (0.96) 20.08 (20.84) 20.09 (21.53) 0.22a (2.68) 20.10 (21.26) 0.01 (0.16) 0.03 (0.50) 0.26a (3.28) 0.20b (2.20) 0.36a 0.16a (3.78) 0.03 (0.13) 0.94a (6.66) 0.90a (4.05) 0.14 (0.62) 0.44b (2.11) 0.42b (2.26) 0.23 (1.13) 0.37c (1.89) 20.20 (21.25) 20.17 (21.00) 0.31 (1.60) 20.12 (20.77) 0.25 (1.32) 0.44a (3.80) 0.58a (3.90) 20.35a (22.28) 0.25a (2.85) 0.47a (4.37) 0.30a (2.88) 0.12 (1.38) 0.22b (2.26) 0.36a (3.11) 0.33a (2.59) 0.18 20.09b (22.54) 20.35 (21.49) 20.35a (23.08) 20.16 (21.39) 0.26b (2.26) 20.12 (21.20) 20.15b (22.11) 20.15c (21.69) 0.06 (0.51) 0.09 (0.88) 0.25a (2.64) 0.40a (3.90) 0.18 (1.62) 0.35a (2.76) 20.10 (21.06) 20.08 (20.78) 20.23c (21.93) 20.06 (20.89) 20.22 (22.66) 20.20b (22.42) 20.13c (21.94) 20.05 (20.79) 0.19b (2.44) 0.05 (0.48) 20.07 0.43a (7.68) 1.07a (3.98) 20.35b (22.27) 0.54c (2.03) 0.46b (2.05) 1.01a (4.62) 0.69a (4.03) 0.74a (4.07) 0.37a (1.88) 0.65a (4.28) 1.06a (6.52) 1.06a (5.75) 0.75a (4.54) 0.54a (2.79) 0.15 (1.16) 0.58a (3.20) 0.61a (3.76) 20.09 (20.95) 0.15 (1.23) 0.16 (1.48) 0.28a (3.01) 0.23b (2.34) 0.26b (2.21) 0.39a (2.81) 0.81a 139 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 167 135 162 149 098 224 211 182 181 238 161 139 156 160 118 144 181 150 166 122 171 152 117 124 (continued) Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 282 Journal of Accounting, Auditing & Finance Table (continued) Year 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 Intercept xit Dxit (14.48) 0.07a (6.12) 0.15a (13.05) 0.06a (5.58) 0.30a (21.61) 0.03a (3.00) 0.28a (22.35) 20.07a (26.21) 0.28a (15.92) 0.10a (6.11) 0.12a (7.97) 20.13a (211.80) 0.44a (29.43) 0.06a (5.00) 0.24a (16.80) 0.08a (7.41) 20.05a (23.56) (4.94) 1.52a,*** (7.89) 1.02a (5.82) 0.79a (4.78) 20.22*** (21.11) 1.96a,*** (11.41) 1.43a,** (7.79) 0.79a (4.27) 22.01a,*** (28.04) 0.45b,** (2.08) 2.88a,*** (13.66) 1.96a,*** (10.56) 20.04*** (20.18) 2.22a,*** (10.72) 0.51b,** (2.05) 1.07a (5.66) 1.31a (5.39) (1.04) 0.55a (3.49) 0.55a (3.67) 0.29c (1.97) 0.50a (2.74) 20.21 (21.31) 0.40b (2.41) 0.09 (0.49) 0.51b (2.11) 0.23 (1.02) 20.35b (21.82) 20.08 (20.49) 1.34a (6.28) 0.46b (2.24) 1.21a (4.62) 0.09 (0.48) 1.03a (4.35) Dqit (3.64) 0.08 (1.11) 0.11 (1.47) 0.26a (3.57) 0.64a (7.42) 0.28a (4.08) 0.40a (5.34) 0.10 (1.24) 0.89a (7.70) 0.81a (7.76) 0.13 (1.24) 0.09 (1.17) 0.41a (3.79) 0.12 (1.32) 0.27b (2.51) 0.34a (3.73) 0.22b (2.10) Dbit Dbit21 H*Dqit (1.32) 0.12 (1.07) 0.18c (1.95) 0.18b (2.10) 0.28a (2.78) 0.19a (2.67) 0.08 (1.06) 0.25a (3.13) 0.24c (1.68) 0.41a (3.10) 0.06 (0.64) 0.09 (0.79) 20.49a (22.64) 20.06 (20.47) 0.00 (20.01) 0.42a (3.83) 20.14 (21.07) (20.66) 20.10 (21.17) 0.17b (2.33) 20.15b (22.19) 20.14 (21.45) 20.32a (23.59) 0.02 (0.29) 20.28a (23.36) 20.31a (22.68) 20.39a (23.39) 20.57a (25.89) 20.37a (23.99) 20.25a (22.58) 20.40a (23.75) 0.19 (1.53) 20.32a (23.37) 0.15 (1.31) (5.32) 0.42a (3.68) 0.35a (3.29) 0.15c (1.69) 0.26c (2.32) 0.10 (1.11) 0.35a (3.61) 0.20b (2.25) 0.46a (3.02) 20.11 (20.76) 0.82a (6.37) 0.13 (1.14) 1.14a (5.86) 0.18 (1.50) 0.53a (3.55) 20.13 (21.20) 0.43a (3.24) R2 149 113 088 080 109 112 063 073 091 194 113 095 140 100 121 126 Panel A reports the regression results of Model (10a) from the annual samples: Rit 5a1b xit 1g Dxit 1h Dqit 1hH H Dqit 1u Dbit 1d DbitÀ1 1eit , where Rit is the annual stock return from the 4th month after the prior fiscal year end to the 3rd month after the current fiscal year end; xit = Xit / Vit21 is the earnings in year t (Xit) scaled by the beginning market value of equity (Vit21); Dxit = (Xit Xit21) / Vit-1 is the earnings change in year t relative to year t scaled by Vit21; Dqit = (qit qit21) is the profitability change of year t relative to year t 1, with qit = Xit / Bit21; Dbit21 = (Bit21 Bit22) / Vit21 is the capital investment in year t scaled by Vit21; and Dbit = (Bit2Bit-1) / Vit21 is the current year’s capital investment scaled by Vit21 H is a dummy variable equal to for firms whose profitability is larger than the annual median level In the row of ‘‘mean’’, the t statistics in parentheses are computed with Fama–MacBeth methodology and adjusted for heteroscedasticity and autocorrelation of six lags with Newey–West approach a b , , and c indicate the coefficient being significantly different from at the 01, 05, and levels, respectively ***, **, and * indicate the coefficient being significantly different from at 01, 05, and levels, respectively (continued) Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 283 Panel B: This panel reports the R2s of Models (10a) and (11) and the incremental explanatory power of balance-sheet information for annual samples Year Mean 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 R2 of Model (11) R2 of Model (10a) Difference in R2 (1) (2) (2)2(1) Vuong’s test 115 126 055 104 140 055 174 182 173 172 215 123 127 126 143 091 136 174 123 147 107 157 131 095 090 130 096 073 049 094 095 042 040 059 151 102 063 132 085 104 116 139 167 135 162 149 098 224 211 182 181 238 161 139 156 160 118 144 181 150 166 122 171 152 117 124 149 113 088 080 109 112 063 073 091 194 113 095 140 100 121 126 024 041 080 057 009 044 050 029 009 009 023 038 012 030 017 027 007 006 027 019 015 014 021 023 034 019 016 015 031 015 017 021 033 032 043 011 032 008 015 018 010 7.31a 4.21*** 7.60*** 6.79*** 2.46** 5.28*** 7.33*** 5.27*** 3.05*** 1.63* 3.67*** 6.03*** 2.49** 6.10*** 4.81*** 5.38*** 1.86* 2.94*** 6.05*** 4.99*** 3.79*** 4.01*** 4.88*** 4.62*** 5.05*** 4.26*** 4.21*** 4.64*** 6.82*** 4.99*** 4.88*** 5.61*** 6.72*** 6.29*** 5.85*** 3.35*** 4.10*** 2.95*** 3.11*** 4.27*** 2.08** Note: Annual regression results and the IEP of Balance-Sheet Information Panel B reports the annual regression R2s of Model (10a) above those of Model (11), Rit 5a1b xit 1g Dxit 1eit ; and the incremental explanatory powers of balance-sheet information a indicates the t statistics for comparing the difference in mean R2 between Model (10a) and (11) being significant at the 0.01 level t statistics is computed with Fama–MacBeth methodology and adjusted for heteroscedasticity and autocorrelation of six lags with Newey–West approach ***,**, and * indicate Vuong’s Z statistics for comparing balance-sheet information integrated model (Model [10a]) with earnings-only model (Model [11]) being significant at the 01, 05, and levels, respectively Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 284 Journal of Accounting, Auditing & Finance IEP; R-squares 0.250 R-squares of earnings-only model IEP of balance-sheet related variables 0.200 0.150 0.100 0.050 0.000 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 2007 year Figure IEP of balance-sheet information This figure plots the IEP of balance-sheet information, computed as the R2 of Model (10a) minus that of Model (11), relative to the R2 of Model (11) Note: IEP = incremental explanatory power average R2 of Model (10a) is significantly higher than that of Model (11) at the 01 level Figure plots the IEP of the balance-sheet variables in Model (10a) against the explanatory power of the earnings variables in Model (11) across the years Rit 5a1b xit 1g Dxit 1h Dqit 1hH H Dqit 1u Dbit 1d DbitÀ1 1eit ; ð10aÞ Rit 5a1b xit 1g Dxit 1eit : ð11Þ To summarize, our empirical results show that the balance sheet–related variables (Dqit, Dbi,t, and Dbit21) generally have significant effects, and they enhance the power to explain stock returns after controlling for earnings and the earnings change The directions of these effects are generally consistent with the theoretical predictions, and their magnitudes are economically important Overall, in both statistical and economic terms, the balance-sheet information improves the explanatory power of the return model relative to that of the earnings only–based benchmark model Complementarity Between Balance-Sheet and Income-Statement Information In this section, we explore how the incremental usefulness of balance sheet–related variables varies over time and in cross sections With the balance sheet and the income statement reporting complementary data about a firm’s operations, we conjecture that the information from the two statements is also complementary in explaining stock returns We Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Huang 285 Table Relationship Between the IEP of Balance-Sheet Information and the Explanatory Power of Earnings Over Time R2 (earnings)t Intercept 0.04a (5.31) 0.06a (7.39) 2.11c (21.90) 2.19a (23.76) Timet Adjusted R2 063 20.08a (24.16) 344 Note: IEP = incremental explanatory power This table provides the result of time-series regressions of the IEP of balance-sheet information on the explanatory power of earnings The specification is as follows: IEP(BS)t 5a0 1a1 R2 (earnings)t 1a2 Timet 1ut where IEP(BS)t is the IEP of balance-sheet information, calculated as the R2 of Model (10a) in year t minus that of Model (11) in year t, R2(earnings)t is the R2 of Model (11), and Timet is time index computed as year minus 1968 For ease of exposition, we multiply the coefficient on Timet by 100 a and c indicate the coefficient being significantly different from at the 01, 05, and levels, respectively therefore explore whether balance sheet–based information is incrementally more useful in situations in which earnings variables are less informative about returns Time-Series Analysis We first take a time-series perspective to examine how the usefulness of balance-sheet information in explaining returns is related to that of earnings information Over time, the explanatory power of earnings variables, denoted as R2(earnings), fluctuates, as does the IEP of information constructed with balance sheet, denoted as IEP(BS) We find a significantly negative correlation between R2(earnings) and IEP(BS), with a Pearson correlation of 2.29 (t = 1.90), which suggests that balance-sheet information complements earnings variables to a greater extent in years in which the latter are less powerful in explaining returns Previous evidence suggests that there may be a time trend in the power of financial statement information to explain returns (see, for example, Collins et al., 1997) To control for a possible time trend, we also run a regression of IEP(BS) on R2(earnings) and a time index (Time = 0, ,39) as follows:17 IEP(BS)t 5a0 1a1 R2 (earnings)t 1a2 Timet 1ut : ð12Þ As reported in Table 5, IEP(BS) is negatively related to R2(earnings), both with and without a time trend The coefficient on R2(earnings) is 20.11 (t = 1.90) without a time index and is 20.19 (t = 3.76) with a time index This provides evidence that balance-sheet information supplements earnings variables more in years in which the latter are less informative about stock returns We also note that the coefficient on the time index is significantly negative, indicating a declining trend in the IEP of balance-sheet information in explaining cross-sectional returns.18 Cross-Sectional Analysis We now turn to the complementarity issue in cross sections Based on both economic intuition and previous findings, we identify three subsamples of firms with earnings that are Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 286 Journal of Accounting, Auditing & Finance considered to be less informative or with future earnings that are more difficult to forecast They are firms with negative (vs positive) earnings, young (vs mature) firms, and firms with a high (vs low) degree of uncertainty about future earnings We examine whether balance-sheet information is incrementally more useful in these subsamples Loss versus profit firms Rational economic behavior implies that a firm’s losses will not be sustained (a loss-making firm will either have to improve its performance or face termination) This suggests that negative earnings are less informative about future cash flows than are positive earnings (Hayn, 1995) Collins, Pincus, and Xie (1999) find that equity book value becomes more important in explaining the stock prices of firms with negative (vs positive) earnings In our context, we posit that balance sheet–based information plays a greater role in explaining stock returns for loss firms than for profit firms Panel A of Table compares the IEP of balance-sheet information between firms with negative and positive earnings.19 The average R2 across the years of Model (11), in which only earnings variables are used, is 4.6% for loss firms and 13.9% for profit firms After adding the balance sheet–related variables, we obtain an average IEP of 4.3% for loss firms In contrast, the IEP of balance-sheet information is 2.1% for profit firms.20 The difference in IEP between the two groups is 2.2% (t = 3.46), significant at the 01 level Thus, balance-sheet information is incrementally more useful in explaining returns for firms with negative earnings than for firms with positive earnings We note that for loss firms, the coefficient on xit is significantly negative and lower than 1, whether or not we include the balance sheet–related variables This might be an indication of investors’ belief that losses will be mean reverting However, for profit firms, the coefficient on xit is significantly positive, with a magnitude significantly greater than 1, indicating that for firms making a profit investors actually place a weight on earnings that is greater than the theoretical value on overall sample as predicted by Ohlson (1995) and Zhang (2000) For loss firms, adding Dqit in regressions removes the effect of Dxit, although the coefficient on Dqit is significantly positive, suggesting that the effect of Dxit is subsumed by that of Dqit in this subsample However, for profit firms, although adding Dqit substantially reduces the effect of Dxit, the latter remains significantly positive together with the coefficient on Dqit Contemporaneous capital investment has a positive coefficient in both firm groups The lagged capital investment is significantly negative, as predicted by Ohlson (1995), for loss firms, but is insignificant for profit firms Young versus mature firms For the purpose of our analysis, young firms refer to firms with a relatively short history of public trading, which are usually at early stages of the life cycle We conjecture that earnings variables are less valuation relevant for young firms and so the balance sheet should play a greater incremental role to supplement earnings information We define a firm in a given year as a ‘‘young’’ firm if it has a listing history of years or less and as a ‘‘mature’’ firm otherwise, and we then divide the annual samples each into subsets of ‘‘young’’ and ‘‘mature’’ firms.21 The results in Panel B of Table show that balance-sheet information explains more of the variations in stock returns for younger firms Across the years, the average IEP of balance-sheet information for young firms is 5.2%, compared with the IEP of 2.3% for mature firms The difference in IEP between the two groups is 3% (t = 2.21), significant at the 05 level For mature firms, adding Dqit in regressions reduces the effect of Dxit (coefficient on Dxit reduces from 0.84 to 0.33 with t statistics of 3.08), whereas for young firms the effect of Dqit on Dxit is even substantial (coefficient on Dxit reduces from 2.02 to 0.99 with t Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 287 Downloaded from jaf.sagepub.com at Taylor's University on December 2, 2012 Intercept xit Dxit 20.02 (20.56) 20.01 (20.42) 0.00 (20.06) 0.02 (0.48) 2.23a (10.84) 2.04a (8.95) 20.83a (23.27) 20.63b (22.24) 1.41a (13.46) 0.70a (7.02) 0.54a (6.53) 0.04 (0.53) 0.07b (2.14) 0.06b (2.03) 0.13b (2.34) 0.13b (2.54) 1.33a (11.76) 1.37a (8.74) 0.48 (0.58) 0.34 (0.41) 0.84a (9.27) 0.33a (3.08) 2.02a (2.71) 0.99c (1.90) 0.44a (2.85) 0.18a (3.10) 0.34a (7.22) 0.47a (7.98) Dqit 0.47b (2.00) 0.10c (1.68) 0.12 (1.60) 0.24a (3.20) Dbit High forecast error firm Medium forecast error firm Low forecast error firm 0.10a (3.27) 0.09a (2.94) 0.10a (2.90) 0.09a (2.93) 0.08a (3.05) 0.08a (3.43) 0.98b (2.59) 0.88b (2.43) 1.32a (5.24) 1.25a (4.33) 1.14a (10.37) 1.10a (6.19) 0.75b (2.48) 0.52b (2.56) 0.84a (4.74) 0.32a (2.62) 0.81a (6.16) 20.01 (20.10) 0.34a (8.33) 0.03 (0.29) 0.04 (0.53) 0.23b (2.06) 0.32a (3.53) 0.26a (3.91) Panel C: IEP of balance-sheet information for low, medium, and high forecast error firms Young firm Mature firm Panel B: IEP of balance-sheet information for young and mature firms Loss firm Profit firm Panel A: IEP of balance-sheet information for loss and profit firms Partition Table IEP of Balance-Sheet Information in Cross Sections 20.23a (23.73) 20.06 (21.14) 20.01 (20.12) 20.16 (20.92) 20.12b (22.43) 20.20a (23.40) 0.05 (1.00) Dbit21 0.34a (4.75) 0.43a (4.76) 20.01 (20.13) 0.42a (5.12) 0.48a (8.42) 0.18 (0.36) 0.07 (1.03) H Dqit 138 098 107 074 081 056 201 149 142 118 089 046 160 139 Average R2 0.040a (6.67) 0.033a (5.02) 0.025a (6.19) 0.052a (3.35) 0.023a (4.67) 0.043a (6.59) 0.021a (7.70) IEP (continued) 0.015b (2.35) 0.030b (2.21) 0.022a (3.46) Difference in IEP